On 2022/04/04 13:58, Christoph Hellwig wrote: > On Sat, Apr 02, 2022 at 08:26:23PM +0900, Tetsuo Handa wrote: >>> git tree: linux-next >>> console output: https://syzkaller.appspot.com/x/log.txt?x=168d578db00000 >>> kernel config: https://syzkaller.appspot.com/x/.config?x=be9183de0824e4d7 >>> dashboard link: https://syzkaller.appspot.com/bug?extid=4f1a237abaf14719db49 >>> compiler: gcc (Debian 10.2.1-6) 10.2.1 20210110, GNU ld (GNU Binutils for Debian) 2.35.2 >>> >>> Unfortunately, I don't have any reproducer for this issue yet. >> >> This is a known problem, and Christoph does not like a mitigation patch. >> https://lkml.kernel.org/r/YgoQBTzb3b0l1SzP@xxxxxxxxxxxxx > > And this report shows why: your patch makes the lock acquisition in > blkdev_fallocate killable. Which would not help with this problem at > all, as it does not come through blkdev_fallocate. My patch proposes filemap_invalidate_lock_killable() and converts only blkdev_fallocate() case as a starting point. Nothing prevents us from converting e.g. blk_ioctl_zeroout() case as well. The "not come through blkdev_fallocate" is bogus. > The problem is that > the loop writing actual zeroes to the disk can take forever, and we > hold the invalidate_lock over that. There can be several locations which cannot be converted to filemap_invalidate_lock_killable() (like we had to use mutex_lock(&lo->lo_mutex) for lo_release() case). But many locations can be converted to filemap_invalidate_lock_killable() (like we use mutex_lock_killable(&lo->lo_mutex) for almost all locations). Being responsive to SIGKILL is good (even if we cannot afford making all locks killable). I made many locations killable. Therefore, I wonder why not?
