[PATCH 1/1] cciss: PCI power management reset for kexec

[Mike Miller]

Patch 1 of 1

This patch provides the better "kick-in-the-pants" on driver load in a
kexec'ed environment. 

I've successfully sanity tested the port in my lab. Randy, please apply and
test. You seem to be able to bring out the worst in the driver. ;-)

Author: Chip Coldwell <coldwell@xxxxxxxxxx>

    CCISS: Use PCI power management to reset the controller

    The kexec kernel resets the CCISS hardware in three steps:

    1. Use PCI power management states to reset the controller
       in the kexec kernel.
    2. Clear the MSI/MSI-X bits in PCI configuration space so
       that MSI initialization in the kexec kernel doesn't fail.
    3. Use the CCISS "No-op" message to determine when the
       controller firmware has recovered from the PCI PM reset.

Signed-off-by: Mike Miller <mike.miller@xxxxxx>

-------------------------------------------------------------------------------
diff --git a/drivers/block/cciss.c b/drivers/block/cciss.c
index 01e6938..ff4a105 100644
--- a/drivers/block/cciss.c
+++ b/drivers/block/cciss.c
@@ -3390,6 +3390,205 @@ static void free_hba(int i)
 	kfree(p);
 }
 
+/* Send a message CDB to the firmware. */
+static __devinit int cciss_message(struct pci_dev *pdev, unsigned char opcode, unsigned char type)
+{
+	typedef struct {
+		CommandListHeader_struct CommandHeader;
+		RequestBlock_struct Request;
+		ErrDescriptor_struct ErrorDescriptor;
+	} Command;
+	static const size_t cmd_sz = sizeof(Command) + sizeof(ErrorInfo_struct);
+	Command *cmd;
+	dma_addr_t paddr64;
+	uint32_t paddr32, tag;
+	void __iomem *vaddr;
+	int i, err;
+
+	vaddr = ioremap_nocache(pci_resource_start(pdev, 0), pci_resource_len(pdev, 0));
+	if (vaddr == NULL)
+		return -ENOMEM;
+
+	/* The Inbound Post Queue only accepts 32-bit physical addresses for the
+	   CCISS commands, so they must be allocated from the lower 4GiB of
+	   memory. */
+	err = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
+	if (err) {
+		iounmap(vaddr);
+		return -ENOMEM;
+	}
+
+	cmd = pci_alloc_consistent(pdev, cmd_sz, &paddr64);
+	if (cmd == NULL) {
+		iounmap(vaddr);
+		return -ENOMEM;
+	}
+
+	/* This must fit, because of the 32-bit consistent DMA mask.  Also,
+	   although there's no guarantee, we assume that the address is at
+	   least 4-byte aligned (most likely, it's page-aligned). */
+	paddr32 = paddr64;
+
+	cmd->CommandHeader.ReplyQueue = 0;
+	cmd->CommandHeader.SGList = 0;
+	cmd->CommandHeader.SGTotal = 0;
+	cmd->CommandHeader.Tag.lower = paddr32;
+	cmd->CommandHeader.Tag.upper = 0;
+	memset(&cmd->CommandHeader.LUN.LunAddrBytes, 0, 8);
+
+	cmd->Request.CDBLen = 16;
+	cmd->Request.Type.Type = TYPE_MSG;
+	cmd->Request.Type.Attribute = ATTR_HEADOFQUEUE;
+	cmd->Request.Type.Direction = XFER_NONE;
+	cmd->Request.Timeout = 0; /* Don't time out */
+	cmd->Request.CDB[0] = opcode;
+	cmd->Request.CDB[1] = type;
+	memset(&cmd->Request.CDB[2], 0, 14); /* the rest of the CDB is reserved */
+
+	cmd->ErrorDescriptor.Addr.lower = paddr32 + sizeof(Command);
+	cmd->ErrorDescriptor.Addr.upper = 0;
+	cmd->ErrorDescriptor.Len = sizeof(ErrorInfo_struct);
+
+	writel(paddr32, vaddr + SA5_REQUEST_PORT_OFFSET);
+
+	for (i = 0; i < 10; i++) {
+		tag = readl(vaddr + SA5_REPLY_PORT_OFFSET);
+		if ((tag & ~3) == paddr32)
+			break;
+		schedule_timeout_uninterruptible(HZ);
+	}
+
+	iounmap(vaddr);
+
+	/* we leak the DMA buffer here ... no choice since the controller could
+	   still complete the command. */
+	if (i == 10) {
+		printk(KERN_ERR "cciss: controller message %02x:%02x timed out\n",
+			opcode, type);
+		return -ETIMEDOUT;
+	}
+
+	pci_free_consistent(pdev, cmd_sz, cmd, paddr64);
+
+	if (tag & 2) {
+		printk(KERN_ERR "cciss: controller message %02x:%02x failed\n",
+			opcode, type);
+		return -EIO;
+	}
+
+	printk(KERN_INFO "cciss: controller message %02x:%02x succeeded\n",
+		opcode, type);
+	return 0;
+}
+
+#define cciss_soft_reset_controller(p) cciss_message(p, 1, 0)
+#define cciss_noop(p) cciss_message(p, 3, 0)
+
+static __devinit int cciss_reset_msi(struct pci_dev *pdev)
+{
+/* the #defines are stolen from drivers/pci/msi.h. */
+#define msi_control_reg(base)		(base + PCI_MSI_FLAGS)
+#define PCI_MSIX_FLAGS_ENABLE		(1 << 15)
+
