linux/mm/mmu_gather.c
Linus Torvalds c47454823b mm: mmu_gather: allow more than one batch of delayed rmaps
Commit 5df397dec7 ("mm: delay page_remove_rmap() until after the TLB has
been flushed") limited the page batching for the mmu gather operation when
a dirty shared page needed to delay rmap removal until after the TLB had
been flushed.

It did so because it needs to walk that array of pages while still holding
the page table lock, and our mmu_gather infrastructure allows for batching
quite a lot of pages.  We may have thousands on pages queued up for
freeing, and we wanted to walk only the last batch if we then added a
dirty page to the queue.

However, when I limited it to one batch, I didn't think of the degenerate
case of the special first batch that is embedded on-stack in the
mmu_gather structure (called "local") and that only has eight entries.

So with the right pattern, that "limit delayed rmap to just one batch"
will trigger over and over in that first small batch, and we'll waste a
lot of time flushing TLB's every eight pages.

And those right patterns are trivially triggered by just having a shared
mappings with lots of adjacent dirty pages.  Like the 'page_fault3'
subtest of the 'will-it-scale' benchmark, that just maps a shared area,
dirties all pages, and unmaps it.  Rinse and repeat.

We still want to limit the batching, but to fix this (easily triggered)
degenerate case, just expand the "only one batch" logic to instead be
"only one batch that isn't the special first on-stack ('local') batch".

That way, when we need to flush the delayed rmaps, we can still limit our
walk to just the last batch - and that first small one.

Link: https://lkml.kernel.org/r/CAHk-=whkL5aM1fR7kYUmhHQHBcMUc-bDoFP7EwYjTxy64DGtvw@mail.gmail.com
Fixes: 5df397dec7 ("mm: delay page_remove_rmap() until after the TLB has been flushed")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Reported-by: kernel test robot <yujie.liu@intel.com>
  Link: https://lore.kernel.org/oe-lkp/202212051534.852804af-yujie.liu@intel.com
Tested-by: Huang, Ying <ying.huang@intel.com>
Tested-by: Hugh Dickins <hughd@google.com>
Cc: Feng Tang <feng.tang@intel.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Nadav Amit <nadav.amit@gmail.com>
Cc: Xing Zhengjun <zhengjun.xing@linux.intel.com>
Cc: "Yin, Fengwei" <fengwei.yin@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-12-11 18:12:21 -08:00

