linux/fs/bcachefs/fs-io-buffered.c

1076 lines
26 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
#ifndef NO_BCACHEFS_FS
#include "bcachefs.h"
#include "alloc_foreground.h"
#include "bkey_buf.h"
#include "fs-io.h"
#include "fs-io-buffered.h"
#include "fs-io-direct.h"
#include "fs-io-pagecache.h"
#include "io_read.h"
#include "io_write.h"
#include <linux/backing-dev.h>
#include <linux/pagemap.h>
#include <linux/writeback.h>
static inline bool bio_full(struct bio *bio, unsigned len)
{
if (bio->bi_vcnt >= bio->bi_max_vecs)
return true;
if (bio->bi_iter.bi_size > UINT_MAX - len)
return true;
return false;
}
/* readpage(s): */
static void bch2_readpages_end_io(struct bio *bio)
{
struct folio_iter fi;
bio_for_each_folio_all(fi, bio)
folio_end_read(fi.folio, bio->bi_status == BLK_STS_OK);
bio_put(bio);
}
struct readpages_iter {
struct address_space *mapping;
unsigned idx;
folios folios;
};
static int readpages_iter_init(struct readpages_iter *iter,
struct readahead_control *ractl)
{
struct folio *folio;
*iter = (struct readpages_iter) { ractl->mapping };
while ((folio = __readahead_folio(ractl))) {
if (!bch2_folio_create(folio, GFP_KERNEL) ||
darray_push(&iter->folios, folio)) {
bch2_folio_release(folio);
ractl->_nr_pages += folio_nr_pages(folio);
ractl->_index -= folio_nr_pages(folio);
return iter->folios.nr ? 0 : -ENOMEM;
}
folio_put(folio);
}
return 0;
}
static inline struct folio *readpage_iter_peek(struct readpages_iter *iter)
{
if (iter->idx >= iter->folios.nr)
return NULL;
return iter->folios.data[iter->idx];
}
static inline void readpage_iter_advance(struct readpages_iter *iter)
{
iter->idx++;
}
static bool extent_partial_reads_expensive(struct bkey_s_c k)
{
struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
struct bch_extent_crc_unpacked crc;
const union bch_extent_entry *i;
bkey_for_each_crc(k.k, ptrs, crc, i)
if (crc.csum_type || crc.compression_type)
return true;
return false;
}
static int readpage_bio_extend(struct btree_trans *trans,
struct readpages_iter *iter,
struct bio *bio,
unsigned sectors_this_extent,
bool get_more)
{
/* Don't hold btree locks while allocating memory: */
bch2_trans_unlock(trans);
while (bio_sectors(bio) < sectors_this_extent &&
bio->bi_vcnt < bio->bi_max_vecs) {
struct folio *folio = readpage_iter_peek(iter);
int ret;
if (folio) {
readpage_iter_advance(iter);
} else {
pgoff_t folio_offset = bio_end_sector(bio) >> PAGE_SECTORS_SHIFT;
if (!get_more)
break;
folio = xa_load(&iter->mapping->i_pages, folio_offset);
if (folio && !xa_is_value(folio))
break;
folio = filemap_alloc_folio(readahead_gfp_mask(iter->mapping), 0);
if (!folio)
break;
if (!__bch2_folio_create(folio, GFP_KERNEL)) {
folio_put(folio);
break;
}
ret = filemap_add_folio(iter->mapping, folio, folio_offset, GFP_KERNEL);
if (ret) {
__bch2_folio_release(folio);
folio_put(folio);
break;
}
folio_put(folio);
}
BUG_ON(folio_sector(folio) != bio_end_sector(bio));
BUG_ON(!bio_add_folio(bio, folio, folio_size(folio), 0));
}
return bch2_trans_relock(trans);
}
static void bchfs_read(struct btree_trans *trans,
struct bch_read_bio *rbio,
subvol_inum inum,
struct readpages_iter *readpages_iter)
{
struct bch_fs *c = trans->c;
struct btree_iter iter;
struct bkey_buf sk;
int flags = BCH_READ_RETRY_IF_STALE|
BCH_READ_MAY_PROMOTE;
int ret = 0;
rbio->c = c;
rbio->start_time = local_clock();
rbio->subvol = inum.subvol;
bch2_bkey_buf_init(&sk);
bch2_trans_begin(trans);
bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
POS(inum.inum, rbio->bio.bi_iter.bi_sector),
BTREE_ITER_slots);
while (1) {
struct bkey_s_c k;
unsigned bytes, sectors, offset_into_extent;
enum btree_id data_btree = BTREE_ID_extents;
bch2_trans_begin(trans);
u32 snapshot;
ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
if (ret)
goto err;
bch2_btree_iter_set_snapshot(&iter, snapshot);
bch2_btree_iter_set_pos(&iter,
POS(inum.inum, rbio->bio.bi_iter.bi_sector));
k = bch2_btree_iter_peek_slot(&iter);
ret = bkey_err(k);
if (ret)
goto err;
offset_into_extent = iter.pos.offset -
bkey_start_offset(k.k);
sectors = k.k->size - offset_into_extent;
bch2_bkey_buf_reassemble(&sk, c, k);
ret = bch2_read_indirect_extent(trans, &data_btree,
&offset_into_extent, &sk);
if (ret)
goto err;
