cciss: add performant mode support for Stars/Sirius

Add a mode of controller operation called Performant Mode.  Even though
cciss has been deprecated in favor of hpsa there are new controllers due
out next year that HP must support in older vendor distros.  Vendors
require all fixes/features be upstream.  These new controllers support
only 16 commands in simple mode but support up to 1024 in performant mode.
This requires us to add this support at this late date.

The performant mode transport minimizes host PCI accesses by performinf
many completions per read.  PCI writes are posted so the host can write
then immediately get off the bus not waiting for the writwe to complete to
the target.  In the context of performant mode the host read out to a
controller pulls all posted writes into host memory ensuring the reply
queue is coherent.
Signed-off-by: Mike Miller <mike.miller@hp.com>
Cc: Stephen M. Cameron <scameron@beardog.cce.hp.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>

drivers/block/cciss.h

@@ -85,8 +85,8 @@ struct ctlr_info
 	int	max_cmd_sgentries;
 	SGDescriptor_struct **cmd_sg_list;
 
-#	define DOORBELL_INT	0
-#	define PERF_MODE_INT	1
+#	define PERF_MODE_INT	0
+#	define DOORBELL_INT	1
 #	define SIMPLE_MODE_INT	2
 #	define MEMQ_MODE_INT	3
 	unsigned int intr[4];
@@ -137,10 +137,27 @@ struct ctlr_info
 	struct list_head scan_list;
 	struct completion scan_wait;
 	struct device dev;
+	/*
+	 * Performant mode tables.
+	 */
+	u32 trans_support;
+	u32 trans_offset;
+	struct TransTable_struct *transtable;
+	unsigned long transMethod;
+
+	/*
+	 * Performant mode completion buffer
+	 */
+	u64 *reply_pool;
+	dma_addr_t reply_pool_dhandle;
+	u64 *reply_pool_head;
+	size_t reply_pool_size;
+	unsigned char reply_pool_wraparound;
+	u32 *blockFetchTable;
 };
 
-/*  Defining the diffent access_menthods */
-/*
+/*  Defining the diffent access_methods
+ *
  * Memory mapped FIFO interface (SMART 53xx cards)
  */
 #define SA5_DOORBELL	0x20
@@ -159,6 +176,15 @@ struct ctlr_info
 #define SA5B_INTR_PENDING	0x04
 #define FIFO_EMPTY		0xffffffff	
 #define CCISS_FIRMWARE_READY	0xffff0000 /* value in scratchpad register */
+/* Perf. mode flags */
+#define SA5_PERF_INTR_PENDING	0x04
+#define SA5_PERF_INTR_OFF	0x05
+#define SA5_OUTDB_STATUS_PERF_BIT	0x01
+#define SA5_OUTDB_CLEAR_PERF_BIT	0x01
+#define SA5_OUTDB_CLEAR         0xA0
+#define SA5_OUTDB_CLEAR_PERF_BIT        0x01
+#define SA5_OUTDB_STATUS        0x9C
+
 
 #define  CISS_ERROR_BIT		0x02
 
@@ -170,8 +196,9 @@ struct ctlr_info
 static void SA5_submit_command( ctlr_info_t *h, CommandList_struct *c) 
 {
 #ifdef CCISS_DEBUG
-	 printk("Sending %x - down to controller\n", c->busaddr );
-#endif /* CCISS_DEBUG */ 
+	printk(KERN_WARNING "cciss%d: Sending %08x - down to controller\n",
+			h->ctlr, c->busaddr);
+#endif /* CCISS_DEBUG */
          writel(c->busaddr, h->vaddr + SA5_REQUEST_PORT_OFFSET);
 	 h->commands_outstanding++;
 	 if ( h->commands_outstanding > h->max_outstanding)
@@ -214,6 +241,20 @@ static void SA5B_intr_mask(ctlr_info_t *h, unsigned long val)
                         h->vaddr + SA5_REPLY_INTR_MASK_OFFSET);
         }
 }
+
+/* Performant mode intr_mask */
+static void SA5_performant_intr_mask(ctlr_info_t *h, unsigned long val)
+{
+	if (val) { /* turn on interrupts */
+		h->interrupts_enabled = 1;
+		writel(0, h->vaddr + SA5_REPLY_INTR_MASK_OFFSET);
+	} else {
+		h->interrupts_enabled = 0;
+		writel(SA5_PERF_INTR_OFF,
+				h->vaddr + SA5_REPLY_INTR_MASK_OFFSET);
+	}
+}
+
 /*
  *  Returns true if fifo is full.  
  * 
@@ -250,6 +291,40 @@ static unsigned long SA5_completed(ctlr_info_t *h)
 	return ( register_value); 
 
