Description
System information
| Type | Version/Name |
|---|---|
| Distribution Name | Archlinux |
| Distribution Version | rolling |
| Linux Kernel | 5.9.10 |
| Architecture | x64 |
| ZFS Version | 8f158ae |
| SPL Version | 8f158ae |
Describe the problem you're observing
Memory stressing a system with mprime results in a panic in arc_wait_for_eviction().
Describe how to reproduce the problem
In a system with 128GB memory, of which 40GB were consumed by ARC running mprime panicked in the assertion of arc_wait_for_eviction(). mprime was run in torture-test mode, type of test was "4. Blend (tests all of the above)" (ie both large and small fast Fourrier transforms). In this test mode mprime tries to consume all available memory.
I believe what happened is that the system was running low on memory so that the if clause at https://github.com/openzfs/zfs/blob/master/module/zfs/arc.c#L4166 was not true and arc_evict_waiters were not woken. However, arc_evict_count was still incremented at https://github.com/openzfs/zfs/blob/master/module/zfs/arc.c#L4164 which led to the panicked assertion in arc_wait_for_eviction().
Just a thought though.
Include any warning/errors/backtraces from the system logs
VERIFY3(last->aew_count > arc_evict_count) failed (850903552 > 868262400)
PANIC at arc.c:5255:arc_wait_for_eviction()
Showing stack for process 1429448
CPU: 4 PID: 1429448 Comm: mprime Tainted: P OE 5.9.10-1-vfio #1
Call Trace:
dump_stack+0x6b/0x83
spl_panic+0xef/0x117 [spl]
? sysvec_call_function+0x36/0x80
? asm_sysvec_call_function+0x12/0x20
arc_wait_for_eviction+0x1db/0x1f0 [zfs]
arc_shrinker_scan+0x36/0xd0 [zfs]
do_shrink_slab+0x146/0x290
shrink_slab+0xd0/0x2f0
shrink_node+0x2c0/0x6e0
do_try_to_free_pages+0xda/0x4c0
try_to_free_pages+0xef/0x1c0
__alloc_pages_slowpath.constprop.0+0x384/0xce0
? tick_nohz_next_event+0x8f/0x180
__alloc_pages_nodemask+0x2e6/0x310
alloc_pages_vma+0x80/0x250
handle_mm_fault+0xebd/0x1930
do_user_addr_fault+0x1b8/0x3f0
exc_page_fault+0x82/0x1a0
? asm_exc_page_fault+0x8/0x30
asm_exc_page_fault+0x1e/0x30