2013-08-14 19:05:12 +00:00
|
|
|
/*
|
|
|
|
* Copyright (C) 2013 Fusion IO. All rights reserved.
|
|
|
|
*
|
|
|
|
* This program is free software; you can redistribute it and/or
|
|
|
|
* modify it under the terms of the GNU General Public
|
|
|
|
* License v2 as published by the Free Software Foundation.
|
|
|
|
*
|
|
|
|
* This program is distributed in the hope that it will be useful,
|
|
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
|
* General Public License for more details.
|
|
|
|
*
|
|
|
|
* You should have received a copy of the GNU General Public
|
|
|
|
* License along with this program; if not, write to the
|
|
|
|
* Free Software Foundation, Inc., 59 Temple Place - Suite 330,
|
|
|
|
* Boston, MA 021110-1307, USA.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/slab.h>
|
|
|
|
#include "btrfs-tests.h"
|
|
|
|
#include "../ctree.h"
|
2015-09-23 18:54:14 +00:00
|
|
|
#include "../disk-io.h"
|
2013-08-14 19:05:12 +00:00
|
|
|
#include "../free-space-cache.h"
|
|
|
|
|
2016-06-01 11:18:24 +00:00
|
|
|
#define BITS_PER_BITMAP (PAGE_SIZE * 8UL)
|
2013-08-14 19:05:12 +00:00
|
|
|
|
|
|
|
/*
|
2016-05-20 01:18:45 +00:00
|
|
|
* This test just does basic sanity checking, making sure we can add an extent
|
2013-08-14 19:05:12 +00:00
|
|
|
* entry and remove space from either end and the middle, and make sure we can
|
|
|
|
* remove space that covers adjacent extent entries.
|
|
|
|
*/
|
|
|
|
static int test_extents(struct btrfs_block_group_cache *cache)
|
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
|
|
|
|
test_msg("Running extent only tests\n");
|
|
|
|
|
|
|
|
/* First just make sure we can remove an entire entry */
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_add_free_space(cache, 0, SZ_4M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error adding initial extents %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, 0, SZ_4M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error removing extent %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, 0, SZ_4M)) {
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Full remove left some lingering space\n");
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Ok edge and middle cases now */
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_add_free_space(cache, 0, SZ_4M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error adding half extent %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, 3 * SZ_1M, SZ_1M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error removing tail end %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, 0, SZ_1M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error removing front end %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, SZ_2M, 4096);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
2013-12-08 15:22:53 +00:00
|
|
|
test_msg("Error removing middle piece %d\n", ret);
|
2013-08-14 19:05:12 +00:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, 0, SZ_1M)) {
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Still have space at the front\n");
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, SZ_2M, 4096)) {
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Still have space in the middle\n");
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, 3 * SZ_1M, SZ_1M)) {
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Still have space at the end\n");
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Cleanup */
|
|
|
|
__btrfs_remove_free_space_cache(cache->free_space_ctl);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2016-06-01 11:18:25 +00:00
|
|
|
static int test_bitmaps(struct btrfs_block_group_cache *cache,
|
|
|
|
u32 sectorsize)
|
2013-08-14 19:05:12 +00:00
|
|
|
{
|
|
|
|
u64 next_bitmap_offset;
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
test_msg("Running bitmap only tests\n");
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, 0, SZ_4M, 1);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't create a bitmap entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, 0, SZ_4M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error removing bitmap full range %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, 0, SZ_4M)) {
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Left some space in bitmap\n");
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, 0, SZ_4M, 1);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add to our bitmap entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, SZ_1M, SZ_2M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't remove middle chunk %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The first bitmap we have starts at offset 0 so the next one is just
|
|
|
|
* at the end of the first bitmap.