398 lines
9.8 KiB
C

#include <linux/gfp.h>
#include <linux/highmem.h>
#include <linux/kernel.h>
#include <linux/mmdebug.h>
#include <linux/mm_types.h>
#include <linux/mm_inline.h>
#include <linux/pagemap.h>
#include <linux/rcupdate.h>
#include <linux/smp.h>
#include <linux/swap.h>
#include <linux/rmap.h>
#include <asm/pgalloc.h>
#include <asm/tlb.h>
#ifndef CONFIG_MMU_GATHER_NO_GATHER
static bool tlb_next_batch(struct mmu_gather *tlb)
{
struct mmu_gather_batch *batch;
/* Limit batching if we have delayed rmaps pending */
if (tlb->delayed_rmap && tlb->active != &tlb->local)
return false;
batch = tlb->active;
if (batch->next) {
tlb->active = batch->next;
return true;
}
if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
return false;
batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
if (!batch)
return false;
tlb->batch_count++;
batch->next = NULL;
batch->nr = 0;
batch->max = MAX_GATHER_BATCH;
tlb->active->next = batch;
tlb->active = batch;
return true;
}
#ifdef CONFIG_SMP
static void tlb_flush_rmap_batch(struct mmu_gather_batch *batch, struct vm_area_struct *vma)
{
for (int i = 0; i < batch->nr; i++) {
struct encoded_page *enc = batch->encoded_pages[i];
if (encoded_page_flags(enc)) {
struct page *page = encoded_page_ptr(enc);
page_remove_rmap(page, vma, false);
}
}
}
/**
* tlb_flush_rmaps - do pending rmap removals after we have flushed the TLB
* @tlb: the current mmu_gather
*
* Note that because of how tlb_next_batch() above works, we will
* never start multiple new batches with pending delayed rmaps, so
* we only need to walk through the current active batch and the
* original local one.
*/
void tlb_flush_rmaps(struct mmu_gather *tlb, struct vm_area_struct *vma)
{
if (!tlb->delayed_rmap)
return;
tlb_flush_rmap_batch(&tlb->local, vma);
if (tlb->active != &tlb->local)
tlb_flush_rmap_batch(tlb->active, vma);
tlb->delayed_rmap = 0;
}
#endif
static void tlb_batch_pages_flush(struct mmu_gather *tlb)
{
struct mmu_gather_batch *batch;
for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
struct encoded_page **pages = batch->encoded_pages;
do {
/*
* limit free batch count when PAGE_SIZE > 4K
*/
unsigned int nr = min(512U, batch->nr);
free_pages_and_swap_cache(pages, nr);
pages += nr;
batch->nr -= nr;
cond_resched();
} while (batch->nr);
}
tlb->active = &tlb->local;
}
static void tlb_batch_list_free(struct mmu_gather *tlb)
{
struct mmu_gather_batch *batch, *next;
for (batch = tlb->local.next; batch; batch = next) {
next = batch->next;
free_pages((unsigned long)batch, 0);
}
tlb->local.next = NULL;
}
bool __tlb_remove_page_size(struct mmu_gather *tlb, struct encoded_page *page, int page_size)
{
struct mmu_gather_batch *batch;
VM_BUG_ON(!tlb->end);
#ifdef CONFIG_MMU_GATHER_PAGE_SIZE
VM_WARN_ON(tlb->page_size != page_size);
#endif
batch = tlb->active;
/*
* Add the page and check if we are full. If so
* force a flush.
*/
batch->encoded_pages[batch->nr++] = page;
if (batch->nr == batch->max) {
if (!tlb_next_batch(tlb))
return true;
batch = tlb->active;
}
VM_BUG_ON_PAGE(batch->nr > batch->max, encoded_page_ptr(page));
return false;
}
#endif /* MMU_GATHER_NO_GATHER */
#ifdef CONFIG_MMU_GATHER_TABLE_FREE
static void __tlb_remove_table_free(struct mmu_table_batch *batch)
{
int i;
for (i = 0; i < batch->nr; i++)
__tlb_remove_table(batch->tables[i]);
free_page((unsigned long)batch);
}
#ifdef CONFIG_MMU_GATHER_RCU_TABLE_FREE
/*
* Semi RCU freeing of the page directories.
*
* This is needed by some architectures to implement software pagetable walkers.
*
* gup_fast() and other software pagetable walkers do a lockless page-table
* walk and therefore needs some synchronization with the freeing of the page
* directories. The chosen means to accomplish that is by disabling IRQs over
* the walk.
*
* Architectures that use IPIs to flush TLBs will then automagically DTRT,
* since we unlink the page, flush TLBs, free the page. Since the disabling of
* IRQs delays the completion of the TLB flush we can never observe an already
* freed page.
*
* Architectures that do not have this (PPC) need to delay the freeing by some
* other means, this is that means.
*
* What we do is batch the freed directory pages (tables) and RCU free them.
* We use the sched RCU variant, as that guarantees that IRQ/preempt disabling
* holds off grace periods.
*
* However, in order to batch these pages we need to allocate storage, this
* allocation is deep inside the MM code and can thus easily fail on memory
* pressure. To guarantee progress we fall back to single table freeing, see
* the implementation of tlb_remove_table_one().
*
*/
static void tlb_remove_table_smp_sync(void *arg)
{
/* Simply deliver the interrupt */
}
void tlb_remove_table_sync_one(void)
{
/*
* This isn't an RCU grace period and hence the page-tables cannot be
* assumed to be actually RCU-freed.
*
