On 5/7/19 4:51 AM, Jan Stancek wrote:
----- Original Message -----
On 5/6/19 2:35 PM, Jan Stancek wrote:
----- Original Message -----
On 5/5/19 7:10 AM, Jan Stancek wrote:
Hi,
I'm seeing userspace program getting stuck on aarch64, on kernels 4.20
and
newer.
It stalls from seconds to hours.
I have simplified it to following scenario (reproducer linked below [1]):
while (1):
spawn Thread 1: mmap, write, munmap
spawn Thread 2: <nothing>
Thread 1 is sporadically getting stuck on write to mapped area.
User-space
is not
moving forward - stdout output stops. Observed CPU usage is however 100%.
At this time, kernel appears to be busy handling page faults (~700k per
second):
# perf top -a -g
- 98.97% 8.30% a.out [.] map_write_unmap
- 23.52% map_write_unmap
- 24.29% el0_sync
- 10.42% do_mem_abort
- 17.81% do_translation_fault
- 33.01% do_page_fault
- 56.18% handle_mm_fault
40.26% __handle_mm_fault
2.19% __ll_sc___cmpxchg_case_acq_4
0.87% mem_cgroup_from_task
- 6.18% find_vma
5.38% vmacache_find
1.35% __ll_sc___cmpxchg_case_acq_8
1.23% __ll_sc_atomic64_sub_return_release
0.78% down_read_trylock
0.93% do_translation_fault
+ 8.30% thread_start
# perf stat -p 8189 -d
^C
Performance counter stats for process id '8189':
984.311350 task-clock (msec) # 1.000 CPUs
utilized
0 context-switches # 0.000 K/sec
0 cpu-migrations # 0.000 K/sec
723,641 page-faults # 0.735 M/sec
2,559,199,434 cycles # 2.600 GHz
711,933,112 instructions # 0.28 insn per
cycle
<not supported> branches
757,658 branch-misses
205,840,557 L1-dcache-loads # 209.121 M/sec
40,561,529 L1-dcache-load-misses # 19.71% of all
L1-dcache hits
<not supported> LLC-loads
<not supported> LLC-load-misses
0.984454892 seconds time elapsed
With some extra traces, it appears looping in page fault for same
address,
over and over:
do_page_fault // mm_flags: 0x55
__do_page_fault
__handle_mm_fault
handle_pte_fault
ptep_set_access_flags
if (pte_same(pte, entry)) // pte: e8000805060f53, entry:
e8000805060f53
I had traces in mmap() and munmap() as well, they don't get hit when
reproducer
hits the bad state.
Notes:
- I'm not able to reproduce this on x86.
- Attaching GDB or strace immediatelly recovers application from stall.
- It also seems to recover faster when system is busy with other tasks.
- MAP_SHARED vs. MAP_PRIVATE makes no difference.
- Turning off THP makes no difference.
- Reproducer [1] usually hits it within ~minute on HW described below.
- Longman mentioned that "When the rwsem becomes reader-owned, it causes
all the spinning writers to go to sleep adding wakeup latency to
the time required to finish the critical sections", but this looks
like busy loop, so I'm not sure if it's related to rwsem issues
identified
in:
https://lore.kernel.org/lkml/20190428212557.13482-2-longman@xxxxxxxxxx/
It sounds possible to me. What the optimization done by the commit ("mm:
mmap: zap pages with read mmap_sem in munmap") is to downgrade write
rwsem to read when zapping pages and page table in munmap() after the
vmas have been detached from the rbtree.
So the mmap(), which is writer, in your test may steal the lock and
execute with the munmap(), which is the reader after the downgrade, in
parallel to break the mutual exclusion.
In this case, the parallel mmap() may map to the same area since vmas
have been detached by munmap(), then mmap() may create the complete same
vmas, and page fault happens on the same vma at the same address.
I'm not sure why gdb or strace could recover this, but they use ptrace
which may acquire mmap_sem to break the parallel inadvertently.
May you please try Waiman's patch to see if it makes any difference?
I don't see any difference in behaviour after applying:
[PATCH-tip v7 01/20] locking/rwsem: Prevent decrement of reader count
before increment
Issue is still easily reproducible for me.
I'm including output of mem_abort_decode() / show_pte() for sample PTE,
that
I see in page fault loop. (I went through all bits, but couldn't find
anything invalid about it)
mem_abort_decode: Mem abort info:
mem_abort_decode: ESR = 0x92000047
mem_abort_decode: Exception class = DABT (lower EL), IL = 32 bits
mem_abort_decode: SET = 0, FnV = 0
mem_abort_decode: EA = 0, S1PTW = 0
mem_abort_decode: Data abort info:
mem_abort_decode: ISV = 0, ISS = 0x00000047
mem_abort_decode: CM = 0, WnR = 1
show_pte: user pgtable: 64k pages, 48-bit VAs, pgdp = 0000000067027567
show_pte: [0000ffff6dff0000] pgd=000000176bae0003
show_pte: , pud=000000176bae0003
show_pte: , pmd=000000174ad60003
show_pte: , pte=00e80008023a0f53
show_pte: , pte_pfn: 8023a
>>> print bin(0x47)
0b1000111
Per D12-2779 (ARM Architecture Reference Manual),
ISS encoding for an exception from an Instruction Abort:
IFSC, bits [5:0], Instruction Fault Status Code
0b000111 Translation fault, level 3
---
My theory is that TLB is getting broken.
