- 05 Nov, 2019 26 commits
-
-
Theodore Ts'o authored
-
Jan Kara authored
Currently we reserve j_max_transaction_buffers / 32 for transaction descriptor blocks. Now that revoke descriptors are accounted for separately this estimate is unnecessarily high and we can actually compute much tighter estimate. In the common case of 32k journal blocks and 4k blocksize this actually reduces the amount of reserved descriptor blocks from 256 to ~25 which allows us to fit more real data into a transaction. Signed-off-by:
Jan Kara <jack@suse.cz> Link: https://lore.kernel.org/r/20191105164437.32602-25-jack@suse.czSigned-off-by:
Theodore Ts'o <tytso@mit.edu>
-
Jan Kara authored
Provide trace event for handle restarts to ease debugging. Reviewed-by:
-
Jan Kara authored
So far we have reserved only relatively high fixed amount of revoke credits for each transaction. We over-reserved by large amount for most cases but when freeing large directories or files with data journalling, the fixed amount is not enough. In fact the worst case estimate is inconveniently large (maximum extent size) for freeing of one extent. We fix this by doing proper estimate of the amount of blocks that need to be revoked when removing blocks from the inode due to truncate or hole punching and otherwise reserve just a small amount of revoke credits for each transaction to accommodate freeing of xattrs block or so. Signed-off-by:
-
Jan Kara authored
Make checking of available credits in jbd2_journal_dirty_metadata() more strict. There should be always enough credits in the handle to write all potential revoke descriptors. Also we warn in case there are not enough credits since this is a bug in the filesystem. Reviewed-by:
-
Jan Kara authored
The credit counter now contains both buffer and revoke descriptor block credits. Rename to counter to h_total_credits to reflect that. No functional change. Reviewed-by:
-
Jan Kara authored
Extend functions for starting, extending, and restarting transaction handles to take number of revoke records handle must be able to accommodate. These functions then make sure transaction has enough credits to be able to store resulting revoke descriptor blocks. Also revoke code tracks number of revoke records created by a handle to catch situation where some place didn't reserve enough space for revoke records. Similarly to standard transaction credits, space for unused reserved revoke records is released when the handle is stopped. On the ext4 side we currently take a simplistic approach of reserving space for 1024 revoke records for any transaction. This grows amount of credits reserved for each handle only by a few and is enough for any normal workload so that we don't hit warnings in jbd2. We will refine the logic in following commits. Signed-off-by:
-
Jan Kara authored
The function is now just a trivial wrapper returning journal->j_max_transaction_buffers. Drop it. Reviewed-by:
-
Jan Kara authored
Currently, journal descriptor blocks were not accounted in transaction->t_outstanding_credits and we were just leaving some slack space in the journal for them (in jbd2_log_space_left() and jbd2_space_needed()). This is making proper accounting (and reservation we want to add) of descriptor blocks difficult so switch to accounting descriptor blocks in transaction->t_outstanding_credits and just reserve the same amount of credits in t_outstanding credits for journal descriptor blocks when creating transaction. Signed-off-by:
-
Jan Kara authored
jbd2__journal_restart() has quite some code that is common with jbd2_journal_stop(). Factor this functionality into stop_this_handle() helper and use it from both functions. Note that this also drops t_handle_lock protection from jbd2__journal_restart() as jbd2_journal_stop() does the same thing without it. Signed-off-by:
-
Jan Kara authored
When we drop last handle from a transaction and journal->j_barrier_count > 0, jbd2_journal_stop() wakes up journal->j_wait_transaction_locked wait queue. This looks pointless - wait for outstanding handles always happens on journal->j_wait_updates waitqueue. journal->j_wait_transaction_locked is used to wait for transaction state changes and by start_this_handle() for waiting until journal->j_barrier_count drops to 0. The first case is clearly irrelevant here since only jbd2 thread changes transaction state. The second case looks related but jbd2_journal_unlock_updates() is responsible for the wakeup in this case. So just drop the wakeup. Reviewed-by:
-
Jan Kara authored
If a transaction is larger than journal->j_max_transaction_buffers, that is a bug and not a trigger for transaction commit. Also the very next attempt to start new handle will start transaction commit anyway. So just remove the pointless check. Arguably, we could start transaction commit whenever the transaction size is *close* to journal->j_max_transaction_buffers. This has a potential to reduce latency of the next jbd2_journal_start() at the cost of somewhat smaller transactions. However for this to have any effect, it would mean that there isn't someone already waiting in jbd2_journal_start() which means metadata load for the fs is pretty light anyway so probably this optimization is not worth it. Reviewed-by:
-
Jan Kara authored
