io_uring: consistent behaviour with linked read/write
v2 cover io_uring: consistent behaviour with linked read/write
Currently submitting multiple read/write for one file with offset = -1 will not behave as if calling read(2)/write(2) multiple times. The offset may be pinned to the same value for each submission (for example if they are punted to the async worker) and so each read/write will have the same offset.
This patch series fixes this.
Patch 1,3 cleans up the code a bit
Patch 2 grabs the file position at execution time, rather than when the job is queued to be run which fixes inconsistincies when jobs are run asynchronously.
Patch 4 increments the file's f_pos when reading it, which fixes inconsistincies with concurrent runs.
A test for this will be submitted to liburing separately.
v2:
* added patch 4 which fixes cases where IOSQE_IO_LINK is not set
Dylan Yudaken (4):
io_uring: remove duplicated calls to io_kiocb_ppos
io_uring: update kiocb->ki_pos at execution time
io_uring: do not recalculate ppos unnecessarily
io_uring: pre-increment f_pos on rw
fs/io_uring.c | 103 ++++++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 88 insertions(+), 15 deletions(-)
base-commit: 754e0b0e35608ed5206d6a67a791563c631cec07
Signature · 1 line
2.30.2
PATCH v2 1/4 io_uring: remove duplicated calls to io_kiocb_ppos
io_kiocb_ppos is called in both branches, and it seems that the compiler does not fuse this. Fusing removes a few bytes from loop_rw_iter.
Before:
$ nm -S fs/io_uring.o | grep loop_rw_iter 0000000000002430 0000000000000124 t loop_rw_iter
After:
$ nm -S fs/io_uring.o | grep loop_rw_iter 0000000000002430 000000000000010d t loop_rw_iter
--- fs/io_uring.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)
fs/io_uring.c +5 -2
@@ -3400,6 +3400,7 @@ static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)struct kiocb *kiocb = &req->rw.kiocb;struct file *file = req->file;ssize_t ret =0 ; +loff_t *ppos;/* * Don't support polled IO through this interface, and we can't
@@ -3412,6 +3413,8 @@ static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter) !(kiocb->ki_filp->f_flags & O_NONBLOCK))return -EAGAIN ; + ppos =io_kiocb_ppos (kiocb); +while (iov_iter_count (iter)) {struct iovec iovec;ssize_t nr;
@@ -3425,10 +3428,10 @@ static ssize_t loop_rw_iter(int rw, struct io_kiocb *req, struct iov_iter *iter)if (rw ==READ ) { nr = file->f_op->read (file, iovec.iov_base, - iovec.iov_len,io_kiocb_ppos (kiocb)); + iovec.iov_len, ppos); }else { nr = file->f_op->write (file, iovec.iov_base, - iovec.iov_len,io_kiocb_ppos (kiocb)); + iovec.iov_len, ppos); }if (nr <0 ) {
Signature · 1 line
2.30.2
PATCH v2 2/4 io_uring: update kiocb->ki_pos at execution time
In reply to Dylan YudakenUpdate kiocb->ki_pos at execution time rather than in io_prep_rw(). io_prep_rw() happens before the job is enqueued to a worker and so the offset might be read multiple times before being executed once.
Ensures that the file position in a set of _linked_ SQEs will be only obtained after earlier SQEs have completed, and so will include their incremented file position.
