tcm: Convert to cpu_relax() for atomic_read() in tight loops

PATCH In review Nicholas A. Bellinger 2 participants 3 messages Sep 9, 2010 – Sep 10, 2010

Subsystems (2)

SCSI SCSI target

lore

Share & tools

lore

Nicholas A. Bellinger

author

2010-09-09 21:29 UTC #0 lore raw

PATCH 1/1 tcm: Convert to cpu_relax() for atomic_read() in tight loops

From: Nicholas Bellinger <[email protected]>

Greetings all,

This patch converts the following functions from within shutdown and release path code to use cpu_relax() instead of msleep() while doing atomic_read() in a tight loop waiting for various PR and MIB reference count(s) to return to zero:

	*) core_update_device_list_for_node() + struct se_dev_entry->pr_ref_count
	*) core_release_port() + struct se_port->sep_tg_pt_ref_cnt
	*) se_dev_stop() + struct se_hba->dev_mib_access_count
	*) __core_scsi3_free_registration() + struct t10_pr_registration->pr_res_holders
	*) core_tpg_wait_for_nacl_pr_ref() + struct se_node_acl->acl_pr_ref_count
	*) core_tpg_wait_for_mib_ref() + struct se_node_acl->mib_ref_count
	*) core_tpg_deregister() + struct se_portal_group->tpg_pr_ref_count
	*) transport_deregister_session() + struct se_session->mib_ref_count

Thanks to Konrad Rzeszutek Wilk for mentioning cpu_relax() usage!

Signed-off-by Nicholas A. Bellinger <[email protected]>

---
 drivers/target/target_core_device.c    |    6 +++---
 drivers/target/target_core_pr.c        |    2 +-
 drivers/target/target_core_tpg.c       |    6 +++---
 drivers/target/target_core_transport.c |    2 +-
 4 files changed, 8 insertions(+), 8 deletions(-)

drivers/target/target_core_device.c +3 -3

index 665b8e2..459f0f8 100644
@@ -532,7 +532,7 @@ int core_update_device_list_for_node(
 	 */
 	spin_unlock_irq(&nacl->device_list_lock);
 	while (atomic_read(&deve->pr_ref_count) != 0)
-		msleep(100);
+		cpu_relax();
 	spin_lock_irq(&nacl->device_list_lock);
 	/*
 	 * Disable struct se_dev_entry LUN ACL mapping

@@ -687,7 +687,7 @@ void core_release_port(struct se_device *dev, struct se_port *port)
 	 */
 	spin_unlock(&dev->se_port_lock);
 	if (atomic_read(&port->sep_tg_pt_ref_cnt))
-		msleep(100);
+		cpu_relax();
 	spin_lock(&dev->se_port_lock);
 
 	core_alua_free_tg_pt_gp_mem(port);

@@ -965,7 +965,7 @@ void se_dev_stop(struct se_device *dev)
 	spin_unlock(&hba->device_lock);
 
 	while (atomic_read(&hba->dev_mib_access_count))
-		msleep(10);
+		cpu_relax();
 }
 
 int se_dev_check_online(struct se_device *dev)

drivers/target/target_core_pr.c +1 -1

index 2cebafa..11118fb 100644
@@ -1289,7 +1289,7 @@ static void __core_scsi3_free_registration(
 		spin_unlock(&pr_tmpl->registration_lock);
 		printk("SPC-3 PR [%s] waiting for pr_res_holders\n",
 				tfo->get_fabric_name());
-		msleep(10);
+		cpu_relax();
 		spin_lock(&pr_tmpl->registration_lock);
 	}

drivers/target/target_core_tpg.c +3 -3

index 75c3900..e598f2b 100644
@@ -315,13 +315,13 @@ EXPORT_SYMBOL(core_tpg_check_initiator_node_acl);
 void core_tpg_wait_for_nacl_pr_ref(struct se_node_acl *nacl)
 {
 	while (atomic_read(&nacl->acl_pr_ref_count) != 0)
-		msleep(100);
+		cpu_relax();
 }
 
 void core_tpg_wait_for_mib_ref(struct se_node_acl *nacl)
 {
 	while (atomic_read(&nacl->mib_ref_count) != 0)
-		msleep(100);
+		cpu_relax();
 }
 
 void core_tpg_clear_object_luns(struct se_portal_group *tpg)

@@ -713,7 +713,7 @@ int core_tpg_deregister(struct se_portal_group *se_tpg)
 	spin_unlock_bh(&se_global->se_tpg_lock);
 
 	while (atomic_read(&se_tpg->tpg_pr_ref_count) != 0)
-		msleep(100);
+		cpu_relax();
 
 	if (se_tpg->se_tpg_type == TRANSPORT_TPG_TYPE_NORMAL)
 		core_tpg_release_virtual_lun0(se_tpg);

drivers/target/target_core_transport.c +1 -1

index 9e7582b..95401cb 100644
@@ -750,7 +750,7 @@ void transport_deregister_session(struct se_session *se_sess)
 	 * scsi_att_intr_port_seq_show()
 	 */
 	while (atomic_read(&se_sess->mib_ref_count) != 0)
-		msleep(100);
+		cpu_relax();
 
 	spin_lock_bh(&se_tpg->session_lock);
 	list_del(&se_sess->sess_list);