Move code in jbd2_journal_stop() around a bit. It removes some unnecessary code duplication and will make factoring out parts common with jbd2__journal_restart() easier. Reviewed-by:
-
Jan Kara authored
jbd2 statistics counting number of blocks logged in a transaction was wrong. It didn't count the commit block and more importantly it didn't count revoke descriptor blocks. Make sure these get properly counted. Reviewed-by:
-
Jan Kara authored
Use the jbd2 accessor function for h_buffer_credits. Reviewed-by:
-
Jan Kara authored
Provide accessor function to get number of credits available in a handle and use it from ext4. Later, computation of available credits won't be so straightforward. Reviewed-by:
-
Jan Kara authored
Provide ext4_journal_ensure_credits_fn() function to ensure transaction has given amount of credits and call helper function to prepare for restarting a transaction. This allows to remove some boilerplate code from various places, add proper error handling for the case where transaction extension or restart fails, and reduces following changes needed for proper revoke record reservation tracking. Signed-off-by:
-
Jan Kara authored
Error cleanup path in ext4_alloc_branch() calls ext4_forget() on freshly allocated indirect blocks with 'metadata' set to 1. This results in generating revoke records for these blocks. However this is unnecessary as the freed blocks are only allocated in the current transaction and thus they will never be journalled. Make this cleanup path similar to e.g. cleanup in ext4_splice_branch() and use ext4_free_blocks() to handle block forgetting by passing EXT4_FREE_BLOCKS_FORGET and not EXT4_FREE_BLOCKS_METADATA to ext4_free_blocks(). This also allows allocating transaction not to reserve any credits for revoke records. Reviewed-by:
-
Jan Kara authored
Use ext4 helper ext4_journal_extend() instead of opencoding it in ext4_try_to_expand_extra_isize(). Reviewed-by:
-
Jan Kara authored
Similarly to directories, EA inodes do only journalled modifications to their data. Change ext4_should_journal_data() to return true for them so that we don't have to special-case them during truncate. Signed-off-by:
-
Jan Kara authored
Estimate for the number of credits needed for final freeing of inode in ext4_evict_inode() was to small. We may modify 4 blocks (inode & sb for orphan deletion, bitmap & group descriptor for inode freeing) and not just 3. [ Fixed minor whitespace nit. -- TYT ] Fixes: e50e5129 ("ext4: xattr-in-inode support") CC: stable@vger.kernel.org Signed-off-by:
-
Jan Kara authored
When ext4_mkdir(), ext4_symlink(), ext4_create(), or ext4_mknod() fail to add entry into directory, it ends up dropping freshly created inode under the running transaction and thus inode truncation happens under that transaction. That breaks assumptions that evict() does not get called from a transaction context and at least in ext4_symlink() case it can result in inode eviction deadlocking in inode_wait_for_writeback() when flush worker finds symlink inode, starts to write it back and blocks on starting a transaction. So change the code in ext4_mkdir() and ext4_add_nondir() to drop inode reference only after the transaction is stopped. We also have to add inode to the orphan list in that case as otherwise the inode would get leaked in case we crash before inode deletion is committed. CC: stable@vger.kernel.org Signed-off-by:
-
Jan Kara authored
Every caller of ext4_add_nondir() marks handle as sync if directory has DIRSYNC set. Move this marking to ext4_add_nondir() so reduce some duplication. Signed-off-by:
-
Jan Kara authored
With 32-bit block numbers, we don't allocate the array for journal buffer heads large enough for corresponding descriptor tags to fill the descriptor block. Thus we end up writing out half-full descriptor blocks to the journal unnecessarily growing the transaction. Fix the logic to allocate the array large enough. Signed-off-by:
-
Jan Kara authored
jbd2_journal_next_log_block() does not look at transaction->t_outstanding_credits. Remove the misleading comment. Reviewed-by:
-
Jan Kara authored
When number of free space in the journal is very low, the arithmetic in jbd2_log_space_left() could underflow resulting in very high number of free blocks and thus triggering assertion failure in transaction commit code complaining there's not enough space in the journal: J_ASSERT(journal->j_free > 1); Properly check for the low number of free blocks. CC: stable@vger.kernel.org Reviewed-by:
-
- 01 Nov, 2019 1 commit
-
-
Theodore Ts'o authored
-
- 22 Oct, 2019 5 commits
-
-
Ritesh Harjani authored
All support is now added for blocksize < pagesize for dioread_nolock. This patch removes those checks which disables dioread_nolock feature for blocksize != pagesize. Signed-off-by:
Ritesh Harjani <riteshh@linux.ibm.com> Link: https://lore.kernel.org/r/20191016073711.4141-6-riteshh@linux.ibm.comSigned-off-by:
-
Ritesh Harjani authored
This patch adds the support for blocksize < pagesize for dioread_nolock feature. Since in case of blocksize < pagesize, we can have multiple small buffers of page as unwritten extents, we need to maintain a vector of these unwritten extents which needs the conversion after the IO is complete. Thus, we maintain a list of tuple <offset, size> pair (io_end_vec) for this & traverse this list to do the unwritten to written conversion. Signed-off-by:
