Re: [PATCH] fat: serialize primary and mirror updates

[Matthias Goergens]

OGAWA Hirofumi <hirofumi@xxxxxxxxxxxxxxxxxx> writes:

> What is the real issue?

Only the data race syzbot reports, "KCSAN: data-race in fat16_ent_put /
fat_mirror_bhs": fat_mirror_bhs() copies a FAT block to the backup FAT
while another task updates an entry in that block.  I have not found any
effect on disk.  fat_mirror_bhs() holds the backup buffer's lock across
the copy, and every writer mirrors the block again after its own update,
so a copy that caught an entry mid-update is replaced by that writer's
copy.  The commit message should have said so, and the patch also lacked
my Signed-off-by: it went out before it was ready, sorry.

That leaves two ways to deal with the report.  Would you prefer marking
the copy as an intended race, data_race() around the memcpy() with a
comment on why it is harmless?  Or the locking in this patch, which
holds fat_lock across fat_sync_bhs() for synchronous inodes and so
makes other allocations wait for that I/O?

Matthias