mirror of
https://github.com/torvalds/linux.git
synced 2024-11-27 06:31:52 +00:00
e4c057d02c
GRU is already using almost the same algorithm as get/put, it even helpfully has a 10 year old comment to make this algorithm common, which is finally happening. There are a few differences and fixes from this conversion: - GRU used rcu not srcu to read the hlist - Unclear how the locking worked to prevent gru_register_mmu_notifier() from running concurrently with gru_drop_mmu_notifier() - this version is safe - GRU had a release function which only set a variable without any locking that skiped the synchronize_srcu during unregister which looks racey, but this makes it reliable via the integrated call_srcu(). - It is unclear if the mmap_sem is actually held when __mmu_notifier_register() was called, lockdep will now warn if this is wrong Link: https://lore.kernel.org/r/20190806231548.25242-5-jgg@ziepe.ca Reviewed-by: Christoph Hellwig <hch@lst.de> Reviewed-by: Dimitri Sivanich <sivanich@hpe.com> Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
319 lines
9.9 KiB
C
319 lines
9.9 KiB
C
// SPDX-License-Identifier: GPL-2.0-or-later
|
|
/*
|
|
* SN Platform GRU Driver
|
|
*
|
|
* MMUOPS callbacks + TLB flushing
|
|
*
|
|
* This file handles emu notifier callbacks from the core kernel. The callbacks
|
|
* are used to update the TLB in the GRU as a result of changes in the
|
|
* state of a process address space. This file also handles TLB invalidates
|
|
* from the GRU driver.
|
|
*
|
|
* Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved.
|
|
*/
|
|
|
|
#include <linux/kernel.h>
|
|
#include <linux/list.h>
|
|
#include <linux/spinlock.h>
|
|
#include <linux/mm.h>
|
|
#include <linux/slab.h>
|
|
#include <linux/device.h>
|
|
#include <linux/hugetlb.h>
|
|
#include <linux/delay.h>
|
|
#include <linux/timex.h>
|
|
#include <linux/srcu.h>
|
|
#include <asm/processor.h>
|
|
#include "gru.h"
|
|
#include "grutables.h"
|
|
#include <asm/uv/uv_hub.h>
|
|
|
|
#define gru_random() get_cycles()
|
|
|
|
/* ---------------------------------- TLB Invalidation functions --------
|
|
* get_tgh_handle
|
|
*
|
|
* Find a TGH to use for issuing a TLB invalidate. For GRUs that are on the
|
|
* local blade, use a fixed TGH that is a function of the blade-local cpu
|
|
* number. Normally, this TGH is private to the cpu & no contention occurs for
|
|
* the TGH. For offblade GRUs, select a random TGH in the range above the
|
|
* private TGHs. A spinlock is required to access this TGH & the lock must be
|
|
* released when the invalidate is completes. This sucks, but it is the best we
|
|
* can do.
|
|
*
|
|
* Note that the spinlock is IN the TGH handle so locking does not involve
|
|
* additional cache lines.
|
|
*
|
|
*/
|
|
static inline int get_off_blade_tgh(struct gru_state *gru)
|
|
{
|
|
int n;
|
|
|
|
n = GRU_NUM_TGH - gru->gs_tgh_first_remote;
|
|
n = gru_random() % n;
|
|
n += gru->gs_tgh_first_remote;
|
|
return n;
|
|
}
|
|
|
|
static inline int get_on_blade_tgh(struct gru_state *gru)
|
|
{
|
|
return uv_blade_processor_id() >> gru->gs_tgh_local_shift;
|
|
}
|
|
|
|
static struct gru_tlb_global_handle *get_lock_tgh_handle(struct gru_state
|
|
*gru)
|
|
{
|
|
struct gru_tlb_global_handle *tgh;
|
|
int n;
|
|
|
|
preempt_disable();
|
|
if (uv_numa_blade_id() == gru->gs_blade_id)
|
|
n = get_on_blade_tgh(gru);
|
|
else
|
|
n = get_off_blade_tgh(gru);
|
|
tgh = get_tgh_by_index(gru, n);
|
|
lock_tgh_handle(tgh);
|
|
|
|
return tgh;
|
|
}
|
|
|
|
static void get_unlock_tgh_handle(struct gru_tlb_global_handle *tgh)
|
|
{
|
|
unlock_tgh_handle(tgh);
|
|
preempt_enable();
|
|
}
|
|
|
|
/*
|
|
* gru_flush_tlb_range
|
|
*
|
|
* General purpose TLB invalidation function. This function scans every GRU in
|
|
* the ENTIRE system (partition) looking for GRUs where the specified MM has
|
|
* been accessed by the GRU. For each GRU found, the TLB must be invalidated OR
|
|
* the ASID invalidated. Invalidating an ASID causes a new ASID to be assigned
|
|
* on the next fault. This effectively flushes the ENTIRE TLB for the MM at the
|
|
* cost of (possibly) a large number of future TLBmisses.
