| Message ID | 1553846032-4451-8-git-send-email-joshi.k@samsung.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Extend write-hint for in-kernel use | expand |
On Mar 29, 2019, at 1:53 AM, Kanchan Joshi <joshi.k@samsung.com> wrote: > > For NAND based SSDs, mixing of data with different life-time reduces > efficiency of internal garbage-collection. During FS operations, series > of journal updates will follow/precede series of data/meta updates, causing > intermixing inside SSD. By passing a write-hint with journal, its write > can be isolated from other data/meta writes, leading to endurance/performance > benefit on SSD. > > This patch introduces "j_writehint" member in JBD2 journal, using which > Ext4 specifies write-hint (as SHORT) for journal The comment here says the "WRITE_LIFE_SHORT" hint is used for the journal, but the code uses WRITE_LIFE_KERN_MIN. However, it seems that "MIN" will be mapped to "NONE" if it exceeds the number of streams available in the underlying device. It would be better to use "SHORT" if there are not enough streams available. It should call blk_queue_stream_limits() to see if there are extra stream IDs available, and fall back to WRITE_LIFE_SHORT if not. Cheers, Andreas > Signed-off-by: Kanchan Joshi <joshi.k@samsung.com> > --- > fs/ext4/ext4_jbd2.h | 1 + > fs/ext4/super.c | 2 ++ > fs/jbd2/commit.c | 11 +++++++---- > fs/jbd2/journal.c | 3 ++- > fs/jbd2/revoke.c | 3 ++- > include/linux/jbd2.h | 8 ++++++++ > 6 files changed, 22 insertions(+), 6 deletions(-) > > diff --git a/fs/ext4/ext4_jbd2.h b/fs/ext4/ext4_jbd2.h > index 15b6dd7..b589ca4 100644 > --- a/fs/ext4/ext4_jbd2.h > +++ b/fs/ext4/ext4_jbd2.h > @@ -16,6 +16,7 @@ > #include <linux/jbd2.h> > #include "ext4.h" > > +#define EXT4_JOURNAL_WRITE_HINT (WRITE_LIFE_KERN_MIN) > #define EXT4_JOURNAL(inode) (EXT4_SB((inode)->i_sb)->s_journal) > > /* Define the number of blocks we need to account to a transaction to > diff --git a/fs/ext4/super.c b/fs/ext4/super.c > index fb12d3c..9c2c73e 100644 > --- a/fs/ext4/super.c > +++ b/fs/ext4/super.c > @@ -4289,6 +4289,8 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent) > > set_task_ioprio(sbi->s_journal->j_task, journal_ioprio); > > + sbi->s_journal->j_writehint = EXT4_JOURNAL_WRITE_HINT; > + > sbi->s_journal->j_commit_callback = ext4_journal_commit_callback; > > no_journal: > diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c > index 2eb55c3..6da4c28 100644 > --- a/fs/jbd2/commit.c > +++ b/fs/jbd2/commit.c > @@ -153,10 +153,12 @@ static int journal_submit_commit_record(journal_t *journal, > > if (journal->j_flags & JBD2_BARRIER && > !jbd2_has_feature_async_commit(journal)) > - ret = submit_bh(REQ_OP_WRITE, > - REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh); > + ret = submit_bh_write_hint(REQ_OP_WRITE, > + REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh, > + journal->j_writehint); > else > - ret = submit_bh(REQ_OP_WRITE, REQ_SYNC, bh); > + ret = submit_bh_write_hint(REQ_OP_WRITE, REQ_SYNC, bh, > + journal->j_writehint); > > *cbh = bh; > return ret; > @@ -711,7 +713,8 @@ void jbd2_journal_commit_transaction(journal_t *journal) > clear_buffer_dirty(bh); > set_buffer_uptodate(bh); > bh->b_end_io = journal_end_buffer_io_sync; > - submit_bh(REQ_OP_WRITE, REQ_SYNC, bh); > + submit_bh_write_hint(REQ_OP_WRITE, REQ_SYNC, > + bh, journal->j_writehint); > } > cond_resched(); > stats.run.rs_blocks_logged += bufs; > diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c > index 8ef6b6d..804dc2c 100644 > --- a/fs/jbd2/journal.c > +++ b/fs/jbd2/journal.c > @@ -1384,7 +1384,8 @@ static int jbd2_write_superblock(journal_t *journal, int