| Message ID | 1469061852-21185-1-git-send-email-bo.li.liu@oracle.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
On 07/20/2016 08:44 PM, Liu Bo wrote: > While processing delayed refs, we may update block group's statistics > and attach it to cur_trans->dirty_bgs, and later writing dirty block > groups will process the list, which happens during > btrfs_commit_transaction(). > > For whatever reason, the transaction is aborted and dirty_bgs > is not processed in cleanup_transaction(), we end up with memory leak > of these dirty block group cache. > > Since btrfs_start_dirty_block_groups() doesn't make it go to the commit > critical section, this also adds the cleanup work inside it. It's the start_drity_block_groups() hunt that worries me a bit: > diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c > index 50bd683..7a35c9d 100644 > --- a/fs/btrfs/extent-tree.c > +++ b/fs/btrfs/extent-tree.c > @@ -3698,6 +3698,8 @@ again: > goto again; > } > spin_unlock(&cur_trans->dirty_bgs_lock); > + } else if (ret < 0) { > + btrfs_cleanup_dirty_bgs(cur_trans, root); > } > > btrfs_free_path(path); > We have checks in here to make sure only one process runs btrfs_start_dirty_block_groups() but that doesn't mean that only one process its messing around with the cache inode. Is there any reason we can't let this cleanup wait for the cleanup_transaction code? btrfs_run_delayed_refs() already aborts when it fails. -chris -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Jul 21, 2016 at 11:32:26AM -0400, Chris Mason wrote: > > > On 07/20/2016 08:44 PM, Liu Bo wrote: > > While processing delayed refs, we may update block group's statistics > > and attach it to cur_trans->dirty_bgs, and later writing dirty block > > groups will process the list, which happens during > > btrfs_commit_transaction(). > > > > For whatever reason, the transaction is aborted and dirty_bgs > > is not processed in cleanup_transaction(), we end up with memory leak > > of these dirty block group cache. > > > > Since btrfs_start_dirty_block_groups() doesn't make it go to the commit > > critical section, this also adds the cleanup work inside it. > > It's the start_drity_block_groups() hunt that worries me a bit: > > > diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c > > index 50bd683..7a35c9d 100644 > > --- a/fs/btrfs/extent-tree.c > > +++ b/fs/btrfs/extent-tree.c > > @@ -3698,6 +3698,8 @@ again: > > goto again; > > } > > spin_unlock(&cur_trans->dirty_bgs_lock); > > + } else if (ret < 0) { > > + btrfs_cleanup_dirty_bgs(cur_trans, root); > > } > > > > btrfs_free_path(path); > > > > We have checks in here to make sure only one process runs > btrfs_start_dirty_block_groups() but that doesn't mean that only one process > its messing around with the cache inode. Is there any reason we can't let > this cleanup wait for the cleanup_transaction code? > > btrfs_run_delayed_refs() already aborts when it fails. update_block_group() is the only producer to add block group cache to dirty_bgs list, and if btrfs_run_delayed_refs() aborts, the transaction is aborted, so seems that there won't be anyone manipulating dirty_bgs list, am I missing? Another point is that when we fail on btrfs_start_dirty_block_groups(), btrfs_commit_transaction() won't get to cleanup_transaction error handling, btrfs_commit_transaction() { ... if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) { ... ret = btrfs_start_dirty_block_groups(trans, root); } if (ret) { btrfs_end_transaction(trans, root); return ret; } ... cleanup_transaction(); } But yes, if we delay the cleanup, we still have a chance to do cleanup in btrfs_error_commit_super(), and I have sent another patch to add several ASSERT()s to check block group related memory leak, with which we'll be warned if anything wrong. I'm OK to remove the part that causes concerns. Thanks, -liubo > > -chris > > -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 07/21/2016 03:03 PM, Liu Bo wrote: > On Thu, Jul 21, 2016 at 11:32:26AM -0400, Chris Mason wrote: >> >> >> On 07/20/2016 08:44 PM, Liu Bo wrote: >>> While processing delayed refs, we may update block group's statistics >>> and attach it to cur_trans->dirty_bgs, and later writing dirty block >>> groups will process the list, which happens during >>> btrfs_commit_transaction(). >>> >>> For whatever reason, the transaction is aborted and dirty_bgs >>> is not processed in cleanup_transaction(), we end up with memory leak >>> of these dirty block group cache. >>> >>> Since btrfs_start_dirty_block_groups() doesn't make it go to the commit >>> critical section, this also adds the cleanup work inside it. >> >> It's the start_drity_block_groups() hunt that worries me a bit: >> >>> diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c >>> index 50bd683..7a35c9d 100644 >>> --- a/fs/btrfs/extent-tree.c >>> +++ b/fs/btrfs/extent-tree.c >>> @@ -3698,6 +3698,8 @@ again: >>> goto again; >>> } >>> spin_unlock(&cur_trans->dirty_bgs_lock); >>> + } else if (ret < 0) { >>> + btrfs_cleanup_dirty_bgs(cur_trans, root); >>> } >>> >>> btrfs_free_path(path); >>> >> >> We have checks in here to make sure only one process runs >> btrfs_start_dirty_block_groups() but that doesn't mean that only one process >> its messing around with the cache inode. Is there any reason we can't let >> this cleanup wait for the cleanup_transaction code? >> >> btrfs_run_delayed_refs() already aborts when it fails. > > update_block_group() is the only producer to add block group cache to > dirty_bgs list, and if btrfs_run_delayed_refs() aborts, the transaction > is aborted, so seems that there won't be anyone manipulating dirty_bgs > list, am I missing? > No, the dirty_bgs processing is safe I think. My concern is with the cache inode which we iput() > Another point is that when we fail on btrfs_start_dirty_block_groups(), > btrfs_commit_transaction() won't get to cleanup_transaction error > handling, Right, because we don't actually finish the commit. Someone will eventually though ;) > > btrfs_commit_transaction() { > ... > if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) { > ... > ret = btrfs_start_dirty_block_groups(trans, root); > } > if (ret) { > btrfs_end_transaction(trans, root); > return ret; > } > ... > cleanup_transaction(); > } > > > But yes, if we delay the cleanup, we still have a chance to do cleanup > in btrfs_error_commit_super(), and I have sent another patch to add > several ASSERT()s to check block group related memory leak, with which > we'll be warned if anything wrong. > > I'm OK to remove the part that causes concerns. Thanks. -chris -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Jul 21, 2016 at 03:17:41PM -0400, Chris Mason wrote: > > > On 07/21/2016 03:03 PM, Liu Bo wrote: > > On Thu, Jul 21, 2016 at 11:32:26AM -0400, Chris Mason wrote: > > > > > > > > > On 07/20/2016 08:44 PM, Liu Bo wrote: > > > > While processing delayed refs, we may update block group's statistics > > > > and attach it to cur_trans->dirty_bgs, and later writing dirty block > > > > groups will process the list, which happens during > > > > btrfs_commit_transaction(). > > > > > > > > For whatever reason, the transaction is aborted and dirty_bgs > > > > is not processed in cleanup_transaction(), we end up with memory leak > > > > of these dirty block group cache. > > > > > > > > Since btrfs_start_dirty_block_groups() doesn't make it go to the commit > > > > critical section, this also adds the cleanup work inside it. > > > > > > It's the start_drity_block_groups() hunt that worries me a bit: > > > > > > > diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c > > > > index 50bd683..7a35c9d 100644 > > > > --- a/fs/btrfs/extent-tree.c > > > > +++ b/fs/btrfs/extent-tree.c > > > > @@ -3698,6 +3698,8 @@ again: > > > > goto again; > > > > } > > > > spin_unlock(&cur_trans->dirty_bgs_lock); > > > > + } else if (ret < 0) { > > > > + btrfs_cleanup_dirty_bgs(cur_trans, root); > > > > } > > > > > > > > btrfs_free_path(path); > > > > > > > > > > We have checks in here to make sure only one process runs > > > btrfs_start_dirty_block_groups() but that doesn't mean that only one process > > > its messing around with the cache inode. Is there any reason we can't let > > > this cleanup wait for the cleanup_transaction code? > > > > > > btrfs_run_delayed_refs() already aborts when it fails. > > > > update_block_group() is the only producer to add block group cache to > > dirty_bgs list, and if btrfs_run_delayed_refs() aborts, the transaction > > is aborted, so seems that there won't be anyone manipulating dirty_bgs > > list, am I missing? > > > > No, the dirty_bgs processing is safe I think. My concern is with the cache > inode which we iput() I think iput() is OK, we're doing iput() on block group cache on the io_bgs list, where all block groups's inodes has been igrab()'d. If others are messing around with our cache inode, they should have their own igrab, too. > > > Another point is that when we fail on btrfs_start_dirty_block_groups(), > > btrfs_commit_transaction() won't get to cleanup_transaction error > > handling, > > Right, because we don't actually finish the commit. Someone will eventually > though ;) Hmm yes, it's possible that there's a concurrent commit transaction running. If that's not true, we may still resort to btrfs_error_commit_super(), other than that, I don't see who could commit/cleanup the transaction after entering into BTRFS_FS_STATE_ERROR state. Thanks, -liubo > > > > > btrfs_commit_transaction() { > > ... > > if (!test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &cur_trans->flags)) { > > ... > > ret = btrfs_start_dirty_block_groups(trans, root); > > } > > if (ret) { > > btrfs_end_transaction(trans, root); > > return ret; > > } > > ... > > cleanup_transaction(); > > } > > > > > > But yes, if we delay the cleanup, we still have a chance to do cleanup > > in btrfs_error_commit_super(), and I have sent another patch to add > > several ASSERT()s to check block group related memory leak, with which > > we'll be warned if anything wrong. > > > > I'm OK to remove the part that causes concerns. > > Thanks. > > -chris > -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Thu, Jul 21, 2016 at 02:33:19PM -0700, Liu Bo wrote: > > > update_block_group() is the only producer to add block group cache to > > > dirty_bgs list, and if btrfs_run_delayed_refs() aborts, the transaction > > > is aborted, so seems that there won't be anyone manipulating dirty_bgs > > > list, am I missing? > > > > > > > No, the dirty_bgs processing is safe I think. My concern is with the cache > > inode which we iput() > > I think iput() is OK, we're doing iput() on block group cache on the io_bgs > list, where all block groups's inodes has been igrab()'d. If others are > messing around with our cache inode, they should have their own igrab, > too. > > > > > > Another point is that when we fail on btrfs_start_dirty_block_groups(), > > > btrfs_commit_transaction() won't get to cleanup_transaction error > > > handling, > > > > Right, because we don't actually finish the commit. Someone will eventually > > though ;) > > Hmm yes, it's possible that there's a concurrent commit transaction > running. If that's not true, we may still resort to > btrfs_error_commit_super(), other than that, I don't see who could > commit/cleanup the transaction after entering into BTRFS_FS_STATE_ERROR > state. What's the resume of this patch? I don't see a followup patch or a (to me) clear yes/no whether to merge it. Please let me know, thanks. -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Hi David, On Fri, Aug 19, 2016 at 01:28:23PM +0200, David Sterba wrote: > On Thu, Jul 21, 2016 at 02:33:19PM -0700, Liu Bo wrote: > > > > update_block_group() is the only producer to add block group cache to > > > > dirty_bgs list, and if btrfs_run_delayed_refs() aborts, the transaction > > > > is aborted, so seems that there won't be anyone manipulating dirty_bgs > > > > list, am I missing? > > > > > > > > > > No, the dirty_bgs processing is safe I think. My concern is with the cache > > > inode which we iput() > > > > I think iput() is OK, we're doing iput() on block group cache on the io_bgs > > list, where all block groups's inodes has been igrab()'d. If others are > > messing around with our cache inode, they should have their own igrab, > > too. > > > > > > > > > Another point is that when we fail on btrfs_start_dirty_block_groups(), > > > > btrfs_commit_transaction() won't get to cleanup_transaction error > > > > handling, > > > > > > Right, because we don't actually finish the commit. Someone will eventually > > > though ;) > > > > Hmm yes, it's possible that there's a concurrent commit transaction > > running. If that's not true, we may still resort to > > btrfs_error_commit_super(), other than that, I don't see who could > > commit/cleanup the