Add support for write life time hints
Subsystems (6)
Jens Axboe
maintainerPATCH v3 1/1 Add support for write life time hints
A new iteration of this patchset, previously known as write streams. As before, this patchset aims at enabling applications split up writes into separate streams, based on the perceived life time of the data written. This is useful for a variety of reasons:
- With NVMe 1.3 compliant devices, the device can expose multiple streams. Separating data written into streams based on life time can drastically reduce the write amplification. This helps device endurance, and increases performance. Testing just performed internally at Facebook with these patches showed up to a 25% reduction in NAND writes in a RocksDB setup.
- Software caching solutions can make more intelligent decisions on how and where to place data.
Contrary to previous patches, we're not exposing numeric stream values anymore. I've previously advocated for just doing a set of hints that makes sense instead. See the coverage from the LSFMM summit this year:
https://lwn.net/Articles/717755/
This patchset attempts to do that. We define 4 flags for the pwritev2 system call:
RWF_WRITE_LIFE_SHORT Data written with this flag is expected to have
a high overwrite rate, or life time.
RWF_WRITE_LIFE_MEDIUM Longer life time than SHORT
RWF_WRITE_LIFE_LONG Longer life time than MEDIUM
RWF_WRITE_LIFE_EXTREME Longer life time than LONG
The idea is that these are relative values, so an application can use them as they see fit. The underlying device can then place data appropriately, or be free to ignore the hint. It's just a hint.
A branch based on current master can be pulled
from here:
git://git.kernel.dk/linux-block write-stream.3
Changes since v2:
- Get rid of bio->bi_stream and replace with four request/bio flags. These map directly to the RWF_WRITE_* flags that the user passes in.
- Cleanup the NVMe stream setting.
- Drivers now responsible for updating the queue stream write counter, as they determine what stream to map a given flag to.
Changes since v1:
- Guard queue stream stats to ensure we don't mess up memory, if bio_stream() ever were to return a larger value than we support.
- NVMe: ensure we set the stream modulo the name space defined count.
- Cleanup the RWF_ and IOCB_ flags. Set aside 4 bits, and just store the stream value in there. This makes the passing of stream ID from RWF_ space to IOCB_ (and IOCB_ to bio) more efficient, and cleans it up in general.
- Kill the block internal definitions of the stream type, we don't need them anymore. See above.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerPATCH 1/11 block: add support for carrying stream information in a bio
No functional changes in this patch, we just add four flags that will be used to denote a stream type, and ensure that we don't merge across different stream types.
--- block/blk-merge.c | 16 ++++++++++++++++ include/linux/blk_types.h | 11 +++++++++++ 2 files changed, 27 insertions(+)
block/blk-merge.c +16 -0
@@ -693,6 +693,14 @@ static struct request *attempt_merge(struct request_queue *q,return NULL ;/* +* Don't allow merge of different streams, or for a stream with +* non-stream IO. +*/ +if ((req->cmd_flags &REQ_WRITE_LIFE_MASK ) != + (next->cmd_flags &REQ_WRITE_LIFE_MASK )) +return NULL ; + +/* * If we are allowed to merge, then append bio list * from next to rq and release next. merge_requests_fn * will have updated segment counts, update sector
@@ -811,6 +819,14 @@ bool blk_rq_merge_ok(struct request *rq, struct bio *bio) !blk_write_same_mergeable(rq->bio, bio))return false; +/* +* Don't allow merge of different streams, or for a stream with +* non-stream IO. +*/ +if ((rq->cmd_flags &REQ_WRITE_LIFE_MASK ) != + (bio->bi_opf &REQ_WRITE_LIFE_MASK )) +return false; +return true; }
include/linux/blk_types.h +11 -0
@@ -201,6 +201,10 @@ enum req_flag_bits { __REQ_PREFLUSH,/* request for cache flush */ __REQ_RAHEAD,/* read ahead, can fail anytime */ __REQ_BACKGROUND,/* background IO */ + __REQ_WRITE_SHORT,/* short life time write */ + __REQ_WRITE_MEDIUM,/* medium life time write */ + __REQ_WRITE_LONG,/* long life time write */ + __REQ_WRITE_EXTREME,/* extremely long life time write */ /* command specific flags for REQ_OP_WRITE_ZEROES: */ __REQ_NOUNMAP, /* do not free blocks when zeroing */
@@ -221,6 +225,13 @@ enum req_flag_bits {#define REQ_PREFLUSH (1ULL << __REQ_PREFLUSH)#define REQ_RAHEAD (1ULL << __REQ_RAHEAD)#define REQ_BACKGROUND (1ULL << __REQ_BACKGROUND) +#define REQ_WRITE_SHORT (1ULL << __REQ_WRITE_SHORT) +#define REQ_WRITE_MEDIUM (1ULL << __REQ_WRITE_MEDIUM) +#define REQ_WRITE_LONG (1ULL << __REQ_WRITE_LONG) +#define REQ_WRITE_EXTREME (1ULL << __REQ_WRITE_EXTREME) + +#define REQ_WRITE_LIFE_MASK (REQ_WRITE_SHORT | REQ_WRITE_MEDIUM | \ + REQ_WRITE_LONG | REQ_WRITE_EXTREME)#define REQ_NOUNMAP (1ULL << __REQ_NOUNMAP)
Signature · 1 line
2.7.4
Jens Axboe
maintainerPATCH 2/11 blk-mq: expose stream write stats through debugfs
In reply to Jens AxboeUseful to verify that things are working the way they should. Reading the file will return number of kb written to each stream. Writing the file will reset the statistics. No care is taken to ensure that we don't race on updates.
Drivers will write to q->stream_writes[] if they handle a stream.