write_flags) > jbd2_superblock_csum_set(journal, sb); > get_bh(bh); > bh->b_end_io = end_buffer_write_sync; > - ret = submit_bh(REQ_OP_WRITE, write_flags, bh); > + ret = submit_bh_write_hint(REQ_OP_WRITE, write_flags, bh, > + journal->j_writehint); > wait_on_buffer(bh); > if (buffer_write_io_error(bh)) { > clear_buffer_write_io_error(bh); > diff --git a/fs/jbd2/revoke.c b/fs/jbd2/revoke.c > index a1143e5..376b1d8 100644 > --- a/fs/jbd2/revoke.c > +++ b/fs/jbd2/revoke.c > @@ -642,7 +642,8 @@ static void flush_descriptor(journal_t *journal, > set_buffer_jwrite(descriptor); > BUFFER_TRACE(descriptor, "write"); > set_buffer_dirty(descriptor); > - write_dirty_buffer(descriptor, REQ_SYNC); > + write_dirty_buffer_with_hint(descriptor, REQ_SYNC, > + journal->j_writehint); > } > #endif > > diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h > index 0f919d5..918f21e 100644 > --- a/include/linux/jbd2.h > +++ b/include/linux/jbd2.h > @@ -1139,6 +1139,14 @@ struct journal_s > */ > __u32 j_csum_seed; > > + /** > + * @j_writehint: > + * > + * write-hint for journal (set by FS). > + */ > + enum rw_hint j_writehint; > + > + > #ifdef CONFIG_DEBUG_LOCK_ALLOC > /** > * @j_trans_commit_map: > -- > 2.7.4 > Cheers, Andreas
On Sat 30-03-19 11:49:54, Andreas Dilger wrote: > On Mar 29, 2019, at 1:53 AM, Kanchan Joshi <joshi.k@samsung.com> wrote: > > > > For NAND based SSDs, mixing of data with different life-time reduces > > efficiency of internal garbage-collection. During FS operations, series > > of journal updates will follow/precede series of data/meta updates, causing > > intermixing inside SSD. By passing a write-hint with journal, its write > > can be isolated from other data/meta writes, leading to endurance/performance > > benefit on SSD. > > > > This patch introduces "j_writehint" member in JBD2 journal, using which > > Ext4 specifies write-hint (as SHORT) for journal > > The comment here says the "WRITE_LIFE_SHORT" hint is used for the journal, > but the code uses WRITE_LIFE_KERN_MIN. However, it seems that "MIN" will > be mapped to "NONE" if it exceeds the number of streams available in the > underlying device. It would be better to use "SHORT" if there are not > enough streams available. > > It should call blk_queue_stream_limits() to see if there are extra stream > IDs available, and fall back to WRITE_LIFE_SHORT if not. I disagree. I'd first keep the behavior implemented in this patch to keep things simple. Later if we decide more smarts are needed when SSDs don't have enough hints available, we can always add them. But this patch either keeps the current behavior (i.e., no hint) or improves the situation by providing a special hint. So it is a clear win. I'm not so convinced using WRITE_LIFE_SHORT is always a win when userspace's idea of "short" is different from the kernel's idea of "short"... Honza > > Cheers, Andreas > > > Signed-off-by: Kanchan Joshi <joshi.k@samsung.com> > > --- > > fs/ext4/ext4_jbd2.h | 1 + > > fs/ext4/super.c | 2 ++ > > fs/jbd2/commit.c | 11 +++++++---- > > fs/jbd2/journal.c | 3 ++- > > fs/jbd2/revoke.c | 3 ++- > > include/linux/jbd2.h | 8 ++++++++ > > 6 files changed, 22 insertions(+), 6 deletions(-) > > > > diff --git a/fs/ext4/ext4_jbd2.h b/fs/ext4/ext4_jbd2.h > > index 15b6dd7..b589ca4 100644 > > --- a/fs/ext4/ext4_jbd2.h > > +++ b/fs/ext4/ext4_jbd2.h > > @@ -16,6 +16,7 @@ > > #include <linux/jbd2.h> > > #include "ext4.h" > > > > +#define EXT4_JOURNAL_WRITE_HINT (WRITE_LIFE_KERN_MIN) > > #define EXT4_JOURNAL(inode) (EXT4_SB((inode)->i_sb)->s_journal) > > > > /* Define the number of blocks we need to account to a transaction to > > diff --git a/fs/ext4/super.c b/fs/ext4/super.c > > index fb12d3c..9c2c73e 