|
|
|
|
*/
|
2016-06-01 11:18:25 +00:00
|
|
|
next_bitmap_offset = (u64)(BITS_PER_BITMAP * sectorsize);
|
2013-08-14 19:05:12 +00:00
|
|
|
|
|
|
|
/* Test a bit straddling two bitmaps */
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, next_bitmap_offset - SZ_2M,
|
|
|
|
SZ_4M, 1);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add space that straddles two bitmaps %d\n",
|
|
|
|
ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, next_bitmap_offset - SZ_1M, SZ_2M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't remove overlapping space %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, next_bitmap_offset - SZ_1M, SZ_2M)) {
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Left some space when removing overlapping\n");
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
__btrfs_remove_free_space_cache(cache->free_space_ctl);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* This is the high grade jackassery */
|
2016-06-01 11:18:25 +00:00
|
|
|
static int test_bitmaps_and_extents(struct btrfs_block_group_cache *cache,
|
|
|
|
u32 sectorsize)
|
2013-08-14 19:05:12 +00:00
|
|
|
{
|
2016-06-01 11:18:25 +00:00
|
|
|
u64 bitmap_offset = (u64)(BITS_PER_BITMAP * sectorsize);
|
2013-08-14 19:05:12 +00:00
|
|
|
int ret;
|
|
|
|
|
|
|
|
test_msg("Running bitmap and extent tests\n");
|
|
|
|
|
|
|
|
/*
|
|
|
|
* First let's do something simple, an extent at the same offset as the
|
|
|
|
* bitmap, but the free space completely in the extent and then
|
|
|
|
* completely in the bitmap.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, SZ_4M, SZ_1M, 1);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't create bitmap entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, 0, SZ_1M, 0);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add extent entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, 0, SZ_1M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't remove extent entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, 0, SZ_1M)) {
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Left remnants after our remove\n");
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Now to add back the extent entry and remove from the bitmap */
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, 0, SZ_1M, 0);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't re-add extent entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, SZ_4M, SZ_1M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't remove from bitmap %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, SZ_4M, SZ_1M)) {
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Left remnants in the bitmap\n");
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Ok so a little more evil, extent entry and bitmap at the same offset,
|
|
|
|
* removing an overlapping chunk.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, SZ_1M, SZ_4M, 1);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add to a bitmap %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, SZ_512K, 3 * SZ_1M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't remove overlapping space %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, SZ_512K, 3 * SZ_1M)) {
|
2014-01-07 12:58:06 +00:00
|
|
|
test_msg("Left over pieces after removing overlapping\n");
|
2013-08-14 19:05:12 +00:00
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
__btrfs_remove_free_space_cache(cache->free_space_ctl);
|
|
|
|
|
|
|
|
/* Now with the extent entry offset into the bitmap */
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, SZ_4M, SZ_4M, 1);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add space to the bitmap %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, SZ_2M, SZ_2M, 0);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add extent to the cache %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, 3 * SZ_1M, SZ_4M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Problem removing overlapping space %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, 3 * SZ_1M, SZ_4M)) {
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Left something behind when removing space");
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This has blown up in the past, the extent entry starts before the
|
|
|
|
* bitmap entry, but we're trying to remove an offset that falls
|
|
|
|
* completely within the bitmap range and is in both the extent entry
|
|
|
|
* and the bitmap entry, looks like this
|
|
|
|
*
|
|
|
|
* [ extent ]
|
|
|
|
* [ bitmap ]
|
|
|
|
* [ del ]
|
|
|
|
*/
|
|
|
|
__btrfs_remove_free_space_cache(cache->free_space_ctl);
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, bitmap_offset + SZ_4M, SZ_4M, 1);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add bitmap %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, bitmap_offset - SZ_1M,
|
|
|
|
5 * SZ_1M, 0);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add extent entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, bitmap_offset + SZ_1M, 5 * SZ_1M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Failed to free our space %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, bitmap_offset + SZ_1M, 5 * SZ_1M)) {
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Left stuff over\n");
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
__btrfs_remove_free_space_cache(cache->free_space_ctl);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This blew up before, we have part of the free space in a bitmap and
|
|
|
|
* then the entirety of the rest of the space in an extent. This used
|
|
|
|
* to return -EAGAIN back from btrfs_remove_extent, make sure this
|
|
|
|
* doesn't happen.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, SZ_1M, SZ_2M, 1);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add bitmap entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, 3 * SZ_1M, SZ_1M, 0);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add extent entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, SZ_1M, 3 * SZ_1M);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error removing bitmap and extent overlapping %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
__btrfs_remove_free_space_cache(cache->free_space_ctl);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
/* Used by test_steal_space_from_bitmap_to_extent(). */
|
|
|
|
static bool test_use_bitmap(struct btrfs_free_space_ctl *ctl,
|
|
|
|
struct btrfs_free_space *info)
|
|
|
|
{
|
|
|
|
return ctl->free_extents > 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Used by test_steal_space_from_bitmap_to_extent(). */
|
|
|
|
static int
|
|
|
|
check_num_extents_and_bitmaps(const struct btrfs_block_group_cache *cache,
|
|
|
|
const int num_extents,
|
|
|
|
const int num_bitmaps)
|
|
|
|
{
|
|
|
|
if (cache->free_space_ctl->free_extents != num_extents) {
|
|
|
|
test_msg("Incorrect # of extent entries in the cache: %d, expected %d\n",
|
|
|
|
cache->free_space_ctl->free_extents, num_extents);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
if (cache->free_space_ctl->total_bitmaps != num_bitmaps) {
|
|
|
|
test_msg("Incorrect # of extent entries in the cache: %d, expected %d\n",
|
|
|
|
cache->free_space_ctl->total_bitmaps, num_bitmaps);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Used by test_steal_space_from_bitmap_to_extent(). */
|
|
|
|
static int check_cache_empty(struct btrfs_block_group_cache *cache)
|
|
|
|
{
|
|
|
|
u64 offset;
|
|
|
|
u64 max_extent_size;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now lets confirm that there's absolutely no free space left to
|
|
|
|
* allocate.