--- block/blk-mq-debugfs.c | 24 ++++++++++++++++++++++++ include/linux/blkdev.h | 3 +++ 2 files changed, 27 insertions(+)
block/blk-mq-debugfs.c +24 -0
@@ -133,6 +133,29 @@ static void print_stat(struct seq_file *m, struct blk_rq_stat *stat) } } +static int queue_streams_show (void *data,struct seq_file *m) +{ +struct request_queue *q = data; +int i; + +for (i =0 ; i <BLK_MAX_STREAM ; i++) +seq_printf (m,"stream%d: %llu\n" , i, q->stream_writes[i]); + +return 0 ; +} + +static ssize_t queue_streams_store (void *data,const char __user *buf, +size_t count,loff_t *ppos) +{ +struct request_queue *q = data; +int i; + +for (i =0 ; i <BLK_MAX_STREAM ; i++) + q->stream_writes[i] =0 ; + +return count; +} +static int queue_poll_stat_show (void *data,struct seq_file *m) {struct request_queue *q = data;
@@ -656,6 +679,7 @@ const struct file_operations blk_mq_debugfs_fops = {static const struct blk_mq_debugfs_attr blk_mq_debugfs_queue_attrs[] = { {"poll_stat" ,0400 , queue_poll_stat_show}, {"state" ,0600 , queue_state_show, queue_state_write}, + {"streams" ,0600 , queue_streams_show, queue_streams_store}, {}, };
include/linux/blkdev.h +3 -0
@@ -586,6 +586,9 @@ struct request_queue {size_t cmd_size;void *rq_alloc_data; + +#define BLK_MAX_STREAM 5 +u64 stream_writes[BLK_MAX_STREAM ]; };#define QUEUE_FLAG_QUEUED 1/* uses generic tag queueing */
Signature · 1 line
2.7.4
Jens Axboe
maintainerPATCH 3/11 fs: add support for an inode to carry stream related data
In reply to Jens AxboeNo functional changes in this patch, just in preparation for allowing applications to pass in hints about data life times for writes.
Pack the i_stream field into a 2-byte hole, so we don't grow the size of the inode.
--- fs/inode.c | 1 + include/linux/fs.h | 9 +++++++++ 2 files changed, 10 insertions(+)
fs/inode.c +1 -0
@@ -149,6 +149,7 @@ int inode_init_always(struct super_block *sb, struct inode *inode) inode->i_blocks =0 ; inode->i_bytes =0 ; inode->i_generation =0 ; + inode->i_stream =0 ; inode->i_pipe =NULL ; inode->i_bdev =NULL ; inode->i_cdev =NULL ;
include/linux/fs.h +9 -0
@@ -591,6 +591,7 @@ struct inode {struct timespec i_ctime;spinlock_t i_lock;/* i_blocks, i_bytes, maybe i_size */ unsigned short i_bytes; +unsigned short i_stream;unsigned int i_blkbits;blkcnt_t i_blocks;
@@ -655,6 +656,14 @@ struct inode {void *i_private;/* fs or device private pointer */ }; +static inline unsigned int inode_streamid (struct inode *inode) +{ +if (inode) +return inode->i_stream; + +return 0 ; +} +static inline unsigned int i_blocksize (const struct inode *node) {return (1 << node->i_blkbits);
Signature · 1 line
2.7.4
Jens Axboe
maintainerPATCH 4/11 fs: add support for allowing applications to pass in write life time hints
In reply to Jens AxboeAdd four flags for the pwritev2(2) system call, allowing an application to give the kernel a hint about what on-media life times can be expected from a given write.
The intent is for these values to be relative to each other, no absolute meaning should be attached to these flag names.
Define IOCB flags to carry this information over, and finally transform them into the block defined stream values.
--- fs/read_write.c | 9 ++++++++- include/linux/fs.h | 12 ++++++++++++ include/uapi/linux/fs.h | 10 ++++++++++ 3 files changed, 30 insertions(+), 1 deletion(-)
fs/read_write.c +8 -1
@@ -678,7 +678,7 @@ static ssize_t do_iter_readv_writev(struct file *filp, struct iov_iter *iter,struct kiocb kiocb;ssize_t ret; -if (flags & ~(RWF_HIPRI |RWF_DSYNC |RWF_SYNC )) +if (flags & ~(RWF_HIPRI |RWF_DSYNC |RWF_SYNC |RWF_WRITE_LIFE_MASK ))return -EOPNOTSUPP ;init_sync_kiocb (&kiocb, filp);
@@ -688,6 +688,13 @@ static ssize_t do_iter_readv_writev(struct file *filp, struct iov_iter *iter, kiocb.ki_flags |=IOCB_DSYNC ;if (flags &RWF_SYNC ) kiocb.ki_flags |= (IOCB_DSYNC |IOCB_SYNC ); +if (flags &RWF_WRITE_LIFE_MASK ) { +struct inode *inode =file_inode (filp); + + inode->i_stream = (flags &RWF_WRITE_LIFE_MASK ) >> +RWF_WRITE_LIFE_SHIFT ; + kiocb.ki_flags |= inode->i_stream <<IOCB_WRITE_LIFE_SHIFT ; + } kiocb.ki_pos = *ppos;if (type ==READ )
include/linux/fs.h +12 -0
@@ -269,6 +269,12 @@ struct writeback_control;#define IOCB_SYNC (1 << 5)#define IOCB_WRITE (1 << 6) +/* +* Steal 4-bits for stream information, this allows 16 valid streams +*/ +#define IOCB_WRITE_LIFE_SHIFT 7 +#define IOCB_WRITE_LIFE_MASK (BIT(7) | BIT(8) | BIT(9) | BIT(10)) +struct kiocb {struct file *ki_filp;loff_t ki_pos;
@@ -292,6 +298,12 @@ static inline void init_sync_kiocb(struct kiocb *kiocb, struct file *filp) }; } +static inline int iocb_streamid (const struct kiocb *iocb) +{ +return (iocb->ki_flags &IOCB_WRITE_LIFE_MASK ) >> +IOCB_WRITE_LIFE_SHIFT ; +} +/* * "descriptor" for what we're up to with a read. * This allows us to use the same read code yet
include/uapi/linux/fs.h +10 -0
@@ -361,4 +361,14 @@ struct fscrypt_key {#define RWF_DSYNC 0x00000002/* per-IO O_DSYNC */ #define RWF_SYNC 0x00000004/* per-IO O_SYNC */ +/* +* Data life time write flags, steal 4 bits for that +*/ +#define RWF_WRITE_LIFE_SHIFT 4 +#define RWF_WRITE_LIFE_MASK 0x000000f0/* 4 bits of stream ID */ +#define RWF_WRITE_LIFE_SHORT (1 << RWF_WRITE_LIFE_SHIFT) +#define RWF_WRITE_LIFE_MEDIUM (2 << RWF_WRITE_LIFE_SHIFT) +#define RWF_WRITE_LIFE_LONG (3 << RWF_WRITE_LIFE_SHIFT) +#define RWF_WRITE_LIFE_EXTREME (4 << RWF_WRITE_LIFE_SHIFT) +#endif /* _UAPI_LINUX_FS_H */
Signature · 1 line
2.7.4
Jens Axboe
maintainerPATCH 5/11 block: add helpers for setting/checking stream validity
In reply to Jens AxboeWe map the RWF_WRITE_* life time flags to the internal flags. Drivers can then, in turn, map those flags to a suitable stream type.
