scsi: iscsi: register sysfs for iscsi workqueue
PATCH v2 1/1 scsi: iscsi: register sysfs for iscsi workqueue
Then users can set cpu affinity through "cpumask" for iscsi workqueues, so as to get performance isolation.
This patch changes the max number of active worker form 1 to 2, since ordered workqueue wont' be allowed to change "cpumask".
--- drivers/scsi/libiscsi.c | 4 +++- drivers/scsi/scsi_transport_iscsi.c | 4 +++- 2 files changed, 6 insertions(+), 2 deletions(-)
drivers/scsi/libiscsi.c +3 -1
@@ -2627,7 +2627,9 @@ struct Scsi_Host *iscsi_host_alloc(struct scsi_host_template *sht,if (xmit_can_sleep) {snprintf (ihost->workq_name,sizeof (ihost->workq_name),"iscsi_q_%d" , shost->host_no); - ihost->workq =create_singlethread_workqueue (ihost->workq_name); + ihost->workq =alloc_workqueue ("%s" , +WQ_SYSFS | __WQ_LEGACY |WQ_MEM_RECLAIM |WQ_UNBOUND , +2 , ihost->workq_name);if (!ihost->workq) goto free_host; }
drivers/scsi/scsi_transport_iscsi.c +3 -1
@@ -4602,7 +4602,9 @@ static __init int iscsi_transport_init(void) goto unregister_flashnode_bus; } - iscsi_eh_timer_workq =create_singlethread_workqueue ("iscsi_eh" ); + iscsi_eh_timer_workq =alloc_workqueue ("%s" , +WQ_SYSFS | __WQ_LEGACY |WQ_MEM_RECLAIM |WQ_UNBOUND , +2 ,"iscsi_eh" );if (!iscsi_eh_timer_workq) { err = -ENOMEM ; goto release_nls;
Signature · 1 line
2.9.5
Friendly ping..
Quoted 43 lines · Bob Liu
Then users can set cpu affinity through "cpumask" for iscsi workqueues, so as to get performance isolation.
This patch changes the max number of active worker form 1 to 2, since ordered workqueue wont' be allowed to change "cpumask".
Signed-off-by Bob Liu <[email protected]>--- drivers/scsi/libiscsi.c | 4 +++- drivers/scsi/scsi_transport_iscsi.c | 4 +++- 2 files changed, 6 insertions(+), 2 deletions(-)drivers/scsi/libiscsi.c +3
-1@@ -2627,7 +2627,9 @@ struct Scsi_Host *iscsi_host_alloc(struct scsi_host_template *sht,if (xmit_can_sleep) {snprintf (ihost->workq_name,sizeof (ihost->workq_name),"iscsi_q_%d" , shost->host_no); - ihost->workq =create_singlethread_workqueue (ihost->workq_name); + ihost->workq =alloc_workqueue ("%s" , +WQ_SYSFS | __WQ_LEGACY |WQ_MEM_RECLAIM |WQ_UNBOUND , +2 , ihost->workq_name);if (!ihost->workq) goto free_host; }drivers/scsi/scsi_transport_iscsi.c +3
-1@@ -4602,7 +4602,9 @@ static __init int iscsi_transport_init(void) goto unregister_flashnode_bus; } - iscsi_eh_timer_workq =create_singlethread_workqueue ("iscsi_eh" ); + iscsi_eh_timer_workq =alloc_workqueue ("%s" , +WQ_SYSFS | __WQ_LEGACY |WQ_MEM_RECLAIM |WQ_UNBOUND , +2 ,"iscsi_eh" );if (!iscsi_eh_timer_workq) { err = -ENOMEM ; goto release_nls;
Quoted 6 lines · Bob Liu
Then users can set cpu affinity through "cpumask" for iscsi workqueues, so as to get performance isolation.
This patch changes the max number of active worker form 1 to 2, since ordered workqueue wont' be allowed to change "cpumask".
Won't having 2 workers break the current ordering guarantees? Did you check that no-one depends on this?