k = bkey_i_to_s_c(sk.k);
sectors = min(sectors, k.k->size - offset_into_extent);
if (readpages_iter) {
ret = readpage_bio_extend(trans, readpages_iter, &rbio->bio, sectors,
extent_partial_reads_expensive(k));
if (ret)
goto err;
}
bytes = min(sectors, bio_sectors(&rbio->bio)) << 9;
swap(rbio->bio.bi_iter.bi_size, bytes);
if (rbio->bio.bi_iter.bi_size == bytes)
flags |= BCH_READ_LAST_FRAGMENT;
bch2_bio_page_state_set(&rbio->bio, k);
bch2_read_extent(trans, rbio, iter.pos,
data_btree, k, offset_into_extent, flags);
if (flags & BCH_READ_LAST_FRAGMENT)
break;
swap(rbio->bio.bi_iter.bi_size, bytes);
bio_advance(&rbio->bio, bytes);
err:
if (ret &&
!bch2_err_matches(ret, BCH_ERR_transaction_restart))
break;
}
bch2_trans_iter_exit(trans, &iter);
if (ret) {
bch_err_inum_offset_ratelimited(c,
iter.pos.inode,
iter.pos.offset << 9,
"read error %i from btree lookup", ret);
rbio->bio.bi_status = BLK_STS_IOERR;
bio_endio(&rbio->bio);
}
bch2_bkey_buf_exit(&sk, c);
}
void bch2_readahead(struct readahead_control *ractl)
{
struct bch_inode_info *inode = to_bch_ei(ractl->mapping->host);
struct bch_fs *c = inode->v.i_sb->s_fs_info;
struct bch_io_opts opts;
struct folio *folio;
struct readpages_iter readpages_iter;
bch2_inode_opts_get(&opts, c, &inode->ei_inode);
int ret = readpages_iter_init(&readpages_iter, ractl);
if (ret)
return;
bch2_pagecache_add_get(inode);
struct btree_trans *trans = bch2_trans_get(c);
while ((folio = readpage_iter_peek(&readpages_iter))) {
unsigned n = min_t(unsigned,
readpages_iter.folios.nr -
readpages_iter.idx,
BIO_MAX_VECS);
struct bch_read_bio *rbio =
rbio_init(bio_alloc_bioset(NULL, n, REQ_OP_READ,
GFP_KERNEL, &c->bio_read),
opts);
readpage_iter_advance(&readpages_iter);
rbio->bio.bi_iter.bi_sector = folio_sector(folio);
rbio->bio.bi_end_io = bch2_readpages_end_io;
BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
bchfs_read(trans, rbio, inode_inum(inode),
&readpages_iter);
bch2_trans_unlock(trans);
}
bch2_trans_put(trans);
bch2_pagecache_add_put(inode);
darray_exit(&readpages_iter.folios);
}
static void bch2_read_single_folio_end_io(struct bio *bio)
{
complete(bio->bi_private);
}
int bch2_read_single_folio(struct folio *folio, struct address_space *mapping)
{
struct bch_inode_info *inode = to_bch_ei(mapping->host);
struct bch_fs *c = inode->v.i_sb->s_fs_info;
struct bch_read_bio *rbio;
struct bch_io_opts opts;
int ret;
DECLARE_COMPLETION_ONSTACK(done);
if (!bch2_folio_create(folio, GFP_KERNEL))
return -ENOMEM;
bch2_inode_opts_get(&opts, c, &inode->ei_inode);
rbio = rbio_init(bio_alloc_bioset(NULL, 1, REQ_OP_READ, GFP_KERNEL, &c->bio_read),
opts);
rbio->bio.bi_private = &done;
rbio->bio.bi_end_io = bch2_read_single_folio_end_io;
rbio->bio.bi_opf = REQ_OP_READ|REQ_SYNC;
rbio->bio.bi_iter.bi_sector = folio_sector(folio);
BUG_ON(!bio_add_folio(&rbio->bio, folio, folio_size(folio), 0));
bch2_trans_run(c, (bchfs_read(trans, rbio, inode_inum(inode), NULL), 0));
wait_for_completion(&done);
ret = blk_status_to_errno(rbio->bio.bi_status);
bio_put(&rbio->bio);
if (ret < 0)
return ret;
folio_mark_uptodate(folio);
return 0;
}
int bch2_read_folio(struct file *file, struct folio *folio)
{
int ret;
ret = bch2_read_single_folio(folio, folio->mapping);
folio_unlock(folio);
return bch2_err_class(ret);
}
/* writepages: */
struct bch_writepage_io {
struct bch_inode_info *inode;
/* must be last: */
struct bch_write_op op;
};
struct bch_writepage_state {
struct bch_writepage_io *io;
struct bch_io_opts opts;
struct bch_folio_sector *tmp;
unsigned tmp_sectors;
};
static inline struct bch_writepage_state bch_writepage_state_init(struct bch_fs *c,
struct bch_inode_info *inode)
{
struct bch_writepage_state ret = { 0 };
bch2_inode_opts_get(&ret.opts, c, &inode->ei_inode);
return ret;
}
/*
* Determine when a writepage io is full. We have to limit writepage bios to a
* single page per bvec (i.e. 1MB with 4k pages) because that is the limit to
* what the bounce path in bch2_write_extent() can handle. In theory we could
* loosen this restriction for non-bounce I/O, but we don't have that context
* here. Ideally, we can up this limit and make it configurable in the future
* when the bounce path can be enhanced to accommodate larger source bios.