--- block/bio.c | 16 ++++++++++++++++ include/linux/bio.h | 1 + include/linux/blk_types.h | 5 +++++ 3 files changed, 22 insertions(+)
block/bio.c +16 -0
@@ -2082,6 +2082,22 @@ void bio_clone_blkcg_association(struct bio *dst, struct bio *src)#endif /* CONFIG_BLK_CGROUP */ +static const unsigned int rwf_write_to_opf_flag[] = { +0 ,REQ_WRITE_SHORT ,REQ_WRITE_MEDIUM ,REQ_WRITE_LONG ,REQ_WRITE_EXTREME +}; + +/* +* 'stream_flags' is one of RWF_WRITE_LIFE_* values +*/ +void bio_set_streamid (struct bio *bio,unsigned int rwf_flags) +{ +if (WARN_ON_ONCE (rwf_flags >=ARRAY_SIZE (rwf_write_to_opf_flag))) +return ; + + bio->bi_opf |= rwf_write_to_opf_flag[rwf_flags]; +} +EXPORT_SYMBOL_GPL (bio_set_streamid); +static void __init biovec_init_slabs (void ) {int i;
include/linux/bio.h +1 -0
@@ -443,6 +443,7 @@ extern struct bio *bio_copy_kern(struct request_queue *, void *, unsigned int, gfp_t, int);extern void bio_set_pages_dirty (struct bio *bio);extern void bio_check_pages_dirty (struct bio *bio); +extern void bio_set_streamid (struct bio *bio,unsigned int rwf_flags);void generic_start_io_acct (int rw,unsigned long sectors,struct hd_struct *part);
include/linux/blk_types.h +5 -0
@@ -323,4 +323,9 @@ struct blk_rq_stat { u64 batch; }; +static inline bool blk_stream_valid (unsigned int opf) +{ +return (opf &REQ_WRITE_LIFE_MASK ) !=0 ; +} +#endif /* __LINUX_BLK_TYPES_H */
Signature · 1 line
2.7.4
Jens Axboe
maintainerPATCH 6/11 fs: add O_DIRECT support for sending down bio stream information
In reply to Jens Axboe--- fs/block_dev.c | 2 ++ fs/direct-io.c | 2 ++ fs/iomap.c | 1 + 3 files changed, 5 insertions(+)
fs/block_dev.c +2 -0
@@ -239,6 +239,7 @@ __blkdev_direct_IO_simple(struct kiocb *iocb, struct iov_iter *iter, should_dirty = true; }else { bio.bi_opf =dio_bio_write_op (iocb); +bio_set_streamid (&bio,iocb_streamid (iocb));task_io_account_write (ret); }
@@ -374,6 +375,7 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)bio_set_pages_dirty (bio); }else { bio->bi_opf =dio_bio_write_op (iocb); +bio_set_streamid (bio,iocb_streamid (iocb));task_io_account_write (bio->bi_iter.bi_size); }
fs/direct-io.c +2 -0
@@ -386,6 +386,8 @@ dio_bio_alloc(struct dio *dio, struct dio_submit *sdio,else bio->bi_end_io = dio_bio_end_io; +bio_set_streamid (bio,iocb_streamid (dio->iocb)); + sdio->bio = bio; sdio->logical_offset_in_bio = sdio->cur_page_fs_offset; }
fs/iomap.c +1 -0
@@ -804,6 +804,7 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,if (dio->flags &IOMAP_DIO_WRITE ) {bio_set_op_attrs (bio,REQ_OP_WRITE ,REQ_SYNC |REQ_IDLE ); +bio_set_streamid (bio,inode_streamid (inode));task_io_account_write (bio->bi_iter.bi_size); }else {bio_set_op_attrs (bio,REQ_OP_READ ,0 );
Signature · 1 line
2.7.4
Jens Axboe
maintainerPATCH 7/11 fs: add support for buffered writeback to pass down stream information
In reply to Jens Axboe--- fs/buffer.c | 14 +++++++++----- fs/mpage.c | 1 + 2 files changed, 10 insertions(+), 5 deletions(-)
fs/buffer.c +9 -5
@@ -49,7 +49,7 @@static int fsync_buffers_list (spinlock_t *lock,struct list_head *list);static int submit_bh_wbc (int op,int op_flags,struct buffer_head *bh, -struct writeback_control *wbc); +unsigned int stream,struct writeback_control *wbc);#define BH_ENTRY (list) list_entry((list), struct buffer_head, b_assoc_buffers)
@@ -1829,7 +1829,8 @@ int __block_write_full_page(struct inode *inode, struct page *page, do { struct buffer_head *next = bh->b_this_page; if (buffer_async_write(bh)) { - submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, wbc); + submit_bh_wbc(REQ_OP_WRITE, write_flags, bh, + inode_streamid(inode), wbc); nr_underway++; } bh = next;
@@ -1883,7 +1884,8 @@ int __block_write_full_page(struct inode *inode, struct page *page,struct buffer_head *next = bh->b_this_page;if (buffer_async_write (bh)) {clear_buffer_dirty (bh); -submit_bh_wbc (REQ_OP_WRITE , write_flags, bh, wbc); +submit_bh_wbc (REQ_OP_WRITE , write_flags, bh, +inode_streamid (inode), wbc); nr_underway++; } bh = next;