--- fs/io_uring.c | 25 +++++++++++++++++-------- 1 file changed, 17 insertions(+), 8 deletions(-)
fs/io_uring.c +17 -8
@@ -3000,14 +3000,6 @@ static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe) req->flags |=io_file_get_flags (file) <<REQ_F_SUPPORT_NOWAIT_BIT ; kiocb->ki_pos =READ_ONCE (sqe->off); -if (kiocb->ki_pos ==-1 ) { -if (!(file->f_mode &FMODE_STREAM )) { - req->flags |=REQ_F_CUR_POS ; - kiocb->ki_pos = file->f_pos; - }else { - kiocb->ki_pos =0 ; - } - } kiocb->ki_flags =iocb_flags (file); ret =kiocb_set_rw_flags (kiocb,READ_ONCE (sqe->rw_flags));if (unlikely (ret))
@@ -3074,6 +3066,19 @@ static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) } } +static inline void +io_kiocb_update_pos (struct io_kiocb *req,struct kiocb *kiocb) +{ +if (kiocb->ki_pos ==-1 ) { +if (!(req->file->f_mode &FMODE_STREAM )) { + req->flags |=REQ_F_CUR_POS ; + kiocb->ki_pos = req->file->f_pos; + }else { + kiocb->ki_pos =0 ; + } + } +} +static void kiocb_done (struct io_kiocb *req,ssize_t ret,unsigned int issue_flags) {
@@ -3662,6 +3667,8 @@ static int io_read(struct io_kiocb *req, unsigned int issue_flags) kiocb->ki_flags &= ~IOCB_NOWAIT ; } +io_kiocb_update_pos (req, kiocb); + ret =rw_verify_area (READ , req->file,io_kiocb_ppos (kiocb), req->result);if (unlikely (ret)) {kfree (iovec);
@@ -3791,6 +3798,8 @@ static int io_write(struct io_kiocb *req, unsigned int issue_flags) kiocb->ki_flags &= ~IOCB_NOWAIT ; } +io_kiocb_update_pos (req, kiocb); + ret =rw_verify_area (WRITE , req->file,io_kiocb_ppos (kiocb), req->result);if (unlikely (ret)) goto out_free;
Signature · 1 line
2.30.2
Jens Axboe
maintainerRe: [PATCH v2 2/4] io_uring: update kiocb->ki_pos at execution time
Quoted 38 lines · Dylan Yudaken
Update
kiocb->ki_posat execution time rather than inio_prep_rw().io_prep_rw()happens before the job is enqueued to a worker and so the offset might be read multiple times before being executed once.Ensures that the file position in a set of _linked_ SQEs will be only obtained after earlier SQEs have completed, and so will include their incremented file position.
Signed-off-by Dylan Yudaken <[email protected]>--- fs/io_uring.c | 25 +++++++++++++++++-------- 1 file changed, 17 insertions(+), 8 deletions(-)fs/io_uring.c +2
-8@@ -3000,14 +3000,6 @@ static int io_prep_rw(struct io_kiocb *req, const struct io_uring_sqe *sqe) req->flags |=io_file_get_flags (file) <<REQ_F_SUPPORT_NOWAIT_BIT ; kiocb->ki_pos =READ_ONCE (sqe->off); -if (kiocb->ki_pos ==-1 ) { -if (!(file->f_mode &FMODE_STREAM )) { - req->flags |=REQ_F_CUR_POS ; - kiocb->ki_pos = file->f_pos; - }else { - kiocb->ki_pos =0 ; - } - } kiocb->ki_flags =iocb_flags (file); ret =kiocb_set_rw_flags (kiocb,READ_ONCE (sqe->rw_flags));if (unlikely (ret))@@ -3074,6 +3066,19 @@ static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) } } +static inline void +io_kiocb_update_pos (struct io_kiocb *req,struct kiocb *kiocb)
fs/io_uring.c static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) · line 3080
Quoted 10 lines
fs/io_uring.c +10
-0+{ +if (kiocb->ki_pos ==-1 ) { +if (!(req->file->f_mode &FMODE_STREAM )) { + req->flags |=REQ_F_CUR_POS ; + kiocb->ki_pos = req->file->f_pos; + }else { + kiocb->ki_pos =0 ; + } + } +}
static inline void io_kiocb_update_pos(struct io_kiocb *req,
struct kiocb *kiocb)
{
}
Can we just drop the kiocb argument? It'll always be req->rw.kiocb. Should generate the same code if you do:
static inline void io_kiocb_update_pos(struct io_kiocb *req)
{
struct kiocb *kiocb = &req->rw.kiocb;
...
}Apart from that minor thing, looks good to me.
Signature · 1 line
Jens Axboe
PATCH v2 3/4 io_uring: do not recalculate ppos unnecessarily
In reply to Dylan YudakenThere is a slight optimisation to be had by calculating the correct pos pointer inside io_kiocb_update_pos and then using that later.
