| Message ID | 20210513164912.5683-1-bvanassche@acm.org (mailing list archive) |
|---|---|
| State | Accepted |
| Headers | show |
| Series | ufs: Increase the usable queue depth | expand |
Hi Bart, On 2021-05-14 00:49, Bart Van Assche wrote: > With the current implementation of the UFS driver active_queues is 1 > instead of 0 if all UFS request queues are idle. That causes > hctx_may_queue() to divide the queue depth by 2 when queueing a request > and hence reduces the usable queue depth. This is interesting. When all UFS queues are idle, in hctx_may_queue(), active_queues reads 1 (users == 1, depth == 32), where is it divided by 2? static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, struct sbitmap_queue *bt) { unsigned int depth, users; .... users = atomic_read(&hctx->tags->active_queues); } if (!users) return true; /* * Allow at least some tags */ depth = max((bt->sb.depth + users - 1) / users, 4U); return __blk_mq_active_requests(hctx) < depth; } Thanks, Can Guo. > > The shared tag set code in the block layer keeps track of the number of > active request queues. blk_mq_tag_busy() is called before a request is > queued onto a hwq and blk_mq_tag_idle() is called some time after the > hwq > became idle. blk_mq_tag_idle() is called from inside > blk_mq_timeout_work(). > Hence, blk_mq_tag_idle() is only called if a timer is associated with > each > request that is submitted to a request queue that shares a tag set with > another request queue. Hence this patch that adds a > blk_mq_start_request() > call in ufshcd_exec_dev_cmd(). This patch doubles the queue depth on my > test setup from 16 to 32. > > In addition to increasing the usable queue depth, also fix the > documentation of the 'timeout' parameter in the header above > ufshcd_exec_dev_cmd(). > > Cc: Can Guo <cang@codeaurora.org> > Cc: Alim Akhtar <alim.akhtar@samsung.com> > Cc: Avri Altman <avri.altman@wdc.com> > Cc: Stanley Chu <stanley.chu@mediatek.com> > Cc: Bean Huo <beanhuo@micron.com> > Cc: Adrian Hunter <adrian.hunter@intel.com> > Fixes: 7252a3603015 ("scsi: ufs: Avoid busy-waiting by eliminating tag > conflicts") > Signed-off-by: Bart Van Assche <bvanassche@acm.org> > --- > drivers/scsi/ufs/ufshcd.c | 5 ++++- > 1 file changed, 4 insertions(+), 1 deletion(-) > > diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c > index c96e36aab989..e669243354da 100644 > --- a/drivers/scsi/ufs/ufshcd.c > +++ b/drivers/scsi/ufs/ufshcd.c > @@ -2838,7 +2838,7 @@ static int ufshcd_wait_for_dev_cmd(struct ufs_hba > *hba, > * ufshcd_exec_dev_cmd - API for sending device management requests > * @hba: UFS hba > * @cmd_type: specifies the type (NOP, Query...) > - * @timeout: time in seconds > + * @timeout: timeout in milliseconds > * > * NOTE: Since there is only one available tag for device management > commands, > * it is expected you hold the hba->dev_cmd.lock mutex. > @@ -2868,6 +2868,9 @@ static int ufshcd_exec_dev_cmd(struct ufs_hba > *hba, > } > tag = req->tag; > WARN_ON_ONCE(!ufshcd_valid_tag(hba, tag)); > + /* Set the timeout such that the SCSI error handler is not activated. > */ > + req->timeout = msecs_to_jiffies(2 * timeout); > + blk_mq_start_request(req); > > init_completion(&wait); > lrbp = &hba->lrb[tag];
On 5/13/21 9:04 PM, Can Guo wrote: > Hi Bart, > > On 2021-05-14 00:49, Bart Van Assche wrote: >> With the current implementation of the UFS driver active_queues is 1 >> instead of 0 if all UFS request queues are idle. That causes >> hctx_may_queue() to divide the queue depth by 2 when queueing a request >> and hence reduces the usable queue depth. > > This is interesting. When all UFS queues are idle, in hctx_may_queue(), > active_queues reads 1 (users == 1, depth == 32), where is it divided by 2? > > static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, > struct sbitmap_queue *bt) > { > unsigned int depth, users; > > .... > users = atomic_read(&hctx->tags->active_queues); > } > > if (!users) > return true; > > /* > * Allow at least some tags > */ > depth = max((bt->sb.depth + users - 1) / users, 4U); > return __blk_mq_active_requests(hctx) < depth; > } Hi Can, If no I/O scheduler has been configured then the active_queues counter is increased from inside blk_get_request() by blk_mq_tag_busy() before hctx_may_queue() is called. So if active_queues == 1 when the UFS device is idle, the active_queues counter will be increased to 2 if a request is submitted to another request queue than hba->cmd_queue. This will cause the hctx_may_queue() calls from inside __blk_mq_alloc_request() and __blk_mq_get_driver_tag() to limit the queue depth to 32 / 2 = 16. If an I/O scheduler has been configured then __blk_mq_get_driver_tag() will be the first function to call blk_mq_tag_busy() while processing a request. The hctx_may_queue() call in __blk_mq_get_driver_tag() will limit the queue depth to 32 / 2 = 16 if an I/O scheduler has been configured. Bart.