@@ -3091,7 +3093,7 @@ void guard_bio_eod(int op, struct bio *bio) }static int submit_bh_wbc (int op,int op_flags,struct buffer_head *bh, -struct writeback_control *wbc) +unsigned int stream,struct writeback_control *wbc) {struct bio *bio;
@@ -3130,6 +3132,8 @@ static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,/* Take care of bh's that straddle the end of the device */ guard_bio_eod (op, bio); +bio_set_streamid (bio, stream); +if (buffer_meta (bh)) op_flags |=REQ_META ;if (buffer_prio (bh))
@@ -3142,7 +3146,7 @@ static int submit_bh_wbc(int op, int op_flags, struct buffer_head *bh,int submit_bh (int op,int op_flags,struct buffer_head *bh) { -return submit_bh_wbc (op, op_flags, bh,NULL ); +return submit_bh_wbc (op, op_flags, bh,0 ,NULL ); }EXPORT_SYMBOL (submit_bh);
fs/mpage.c +1 -0
@@ -614,6 +614,7 @@ static int __mpage_writepage(struct page *page, struct writeback_control *wbc, goto confused;wbc_init_bio (wbc, bio); +bio_set_streamid (bio,inode_streamid (inode)); }/*
Signature · 1 line
2.7.4
Jens Axboe
maintainerPATCH 8/11 ext4: add support for passing in stream information for buffered writes
In reply to Jens Axboe--- fs/ext4/page-io.c | 2 ++ 1 file changed, 2 insertions(+)
fs/ext4/page-io.c +2 -0
@@ -350,6 +350,7 @@ void ext4_io_submit(struct ext4_io_submit *io)int io_op_flags = io->io_wbc->sync_mode ==WB_SYNC_ALL ?REQ_SYNC :0 ;bio_set_op_attrs (io->io_bio,REQ_OP_WRITE , io_op_flags); +bio_set_streamid (io->io_bio,inode_streamid (io->io_end->inode));submit_bio (io->io_bio); } io->io_bio =NULL ;
@@ -396,6 +397,7 @@ static int io_submit_add_bh(struct ext4_io_submit *io, ret =io_submit_init_bio (io, bh);if (ret)return ret; +bio_set_streamid (io->io_bio,inode_streamid (inode)); } ret =bio_add_page (io->io_bio, page, bh->b_size,bh_offset (bh));if (ret != bh->b_size)
Signature · 1 line
2.7.4
Jens Axboe
maintainerPATCH 9/11 xfs: add support for passing in stream information for buffered writes
In reply to Jens Axboe--- fs/xfs/xfs_aops.c | 2 ++ 1 file changed, 2 insertions(+)
fs/xfs/xfs_aops.c +2 -0
@@ -505,6 +505,7 @@ xfs_submit_ioend(return status; } +bio_set_streamid (ioend->io_bio,inode_streamid (ioend->io_inode));submit_bio (ioend->io_bio);return 0 ; }
@@ -564,6 +565,7 @@ xfs_chain_bio(bio_chain (ioend->io_bio, new);bio_get (ioend->io_bio);/* for xfs_destroy_ioend */ ioend->io_bio->bi_opf =REQ_OP_WRITE |wbc_to_write_flags (wbc); +bio_set_streamid (ioend->io_bio,inode_streamid (ioend->io_inode));submit_bio (ioend->io_bio); ioend->io_bio = new; }
Signature · 1 line
2.7.4
Jens Axboe
maintainerPATCH 10/11 btrfs: add support for passing in stream information for buffered writes
In reply to Jens Axboe--- fs/btrfs/extent_io.c | 1 + 1 file changed, 1 insertion(+)
fs/btrfs/extent_io.c +1 -0
@@ -2827,6 +2827,7 @@ static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree, bio->bi_end_io = end_io_func; bio->bi_private = tree;bio_set_op_attrs (bio, op, op_flags); +bio_set_streamid (bio,inode_streamid (page->mapping->host));if (wbc) {wbc_init_bio (wbc, bio);wbc_account_io (wbc, page, page_size);
Signature · 1 line
2.7.4
Jens Axboe
maintainerPATCH 11/11 nvme: add support for streams and directives
In reply to Jens AxboeThis adds support for Directives in NVMe, particular for the Streams directive. Support for Directives is a new feature in NVMe 1.3. It allows a user to pass in information about where to store the data, so that it the device can do so most effiently. If an application is managing and writing data with different life times, mixing differently retentioned data onto the same locations on flash can cause write amplification to grow. This, in turn, will reduce performance and life time of the device.
We default to allocating 4 streams per name space, but it is configurable with the 'streams_per_ns' module option. If a write stream is set in a write, flag is as such before sending it to the device.
Some debug stuff in this patch, dumping streams ID params when we load nvme.
