On Fri, Feb 24, 2017 at 2:06 AM, Qu Wenruo <quwen...@cn.fujitsu.com> wrote:
> If run btrfs/125 with nospace_cache or space_cache=v2 mount option,
> btrfs will block with the following backtrace:
Happens with btrfs/124 without any mount options too.
>
> Call Trace:
> __schedule+0x2d4/0xae0
> schedule+0x3d/0x90
> btrfs_start_ordered_extent+0x160/0x200 [btrfs]
> ? wake_atomic_t_function+0x60/0x60
> btrfs_run_ordered_extent_work+0x25/0x40 [btrfs]
> btrfs_scrubparity_helper+0x1c1/0x620 [btrfs]
> btrfs_flush_delalloc_helper+0xe/0x10 [btrfs]
> process_one_work+0x2af/0x720
> ? process_one_work+0x22b/0x720
> worker_thread+0x4b/0x4f0
> kthread+0x10f/0x150
> ? process_one_work+0x720/0x720
> ? kthread_create_on_node+0x40/0x40
> ret_from_fork+0x2e/0x40
>
> The direct cause is the error handler in run_delalloc_nocow() doesn't
> handle error from btrfs_reloc_clone_csums() well.
>
> The error handler of run_delalloc_nocow() will clear dirty and finish IO
> for the pages in that extent.
> However we have already inserted one ordered extent.
> And that ordered extent is relying on endio hooks to wait all its pages
> to finish, while only the first page will finish.
>
> This makes that ordered extent never finish, so blocking the file
> system.
>
> Although the root cause is still in RAID5/6, it won't hurt to fix the
> error routine first.
This is a higher level problem independent of the raid level being
used. So this changelog gives the wrong idea that this happens only
with raid5/6.
>
> This patch will slightly modify one existing function,
> btrfs_endio_direct_write_update_ordered() to handle free space inode,
> and skip releasing metadata, which will be handled by
> extent_clear_unlock_delalloc().
>
> And use it as base to implement one inline function,
> btrfs_cleanup_ordered_extents() to handle the error in
> run_delalloc_nocow() and cow_file_range().
>
> Also, extent_clear_unlock_delalloc() will handle all the metadata
> release, so btrfs_cleanup_ordered_extents() doesn't need to do it.
>
> For compression, it's calling writepage_end_io_hook() itself to handle
> its error, and any submitted ordered extent will have its bio submitted,
> so no need to worry about compression part.
>
> Suggested-by: Filipe Manana <fdman...@suse.com>
> Signed-off-by: Qu Wenruo <quwen...@cn.fujitsu.com>
> ---
> v2:
> Add BTRFS_ORDERED_SKIP_METADATA flag to avoid double reducing
> outstanding extents, which is already done by
> extent_clear_unlock_delalloc() with EXTENT_DO_ACCOUNT control bit
> v3:
> Skip first page to avoid underflow ordered->bytes_left.
> Fix range passed in cow_file_range() which doesn't cover the whole
> extent.
> Expend extent_clear_unlock_delalloc() range to allow them to handle
> metadata release.
> v4:
> Don't use extra bit to skip metadata freeing for ordered extent,
> but only handle btrfs_reloc_clone_csums() error just before processing
> next extent.
> This makes error handle much easier for run_delalloc_nocow().
> ---
> fs/btrfs/extent_io.c | 1 -
> fs/btrfs/inode.c | 112
> +++++++++++++++++++++++++++++++++++++++++++--------
> 2 files changed, 96 insertions(+), 17 deletions(-)
>
> diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c
> index 4ac383a3a649..a14d1b0840c5 100644
> --- a/fs/btrfs/extent_io.c
> +++ b/fs/btrfs/extent_io.c
> @@ -3258,7 +3258,6 @@ static noinline_for_stack int writepage_delalloc(struct
> inode *inode,
> delalloc_end,
> &page_started,
> nr_written);
> - /* File system has been set read-only */
This is unrelated to the problem you're solving. Should be an
independent and unrelated patch.
> if (ret) {
> SetPageError(page);
> /* fill_delalloc should be return < 0 for error
> diff --git a/fs/btrfs/inode.c b/fs/btrfs/inode.c
> index 1e861a063721..b98a92807aa2 100644
> --- a/fs/btrfs/inode.c
> +++ b/fs/btrfs/inode.c
> @@ -116,6 +116,34 @@ static struct extent_map *create_pinned_em(struct inode
> *inode, u64 start,
>
> static int btrfs_dirty_inode(struct inode *inode);
>
> +
> +static void __endio_write_update_ordered(struct inode *inode,
> + const u64 offset, const u64 bytes,
> + bool uptodate, bool cleanup);
Offset and bytes are const, why not the booleans as well? It's
inconsistent and confusing.
