[ofa-general][PATCH 4/4] SRP fail-over faster

public inbox for linux-rdma@vger.kernel.org
 help / color / mirror / Atom feed

* [ofa-general][PATCH 4/4] SRP fail-over faster
@ 2009-10-12 22:57 Vu Pham
       [not found] ` <4AD3B466.8060908-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
  0 siblings, 1 reply; 4+ messages in thread
From: Vu Pham @ 2009-10-12 22:57 UTC (permalink / raw)
  To: Linux RDMA list

[-- Attachment #1: Type: text/plain, Size: 1 bytes --]



[-- Attachment #2: [ofa-general][PATCH 4/4] SRP fail-over faster,.eml --]
[-- Type: message/rfc822, Size: 4725 bytes --]

[-- Attachment #2.1.1: Type: text/plain, Size: 283 bytes --]

Handling async local port events: port err, active, lid change...
Upon local port err, it will set up timer srp_dev_loss_tmo seconds to 
reconnect. If local port active and there is timer, it will delete the timer

Signed-off-by: Vu Pham <vu-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>

[-- Attachment #2.1.2: srp_4_async_event_handler.patch --]
[-- Type: text/plain, Size: 3489 bytes --]

Index: ofed_kernel/drivers/infiniband/ulp/srp/ib_srp.c
===================================================================
--- ofed_kernel.orig/drivers/infiniband/ulp/srp/ib_srp.c
+++ ofed_kernel/drivers/infiniband/ulp/srp/ib_srp.c
@@ -2071,6 +2120,77 @@
 	return NULL;
 }
 
+static void srp_event_handler(struct ib_event_handler *handler,
+			      struct ib_event *event)
+{
+	struct srp_device *srp_dev =
+	    ib_get_client_data(event->device, &srp_client);
+	struct srp_host *host, *tmp_host;
+	struct srp_target_port *target, *tmp_target;
+
+	if (!srp_dev || srp_dev->dev != event->device)
+		return;
+
+	printk(KERN_WARNING PFX "ASYNC event= %d on device= %s\n",
+		event->event, srp_dev->dev->name);
+
+	switch (event->event) {
+	case IB_EVENT_PORT_ERR:
+		list_for_each_entry_safe(host, tmp_host,
+					 &srp_dev->dev_list, list) {
+			if (event->element.port_num == host->port) {
+				spin_lock(&host->target_lock);
+				list_for_each_entry_safe(target, tmp_target,
+							 &host->target_list, list) {
+					unsigned long flags;
+
+					spin_lock_irqsave(target->scsi_host->host_lock,
+							  flags);
+					if (!target->qp_in_error &&
+					    target->state == SRP_TARGET_LIVE)
+						srp_qp_err_add_timer(target,
+								     srp_dev_loss_tmo);
+					spin_unlock_irqrestore(target->scsi_host->host_lock,
+							       flags);
+				}
+				spin_unlock(&host->target_lock);
+			}
+		}
+		break;
+	case IB_EVENT_PORT_ACTIVE:
+	case IB_EVENT_LID_CHANGE:
+	case IB_EVENT_PKEY_CHANGE:
+	case IB_EVENT_SM_CHANGE:
+		list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list,
+					 list) {
+			if (event->element.port_num == host->port) {
+				spin_lock(&host->target_lock);
+				list_for_each_entry_safe(target, tmp_target,
+							 &host->target_list, list) {
+					unsigned long flags;
+
+					spin_lock_irqsave(target->scsi_host->host_lock,
+							  flags);
+					if (timer_pending(&target->qp_err_timer)
+					    && !target->qp_in_error) {
+						shost_printk(KERN_WARNING PFX,
+							     target->scsi_host,
+							     "delete qp_in_err timer\n");
+						del_timer(&target->qp_err_timer);
+					}
+					spin_unlock_irqrestore(target->scsi_host->host_lock,
+							       flags);
+				}
+				spin_unlock(&host->target_lock);
+			}
+		}
+		break;
+	default:
+		break;
+	}
+
+}
+
 static void srp_add_one(struct ib_device *device)
 {
 	struct srp_device *srp_dev;
@@ -2116,6 +2228,11 @@
 	if (IS_ERR(srp_dev->mr))
 		goto err_pd;
 
