| Message ID | 20200313192848.140759-1-josef@toxicpanda.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | btrfs: force chunk allocation if our global rsv is larger than metadata | expand |
On 13.03.20 г. 21:28 ч., Josef Bacik wrote: > Nikolay noticed a bunch of test failures with my global rsv steal > patches. At first he thought they were introduced by them, but they've > been failing for a while with 64k nodes. > > The problem is with 64k nodes we have a global reserve that calculates > out to 13mib on a freshly made file system, which only has 8mib of > metadata space. Because of changes I previously made we no longer > account for the global reserve in the overcommit logic, which means we > correctly allow overcommit to happen even though we are already > overcommitted. > > However in some corner cases, for example btrfs/170, we will allocate > the entire file system up with data chunks before we have enough space > pressure to allocate a metadata chunk. Then once the fs is full we > ENOSPC out because we cannot overcommit and the global reserve is taking > up all of the available space. > > The most ideal way to deal with this is to change our space reservation > stuff to take into account the height of the tree's that we're > modifying, so that our global reserve calculation does not end up so > obscenely large. > > However that is a huuuuuuge undertaking. Instead fix this by forcing a > chunk allocation if the global reserve is larger than the total metadata > space. This gives us essentially the same behavior that happened > before, we get a chunk allocated and these tests can pass. > > This is meant to be a stop-gap measure until we can tackle the "tree > height only" project. > > Fixes: 0096420adb03 ("btrfs: do not account global reserve in can_overcommit") > Signed-off-by: Josef Bacik <josef@toxicpanda.com> Tested-by: Nikolay Borisov <nborisov@suse.com> Reviewed-by: Nikolay Borisov <nborisov@suse.com>
On Fri, Mar 13, 2020 at 03:28:48PM -0400, Josef Bacik wrote: > Nikolay noticed a bunch of test failures with my global rsv steal > patches. At first he thought they were introduced by them, but they've > been failing for a while with 64k nodes. > > The problem is with 64k nodes we have a global reserve that calculates > out to 13mib on a freshly made file system, which only has 8mib of > metadata space. Because of changes I previously made we no longer > account for the global reserve in the overcommit logic, which means we > correctly allow overcommit to happen even though we are already > overcommitted. > > However in some corner cases, for example btrfs/170, we will allocate > the entire file system up with data chunks before we have enough space > pressure to allocate a metadata chunk. Then once the fs is full we > ENOSPC out because we cannot overcommit and the global reserve is taking > up all of the available space. > > The most ideal way to deal with this is to change our space reservation > stuff to take into account the height of the tree's that we're > modifying, so that our global reserve calculation does not end up so > obscenely large. > > However that is a huuuuuuge undertaking. Instead fix this by forcing a > chunk allocation if the global reserve is larger than the total metadata > space. This gives us essentially the same behavior that happened > before, we get a chunk allocated and these tests can pass. > > This is meant to be a stop-gap measure until we can tackle the "tree > height only" project. > > Fixes: 0096420adb03 ("btrfs: do not account global reserve in can_overcommit") > Signed-off-by: Josef Bacik <josef@toxicpanda.com> Added to misc-next, thanks.
diff --git a/fs/btrfs/block-rsv.c b/fs/btrfs/block-rsv.c index e46dc3688983..2e33a31dfc8e 100644 --- a/fs/btrfs/block-rsv.c +++ b/fs/btrfs/block-rsv.c @@ -5,6 +5,7 @@ #include "block-rsv.h" #include "space-info.h" #include "transaction.h" +#include "block-group.h" /* * HOW DO BLOCK RESERVES WORK @@ -405,6 +406,8 @@ void btrfs_update_global_block_rsv(struct btrfs_fs_info *fs_info) else block_rsv->full = 0; + if (block_rsv->size >= sinfo->total_bytes) + sinfo->force_alloc = CHUNK_ALLOC_FORCE; spin_unlock(&block_rsv->lock); spin_unlock(&sinfo->lock); } diff --git a/fs/btrfs/transaction.c b/fs/btrfs/transaction.c index d171fd52c82b..304606c911e8 100644 --- a/fs/btrfs/transaction.c +++ b/fs/btrfs/transaction.c @@ -21,6 +21,7 @@ #include "dev-replace.h" #include "qgroup.h" #include "block-group.h" +#include "space-info.h" #define BTRFS_ROOT_TRANS_TAG 0 @@ -519,6 +520,7 @@ start_transaction(struct btrfs_root *root, unsigned int num_items, u64 num_bytes = 0; u64 qgroup_reserved = 0; bool reloc_reserved = false; + bool do_chunk_alloc = false; int ret; /* Send isn't supposed to start transactions. */ @@ -581,6 +583,9 @@ start_transaction(struct btrfs_root *root, unsigned int num_items, delayed_refs_bytes); num_bytes -= delayed_refs_bytes; } + + if (rsv->space_info->force_alloc) + do_chunk_alloc = true; } else if (num_items == 0 && flush == BTRFS_RESERVE_FLUSH_ALL && !delayed_refs_rsv->full) { /* @@ -663,6 +668,19 @@ start_transaction(struct btrfs_root *root, unsigned int num_items, if (!current->journal_info) current->journal_info = h; + + /* + * If the space_info is marked ALLOC_FORCE then we'll get upgraded to + * ALLOC_FORCE the first run through, and then we won't allocate for + * anybody else who races in later. We don't care about the return + * value here. + */ + if (do_chunk_alloc && num_bytes) { + u64 flags = h->block_rsv->space_info->flags; + btrfs_chunk_alloc(h, btrfs_get_alloc_profile(fs_info, flags), + CHUNK_ALLOC_NO_FORCE); + } + return h; join_fail:
Nikolay noticed a bunch of test failures with my global rsv steal patches. At first he thought they were introduced by them, but they've been failing for a while with 64k nodes. The problem is with 64k nodes we have a global reserve that calculates out to 13mib on a freshly made file system, which only has 8mib of metadata space. Because of changes I previously made we no longer account for the global reserve in the overcommit logic, which means we correctly allow overcommit to happen even though we are already overcommitted. However in some corner cases, for example btrfs/170, we will allocate the entire file system up with data chunks before we have enough space pressure to allocate a metadata chunk. Then once the fs is full we ENOSPC out because we cannot overcommit and the global reserve is taking up all of the available space. The most ideal way to deal with this is to change our space reservation stuff to take into account the height of the tree's that we're modifying, so that our global reserve calculation does not end up so obscenely large. However that is a huuuuuuge undertaking. Instead fix this by forcing a chunk allocation if the global reserve is larger than the total metadata space. This gives us essentially the same behavior that happened before, we get a chunk allocated and these tests can pass. This is meant to be a stop-gap measure until we can tackle the "tree height only" project. Fixes: 0096420adb03 ("btrfs: do not account global reserve in can_overcommit") Signed-off-by: Josef Bacik <josef@toxicpanda.com> --- fs/btrfs/block-rsv.c | 3 +++ fs/btrfs/transaction.c | 18 ++++++++++++++++++ 2 files changed, 21 insertions(+)