Theory continued:
unmap_region() is batching updates to TLB (for vmas and page tables).
And at the same time another thread handles page fault for same mm,
which increases "tlb_flush_pending".
tlb_finish_mmu() called from unmap_region() will thus set 'force = 1'.
And arch_tlb_finish_mmu() will in turn reset TLB range, presumably making
it smaller then it would be if force == 0.
Change below appears to fix it:
diff --git a/mm/mmu_gather.c b/mm/mmu_gather.c
index f2f03c655807..a4cef21bd62b 100644
--- a/mm/mmu_gather.c
+++ b/mm/mmu_gather.c
@@ -93,7 +93,7 @@ void arch_tlb_finish_mmu(struct mmu_gather *tlb,
struct mmu_gather_batch *batch, *next;
if (force) {
- __tlb_reset_range(tlb);
__tlb_adjust_range(tlb, start, end - start);
I don't get why the change could fix it? __tlb_reset_range() just reset
start and end to TASK_SIZE and 0, then __tlb_adjust_range() set proper
start and end. I don't get why "force" flush smaller range?
}
I made a dummy kernel module that exports debugfs file, which on read
triggers:
flush_tlb_all();
Any time reproducer stalls and I read debugfs file, it recovers
immediately and resumes printing to stdout.
That commit doesn't change anything about TLB flush, just move zapping
pages under read mmap_sem as what MADV_DONTNEED does.
I don't have aarch64 board to reproduce and debug it. And, I'm not
familiar with aarch64 architecture either. But, some history told me the
parallel zapping page may run into stale TLB and defer a flush meaning
that this call may observe pte_none and fails to flush the TLB. But,
this has been solved by commit 56236a59556c ("mm: refactor TLB gathering
API") and 99baac21e458 ("mm: fix MADV_[FREE|DONTNEED] TLB flush miss
problem").
For more detail, please refer to commit 4647706ebeee ("mm: always flush
VMA ranges affected by zap_page_range"). Copied Mel and Rik in this
thread. Also added Will Deacon and Catalin Marinas, who are aarch64
maintainers, in this loop
Thanks
But, your test (triggering TLB flush) does demonstrate TLB flush is
*not* done properly at some point as expected for aarch64. Could you
please give the below patch a try?
Your patch also fixes my reproducer.
Thanks for testing it.
diff --git a/mm/memory.c b/mm/memory.c
index ab650c2..ef41ad5 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -1336,8 +1336,10 @@ void unmap_vmas(struct mmu_gather *tlb,
mmu_notifier_range_init(&range, vma->vm_mm, start_addr, end_addr);
mmu_notifier_invalidate_range_start(&range);
- for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
+ for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
+ flush_tlb_range(vma, start_addr, end_addr);
+ }
mmu_notifier_invalidate_range_end(&range);
}
- I tried 2 different aarch64 systems so far: APM X-Gene CPU Potenza A3
and
Qualcomm 65-LA-115-151.
I can reproduce it on both with v5.1-rc7. It's easier to reproduce
on latter one (for longer periods of time), which has 46 CPUs.
- Sample output of reproducer on otherwise idle system:
# ./a.out
[00000314] map_write_unmap took: 26305 ms
[00000867] map_write_unmap took: 13642 ms
[00002200] map_write_unmap took: 44237 ms
[00002851] map_write_unmap took: 992 ms
[00004725] map_write_unmap took: 542 ms
[00006443] map_write_unmap took: 5333 ms
[00006593] map_write_unmap took: 21162 ms
[00007435] map_write_unmap took: 16982 ms
[00007488] map_write unmap took: 13 ms^C
I ran a bisect, which identified following commit as first bad one:
dd2283f2605e ("mm: mmap: zap pages with read mmap_sem in munmap")
I can also make the issue go away with following change:
diff --git a/mm/mmap.c b/mm/mmap.c
index 330f12c17fa1..13ce465740e2 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -2844,7 +2844,7 @@ EXPORT_SYMBOL(vm_munmap);
SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
{
profile_munmap(addr);
- return __vm_munmap(addr, len, true);
+ return __vm_munmap(addr, len, false);
}
# cat /proc/cpuinfo | head
processor : 0
BogoMIPS : 40.00
Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 cpuid
asimdrdm
CPU implementer : 0x51
CPU architecture: 8
CPU variant : 0x0
CPU part : 0xc00
CPU revision : 1
# numactl -H
available: 1 nodes (0)
node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
23
24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45
node 0 size: 97938 MB
node 0 free: 95732 MB
node distances:
node 0
0: 10
Regards,
Jan
[1]
https://github.com/jstancek/reproducers/blob/master/kernel/page_fault_stall/mmap5.c
[2]
https://github.com/jstancek/reproducers/blob/master/kernel/page_fault_stall/config