|
|
|
|
*/
|
|
|
|
if (cache->free_space_ctl->free_space != 0) {
|
|
|
|
test_msg("Cache free space is not 0\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* And any allocation request, no matter how small, should fail now. */
|
|
|
|
offset = btrfs_find_space_for_alloc(cache, 0, 4096, 0,
|
|
|
|
&max_extent_size);
|
|
|
|
if (offset != 0) {
|
|
|
|
test_msg("Space allocation did not fail, returned offset: %llu",
|
|
|
|
offset);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* And no extent nor bitmap entries in the cache anymore. */
|
|
|
|
return check_num_extents_and_bitmaps(cache, 0, 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Before we were able to steal free space from a bitmap entry to an extent
|
|
|
|
* entry, we could end up with 2 entries representing a contiguous free space.
|
|
|
|
* One would be an extent entry and the other a bitmap entry. Since in order
|
|
|
|
* to allocate space to a caller we use only 1 entry, we couldn't return that
|
|
|
|
* whole range to the caller if it was requested. This forced the caller to
|
|
|
|
* either assume ENOSPC or perform several smaller space allocations, which
|
|
|
|
* wasn't optimal as they could be spread all over the block group while under
|
|
|
|
* concurrency (extra overhead and fragmentation).
|
|
|
|
*
|
2016-05-20 01:18:45 +00:00
|
|
|
* This stealing approach is beneficial, since we always prefer to allocate
|
|
|
|
* from extent entries, both for clustered and non-clustered allocation
|
|
|
|
* requests.
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
*/
|
|
|
|
static int
|
2016-06-01 11:18:25 +00:00
|
|
|
test_steal_space_from_bitmap_to_extent(struct btrfs_block_group_cache *cache,
|
|
|
|
u32 sectorsize)
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
u64 offset;
|
|
|
|
u64 max_extent_size;
|
2015-11-19 10:42:28 +00:00
|
|
|
const struct btrfs_free_space_op test_free_space_ops = {
|
2015-11-19 10:42:24 +00:00
|
|
|
.recalc_thresholds = cache->free_space_ctl->op->recalc_thresholds,
|
|
|
|
.use_bitmap = test_use_bitmap,
|
|
|
|
};
|
2015-11-19 10:42:28 +00:00
|
|
|
const struct btrfs_free_space_op *orig_free_space_ops;
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
|
|
|
|
test_msg("Running space stealing from bitmap to extent\n");
|
|
|
|
|
|
|
|
/*
|
|
|
|
* For this test, we want to ensure we end up with an extent entry
|
|
|
|
* immediately adjacent to a bitmap entry, where the bitmap starts
|
|
|
|
* at an offset where the extent entry ends. We keep adding and
|
|
|
|
* removing free space to reach into this state, but to get there
|
|
|
|
* we need to reach a point where marking new free space doesn't
|
|
|
|
* result in adding new extent entries or merging the new space
|
|
|
|
* with existing extent entries - the space ends up being marked
|
|
|
|
* in an existing bitmap that covers the new free space range.
|
|
|
|
*
|
|
|
|
* To get there, we need to reach the threshold defined set at
|
|
|
|
* cache->free_space_ctl->extents_thresh, which currently is
|
|
|
|
* 256 extents on a x86_64 system at least, and a few other
|
|
|
|
* conditions (check free_space_cache.c). Instead of making the
|
|
|
|
* test much longer and complicated, use a "use_bitmap" operation
|
|
|
|
* that forces use of bitmaps as soon as we have at least 1
|
|
|
|
* extent entry.