100644 > > --- a/fs/ext4/super.c > > +++ b/fs/ext4/super.c > > @@ -4289,6 +4289,8 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent) > > > > set_task_ioprio(sbi->s_journal->j_task, journal_ioprio); > > > > + sbi->s_journal->j_writehint = EXT4_JOURNAL_WRITE_HINT; > > + > > sbi->s_journal->j_commit_callback = ext4_journal_commit_callback; > > > > no_journal: > > diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c > > index 2eb55c3..6da4c28 100644 > > --- a/fs/jbd2/commit.c > > +++ b/fs/jbd2/commit.c > > @@ -153,10 +153,12 @@ static int journal_submit_commit_record(journal_t *journal, > > > > if (journal->j_flags & JBD2_BARRIER && > > !jbd2_has_feature_async_commit(journal)) > > - ret = submit_bh(REQ_OP_WRITE, > > - REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh); > > + ret = submit_bh_write_hint(REQ_OP_WRITE, > > + REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh, > > + journal->j_writehint); > > else > > - ret = submit_bh(REQ_OP_WRITE, REQ_SYNC, bh); > > + ret = submit_bh_write_hint(REQ_OP_WRITE, REQ_SYNC, bh, > > + journal->j_writehint); > > > > *cbh = bh; > > return ret; > > @@ -711,7 +713,8 @@ void jbd2_journal_commit_transaction(journal_t *journal) > > clear_buffer_dirty(bh); > > set_buffer_uptodate(bh); > > bh->b_end_io = journal_end_buffer_io_sync; > > - submit_bh(REQ_OP_WRITE, REQ_SYNC, bh); > > + submit_bh_write_hint(REQ_OP_WRITE, REQ_SYNC, > > + bh, journal->j_writehint); > > } > > cond_resched(); > > stats.run.rs_blocks_logged += bufs; > > diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c > > index 8ef6b6d..804dc2c 100644 > > --- a/fs/jbd2/journal.c > > +++ b/fs/jbd2/journal.c > > @@ -1384,7 +1384,8 @@ static int jbd2_write_superblock(journal_t *journal, int write_flags) > > jbd2_superblock_csum_set(journal, sb); > > get_bh(bh); > > bh->b_end_io = end_buffer_write_sync; > > - ret = submit_bh(REQ_OP_WRITE, write_flags, bh); > > + ret = submit_bh_write_hint(REQ_OP_WRITE, write_flags, bh, > > + journal->j_writehint); > > wait_on_buffer(bh); > > if (buffer_write_io_error(bh)) { > > clear_buffer_write_io_error(bh); > > diff --git a/fs/jbd2/revoke.c b/fs/jbd2/revoke.c > > index a1143e5..376b1d8 100644 > > --- a/fs/jbd2/revoke.c > > +++ b/fs/jbd2/revoke.c > > @@ -642,7 +642,8 @@ static void flush_descriptor(journal_t *journal, > > set_buffer_jwrite(descriptor); > > BUFFER_TRACE(descriptor, "write"); > > set_buffer_dirty(descriptor); > > - write_dirty_buffer(descriptor, REQ_SYNC); > > + write_dirty_buffer_with_hint(descriptor, REQ_SYNC, > > + journal->j_writehint); > > } > > #endif > > > > diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h > > index 0f919d5..918f21e 100644 > > --- a/include/linux/jbd2.h > > +++ b/include/linux/jbd2.h > > @@ -1139,6 +1139,14 @@ struct journal_s > > */ > > __u32 j_csum_seed; > > > > + /** > > + * @j_writehint: > > + * > > + * write-hint for journal (set by FS). > > + */ > > + enum rw_hint j_writehint; > > + > > + > > #ifdef CONFIG_DEBUG_LOCK_ALLOC > > /** > > * @j_trans_commit_map: > > -- > > 2.7.4 > > > > > Cheers, Andreas > > > > >
Kanchan, > For NAND based SSDs, mixing of data with different life-time reduces > efficiency of internal garbage-collection. During FS operations, > series of journal updates will follow/precede series of data/meta > updates, causing intermixing inside SSD. By passing a write-hint with > journal, its write can be isolated from other data/meta writes, > leading to endurance/performance benefit on SSD. Why not just introduce REQ_JOURNAL and let the device driver decide how to turn that into something appropriate for the device? That's what I'll need for SCSI. Existing SCSI streams are not a good fit.