transaction after entering into BTRFS_FS_STATE_ERROR > > state. > > What's the resume of this patch? I don't see a followup patch or a (to > me) clear yes/no whether to merge it. Please let me know, thanks. I'm going to update the patch to remove btrfs_start_dirty_block_groups() part. Thanks, -liubo -- To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c index db53eb8..4c110de 100644 --- a/fs/btrfs/disk-io.c +++ b/fs/btrfs/disk-io.c @@ -4434,9 +4434,80 @@ again: return 0; } +static void btrfs_cleanup_bg_io(struct btrfs_block_group_cache *cache) +{ + struct inode *inode; + + inode = cache->io_ctl.inode; + if (inode) { + invalidate_inode_pages2(inode->i_mapping); + BTRFS_I(inode)->generation = 0; + cache->io_ctl.inode = NULL; + iput(inode); + } + btrfs_put_block_group(cache); +} + +void btrfs_cleanup_dirty_bgs(struct btrfs_transaction *cur_trans, + struct btrfs_root *root) +{ + struct btrfs_block_group_cache *cache; + + spin_lock(&cur_trans->dirty_bgs_lock); + while (!list_empty(&cur_trans->dirty_bgs)) { + cache = list_first_entry(&cur_trans->dirty_bgs, + struct btrfs_block_group_cache, + dirty_list); + if (!cache) { + btrfs_err(root->fs_info, + "orphan block group dirty_bgs list"); + spin_unlock(&cur_trans->dirty_bgs_lock); + return; + } + + if (!list_empty(&cache->io_list)) { + spin_unlock(&cur_trans->dirty_bgs_lock); + list_del_init(&cache->io_list); + btrfs_cleanup_bg_io(cache); + spin_lock(&cur_trans->dirty_bgs_lock); + } + + list_del_init(&cache->dirty_list); + spin_lock(&cache->lock); + cache->disk_cache_state = BTRFS_DC_ERROR; + spin_unlock(&cache->lock); + + spin_unlock(&cur_trans->dirty_bgs_lock); + btrfs_put_block_group(cache); + spin_lock(&cur_trans->dirty_bgs_lock); + } + spin_unlock(&cur_trans->dirty_bgs_lock); + + while (!list_empty(&cur_trans->io_bgs)) { + cache = list_first_entry(&cur_trans->io_bgs, + struct btrfs_block_group_cache, + io_list); + if (!cache) { + btrfs_err(root->fs_info, + "orphan block group on io_bgs list"); + return; + } + + list_del_init(&cache->io_list); + spin_lock(&cache->lock); + cache->disk_cache_state = BTRFS_DC_ERROR; + spin_unlock(&cache->lock); + btrfs_cleanup_bg_io(cache); + } +} + void btrfs_cleanup_one_transaction(struct btrfs_transaction *cur_trans, struct btrfs_root *root) { + btrfs_cleanup_dirty_bgs(cur_trans, root); + ASSERT(list_empty(&cur_trans->dirty_bgs)); + ASSERT(list_empty(&cur_trans->io_bgs)); + btrfs_destroy_delayed_refs(cur_trans, root); cur_trans->state = TRANS_STATE_COMMIT_START; diff --git a/fs/btrfs/disk-io.h b/fs/btrfs/disk-io.h index acba821..6201663 100644 --- a/fs/btrfs/disk-io.h +++ b/fs/btrfs/disk-io.h @@ -133,6 +133,8 @@ int btrfs_init_log_root_tree(struct btrfs_trans_handle *trans, struct btrfs_fs_info *fs_info); int btrfs_add_log_tree(struct btrfs_trans_handle *trans, struct btrfs_root *root); +void btrfs_cleanup_dirty_bgs(struct btrfs_transaction *trans, + struct btrfs_root *root); void btrfs_cleanup_one_transaction(struct btrfs_transaction *trans, struct btrfs_root *root); struct btrfs_root *btrfs_create_tree(struct btrfs_trans_handle *trans, diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c index 50bd683..7a35c9d 100644 --- a/fs/btrfs/extent-tree.c +++ b/fs/btrfs/extent-tree.c @@ -3698,6 +3698,8 @@ again: goto again; } spin_unlock(&cur_trans->dirty_bgs_lock); + } else if (ret < 0) { + btrfs_cleanup_dirty_bgs(cur_trans, root); } btrfs_free_path(path);
While processing delayed refs, we may update block group's statistics and attach it to cur_trans->dirty_bgs, and later writing dirty block groups will process the list, which happens during btrfs_commit_transaction(). For whatever reason, the transaction is aborted and dirty_bgs is not processed in cleanup_transaction(), we end up with memory leak of these dirty block group cache. Since btrfs_start_dirty_block_groups() doesn't make it go to the commit critical section, this also adds the cleanup work inside it. Signed-off-by: Liu Bo <bo.li.liu@oracle.com> --- v2: - remove the 'return' when dirty_bgs is empty because there might be block group in list io_bgs. - remove unnecessary '\n' in btrfs_err(). - more commit log. fs/btrfs/disk-io.c | 71 ++++++++++++++++++++++++++++++++++++++++++++++++++ fs/btrfs/disk-io.h | 2 ++ fs/btrfs/extent-tree.c | 2 ++ 3 files changed, 75 insertions(+)