It seems code size drops by a bit:000000000000a1b0 0000000000000400 t io_read000000000000a5b0 0000000000000319 t io_write
vs000000000000a1b0 00000000000003f6 t io_read000000000000a5b0 0000000000000310 t io_write
--- fs/io_uring.c | 18 ++++++++++++------ 1 file changed, 12 insertions(+), 6 deletions(-)
fs/io_uring.c +12 -6
@@ -3066,17 +3066,21 @@ static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) } } -static inline void +static inline loff_t *io_kiocb_update_pos (struct io_kiocb *req,struct kiocb *kiocb) { +bool is_stream = req->file->f_mode &FMODE_STREAM ;if (kiocb->ki_pos ==-1 ) { -if (!(req->file->f_mode &FMODE_STREAM )) { +if (!is_stream) { req->flags |=REQ_F_CUR_POS ; kiocb->ki_pos = req->file->f_pos; +return &kiocb->ki_pos; }else { kiocb->ki_pos =0 ; +return NULL ; } } +return is_stream ?NULL : &kiocb->ki_pos; } static void kiocb_done(struct io_kiocb *req, ssize_t ret,
@@ -3637,6 +3641,7 @@ static int io_read(struct io_kiocb *req, unsigned int issue_flags)bool force_nonblock = issue_flags &IO_URING_F_NONBLOCK ;struct io_async_rw *rw;ssize_t ret, ret2; +loff_t *ppos;if (!req_has_async_data (req)) { ret =io_import_iovec (READ , req, &iovec, s, issue_flags);
@@ -3667,9 +3672,9 @@ static int io_read(struct io_kiocb *req, unsigned int issue_flags) kiocb->ki_flags &= ~IOCB_NOWAIT ; } -io_kiocb_update_pos (req, kiocb); + ppos =io_kiocb_update_pos (req, kiocb); - ret =rw_verify_area (READ , req->file,io_kiocb_ppos (kiocb), req->result); + ret =rw_verify_area (READ , req->file, ppos, req->result);if (unlikely (ret)) {kfree (iovec);return ret;
@@ -3768,6 +3773,7 @@ static int io_write(struct io_kiocb *req, unsigned int issue_flags)struct kiocb *kiocb = &req->rw.kiocb;bool force_nonblock = issue_flags &IO_URING_F_NONBLOCK ;ssize_t ret, ret2; +loff_t *ppos;if (!req_has_async_data (req)) { ret =io_import_iovec (WRITE , req, &iovec, s, issue_flags);
@@ -3798,9 +3804,9 @@ static int io_write(struct io_kiocb *req, unsigned int issue_flags) kiocb->ki_flags &= ~IOCB_NOWAIT ; } -io_kiocb_update_pos (req, kiocb); + ppos =io_kiocb_update_pos (req, kiocb); - ret =rw_verify_area (WRITE , req->file,io_kiocb_ppos (kiocb), req->result); + ret =rw_verify_area (WRITE , req->file, ppos, req->result);if (unlikely (ret)) goto out_free;
Signature · 1 line
2.30.2
PATCH v2 4/4 io_uring: pre-increment f_pos on rw
In reply to Dylan YudakenIn read/write ops, preincrement f_pos when no offset is specified, and then attempt fix up the position after IO completes if it completed less than expected. This fixes the problem where multiple queued up IO will all obtain the same f_pos, and so perform the same read/write.
This is still not as consistent as sync r/w, as it is able to advance the file offset past the end of the file. It seems it would be quite a performance hit to work around this limitation - such as by keeping track of concurrent operations - and the downside does not seem to be too problematic.
The attempt to fix up the f_pos after will at least mean that in situations where a single operation is run, then the position will be consistent.
