The do_get_write_access()(fs/jbd2/transaction.c) may have to wait on the buffer that is being flushed out by the committing transaction and has not yet made the copy-out by the current transaction. When such condition is true, do_get_write_access() has to wait until BH_Shadow is cleared, which could take a very long time because of waiting on IO. In our research*, we observed that this situation become worse on parallel-transactional workloads (Filbebench** OLTP and IOzone*** write tests) to the ext4 filesystem. This patch forces copy-out to enable, and can be configured on the fly via /sys/fs/ext4/<device>/force_copyout. Our experimental results*, with force copy-out, showed ~25% speedup on Filebench OLTP and ~3X speedup on IOzone. *Hung-Ching Chang, Bo Li, Godmar Back, Ali R. Butt, Kirk W. Cameron, "LUC: Limiting the Unintended Consequences of Power Scaling on Parallel Transaction-oriented Workloads," accepted by in proceedings of 29th IEEE International Parallel & Distributed Processing Symposium (IPDPS), Hyderabad, INDIA, 2015. **http://filebench.sourceforge.net/ ***http://www.iozone.org Signed-off-by: Hung-Ching Chang <hcchang@xxxxxx> --- Documentation/filesystems/ext4.txt | 11 +++++++++++ fs/ext4/ext4.h | 6 ++++++ fs/ext4/super.c | 26 ++++++++++++++++++++++++++ fs/jbd2/journal.c | 2 ++ include/linux/jbd2.h | 7 +++++++ 5 files changed, 52 insertions(+), 0 deletions(-) diff --git a/Documentation/filesystems/ext4.txt b/Documentation/filesystems/ext4.txt index 6c0108e..8282959 100644 --- a/Documentation/filesystems/ext4.txt +++ b/Documentation/filesystems/ext4.txt @@ -510,6 +510,17 @@ Files in /sys/fs/ext4/<devname> in the file system. If there is not enough space for the reserved space when mounting the file mount will _not_ fail. + + force_copyout=<0|1(*)> This enables/disables the force copyout in the + jbd code. force_copyout=0 disables, + force_copyout=1 enables. When enabled, the + committing transaction makes copy-outs for the + metadata for IO submission. This removes + potential locks when the current transaction + attempts to modify the metadata, which are also + owned by the committing transaction and are + being flushed out to non-volatile storage, but + copy-outs have not yet made for these metadata. .............................................................................. Ioctls diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h index f63c3d5..585836f 100644 --- a/fs/ext4/ext4.h +++ b/fs/ext4/ext4.h @@ -1247,6 +1247,7 @@ struct ext4_sb_info { unsigned long s_commit_interval; u32 s_max_batch_time; u32 s_min_batch_time; + u32 s_force_copyout; struct block_device *journal_bdev; #ifdef CONFIG_QUOTA char *s_qf_names[EXT4_MAXQUOTAS]; /* Names of quota files with journalled quota */ @@ -1620,6 +1621,11 @@ static inline void ext4_clear_state_flags(struct ext4_inode_info *ei) #define EXT4_DEF_MAX_BATCH_TIME 15000 /* 15ms */ /* + * Default force copyout is set to enable + */ +#define EXT4_DEF_FORCE_COPYOUT 1 + +/* * Minimum number of groups in a flexgroup before we separate out * directories into the first block group of a flexgroup */ diff --git a/fs/ext4/super.c b/fs/ext4/super.c index e061e66..e09f7d2 100644 --- a/fs/ext4/super.c +++ b/fs/ext4/super.c @@ -2508,6 +2508,27 @@ static ssize_t inode_readahead_blks_store(struct ext4_attr *a, return count; } +static ssize_t force_copyout_store(struct ext4_attr *a, + struct ext4_sb_info *sbi, + const char *buf, size_t count) +{ + journal_t *journal = sbi->s_journal; + unsigned long t; + int ret; + + ret = kstrtoul(skip_spaces(buf), 0, &t); + if (ret) + return ret; + + if (t >= 2) + return -EINVAL; + + sbi->s_force_copyout = t; + if (journal) + journal->j_force_copyout = sbi->s_force_copyout; + return count; +} + static ssize_t sbi_ui_show(struct ext4_attr *a, struct ext4_sb_info *sbi, char *buf) { @@ -2652,6 +2673,8 @@ EXT4_RW_ATTR_SBI_UI(msg_ratelimit_burst, s_msg_ratelimit_state.burst); EXT4_RO_ATTR_ES_UI(errors_count, s_error_count); EXT4_RO_ATTR_ES_UI(first_error_time, s_first_error_time); EXT4_RO_ATTR_ES_UI(last_error_time, s_last_error_time); +EXT4_ATTR_OFFSET(force_copyout, 0644, sbi_ui_show, + force_copyout_store, s_force_copyout); static struct attribute *ext4_attrs[] = { ATTR_LIST(delayed_allocation_blocks), @@ -2678,6 +2701,7 @@ static struct attribute *ext4_attrs[] = { ATTR_LIST(errors_count), ATTR_LIST(first_error_time), ATTR_LIST(last_error_time), + ATTR_LIST(force_copyout), NULL, }; @@ -3444,6 +3468,7 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent) sb->s_fs_info = sbi; sbi->s_sb = sb; sbi->s_inode_readahead_blks = EXT4_DEF_INODE_READAHEAD_BLKS; + sbi->s_force_copyout = EXT4_DEF_FORCE_COPYOUT; sbi->s_sb_block = sb_block; if (sb->s_bdev->bd_part) sbi->s_sectors_written_start = @@ -4321,6 +4346,7 @@ static void ext4_init_journal_params(struct super_block *sb, journal_t *journal) journal->j_commit_interval = sbi->s_commit_interval; journal->j_min_batch_time = sbi->s_min_batch_time; journal->j_max_batch_time = sbi->s_max_batch_time; + journal->j_force_copyout = sbi->s_force_copyout; write_lock(&journal->j_state_lock); if (test_opt(sb, BARRIER)) diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c index b96bd80..2637234 100644 --- a/fs/jbd2/journal.c +++ b/fs/jbd2/journal.c @@ -396,6 +396,8 @@ repeat: new_page = virt_to_page(jh_in->b_frozen_data); new_offset = offset_in_page(jh_in->b_frozen_data); } else { + if (journal->j_force_copyout == 1) + need_copy_out = 1; /* pessimistically copy data out */ new_page = jh2bh(jh_in)->b_page; new_offset = offset_in_page(jh2bh(jh_in)->b_data); } diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h index 20e7f78..eb72346 100644 --- a/include/linux/jbd2.h +++ b/include/linux/jbd2.h @@ -968,6 +968,13 @@ struct journal_s u32 j_min_batch_time; u32 j_max_batch_time; + /* + * when force_copyout is set to enable(1), the committing + * transaction makes copy-outs for the metadata and uses the copied + * buffers for IO submission. + */ + u32 j_force_copyout; + /* This function is called when a transaction is closed */ void (*j_commit_callback)(journal_t *, transaction_t *); -- 1.7.1 -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