|
|
*
|
|
* The current algorithm is optimized based on the following (somewhat true)
|
|
* assumptions:
|
|
* - GRU contexts are not loaded into a GRU unless a reference is made to
|
|
* the data segment or control block (this is true, not an assumption).
|
|
* If a DS/CB is referenced, the user will also issue instructions that
|
|
* cause TLBmisses. It is not necessary to optimize for the case where
|
|
* contexts are loaded but no instructions cause TLB misses. (I know
|
|
* this will happen but I'm not optimizing for it).
|
|
* - GRU instructions to invalidate TLB entries are SLOOOOWWW - normally
|
|
* a few usec but in unusual cases, it could be longer. Avoid if
|
|
* possible.
|
|
* - intrablade process migration between cpus is not frequent but is
|
|
* common.
|
|
* - a GRU context is not typically migrated to a different GRU on the
|
|
* blade because of intrablade migration
|
|
* - interblade migration is rare. Processes migrate their GRU context to
|
|
* the new blade.
|
|
* - if interblade migration occurs, migration back to the original blade
|
|
* is very very rare (ie., no optimization for this case)
|
|
* - most GRU instruction operate on a subset of the user REGIONS. Code
|
|
* & shared library regions are not likely targets of GRU instructions.
|
|
*
|
|
* To help improve the efficiency of TLB invalidation, the GMS data
|
|
* structure is maintained for EACH address space (MM struct). The GMS is
|
|
* also the structure that contains the pointer to the mmu callout
|
|
* functions. This structure is linked to the mm_struct for the address space
|
|
* using the mmu "register" function. The mmu interfaces are used to
|
|
* provide the callbacks for TLB invalidation. The GMS contains:
|
|
*
|
|
* - asid[maxgrus] array. ASIDs are assigned to a GRU when a context is
|
|
* loaded into the GRU.
|
|
* - asidmap[maxgrus]. bitmap to make it easier to find non-zero asids in
|
|
* the above array
|
|
* - ctxbitmap[maxgrus]. Indicates the contexts that are currently active
|
|
* in the GRU for the address space. This bitmap must be passed to the
|
|
* GRU to do an invalidate.
|
|
*
|
|
* The current algorithm for invalidating TLBs is:
|
|
* - scan the asidmap for GRUs where the context has been loaded, ie,
|
|
* asid is non-zero.
|
|
* - for each gru found:
|
|
* - if the ctxtmap is non-zero, there are active contexts in the
|
|
* GRU. TLB invalidate instructions must be issued to the GRU.