--- fs/io_uring.c | 81 ++++++++++++++++++++++++++++++++++++++++++--------- 1 file changed, 68 insertions(+), 13 deletions(-)
fs/io_uring.c +68 -13
@@ -3066,21 +3066,71 @@ static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) } } -static inline loff_t * -io_kiocb_update_pos (struct io_kiocb *req,struct kiocb *kiocb) +static inline bool +io_kiocb_update_pos (struct io_kiocb *req,struct kiocb *kiocb, +loff_t **ppos,u64 expected,bool force_nonblock) {bool is_stream = req->file->f_mode &FMODE_STREAM ;if (kiocb->ki_pos ==-1 ) {if (!is_stream) { - req->flags |=REQ_F_CUR_POS ; + *ppos = &kiocb->ki_pos; +WARN_ON (req->flags &REQ_F_CUR_POS ); +if (req->file->f_mode &FMODE_ATOMIC_POS ) { +if (force_nonblock) { +if (!mutex_trylock (&req->file->f_pos_lock)) +return true; + }else { +mutex_lock (&req->file->f_pos_lock); + } + } kiocb->ki_pos = req->file->f_pos; -return &kiocb->ki_pos; + req->flags |=REQ_F_CUR_POS ; + req->file->f_pos += expected; +if (req->file->f_mode &FMODE_ATOMIC_POS ) +mutex_unlock (&req->file->f_pos_lock); +return false; }else { kiocb->ki_pos =0 ; -return NULL ; + *ppos =NULL ; +return false; } } -return is_stream ?NULL : &kiocb->ki_pos; + *ppos = is_stream ?NULL : &kiocb->ki_pos; +return false; +} + +static inline void +io_kiocb_done_pos (struct io_kiocb *req,struct kiocb *kiocb,u64 actual) +{ +u64 expected; + +if (likely (!(req->flags &REQ_F_CUR_POS ))) +return ; + + expected = req->rw.len; +if (actual >= expected) +return ; + +/* +* It's not definitely safe to lock here, and the assumption is, +* that if we cannot lock the position that it will be changing, +* and if it will be changing - then we can't update it anyway +*/ +if (req->file->f_mode &FMODE_ATOMIC_POS + && !mutex_trylock (&req->file->f_pos_lock)) +return ; + +/* +* now we want to move the pointer, but only if everything is consistent +* with how we left it originally +*/ +if (req->file->f_pos == kiocb->ki_pos + (expected - actual)) + req->file->f_pos = kiocb->ki_pos; + +/* else something else messed with f_pos and we can't do anything */ + +if (req->file->f_mode &FMODE_ATOMIC_POS ) +mutex_unlock (&req->file->f_pos_lock); } static void kiocb_done(struct io_kiocb *req, ssize_t ret,
@@ -3096,8 +3146,7 @@ static void kiocb_done(struct io_kiocb *req, ssize_t ret, ret += io->bytes_done; } -if (req->flags &REQ_F_CUR_POS ) - req->file->f_pos = req->rw.kiocb.ki_pos; +io_kiocb_done_pos (req, &req->rw.kiocb, ret >=0 ? ret :0 );if (ret >=0 && (req->rw.kiocb.ki_complete == io_complete_rw))__io_complete_rw (req, ret, issue_flags);else
@@ -3662,21 +3711,23 @@ static int io_read(struct io_kiocb *req, unsigned int issue_flags)if (force_nonblock) {/* If the file doesn't support async, just async punt */ -if (unlikely (!io_file_supports_nowait (req))) { +if (unlikely (!io_file_supports_nowait (req) || +io_kiocb_update_pos (req, kiocb, &ppos, + req->rw.len, true))) { ret =io_setup_async_rw (req, iovec, s, true);return ret ?: -EAGAIN ; } kiocb->ki_flags |=IOCB_NOWAIT ; }else { +io_kiocb_update_pos (req, kiocb, &ppos, req->rw.len, false);/* Ensure we clear previously set non-block flag */ kiocb->ki_flags &= ~IOCB_NOWAIT ; } - ppos =io_kiocb_update_pos (req, kiocb); - ret =rw_verify_area (READ , req->file, ppos, req->result);if (unlikely (ret)) {kfree (iovec); +io_kiocb_done_pos (req, kiocb,0 );return ret; }
@@ -3798,14 +3849,17 @@ static int io_write(struct io_kiocb *req, unsigned int issue_flags) (req->flags & REQ_F_ISREG)) goto copy_iov; +/* if we cannot lock the file position then punt */ +if (unlikely (io_kiocb_update_pos (req, kiocb, &ppos, req->rw.len, true))) + goto copy_iov; + kiocb->ki_flags |=IOCB_NOWAIT ; }else { +io_kiocb_update_pos (req, kiocb, &ppos, req->rw.len, false);/* Ensure we clear previously set non-block flag */ kiocb->ki_flags &= ~IOCB_NOWAIT ; } - ppos =io_kiocb_update_pos (req, kiocb); - ret =rw_verify_area (WRITE , req->file, ppos, req->result);if (unlikely (ret)) goto out_free;
@@ -3858,6 +3912,7 @@ static int io_write(struct io_kiocb *req, unsigned int issue_flags)return ret ?: -EAGAIN ; } out_free: +io_kiocb_done_pos (req, kiocb,0 );/* it's reportedly faster than delegating the null check to kfree() */ if (iovec)kfree (iovec);
Signature · 1 line
2.30.2
Re: [PATCH v2 4/4] io_uring: pre-increment f_pos on rw
Quoted 26 lines · Dylan Yudaken
In read/write ops, preincrement
f_poswhen no offset is specified, and then attempt fix up the position after IO completes if it completed less than expected. This fixes the problem where multiple queued up IO will all obtain the samef_pos, and so perform the same read/write.This is still not as consistent as sync r/w, as it is able to advance the file offset past the end of the file. It seems it would be quite a performance hit to work around this limitation - such as by keeping track of concurrent operations - and the downside does not seem to be too problematic.