*/
static inline bool bch_io_full(struct bch_writepage_io *io, unsigned len)
{
struct bio *bio = &io->op.wbio.bio;
return bio_full(bio, len) ||
(bio->bi_iter.bi_size + len > BIO_MAX_VECS * PAGE_SIZE);
}
static void bch2_writepage_io_done(struct bch_write_op *op)
{
struct bch_writepage_io *io =
container_of(op, struct bch_writepage_io, op);
struct bch_fs *c = io->op.c;
struct bio *bio = &io->op.wbio.bio;
struct folio_iter fi;
unsigned i;
if (io->op.error) {
set_bit(EI_INODE_ERROR, &io->inode->ei_flags);
bio_for_each_folio_all(fi, bio) {
struct bch_folio *s;
mapping_set_error(fi.folio->mapping, -EIO);
s = __bch2_folio(fi.folio);
spin_lock(&s->lock);
for (i = 0; i < folio_sectors(fi.folio); i++)
s->s[i].nr_replicas = 0;
spin_unlock(&s->lock);
}
}
if (io->op.flags & BCH_WRITE_WROTE_DATA_INLINE) {
bio_for_each_folio_all(fi, bio) {
struct bch_folio *s;
s = __bch2_folio(fi.folio);
spin_lock(&s->lock);
for (i = 0; i < folio_sectors(fi.folio); i++)
s->s[i].nr_replicas = 0;
spin_unlock(&s->lock);
}
}
/*
* racing with fallocate can cause us to add fewer sectors than
* expected - but we shouldn't add more sectors than expected:
*/
WARN_ON_ONCE(io->op.i_sectors_delta > 0);
/*
* (error (due to going RO) halfway through a page can screw that up
* slightly)
* XXX wtf?
BUG_ON(io->op.op.i_sectors_delta >= PAGE_SECTORS);
*/
/*
* The writeback flag is effectively our ref on the inode -
* fixup i_blocks before calling folio_end_writeback:
*/
bch2_i_sectors_acct(c, io->inode, NULL, io->op.i_sectors_delta);
bio_for_each_folio_all(fi, bio) {
struct bch_folio *s = __bch2_folio(fi.folio);
if (atomic_dec_and_test(&s->write_count))
folio_end_writeback(fi.folio);
}
bio_put(&io->op.wbio.bio);
}
static void bch2_writepage_do_io(struct bch_writepage_state *w)
{
struct bch_writepage_io *io = w->io;
w->io = NULL;
closure_call(&io->op.cl, bch2_write, NULL, NULL);
}
/*
* Get a bch_writepage_io and add @page to it - appending to an existing one if
* possible, else allocating a new one:
*/
static void bch2_writepage_io_alloc(struct bch_fs *c,
struct writeback_control *wbc,
struct bch_writepage_state *w,
struct bch_inode_info *inode,
u64 sector,
unsigned nr_replicas)
{
struct bch_write_op *op;
w->io = container_of(bio_alloc_bioset(NULL, BIO_MAX_VECS,
REQ_OP_WRITE,
GFP_KERNEL,
&c->writepage_bioset),
struct bch_writepage_io, op.wbio.bio);
w->io->inode = inode;
op = &w->io->op;
bch2_write_op_init(op, c, w->opts);
op->target = w->opts.foreground_target;
op->nr_replicas = nr_replicas;
op->res.nr_replicas = nr_replicas;
op->write_point = writepoint_hashed(inode->ei_last_dirtied);
op->subvol = inode->ei_inum.subvol;
op->pos = POS(inode->v.i_ino, sector);
op->end_io = bch2_writepage_io_done;
op->devs_need_flush = &inode->ei_devs_need_flush;
op->wbio.bio.bi_iter.bi_sector = sector;
op->wbio.bio.bi_opf = wbc_to_write_flags(wbc);
}
static int __bch2_writepage(struct folio *folio,
struct writeback_control *wbc,
void *data)
{
struct bch_inode_info *inode = to_bch_ei(folio->mapping->host);
struct bch_fs *c = inode->v.i_sb->s_fs_info;
struct bch_writepage_state *w = data;
struct bch_folio *s;
unsigned i, offset, f_sectors, nr_replicas_this_write = U32_MAX;
loff_t i_size = i_size_read(&inode->v);
int ret;
EBUG_ON(!folio_test_uptodate(folio));
/* Is the folio fully inside i_size? */
if (folio_end_pos(folio) <= i_size)
goto do_io;
/* Is the folio fully outside i_size? (truncate in progress) */
if (folio_pos(folio) >= i_size) {
folio_unlock(folio);
return 0;
}
/*
* The folio straddles i_size. It must be zeroed out on each and every
* writepage invocation because it may be mmapped. "A file is mapped
* in multiples of the folio size. For a file that is not a multiple of
* the folio size, the remaining memory is zeroed when mapped, and
* writes to that region are not written out to the file."
