commit 7a30df49f63ad92318ddf1f7498d1129a77dd4bd upstream
A few new fields were added to mmu_gather to make TLB flush smarter for
huge page by telling what level of page table is changed.
__tlb_reset_range() is used to reset all these page table state to
unchanged, which is called by TLB flush for parallel mapping changes for
the same range under non-exclusive lock (i.e. read mmap_sem).
Before commit dd2283f2605e ("mm: mmap: zap pages with read mmap_sem in
munmap"), the syscalls (e.g. MADV_DONTNEED, MADV_FREE) which may update
PTEs in parallel don't remove page tables. But, the forementioned
commit may do munmap() under read mmap_sem and free page tables. This
may result in program hang on aarch64 reported by Jan Stancek. The
problem could be reproduced by his test program with slightly modified
below.
---8<---
static int map_size = 4096;
static int num_iter = 500;
static long threads_total;
static void *distant_area;
void *map_write_unmap(void *ptr)
{
int *fd = ptr;
unsigned char *map_address;
int i, j = 0;
for (i = 0; i < num_iter; i++) {
map_address = mmap(distant_area, (size_t) map_size, PROT_WRITE | PROT_READ,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
if (map_address == MAP_FAILED) {
perror("mmap");
exit(1);
}
for (j = 0; j < map_size; j++)
map_address[j] = 'b';
if (munmap(map_address, map_size) == -1) {
perror("munmap");
exit(1);
}
}
return NULL;
}
void *dummy(void *ptr)
{
return NULL;
}
int main(void)
{
pthread_t thid[2];
/* hint for mmap in map_write_unmap() */
distant_area = mmap(0, DISTANT_MMAP_SIZE, PROT_WRITE | PROT_READ,
MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
munmap(distant_area, (size_t)DISTANT_MMAP_SIZE);
distant_area += DISTANT_MMAP_SIZE / 2;
while (1) {
pthread_create(&thid[0], NULL, map_write_unmap, NULL);
pthread_create(&thid[1], NULL, dummy, NULL);
pthread_join(thid[0], NULL);
pthread_join(thid[1], NULL);
}
}
---8<---
The program may bring in parallel execution like below:
t1 t2
munmap(map_address)
downgrade_write(&mm->mmap_sem);
unmap_region()
tlb_gather_mmu()
inc_tlb_flush_pending(tlb->mm);
free_pgtables()
tlb->freed_tables = 1
tlb->cleared_pmds = 1
pthread_exit()
madvise(thread_stack, 8M, MADV_DONTNEED)
zap_page_range()
tlb_gather_mmu()
inc_tlb_flush_pending(tlb->mm);
tlb_finish_mmu()
if (mm_tlb_flush_nested(tlb->mm))
__tlb_reset_range()
__tlb_reset_range() would reset freed_tables and cleared_* bits, but this
may cause inconsistency for munmap() which do free page tables. Then it
may result in some architectures, e.g. aarch64, may not flush TLB
completely as expected to have stale TLB entries remained.
Use fullmm flush since it yields much better performance on aarch64 and
non-fullmm doesn't yields significant difference on x86.
The original proposed fix came from Jan Stancek who mainly debugged this
issue, I just wrapped up everything together.
Jan's testing results:
v5.2-rc2-24-gbec7550cca10
--------------------------
mean stddev
real 37.382 2.780
user 1.420 0.078
sys 54.658 1.855
v5.2-rc2-24-gbec7550cca10 + "mm: mmu_gather: remove __tlb_reset_range() for force flush"
---------------------------------------------------------------------------------------_
mean stddev
real 37.119 2.105
user 1.548 0.087
sys 55.698 1.357
[akpm@xxxxxxxxxxxxxxxxxxxx: coding-style fixes]
Link: http://lkml.kernel.org/r/1558322252-113575-1-git-send-email-yang.shi@xxxxxxxxxxxxxxxxx
Fixes: dd2283f2605e ("mm: mmap: zap pages with read mmap_sem in munmap")
Signed-off-by: Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx>
Signed-off-by: Jan Stancek <jstancek@xxxxxxxxxx>
Reported-by: Jan Stancek <jstancek@xxxxxxxxxx>
Tested-by: Jan Stancek <jstancek@xxxxxxxxxx>
Suggested-by: Will Deacon <will.deacon@xxxxxxx>
Tested-by: Will Deacon <will.deacon@xxxxxxx>
Acked-by: Will Deacon <will.deacon@xxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Nick Piggin <npiggin@xxxxxxxxx>
Cc: "Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxxxxx>
Cc: Nadav Amit <namit@xxxxxxxxxx>
Cc: Minchan Kim <minchan@xxxxxxxxxx>
Cc: Mel Gorman <mgorman@xxxxxxx>
Cc: <stable@xxxxxxxxxxxxxxx> [4.20+]
Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Signed-off-by: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
---
mm/mmu_gather.c | 24 +++++++++++++++++++-----
1 file changed, 19 insertions(+), 5 deletions(-)
diff --git a/mm/mmu_gather.c b/mm/mmu_gather.c
index f2f03c6..a58bd0d 100644
--- a/mm/mmu_gather.c
+++ b/mm/mmu_gather.c
@@ -93,8 +93,17 @@ void arch_tlb_finish_mmu(struct mmu_gather *tlb,
struct mmu_gather_batch *batch, *next;
if (force) {
+ /*
+ * 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_adjust_range(tlb, start, end - start);
+ tlb->freed_tables = 1;
}
tlb_flush_mmu(tlb);
@@ -249,10 +258,15 @@ 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_sem read-side) but defer TLB
- * flush by batching, a thread has stable TLB entry can fail to flush
- * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
- * forcefully if we detect parallel PTE batching threads.
+ * under non-exclusive lock (e.g., mmap_sem 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.
*/
bool force = mm_tlb_flush_nested(tlb->mm);
--
1.8.3.1