|
|
|
|
*/
|
2015-11-19 10:42:24 +00:00
|
|
|
orig_free_space_ops = cache->free_space_ctl->op;
|
|
|
|
cache->free_space_ctl->op = &test_free_space_ops;
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Extent entry covering free space range [128Mb - 256Kb, 128Mb - 128Kb[
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, SZ_128M - SZ_256K, SZ_128K, 0);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add extent entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Bitmap entry covering free space range [128Mb + 512Kb, 256Mb[ */
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, SZ_128M + SZ_512K,
|
|
|
|
SZ_128M - SZ_512K, 1);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add bitmap entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
ret = check_num_extents_and_bitmaps(cache, 2, 1);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now make only the first 256Kb of the bitmap marked as free, so that
|
|
|
|
* we end up with only the following ranges marked as free space:
|
|
|
|
*
|
|
|
|
* [128Mb - 256Kb, 128Mb - 128Kb[
|
|
|
|
* [128Mb + 512Kb, 128Mb + 768Kb[
|
|
|
|
*/
|
|
|
|
ret = btrfs_remove_free_space(cache,
|
2015-12-14 16:42:10 +00:00
|
|
|
SZ_128M + 768 * SZ_1K,
|
|
|
|
SZ_128M - 768 * SZ_1K);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Failed to free part of bitmap space %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Confirm that only those 2 ranges are marked as free. */
|
2015-12-14 16:42:10 +00:00
|
|
|
if (!test_check_exists(cache, SZ_128M - SZ_256K, SZ_128K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Free space range missing\n");
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
2015-12-14 16:42:10 +00:00
|
|
|
if (!test_check_exists(cache, SZ_128M + SZ_512K, SZ_256K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Free space range missing\n");
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Confirm that the bitmap range [128Mb + 768Kb, 256Mb[ isn't marked
|
|
|
|
* as free anymore.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, SZ_128M + 768 * SZ_1K,
|
|
|
|
SZ_128M - 768 * SZ_1K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Bitmap region not removed from space cache\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Confirm that the region [128Mb + 256Kb, 128Mb + 512Kb[, which is
|
|
|
|
* covered by the bitmap, isn't marked as free.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, SZ_128M + SZ_256K, SZ_256K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Invalid bitmap region marked as free\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Confirm that the region [128Mb, 128Mb + 256Kb[, which is covered
|
|
|
|
* by the bitmap too, isn't marked as free either.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, SZ_128M, SZ_256K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Invalid bitmap region marked as free\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now lets mark the region [128Mb, 128Mb + 512Kb[ as free too. But,
|
|
|
|
* lets make sure the free space cache marks it as free in the bitmap,
|
|
|
|
* and doesn't insert a new extent entry to represent this region.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_add_free_space(cache, SZ_128M, SZ_512K);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error adding free space: %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
/* Confirm the region is marked as free. */
|
2015-12-14 16:42:10 +00:00
|
|
|
if (!test_check_exists(cache, SZ_128M, SZ_512K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Bitmap region not marked as free\n");
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Confirm that no new extent entries or bitmap entries were added to
|
|
|
|
* the cache after adding that free space region.
|
|
|
|
*/
|
|
|
|
ret = check_num_extents_and_bitmaps(cache, 2, 1);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now lets add a small free space region to the right of the previous
|
|
|
|
* one, which is not contiguous with it and is part of the bitmap too.
|
|
|
|
* The goal is to test that the bitmap entry space stealing doesn't
|
|
|
|
* steal this space region.
|
|
|
|
*/
|
2016-06-01 11:18:25 +00:00
|
|
|
ret = btrfs_add_free_space(cache, SZ_128M + SZ_16M, sectorsize);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error adding free space: %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Confirm that no new extent entries or bitmap entries were added to
|
|
|
|
* the cache after adding that free space region.