+	int pos;
+	u16 control = 0;
+
+	pos = pci_find_capability(pdev, PCI_CAP_ID_MSI);
+	if (pos) {
+		pci_read_config_word(pdev, msi_control_reg(pos), &control);
+		if (control & PCI_MSI_FLAGS_ENABLE) {
+			printk(KERN_INFO "cciss: resetting MSI\n");
+			pci_write_config_word(pdev, msi_control_reg(pos), control & ~PCI_MSI_FLAGS_ENABLE);
+		}
+	}
+
+	pos = pci_find_capability(pdev, PCI_CAP_ID_MSIX);
+	if (pos) {
+		pci_read_config_word(pdev, msi_control_reg(pos), &control);
+		if (control & PCI_MSIX_FLAGS_ENABLE) {
+			printk(KERN_INFO "cciss: resetting MSI-X\n");
+			pci_write_config_word(pdev, msi_control_reg(pos), control & ~PCI_MSIX_FLAGS_ENABLE);
+		}
+	}
+
+	return 0;
+}
+
+/* This does a hard reset of the controller using PCI power management
+ * states. */
+static __devinit int cciss_hard_reset_controller(struct pci_dev *pdev)
+{
+	u16 pmcsr, saved_config_space[32];
+	int i, pos;
+
+	printk(KERN_INFO "cciss: using PCI PM to reset controller\n");
+
+	/* This is very nearly the same thing as
+
+	   pci_save_state(pci_dev);
+	   pci_set_power_state(pci_dev, PCI_D3hot);
+	   pci_set_power_state(pci_dev, PCI_D0);
+	   pci_restore_state(pci_dev);
+
+	   but we can't use these nice canned kernel routines on
+	   kexec, because they also check the MSI/MSI-X state in PCI
+	   configuration space and do the wrong thing when it is
+	   set/cleared.  Also, the pci_save/restore_state functions
+	   violate the ordering requirements for restoring the
+	   configuration space from the CCISS document (see the
+	   comment below).  So we roll our own .... */
+
+	for (i = 0; i < 32; i++)
+		pci_read_config_word(pdev, 2*i, &saved_config_space[i]);
+
+	pos = pci_find_capability(pdev, PCI_CAP_ID_PM);
+	if (pos == 0) {
+		printk(KERN_ERR "cciss_reset_controller: PCI PM not supported\n");
+		return -ENODEV;
+	}
+
+	/* Quoting from the Open CISS Specification: "The Power
+	 * Management Control/Status Register (CSR) controls the power
+	 * state of the device.  The normal operating state is D0,
+	 * CSR=00h.  The software off state is D3, CSR=03h.  To reset
+	 * the controller, place the interface device in D3 then to
+	 * D0, this causes a secondary PCI reset which will reset the
+	 * controller." */
+
+	/* enter the D3hot power management state */
+	pci_read_config_word(pdev, pos + PCI_PM_CTRL, &pmcsr);
+	pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
+	pmcsr |= PCI_D3hot;
+	pci_write_config_word(pdev, pos + PCI_PM_CTRL, pmcsr);
+
+	set_current_state(TASK_UNINTERRUPTIBLE);
+	schedule_timeout(HZ >> 1);
+
+	/* enter the D0 power management state */
+	pmcsr &= ~PCI_PM_CTRL_STATE_MASK;
+	pmcsr |= PCI_D0;
+	pci_write_config_word(pdev, pos + PCI_PM_CTRL, pmcsr);
+
+	set_current_state(TASK_UNINTERRUPTIBLE);
+	schedule_timeout(HZ >> 1);
+
+	/* Restore the PCI configuration space.  The Open CISS
+	 * Specification says, "Restore the PCI Configuration
+	 * Registers, offsets 00h through 60h. It is important to
+	 * restore the command register, 16-bits at offset 04h,
+	 * last. Do not restore the configuration status register,
+	 * 16-bits at offset 06h."  Note that the offset is 2*i. */
+	for (i = 0; i < 32; i++) {
+		if (i == 2 || i == 3)
+			continue;
+		pci_write_config_word(pdev, 2*i, saved_config_space[i]);
+	}
+	wmb();
+	pci_write_config_word(pdev, 4, saved_config_space[2]);
+
+	return 0;
+}
+
 /*
  *  This is it.  Find all the controllers and register them.  I really hate
  *  stealing all these major device numbers.
@@ -3404,6 +3603,24 @@ static int __devinit cciss_init_one(struct pci_dev *pdev,
 	int dac, return_code;
 	InquiryData_struct *inq_buff = NULL;
 
+	if (reset_devices) {
+		/* Reset the controller with a PCI power-cycle */
+		if (cciss_hard_reset_controller(pdev) || cciss_reset_msi(pdev))
+			return -ENODEV;
+
+		/* Some devices (notably the HP Smart Array 5i Controller)
+		   need a little pause here */
+		schedule_timeout_uninterruptible(30*HZ);
+
+		/* Now try to get the controller to respond to a no-op */
+		for (i=0; i<12; i++) {
+			if (cciss_noop(pdev) == 0)
+				break;
+			else
+				printk("cciss: no-op failed%s\n", (i < 11 ? "; re-trying" : ""));
+		}
+	}
+
 	i = alloc_cciss_hba();
 	if (i < 0)
 		return -1;
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/