Hi Martin, > Why not just introduce REQ_JOURNAL and let the device driver decide how to turn that into something appropriate for the device? It began with that kind of thought/goal i.e. introduce something just for FS journal. But it seems to have evolved for good. Current approach extends write-hint infra so that whole thing becomes extensible for other kind of use-cases (than FS journal) as well. Also in this approach, driver will do little, while block-layer will do majority of the work. > That's what I'll need for SCSI. Existing SCSI streams are not a good fit. Do you see that it's difficult for SCSI to use write-hint infrastructure for streams? -----Original Message----- From: Martin K. Petersen [mailto:martin.petersen@oracle.com] Sent: Wednesday, April 03, 2019 8:28 AM To: Kanchan Joshi <joshi.k@samsung.com> Cc: linux-kernel@vger.kernel.org; linux-block@vger.kernel.org; linux-nvme@lists.infradead.org; linux-fsdevel@vger.kernel.org; linux-ext4@vger.kernel.org; axboe@fb.com; prakash.v@samsung.com; anshul@samsung.com; joshiiitr@gmail.com Subject: Re: [PATCH v3 7/7] fs/ext4,jbd2: add support for passing write-hint with journal Kanchan, > For NAND based SSDs, mixing of data with different life-time reduces > efficiency of internal garbage-collection. During FS operations, > series of journal updates will follow/precede series of data/meta > updates, causing intermixing inside SSD. By passing a write-hint with > journal, its write can be isolated from other data/meta writes, > leading to endurance/performance benefit on SSD. Why not just introduce REQ_JOURNAL and let the device driver decide how to turn that into something appropriate for the device? That's what I'll need for SCSI. Existing SCSI streams are not a good fit.