Quoted 11 lines
Signed-off-by Bob Liu <[email protected]>--- drivers/scsi/libiscsi.c | 4 +++- drivers/scsi/scsi_transport_iscsi.c | 4 +++- 2 files changed, 6 insertions(+), 2 deletions(-)drivers/scsi/libiscsi.c +0
-0
drivers/scsi/libiscsi.c struct Scsi_Host *iscsi_host_alloc(struct scsi_host_template *sht, · line 2632
Quoted 7 lines
drivers/scsi/libiscsi.c +3
-1@@ -2627,7 +2627,9 @@ struct Scsi_Host *iscsi_host_alloc(struct scsi_host_template *sht,if (xmit_can_sleep) {snprintf (ihost->workq_name,sizeof (ihost->workq_name),"iscsi_q_%d" , shost->host_no); - ihost->workq =create_singlethread_workqueue (ihost->workq_name); + ihost->workq =alloc_workqueue ("%s" , +WQ_SYSFS | __WQ_LEGACY |WQ_MEM_RECLAIM |WQ_UNBOUND , +2 , ihost->workq_name);
From looking at the documentation, __WQ_LEGACY doesn't seem to belong here.
Quoted 19 lines
drivers/scsi/libiscsi.c +0
-0if (!ihost->workq) goto free_host; }drivers/scsi/scsi_transport_iscsi.c +3
-1@@ -4602,7 +4602,9 @@ static __init int iscsi_transport_init(void) goto unregister_flashnode_bus; } - iscsi_eh_timer_workq =create_singlethread_workqueue ("iscsi_eh" ); + iscsi_eh_timer_workq =alloc_workqueue ("%s" , +WQ_SYSFS | __WQ_LEGACY |WQ_MEM_RECLAIM |WQ_UNBOUND , +2 ,"iscsi_eh" );if (!iscsi_eh_timer_workq) { err = -ENOMEM ; goto release_nls;
Sorry forgot Cc iscsi maintainers.
Quoted 9 lines · Julian Wiedmann
Quoted 6 lines · Bob Liu
Then users can set cpu affinity through "cpumask" for iscsi workqueues, so as to get performance isolation.
This patch changes the max number of active worker form 1 to 2, since ordered workqueue wont' be allowed to change "cpumask".
Won't having 2 workers break the current ordering guarantees?
Yes.
Quoted 2 lines
Did you check that no-one depends on this?
Not quite sure, I'd better add [rfc] to this patch.
But this is the easiest way to set cpu affinity of iscsi workqueue.
Quoted 6 lines
Quoted 6 lines
Signed-off-by Bob Liu <[email protected]>--- drivers/scsi/libiscsi.c | 4 +++- drivers/scsi/scsi_transport_iscsi.c | 4 +++- 2 files changed, 6 insertions(+), 2 deletions(-)
Quoted 15 lines
drivers/scsi/libiscsi.c struct Scsi_Host *iscsi_host_alloc(struct scsi_host_template *sht, · line 2632
Quoted 12 lines
drivers/scsi/libiscsi.c +3
-1@@ -2627,7 +2627,9 @@ struct Scsi_Host *iscsi_host_alloc(struct scsi_host_template *sht,if (xmit_can_sleep) {snprintf (ihost->workq_name,sizeof (ihost->workq_name),"iscsi_q_%d" , shost->host_no); - ihost->workq =create_singlethread_workqueue (ihost->workq_name); + ihost->workq =alloc_workqueue ("%s" , +WQ_SYSFS | __WQ_LEGACY |WQ_MEM_RECLAIM |WQ_UNBOUND , +2 , ihost->workq_name);From looking at the documentation,
__WQ_LEGACYdoesn't seem to belong here.
Indeed, thanks for the review.
Quoted 20 lines
Quoted 19 lines
drivers/scsi/libiscsi.c +0
-0if (!ihost->workq) goto free_host; }drivers/scsi/scsi_transport_iscsi.c +3
-1@@ -4602,7 +4602,9 @@ static __init int iscsi_transport_init(void) goto unregister_flashnode_bus; } - iscsi_eh_timer_workq =create_singlethread_workqueue ("iscsi_eh" ); + iscsi_eh_timer_workq =alloc_workqueue ("%s" , +WQ_SYSFS | __WQ_LEGACY |WQ_MEM_RECLAIM |WQ_UNBOUND , +2 ,"iscsi_eh" );if (!iscsi_eh_timer_workq) { err = -ENOMEM ; goto release_nls;