--- drivers/nvme/host/core.c | 146 +++++++++++++++++++++++++++++++++++++++++++++++ drivers/nvme/host/nvme.h | 1 + include/linux/nvme.h | 48 ++++++++++++++++ 3 files changed, 195 insertions(+)
drivers/nvme/host/core.c +146 -0
@@ -65,6 +65,10 @@ static bool force_apst;module_param (force_apst, bool,0644 );MODULE_PARM_DESC (force_apst,"allow APST for newly enumerated devices even if quirked off" ); +static char streams_per_ns =4 ; +module_param (streams_per_ns, byte,0644 ); +MODULE_PARM_DESC (streams_per_ns,"if available, allocate this many streams per NS" ); +static LIST_HEAD (nvme_ctrl_list);static DEFINE_SPINLOCK (dev_list_lock);
@@ -331,9 +335,34 @@ static inline int nvme_setup_discard(struct nvme_ns *ns, struct request *req,return BLK_MQ_RQ_QUEUE_OK ; } +static unsigned int nvme_get_write_stream (struct nvme_ns *ns, +struct request *req) +{ +unsigned int streamid =0 ; + +if (req_op (req) !=REQ_OP_WRITE || !blk_stream_valid (req->cmd_flags) || + !ns->nr_streams) +return 0 ; + +if (req->cmd_flags &REQ_WRITE_SHORT ) + streamid =1 ; +else if (req->cmd_flags &REQ_WRITE_MEDIUM ) + streamid =2 ; +else if (req->cmd_flags &REQ_WRITE_LONG ) + streamid =3 ; +else if (req->cmd_flags &REQ_WRITE_EXTREME ) + streamid =4 ; + +if (streamid <BLK_MAX_STREAM ) + req->q->stream_writes[streamid] +=blk_rq_bytes (req) >>9 ; + +return (streamid % (ns->nr_streams +1 )); +} +static inline void nvme_setup_rw (struct nvme_ns *ns,struct request *req,struct nvme_command *cmnd) { +unsigned int stream;u16 control =0 ;u32 dsmgmt =0 ;
@@ -351,6 +380,12 @@ static inline void nvme_setup_rw(struct nvme_ns *ns, struct request *req, cmnd->rw.slba =cpu_to_le64 (nvme_block_nr (ns,blk_rq_pos (req))); cmnd->rw.length =cpu_to_le16 ((blk_rq_bytes (req) >> ns->lba_shift) -1 ); + stream =nvme_get_write_stream (ns, req); +if (stream) { + control |=NVME_RW_DTYPE_STREAMS ; + dsmgmt |= (stream <<16 ); + } +if (ns->ms) {switch (ns->pi_type) {case NVME_NS_DPS_PI_TYPE3 :
@@ -1073,6 +1108,109 @@ static int nvme_revalidate_disk(struct gendisk *disk)return 0 ; } +static int nvme_enable_streams (struct nvme_ns *ns) +{ +struct nvme_command c; + +memset (&c,0 ,sizeof (c)); + + c.directive.opcode = nvme_admin_directive_send; + c.directive.nsid =cpu_to_le32 (ns->ns_id); + c.directive.doper =NVME_DIR_SND_ID_OP_ENABLE ; + c.directive.dtype =NVME_DIR_IDENTIFY ; + c.directive.tdtype =NVME_DIR_STREAMS ; + c.directive.endir =NVME_DIR_ENDIR ; + +return nvme_submit_sync_cmd (ns->ctrl->admin_q, &c,NULL ,0 ); +} + +static int nvme_streams_params (struct nvme_ns *ns) +{ +struct nvme_ctrl *ctrl = ns->ctrl; +struct streams_directive_params s; +struct nvme_command c; +int ret; + +memset (&c,0 ,sizeof (c)); +memset (&s,0 ,sizeof (s)); + + c.directive.opcode = nvme_admin_directive_recv; + c.directive.nsid =cpu_to_le32 (ns->ns_id); + c.directive.numd =sizeof (s); + c.directive.doper =NVME_DIR_RCV_ST_OP_PARAM ; + c.directive.dtype =NVME_DIR_STREAMS ; + + ret =nvme_submit_sync_cmd (ctrl->admin_q, &c, &s,sizeof (s)); +if (ret) +return ret; + + s.msl =le16_to_cpu (s.msl); + s.nssa =le16_to_cpu (s.nssa); + s.nsso =le16_to_cpu (s.nsso); + s.sws =le32_to_cpu (s.sws); + s.sgs =le16_to_cpu (s.sgs); + s.nsa =le16_to_cpu (s.nsa); + s.nso =le16_to_cpu (s.nso); + +dev_info (ctrl->device,"streams: msl=%u, nssa=%u, nsso=%u, sws=%u " +"sgs=%u, nsa=%u, nso=%u\n" , s.msl, s.nssa, + s.nsso, s.sws, s.sgs, s.nsa, s.nso); +return 0 ; +} + +static int nvme_streams_allocate (struct nvme_ns *ns,unsigned int streams) +{ +struct nvme_command c; + +memset (&c,0 ,sizeof (c)); + + c.directive.opcode = nvme_admin_directive_recv; + c.directive.nsid =cpu_to_le32 (ns->ns_id); + c.directive.doper =NVME_DIR_RCV_ST_OP_RESOURCE ; + c.directive.dtype =NVME_DIR_STREAMS ; + c.directive.endir = streams; + +return nvme_submit_sync_cmd (ns->ctrl->admin_q, &c,NULL ,0 ); +} + +static int nvme_streams_deallocate (struct nvme_ns *ns) +{ +struct nvme_command c; + +memset (&c,0 ,sizeof (c)); + + c.directive.opcode = nvme_admin_directive_send; + c.directive.nsid =cpu_to_le32 (ns->ns_id); + c.directive.doper =NVME_DIR_SND_ST_OP_REL_RSC ; + c.directive.dtype =NVME_DIR_STREAMS ; + +return nvme_submit_sync_cmd (ns->ctrl->admin_q, &c,NULL ,0 ); +} + +static void nvme_config_streams (struct nvme_ns *ns) +{ +int ret; + + ret =nvme_enable_streams (ns); +if (ret) +return ; + + ret =nvme_streams_params (ns); +if (ret) +return ; + + ret =nvme_streams_allocate (ns, streams_per_ns); +if (ret) +return ; + + ret =nvme_streams_params (ns); +if (ret) +return ; + + ns->nr_streams = streams_per_ns; +dev_info (ns->ctrl->device,"successfully enabled streams\n" ); +} +static char nvme_pr_type (enum pr_type type) {switch (type) {