+	INIT_IB_EVENT_HANDLER(&srp_dev->event_handler, srp_dev->dev,
+			      srp_event_handler);
+	if (ib_register_event_handler(&srp_dev->event_handler))
+		goto err_pd;
+
 	memset(&fmr_param, 0, sizeof fmr_param);
 	fmr_param.pool_size	    = SRP_FMR_POOL_SIZE;
 	fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
@@ -2160,6 +2284,8 @@
 
 	srp_dev = ib_get_client_data(device, &srp_client);
 
+	ib_unregister_event_handler(&srp_dev->event_handler);
+
 	list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
 		device_unregister(&host->dev);
 		/*
Index: ofed_kernel/drivers/infiniband/ulp/srp/ib_srp.h
===================================================================
--- ofed_kernel.orig/drivers/infiniband/ulp/srp/ib_srp.h
+++ ofed_kernel/drivers/infiniband/ulp/srp/ib_srp.h
@@ -88,6 +88,7 @@ struct srp_device {
 	struct ib_device       *dev;
 	struct ib_pd	       *pd;
 	struct ib_mr	       *mr;
+	struct ib_event_handler event_handler;
 	struct ib_fmr_pool     *fmr_pool;
 	int			fmr_page_shift;
 	int			fmr_page_size;

^ permalink raw reply	[flat|nested] 4+ messages in thread

[parent not found: <4AD3B466.8060908-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>]

* Re: [ofa-general][PATCH 4/4] SRP fail-over faster
       [not found] ` <4AD3B466.8060908-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
@ 2009-10-13 11:09   ` Bart Van Assche
       [not found]     ` <e2e108260910130409q78e3edbcndd64b7cf419705b9-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
  0 siblings, 1 reply; 4+ messages in thread
From: Bart Van Assche @ 2009-10-13 11:09 UTC (permalink / raw)
  To: Vu Pham; +Cc: Linux RDMA list

On Tue, Oct 13, 2009 at 12:57 AM, Vu Pham <vuhuong@mellanox.com> wrote:
>
>
> Handling async local port events: port err, active, lid change...
> Upon local port err, it will set up timer srp_dev_loss_tmo seconds to reconnect. If local port active and there is timer, it will delete the timer
>
> Signed-off-by: Vu Pham <vu@mellanox.com>
>
> Index: ofed_kernel/drivers/infiniband/ulp/srp/ib_srp.c
> ===================================================================
> --- ofed_kernel.orig/drivers/infiniband/ulp/srp/ib_srp.c
> +++ ofed_kernel/drivers/infiniband/ulp/srp/ib_srp.c
> @@ -2071,6 +2120,77 @@
>        return NULL;
>  }
>
> +static void srp_event_handler(struct ib_event_handler *handler,
> +                             struct ib_event *event)
> +{
> +       struct srp_device *srp_dev =
> +           ib_get_client_data(event->device, &srp_client);
> +       struct srp_host *host, *tmp_host;
> +       struct srp_target_port *target, *tmp_target;
> +
> +       if (!srp_dev || srp_dev->dev != event->device)
> +               return;
> +
> +       printk(KERN_WARNING PFX "ASYNC event= %d on device= %s\n",
> +               event->event, srp_dev->dev->name);
> +
> +       switch (event->event) {
> +       case IB_EVENT_PORT_ERR:
> +               list_for_each_entry_safe(host, tmp_host,
> +                                        &srp_dev->dev_list, list) {
> +                       if (event->element.port_num == host->port) {
> +                               spin_lock(&host->target_lock);

Can srp_remove_work() be executed concurrently with
srp_event_handler() ? In that case the above code isn't safe and the
spin_lock(&host->target_lock) should be moved to just before
list_for_each_entry_safe(). The current implementation can trigger
reading deallocated memory.

