block/file-posix.c: Use pwritev2() with RWF_DSYNC for FUA - update
v2 cover block/file-posix.c: Use pwritev2() with RWF_DSYNC for FUA - update
The testing with "-t writeback" works for turning on enable_write_cache. I renamed the function to qemu_pwritev_fua() and fixed any typos.
I moved the handle_aiocb_flush() into the qemu_pwritev_fua() and removed from the previously todo seciont. Initially I thought of only passing aiocb, but then I was not sure whethe I could derive buf from aiocb, so I added arguments for iovec and iovcnt into qemu_pwritev_fua().
For handling buf in handle_aiocb_rw_linear(), I created iovec and passed its reference. I assumed that there will be only one buffer/iovec, so I passed 1 for iovcnt.
Pinku Deb Nath (2):
block/file-posix.c: Use pwritev2() with RWF_DSYNC for FUA block/file-posix.c: Use pwritev2() with RWF_DSYNC for FUA - update
block/file-posix.c | 54 +++++++++++++++++++++++++++++++++++----------- 1 file changed, 42 insertions(+), 12 deletions(-)
Signature · 1 line
2.43.0
PATCH v2 2/2 block/file-posix.c: Use pwritev2() with RWF_DSYNC for FUA - update
The testing with "-t writeback" works for turning on enable_write_cache. I renamed the function to qemu_pwritev_fua() and fixed any typos.
I moved the handle_aiocb_flush() into the qemu_pwritev_fua() and removed from the previously todo seciont. Initially I thought of only passing aiocb, but then I was not sure whethe I could derive buf from aiocb, so I added arguments for iovec and iovcnt into qemu_pwritev_fua().
For handling buf in handle_aiocb_rw_linear(), I created iovec and passed its reference. I assumed that there will be only one buffer/iovec, so I passed 1 for iovcnt.
--- block/file-posix.c | 38 +++++++++++++++++++++----------------- 1 file changed, 21 insertions(+), 17 deletions(-)
block/file-posix.c +21 -17
@@ -1676,12 +1676,24 @@ qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) }static ssize_t -qemu_pwrite_fua (int fd,const struct iovec *iov,int nr_iov,off_t offset) +qemu_pwritev_fua (const RawPosixAIOData *aiocb,struct iovec *iov,int iovcnt) {#ifdef RWF_DSYNC -return pwritev2 (fd, iov, nr_iov, offset,RWF_DSYNC ); +return pwritev2 (aiocb->aio_fildes, + iov, + iovcnt, + aiocb->aio_offset, +RWF_DSYNC );#else -return pwritev2 (fd, iov, nr_iov, offset,0 ); +ssize_t len =pwritev2 (aiocb->aio_fildes, + iov, + iovcnt, + aiocb->aio_offset, +0 ); +if (len ==0 ) { + len =handle_aiocb_flush (aiocb); + } +return len;#endif }
@@ -1710,10 +1722,7 @@ static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb) len = RETRY_ON_EINTR( (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) ? (aiocb->flags & BDRV_REQ_FUA) ? - qemu_pwrite_fua(aiocb->aio_fildes, - aiocb->io.iov, - aiocb->io.niov, - aiocb->aio_offset) : + qemu_pwritev_fua(aiocb, aiocb->io.iov, aiocb->io.niov) : qemu_pwritev(aiocb->aio_fildes, aiocb->io.iov, aiocb->io.niov,
@@ -1744,10 +1753,11 @@ static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)while (offset < aiocb->aio_nbytes) {if (aiocb->aio_type & (QEMU_AIO_WRITE |QEMU_AIO_ZONE_APPEND )) {if (aiocb->flags &BDRV_REQ_FUA ) { - len =qemu_pwrite_fua (aiocb->aio_fildes, - aiocb->io.iov, - aiocb->io.niov, - aiocb->aio_offset); +struct iovec iov = { + .iov_base = buf, + .iov_len = aiocb->aio_nbytes - offset, + }; + len =qemu_pwritev_fua (aiocb, &iov,1 ); }else { len = pwrite(aiocb->aio_fildes, (const char *)buf + offset,
@@ -2567,12 +2577,6 @@ static int coroutine_fn raw_co_prw(BlockDriverState *bs, int64_t *offset_ptr,assert (qiov->size == bytes); ret =raw_thread_pool_submit (handle_aiocb_rw, &acb); -#ifndef RWD_DSYNC -if (ret ==0 && (flags &BDRV_REQ_FUA )) { -/* TODO Use pwritev2() instead if it's available */ - ret =raw_co_flush_to_disk (bs); - } -#endif goto out;/* Avoid the compiler err of unused label */ out:
Signature · 1 line
2.43.0
PATCH v2 1/2 block/file-posix.c: Use pwritev2() with RWF_DSYNC for FUA
In reply to Pinku Deb NathFull Unit Access (FUA) is an optimization where a disk write with the flag set will be persisted to disk immediately instead of potentially remaining in the disk's write cache. This commit address the todo task for using pwritev2() with RWF_DSYNC in the thread pool section of raw_co_prw(), if pwritev2 with RWF_DSYNC is available in the host, which is alway for Linux kernel >= 4.7. The intent for FUA is indicated with the BDRV_REQ_FUA flag. The old code paths are preserved in case BDRV_REQ_FUA is off or pwritev2() with RWF_DSYNC is not available.