*/
folio_zero_segment(folio,
i_size - folio_pos(folio),
folio_size(folio));
do_io:
f_sectors = folio_sectors(folio);
s = bch2_folio(folio);
if (f_sectors > w->tmp_sectors) {
kfree(w->tmp);
bcachefs: Fix incorrect gfp flags fixes: 00488 WARNING: CPU: 9 PID: 194 at mm/page_alloc.c:4410 __alloc_pages_noprof+0x1818/0x1888 00488 Modules linked in: 00488 CPU: 9 UID: 0 PID: 194 Comm: kworker/u66:1 Not tainted 6.11.0-rc1-ktest-g18fa10d6495f #2931 00488 Hardware name: linux,dummy-virt (DT) 00488 Workqueue: writeback wb_workfn (flush-bcachefs-2) 00488 pstate: 20001005 (nzCv daif -PAN -UAO -TCO -DIT +SSBS BTYPE=--) 00488 pc : __alloc_pages_noprof+0x1818/0x1888 00488 lr : __alloc_pages_noprof+0x5f4/0x1888 00488 sp : ffffff80ccd8ed00 00488 x29: ffffff80ccd8ed00 x28: 0000000000000000 x27: dfffffc000000000 00488 x26: 0000000000000010 x25: 0000000000000002 x24: 0000000000000000 00488 x23: 0000000000000000 x22: 1ffffff0199b1dbe x21: ffffff80cc680900 00488 x20: 0000000000000000 x19: ffffff80ccd8eed0 x18: 0000000000000000 00488 x17: ffffff80cc58a010 x16: dfffffc000000000 x15: 1ffffff00474e518 00488 x14: 1ffffff00474e518 x13: 1ffffff00474e518 x12: ffffffb8104701b9 00488 x11: 1ffffff8104701b8 x10: ffffffb8104701b8 x9 : ffffffc08043cde8 00488 x8 : 00000047efb8fe48 x7 : ffffff80ccd8ee20 x6 : 0000000000048000 00488 x5 : 1ffffff810470138 x4 : 0000000000000050 x3 : 1ffffff0199b1d94 00488 x2 : ffffffb0199b1d94 x1 : 0000000000000001 x0 : ffffffc082387448 00488 Call trace: 00488 __alloc_pages_noprof+0x1818/0x1888 00488 new_slab+0x284/0x2f0 00488 ___slab_alloc+0x208/0x8e0 00488 __kmalloc_noprof+0x328/0x340 00488 __bch2_writepage+0x106c/0x1830 00488 write_cache_pages+0xa0/0xe8 due to __GFP_NOFAIL without allowing reclaim Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-08-17 21:38:43 +00:00
w->tmp = kcalloc(f_sectors, sizeof(struct bch_folio_sector), GFP_NOFS|__GFP_NOFAIL);
w->tmp_sectors = f_sectors;
}
/*
* Things get really hairy with errors during writeback:
*/
ret = bch2_get_folio_disk_reservation(c, inode, folio, false);
BUG_ON(ret);
/* Before unlocking the page, get copy of reservations: */
spin_lock(&s->lock);
memcpy(w->tmp, s->s, sizeof(struct bch_folio_sector) * f_sectors);
for (i = 0; i < f_sectors; i++) {
if (s->s[i].state < SECTOR_dirty)
continue;
nr_replicas_this_write =
min_t(unsigned, nr_replicas_this_write,
s->s[i].nr_replicas +
s->s[i].replicas_reserved);
}
for (i = 0; i < f_sectors; i++) {
if (s->s[i].state < SECTOR_dirty)
continue;
s->s[i].nr_replicas = w->opts.compression
? 0 : nr_replicas_this_write;
s->s[i].replicas_reserved = 0;
bch2_folio_sector_set(folio, s, i, SECTOR_allocated);
}
spin_unlock(&s->lock);
BUG_ON(atomic_read(&s->write_count));
atomic_set(&s->write_count, 1);
BUG_ON(folio_test_writeback(folio));
folio_start_writeback(folio);
folio_unlock(folio);
offset = 0;
while (1) {
unsigned sectors = 0, dirty_sectors = 0, reserved_sectors = 0;
u64 sector;
while (offset < f_sectors &&
w->tmp[offset].state < SECTOR_dirty)
offset++;
if (offset == f_sectors)
break;
while (offset + sectors < f_sectors &&
w->tmp[offset + sectors].state >= SECTOR_dirty) {
reserved_sectors += w->tmp[offset + sectors].replicas_reserved;
dirty_sectors += w->tmp[offset + sectors].state == SECTOR_dirty;
sectors++;
}
BUG_ON(!sectors);
sector = folio_sector(folio) + offset;
if (w->io &&
(w->io->op.res.nr_replicas != nr_replicas_this_write ||
bch_io_full(w->io, sectors << 9) ||
bio_end_sector(&w->io->op.wbio.bio) != sector))
bch2_writepage_do_io(w);
if (!w->io)
bch2_writepage_io_alloc(c, wbc, w, inode, sector,
nr_replicas_this_write);
atomic_inc(&s->write_count);
BUG_ON(inode != w->io->inode);
BUG_ON(!bio_add_folio(&w->io->op.wbio.bio, folio,
sectors << 9, offset << 9));
/* Check for writing past i_size: */
WARN_ONCE((bio_end_sector(&w->io->op.wbio.bio) << 9) >
round_up(i_size, block_bytes(c)) &&
!test_bit(BCH_FS_emergency_ro, &c->flags),
"writing past i_size: %llu > %llu (unrounded %llu)\n",
bio_end_sector(&w->io->op.wbio.bio) << 9,
round_up(i_size, block_bytes(c)),
i_size);
w->io->op.res.sectors += reserved_sectors;
w->io->op.i_sectors_delta -= dirty_sectors;
w->io->op.new_i_size = i_size;
offset += sectors;
}
if (atomic_dec_and_test(&s->write_count))
folio_end_writeback(folio);
return 0;
}
int bch2_writepages(struct address_space *mapping, struct writeback_control *wbc)
{
struct bch_fs *c = mapping->host->i_sb->s_fs_info;
struct bch_writepage_state w =
bch_writepage_state_init(c, to_bch_ei(mapping->host));
struct blk_plug plug;
int ret;
blk_start_plug(&plug);
ret = write_cache_pages(mapping, wbc, __bch2_writepage, &w);
if (w.io)
bch2_writepage_do_io(&w);
blk_finish_plug(&plug);
kfree(w.tmp);
return bch2_err_class(ret);
}
/* buffered writes: */
