[SCSI] libsas: Handle SCSI commands that complete with failure codes
authorDarrick J. Wong <djwong@us.ibm.com>
Fri, 26 Jan 2007 22:08:52 +0000 (14:08 -0800)
committerJames Bottomley <jejb@mulgrave.il.steeleye.com>
Sat, 27 Jan 2007 16:06:51 +0000 (10:06 -0600)
This patch moves the code that handles SAS failures out of the main EH
function and into a separate function.  It also detects commands that have
no sas_task (i.e. they completed, but with error data) and sends them into
scsi_error for processing.  This allows us to handle SCSI errors (and
enables auto-spinup as a side effect) instead of dropping them on the
floor and falling into an infinite loop.  It also requires the
implementation of a device reset function, which the SAS failure code has
been modified to employ for REQ_DEVICE_RESET.

Signed-off-by: Darrick J. Wong <djwong@us.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@SteelEye.com>
drivers/scsi/libsas/sas_scsi_host.c
include/scsi/libsas.h

index 9ffe7605fb9d6d2d075030ea35b33fbd65b75c07..eac1d2dde743daf23016d2a1ce372485a6725131 100644 (file)
@@ -34,6 +34,7 @@
 #include <scsi/scsi_transport_sas.h>
 #include "../scsi_sas_internal.h"
 #include "../scsi_transport_api.h"
+#include "../scsi_priv.h"
 
 #include <linux/err.h>
 #include <linux/blkdev.h>
@@ -396,54 +397,80 @@ static int sas_recover_I_T(struct domain_device *dev)
        return res;
 }
 
-static int eh_reset_phy_helper(struct sas_phy *phy)
+/* Find the sas_phy that's attached to this device */
+struct sas_phy *find_local_sas_phy(struct domain_device *dev)
 {
-       int tmf_resp;
+       struct domain_device *pdev = dev->parent;
+       struct ex_phy *exphy = NULL;
+       int i;
+
+       /* Directly attached device */
+       if (!pdev)
+               return dev->port->phy;
 
-       tmf_resp = sas_phy_reset(phy, 1);
-       if (tmf_resp)
-               SAS_DPRINTK("Hard reset of phy %d failed 0x%x\n",
-                           phy->identify.phy_identifier,
-                           tmf_resp);
+       /* Otherwise look in the expander */
+       for (i = 0; i < pdev->ex_dev.num_phys; i++)
+               if (!memcmp(dev->sas_addr,
+                           pdev->ex_dev.ex_phy[i].attached_sas_addr,
+                           SAS_ADDR_SIZE)) {
+                       exphy = &pdev->ex_dev.ex_phy[i];
+                       break;
+               }
 
-       return tmf_resp;
+       BUG_ON(!exphy);
+       return exphy->phy;
 }
 
-void sas_scsi_recover_host(struct Scsi_Host *shost)
+/* Attempt to send a target reset message to a device */
+int sas_eh_device_reset_handler(struct scsi_cmnd *cmd)
+{
+       struct domain_device *dev = cmd_to_domain_dev(cmd);
+       struct sas_phy *phy = find_local_sas_phy(dev);
+       int res;
+
+       res = sas_phy_reset(phy, 1);
+       if (res)
+               SAS_DPRINTK("Device reset of %s failed 0x%x\n",
+                           phy->dev.kobj.k_name,
+                           res);
+       if (res == TMF_RESP_FUNC_SUCC || res == TMF_RESP_FUNC_COMPLETE)
+               return SUCCESS;
+
+       return FAILED;
+}
+
+/* Try to reset a device */
+static int try_to_reset_cmd_device(struct Scsi_Host *shost,
+                                  struct scsi_cmnd *cmd)
+{
+       if (!shost->hostt->eh_device_reset_handler)
+               return FAILED;
+
+       return shost->hostt->eh_device_reset_handler(cmd);
+}
+
+static int sas_eh_handle_sas_errors(struct Scsi_Host *shost,
+                                   struct list_head *work_q,
+                                   struct list_head *done_q)
 {
-       struct sas_ha_struct *ha = SHOST_TO_SAS_HA(shost);
-       unsigned long flags;
-       LIST_HEAD(error_q);
        struct scsi_cmnd *cmd, *n;
        enum task_disposition res = TASK_IS_DONE;
        int tmf_resp, need_reset;
        struct sas_internal *i = to_sas_internal(shost->transportt);
-       struct sas_phy *task_sas_phy = NULL;
-
-       spin_lock_irqsave(shost->host_lock, flags);
-       list_splice_init(&shost->eh_cmd_q, &error_q);
-       spin_unlock_irqrestore(shost->host_lock, flags);
-
-       SAS_DPRINTK("Enter %s\n", __FUNCTION__);
+       unsigned long flags;
+       struct sas_ha_struct *ha = SHOST_TO_SAS_HA(shost);
 