Also leaving a blank line before the declaration of next function
would make the code easier to the eyes.
> +static inline void btrfs_endio_direct_write_update_ordered(struct inode
> *inode,
> + const u64 offset,
> + const u64 bytes,
> + const int uptodate)
Now uptodate is an int and not bool as in the previous call that it
ends up calling. At least within the same patch we should be
consistent.
> +{
> + return __endio_write_update_ordered(inode, offset, bytes, uptodate,
> false);
> +}
> +
> +/*
> + * Cleanup all submitted ordered extent in specified range to handle error
extent -> extents
> + * in cow_file_range() and run_delalloc_nocow().
> + * Compression handles error and ordered extent submission all by themselves,
all by themselves -> by itself
> + * so no need to call this function.
> + *
> + * NOTE: caller must ensure extent_clear_unlock_delalloc() in error handler
> + * doesn't cover any range of submitted ordered extent.
> + * Or we will double free metadata for submitted ordered extent.
> + */
> +static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
> + u64 offset, u64 bytes)
> +{
> + return __endio_write_update_ordered(inode, offset, bytes, false,
> true);
> +}
> +
> #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
> void btrfs_test_inode_set_ops(struct inode *inode)
> {
> @@ -950,6 +978,7 @@ static noinline int cow_file_range(struct inode *inode,
> u64 disk_num_bytes;
> u64 cur_alloc_size;
> u64 blocksize = fs_info->sectorsize;
> + u64 orig_start = start;
> struct btrfs_key ins;
> struct extent_map *em;
> struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
> @@ -1052,15 +1081,22 @@ static noinline int cow_file_range(struct inode
> *inode,
> BTRFS_DATA_RELOC_TREE_OBJECTID) {
> ret = btrfs_reloc_clone_csums(inode, start,
> cur_alloc_size);
> + /*
> + * Only drop cache here, and process as normal.
> + *
> + * We must not allow extent_clear_unlock_delalloc()
> + * to free meta of this ordered extent, as its
> + * meta should be freed by btrfs_finish_ordered_io().
> + *
> + * So we must continue until processing next extent
> + */
> if (ret)
> - goto out_drop_extent_cache;
> + btrfs_drop_extent_cache(inode, start,
> + start + ram_size - 1, 0);
> }
>
> btrfs_dec_block_group_reservations(fs_info, ins.objectid);
>
> - if (disk_num_bytes < cur_alloc_size)
> - break;
> -
If I understanding correctly you're removing this just to make sure we
call extent_clear_unlock_delalloc() below.
However later on, we should still break if this condition is true,
otherwise this code is not equivalent to what we had before.
> /* we're not doing compressed IO, don't unlock the first
> * page (which the caller expects to stay locked), don't
> * clear any dirty bits and don't set any writeback bits
> @@ -1076,10 +1112,21 @@ static noinline int cow_file_range(struct inode
> *inode,
> delalloc_end, locked_page,
> EXTENT_LOCKED | EXTENT_DELALLOC,
> op);
> - disk_num_bytes -= cur_alloc_size;
> + if (disk_num_bytes < cur_alloc_size)
> + disk_num_bytes = 0;
> + else
> + disk_num_bytes -= cur_alloc_size;
> num_bytes -= cur_alloc_size;
> alloc_hint = ins.objectid + ins.offset;
> start += cur_alloc_size;
> +
> + /*
> + * btrfs_reloc_clone_csums() error, now
> + * extent_clear_unlock_delalloc() in out_unlock() won't
> + * double free metadata of current oredered extent.
in out_unlock() -> at out_lock label
(it's not a function, should not have the parenthesis following the name)
oredered -> ordered
> + */
> + if (ret)
> + goto out_reserve;
Jumping to this label will make us decrement block group reservations
twice - i.e. btrfs_dec_block_group_reservations() is now called twice
instead of once, allowing for an underflow to happen.