|
|
|
|
*/
|
|
|
|
ret = check_num_extents_and_bitmaps(cache, 2, 1);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now mark the region [128Mb - 128Kb, 128Mb[ as free too. This will
|
|
|
|
* expand the range covered by the existing extent entry that represents
|
|
|
|
* the free space [128Mb - 256Kb, 128Mb - 128Kb[.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_add_free_space(cache, SZ_128M - SZ_128K, SZ_128K);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error adding free space: %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
/* Confirm the region is marked as free. */
|
2015-12-14 16:42:10 +00:00
|
|
|
if (!test_check_exists(cache, SZ_128M - SZ_128K, SZ_128K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Extent region not marked as free\n");
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Confirm that our extent entry didn't stole all free space from the
|
|
|
|
* bitmap, because of the small 4Kb free space region.
|
|
|
|
*/
|
|
|
|
ret = check_num_extents_and_bitmaps(cache, 2, 1);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* So now we have the range [128Mb - 256Kb, 128Mb + 768Kb[ as free
|
|
|
|
* space. Without stealing bitmap free space into extent entry space,
|
|
|
|
* we would have all this free space represented by 2 entries in the
|
|
|
|
* cache:
|
|
|
|
*
|
|
|
|
* extent entry covering range: [128Mb - 256Kb, 128Mb[
|
|
|
|
* bitmap entry covering range: [128Mb, 128Mb + 768Kb[
|
|
|
|
*
|
|
|
|
* Attempting to allocate the whole free space (1Mb) would fail, because
|
|
|
|
* we can't allocate from multiple entries.
|
|
|
|
* With the bitmap free space stealing, we get a single extent entry
|
|
|
|
* that represents the 1Mb free space, and therefore we're able to
|
|
|
|
* allocate the whole free space at once.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
if (!test_check_exists(cache, SZ_128M - SZ_256K, SZ_1M)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Expected region not marked as free\n");
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
|
|
|
|
2016-06-01 11:18:25 +00:00
|
|
|
if (cache->free_space_ctl->free_space != (SZ_1M + sectorsize)) {
|
|
|
|
test_msg("Cache free space is not 1Mb + %u\n", sectorsize);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
offset = btrfs_find_space_for_alloc(cache,
|
2015-12-14 16:42:10 +00:00
|
|
|
0, SZ_1M, 0,
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
&max_extent_size);
|
2015-12-14 16:42:10 +00:00
|
|
|
if (offset != (SZ_128M - SZ_256K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Failed to allocate 1Mb from space cache, returned offset is: %llu\n",
|
|
|
|
offset);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
2016-06-01 11:18:25 +00:00
|
|
|
/*
|
|
|
|
* All that remains is a sectorsize free space region in a bitmap.
|
|
|
|
* Confirm.
|
|
|
|
*/
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
ret = check_num_extents_and_bitmaps(cache, 1, 1);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
2016-06-01 11:18:25 +00:00
|
|
|
if (cache->free_space_ctl->free_space != sectorsize) {
|
|
|
|
test_msg("Cache free space is not %u\n", sectorsize);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
offset = btrfs_find_space_for_alloc(cache,
|
2016-06-01 11:18:25 +00:00
|
|
|
0, sectorsize, 0,
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
&max_extent_size);
|
2015-12-14 16:42:10 +00:00
|
|
|
if (offset != (SZ_128M + SZ_16M)) {
|
2016-06-01 11:18:25 +00:00
|
|
|
test_msg("Failed to allocate %u, returned offset : %llu\n",
|
|
|
|
sectorsize, offset);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
ret = check_cache_empty(cache);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
__btrfs_remove_free_space_cache(cache->free_space_ctl);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now test a similar scenario, but where our extent entry is located
|
|
|
|
* to the right of the bitmap entry, so that we can check that stealing
|
|
|
|
* space from a bitmap to the front of an extent entry works.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Extent entry covering free space range [128Mb + 128Kb, 128Mb + 256Kb[
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, SZ_128M + SZ_128K, SZ_128K, 0);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add extent entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Bitmap entry covering free space range [0, 128Mb - 512Kb[ */
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = test_add_free_space_entry(cache, 0, SZ_128M - SZ_512K, 1);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Couldn't add bitmap entry %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
ret = check_num_extents_and_bitmaps(cache, 2, 1);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now make only the last 256Kb of the bitmap marked as free, so that
|
|
|
|