On 2021-05-14 12:04, Can Guo wrote: > Hi Bart, > > On 2021-05-14 00:49, Bart Van Assche wrote: >> With the current implementation of the UFS driver active_queues is 1 >> instead of 0 if all UFS request queues are idle. That causes >> hctx_may_queue() to divide the queue depth by 2 when queueing a >> request >> and hence reduces the usable queue depth. > > This is interesting. When all UFS queues are idle, in hctx_may_queue(), > active_queues reads 1 (users == 1, depth == 32), where is it divided by > 2? > Are you saying that if we queue a new request on one of the UFS request queues, since the active_queues is always above 0, we can never use the full queue depth? If so, then I agree. Reviewed-by: Can Guo <cang@codeaurora.org> > static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, > struct sbitmap_queue *bt) > { > unsigned int depth, users; > > .... > users = atomic_read(&hctx->tags->active_queues); > } > > if (!users) > return true; > > /* > * Allow at least some tags > */ > depth = max((bt->sb.depth + users - 1) / users, 4U); > return __blk_mq_active_requests(hctx) < depth; > } > > Thanks, > Can Guo. > >> >> The shared tag set code in the block layer keeps track of the number >> of >> active request queues. blk_mq_tag_busy() is called before a request is >> queued onto a hwq and blk_mq_tag_idle() is called some time after the >> hwq >> became idle. blk_mq_tag_idle() is called from inside >> blk_mq_timeout_work(). >> Hence, blk_mq_tag_idle() is only called if a timer is associated with >> each >> request that is submitted to a request queue that shares a tag set >> with >> another request queue. Hence this patch that adds a >> blk_mq_start_request() >> call in ufshcd_exec_dev_cmd(). This patch doubles the queue depth on >> my >> test setup from 16 to 32. >> >> In addition to increasing the usable queue depth, also fix the >> documentation of the 'timeout' parameter in the header above >> ufshcd_exec_dev_cmd(). >> >> Cc: Can Guo <cang@codeaurora.org> >> Cc: Alim Akhtar <alim.akhtar@samsung.com> >> Cc: Avri Altman <avri.altman@wdc.com> >> Cc: Stanley Chu <stanley.chu@mediatek.com> >> Cc: Bean Huo <beanhuo@micron.com> >> Cc: Adrian Hunter <adrian.hunter@intel.com> >> Fixes: 7252a3603015 ("scsi: ufs: Avoid busy-waiting by eliminating tag >> conflicts") >> Signed-off-by: Bart Van Assche <bvanassche@acm.org> >> --- >> drivers/scsi/ufs/ufshcd.c | 5 ++++- >> 1 file changed, 4 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c >> index c96e36aab989..e669243354da 100644 >> --- a/drivers/scsi/ufs/ufshcd.c >> +++ b/drivers/scsi/ufs/ufshcd.c >> @@ -2838,7 +2838,7 @@ static int ufshcd_wait_for_dev_cmd(struct >> ufs_hba *hba, >> * ufshcd_exec_dev_cmd - API for sending device management requests >> * @hba: UFS hba >> * @cmd_type: specifies the type (NOP, Query...) >> - * @timeout: time in seconds >> + * @timeout: timeout in milliseconds >> * >> * NOTE: Since there is only one available tag for device management >> commands, >> * it is expected you hold the hba->dev_cmd.lock mutex. >> @@ -2868,6 +2868,9 @@ static int ufshcd_exec_dev_cmd(struct ufs_hba >> *hba, >> } >> tag = req->tag; >> WARN_ON_ONCE(!ufshcd_valid_tag(hba, tag)); >> + /* Set the timeout such that the SCSI error handler is not >> activated. */ >> + req->timeout = msecs_to_jiffies(2 * timeout); >> + blk_mq_start_request(req); >> >> init_completion(&wait); >> lrbp = &hba->lrb[tag];