-
Ritesh Harjani authored
This patch refactors mpage_map_and_submit_buffers to take out the page buffers processing, as a separate function. This will be required to add support for blocksize < pagesize for dioread_nolock feature. No functionality change in this patch. Signed-off-by:
-
Ritesh Harjani authored
This patch just brings in the API for conversion of unwritten io_end_vec extents which will be required for blocksize < pagesize support for dioread_nolock feature. No functional changes in this patch. Signed-off-by:
-
Ritesh Harjani authored
Let's keep uniform naming convention for ext4_submit_io (io) & ext4_end_io_t (io_end) structures, to avoid any confusion. No functionality change in this patch. Signed-off-by:
-
- 21 Oct, 2019 7 commits
-
-
Thomas Gleixner authored
On PREEMPT_RT bit-spinlocks have the same semantics as on PREEMPT_RT=n, i.e. they disable preemption. That means functions which are not safe to be called in preempt disabled context on RT trigger a might_sleep() assert. The journal head bit spinlock is mostly held for short code sequences with trivial RT safe functionality, except for one place: jbd2_journal_put_journal_head() invokes __journal_remove_journal_head() with the journal head bit spinlock held. __journal_remove_journal_head() invokes kmem_cache_free() which must not be called with preemption disabled on RT. Jan suggested to rework the removal function so the actual free happens outside the bit-spinlocked region. Split it into two parts: - Do the sanity checks and the buffer head detach under the lock - Do the actual free after dropping the lock There is error case handling in the free part which needs to dereference the b_size field of the now detached buffer head. Due to paranoia (caused by ignorance) the size is retrieved in the detach function and handed into the free function. Might be over-engineered, but better safe than sorry. This makes the journal head bit-spinlock usage RT compliant and also avoids nested locking which is not covered by lockdep. Suggested-by:
Jan Kara <jack@suse.com> Signed-off-by:
Thomas Gleixner <tglx@linutronix.de> Cc: linux-ext4@vger.kernel.org Cc: "Theodore Ts'o" <tytso@mit.edu> Cc: Jan Kara <jack@suse.com> Signed-off-by:
-
Thomas Gleixner authored
Bit-spinlocks are problematic on PREEMPT_RT if functions which might sleep on RT, e.g. spin_lock(), alloc/free(), are invoked inside the lock held region because bit spinlocks disable preemption even on RT. A first attempt was to replace state lock with a spinlock placed in struct buffer_head and make the locking conditional on PREEMPT_RT and DEBUG_BIT_SPINLOCKS. Jan pointed out that there is a 4 byte hole in struct journal_head where a regular spinlock fits in and he would not object to convert the state lock to a spinlock unconditionally. Aside of solving the RT problem, this also gains lockdep coverage for the journal head state lock (bit-spinlocks are not covered by lockdep as it's hard to fit a lockdep map into a single bit). The trivial change would have been to convert the jbd_*lock_bh_state() inlines, but that comes with the downside that these functions take a buffer head pointer which needs to be converted to a journal head pointer which adds another level of indirection. As almost all functions which use this lock have a journal head pointer readily available, it makes more sense to remove the lock helper inlines and write out spin_*lock() at all call sites. Fixup all locking comments as well. Suggested-by:
-
Jan Kara authored
jbd2_journal_forget() jumps to 'not_jbd' branch which calls __bforget() in cases where the buffer is clean which is pointless. In case of failed assertion, it can be even argued that it is safer not to touch buffer's dirty bits. Also logically it makes more sense to just jump to 'drop' and that will make logic also simpler when we switch bh_state_lock to a spinlock. Signed-off-by:
-
Jan Kara authored
We have cleared both dirty & jbddirty bits from the bh. So there's no difference between bforget() and brelse(). Thus there's no point jumping to no_jbd branch. Signed-off-by:
-
Jan Kara authored
__jbd2_journal_unfile_buffer() and __jbd2_journal_refile_buffer() drop transaction's jh reference when they remove jh from a transaction. This will be however inconvenient once we move state lock into journal_head itself as we still need to unlock it and we'd need to grab jh reference just for that. Move dropping of jh reference out of these functions into the few callers. Signed-off-by:
-
Thomas Gleixner authored
No users. Signed-off-by:
-
Thomas Gleixner authored
journal_unmap_buffer() checks first whether the buffer head is a journal. If so it takes locks and then invokes jbd2_journal_grab_journal_head() followed by another check whether this is journal head buffer. The double checking is pointless. Replace the initial check with jbd2_journal_grab_journal_head() which alredy checks whether the buffer head is actually a journal. Allows also early access to the journal head pointer for the upcoming conversion of state lock to a regular spinlock. Signed-off-by:
-
- 13 Oct, 2019 1 commit
-
-
Linus Torvalds authored
-