The attempt to fix up the
f_posafter will at least mean that in situations where a single operation is run, then the position will be consistent.Co-developed-by Jens Axboe <[email protected]>Signed-off-by Jens Axboe <[email protected]>Signed-off-by Dylan Yudaken <[email protected]>--- fs/io_uring.c | 81 ++++++++++++++++++++++++++++++++++++++++++--------- 1 file changed, 68 insertions(+), 13 deletions(-)fs/io_uring.c +0
-0@@ -3066,21 +3066,71 @@ static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)
[...]
fs/io_uring.c static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) · line 3103
Quoted 10 lines
fs/io_uring.c +7
-1+return false; } } -return is_stream ?NULL : &kiocb->ki_pos; + *ppos = is_stream ?NULL : &kiocb->ki_pos; +return false; +} + +static inline void +io_kiocb_done_pos (struct io_kiocb *req,struct kiocb *kiocb,u64 actual)
That's a lot of inlining, I wouldn't be surprised if the compiler will even refuse to do that.
io_kiocb_done_pos() {
// rest of it
}inline io_kiocb_done_pos() {
if (!(flags & CUR_POS));
return;
__io_kiocb_done_pos();
}io_kiocb_update_pos() is huge as well
Quoted 15 lines
fs/io_uring.c +15
-0+{ +u64 expected; + +if (likely (!(req->flags &REQ_F_CUR_POS ))) +return ; + + expected = req->rw.len; +if (actual >= expected) +return ; + +/* +* It's not definitely safe to lock here, and the assumption is, +* that if we cannot lock the position that it will be changing, +* and if it will be changing - then we can't update it anyway +*/
fs/io_uring.c static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) · line 3128
Quoted 10 lines
fs/io_uring.c +10
-0+if (req->file->f_mode &FMODE_ATOMIC_POS + && !mutex_trylock (&req->file->f_pos_lock)) +return ; + +/* +* now we want to move the pointer, but only if everything is consistent +* with how we left it originally +*/ +if (req->file->f_pos == kiocb->ki_pos + (expected - actual)) + req->file->f_pos = kiocb->ki_pos;
I wonder, is it good enough / safe to just assign it considering that the request was executed outside of locks? vfs_seek()?
fs/io_uring.c static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret) · line 3134
Quoted 6 lines
fs/io_uring.c +5
-0+ +/* else something else messed with f_pos and we can't do anything */ + +if (req->file->f_mode &FMODE_ATOMIC_POS ) +mutex_unlock (&req->file->f_pos_lock); }
Do we even care about races while reading it? E.g. pos = READ_ONCE();
fs/io_uring.c static int io_read(struct io_kiocb *req, unsigned int issue_flags) · line 3730
Quoted 7 lines
fs/io_uring.c +1
-2- ppos =io_kiocb_update_pos (req, kiocb); - ret =rw_verify_area (READ , req->file, ppos, req->result);if (unlikely (ret)) {kfree (iovec); +io_kiocb_done_pos (req, kiocb,0 );
Why do we update it on failure?