On 2021-05-14 12:19, Bart Van Assche wrote: > On 5/13/21 9:04 PM, Can Guo wrote: >> Hi Bart, >> >> On 2021-05-14 00:49, Bart Van Assche wrote: >>> With the current implementation of the UFS driver active_queues is 1 >>> instead of 0 if all UFS request queues are idle. That causes >>> hctx_may_queue() to divide the queue depth by 2 when queueing a >>> request >>> and hence reduces the usable queue depth. >> >> This is interesting. When all UFS queues are idle, in >> hctx_may_queue(), >> active_queues reads 1 (users == 1, depth == 32), where is it divided >> by 2? >> >> static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, >> struct sbitmap_queue *bt) >> { >> unsigned int depth, users; >> >> .... >> users = atomic_read(&hctx->tags->active_queues); >> } >> >> if (!users) >> return true; >> >> /* >> * Allow at least some tags >> */ >> depth = max((bt->sb.depth + users - 1) / users, 4U); >> return __blk_mq_active_requests(hctx) < depth; >> } > > Hi Can, > > If no I/O scheduler has been configured then the active_queues counter > is increased from inside blk_get_request() by blk_mq_tag_busy() before > hctx_may_queue() is called. So if active_queues == 1 when the UFS > device > is idle, the active_queues counter will be increased to 2 if a request > is submitted to another request queue than hba->cmd_queue. This will > cause the hctx_may_queue() calls from inside __blk_mq_alloc_request() > and __blk_mq_get_driver_tag() to limit the queue depth to 32 / 2 = 16. > > If an I/O scheduler has been configured then __blk_mq_get_driver_tag() > will be the first function to call blk_mq_tag_busy() while processing a > request. The hctx_may_queue() call in __blk_mq_get_driver_tag() will > limit the queue depth to 32 / 2 = 16 if an I/O scheduler has been > configured. > > Bart. Yes, I just figured out what you are saying from the commit message and gave my reviewed-by tag. Thanks for the explanation and the fix. Regards, Can Guo.
On 2021-05-14 12:24, Can Guo wrote: > On 2021-05-14 12:19, Bart Van Assche wrote: >> On 5/13/21 9:04 PM, Can Guo wrote: >>> Hi Bart, >>> >>> On 2021-05-14 00:49, Bart Van Assche wrote: >>>> With the current implementation of the UFS driver active_queues is 1 >>>> instead of 0 if all UFS request queues are idle. That causes >>>> hctx_may_queue() to divide the queue depth by 2 when queueing a >>>> request >>>> and hence reduces the usable queue depth. >>> >>> This is interesting. When all UFS queues are idle, in >>> hctx_may_queue(), >>> active_queues reads 1 (users == 1, depth == 32), where is it divided >>> by 2? >>> >>> static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx, >>> struct sbitmap_queue *bt) >>> { >>> unsigned int depth, users; >>> >>> .... >>> users = atomic_read(&hctx->tags->active_queues); >>> } >>> >>> if (!users) >>> return true; >>> >>> /* >>> * Allow at least some tags >>> */ >>> depth = max((bt->sb.depth + users - 1) / users, 4U); >>> return __blk_mq_active_requests(hctx) < depth; >>> } >> >> Hi Can, >> >> If no I/O scheduler has been configured then the active_queues counter >> is increased from inside blk_get_request() by blk_mq_tag_busy() before >> hctx_may_queue() is called. So if active_queues == 1 when the UFS >> device >> is idle, the active_queues counter will be increased to 2 if a request >> is submitted to another request queue than hba->cmd_queue. This will >> cause the hctx_may_queue() calls from inside __blk_mq_alloc_request() >> and __blk_mq_get_driver_tag() to limit the queue depth to 32 / 2 = 16. >> >> If an I/O scheduler has been configured then __blk_mq_get_driver_tag() >> will be the first function to call blk_mq_tag_busy() while processing >> a >> request. The hctx_may_queue() call in __blk_mq_get_driver_tag() will >> limit the queue depth to 32 / 2 = 16 if an I/O scheduler has been >> configured. >> >> Bart. > > Yes, I just figured out what you are saying from the commit message and > gave my reviewed-by tag. Thanks for the explanation and the fix. > > Regards, > Can Guo. We definitely need to have fix present on Android12-5.10, because performance may be impacted without it... Thanks, Can Guo.