And btrfs_free_reserved_extent() will be called twice for the reserved
extent, once at the "out_reserve" label and once when the ordered
extent finishes (with the error bit set), at
btrfs_finish_ordered_io(). This is particularly dangerous, as we are
in a transaction-less context, so that means once the first free call
happens, the same space can be allocated elsewhere by another task and
then when the ordered extent finishes we incorrectly free that space
again.
> }
> out:
> return ret;
> @@ -1096,6 +1143,7 @@ static noinline int cow_file_range(struct inode *inode,
> EXTENT_DELALLOC | EXTENT_DEFRAG,
> PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
> PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
> + btrfs_cleanup_ordered_extents(inode, orig_start, end - orig_start +
> 1);
> goto out;
> }
>
> @@ -1496,15 +1544,14 @@ static noinline int run_delalloc_nocow(struct inode
> *inode,
> BUG_ON(ret); /* -ENOMEM */
>
> if (root->root_key.objectid ==
> - BTRFS_DATA_RELOC_TREE_OBJECTID) {
> + BTRFS_DATA_RELOC_TREE_OBJECTID)
> + /*
> + * Error handled later, as we must prevent
> + * extent_clear_unlock_delalloc() in error handler
> + * from freeing metadata of submitted ordered extent.
> + */
> ret = btrfs_reloc_clone_csums(inode, cur_offset,
> num_bytes);
> - if (ret) {
> - if (!nolock && nocow)
> - btrfs_end_write_no_snapshoting(root);
> - goto error;
> - }
> - }
>
> extent_clear_unlock_delalloc(inode, cur_offset,
> cur_offset + num_bytes - 1, end,
> @@ -1516,6 +1563,14 @@ static noinline int run_delalloc_nocow(struct inode
> *inode,
> if (!nolock && nocow)
> btrfs_end_write_no_snapshoting(root);
> cur_offset = extent_end;
> +
> + /*
> + * btrfs_reloc_clone_csums() error, now we're OK to call error
> + * handler, as metadata for submitted ordered extent will only
> + * be freed by btrfs_finish_ordered_io().
> + */
> + if (ret)
> + goto error;
> if (cur_offset > end)
> break;
> }
> @@ -1546,6 +1601,12 @@ static noinline int run_delalloc_nocow(struct inode
> *inode,
> PAGE_CLEAR_DIRTY |
> PAGE_SET_WRITEBACK |
> PAGE_END_WRITEBACK);
> + /*
> + * It's possible that last ordered extent covered the last part
> + * but failed. In that case we still need to clean them up.
> + */
Confusing comment. You mention ordered extent (singular) but then
later mention "clean them up" (plural). Also saying that an ordered
extent failed gives the idea that the IO failed (i.e. bio finished
with an error), while what you want to say is that we created one or
more ordered extents but failed somewhere before submitting bios for
them. I would just remove the comment, or the very least make it
understandable.
> + if (ret)
> + btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
> btrfs_free_path(path);
> return ret;
> }
> @@ -8185,17 +8246,36 @@ static void btrfs_endio_direct_read(struct bio *bio)
> bio_put(bio);
> }
>
> -static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
> - const u64 offset,
> - const u64 bytes,
> - const int uptodate)
> +static void __endio_write_update_ordered(struct inode *inode,
> + const u64 offset, const u64 bytes,
> + bool uptodate, bool cleanup)
> {
> struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
> struct btrfs_ordered_extent *ordered = NULL;
> + struct btrfs_workqueue *wq;
> + btrfs_work_func_t func;
> u64 ordered_offset = offset;
> u64 ordered_bytes = bytes;
> int ret;
>
> + if (btrfs_is_free_space_inode(inode)) {
> + wq = fs_info->endio_freespace_worker;
> + func = btrfs_freespace_write_helper;
> + } else {
> + wq = fs_info->endio_write_workers;
> + func = btrfs_endio_write_helper;
> + }
What is going on here? wq and func are set but not used anywhere.
> +
> + /*
> + * In cleanup case, the first page of the range will be handled
> + * by end_extent_writepage() under done tag of __extent_writepage().
This sentence is confusing, "under done tag"?
Just say:
In cleanup case, the first page of the range will be handled by
end_extent_writepage() when called from __extent_writepage().
thanks
> + *
> + * So we must skip first page, or we will underflow
> ordered->bytes_left
> + */
> + if (cleanup) {
> + ordered_offset += PAGE_SIZE;
> + ordered_bytes -= PAGE_SIZE;
> + }
> again:
> ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
> &ordered_offset,
> --
> 2.11.1
>
>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