> +                               list_for_each_entry_safe(target, tmp_target,
> +                                                        &host->target_list, list) {
> +                                       unsigned long flags;
> +
> +                                       spin_lock_irqsave(target->scsi_host->host_lock,
> +                                                         flags);
> +                                       if (!target->qp_in_error &&
> +                                           target->state == SRP_TARGET_LIVE)
> +                                               srp_qp_err_add_timer(target,
> +                                                                    srp_dev_loss_tmo);
> +                                       spin_unlock_irqrestore(target->scsi_host->host_lock,
> +                                                              flags);
> +                               }
> +                               spin_unlock(&host->target_lock);
> +                       }
> +               }
> +               break;
> +       case IB_EVENT_PORT_ACTIVE:
> +       case IB_EVENT_LID_CHANGE:
> +       case IB_EVENT_PKEY_CHANGE:
> +       case IB_EVENT_SM_CHANGE:
> +               list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list,
> +                                        list) {
> +                       if (event->element.port_num == host->port) {
> +                               spin_lock(&host->target_lock);

A remark similar to the above: this code isn't safe against concurrent
calls of srp_remove_one().

> +                               list_for_each_entry_safe(target, tmp_target,
> +                                                        &host->target_list, list) {
> +                                       unsigned long flags;
> +
> +                                       spin_lock_irqsave(target->scsi_host->host_lock,
> +                                                         flags);
> +                                       if (timer_pending(&target->qp_err_timer)
> +                                           && !target->qp_in_error) {
> +                                               shost_printk(KERN_WARNING PFX,
> +                                                            target->scsi_host,
> +                                                            "delete qp_in_err timer\n");
> +                                               del_timer(&target->qp_err_timer);
> +                                       }
> +                                       spin_unlock_irqrestore(target->scsi_host->host_lock,
> +                                                              flags);
> +                               }
> +                               spin_unlock(&host->target_lock);
> +                       }
> +               }
> +               break;
> +       default:
> +               break;
> +       }
> +
> +}
> +
>  static void srp_add_one(struct ib_device *device)
>  {
>        struct srp_device *srp_dev;
> @@ -2116,6 +2228,11 @@
>        if (IS_ERR(srp_dev->mr))
>                goto err_pd;
>
> +       INIT_IB_EVENT_HANDLER(&srp_dev->event_handler, srp_dev->dev,
> +                             srp_event_handler);
> +       if (ib_register_event_handler(&srp_dev->event_handler))
> +               goto err_pd;
> +
>        memset(&fmr_param, 0, sizeof fmr_param);
>        fmr_param.pool_size         = SRP_FMR_POOL_SIZE;
>        fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
> @@ -2160,6 +2284,8 @@
>
>        srp_dev = ib_get_client_data(device, &srp_client);
>
> +       ib_unregister_event_handler(&srp_dev->event_handler);
> +
>        list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
>                device_unregister(&host->dev);
>                /*
> Index: ofed_kernel/drivers/infiniband/ulp/srp/ib_srp.h
> ===================================================================
> --- ofed_kernel.orig/drivers/infiniband/ulp/srp/ib_srp.h
> +++ ofed_kernel/drivers/infiniband/ulp/srp/ib_srp.h
> @@ -88,6 +88,7 @@ struct srp_device {
>        struct ib_device       *dev;
>        struct ib_pd           *pd;
>        struct ib_mr           *mr;
> +       struct ib_event_handler event_handler;
>        struct ib_fmr_pool     *fmr_pool;
>        int                     fmr_page_shift;
>        int                     fmr_page_size;
>
N‹§²æìr¸›yúèšØb²X¬¶Ç§vØ^–)Þº{.nÇ+‰·¥Š{±ÙšŠ{ayº\x1dÊ‡Ú™ë,j\a¢f£¢·hš‹»öì\x17/oSc¾™Ú³9˜uÀ¦æå‰È&jw¨®\x03(éšŽŠÝ¢j"ú\x1a¶^[m§ÿïêäz¹Þ–Šàþf£¢·hšˆ§~ˆmš