During testing, I observed that the BDRV_REQ_FUA is always turned on when blk->enable_write_cache is not set in block/block-backend.c, so I commented this section off during testing:
https://gitlab.com/qemu-project/qemu/-/blob/master/block/block-backend.c?ref_type=heads#L1432-1434
--- block/file-posix.c | 42 ++++++++++++++++++++++++++++++++++-------- 1 file changed, 34 insertions(+), 8 deletions(-)
block/file-posix.c +34 -8
@@ -229,6 +229,7 @@ typedef struct RawPosixAIOData {unsigned long op; } zone_mgmt; }; +BdrvRequestFlags flags; } RawPosixAIOData;#if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
@@ -1674,6 +1675,16 @@ qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) return pwritev(fd, iov, nr_iov, offset); } +static ssize_t +qemu_pwrite_fua (int fd,const struct iovec *iov,int nr_iov,off_t offset) +{ +#ifdef RWF_DSYNC +return pwritev2 (fd, iov, nr_iov, offset,RWF_DSYNC ); +#else +return pwritev2 (fd, iov, nr_iov, offset,0 ); +#endif +} +#else static bool preadv_present = false;
@@ -1698,10 +1709,15 @@ static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb) len =RETRY_ON_EINTR ( (aiocb->aio_type & (QEMU_AIO_WRITE |QEMU_AIO_ZONE_APPEND )) ? -qemu_pwritev (aiocb->aio_fildes, - aiocb->io.iov, - aiocb->io.niov, - aiocb->aio_offset) : + (aiocb->flags &BDRV_REQ_FUA ) ? +qemu_pwrite_fua (aiocb->aio_fildes, + aiocb->io.iov, + aiocb->io.niov, + aiocb->aio_offset) : +qemu_pwritev (aiocb->aio_fildes, + aiocb->io.iov, + aiocb->io.niov, + aiocb->aio_offset) : qemu_preadv(aiocb->aio_fildes, aiocb->io.iov, aiocb->io.niov,
@@ -1727,10 +1743,17 @@ static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)while (offset < aiocb->aio_nbytes) {if (aiocb->aio_type & (QEMU_AIO_WRITE |QEMU_AIO_ZONE_APPEND )) { - len =pwrite (aiocb->aio_fildes, - (const char *)buf + offset, - aiocb->aio_nbytes - offset, - aiocb->aio_offset + offset); +if (aiocb->flags &BDRV_REQ_FUA ) { + len =qemu_pwrite_fua (aiocb->aio_fildes, + aiocb->io.iov, + aiocb->io.niov, + aiocb->aio_offset); + }else { + len =pwrite (aiocb->aio_fildes, + (const char *)buf + offset, + aiocb->aio_nbytes - offset, + aiocb->aio_offset + offset); + } }else { len = pread(aiocb->aio_fildes, buf + offset,
@@ -2539,14 +2562,17 @@ static int coroutine_fn raw_co_prw(BlockDriverState *bs, int64_t *offset_ptr, .iov = qiov->iov, .niov = qiov->niov, }, + .flags = flags, };assert (qiov->size == bytes); ret =raw_thread_pool_submit (handle_aiocb_rw, &acb); +#ifndef RWD_DSYNC if (ret ==0 && (flags &BDRV_REQ_FUA )) {/* TODO Use pwritev2() instead if it's available */ ret =raw_co_flush_to_disk (bs); } +#endif goto out;/* Avoid the compiler err of unused label */ out:
Signature · 1 line
2.43.0
Quoted 33 lines · Pinku Deb Nath
The testing with "-t writeback" works for turning on
enable_write_cache. I renamed the function toqemu_pwritev_fua()and fixed any typos.I moved the
handle_aiocb_flush()into theqemu_pwritev_fua()and removed from the previously todo seciont. Initially I thought of only passing aiocb, but then I was not sure whethe I could derive buf from aiocb, so I added arguments for iovec and iovcnt intoqemu_pwritev_fua().For handling buf in
handle_aiocb_rw_linear(), I created iovec and passed its reference. I assumed that there will be only one buffer/iovec, so I passed 1 for iovcnt.Signed-off-by Pinku Deb Nath <[email protected]>--- block/file-posix.c | 38 +++++++++++++++++++++----------------- 1 file changed, 21 insertions(+), 17 deletions(-)block/file-posix.c +3
-2@@ -1676,12 +1676,24 @@ qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) }static ssize_t -qemu_pwrite_fua (int fd,const struct iovec *iov,int nr_iov,off_t offset) +qemu_pwritev_fua (const RawPosixAIOData *aiocb,struct iovec *iov,int iovcnt) {#ifdef RWF_DSYNC -return pwritev2 (fd, iov, nr_iov, offset,RWF_DSYNC ); + return pwritev2(aiocb->aio_fildes, + iov,
block/file-posix.c qemu_pwritev(int fd, const struct iovec *iov, int nr_iov, off_t offset) · line 1692
Quoted 10 lines
block/file-posix.c +8
-1+ iovcnt, + aiocb->aio_offset, + RWF_DSYNC);#else -return pwritev2 (fd, iov, nr_iov, offset,0 ); +ssize_t len =pwritev2 (aiocb->aio_fildes, + iov, + iovcnt, + aiocb->aio_offset, +0 );
On a non-Linux host pwritev2(2) will not exist. Please take a look at how qemu_preadv() is integrated (including the !CONFIG_PREADV case) and decide on a solution that works on non-Linux hosts.