Signature · 1 line

1.5.6.5

Jiri Slaby

2010-09-10 06:32 UTC #1 lore raw

Quoted 32 lines · Nicholas A. Bellinger

From: Nicholas Bellinger <[email protected]>
Greetings all,
This patch converts the following functions from within shutdown and release path code to use cpu_relax() instead of msleep() while doing atomic_read() in a tight loop waiting for various PR and MIB reference count(s) to return to zero:
	*) core_update_device_list_for_node() + struct se_dev_entry->pr_ref_count
	*) core_release_port() + struct se_port->sep_tg_pt_ref_cnt
	*) se_dev_stop() + struct se_hba->dev_mib_access_count
	*) __core_scsi3_free_registration() + struct t10_pr_registration->pr_res_holders
	*) core_tpg_wait_for_nacl_pr_ref() + struct se_node_acl->acl_pr_ref_count
	*) core_tpg_wait_for_mib_ref() + struct se_node_acl->mib_ref_count
	*) core_tpg_deregister() + struct se_portal_group->tpg_pr_ref_count
	*) transport_deregister_session() + struct se_session->mib_ref_count
Thanks to Konrad Rzeszutek Wilk for mentioning cpu_relax() usage!
Signed-off-by Nicholas A. Bellinger <[email protected]>
---
 drivers/target/target_core_device.c    |    6 +++---
 drivers/target/target_core_pr.c        |    2 +-
 drivers/target/target_core_tpg.c       |    6 +++---
 drivers/target/target_core_transport.c |    2 +-
 4 files changed, 8 insertions(+), 8 deletions(-)
drivers/target/target_core_device.c +0 -0
index 665b8e2..459f0f8 100644

drivers/target/target_core_device.c int core_update_device_list_for_node( · line 538

Quoted 8 lines

drivers/target/target_core_device.c +1 -1

@@ -532,7 +532,7 @@ int core_update_device_list_for_node(
 	 */
 	spin_unlock_irq(&nacl->device_list_lock);
 	while (atomic_read(&deve->pr_ref_count) != 0)
-		msleep(100);
+		cpu_relax();
 	spin_lock_irq(&nacl->device_list_lock);
 	/*
 	 * Disable struct se_dev_entry LUN ACL mapping

Why would you want to convert those to busy-waiting? This should rather be converted to wait_event/wake_up, right?

thanks,

Signature · 1 line

js

Nicholas A. Bellinger

author

2010-09-10 19:17 UTC #2 lore raw

Quoted 28 lines · Jiri Slaby

Quoted 27 lines · Nicholas A. Bellinger
From: Nicholas Bellinger <[email protected]>
Greetings all,
This patch converts the following functions from within shutdown and release path code to use cpu_relax() instead of msleep() while doing atomic_read() in a tight loop waiting for various PR and MIB reference count(s) to return to zero:
	*) core_update_device_list_for_node() + struct se_dev_entry->pr_ref_count
	*) core_release_port() + struct se_port->sep_tg_pt_ref_cnt
	*) se_dev_stop() + struct se_hba->dev_mib_access_count
	*) __core_scsi3_free_registration() + struct t10_pr_registration->pr_res_holders
	*) core_tpg_wait_for_nacl_pr_ref() + struct se_node_acl->acl_pr_ref_count
	*) core_tpg_wait_for_mib_ref() + struct se_node_acl->mib_ref_count
	*) core_tpg_deregister() + struct se_portal_group->tpg_pr_ref_count
	*) transport_deregister_session() + struct se_session->mib_ref_count
Thanks to Konrad Rzeszutek Wilk for mentioning cpu_relax() usage!
Signed-off-by Nicholas A. Bellinger <[email protected]>
---
 drivers/target/target_core_device.c    |    6 +++---
 drivers/target/target_core_pr.c        |    2 +-
 drivers/target/target_core_tpg.c       |    6 +++---
 drivers/target/target_core_transport.c |    2 +-
 4 files changed, 8 insertions(+), 8 deletions(-)

Quoted 16 lines

drivers/target/target_core_device.c int core_update_device_list_for_node( · line 538
Quoted 13 lines
drivers/target/target_core_device.c +1 -1
index 665b8e2..459f0f8 100644
@@ -532,7 +532,7 @@ int core_update_device_list_for_node(
 	 */
 	spin_unlock_irq(&nacl->device_list_lock);
 	while (atomic_read(&deve->pr_ref_count) != 0)
-		msleep(100);
+		cpu_relax();
 	spin_lock_irq(&nacl->device_list_lock);
 	/*
 	 * Disable struct se_dev_entry LUN ACL mapping
Why would you want to convert those to busy-waiting? This should rather be converted to wait_event/wake_up, right?

Hi Jiri,

I agree that this is the one scenario within the msleep() -> cpu_relax() conversion patch that involves something besides a simple shortly held 'non CDB I/O emulation path context ref/deref', and might probably be better served by using wait_event/wake_up.

I will take a look at a conversion to use wait_queue_head_t in the near future.

Thanks for your comments Jiri!

--nab