int bch2_write_begin(struct file *file, struct address_space *mapping,
loff_t pos, unsigned len,
struct folio **foliop, void **fsdata)
{
struct bch_inode_info *inode = to_bch_ei(mapping->host);
struct bch_fs *c = inode->v.i_sb->s_fs_info;
struct bch2_folio_reservation *res;
struct folio *folio;
unsigned offset;
int ret = -ENOMEM;
res = kmalloc(sizeof(*res), GFP_KERNEL);
if (!res)
return -ENOMEM;
bch2_folio_reservation_init(c, inode, res);
*fsdata = res;
bch2_pagecache_add_get(inode);
folio = __filemap_get_folio(mapping, pos >> PAGE_SHIFT,
FGP_WRITEBEGIN | fgf_set_order(len),
mapping_gfp_mask(mapping));
if (IS_ERR_OR_NULL(folio))
goto err_unlock;
offset = pos - folio_pos(folio);
len = min_t(size_t, len, folio_end_pos(folio) - pos);
if (folio_test_uptodate(folio))
goto out;
/* If we're writing entire folio, don't need to read it in first: */
if (!offset && len == folio_size(folio))
goto out;
if (!offset && pos + len >= inode->v.i_size) {
folio_zero_segment(folio, len, folio_size(folio));
flush_dcache_folio(folio);
goto out;
}
if (folio_pos(folio) >= inode->v.i_size) {
folio_zero_segments(folio, 0, offset, offset + len, folio_size(folio));
flush_dcache_folio(folio);
goto out;
}
readpage:
ret = bch2_read_single_folio(folio, mapping);
if (ret)
goto err;
out:
ret = bch2_folio_set(c, inode_inum(inode), &folio, 1);
if (ret)
goto err;
ret = bch2_folio_reservation_get(c, inode, folio, res, offset, len);
if (ret) {
if (!folio_test_uptodate(folio)) {
/*
* If the folio hasn't been read in, we won't know if we
* actually need a reservation - we don't actually need
* to read here, we just need to check if the folio is
* fully backed by uncompressed data:
*/
goto readpage;
}
goto err;
}
*foliop = folio;
return 0;
err:
folio_unlock(folio);
folio_put(folio);
err_unlock:
bch2_pagecache_add_put(inode);
kfree(res);
*fsdata = NULL;
return bch2_err_class(ret);
}
int bch2_write_end(struct file *file, struct address_space *mapping,
loff_t pos, unsigned len, unsigned copied,
struct folio *folio, void *fsdata)
{
struct bch_inode_info *inode = to_bch_ei(mapping->host);
struct bch_fs *c = inode->v.i_sb->s_fs_info;
struct bch2_folio_reservation *res = fsdata;
unsigned offset = pos - folio_pos(folio);
lockdep_assert_held(&inode->v.i_rwsem);
BUG_ON(offset + copied > folio_size(folio));
if (unlikely(copied < len && !folio_test_uptodate(folio))) {
/*
* The folio needs to be read in, but that would destroy
* our partial write - simplest thing is to just force
* userspace to redo the write:
*/
folio_zero_range(folio, 0, folio_size(folio));
flush_dcache_folio(folio);
copied = 0;
}
spin_lock(&inode->v.i_lock);
if (pos + copied > inode->v.i_size)
i_size_write(&inode->v, pos + copied);
spin_unlock(&inode->v.i_lock);
if (copied) {
if (!folio_test_uptodate(folio))
folio_mark_uptodate(folio);
bch2_set_folio_dirty(c, inode, folio, res, offset, copied);
inode->ei_last_dirtied = (unsigned long) current;
}
folio_unlock(folio);
folio_put(folio);
bch2_pagecache_add_put(inode);
bch2_folio_reservation_put(c, inode, res);
kfree(res);
return copied;
}
static noinline void folios_trunc(folios *fs, struct folio **fi)
{
while (fs->data + fs->nr > fi) {
struct folio *f = darray_pop(fs);
folio_unlock(f);
folio_put(f);
}
}
static int __bch2_buffered_write(struct bch_inode_info *inode,
struct address_space *mapping,
struct iov_iter *iter,
bcachefs: Revert lockless buffered IO path We had a report of data corruption on nixos when building installer images. https://github.com/NixOS/nixpkgs/pull/321055#issuecomment-2184131334 It seems that writes are being dropped, but only when issued by QEMU, and possibly only in snapshot mode. It's undetermined if it's write calls are being dropped or dirty folios. Further testing, via minimizing the original patch to just the change that skips the inode lock on non appends/truncates, reveals that it really is just not taking the inode lock that causes the corruption: it has nothing to do with the other logic changes for preserving write atomicity in corner cases. It's also kernel config dependent: it doesn't reproduce with the minimal kernel config that ktest uses, but it does reproduce with nixos's distro config. Bisection the kernel config initially pointer the finger at page migration or compaction, but it appears that was erroneous; we haven't yet determined what kernel config option actually triggers it. Sadly it appears this will have to be reverted since we're getting too close to release and my plate is full, but we'd _really_ like to fully debug it. My suspicion is that this patch is exposing a preexisting bug - the inode lock actually covers very little in IO paths, and we have a different lock (the pagecache add lock) that guards against races with truncate here. Fixes: 7e64c86cdc6c ("bcachefs: Buffered write path now can avoid the inode lock") Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-08-31 21:44:51 +00:00