[...]
fs/io_uring.c static int io_write(struct io_kiocb *req, unsigned int issue_flags) · line 3915
Quoted 10 lines
fs/io_uring.c +1
-2- ppos =io_kiocb_update_pos (req, kiocb); - ret =rw_verify_area (WRITE , req->file, ppos, req->result);if (unlikely (ret)) goto out_free;@@ -3858,6 +3912,7 @@ static int io_write(struct io_kiocb *req, unsigned int issue_flags)return ret ?: -EAGAIN ; } out_free: +io_kiocb_done_pos (req, kiocb,0 );
Looks weird. It appears we don't need it on failure and successes are covered by kiocb_done() / ->ki_complete
Quoted 3 lines
fs/io_uring.c +0
-0/* it's reportedly faster than delegating the null check to kfree() */ if (iovec)kfree (iovec);
Signature · 1 line
Pavel Begunkov
Re: [PATCH v2 4/4] io_uring: pre-increment f_pos on rw
Quoted 99 lines · Pavel Begunkov
Quoted 27 lines · Dylan Yudaken
In read/write ops, preincrement
f_poswhen no offset is specified, and then attempt fix up the position after IO completes if it completed less than expected. This fixes the problem where multiple queued up IO will all obtain the samef_pos, and so perform the same read/write.This is still not as consistent as sync r/w, as it is able to advance the file offset past the end of the file. It seems it would be quite a performance hit to work around this limitation - such as by keeping track of concurrent operations - and the downside does not seem to be too problematic.
The attempt to fix up the
f_posafter will at least mean that in situations where a single operation is run, then the position will be consistent.Co-developed-by Jens Axboe <[email protected]>Signed-off-by Jens Axboe <[email protected]>Signed-off-by Dylan Yudaken <[email protected]>---
fs/io_uring.c| 81 ++++++++++++++++++++++++++++++++++++++++++---------1 file changed, 68 insertions(+), 13 deletions(-)fs/io_uring.c +0
-0@@ -3066,21 +3066,71 @@ static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)[...]
Quoted 11 lines
+ return false;
}
}
- returnis_stream? NULL :&kiocb->ki_pos;
+ *ppos =is_stream? NULL :&kiocb->ki_pos;
+ return false;fs/io_uring.c +4
-0+} + +static inline void +io_kiocb_done_pos(struct io_kiocb *req, struct kiocb *kiocb, u64actual)
That's a lot of inlining, I wouldn't be surprised if the compiler will even refuse to do that.
io_kiocb_done_pos(){
// rest of it
}inline
io_kiocb_done_pos(){
if (!(flags &CUR_POS));
return;
__io_kiocb_done_pos();
}
io_kiocb_update_pos()is huge as wellQuoted 21 lines
fs/io_uring.c +21
-0+{ + u64 expected; + + if (likely(!(req->flags & REQ_F_CUR_POS))) + return; + + expected = req->rw.len; + if (actual >= expected) + return; + + /* + * It's not definitely safe to lock here, and the assumption is, + * that if we cannot lock the position that it will be changing, + * and if it will be changing - then we can't update it anyway + */ + if (req->file->f_mode & FMODE_ATOMIC_POS + && !mutex_trylock(&req->file->f_pos_lock)) + return; + + /* +* now we want to move the pointer, but only if everything is Quoted 5 lines
consistent
+ * with how we left it originally
+ */
+ if (req->file->f_pos==kiocb->ki_pos+ (expected - actual)) +req->file->f_pos=kiocb->ki_pos;I wonder, is it good enough / safe to just assign it considering that the request was executed outside of locks?
vfs_seek()?Quoted 7 lines
+
+ /* else something else messed withf_posand we can't do anything */
+
+ if (req->file->f_mode&FMODE_ATOMIC_POS)
+mutex_unlock(&req->file->f_pos_lock);
}Do we even care about races while reading it? E.g. pos =
READ_ONCE();
Quoted 8 lines
Quoted 6 lines
- ppos =
io_kiocb_update_pos(req, kiocb);
-
ret =rw_verify_area(READ, req->file, ppos, req->result); if (unlikely(ret)) {
kfree(iovec);
+io_kiocb_done_pos(req, kiocb, 0);Why do we update it on failure?
It seems like a fallback, if no pos change, fallback file->f_pos to the original place
Quoted 22 lines
[...]
