The code in fscache_retrieval_complete is using atomic_sub followed by an
atomic_read:
atomic_sub(n_pages, &op->n_pages);
if (atomic_read(&op->n_pages) <= 0)
fscache_op_complete(&op->op, true);
This causes two threads doing a decrement of n_pages to race with each
other seeing the op->refcount 0 at same time - and they end up calling
fscache_op_complete() in both the threads leading to an assertion failure.
Fix this by using atomic_sub_return_relaxed() instead of two calls. Note
that I'm using 'relaxed' rather than, say, 'release' as there aren't
multiple variables that appear to need ordering across the release.
The oops looks something like:
FS-Cache: Assertion failed
FS-Cache: 0 > 0 is false
...
kernel BUG at /usr/src/linux-4.4.0/fs/fscache/operation.c:449!
...
Workqueue: fscache_operation fscache_op_work_func [fscache]
...
RIP: 0010:[<ffffffffc037eacd>] fscache_op_complete+0x10d/0x180 [fscache]
...
Call Trace:
[<ffffffffc1464cf9>] cachefiles_read_copier+0x3a9/0x410 [cachefiles]
[<ffffffffc037e272>] fscache_op_work_func+0x22/0x50 [fscache]
[<ffffffff81096da0>] process_one_work+0x150/0x3f0
[<ffffffff8109751a>] worker_thread+0x11a/0x470
[<ffffffff81808e59>] ? __schedule+0x359/0x980
[<ffffffff81097400>] ? rescuer_thread+0x310/0x310
[<ffffffff8109cdd6>] kthread+0xd6/0xf0
[<ffffffff8109cd00>] ? kthread_park+0x60/0x60
[<ffffffff8180d0cf>] ret_from_fork+0x3f/0x70
[<ffffffff8109cd00>] ? kthread_park+0x60/0x60
This seen this in 4.4.x kernels and the same bug affects fscache in latest
upstreams kernels.
Fixes:
|
||
---|---|---|
.. | ||
acpi | ||
asm-generic | ||
clocksource | ||
crypto | ||
drm | ||
dt-bindings | ||
keys | ||
kvm | ||
linux | ||
math-emu | ||
media | ||
memory | ||
misc | ||
net | ||
pcmcia | ||
ras | ||
rdma | ||
scsi | ||
soc | ||
sound | ||
target | ||
trace | ||
uapi | ||
video | ||
xen |