@@ -1606,6 +1744,9 @@ int nvme_init_identify(struct nvme_ctrl *ctrl) ctrl->sgls =le32_to_cpu (id->sgls); ctrl->kas =le16_to_cpu (id->kas); +if (ctrl->oacs &NVME_CTRL_OACS_DIRECTIVES ) +dev_info (ctrl->dev,"supports directives\n" ); + ctrl->npss = id->npss; prev_apsta = ctrl->apsta;if (ctrl->quirks &NVME_QUIRK_NO_APST ) {
@@ -2060,6 +2201,9 @@ static void nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid) goto out_free_id; } +if (ctrl->oacs &NVME_CTRL_OACS_DIRECTIVES ) +nvme_config_streams (ns); + disk =alloc_disk_node (0 , node);if (!disk) goto out_free_id;
@@ -2112,6 +2256,8 @@ static void nvme_ns_remove(struct nvme_ns *ns) &nvme_ns_attr_group);if (ns->ndev)nvme_nvm_unregister_sysfs (ns); +if (ns->nr_streams) +nvme_streams_deallocate (ns);del_gendisk (ns->disk);blk_cleanup_queue (ns->queue); }
drivers/nvme/host/nvme.h +1 -0
@@ -192,6 +192,7 @@ struct nvme_ns {u8 uuid[16 ];unsigned ns_id; +unsigned nr_streams;int lba_shift;u16 ms;bool ext;
include/linux/nvme.h +48 -0
90 lines not on this page · raw message
Signature · 1 line
2.7.4
Re: [PATCH 03/11] fs: add support for an inode to carry stream related data
In reply to Jens AxboeQuoted 6 lines · Jens Axboe
No functional changes in this patch, just in preparation for allowing applications to pass in hints about data life times for writes.
Pack the
i_streamfield into a 2-byte hole, so we don't grow the size of the inode.
Can't we find space for 4 flags somewhere else?
Re: [PATCH 04/11] fs: add support for allowing applications to pass in write life time hints
In reply to Jens AxboeQuoted 9 lines · Jens Axboe
Add four flags for the pwritev2(2) system call, allowing an application to give the kernel a hint about what on-media life times can be expected from a given write.
The intent is for these values to be relative to each other, no absolute meaning should be attached to these flag names.
Define IOCB flags to carry this information over, and finally transform them into the block defined stream values.
Note that we also have the RWF_NOWAIT flag pending, with which this would clash. Maybe move it to the end of the space.
Re: [PATCH 05/11] block: add helpers for setting/checking stream validity
In reply to Jens AxboeQuoted 8 lines
block/bio.c +5
-0+static const unsigned int rwf_write_to_opf_flag[] = { +0 ,REQ_WRITE_SHORT ,REQ_WRITE_MEDIUM ,REQ_WRITE_LONG ,REQ_WRITE_EXTREME +}; + +/*
block/bio.c void bio_clone_blkcg_association(struct bio *dst, struct bio *src) · line 2099
Quoted 10 lines
block/bio.c +10
-0+* 'stream_flags' is one of RWF_WRITE_LIFE_* values +*/ +void bio_set_streamid (struct bio *bio,unsigned int rwf_flags) +{ +if (WARN_ON_ONCE (rwf_flags >=ARRAY_SIZE (rwf_write_to_opf_flag))) +return ; + + bio->bi_opf |= rwf_write_to_opf_flag[rwf_flags]; +} +EXPORT_SYMBOL_GPL (bio_set_streamid);
I'd move the bio->bi_opf assignment outş for a call like:
bio->bi_opf |= bio_op_write_bucket(flags);
and preferably move the flags masking / shifting into the helper as well.
include/linux/blk_types.h struct blk_rq_stat { · line 329
Quoted 6 lines
};
include/linux/blk_types.h +4
-0+static inline bool blk_stream_valid (unsigned int opf) +{ +return (opf &REQ_WRITE_LIFE_MASK ) !=0 ; +}
Replace the stream name here with lifetime or similar as well?
Re: [PATCH 11/11] nvme: add support for streams and directives
In reply to Jens Axboedrivers/nvme/host/core.c static inline int nvme_setup_discard(struct nvme_ns *ns, struct request *req, · line 345
Quoted 8 lines
drivers/nvme/host/core.c +8
-0+static unsigned int nvme_get_write_stream (struct nvme_ns *ns, +struct request *req) +{ +unsigned int streamid =0 ; + +if (req_op (req) !=REQ_OP_WRITE || !blk_stream_valid (req->cmd_flags) || + !ns->nr_streams) +return 0 ;
Might make more sense to do this check in the caller?
drivers/nvme/host/core.c static inline int nvme_setup_discard(struct nvme_ns *ns, struct request *req, · line 356
Quoted 11 lines
drivers/nvme/host/core.c +11
-0+ +if (req->cmd_flags &REQ_WRITE_SHORT ) + streamid =1 ; +else if (req->cmd_flags &REQ_WRITE_MEDIUM ) + streamid =2 ; +else if (req->cmd_flags &REQ_WRITE_LONG ) + streamid =3 ; +else if (req->cmd_flags &REQ_WRITE_EXTREME ) + streamid =4 ; + +if (streamid <BLK_MAX_STREAM )
Can happen per the index above.