Quoted 11 lines
- ppos =
io_kiocb_update_pos(req, kiocb);
-
ret =rw_verify_area(WRITE, req->file, ppos, req->result); if (unlikely(ret))
gotoout_free;fs/io_uring.c +0
-0@@ -3858,6 +3912,7 @@ static int io_write(struct io_kiocb *req, unsigned int issue_flags) return ret ?: -EAGAIN; }
out_free:
+io_kiocb_done_pos(req, kiocb, 0);Looks weird. It appears we don't need it on failure and successes are covered by
kiocb_done()/ ->ki_completeQuoted 3 lines
/* it's reportedly faster than delegating the null check to
kfree()*/ if (iovec)
kfree(iovec);
Re: [PATCH v2 4/4] io_uring: pre-increment f_pos on rw
In reply to Pavel BegunkovQuoted 58 lines · Pavel Begunkov
Quoted 35 lines · Dylan Yudaken
In read/write ops, preincrement
f_poswhen no offset is specified, and
then attempt fix up the position after IO completes if it completed less
than expected. This fixes the problem where multiple queued up IO will all
obtain the samef_pos, and so perform the same read/write.This is still not as consistent as sync r/w, as it is able to advance the
file offset past the end of the file. It seems it would be quite a performance hit to work around this limitation - such as by keeping track
of concurrent operations - and the downside does not seem to be too problematic.The attempt to fix up the
f_posafter will at least mean that in situations
where a single operation is run, then the position will be consistent.Co-developed-by Jens Axboe <[email protected]>Signed-off-by Jens Axboe <[email protected]>Signed-off-by Dylan Yudaken <[email protected]>---
fs/io_uring.c| 81 ++++++++++++++++++++++++++++++++++++++++++----
-----1 file changed, 68 insertions(+), 13 deletions(-)fs/io_uring.c +0
-0@@ -3066,21 +3066,71 @@ static inline void io_rw_done(struct kiocb *kiocb, ssize_t ret)[...]
Quoted 11 lines
+ return false;
}
}
- returnis_stream? NULL :&kiocb->ki_pos; + *ppos =is_stream? NULL :&kiocb->ki_pos; + return false;fs/io_uring.c +4
-0+} + +static inline void +io_kiocb_done_pos(struct io_kiocb *req, struct kiocb *kiocb, u64actual)
That's a lot of inlining, I wouldn't be surprised if the compiler will even refuse to do that.
io_kiocb_done_pos(){
// rest of it
}
Quoted 7 lines
inline
io_kiocb_done_pos(){
if (!(flags &CUR_POS));
return;
__io_kiocb_done_pos();
}
io_kiocb_update_pos()is huge as well
Good idea, will split the slower paths out.
Quoted 26 lines
Quoted 26 lines
fs/io_uring.c +12
-0+{ + u64 expected; + + if (likely(!(req->flags & REQ_F_CUR_POS))) + return; + + expected = req->rw.len; + if (actual >= expected) + return; + + /* +* It's not definitely safe to lock here, and the assumption is,
+ * that if we cannot lock the position that it will be changing,
+ * and if it will be changing - then we can't update it anyway
+ */
+ if (req->file->f_mode&FMODE_ATOMIC_POS
+ && !mutex_trylock(&req->file->f_pos_lock)) + return;
+
+ /*
+ * now we want to move the pointer, but only if everything is consistent
Quoted 8 lines
Quoted 5 lines
+ * with how we left it originally
+ */
+ if (req->file->f_pos==kiocb->ki_pos+ (expected - actual))
+ req->file->f_pos =kiocb->ki_pos;I wonder, is it good enough / safe to just assign it considering that the request was executed outside of locks?
vfs_seek()?
No I do not think so - in the case of multiple r/w the same thing will happen, even with no vfs_seek().
Quoted 11 lines
Quoted 8 lines
+
+ /* else something else messed withf_posand we can't do anything */
+
+ if (req->file->f_mode&FMODE_ATOMIC_POS)
+ mutex_unlock(&req->file->f_pos_lock); }Do we even care about races while reading it? E.g. pos =
READ_ONCE();
I think so - if I remove all the locks the test cases fail.
Quoted 33 lines
Quoted 8 lines
- ppos =io_kiocb_update_pos(req, kiocb);
-
ret =rw_verify_area(READ, req->file, ppos, req->result); if (unlikely(ret)) {
kfree(iovec);
+ io_kiocb_done_pos(req, kiocb, 0);Why do we update it on failure?
[...]