^ permalink raw reply	[flat|nested] 4+ messages in thread

[parent not found: <e2e108260910130409q78e3edbcndd64b7cf419705b9-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>]

* Re: [ofa-general][PATCH 4/4] SRP fail-over faster
       [not found]     ` <e2e108260910130409q78e3edbcndd64b7cf419705b9-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
@ 2009-10-14 21:11       ` Vu Pham
  2009-10-22 23:17       ` Vu Pham
  1 sibling, 0 replies; 4+ messages in thread
From: Vu Pham @ 2009-10-14 21:11 UTC (permalink / raw)
  To: Bart Van Assche; +Cc: Linux RDMA list

Bart Van Assche wrote:
>
>> +
>> +       switch (event->event) {
>> +       case IB_EVENT_PORT_ERR:
>> +               list_for_each_entry_safe(host, tmp_host,
>> +                                        &srp_dev->dev_list, list) {
>> +                       if (event->element.port_num == host->port) {
>> +                               spin_lock(&host->target_lock);
>>     
>
> Can srp_remove_work() be executed concurrently with
> srp_event_handler() ? In that case the above code isn't safe and the
> spin_lock(&host->target_lock) should be moved to just before
> list_for_each_entry_safe(). The current implementation can trigger
> reading deallocated memory.
>
>   
>> +                               list_for_each_entry_safe(target, tmp_target,
>> +                                                        &host->target_list, list) {
>> +                                       unsigned long flags;
>> +
>> +                                       spin_lock_irqsave(target->scsi_host->host_lock,
>> +                                                         flags);
>> +                                       if (!target->qp_in_error &&
>> +                                           target->state == SRP_TARGET_LIVE)
>> +                                               srp_qp_err_add_timer(target,
>> +                                                                    srp_dev_loss_tmo);
>> +                                       spin_unlock_irqrestore(target->scsi_host->host_lock,
>> +                                                              flags);
>> +                               }
>> +                               spin_unlock(&host->target_lock);
>> +                       }
>> +               }
>> +               break;
>> +       case IB_EVENT_PORT_ACTIVE:
>> +       case IB_EVENT_LID_CHANGE:
>> +       case IB_EVENT_PKEY_CHANGE:
>> +       case IB_EVENT_SM_CHANGE:
>> +               list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list,
>> +                                        list) {
>> +                       if (event->element.port_num == host->port) {
>> +                               spin_lock(&host->target_lock);
>>     
>
> A remark similar to the above: this code isn't safe against concurrent
> calls of srp_remove_one().
>
>   
I did not pay attention to the races created by unloading module (same 
for patch 3/4). I'll rework patches and address these races.

thanks,
-vu
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [ofa-general][PATCH 4/4] SRP fail-over faster
       [not found]     ` <e2e108260910130409q78e3edbcndd64b7cf419705b9-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
  2009-10-14 21:11       ` Vu Pham
@ 2009-10-22 23:17       ` Vu Pham
  1 sibling, 0 replies; 4+ messages in thread
From: Vu Pham @ 2009-10-22 23:17 UTC (permalink / raw)
  To: Bart Van Assche; +Cc: Linux RDMA list

[-- Attachment #1: Type: text/plain, Size: 1278 bytes --]