loff_t pos, unsigned len)
{
struct bch_fs *c = inode->v.i_sb->s_fs_info;
struct bch2_folio_reservation res;
folios fs;
struct folio *f;
unsigned copied = 0, f_offset, f_copied;
u64 end = pos + len, f_pos, f_len;
loff_t last_folio_pos = inode->v.i_size;
int ret = 0;
BUG_ON(!len);
bch2_folio_reservation_init(c, inode, &res);
darray_init(&fs);
ret = bch2_filemap_get_contig_folios_d(mapping, pos, end,
FGP_WRITEBEGIN | fgf_set_order(len),
mapping_gfp_mask(mapping), &fs);
if (ret)
goto out;
BUG_ON(!fs.nr);
f = darray_first(fs);
if (pos != folio_pos(f) && !folio_test_uptodate(f)) {
ret = bch2_read_single_folio(f, mapping);
if (ret)
goto out;
}
f = darray_last(fs);
end = min(end, folio_end_pos(f));
last_folio_pos = folio_pos(f);
if (end != folio_end_pos(f) && !folio_test_uptodate(f)) {
if (end >= inode->v.i_size) {
folio_zero_range(f, 0, folio_size(f));
} else {
ret = bch2_read_single_folio(f, mapping);
if (ret)
goto out;
}
}
ret = bch2_folio_set(c, inode_inum(inode), fs.data, fs.nr);
if (ret)
goto out;
f_pos = pos;
f_offset = pos - folio_pos(darray_first(fs));
darray_for_each(fs, fi) {
ssize_t f_reserved;
f = *fi;
f_len = min(end, folio_end_pos(f)) - f_pos;
f_reserved = bch2_folio_reservation_get_partial(c, inode, f, &res, f_offset, f_len);
if (unlikely(f_reserved != f_len)) {
if (f_reserved < 0) {
if (f == darray_first(fs)) {
ret = f_reserved;
goto out;
}
folios_trunc(&fs, fi);
end = min(end, folio_end_pos(darray_last(fs)));
} else {
folios_trunc(&fs, fi + 1);
end = f_pos + f_reserved;
}
break;
}
f_pos = folio_end_pos(f);
f_offset = 0;
}
if (mapping_writably_mapped(mapping))
darray_for_each(fs, fi)
flush_dcache_folio(*fi);
f_pos = pos;
f_offset = pos - folio_pos(darray_first(fs));
darray_for_each(fs, fi) {
f = *fi;
f_len = min(end, folio_end_pos(f)) - f_pos;
f_copied = copy_folio_from_iter_atomic(f, f_offset, f_len, iter);
if (!f_copied) {
folios_trunc(&fs, fi);
break;
}
if (!folio_test_uptodate(f) &&
f_copied != folio_size(f) &&
pos + copied + f_copied < inode->v.i_size) {
iov_iter_revert(iter, f_copied);
folio_zero_range(f, 0, folio_size(f));
folios_trunc(&fs, fi);
break;
}
flush_dcache_folio(f);
copied += f_copied;
if (f_copied != f_len) {
folios_trunc(&fs, fi + 1);
break;
}
f_pos = folio_end_pos(f);
f_offset = 0;
}
if (!copied)
goto out;
end = pos + copied;
spin_lock(&inode->v.i_lock);
bcachefs: Revert lockless buffered IO path We had a report of data corruption on nixos when building installer images. https://github.com/NixOS/nixpkgs/pull/321055#issuecomment-2184131334 It seems that writes are being dropped, but only when issued by QEMU, and possibly only in snapshot mode. It's undetermined if it's write calls are being dropped or dirty folios. Further testing, via minimizing the original patch to just the change that skips the inode lock on non appends/truncates, reveals that it really is just not taking the inode lock that causes the corruption: it has nothing to do with the other logic changes for preserving write atomicity in corner cases. It's also kernel config dependent: it doesn't reproduce with the minimal kernel config that ktest uses, but it does reproduce with nixos's distro config. Bisection the kernel config initially pointer the finger at page migration or compaction, but it appears that was erroneous; we haven't yet determined what kernel config option actually triggers it. Sadly it appears this will have to be reverted since we're getting too close to release and my plate is full, but we'd _really_ like to fully debug it. My suspicion is that this patch is exposing a preexisting bug - the inode lock actually covers very little in IO paths, and we have a different lock (the pagecache add lock) that guards against races with truncate here. Fixes: 7e64c86cdc6c ("bcachefs: Buffered write path now can avoid the inode lock") Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-08-31 21:44:51 +00:00
if (end > inode->v.i_size)
i_size_write(&inode->v, end);
spin_unlock(&inode->v.i_lock);
f_pos = pos;
f_offset = pos - folio_pos(darray_first(fs));
darray_for_each(fs, fi) {
f = *fi;
f_len = min(end, folio_end_pos(f)) - f_pos;
if (!folio_test_uptodate(f))
folio_mark_uptodate(f);
bch2_set_folio_dirty(c, inode, f, &res, f_offset, f_len);
f_pos = folio_end_pos(f);
f_offset = 0;
}
inode->ei_last_dirtied = (unsigned long) current;
out:
darray_for_each(fs, fi) {
folio_unlock(*fi);
folio_put(*fi);
}
/*
* If the last folio added to the mapping starts beyond current EOF, we
* performed a short write but left around at least one post-EOF folio.
* Clean up the mapping before we return.