* It is however sufficient for software page-table walkers that rely on
* IRQ disabling.
*/
smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
}
static void tlb_remove_table_rcu(struct rcu_head *head)
{
__tlb_remove_table_free(container_of(head, struct mmu_table_batch, rcu));
}
static void tlb_remove_table_free(struct mmu_table_batch *batch)
{
call_rcu(&batch->rcu, tlb_remove_table_rcu);
}
#else /* !CONFIG_MMU_GATHER_RCU_TABLE_FREE */
static void tlb_remove_table_free(struct mmu_table_batch *batch)
{
__tlb_remove_table_free(batch);
}
#endif /* CONFIG_MMU_GATHER_RCU_TABLE_FREE */
/*
* If we want tlb_remove_table() to imply TLB invalidates.
*/
static inline void tlb_table_invalidate(struct mmu_gather *tlb)
{
if (tlb_needs_table_invalidate()) {
/*
* Invalidate page-table caches used by hardware walkers. Then
* we still need to RCU-sched wait while freeing the pages
* because software walkers can still be in-flight.
*/
tlb_flush_mmu_tlbonly(tlb);
}
}
static void tlb_remove_table_one(void *table)
{
tlb_remove_table_sync_one();
__tlb_remove_table(table);
}
static void tlb_table_flush(struct mmu_gather *tlb)
{
struct mmu_table_batch **batch = &tlb->batch;
if (*batch) {
tlb_table_invalidate(tlb);
tlb_remove_table_free(*batch);
*batch = NULL;
}
}
void tlb_remove_table(struct mmu_gather *tlb, void *table)
{
struct mmu_table_batch **batch = &tlb->batch;
if (*batch == NULL) {
*batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
if (*batch == NULL) {
tlb_table_invalidate(tlb);
tlb_remove_table_one(table);
return;
}
(*batch)->nr = 0;
}
(*batch)->tables[(*batch)->nr++] = table;
if ((*batch)->nr == MAX_TABLE_BATCH)
tlb_table_flush(tlb);
}
static inline void tlb_table_init(struct mmu_gather *tlb)
{
tlb->batch = NULL;
}
#else /* !CONFIG_MMU_GATHER_TABLE_FREE */
static inline void tlb_table_flush(struct mmu_gather *tlb) { }
static inline void tlb_table_init(struct mmu_gather *tlb) { }
#endif /* CONFIG_MMU_GATHER_TABLE_FREE */
static void tlb_flush_mmu_free(struct mmu_gather *tlb)
{
tlb_table_flush(tlb);
#ifndef CONFIG_MMU_GATHER_NO_GATHER
tlb_batch_pages_flush(tlb);
#endif
}
void tlb_flush_mmu(struct mmu_gather *tlb)
{
tlb_flush_mmu_tlbonly(tlb);
tlb_flush_mmu_free(tlb);
}
static void __tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm,
bool fullmm)
{
tlb->mm = mm;
tlb->fullmm = fullmm;
#ifndef CONFIG_MMU_GATHER_NO_GATHER
tlb->need_flush_all = 0;
tlb->local.next = NULL;
tlb->local.nr = 0;
tlb->local.max = ARRAY_SIZE(tlb->__pages);
tlb->active = &tlb->local;
tlb->batch_count = 0;
#endif
tlb->delayed_rmap = 0;
tlb_table_init(tlb);
#ifdef CONFIG_MMU_GATHER_PAGE_SIZE
tlb->page_size = 0;
#endif
__tlb_reset_range(tlb);
inc_tlb_flush_pending(tlb->mm);
}
/**
* tlb_gather_mmu - initialize an mmu_gather structure for page-table tear-down
* @tlb: the mmu_gather structure to initialize
* @mm: the mm_struct of the target address space
*
* Called to initialize an (on-stack) mmu_gather structure for page-table
* tear-down from @mm.
*/
void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm)
{
__tlb_gather_mmu(tlb, mm, false);
}
/**
* tlb_gather_mmu_fullmm - initialize an mmu_gather structure for page-table tear-down
* @tlb: the mmu_gather structure to initialize
* @mm: the mm_struct of the target address space
*
* In this case, @mm is without users and we're going to destroy the
* full address space (exit/execve).
*
* Called to initialize an (on-stack) mmu_gather structure for page-table
* tear-down from @mm.
*/
void tlb_gather_mmu_fullmm(struct mmu_gather *tlb, struct mm_struct *mm)
{
__tlb_gather_mmu(tlb, mm, true);
}
/**
* tlb_finish_mmu - finish an mmu_gather structure
* @tlb: the mmu_gather structure to finish
*
* Called at the end of the shootdown operation to free up any resources that
* were required.
*/
void tlb_finish_mmu(struct mmu_gather *tlb)
{
/*
* If there are parallel threads are doing PTE changes on same range
* under non-exclusive lock (e.g., mmap_lock read-side) but defer TLB
* flush by batching, one thread may end up seeing inconsistent PTEs
* and result in having stale TLB entries. So flush TLB forcefully
* if we detect parallel PTE batching threads.
*
* However, some syscalls, e.g. munmap(), may free page tables, this
* needs force flush everything in the given range. Otherwise this
* may result in having stale TLB entries for some architectures,
* e.g. aarch64, that could specify flush what level TLB.
*/
if (mm_tlb_flush_nested(tlb->mm)) {
/*
* The aarch64 yields better performance with fullmm by
* avoiding multiple CPUs spamming TLBI messages at the
* same time.
*
* On x86 non-fullmm doesn't yield significant difference
* against fullmm.
*/
tlb->fullmm = 1;
__tlb_reset_range(tlb);
tlb->freed_tables = 1;
}
tlb_flush_mmu(tlb);
#ifndef CONFIG_MMU_GATHER_NO_GATHER
tlb_batch_list_free(tlb);
#endif
dec_tlb_flush_pending(tlb->mm);
}