On Thu, 13 May 2021 09:49:12 -0700, Bart Van Assche wrote: > With the current implementation of the UFS driver active_queues is 1 > instead of 0 if all UFS request queues are idle. That causes > hctx_may_queue() to divide the queue depth by 2 when queueing a request > and hence reduces the usable queue depth. > > The shared tag set code in the block layer keeps track of the number of > active request queues. blk_mq_tag_busy() is called before a request is > queued onto a hwq and blk_mq_tag_idle() is called some time after the hwq > became idle. blk_mq_tag_idle() is called from inside blk_mq_timeout_work(). > Hence, blk_mq_tag_idle() is only called if a timer is associated with each > request that is submitted to a request queue that shares a tag set with > another request queue. Hence this patch that adds a blk_mq_start_request() > call in ufshcd_exec_dev_cmd(). This patch doubles the queue depth on my > test setup from 16 to 32. > > [...] Applied to 5.13/scsi-fixes, thanks! [1/1] ufs: Increase the usable queue depth https://git.kernel.org/mkp/scsi/c/d0b2b70eb12e
Hi Bart, On 2021-05-14 00:49, Bart Van Assche wrote: > With the current implementation of the UFS driver active_queues is 1 > instead of 0 if all UFS request queues are idle. That causes > hctx_may_queue() to divide the queue depth by 2 when queueing a request > and hence reduces the usable queue depth. > > The shared tag set code in the block layer keeps track of the number of > active request queues. blk_mq_tag_busy() is called before a request is > queued onto a hwq and blk_mq_tag_idle() is called some time after the > hwq > became idle. blk_mq_tag_idle() is called from inside > blk_mq_timeout_work(). > Hence, blk_mq_tag_idle() is only called if a timer is associated with > each > request that is submitted to a request queue that shares a tag set with > another request queue. Hence this patch that adds a > blk_mq_start_request() > call in ufshcd_exec_dev_cmd(). This patch doubles the queue depth on my > test setup from 16 to 32. > > In addition to increasing the usable queue depth, also fix the > documentation of the 'timeout' parameter in the header above > ufshcd_exec_dev_cmd(). > > Cc: Can Guo <cang@codeaurora.org> > Cc: Alim Akhtar <alim.akhtar@samsung.com> > Cc: Avri Altman <avri.altman@wdc.com> > Cc: Stanley Chu <stanley.chu@mediatek.com> > Cc: Bean Huo <beanhuo@micron.com> > Cc: Adrian Hunter <adrian.hunter@intel.com> > Fixes: 7252a3603015 ("scsi: ufs: Avoid busy-waiting by eliminating tag > conflicts") > Signed-off-by: Bart Van Assche <bvanassche@acm.org> > --- > drivers/scsi/ufs/ufshcd.c | 5 ++++- > 1 file changed, 4 insertions(+), 1 deletion(-) > > diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c > index c96e36aab989..e669243354da 100644 > --- a/drivers/scsi/ufs/ufshcd.c > +++ b/drivers/scsi/ufs/ufshcd.c > @@ -2838,7 +2838,7 @@ static int ufshcd_wait_for_dev_cmd(struct ufs_hba > *hba, > * ufshcd_exec_dev_cmd - API for sending device management requests > * @hba: UFS hba > * @cmd_type: specifies the type (NOP, Query...) > - * @timeout: time in seconds > + * @timeout: timeout in milliseconds > * > * NOTE: Since there is only one available tag for device management > commands, > * it is expected you hold the hba->dev_cmd.lock mutex. > @@ -2868,6 +2868,9 @@ static int ufshcd_exec_dev_cmd(struct