-       /* All tasks on this list were marked SAS_TASK_STATE_ABORTED
-        * by sas_scsi_timed_out() callback.
-        */
 Again:
-       SAS_DPRINTK("going over list...\n");
-       list_for_each_entry_safe(cmd, n, &error_q, eh_entry) {
+       list_for_each_entry_safe(cmd, n, work_q, eh_entry) {
                struct sas_task *task = TO_SAS_TASK(cmd);
-               list_del_init(&cmd->eh_entry);
 
-               if (!task) {
-                       SAS_DPRINTK("%s: taskless cmd?!\n", __FUNCTION__);
+               if (!task)
                        continue;
-               }
+
+               list_del_init(&cmd->eh_entry);
 
                spin_lock_irqsave(&task->task_state_lock, flags);
                need_reset = task->task_state_flags & SAS_TASK_NEED_DEV_RESET;
-               if (need_reset)
-                       task_sas_phy = task->dev->port->phy;
                spin_unlock_irqrestore(&task->task_state_lock, flags);
 
                SAS_DPRINTK("trying to find task 0x%p\n", task);
@@ -457,14 +484,14 @@ Again:
                                    task);
                        task->task_done(task);
                        if (need_reset)
-                               eh_reset_phy_helper(task_sas_phy);
+                               try_to_reset_cmd_device(shost, cmd);
                        continue;
                case TASK_IS_ABORTED:
                        SAS_DPRINTK("%s: task 0x%p is aborted\n",
                                    __FUNCTION__, task);
                        task->task_done(task);
                        if (need_reset)
-                               eh_reset_phy_helper(task_sas_phy);
+                               try_to_reset_cmd_device(shost, cmd);
                        continue;
                case TASK_IS_AT_LU:
                        SAS_DPRINTK("task 0x%p is at LU: lu recover\n", task);
@@ -476,8 +503,8 @@ Again:
                                            cmd->device->lun);
                                task->task_done(task);
                                if (need_reset)
-                                       eh_reset_phy_helper(task_sas_phy);
-                               sas_scsi_clear_queue_lu(&error_q, cmd);
+                                       try_to_reset_cmd_device(shost, cmd);
+                               sas_scsi_clear_queue_lu(work_q, cmd);
                                goto Again;
                        }
                        /* fallthrough */
@@ -491,8 +518,8 @@ Again:
                                            SAS_ADDR(task->dev->sas_addr));
                                task->task_done(task);
                                if (need_reset)
-                                       eh_reset_phy_helper(task_sas_phy);
-                               sas_scsi_clear_queue_I_T(&error_q, task->dev);
+                                       try_to_reset_cmd_device(shost, cmd);
+                               sas_scsi_clear_queue_I_T(work_q, task->dev);
                                goto Again;
                        }
                        /* Hammer time :-) */
@@ -506,8 +533,8 @@ Again:
                                                    "succeeded\n", port->id);
                                        task->task_done(task);
                                        if (need_reset)
-                                               eh_reset_phy_helper(task_sas_phy);
-                                       sas_scsi_clear_queue_port(&error_q,
+                                               try_to_reset_cmd_device(shost, cmd);
+                                       sas_scsi_clear_queue_port(work_q,
                                                                  port);
                                        goto Again;
                                }
@@ -520,7 +547,7 @@ Again:
                                                    "succeeded\n");
                                        task->task_done(task);
                                        if (need_reset)
-                                               eh_reset_phy_helper(task_sas_phy);
+                                               try_to_reset_cmd_device(shost, cmd);
                                        goto out;
                                }
                        }
@@ -535,21 +562,53 @@ Again:
 