Bart Van Assche wrote:
>
>>
>> +static void srp_event_handler(struct ib_event_handler *handler,
>> +                             struct ib_event *event)
>> +{
>> +       struct srp_device *srp_dev =
>> +           ib_get_client_data(event->device, &srp_client);
>> +       struct srp_host *host, *tmp_host;
>> +       struct srp_target_port *target, *tmp_target;
>> +
>> +       if (!srp_dev || srp_dev->dev != event->device)
>> +               return;
>> +
>> +       printk(KERN_WARNING PFX "ASYNC event= %d on device= %s\n",
>> +               event->event, srp_dev->dev->name);
>> +
>> +       switch (event->event) {
>> +       case IB_EVENT_PORT_ERR:
>> +               list_for_each_entry_safe(host, tmp_host,
>> +                                        &srp_dev->dev_list, list) {
>> +                       if (event->element.port_num == host->port) {
>> +                               spin_lock(&host->target_lock);
>>     
>
> Can srp_remove_work() be executed concurrently with
> srp_event_handler() ? In that case the above code isn't safe and the
> spin_lock(&host->target_lock) should be moved to just before
> list_for_each_entry_safe(). The current implementation can trigger
> reading deallocated memory.
>
>   
Here is the updated patch to address the race


[-- Attachment #2: srp_4_async_event_handler.patch --]
[-- Type: text/plain, Size: 4754 bytes --]

Handling async local port events: port error, active, lid change...
Upon local port err, it will set up timer device_loss_timeout seconds to
reconnect. If local port active and there is timer, it will delete the timer

Signed-off-by: Vu Pham <vu-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org> 


 drivers/infiniband/ulp/srp/ib_srp.c |   85 +++++++++++++++++++++++++++++++++++
 drivers/infiniband/ulp/srp/ib_srp.h |    2 +
 2 files changed, 87 insertions(+), 0 deletions(-)

diff --git a/drivers/infiniband/ulp/srp/ib_srp.c b/drivers/infiniband/ulp/srp/ib_srp.c
index 12404d5..88dbc17 100644
--- a/drivers/infiniband/ulp/srp/ib_srp.c
+++ b/drivers/infiniband/ulp/srp/ib_srp.c
@@ -2058,6 +2058,81 @@ free_host:
 	return NULL;
 }
 
+static void srp_event_handler(struct ib_event_handler *handler,
+			      struct ib_event *event)
+{
+	struct srp_device *srp_dev =
+	    ib_get_client_data(event->device, &srp_client);
+	struct srp_host *host, *tmp_host;
+	struct srp_target_port *target, *tmp_target;
+
+	if (!srp_dev || srp_dev->dev != event->device)
+		return;
+
+	printk(KERN_WARNING PFX "ASYNC event= %d on device= %s\n",
+		event->event, srp_dev->dev->name);
+
+	switch (event->event) {
+	case IB_EVENT_PORT_ERR:
+		spin_lock(&srp_dev->dev_lock);
+		list_for_each_entry_safe(host, tmp_host,
+					 &srp_dev->dev_list, list) {
+			if (event->element.port_num == host->port) {
+				spin_lock(&host->target_lock);
+				list_for_each_entry_safe(target, tmp_target,
+							 &host->target_list, list) {
+					unsigned long flags;
+
+					spin_lock_irqsave(target->scsi_host->host_lock,
+							  flags);
+					if (!target->qp_in_error &&
+					    target->state == SRP_TARGET_LIVE)
+						srp_qp_err_add_timer(target,
+								     target->device_loss_timeout);
+					spin_unlock_irqrestore(target->scsi_host->host_lock,
+							       flags);
+				}
+				spin_unlock(&host->target_lock);
+			}
+		}
+		spin_unlock(&srp_dev->dev_lock);
+		break;
+	case IB_EVENT_PORT_ACTIVE:
+	case IB_EVENT_LID_CHANGE:
+	case IB_EVENT_PKEY_CHANGE:
+	case IB_EVENT_SM_CHANGE:
+		spin_lock(&srp_dev->dev_lock);
+		list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list,
+					 list) {
+			if (event->element.port_num == host->port) {
+				spin_lock(&host->target_lock);
+				list_for_each_entry_safe(target, tmp_target,
+							 &host->target_list, list) {
+					unsigned long flags;
+
+					spin_lock_irqsave(target->scsi_host->host_lock,
+							  flags);
+					if (timer_pending(&target->qp_err_timer)
+					    && !target->qp_in_error) {
+						shost_printk(KERN_WARNING PFX,
+							     target->scsi_host,
+							     "delete qp_in_err timer\n");
+						del_timer(&target->qp_err_timer);
+					}
+					spin_unlock_irqrestore(target->scsi_host->host_lock,
+							       flags);
+				}
+				spin_unlock(&host->target_lock);
+			}
+		}
+		spin_unlock(&srp_dev->dev_lock);
+		break;
+	default:
+		break;
+	}
+
+}
+
 static void srp_add_one(struct ib_device *device)
 {
 	struct srp_device *srp_dev;
@@ -2090,6 +2165,7 @@ static void srp_add_one(struct ib_device *device)
 	srp_dev->fmr_page_mask  = ~((u64) srp_dev->fmr_page_size - 1);
 
