On 6/2/20 8:22 PM, Hillf Danton wrote:
Tue, 02 Jun 2020 04:20:16 -0700
syzbot found the following crash on:
HEAD commit: 0e21d462 Add linux-next specific files for 20200602
git tree: linux-next
console output: https://syzkaller.appspot.com/x/log.txt?x=127233ee100000
kernel config: https://syzkaller.appspot.com/x/.config?x=ecc1aef35f550ee3
dashboard link: https://syzkaller.appspot.com/bug?extid=82f324bb69744c5f6969
compiler: gcc (GCC) 9.0.0 20181231 (experimental)
IMPORTANT: if you fix the bug, please add the following tag to the commit:
Reported-by: syzbot+82f324bb69744c5f6969@xxxxxxxxxxxxxxxxxxxxxxxxx
BUG: using smp_processor_id() in preemptible [00000000] code: syz-fuzzer/6792
caller is ext4_mb_new_blocks+0xa4d/0x3b70 fs/ext4/mballoc.c:4711
Fix 42f56b7a4a7d ("ext4: mballoc: introduce pcpu seqcnt for freeing PA
to improve ENOSPC handling") by redefining discard_pa_seq to be a simple
regular sequence counter to axe the need of percpu operation.
Why remove percpu seqcnt? IIUC, percpu are much better in case of a
multi-threaded use case which could run and allocate blocks in parallel.
Whereas a updating a simple variable across different cpus may lead to
cacheline bouncing problem.
Since in this case we can very well have a use case of multiple threads
trying to allocate blocks at the same time, so why change this to a
simple seqcnt from percpu seqcnt?
-ritesh