* we end up with only the following ranges marked as free space:
|
|
|
|
*
|
|
|
|
* [128Mb + 128b, 128Mb + 256Kb[
|
|
|
|
* [128Mb - 768Kb, 128Mb - 512Kb[
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_remove_free_space(cache, 0, SZ_128M - 768 * SZ_1K);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Failed to free part of bitmap space %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Confirm that only those 2 ranges are marked as free. */
|
2015-12-14 16:42:10 +00:00
|
|
|
if (!test_check_exists(cache, SZ_128M + SZ_128K, SZ_128K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Free space range missing\n");
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
2015-12-14 16:42:10 +00:00
|
|
|
if (!test_check_exists(cache, SZ_128M - 768 * SZ_1K, SZ_256K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Free space range missing\n");
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Confirm that the bitmap range [0, 128Mb - 768Kb[ isn't marked
|
|
|
|
* as free anymore.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, 0, SZ_128M - 768 * SZ_1K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Bitmap region not removed from space cache\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Confirm that the region [128Mb - 512Kb, 128Mb[, which is
|
|
|
|
* covered by the bitmap, isn't marked as free.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
if (test_check_exists(cache, SZ_128M - SZ_512K, SZ_512K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Invalid bitmap region marked as free\n");
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now lets mark the region [128Mb - 512Kb, 128Mb[ as free too. But,
|
|
|
|
* lets make sure the free space cache marks it as free in the bitmap,
|
|
|
|
* and doesn't insert a new extent entry to represent this region.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_add_free_space(cache, SZ_128M - SZ_512K, SZ_512K);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error adding free space: %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
/* Confirm the region is marked as free. */
|
2015-12-14 16:42:10 +00:00
|
|
|
if (!test_check_exists(cache, SZ_128M - SZ_512K, SZ_512K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Bitmap region not marked as free\n");
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Confirm that no new extent entries or bitmap entries were added to
|
|
|
|
* the cache after adding that free space region.
|
|
|
|
*/
|
|
|
|
ret = check_num_extents_and_bitmaps(cache, 2, 1);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now lets add a small free space region to the left of the previous
|
|
|
|
* one, which is not contiguous with it and is part of the bitmap too.
|
|
|
|
* The goal is to test that the bitmap entry space stealing doesn't
|
|
|
|
* steal this space region.
|
|
|
|
*/
|
2016-06-01 11:18:25 +00:00
|
|
|
ret = btrfs_add_free_space(cache, SZ_32M, 2 * sectorsize);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error adding free space: %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Now mark the region [128Mb, 128Mb + 128Kb[ as free too. This will
|
|
|
|
* expand the range covered by the existing extent entry that represents
|
|
|
|
* the free space [128Mb + 128Kb, 128Mb + 256Kb[.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
ret = btrfs_add_free_space(cache, SZ_128M, SZ_128K);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
if (ret) {
|
|
|
|
test_msg("Error adding free space: %d\n", ret);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
/* Confirm the region is marked as free. */
|
2015-12-14 16:42:10 +00:00
|
|
|
if (!test_check_exists(cache, SZ_128M, SZ_128K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Extent region not marked as free\n");
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Confirm that our extent entry didn't stole all free space from the
|
2016-06-01 11:18:25 +00:00
|
|
|
* bitmap, because of the small 2 * sectorsize free space region.
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
*/
|
|
|
|
ret = check_num_extents_and_bitmaps(cache, 2, 1);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* So now we have the range [128Mb - 768Kb, 128Mb + 256Kb[ as free
|
|
|
|
* space. Without stealing bitmap free space into extent entry space,
|
|
|
|
* we would have all this free space represented by 2 entries in the
|
|
|
|
* cache:
|
|
|
|
*
|
|
|
|
* extent entry covering range: [128Mb, 128Mb + 256Kb[
|
|
|
|
* bitmap entry covering range: [128Mb - 768Kb, 128Mb[
|
|
|
|
*
|
|
|
|
* Attempting to allocate the whole free space (1Mb) would fail, because
|
|
|
|
* we can't allocate from multiple entries.
|
|
|
|
* With the bitmap free space stealing, we get a single extent entry
|
|
|
|
* that represents the 1Mb free space, and therefore we're able to
|
|
|
|
* allocate the whole free space at once.