*/
if (last_folio_pos >= inode->v.i_size)
truncate_pagecache(&inode->v, inode->v.i_size);
darray_exit(&fs);
bch2_folio_reservation_put(c, inode, &res);
return copied ?: ret;
}
static ssize_t bch2_buffered_write(struct kiocb *iocb, struct iov_iter *iter)
{
struct file *file = iocb->ki_filp;
struct address_space *mapping = file->f_mapping;
struct bch_inode_info *inode = file_bch_inode(file);
bcachefs: Revert lockless buffered IO path We had a report of data corruption on nixos when building installer images. https://github.com/NixOS/nixpkgs/pull/321055#issuecomment-2184131334 It seems that writes are being dropped, but only when issued by QEMU, and possibly only in snapshot mode. It's undetermined if it's write calls are being dropped or dirty folios. Further testing, via minimizing the original patch to just the change that skips the inode lock on non appends/truncates, reveals that it really is just not taking the inode lock that causes the corruption: it has nothing to do with the other logic changes for preserving write atomicity in corner cases. It's also kernel config dependent: it doesn't reproduce with the minimal kernel config that ktest uses, but it does reproduce with nixos's distro config. Bisection the kernel config initially pointer the finger at page migration or compaction, but it appears that was erroneous; we haven't yet determined what kernel config option actually triggers it. Sadly it appears this will have to be reverted since we're getting too close to release and my plate is full, but we'd _really_ like to fully debug it. My suspicion is that this patch is exposing a preexisting bug - the inode lock actually covers very little in IO paths, and we have a different lock (the pagecache add lock) that guards against races with truncate here. Fixes: 7e64c86cdc6c ("bcachefs: Buffered write path now can avoid the inode lock") Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-08-31 21:44:51 +00:00
loff_t pos = iocb->ki_pos;
ssize_t written = 0;
int ret = 0;
bch2_pagecache_add_get(inode);
do {
unsigned offset = pos & (PAGE_SIZE - 1);
unsigned bytes = iov_iter_count(iter);
again:
/*
* Bring in the user page that we will copy from _first_.
* Otherwise there's a nasty deadlock on copying from the
* same page as we're writing to, without it being marked
* up-to-date.
*
* Not only is this an optimisation, but it is also required
* to check that the address is actually valid, when atomic
* usercopies are used, below.
*/
if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
bytes = min_t(unsigned long, iov_iter_count(iter),
PAGE_SIZE - offset);
if (unlikely(fault_in_iov_iter_readable(iter, bytes))) {
ret = -EFAULT;
break;
}
}
if (unlikely(fatal_signal_pending(current))) {
ret = -EINTR;
break;
}
bcachefs: Revert lockless buffered IO path We had a report of data corruption on nixos when building installer images. https://github.com/NixOS/nixpkgs/pull/321055#issuecomment-2184131334 It seems that writes are being dropped, but only when issued by QEMU, and possibly only in snapshot mode. It's undetermined if it's write calls are being dropped or dirty folios. Further testing, via minimizing the original patch to just the change that skips the inode lock on non appends/truncates, reveals that it really is just not taking the inode lock that causes the corruption: it has nothing to do with the other logic changes for preserving write atomicity in corner cases. It's also kernel config dependent: it doesn't reproduce with the minimal kernel config that ktest uses, but it does reproduce with nixos's distro config. Bisection the kernel config initially pointer the finger at page migration or compaction, but it appears that was erroneous; we haven't yet determined what kernel config option actually triggers it. Sadly it appears this will have to be reverted since we're getting too close to release and my plate is full, but we'd _really_ like to fully debug it. My suspicion is that this patch is exposing a preexisting bug - the inode lock actually covers very little in IO paths, and we have a different lock (the pagecache add lock) that guards against races with truncate here. Fixes: 7e64c86cdc6c ("bcachefs: Buffered write path now can avoid the inode lock") Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-08-31 21:44:51 +00:00
ret = __bch2_buffered_write(inode, mapping, iter, pos, bytes);
if (unlikely(ret < 0))
break;
cond_resched();
if (unlikely(ret == 0)) {
/*
* If we were unable to copy any data at all, we must
* fall back to a single segment length write.
*
* If we didn't fallback here, we could livelock
* because not all segments in the iov can be copied at
* once without a pagefault.
*/
bytes = min_t(unsigned long, PAGE_SIZE - offset,
iov_iter_single_seg_count(iter));
goto again;
}
pos += ret;
written += ret;
ret = 0;
balance_dirty_pages_ratelimited(mapping);
} while (iov_iter_count(iter));
bcachefs: Revert lockless buffered IO path We had a report of data corruption on nixos when building installer images. https://github.com/NixOS/nixpkgs/pull/321055#issuecomment-2184131334 It seems that writes are being dropped, but only when issued by QEMU, and possibly only in snapshot mode. It's undetermined if it's write calls are being dropped or dirty folios. Further testing, via minimizing the original patch to just the change that skips the inode lock on non appends/truncates, reveals that it really is just not taking the inode lock that causes the corruption: it has nothing to do with the other logic changes for preserving write atomicity in corner cases. It's also kernel config dependent: it doesn't reproduce with the minimal kernel config that ktest uses, but it does reproduce with nixos's distro config. Bisection the kernel config initially pointer the finger at page migration or compaction, but it appears that was erroneous; we haven't yet determined what kernel config option actually triggers it. Sadly it appears this will have to be reverted since we're getting too close to release and my plate is full, but we'd _really_ like to fully debug it. My suspicion is that this patch is exposing a preexisting bug - the inode lock actually covers very little in IO paths, and we have a different lock (the pagecache add lock) that guards against races with truncate here. Fixes: 7e64c86cdc6c ("bcachefs: Buffered write path now can avoid the inode lock") Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-08-31 21:44:51 +00:00
bch2_pagecache_add_put(inode);
