Hi, Chris Mason > > > > > > > reada create 2 works for each level of tree in recursion. > > > > > > > > > > > > > > In case of a tree having many levels, the number of created > > > > > > > works is 2^level_of_tree. > > > > > > > Actually we don't need so many works in parallel, this patch > > > > > > > limit max works to BTRFS_MAX_MIRRORS * 2. > > > > > > > > > > > > Hi, > > > > > > > > > > > > I don't think you end up calling atomic_dec() for every time > > > > > > that > > > > > > reada_start_machine() is called. Also, I'd rather not have a > > > > > > global static variable to limit the parallel workers, when we > > > > > > have more than one FS mounted it'll end up limiting things too much. > > > > > > > > > > > > With this patch applied, I'm seeing deadlocks during btrfs/066. > You > > > > > > have to run the scrub tests as well, basically we're just > > > > > > getting fsstress run alongside scrub. > > > > > > > > > > > > I'll run a few more times with it reverted to make sure, but I > > > > > > think it's the root cause. > > > > > > > > > > I spoke too soon, it ended up deadlocking a few tests later. > > > > > > > > > In logic, even if the calculation of atomic_dec() in this patch > > > > having bug, in worst condition, reada will works in single-thread > > > > mode, and will not introduce deadlock. > > > > > > > > And by looking the backtrace in this mail, maybe it is caused by > > > > reada_control->elems in someplace of this patchset. > > > > > > > > I recheck xfstests/066 in both vm and physical machine, on top of > > > > my pull-request git today, with btrfs-progs 4.4 for many times, > > > > but had not > > > triggered the bug. > > > > > > Just running 066 alone doesn't trigger it for me. I have to run > > > everything from > > > 00->066. > > > > > > My setup is 5 drives. I use a script to carve them up into logical > > > volumes, 5 for the test device and 5 for the scratch pool. I think > > > it should reproduce with a single drive, if you still can't trigger I'll confirm > that. > > > > > > > > > > > Could you tell me your test environment(TEST_DEV size, mount > > > > option), and odds of fails in btrfs/066? > > > > > > 100% odds of failing, one time it made it up to btrfs/072. I think > > > more important than the drive setup is that I have all the debugging on. > > > CONFIG_DEBUG_PAGEALLOC, spinlock debugging, mutex debugging and > lock > > > dep enabled. > > > > > Thanks for your answer. > > > > But unfortunately I hadn't reproduce the dead_lock in above way today... > > Now I queued loop of above reproduce script in more nodes, and hopes > > it can happen in this weekend. > > > > And by reviewing code, I found a problem which can introduce similar > > bad result in logic, and made a patch for it. > > [PATCH] [RFC] btrfs: reada: avoid undone reada extents in > > btrfs_reada_wait > > > > Because it is only a problem in logic, but rarely happened, I only > > confirmed no-problem after patch applied. > > > > Sorry for increased your works, could you apply this patch and test is > > it works? > > No problem, I'll try the patch and see if I can get a more reliable way to > reproduce if it doesn't fix things. Thanks! > I rebased following branch: https://github.com/zhaoleidd/btrfs.git integration-4.5 With updated patch to fix btrfs/066 bug. Bug reason is descripted in changelog of: btrfs: reada: avoid undone reada extents in btrfs_reada_wait Test: 1: In the node which can repgoduce btrfs/066 bug, Confirmed HAVING_BUG before patch, and NO_BUG after patch. 2: Run xfstests's btrfs group, confirmed no regression. Most patchs in this branch are for reada, except this one for NO_SPACE bug: btrfs: Continue write in case of can_not_nocow Cound you consider merging it in suitable time? Thanks Zhaolei > -chris > -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
