Thank you very much for your reply :)
Inspired by your reply，we successfully reproduced the bug.
The test steps:
1.swapon /dev/zram0
2.add some memory pressure by stress-ng
3.calling swapoff /dev/zram0 in the do_swap_page function (this changed
the source code)
4.bug occured in the same place.
After testing, this patch solves the bug.
Finally, there is a small question. Why linux5.10 revert this patch
(2799e77529c2)?
We found that to fix this bug, the following patches may be required:
efa33fc7f6e mm/shmem: fix shmem_swapin() race with swapoff
5c046235a826 mm/swap: remove confusing checking for non_swap_entry() in
swap_ra_info()
2799e77529c2 swap: fix do_swap_page() race with swapoff
63d8620ecf93 mm/swapfile: use percpu_ref to serialize against concurrent
swapoff
seem like all this patchset is needed except commit 5c046235a826
("mm/swap: remove confusing checking for non_swap_entry() in
swap_ra_info()")
Best Regards,
Xia, longlong
在 2022/11/28 9:08, Huang, Ying 写道:
Hi,
xialonglong <xialonglong1@xxxxxxxxxx> writes:
A panic occur in the linux 5.10 we meet it only once it seems that
there is no special changes between 5.10 and upsteam about swap_cgroup.
The test is based on QEMU with 64GB memory, one 2GB zram device as
swap area.
The test steps:
1.swapoff -a
2.add some memory pressure by stress-ng
3.while (2 minutes) {
swapoff /dev/zram0
swapon /dev/zram0
sleep 3
}
4. swapon -a
Preliminary analysis showed that the swap entry point to a swap area
which have already been swapoff, and no other obvious clues, still
trying to reproduce it.
We have a patch as follows to fix a similar issue,
2799e77529c2a25492a4395db93996e3dacd762d
Author: Miaohe Lin <linmiaohe@xxxxxxxxxx>
AuthorDate: Mon Jun 28 19:36:50 2021 -0700
Commit: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
CommitDate: Tue Jun 29 10:53:49 2021 -0700
swap: fix do_swap_page() race with swapoff
When I was investigating the swap code, I found the below possible race
window:
CPU 1 CPU 2
----- -----
do_swap_page
if (data_race(si->flags & SWP_SYNCHRONOUS_IO)
swap_readpage
if (data_race(sis->flags & SWP_FS_OPS)) {
swapoff
..
p->swap_file = NULL;
..
struct file *swap_file = sis->swap_file;
struct address_space *mapping = swap_file->f_mapping;[oops!]
Note that for the pages that are swapped in through swap cache, this isn't
an issue. Because the page is locked, and the swap entry will be marked
with SWAP_HAS_CACHE, so swapoff() can not proceed until the page has been
unlocked.
Fix this race by using get/put_swap_device() to guard against concurrent
swapoff.
Can you check whether that can fix your issue?
Best Regards,
Huang, Ying
Any known issue about this feature, or any advise will be appreciated.
Here are the panic log,
Unable to handle kernel NULL pointer dereference at virtual address
0000000000000740
Mem abort info:
ESR = 0x96000004
EC = 0x25: DABT (current EL), IL = 32 bits SET = 0, FnV = 0 EA = 0,
S1PTW = 0 Data abort info:
ISV = 0, ISS = 0x00000004
CM = 0, WnR = 0
user pgtable: 4k pages, 48-bit VAs, pgdp=000000010ae6e000
pgd=0000000000000000, p4d=0000000000000000 Internal error: Oops:
96000004 [#1] SMP Hardware name: QEMU KVM Virtual Machine, BIOS 0.0.0
02/06/2015
pstate: 00000005 (nzcv daif -PAN -UAO -TCO BTYPE=--)
pc : lookup_swap_cgroup_id+0x38/0x50
lr : mem_cgroup_charge+0x9c/0x424
sp : ffff800102f63bc0
x29: ffff800102f63bc0 x28: ffff0000d0d64d00
x27: 0000000000000000 x26: 0000000000000007
x25: ffff0000018c86a8 x24: ffff0000018c8640
x23: 0000000000000cc0 x22: 0000000000000001
x21: 0000000000000001 x20: ffff800102f63d28
x19: fffffe000373cb40 x18: 0000000000000000
x17: 0000000000000000 x16: ffff8001004715a4
x15: 00000000ffffffff x14: 0000000000003000
x13: 00000000ffffffff x12: 0000000000000040
x11: ffff0000c0403478 x10: ffff0000c040347a
x9 : ffff8001003e957c x8 : 000000000009dddd
x7 : 0000000000000600 x6 : 00000000000000e8
x5 : 0000020000200000 x4 : ffff000000000000
x3 : ffff800101f4c030 x2 : 0000000000000000
x1 : 00000000000001e4 x0 : 0000000000000000
Call trace:
lookup_swap_cgroup_id+0x38/0x50
do_swap_page+0xa64/0xc04
handle_pte_fault+0x1c8/0x214
__handle_mm_fault+0x1b0/0x380
handle_mm_fault+0xf4/0x284
do_page_fault+0x188/0x474
do_translation_fault+0xb8/0xe4
do_mem_abort+0x48/0xb0
el0_da+0x44/0x80
el0_sync_handler+0x88/0xb4
el0_sync+0x160/0x180
<lookup_swap_cgroup_id>:?????? mov?? x9, x30
<lookup_swap_cgroup_id+0x4>:???? nop
<lookup_swap_cgroup_id+0x8>:????
lsr?? x2, x0, #58 SWP_TYPE_SHIFT == 58? x2 =
swp_type
<lookup_swap_cgroup_id+0xc>:????
adrp? x1, 0xffff800101f4c000
<memcg_sockets_enabled_key+0x8>
<lookup_swap_cgroup_id+0x10>:???
add?? x3, x1, #0x30????
x3 == swap_cgroup_ctrl
<lookup_swap_cgroup_id+0x14>:??? ubfx? x6, x0, #11, #47
<lookup_swap_cgroup_id+0x18>:??? add?? x2, x2, x2, lsl #1
<lookup_swap_cgroup_id+0x1c>:??? ubfiz? x1, x0, #1, #11
<lookup_swap_cgroup_id+0x20>:???
mov?? x5,
#0x200000?????????
// #2097152
<lookup_swap_cgroup_id+0x24>:???
mov?? x4,
#0xffff000000000000???? //
#-281474976710656
<lookup_swap_cgroup_id+0x28>:??? movk? x5, #0x200, lsl #32
<lookup_swap_cgroup_id+0x2c>:??? hint? #0x19
<lookup_swap_cgroup_id+0x30>:???
ldr?? x0, [x3,x2,lsl #3] x3=ffff800101f4c030, x0 = 0
<lookup_swap_cgroup_id+0x34>:??? hint? #0x1d
<lookup_swap_cgroup_id+0x38>:???
ldr?? x0, [x0,x6,lsl #3] x0 = 0 + 0xe8 * 8 == 0x740
<lookup_swap_cgroup_id+0x3c>:??? add?? x0, x0, x5
<lookup_swap_cgroup_id+0x40>:??? lsr?? x0, x0, #6
<lookup_swap_cgroup_id+0x44>:??? add?? x0, x1, x0, lsl #12
<lookup_swap_cgroup_id+0x48>:??? ldrh? w0, [x0,x4]
<lookup_swap_cgroup_id+0x4c>:??? ret