                        task->task_done(task);
                        if (need_reset)
-                               eh_reset_phy_helper(task_sas_phy);
+                               try_to_reset_cmd_device(shost, cmd);
                        goto clear_q;
                }
        }
 out:
-       scsi_eh_flush_done_q(&ha->eh_done_q);
-       SAS_DPRINTK("--- Exit %s\n", __FUNCTION__);
-       return;
+       return list_empty(work_q);
 clear_q:
        SAS_DPRINTK("--- Exit %s -- clear_q\n", __FUNCTION__);
-       list_for_each_entry_safe(cmd, n, &error_q, eh_entry) {
+       list_for_each_entry_safe(cmd, n, work_q, eh_entry) {
                struct sas_task *task = TO_SAS_TASK(cmd);
                list_del_init(&cmd->eh_entry);
                task->task_done(task);
        }
+       return list_empty(work_q);
+}
+
+void sas_scsi_recover_host(struct Scsi_Host *shost)
+{
+       struct sas_ha_struct *ha = SHOST_TO_SAS_HA(shost);
+       unsigned long flags;
+       LIST_HEAD(eh_work_q);
+
+       spin_lock_irqsave(shost->host_lock, flags);
+       list_splice_init(&shost->eh_cmd_q, &eh_work_q);
+       spin_unlock_irqrestore(shost->host_lock, flags);
+
+       SAS_DPRINTK("Enter %s\n", __FUNCTION__);
+       /*
+        * Deal with commands that still have SAS tasks (i.e. they didn't
+        * complete via the normal sas_task completion mechanism)
+        */
+       if (sas_eh_handle_sas_errors(shost, &eh_work_q, &ha->eh_done_q))
+               goto out;
+
+       /*
+        * Now deal with SCSI commands that completed ok but have a an error
+        * code (and hopefully sense data) attached.  This is roughly what
+        * scsi_unjam_host does, but we skip scsi_eh_abort_cmds because any
+        * command we see here has no sas_task and is thus unknown to the HA.
+        */
+       if (!scsi_eh_get_sense(&eh_work_q, &ha->eh_done_q))
+               scsi_eh_ready_devs(shost, &eh_work_q, &ha->eh_done_q);
+
+out:
+       scsi_eh_flush_done_q(&ha->eh_done_q);
+       SAS_DPRINTK("--- Exit %s\n", __FUNCTION__);
+       return;
 }
 
 enum scsi_eh_timer_return sas_scsi_timed_out(struct scsi_cmnd *cmd)
@@ -914,3 +973,4 @@ EXPORT_SYMBOL_GPL(__sas_task_abort);
 EXPORT_SYMBOL_GPL(sas_task_abort);
 EXPORT_SYMBOL_GPL(sas_phy_reset);
 EXPORT_SYMBOL_GPL(sas_phy_enable);
+EXPORT_SYMBOL_GPL(sas_eh_device_reset_handler);
index ca393929c10b1d2fe37ac649ecbaa4019e28db94..b200233cc6f45e8cdc91221a8a546b9413dd6932 100644 (file)
@@ -660,5 +660,6 @@ void sas_init_dev(struct domain_device *);
 
 void sas_task_abort(struct sas_task *);
 int __sas_task_abort(struct sas_task *);
+int sas_eh_device_reset_handler(struct scsi_cmnd *cmd);
 
 #endif /* _SASLIB_H_ */