|
|
* - if the ctxtmap is zero, no context is active. Set the ASID to
|
|
* zero to force a full TLB invalidation. This is fast but will
|
|
* cause a lot of TLB misses if the context is reloaded onto the
|
|
* GRU
|
|
*
|
|
*/
|
|
|
|
void gru_flush_tlb_range(struct gru_mm_struct *gms, unsigned long start,
|
|
unsigned long len)
|
|
{
|
|
struct gru_state *gru;
|
|
struct gru_mm_tracker *asids;
|
|
struct gru_tlb_global_handle *tgh;
|
|
unsigned long num;
|
|
int grupagesize, pagesize, pageshift, gid, asid;
|
|
|
|
/* ZZZ TODO - handle huge pages */
|
|
pageshift = PAGE_SHIFT;
|
|
pagesize = (1UL << pageshift);
|
|
grupagesize = GRU_PAGESIZE(pageshift);
|
|
num = min(((len + pagesize - 1) >> pageshift), GRUMAXINVAL);
|
|
|
|
STAT(flush_tlb);
|
|
gru_dbg(grudev, "gms %p, start 0x%lx, len 0x%lx, asidmap 0x%lx\n", gms,
|
|
start, len, gms->ms_asidmap[0]);
|
|
|
|
spin_lock(&gms->ms_asid_lock);
|
|
for_each_gru_in_bitmap(gid, gms->ms_asidmap) {
|
|
STAT(flush_tlb_gru);
|
|
gru = GID_TO_GRU(gid);
|
|
asids = gms->ms_asids + gid;
|
|
asid = asids->mt_asid;
|
|
if (asids->mt_ctxbitmap && asid) {
|
|
STAT(flush_tlb_gru_tgh);
|
|
asid = GRUASID(asid, start);
|
|
gru_dbg(grudev,
|
|
" FLUSH gruid %d, asid 0x%x, vaddr 0x%lx, vamask 0x%x, num %ld, cbmap 0x%x\n",
|
|
gid, asid, start, grupagesize, num, asids->mt_ctxbitmap);
|
|
tgh = get_lock_tgh_handle(gru);
|
|
tgh_invalidate(tgh, start, ~0, asid, grupagesize, 0,
|
|
num - 1, asids->mt_ctxbitmap);
|
|
get_unlock_tgh_handle(tgh);
|
|
} else {
|
|
STAT(flush_tlb_gru_zero_asid);
|
|
asids->mt_asid = 0;
|
|
__clear_bit(gru->gs_gid, gms->ms_asidmap);
|
|
gru_dbg(grudev,
|
|
" CLEARASID gruid %d, asid 0x%x, cbtmap 0x%x, asidmap 0x%lx\n",
|
|
gid, asid, asids->mt_ctxbitmap,
|
|
gms->ms_asidmap[0]);
|
|
}
|
|
}
|
|
spin_unlock(&gms->ms_asid_lock);
|
|
}
|
|
|
|
/*
|
|
* Flush the entire TLB on a chiplet.
|
|
*/
|
|
void gru_flush_all_tlb(struct gru_state *gru)
|
|
{
|
|
struct gru_tlb_global_handle *tgh;
|
|
|
|
gru_dbg(grudev, "gid %d\n", gru->gs_gid);
|
|
tgh = get_lock_tgh_handle(gru);
|
|
tgh_invalidate(tgh, 0, ~0, 0, 1, 1, GRUMAXINVAL - 1, 0xffff);
|
|
get_unlock_tgh_handle(tgh);
|
|
}
|
|
|
|
/*
|
|
* MMUOPS notifier callout functions
|
|
*/
|
|
static int gru_invalidate_range_start(struct mmu_notifier *mn,
|
|
const struct mmu_notifier_range *range)
|
|
{
|
|
struct gru_mm_struct *gms = container_of(mn, struct gru_mm_struct,
|
|
ms_notifier);
|
|
|
|
STAT(mmu_invalidate_range);
|
|
atomic_inc(&gms->ms_range_active);
|
|
gru_dbg(grudev, "gms %p, start 0x%lx, end 0x%lx, act %d\n", gms,
|
|
range->start, range->end, atomic_read(&gms->ms_range_active));
|
|
gru_flush_tlb_range(gms, range->start, range->end - range->start);
|
|
|
|
return 0;
|
|
}
|
|
|
|