Quoted 22 lines
block/file-posix.c +5
-4+if (len ==0 ) { + len =handle_aiocb_flush (aiocb); + } +return len;#endif }@@ -1710,10 +1722,7 @@ static ssize_t handle_aiocb_rw_vector(RawPosixAIOData *aiocb) len = RETRY_ON_EINTR( (aiocb->aio_type & (QEMU_AIO_WRITE | QEMU_AIO_ZONE_APPEND)) ? (aiocb->flags & BDRV_REQ_FUA) ? - qemu_pwrite_fua(aiocb->aio_fildes, - aiocb->io.iov, - aiocb->io.niov, - aiocb->aio_offset) : + qemu_pwritev_fua(aiocb, aiocb->io.iov, aiocb->io.niov) : qemu_pwritev(aiocb->aio_fildes, aiocb->io.iov, aiocb->io.niov,@@ -1744,10 +1753,11 @@ static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf)while (offset < aiocb->aio_nbytes) {if (aiocb->aio_type & (QEMU_AIO_WRITE |QEMU_AIO_ZONE_APPEND )) {
block/file-posix.c static ssize_t handle_aiocb_rw_linear(RawPosixAIOData *aiocb, char *buf) · line 1760
Quoted 10 lines
block/file-posix.c +5
-4if (aiocb->flags &BDRV_REQ_FUA ) { - len =qemu_pwrite_fua (aiocb->aio_fildes, - aiocb->io.iov, - aiocb->io.niov, - aiocb->aio_offset); +struct iovec iov = { + .iov_base = buf, + .iov_len = aiocb->aio_nbytes - offset, + }; + len =qemu_pwritev_fua (aiocb, &iov,1 );
The else branch takes offset into account. Here aiocb is passed in assuming it's the first iteration of the while (offset < aiocb->aio_nbytes) loop. On subsequent iterations the wrong values will be used because offset has changed.
Perhaps it's easier to pass in the individual parameters (fd, offset, etc) instead of passing in aiocb.
Quoted 19 lines
block/file-posix.c +0
-6}else { len = pwrite(aiocb->aio_fildes, (const char *)buf + offset,@@ -2567,12 +2577,6 @@ static int coroutine_fn raw_co_prw(BlockDriverState *bs, int64_t *offset_ptr,assert (qiov->size == bytes); ret =raw_thread_pool_submit (handle_aiocb_rw, &acb); -#ifndef RWD_DSYNC -if (ret ==0 && (flags &BDRV_REQ_FUA )) { -/* TODO Use pwritev2() instead if it's available */ - ret =raw_co_flush_to_disk (bs); - } -#endif goto out;/* Avoid the compiler err of unused label */ out:Signature · 1 line
2.43.0
Quoted 13 lines · Pinku Deb Nath
The testing with "-t writeback" works for turning on
enable_write_cache. I renamed the function toqemu_pwritev_fua()and fixed any typos.I moved the
handle_aiocb_flush()into theqemu_pwritev_fua()and removed from the previously todo seciont. Initially I thought of only passing aiocb, but then I was not sure whethe I could derive buf from aiocb, so I added arguments for iovec and iovcnt intoqemu_pwritev_fua().For handling buf in
handle_aiocb_rw_linear(), I created iovec and passed its reference. I assumed that there will be only one buffer/iovec, so I passed 1 for iovcnt.
Quoted 8 lines
Signed-off-by Pinku Deb Nath <[email protected]>Pinku Deb Nath (2):
block/file-posix.c: Usepwritev2()withRWF_DSYNCfor FUAblock/file-posix.c: Usepwritev2()withRWF_DSYNCfor FUA - updateblock/file-posix.c | 54 +++++++++++++++++++++++++++++++++++----------- 1 file changed, 42 insertions(+), 12 deletions(-)
Thanks for sending this updates patch series. Please squash changes in the future instead of appending them as separate commits. This means editing previous commits (e.g. git rebase -i master) so that they contain changes made after code review.
So if commit 1 is '+ printf("foo\n")', then instead of adding commit 2 to add a semi-colon to the end of the line, just edit the commit so it is '+ printf("foo\n");' in v2 of your patch.
One reason to squash changes is so that git-bisect(1) works. Without squashing, there will be intermediate commits that are broken and maybe don't even compile. git-bisect(1) is only usable when each commit compiles and passes tests.
Reviews also tend to prefer to see the final state of commits so they don't have to review every incremental edit that was made (often replacing code they already reviewed). It saves them time.
Thanks,
Stefan