|
|
|
|
*/
|
2015-12-14 16:42:10 +00:00
|
|
|
if (!test_check_exists(cache, SZ_128M - 768 * SZ_1K, SZ_1M)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Expected region not marked as free\n");
|
|
|
|
return -ENOENT;
|
|
|
|
}
|
|
|
|
|
2016-06-01 11:18:25 +00:00
|
|
|
if (cache->free_space_ctl->free_space != (SZ_1M + 2 * sectorsize)) {
|
|
|
|
test_msg("Cache free space is not 1Mb + %u\n", 2 * sectorsize);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
2015-12-14 16:42:10 +00:00
|
|
|
offset = btrfs_find_space_for_alloc(cache, 0, SZ_1M, 0,
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
&max_extent_size);
|
2015-12-14 16:42:10 +00:00
|
|
|
if (offset != (SZ_128M - 768 * SZ_1K)) {
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
test_msg("Failed to allocate 1Mb from space cache, returned offset is: %llu\n",
|
|
|
|
offset);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
2016-06-01 11:18:25 +00:00
|
|
|
/*
|
|
|
|
* All that remains is 2 * sectorsize free space region
|
|
|
|
* in a bitmap. Confirm.
|
|
|
|
*/
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
ret = check_num_extents_and_bitmaps(cache, 1, 1);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
2016-06-01 11:18:25 +00:00
|
|
|
if (cache->free_space_ctl->free_space != 2 * sectorsize) {
|
|
|
|
test_msg("Cache free space is not %u\n", 2 * sectorsize);
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
offset = btrfs_find_space_for_alloc(cache,
|
2016-06-01 11:18:25 +00:00
|
|
|
0, 2 * sectorsize, 0,
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
&max_extent_size);
|
2015-12-14 16:42:10 +00:00
|
|
|
if (offset != SZ_32M) {
|
2016-06-01 11:18:25 +00:00
|
|
|
test_msg("Failed to allocate %u, offset: %llu\n",
|
|
|
|
2 * sectorsize,
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
offset);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
ret = check_cache_empty(cache);
|
|
|
|
if (ret)
|
|
|
|
return ret;
|
|
|
|
|
2015-11-19 10:42:24 +00:00
|
|
|
cache->free_space_ctl->op = orig_free_space_ops;
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
__btrfs_remove_free_space_cache(cache->free_space_ctl);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2016-06-01 11:18:25 +00:00
|
|
|
int btrfs_test_free_space_cache(u32 sectorsize, u32 nodesize)
|
2013-08-14 19:05:12 +00:00
|
|
|
{
|
2016-06-20 18:14:09 +00:00
|
|
|
struct btrfs_fs_info *fs_info;
|
2013-08-14 19:05:12 +00:00
|
|
|
struct btrfs_block_group_cache *cache;
|
2015-09-23 18:54:14 +00:00
|
|
|
struct btrfs_root *root = NULL;
|
|
|
|
int ret = -ENOMEM;
|
2013-08-14 19:05:12 +00:00
|
|
|
|
|
|
|
test_msg("Running btrfs free space cache tests\n");
|
2016-06-15 13:22:56 +00:00
|
|
|
fs_info = btrfs_alloc_dummy_fs_info(nodesize, sectorsize);
|
|
|
|
if (!fs_info)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
2013-08-14 19:05:12 +00:00
|
|
|
|
2016-06-01 11:18:29 +00:00
|
|
|
/*
|
|
|
|
* For ppc64 (with 64k page size), bytes per bitmap might be
|
|
|
|
* larger than 1G. To make bitmap test available in ppc64,
|
|
|
|
* alloc dummy block group whose size cross bitmaps.
|
|
|
|
*/
|
2016-06-15 13:22:56 +00:00
|
|
|
cache = btrfs_alloc_dummy_block_group(fs_info,
|
|
|
|
BITS_PER_BITMAP * sectorsize + PAGE_SIZE);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (!cache) {
|
|
|
|
test_msg("Couldn't run the tests\n");
|
2016-06-15 13:22:56 +00:00
|
|
|
btrfs_free_dummy_fs_info(fs_info);
|
2013-08-14 19:05:12 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2016-06-15 13:22:56 +00:00
|
|
|
root = btrfs_alloc_dummy_root(fs_info);
|
2016-06-20 18:14:09 +00:00
|
|
|
if (IS_ERR(root)) {
|
|
|
|
ret = PTR_ERR(root);
|
2015-09-23 18:54:14 +00:00
|
|
|
goto out;
|
2016-06-20 18:14:09 +00:00
|
|
|
}
|
2015-09-23 18:54:14 +00:00
|
|
|
|
|
|
|
root->fs_info->extent_root = root;
|
|
|
|
|
2013-08-14 19:05:12 +00:00
|
|
|
ret = test_extents(cache);
|
|
|
|
if (ret)
|
|
|
|
goto out;
|
2016-06-01 11:18:25 +00:00
|
|
|
ret = test_bitmaps(cache, sectorsize);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret)
|
|
|
|
goto out;
|
2016-06-01 11:18:25 +00:00
|
|
|
ret = test_bitmaps_and_extents(cache, sectorsize);
|
2013-08-14 19:05:12 +00:00
|
|
|
if (ret)
|
|
|
|
goto out;
|
Btrfs: improve free space cache management and space allocation
While under random IO, a block group's free space cache eventually reaches
a state where it has a mix of extent entries and bitmap entries representing
free space regions.