ufs_hba > *hba, > } > tag = req->tag; > WARN_ON_ONCE(!ufshcd_valid_tag(hba, tag)); > + /* Set the timeout such that the SCSI error handler is not activated. > */ > + req->timeout = msecs_to_jiffies(2 * timeout); > + blk_mq_start_request(req); > > init_completion(&wait); > lrbp = &hba->lrb[tag]; We found a regression after this change gets merged - schedule blk_mq_get_tag __blk_mq_alloc_request blk_get_request ufshcd_exec_dev_cmd ufshcd_query_flag ufshcd_wb_ctrl ufshcd_devfreq_scale ufshcd_devfreq_target devfreq_set_target update_devfreq devfreq_monitor process_one_work worker_thread kthread ret_from_fork Since ufshcd_devfreq_scale() blocks scsi requests, when ufshcd_wb_ctrl() runs, if it cannot get a free tag (all tags are taken by normal requests), then ufshcd_devfreq_scale() gets stuck, thus scsi layer stays blocked, which leads to I/O hung. Maybe consider unblocking scsi requests before call ufshcd_wb_ctrl()? Thanks, Can Guo.
diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c index c96e36aab989..e669243354da 100644 --- a/drivers/scsi/ufs/ufshcd.c +++ b/drivers/scsi/ufs/ufshcd.c @@ -2838,7 +2838,7 @@ static int ufshcd_wait_for_dev_cmd(struct ufs_hba *hba, * ufshcd_exec_dev_cmd - API for sending device management requests * @hba: UFS hba * @cmd_type: specifies the type (NOP, Query...) - * @timeout: time in seconds + * @timeout: timeout in milliseconds * * NOTE: Since there is only one available tag for device management commands, * it is expected you hold the hba->dev_cmd.lock mutex. @@ -2868,6 +2868,9 @@ static int ufshcd_exec_dev_cmd(struct ufs_hba *hba, } tag = req->tag; WARN_ON_ONCE(!ufshcd_valid_tag(hba, tag)); + /* Set the timeout such that the SCSI error handler is not activated. */ + req->timeout = msecs_to_jiffies(2 * timeout); + blk_mq_start_request(req); init_completion(&wait); lrbp = &hba->lrb[tag];
With the current implementation of the UFS driver active_queues is 1 instead of 0 if all UFS request queues are idle. That causes hctx_may_queue() to divide the queue depth by 2 when queueing a request and hence reduces the usable queue depth. The shared tag set code in the block layer keeps track of the number of active request queues. blk_mq_tag_busy() is called before a request is queued onto a hwq and blk_mq_tag_idle() is called some time after the hwq became idle. blk_mq_tag_idle() is called from inside blk_mq_timeout_work(). Hence, blk_mq_tag_idle() is only called if a timer is associated with each request that is submitted to a request queue that shares a tag set with another request queue. Hence this patch that adds a blk_mq_start_request() call in ufshcd_exec_dev_cmd(). This patch doubles the queue depth on my test setup from 16 to 32. In addition to increasing the usable queue depth, also fix the documentation of the 'timeout' parameter in the header above ufshcd_exec_dev_cmd(). Cc: Can Guo <cang@codeaurora.org> Cc: Alim Akhtar <alim.akhtar@samsung.com> Cc: Avri Altman <avri.altman@wdc.com> Cc: Stanley Chu <stanley.chu@mediatek.com> Cc: Bean Huo <beanhuo@micron.com> Cc: Adrian Hunter <adrian.hunter@intel.com> Fixes: 7252a3603015 ("scsi: ufs: Avoid busy-waiting by eliminating tag conflicts") Signed-off-by: Bart Van Assche <bvanassche@acm.org> --- drivers/scsi/ufs/ufshcd.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)