drivers/nvme/host/core.c static inline int nvme_setup_discard(struct nvme_ns *ns, struct request *req, · line 359
Quoted 3 lines
drivers/nvme/host/core.c +3
-0+ req->q->stream_writes[streamid] +=blk_rq_bytes (req) >>9 ; + +return (streamid % (ns->nr_streams +1 ));
Should we do smarted collapsing? e.g. short + medium and long + extreme for two? What for three? Does one extra stream make sense in this scheme?
drivers/nvme/host/core.c static int nvme_revalidate_disk(struct gendisk *disk) · line 1157
Quoted 3 lines
drivers/nvme/host/core.c +3
-0+dev_info (ctrl->device,"streams: msl=%u, nssa=%u, nsso=%u, sws=%u " +"sgs=%u, nsa=%u, nso=%u\n" , s.msl, s.nssa, + s.nsso, s.sws, s.sgs, s.nsa, s.nso);
Way to chatty.
drivers/nvme/host/core.c int nvme_init_identify(struct nvme_ctrl *ctrl) · line 1748
Quoted 2 lines
drivers/nvme/host/core.c +2
-0+if (ctrl->oacs &NVME_CTRL_OACS_DIRECTIVES ) +dev_info (ctrl->dev,"supports directives\n" );
Same. Use nvme-cli for that sort of info.
drivers/nvme/host/core.c static void nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid) · line 2205
Quoted 9 lines
drivers/nvme/host/core.c +2
-0ctrl->npss = id->npss; prev_apsta = ctrl->apsta;if (ctrl->quirks &NVME_QUIRK_NO_APST ) {@@ -2060,6 +2201,9 @@ static void nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid) goto out_free_id; } +if (ctrl->oacs &NVME_CTRL_OACS_DIRECTIVES ) +nvme_config_streams (ns);
This sets aside four streams on any device that supports them, and will probably kill performance on them unless you have a workload that actually uses those streams. I think they need to be allocated lazily.
Jens Axboe
maintainerRe: [PATCH 03/11] fs: add support for an inode to carry stream related data
In reply to Christoph HellwigQuoted 9 lines · Christoph Hellwig
Quoted 6 lines · Jens Axboe
No functional changes in this patch, just in preparation for allowing applications to pass in hints about data life times for writes.
Pack the
i_streamfield into a 2-byte hole, so we don't grow the size of the inode.Can't we find space for 4 flags somewhere else?
Where? I can stuff them in i_flags? But apart from that, stealing space in a hole seems the sanest.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerRe: [PATCH 04/11] fs: add support for allowing applications to pass in write life time hints
In reply to Christoph HellwigQuoted 5 lines · Christoph Hellwig
Quoted 4 lines · Jens Axboe
Add four flags for the pwritev2(2) system call, allowing an application to give the kernel a hint about what on-media life times can be expected from a given write.
Quoted 8 lines
Quoted 5 lines
The intent is for these values to be relative to each other, no absolute meaning should be attached to these flag names.
Define IOCB flags to carry this information over, and finally transform them into the block defined stream values.
Note that we also have the
RWF_NOWAITflag pending, with which this would clash. Maybe move it to the end of the space.
That's a trivial thing to resolve. Whomever goes last can resolve that. I can put them at the end if you feel strongly about it, I personally don't really care.
Signature · 1 line
Jens Axboe
Re: [PATCH 01/11] block: add support for carrying stream information in a bio
In reply to Jens AxboeBtw, I think these could also easily map to DSM field in the NVMe write command, except that these unfortunately mix in read information as well.
include/linux/blk_types.h enum req_flag_bits { · line 204
Quoted 1 line
include/linux/blk_types.h +1
-0+ __REQ_WRITE_SHORT, /* short life time write */
-> Frequent writes and infrequent reads to the LBA range indicated.
or
-> Frequent writes and frequent reads to the LBA range indicated.
include/linux/blk_types.h enum req_flag_bits { · line 205
Quoted 1 line
include/linux/blk_types.h +1
-0+ __REQ_WRITE_MEDIUM, /* medium life time write */
-> Typical number of reads and writes expected for this LBA range.
include/linux/blk_types.h enum req_flag_bits { · line 206
Quoted 1 line
include/linux/blk_types.h +1
-0+ __REQ_WRITE_LONG, /* long life time write */
-> Infrequent writes and infrequent reads to the LBA range indicated.
or
-> Infrequent writes and frequent reads to the LBA range indicated.
include/linux/blk_types.h enum req_flag_bits { · line 207
Quoted 1 line
include/linux/blk_types.h +1
-0+ __REQ_WRITE_EXTREME, /* extremely long life time write */
-> One time write. E.g. command is due to virus scan, backup, file copy, or archive.
Jens Axboe
maintainerRe: [PATCH 05/11] block: add helpers for setting/checking stream validity
In reply to Christoph HellwigQuoted 21 lines · Christoph Hellwig
Quoted 8 lines
block/bio.c +5
-0+static const unsigned int rwf_write_to_opf_flag[] = { +0 ,REQ_WRITE_SHORT ,REQ_WRITE_MEDIUM ,REQ_WRITE_LONG ,REQ_WRITE_EXTREME +}; + +/* block/bio.c void bio_clone_blkcg_association(struct bio *dst, struct bio *src) · line 2099
Quoted 10 lines
block/bio.c +10
-0+* 'stream_flags' is one of RWF_WRITE_LIFE_* values +*/ +void bio_set_streamid (struct bio *bio,unsigned int rwf_flags) +{ +if (WARN_ON_ONCE (rwf_flags >=ARRAY_SIZE (rwf_write_to_opf_flag))) +return ; + + bio->bi_opf |= rwf_write_to_opf_flag[rwf_flags]; +} +EXPORT_SYMBOL_GPL (bio_set_streamid);I'd move the
bio->bi_opfassignment outş for a call like:
Quoted 4 lines
bio->bi_opf |= bio_op_write_bucket(flags);and preferably move the flags masking / shifting into the helper as well.
OK, I can do that.