static void gru_invalidate_range_end(struct mmu_notifier *mn,
|
|
const struct mmu_notifier_range *range)
|
|
{
|
|
struct gru_mm_struct *gms = container_of(mn, struct gru_mm_struct,
|
|
ms_notifier);
|
|
|
|
/* ..._and_test() provides needed barrier */
|
|
(void)atomic_dec_and_test(&gms->ms_range_active);
|
|
|
|
wake_up_all(&gms->ms_wait_queue);
|
|
gru_dbg(grudev, "gms %p, start 0x%lx, end 0x%lx\n",
|
|
gms, range->start, range->end);
|
|
}
|
|
|
|
static struct mmu_notifier *gru_alloc_notifier(struct mm_struct *mm)
|
|
{
|
|
struct gru_mm_struct *gms;
|
|
|
|
gms = kzalloc(sizeof(*gms), GFP_KERNEL);
|
|
if (!gms)
|
|
return ERR_PTR(-ENOMEM);
|
|
STAT(gms_alloc);
|
|
spin_lock_init(&gms->ms_asid_lock);
|
|
init_waitqueue_head(&gms->ms_wait_queue);
|
|
|
|
return &gms->ms_notifier;
|
|
}
|
|
|
|
static void gru_free_notifier(struct mmu_notifier *mn)
|
|
{
|
|
kfree(container_of(mn, struct gru_mm_struct, ms_notifier));
|
|
STAT(gms_free);
|
|
}
|
|
|
|
static const struct mmu_notifier_ops gru_mmuops = {
|
|
.invalidate_range_start = gru_invalidate_range_start,
|
|
.invalidate_range_end = gru_invalidate_range_end,
|
|
.alloc_notifier = gru_alloc_notifier,
|
|
.free_notifier = gru_free_notifier,
|
|
};
|
|
|
|
struct gru_mm_struct *gru_register_mmu_notifier(void)
|
|
{
|
|
struct mmu_notifier *mn;
|
|
|
|
mn = mmu_notifier_get_locked(&gru_mmuops, current->mm);
|
|
if (IS_ERR(mn))
|
|
return ERR_CAST(mn);
|
|
|
|
return container_of(mn, struct gru_mm_struct, ms_notifier);
|
|
}
|
|
|
|
void gru_drop_mmu_notifier(struct gru_mm_struct *gms)
|
|
{
|
|
mmu_notifier_put(&gms->ms_notifier);
|
|
}
|
|
|
|
/*
|
|
* Setup TGH parameters. There are:
|
|
* - 24 TGH handles per GRU chiplet
|
|
* - a portion (MAX_LOCAL_TGH) of the handles are reserved for
|
|
* use by blade-local cpus
|
|
* - the rest are used by off-blade cpus. This usage is
|
|
* less frequent than blade-local usage.
|
|
*
|
|
* For now, use 16 handles for local flushes, 8 for remote flushes. If the blade
|
|
* has less tan or equal to 16 cpus, each cpu has a unique handle that it can
|
|
* use.
|
|
*/
|
|
#define MAX_LOCAL_TGH 16
|
|
|
|
void gru_tgh_flush_init(struct gru_state *gru)
|
|
{
|
|
int cpus, shift = 0, n;
|
|
|
|
cpus = uv_blade_nr_possible_cpus(gru->gs_blade_id);
|
|
|
|
/* n = cpus rounded up to next power of 2 */
|
|
if (cpus) {
|
|
n = 1 << fls(cpus - 1);
|
|
|
|
/*
|
|
* shift count for converting local cpu# to TGH index
|
|
* 0 if cpus <= MAX_LOCAL_TGH,
|
|
* 1 if cpus <= 2*MAX_LOCAL_TGH,
|
|
* etc
|
|
*/
|
|
shift = max(0, fls(n - 1) - fls(MAX_LOCAL_TGH - 1));
|
|
}
|
|
gru->gs_tgh_local_shift = shift;
|
|
|
|
/* first starting TGH index to use for remote purges */
|
|
gru->gs_tgh_first_remote = (cpus + (1 << shift) - 1) >> shift;
|
|
|
|
}
|