 	INIT_LIST_HEAD(&srp_dev->dev_list);
+	spin_lock_init(&srp_dev->dev_lock);
 
 	srp_dev->dev = device;
 	srp_dev->pd  = ib_alloc_pd(device);
@@ -2103,6 +2179,11 @@ static void srp_add_one(struct ib_device *device)
 	if (IS_ERR(srp_dev->mr))
 		goto err_pd;
 
+	INIT_IB_EVENT_HANDLER(&srp_dev->event_handler, srp_dev->dev,
+			      srp_event_handler);
+	if (ib_register_event_handler(&srp_dev->event_handler))
+		goto err_pd;
+
 	memset(&fmr_param, 0, sizeof fmr_param);
 	fmr_param.pool_size	    = SRP_FMR_POOL_SIZE;
 	fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
@@ -2154,6 +2235,9 @@ static void srp_remove_one(struct ib_device *device)
 
 	srp_dev = ib_get_client_data(device, &srp_client);
 
+	ib_unregister_event_handler(&srp_dev->event_handler);
+
+	spin_lock(&srp_dev->dev_lock);
 	list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
 		device_unregister(&host->dev);
 		/*
@@ -2193,6 +2277,7 @@ static void srp_remove_one(struct ib_device *device)
 
 		kfree(host);
 	}
+	spin_unlock(&srp_dev->dev_lock);
 
 	if (srp_dev->fmr_pool)
 		ib_destroy_fmr_pool(srp_dev->fmr_pool);
diff --git a/drivers/infiniband/ulp/srp/ib_srp.h b/drivers/infiniband/ulp/srp/ib_srp.h
index daa4bf7..74d1f09 100644
--- a/drivers/infiniband/ulp/srp/ib_srp.h
+++ b/drivers/infiniband/ulp/srp/ib_srp.h
@@ -86,8 +86,10 @@ enum srp_request_type {
 struct srp_device {
 	struct list_head	dev_list;
 	struct ib_device       *dev;
+	spinlock_t		dev_lock;
 	struct ib_pd	       *pd;
 	struct ib_mr	       *mr;
+	struct ib_event_handler event_handler;
 	struct ib_fmr_pool     *fmr_pool;
 	int			fmr_page_shift;
 	int			fmr_page_size;

^ permalink raw reply related	[flat|nested] 4+ messages in thread

end of thread, other threads:[~2009-10-22 23:17 UTC | newest]

Thread overview: 4+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2009-10-12 22:57 [ofa-general][PATCH 4/4] SRP fail-over faster Vu Pham
     [not found] ` <4AD3B466.8060908-VPRAkNaXOzVWk0Htik3J/w@public.gmane.org>
2009-10-13 11:09   ` Bart Van Assche
     [not found]     ` <e2e108260910130409q78e3edbcndd64b7cf419705b9-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org>
2009-10-14 21:11       ` Vu Pham
2009-10-22 23:17       ` Vu Pham

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox