cciss: do soft reset if hard reset is broken
authorStephen M. Cameron <scameron@beardog.cce.hp.com>
Tue, 3 May 2011 19:53:52 +0000 (14:53 -0500)
committerJens Axboe <jaxboe@fusionio.com>
Fri, 6 May 2011 14:23:56 +0000 (08:23 -0600)
on driver load, if reset_devices is set, and the hard reset
attempts fail, try to bring up the controller to the point that
a command can be sent, and send it a soft reset command, then
after the reset undo whatever driver initialization was done to get
it to the point to take a command, and re-do it after the reset.

This is to get kdump to work on all the "non-resettable" controllers
(except 64xx controllers which can't be reset due to the potentially
shared cache module.)

Signed-off-by: Stephen M. Cameron <scameron@beardog.cce.hp.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
drivers/block/cciss.c

index df58c59d9031d24bb029eb96d50aae79f5347b66..23b0ba49300a054defe0a4a913c6769a87b87683 100644 (file)
@@ -2477,6 +2477,31 @@ static int deregister_disk(ctlr_info_t *h, int drv_index,
        return 0;
 }
 
+static int __devinit cciss_send_reset(ctlr_info_t *h, unsigned char *scsi3addr,
+       u8 reset_type)
+{
+       CommandList_struct *c;
+       int return_status;
+
+       c = cmd_alloc(h);
+       if (!c)
+               return -ENOMEM;
+       return_status = fill_cmd(h, c, CCISS_RESET_MSG, NULL, 0, 0,
+               CTLR_LUNID, TYPE_MSG);
+       c->Request.CDB[1] = reset_type; /* fill_cmd defaults to target reset */
+       if (return_status != IO_OK) {
+               cmd_special_free(h, c);
+               return return_status;
+       }
+       c->waiting = NULL;
+       enqueue_cmd_and_start_io(h, c);
+       /* Don't wait for completion, the reset won't complete.  Don't free
+        * the command either.  This is the last command we will send before
+        * re-initializing everything, so it doesn't matter and won't leak.
+        */
+       return 0;
+}
+
 static int fill_cmd(ctlr_info_t *h, CommandList_struct *c, __u8 cmd, void *buff,
                size_t size, __u8 page_code, unsigned char *scsi3addr,
                int cmd_type)
@@ -3463,6 +3488,63 @@ static inline u32 process_nonindexed_cmd(ctlr_info_t *h, u32 raw_tag)
        return next_command(h);
 }
 
+/* Some controllers, like p400, will give us one interrupt
+ * after a soft reset, even if we turned interrupts off.
+ * Only need to check for this in the cciss_xxx_discard_completions
+ * functions.
+ */
+static int ignore_bogus_interrupt(ctlr_info_t *h)
+{
+       if (likely(!reset_devices))
+               return 0;
+
+       if (likely(h->interrupts_enabled))
+               return 0;
+
+       dev_info(&h->pdev->dev, "Received interrupt while interrupts disabled "
+               "(known firmware bug.)  Ignoring.\n");
+
+       return 1;
+}
+
+static irqreturn_t cciss_intx_discard_completions(int irq, void *dev_id)
+{
+       ctlr_info_t *h = dev_id;
+       unsigned long flags;
+       u32 raw_tag;
+
+       if (ignore_bogus_interrupt(h))
+               return IRQ_NONE;
+
+       if (interrupt_not_for_us(h))
+               return IRQ_NONE;
+       spin_lock_irqsave(&h->lock, flags);
+       while (interrupt_pending(h)) {
+               raw_tag = get_next_completion(h);
+               while (raw_tag != FIFO_EMPTY)
+                       raw_tag = next_command(h);
+       }
+       spin_unlock_irqrestore(&h->lock, flags);
+       return IRQ_HANDLED;
+}
+
+static irqreturn_t cciss_msix_discard_completions(int irq, void *dev_id)
+{
+       ctlr_info_t *h = dev_id;
+       unsigned long flags;
+       u32 raw_tag;
+
+       if (ignore_bogus_interrupt(h))
+               return IRQ_NONE;
+
+       spin_lock_irqsave(&h->lock, flags);
+       raw_tag = get_next_completion(h);
+       while (raw_tag != FIFO_EMPTY)
+               raw_tag = next_command(h);
+       spin_unlock_irqrestore(&h->lock, flags);
+       return IRQ_HANDLED;
+}
+
 static irqreturn_t do_cciss_intx(int irq, void *dev_id)
 {
        ctlr_info_t *h = dev_id;
@@ -4380,7 +4462,6 @@ static __devinit int cciss_message(struct pci_dev *pdev, unsigned char opcode, u
        return 0;
 }
 
-#define cciss_soft_reset_controller(p) cciss_message(p, 1, 0)
 #define cciss_noop(p) cciss_message(p, 3, 0)
 
 static int cciss_controller_hard_reset(struct pci_dev *pdev,
@@ -4591,13 +4672,17 @@ static __devinit int cciss_kdump_hard_reset_controller(struct pci_dev *pdev)
        /* Wait for board to become not ready, then ready. */
        dev_info(&pdev->dev, "Waiting for board to reset.\n");
        rc = cciss_wait_for_board_state(pdev, vaddr, BOARD_NOT_READY);
-       if (rc) /* Don't bail, might be E500, etc. which can't be reset */
-               dev_warn(&pdev->dev,
-                       "failed waiting for board to reset\n");
+       if (rc) {
+               dev_warn(&pdev->dev, "Failed waiting for board to hard reset."
+                               "  Will try soft reset.\n");
+               rc = -ENOTSUPP; /* Not expected, but try soft reset later */
+               goto unmap_cfgtable;
+       }
        rc = cciss_wait_for_board_state(pdev, vaddr, BOARD_READY);
        if (rc) {
                dev_warn(&pdev->dev,
-                       "failed waiting for board to become ready\n");
+                       "failed waiting for board to become ready "
+                       "after hard reset\n");
                goto unmap_cfgtable;
        }
 
@@ -4605,16 +4690,13 @@ static __devinit int cciss_kdump_hard_reset_controller(struct pci_dev *pdev)
        if (rc < 0)
                goto unmap_cfgtable;
        if (rc) {
-               dev_warn(&pdev->dev, "Unable to successfully reset controller,"
-                       " Ignoring controller.\n");
-               rc = -ENODEV;
-               goto unmap_cfgtable;
+               dev_warn(&pdev->dev, "Unable to successfully hard reset "
+                       "controller. Will try soft reset.\n");
+               rc = -ENOTSUPP; /* Not expected, but try soft reset later */
        } else {
-               dev_info(&pdev->dev, "board ready.\n");
+               dev_info(&pdev->dev, "Board ready after hard reset.\n");
        }
 
-       dev_info(&pdev->dev, "board ready.\n");
-
 unmap_cfgtable:
        iounmap(cfgtable);
 
@@ -4639,7 +4721,7 @@ static __devinit int cciss_init_reset_devices(struct pci_dev *pdev)
         * due to concerns about shared bbwc between 6402/6404 pair.
         */
        if (rc == -ENOTSUPP)
-               return 0; /* just try to do the kdump anyhow. */
+               return rc; /* just try to do the kdump anyhow. */
        if (rc)
                return -ENODEV;
 
@@ -4745,6 +4827,60 @@ static int cciss_request_irq(ctlr_info_t *h,
        return -1;
 }
 
+static int __devinit cciss_kdump_soft_reset(ctlr_info_t *h)
+{
+       if (cciss_send_reset(h, CTLR_LUNID, CCISS_RESET_TYPE_CONTROLLER)) {
+               dev_warn(&h->pdev->dev, "Resetting array controller failed.\n");
+               return -EIO;
+       }
+
+       dev_info(&h->pdev->dev, "Waiting for board to soft reset.\n");
+       if (cciss_wait_for_board_state(h->pdev, h->vaddr, BOARD_NOT_READY)) {
+               dev_warn(&h->pdev->dev, "Soft reset had no effect.\n");
+               return -1;
+       }
+
+       dev_info(&h->pdev->dev, "Board reset, awaiting READY status.\n");
+       if (cciss_wait_for_board_state(h->pdev, h->vaddr, BOARD_READY)) {
+               dev_warn(&h->pdev->dev, "Board failed to become ready "
+                       "after soft reset.\n");
+               return -1;
+       }
+
+       return 0;
+}
+
+static void cciss_undo_allocations_after_kdump_soft_reset(ctlr_info_t *h)
+{
+       int ctlr = h->ctlr;
+
+       free_irq(h->intr[PERF_MODE_INT], h);
+#ifdef CONFIG_PCI_MSI
+       if (h->msix_vector)
+               pci_disable_msix(h->pdev);
+       else if (h->msi_vector)
+               pci_disable_msi(h->pdev);
+#endif /* CONFIG_PCI_MSI */
+       cciss_free_sg_chain_blocks(h->cmd_sg_list, h->nr_cmds);
+       cciss_free_scatterlists(h);
+       cciss_free_cmd_pool(h);
+       kfree(h->blockFetchTable);
+       if (h->reply_pool)
+               pci_free_consistent(h->pdev, h->max_commands * sizeof(__u64),
+                               h->reply_pool, h->reply_pool_dhandle);
+       if (h->transtable)
+               iounmap(h->transtable);
+       if (h->cfgtable)
+               iounmap(h->cfgtable);
+       if (h->vaddr)
+               iounmap(h->vaddr);
+       unregister_blkdev(h->major, h->devname);
+       cciss_destroy_hba_sysfs_entry(h);
+       pci_release_regions(h->pdev);
+       kfree(h);
+       hba[ctlr] = NULL;
+}
+
 /*
  *  This is it.  Find all the controllers and register them.  I really hate
  *  stealing all these major device numbers.
@@ -4756,13 +4892,27 @@ static int __devinit cciss_init_one(struct pci_dev *pdev,
        int i;
        int j = 0;
        int rc;
+       int try_soft_reset = 0;
        int dac, return_code;
        InquiryData_struct *inq_buff;
        ctlr_info_t *h;
+       unsigned long flags;
 
        rc = cciss_init_reset_devices(pdev);
-       if (rc)
-               return rc;
+       if (rc) {
+               if (rc != -ENOTSUPP)
+                       return rc;
+               /* If the reset fails in a particular way (it has no way to do
+                * a proper hard reset, so returns -ENOTSUPP) we can try to do
+                * a soft reset once we get the controller configured up to the
+                * point that it can accept a command.
+                */
+               try_soft_reset = 1;
+               rc = 0;
+       }
+
+reinit_after_soft_reset:
+
        i = alloc_cciss_hba(pdev);
        if (i < 0)
                return -1;
@@ -4852,6 +5002,62 @@ static int __devinit cciss_init_one(struct pci_dev *pdev,
                h->gendisk[j] = NULL;
        }
 
+       /* At this point, the controller is ready to take commands.
+        * Now, if reset_devices and the hard reset didn't work, try
+        * the soft reset and see if that works.
+        */
+       if (try_soft_reset) {
+
+               /* This is kind of gross.  We may or may not get a completion
+                * from the soft reset command, and if we do, then the value
+                * from the fifo may or may not be valid.  So, we wait 10 secs
+                * after the reset throwing away any completions we get during
+                * that time.  Unregister the interrupt handler and register
+                * fake ones to scoop up any residual completions.
+                */
+               spin_lock_irqsave(&h->lock, flags);
+               h->access.set_intr_mask(h, CCISS_INTR_OFF);
+               spin_unlock_irqrestore(&h->lock, flags);
+               free_irq(h->intr[PERF_MODE_INT], h);
+               rc = cciss_request_irq(h, cciss_msix_discard_completions,
+                                       cciss_intx_discard_completions);
+               if (rc) {
+                       dev_warn(&h->pdev->dev, "Failed to request_irq after "
+                               "soft reset.\n");
+                       goto clean4;
+               }
+
+               rc = cciss_kdump_soft_reset(h);
+               if (rc) {
+                       dev_warn(&h->pdev->dev, "Soft reset failed.\n");
+                       goto clean4;
+               }
+
+               dev_info(&h->pdev->dev, "Board READY.\n");
+               dev_info(&h->pdev->dev,
+                       "Waiting for stale completions to drain.\n");
+               h->access.set_intr_mask(h, CCISS_INTR_ON);
+               msleep(10000);
+               h->access.set_intr_mask(h, CCISS_INTR_OFF);
+
+               rc = controller_reset_failed(h->cfgtable);
+               if (rc)
+                       dev_info(&h->pdev->dev,
+                               "Soft reset appears to have failed.\n");
+
+               /* since the controller's reset, we have to go back and re-init
+                * everything.  Easiest to just forget what we've done and do it
+                * all over again.
+                */
+               cciss_undo_allocations_after_kdump_soft_reset(h);
+               try_soft_reset = 0;
+               if (rc)
+                       /* don't go to clean4, we already unallocated */
+                       return -ENODEV;
+
+               goto reinit_after_soft_reset;
+       }
+
        cciss_scsi_setup(h);
 
        /* Turn the interrupts on so we can service requests */