As later free space regions are returned to the cache, some of them are merged
with existing extent entries if they are contiguous with them. But others are
not merged, because despite the existence of adjacent free space regions in
the cache, the merging doesn't happen because the existing free space regions
are represented in bitmap extents. Even when new free space regions are merged
with existing extent entries (enlarging the free space range they represent),
we create chances of having after an enlarged region that is contiguous with
some other region represented in a bitmap entry.
Both clustered and non-clustered space allocation work by iterating over our
extent and bitmap entries and skipping any that represents a region smaller
then the allocation request (and giving preference to extent entries before
bitmap entries). By having a contiguous free space region that is represented
by 2 (or more) entries (mix of extent and bitmap entries), we end up not
satisfying an allocation request with a size larger than the size of any of
the entries but no larger than the sum of their sizes. Making the caller assume
we're under a ENOSPC condition or force it to allocate multiple smaller space
regions (as we do for file data writes), which adds extra overhead and more
chances of causing fragmentation due to the smaller regions being all spread
apart from each other (more likely when under concurrency).
For example, if we have the following in the cache:
* extent entry representing free space range: [128Mb - 256Kb, 128Mb[
* bitmap entry covering the range [128Mb, 256Mb[, but only with the bits
representing the range [128Mb, 128Mb + 768Kb[ set - that is, only that
space in this 128Mb area is marked as free
An allocation request for 1Mb, starting at offset not greater than 128Mb - 256Kb,
would fail before, despite the existence of such contiguous free space area in the
cache. The caller could only allocate up to 768Kb of space at once and later another
256Kb (or vice-versa). In between each smaller allocation request, another task
working on a different file/inode might come in and take that space, preventing the
former task of getting a contiguous 1Mb region of free space.
Therefore this change implements the ability to move free space from bitmap
entries into existing and new free space regions represented with extent
entries. This is done when a space region is added to the cache.
A test was added to the sanity tests that explains in detail the issue too.
Some performance test results with compilebench on a 4 cores machine, with
32Gb of ram and using an HDD follow.
Test: compilebench -D /mnt -i 30 -r 1000 --makej
Before this change:
intial create total runs 30 avg 69.02 MB/s (user 0.28s sys 0.57s)
compile total runs 30 avg 314.96 MB/s (user 0.12s sys 0.25s)
read compiled tree total runs 3 avg 27.14 MB/s (user 1.52s sys 0.90s)
delete compiled tree total runs 30 avg 3.14 seconds (user 0.15s sys 0.66s)
After this change:
intial create total runs 30 avg 68.37 MB/s (user 0.29s sys 0.55s)
compile total runs 30 avg 382.83 MB/s (user 0.12s sys 0.24s)
read compiled tree total runs 3 avg 27.82 MB/s (user 1.45s sys 0.97s)
delete compiled tree total runs 30 avg 3.18 seconds (user 0.17s sys 0.65s)
Signed-off-by: Filipe Manana <fdmanana@suse.com>
Signed-off-by: Chris Mason <clm@fb.com>
2014-08-29 12:35:13 +00:00
|
|
|
|
2016-06-01 11:18:25 +00:00
|
|
|
ret = test_steal_space_from_bitmap_to_extent(cache, sectorsize);
|
2013-08-14 19:05:12 +00:00
|
|
|
out:
|
2015-09-30 03:50:36 +00:00
|
|
|
btrfs_free_dummy_block_group(cache);
|
2015-09-23 18:54:14 +00:00
|
|
|
btrfs_free_dummy_root(root);
|
2016-06-20 18:14:09 +00:00
|
|
|
btrfs_free_dummy_fs_info(fs_info);
|
2013-08-14 19:05:12 +00:00
|
|
|
test_msg("Free space cache tests finished\n");
|
|
|
|
return ret;
|
|
|
|
}
|