diff --git a/fs/ext4/ext4_jbd2.h b/fs/ext4/ext4_jbd2.h index 15b6dd7..b589ca4 100644 --- a/fs/ext4/ext4_jbd2.h +++ b/fs/ext4/ext4_jbd2.h @@ -16,6 +16,7 @@ #include <linux/jbd2.h> #include "ext4.h" +#define EXT4_JOURNAL_WRITE_HINT (WRITE_LIFE_KERN_MIN) #define EXT4_JOURNAL(inode) (EXT4_SB((inode)->i_sb)->s_journal) /* Define the number of blocks we need to account to a transaction to diff --git a/fs/ext4/super.c b/fs/ext4/super.c index fb12d3c..9c2c73e 100644 --- a/fs/ext4/super.c +++ b/fs/ext4/super.c @@ -4289,6 +4289,8 @@ static int ext4_fill_super(struct super_block *sb, void *data, int silent) set_task_ioprio(sbi->s_journal->j_task, journal_ioprio); + sbi->s_journal->j_writehint = EXT4_JOURNAL_WRITE_HINT; + sbi->s_journal->j_commit_callback = ext4_journal_commit_callback; no_journal: diff --git a/fs/jbd2/commit.c b/fs/jbd2/commit.c index 2eb55c3..6da4c28 100644 --- a/fs/jbd2/commit.c +++ b/fs/jbd2/commit.c @@ -153,10 +153,12 @@ static int journal_submit_commit_record(journal_t *journal, if (journal->j_flags & JBD2_BARRIER && !jbd2_has_feature_async_commit(journal)) - ret = submit_bh(REQ_OP_WRITE, - REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh); + ret = submit_bh_write_hint(REQ_OP_WRITE, + REQ_SYNC | REQ_PREFLUSH | REQ_FUA, bh, + journal->j_writehint); else - ret = submit_bh(REQ_OP_WRITE, REQ_SYNC, bh); + ret = submit_bh_write_hint(REQ_OP_WRITE, REQ_SYNC, bh, + journal->j_writehint); *cbh = bh; return ret; @@ -711,7 +713,8 @@ void jbd2_journal_commit_transaction(journal_t *journal) clear_buffer_dirty(bh); set_buffer_uptodate(bh); bh->b_end_io = journal_end_buffer_io_sync; - submit_bh(REQ_OP_WRITE, REQ_SYNC, bh); + submit_bh_write_hint(REQ_OP_WRITE, REQ_SYNC, + bh, journal->j_writehint); } cond_resched(); stats.run.rs_blocks_logged += bufs; diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c index 8ef6b6d..804dc2c 100644 --- a/fs/jbd2/journal.c +++ b/fs/jbd2/journal.c @@ -1384,7 +1384,8 @@ static int jbd2_write_superblock(journal_t *journal, int write_flags) jbd2_superblock_csum_set(journal, sb); get_bh(bh); bh->b_end_io = end_buffer_write_sync; - ret = submit_bh(REQ_OP_WRITE, write_flags, bh); + ret = submit_bh_write_hint(REQ_OP_WRITE, write_flags, bh, + journal->j_writehint); wait_on_buffer(bh); if (buffer_write_io_error(bh)) { clear_buffer_write_io_error(bh); diff --git a/fs/jbd2/revoke.c b/fs/jbd2/revoke.c index a1143e5..376b1d8 100644 --- a/fs/jbd2/revoke.c +++ b/fs/jbd2/revoke.c @@ -642,7 +642,8 @@ static void flush_descriptor(journal_t *journal, set_buffer_jwrite(descriptor); BUFFER_TRACE(descriptor, "write"); set_buffer_dirty(descriptor); - write_dirty_buffer(descriptor, REQ_SYNC); + write_dirty_buffer_with_hint(descriptor, REQ_SYNC, + journal->j_writehint); } #endif diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h index 0f919d5..918f21e 100644 --- a/include/linux/jbd2.h +++ b/include/linux/jbd2.h @@ -1139,6 +1139,14 @@ struct journal_s */ __u32 j_csum_seed; + /** + * @j_writehint: + * + * write-hint for journal (set by FS). + */ + enum rw_hint j_writehint; + + #ifdef CONFIG_DEBUG_LOCK_ALLOC /** * @j_trans_commit_map:
For NAND based SSDs, mixing of data with different life-time reduces efficiency of internal garbage-collection. During FS operations, series of journal updates will follow/precede series of data/meta updates, causing intermixing inside SSD. By passing a write-hint with journal, its write can be isolated from other data/meta writes, leading to endurance/performance benefit on SSD. This patch introduces "j_writehint" member in JBD2 journal, using which Ext4 specifies write-hint (as SHORT) for journal Signed-off-by: Kanchan Joshi <joshi.k@samsung.com> --- fs/ext4/ext4_jbd2.h | 1 + fs/ext4/super.c | 2 ++ fs/jbd2/commit.c | 11 +++++++---- fs/jbd2/journal.c | 3 ++- fs/jbd2/revoke.c | 3 ++- include/linux/jbd2.h | 8 ++++++++ 6 files changed, 22 insertions(+), 6 deletions(-)