Quoted 6 lines
include/linux/blk_types.h struct blk_rq_stat { · line 329
Quoted 4 lines
include/linux/blk_types.h +4
-0+static inline bool blk_stream_valid (unsigned int opf) +{ +return (opf &REQ_WRITE_LIFE_MASK ) !=0 ; +}Replace the stream name here with lifetime or similar as well?
Yeah, I thought about that after sending v3 out. I should do that everywhere.
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerRe: [PATCH 11/11] nvme: add support for streams and directives
In reply to Christoph HellwigQuoted 10 lines · Christoph Hellwig
drivers/nvme/host/core.c static inline int nvme_setup_discard(struct nvme_ns *ns, struct request *req, · line 345
Quoted 8 lines
drivers/nvme/host/core.c +8
-0+static unsigned int nvme_get_write_stream (struct nvme_ns *ns, +struct request *req) +{ +unsigned int streamid =0 ; + +if (req_op (req) !=REQ_OP_WRITE || !blk_stream_valid (req->cmd_flags) || + !ns->nr_streams) +return 0 ;Might make more sense to do this check in the caller?
OK, will fix up.
Quoted 4 lines
Quoted 4 lines
drivers/nvme/host/core.c +4
-0+if (req->cmd_flags &REQ_WRITE_SHORT ) + streamid =1 ; +else if (req->cmd_flags &REQ_WRITE_MEDIUM ) + streamid =2 ;
Quoted 8 lines
drivers/nvme/host/core.c static inline int nvme_setup_discard(struct nvme_ns *ns, struct request *req, · line 356
Quoted 6 lines
drivers/nvme/host/core.c +6
-0+else if (req->cmd_flags &REQ_WRITE_LONG ) + streamid =3 ; +else if (req->cmd_flags &REQ_WRITE_EXTREME ) + streamid =4 ; + +if (streamid <BLK_MAX_STREAM )Can happen per the index above.
True, that's a leftover from the previous version.
Quoted 7 lines
drivers/nvme/host/core.c static inline int nvme_setup_discard(struct nvme_ns *ns, struct request *req, · line 359
Quoted 3 lines
drivers/nvme/host/core.c +3
-0+ req->q->stream_writes[streamid] +=blk_rq_bytes (req) >>9 ; + +return (streamid % (ns->nr_streams +1 ));Should we do smarted collapsing? e.g. short + medium and long + extreme for two? What for three? Does one extra stream make sense in this scheme?
Collapsing is probably saner than round-robin. I'd tend to collapse on the longer life time end of things, logically that would make the most sense.
Quoted 5 lines
drivers/nvme/host/core.c static int nvme_revalidate_disk(struct gendisk *disk) · line 1157
Quoted 3 lines
drivers/nvme/host/core.c +3
-0+dev_info (ctrl->device,"streams: msl=%u, nssa=%u, nsso=%u, sws=%u " +"sgs=%u, nsa=%u, nso=%u\n" , s.msl, s.nssa, + s.nsso, s.sws, s.sgs, s.nsa, s.nso);Way to chatty.
Sure, that's mentioned in the changelog, that stuff will go.
Quoted 4 lines
drivers/nvme/host/core.c int nvme_init_identify(struct nvme_ctrl *ctrl) · line 1748
Quoted 2 lines
drivers/nvme/host/core.c +2
-0+if (ctrl->oacs &NVME_CTRL_OACS_DIRECTIVES ) +dev_info (ctrl->dev,"supports directives\n" );Same. Use nvme-cli for that sort of info.
Ditto
Quoted 7 lines
Quoted 7 lines
drivers/nvme/host/core.c +0
-0ctrl->npss = id->npss; prev_apsta = ctrl->apsta;if (ctrl->quirks &NVME_QUIRK_NO_APST ) {@@ -2060,6 +2201,9 @@ static void nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid) goto out_free_id; }
Quoted 7 lines
drivers/nvme/host/core.c static void nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid) · line 2205
Quoted 2 lines
drivers/nvme/host/core.c +2
-0+if (ctrl->oacs &NVME_CTRL_OACS_DIRECTIVES ) +nvme_config_streams (ns);This sets aside four streams on any device that supports them, and will probably kill performance on them unless you have a workload that actually uses those streams. I think they need to be allocated lazily.
That's a good point, I have been thinking about how best to handle this. I don't want an API for this, but doing it lazy would be fine. When we see a write with a life time attached, kick off background setup of streams. Until that's done, don't use any streams. Once setup, we'll mark it as we currently do now.
How's that?
Signature · 1 line
Jens Axboe
Jens Axboe
maintainerRe: [PATCH 01/11] block: add support for carrying stream information in a bio
In reply to Christoph HellwigQuoted 3 lines · Christoph Hellwig
Btw, I think these could also easily map to DSM field in the NVMe write command, except that these unfortunately mix in read information as well.
But that's the problem, they are read/write mixed flags. I'd much rather keep them separate. If some application finds it useful to specify read access patterns, we should have separate flags for those imho.
Signature · 1 line
Jens Axboe
Re: [PATCH 10/11] btrfs: add support for passing in stream information for buffered writes
In reply to Jens AxboeQuoted 2 lines · Jens Axboe
Reviewed-by Andreas Dilger <[email protected]>Signed-off-by Jens Axboe <[email protected]>
Thanks Jens!
Quoted 19 lines
---
fs/btrfs/extent_io.c| 1 +1 file changed, 1 insertion(+)fs/btrfs/extent_io.c +1
-0@@ -2827,6 +2827,7 @@ static int submit_extent_page(int op, int op_flags, struct extent_io_tree *tree, bio->bi_end_io = end_io_func; bio->bi_private = tree;bio_set_op_attrs (bio, op, op_flags); +bio_set_streamid (bio,inode_streamid (page->mapping->host));if (wbc) {wbc_init_bio (wbc, bio);wbc_account_io (wbc, page, page_size);Signature · 1 line
2.7.4