Quoted 11 lines
- ppos =
io_kiocb_update_pos(req, kiocb);
-
ret =rw_verify_area(WRITE, req->file, ppos, req->result); if (unlikely(ret))
gotoout_free;fs/io_uring.c +0
-0@@ -3858,6 +3912,7 @@ static int io_write(struct io_kiocb *req, unsigned int issue_flags) return ret ?: -EAGAIN; }
out_free:
+ io_kiocb_done_pos(req, kiocb, 0);Looks weird. It appears we don't need it on failure and successes are covered by
kiocb_done()/ ->ki_completeQuoted 4 lines
/* it's reportedly faster than delegating the null check to
kfree()*/
if (iovec)
kfree(iovec);
Re: [PATCH v2 4/4] io_uring: pre-increment f_pos on rw
In reply to Dylan YudakenQuoted 5 lines · Dylan Yudaken
In read/write ops, preincrement
f_poswhen no offset is specified, and then attempt fix up the position after IO completes if it completed less than expected. This fixes the problem where multiple queued up IO will all obtain the samef_pos, and so perform the same read/write.
Quoted 9 lines
This is still not as consistent as sync r/w, as it is able to advance the file offset past the end of the file. It seems it would be quite a performance hit to work around this limitation - such as by keeping track of concurrent operations - and the downside does not seem to be too problematic.
The attempt to fix up the
f_posafter will at least mean that in situations where a single operation is run, then the position will be consistent.
It's a little bit weird, when a read req returns x bytes read while f_pos
moves ahead y bytes where x isn't equal to y. Don't know if this causes
problems..
Re: [PATCH v2 4/4] io_uring: pre-increment f_pos on rw
Quoted 23 lines · Hao Xu
Quoted 16 lines · Dylan Yudaken
In read/write ops, preincrement
f_poswhen no offset is specified, and
then attempt fix up the position after IO completes if it completed less
than expected. This fixes the problem where multiple queued up IO will all
obtain the samef_pos, and so perform the same read/write.This is still not as consistent as sync r/w, as it is able to advance the
file offset past the end of the file. It seems it would be quite a performance hit to work around this limitation - such as by keeping track
of concurrent operations - and the downside does not seem to be too problematic.Quoted 5 lines
The attempt to fix up the
f_posafter will at least mean that in situations
where a single operation is run, then the position will be consistent.
Quoted 8 lines
It's a little bit weird, when a read req returns x bytes read while
f_posmoves ahead y bytes where x isn't equal to y. Don't know if this causes
problems..
It seems to be ok - as in nothing crashes when f_pos is past the end of the file - but I really am not an expert on these things so am happy to receive feedback on this.
Jens Axboe
maintainerQuoted 10 lines · Dylan Yudaken
Currently submitting multiple read/write for one file with offset = -1 will not behave as if calling read(2)/write(2) multiple times. The offset may be pinned to the same value for each submission (for example if they are punted to the async worker) and so each read/write will have the same offset.
This patch series fixes this.
Patch 1,3 cleans up the code a bit
Quoted 7 lines
Patch 2 grabs the file position at execution time, rather than when the job is queued to be run which fixes inconsistincies when jobs are run asynchronously.
Patch 4 increments the file's
f_poswhen reading it, which fixes inconsistincies with concurrent runs.A test for this will be submitted to liburing separately.
Looks good to me, but the patch 2 change will bubble through to patch 3 and 4 as well. Care to respin a v3?
Signature · 1 line
Jens Axboe
Quoted 20 lines · Jens Axboe
Quoted 19 lines · Dylan Yudaken
Currently submitting multiple read/write for one file with offset = -1 will
not behave as if calling read(2)/write(2) multiple times. The offset may be
pinned to the same value for each submission (for example if they are
punted to the async worker) and so each read/write will have the same
offset.This patch series fixes this.
Patch 1,3 cleans up the code a bit
Patch 2 grabs the file position at execution time, rather than when the job
is queued to be run which fixes inconsistincies when jobs are run asynchronously.
Quoted 9 lines
Quoted 4 lines
Patch 4 increments the file's
f_poswhen reading it, which fixes inconsistincies with concurrent runs.A test for this will be submitted to liburing separately.
Looks good to me, but the patch 2 change will bubble through to patch 3
and 4 as well. Care to respin a v3?
Yes sure - will do it combined with the test v3