bcachefs: Revert lockless buffered IO path We had a report of data corruption on nixos when building installer images. https://github.com/NixOS/nixpkgs/pull/321055#issuecomment-2184131334 It seems that writes are being dropped, but only when issued by QEMU, and possibly only in snapshot mode. It's undetermined if it's write calls are being dropped or dirty folios. Further testing, via minimizing the original patch to just the change that skips the inode lock on non appends/truncates, reveals that it really is just not taking the inode lock that causes the corruption: it has nothing to do with the other logic changes for preserving write atomicity in corner cases. It's also kernel config dependent: it doesn't reproduce with the minimal kernel config that ktest uses, but it does reproduce with nixos's distro config. Bisection the kernel config initially pointer the finger at page migration or compaction, but it appears that was erroneous; we haven't yet determined what kernel config option actually triggers it. Sadly it appears this will have to be reverted since we're getting too close to release and my plate is full, but we'd _really_ like to fully debug it. My suspicion is that this patch is exposing a preexisting bug - the inode lock actually covers very little in IO paths, and we have a different lock (the pagecache add lock) that guards against races with truncate here. Fixes: 7e64c86cdc6c ("bcachefs: Buffered write path now can avoid the inode lock") Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-08-31 21:44:51 +00:00
return written ? written : ret;
}
bcachefs: Revert lockless buffered IO path We had a report of data corruption on nixos when building installer images. https://github.com/NixOS/nixpkgs/pull/321055#issuecomment-2184131334 It seems that writes are being dropped, but only when issued by QEMU, and possibly only in snapshot mode. It's undetermined if it's write calls are being dropped or dirty folios. Further testing, via minimizing the original patch to just the change that skips the inode lock on non appends/truncates, reveals that it really is just not taking the inode lock that causes the corruption: it has nothing to do with the other logic changes for preserving write atomicity in corner cases. It's also kernel config dependent: it doesn't reproduce with the minimal kernel config that ktest uses, but it does reproduce with nixos's distro config. Bisection the kernel config initially pointer the finger at page migration or compaction, but it appears that was erroneous; we haven't yet determined what kernel config option actually triggers it. Sadly it appears this will have to be reverted since we're getting too close to release and my plate is full, but we'd _really_ like to fully debug it. My suspicion is that this patch is exposing a preexisting bug - the inode lock actually covers very little in IO paths, and we have a different lock (the pagecache add lock) that guards against races with truncate here. Fixes: 7e64c86cdc6c ("bcachefs: Buffered write path now can avoid the inode lock") Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-08-31 21:44:51 +00:00
ssize_t bch2_write_iter(struct kiocb *iocb, struct iov_iter *from)
{
bcachefs: Revert lockless buffered IO path We had a report of data corruption on nixos when building installer images. https://github.com/NixOS/nixpkgs/pull/321055#issuecomment-2184131334 It seems that writes are being dropped, but only when issued by QEMU, and possibly only in snapshot mode. It's undetermined if it's write calls are being dropped or dirty folios. Further testing, via minimizing the original patch to just the change that skips the inode lock on non appends/truncates, reveals that it really is just not taking the inode lock that causes the corruption: it has nothing to do with the other logic changes for preserving write atomicity in corner cases. It's also kernel config dependent: it doesn't reproduce with the minimal kernel config that ktest uses, but it does reproduce with nixos's distro config. Bisection the kernel config initially pointer the finger at page migration or compaction, but it appears that was erroneous; we haven't yet determined what kernel config option actually triggers it. Sadly it appears this will have to be reverted since we're getting too close to release and my plate is full, but we'd _really_ like to fully debug it. My suspicion is that this patch is exposing a preexisting bug - the inode lock actually covers very little in IO paths, and we have a different lock (the pagecache add lock) that guards against races with truncate here. Fixes: 7e64c86cdc6c ("bcachefs: Buffered write path now can avoid the inode lock") Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-08-31 21:44:51 +00:00
struct file *file = iocb->ki_filp;
struct bch_inode_info *inode = file_bch_inode(file);
ssize_t ret;
if (iocb->ki_flags & IOCB_DIRECT) {
ret = bch2_direct_write(iocb, from);
goto out;
}
inode_lock(&inode->v);
ret = generic_write_checks(iocb, from);
if (ret <= 0)
goto unlock;
ret = file_remove_privs(file);
if (ret)
goto unlock;
ret = file_update_time(file);
if (ret)
goto unlock;
ret = bch2_buffered_write(iocb, from);
if (likely(ret > 0))
iocb->ki_pos += ret;
unlock:
inode_unlock(&inode->v);
bcachefs: Revert lockless buffered IO path We had a report of data corruption on nixos when building installer images. https://github.com/NixOS/nixpkgs/pull/321055#issuecomment-2184131334 It seems that writes are being dropped, but only when issued by QEMU, and possibly only in snapshot mode. It's undetermined if it's write calls are being dropped or dirty folios. Further testing, via minimizing the original patch to just the change that skips the inode lock on non appends/truncates, reveals that it really is just not taking the inode lock that causes the corruption: it has nothing to do with the other logic changes for preserving write atomicity in corner cases. It's also kernel config dependent: it doesn't reproduce with the minimal kernel config that ktest uses, but it does reproduce with nixos's distro config. Bisection the kernel config initially pointer the finger at page migration or compaction, but it appears that was erroneous; we haven't yet determined what kernel config option actually triggers it. Sadly it appears this will have to be reverted since we're getting too close to release and my plate is full, but we'd _really_ like to fully debug it. My suspicion is that this patch is exposing a preexisting bug - the inode lock actually covers very little in IO paths, and we have a different lock (the pagecache add lock) that guards against races with truncate here. Fixes: 7e64c86cdc6c ("bcachefs: Buffered write path now can avoid the inode lock") Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
2024-08-31 21:44:51 +00:00
if (ret > 0)
ret = generic_write_sync(iocb, ret);
out:
return bch2_err_class(ret);
}
void bch2_fs_fs_io_buffered_exit(struct bch_fs *c)
{
bioset_exit(&c->writepage_bioset);
}
int bch2_fs_fs_io_buffered_init(struct bch_fs *c)
{
if (bioset_init(&c->writepage_bioset,
4, offsetof(struct bch_writepage_io, op.wbio.bio),
BIOSET_NEED_BVECS))
return -BCH_ERR_ENOMEM_writepage_bioset_init;
return 0;
}
#endif /* NO_BCACHEFS_FS */