 }
+
+/* Performant mode command completed */
+static unsigned long SA5_performant_completed(ctlr_info_t *h)
+{
+	unsigned long register_value = FIFO_EMPTY;
+
+	/* flush the controller write of the reply queue by reading
+	 * outbound doorbell status register.
+	 */
+	register_value = readl(h->vaddr + SA5_OUTDB_STATUS);
+	/* msi auto clears the interrupt pending bit. */
+	if (!(h->msi_vector || h->msix_vector)) {
+		writel(SA5_OUTDB_CLEAR_PERF_BIT, h->vaddr + SA5_OUTDB_CLEAR);
+		/* Do a read in order to flush the write to the controller
+		 * (as per spec.)
+		 */
+		register_value = readl(h->vaddr + SA5_OUTDB_STATUS);
+	}
+
+	if ((*(h->reply_pool_head) & 1) == (h->reply_pool_wraparound)) {
+		register_value = *(h->reply_pool_head);
+		(h->reply_pool_head)++;
+		h->commands_outstanding--;
+	} else {
+		register_value = FIFO_EMPTY;
+	}
+	/* Check for wraparound */
+	if (h->reply_pool_head == (h->reply_pool + h->max_commands)) {
+		h->reply_pool_head = h->reply_pool;
+		h->reply_pool_wraparound ^= 1;
+	}
+
+	return register_value;
+}
 /*
  *	Returns true if an interrupt is pending.. 
  */
@@ -280,6 +355,20 @@ static bool SA5B_intr_pending(ctlr_info_t *h)
         return 0 ;
 }
 
+static bool SA5_performant_intr_pending(ctlr_info_t *h)
+{
+	unsigned long register_value = readl(h->vaddr + SA5_INTR_STATUS);
+
+	if (!register_value)
+		return false;
+
+	if (h->msi_vector || h->msix_vector)
+		return true;
+
+	/* Read outbound doorbell to flush */
+	register_value = readl(h->vaddr + SA5_OUTDB_STATUS);
+	return register_value & SA5_OUTDB_STATUS_PERF_BIT;
+}
 
 static struct access_method SA5_access = {
 	SA5_submit_command,
@@ -297,6 +386,14 @@ static struct access_method SA5B_access = {
         SA5_completed,
 };
 
+static struct access_method SA5_performant_access = {
+	SA5_submit_command,
+	SA5_performant_intr_mask,
+	SA5_fifo_full,
+	SA5_performant_intr_pending,
+	SA5_performant_completed,
+};
+
 struct board_type {
 	__u32	board_id;
 	char	*product_name;

drivers/block/cciss_cmd.h

@@ -54,6 +54,7 @@
 #define CFGTBL_AccCmds          0x00000001l
 
 #define CFGTBL_Trans_Simple     0x00000002l
+#define CFGTBL_Trans_Performant 0x00000004l
 
 #define CFGTBL_BusType_Ultra2   0x00000001l
 #define CFGTBL_BusType_Ultra3   0x00000002l
@@ -173,12 +174,15 @@ typedef struct _SGDescriptor_struct {
  * PAD_64 can be adjusted independently as needed for 32-bit
  * and 64-bits systems.
  */
-#define COMMANDLIST_ALIGNMENT (8)
+#define COMMANDLIST_ALIGNMENT (32)
 #define IS_64_BIT ((sizeof(long) - 4)/4)
 #define IS_32_BIT (!IS_64_BIT)
-#define PAD_32 (0)
+#define PAD_32 (32)
 #define PAD_64 (4)
 #define PADSIZE (IS_32_BIT * PAD_32 + IS_64_BIT * PAD_64)
+#define DIRECT_LOOKUP_BIT 0x10
+#define DIRECT_LOOKUP_SHIFT 5
+
 typedef struct _CommandList_struct {
   CommandListHeader_struct Header;
   RequestBlock_struct      Request;
@@ -195,7 +199,7 @@ typedef struct _CommandList_struct {
   struct completion *waiting;
   int	 retry_count;
   void * scsi_cmd;
-  char   pad[PADSIZE];
+  char pad[PADSIZE];
 } CommandList_struct;
 
 /* Configuration Table Structure */
@@ -209,12 +213,15 @@ typedef struct _HostWrite_struct {
 typedef struct _CfgTable_struct {
   BYTE             Signature[4];
   DWORD            SpecValence;
+#define SIMPLE_MODE	0x02
+#define PERFORMANT_MODE	0x04
+#define MEMQ_MODE	0x08
   DWORD            TransportSupport;
   DWORD            TransportActive;
   HostWrite_struct HostWrite;
   DWORD            CmdsOutMax;
   DWORD            BusTypes;
-  DWORD            Reserved; 
+  DWORD            TransMethodOffset;
   BYTE             ServerName[16];
   DWORD            HeartBeat;
   DWORD            SCSI_Prefetch;
@@ -222,6 +229,25 @@ typedef struct _CfgTable_struct {
   DWORD            MaxLogicalUnits;
   DWORD            MaxPhysicalDrives;
   DWORD            MaxPhysicalDrivesPerLogicalUnit;
+  DWORD            MaxPerformantModeCommands;
 } CfgTable_struct;
+
+struct TransTable_struct {
+  u32 BlockFetch0;
+  u32 BlockFetch1;
+  u32 BlockFetch2;
+  u32 BlockFetch3;
+  u32 BlockFetch4;
+  u32 BlockFetch5;
+  u32 BlockFetch6;
+  u32 BlockFetch7;
+  u32 RepQSize;
+  u32 RepQCount;
+  u32 RepQCtrAddrLow32;
+  u32 RepQCtrAddrHigh32;
+  u32 RepQAddr0Low32;
+  u32 RepQAddr0High32;
+};
+
 #pragma pack()	 
 #endif /* CCISS_CMD_H */

drivers/block/cciss_scsi.c

@@ -93,8 +93,8 @@ static struct scsi_host_template cciss_driver_template = {
 
 #pragma pack(1)
 
-#define SCSI_PAD_32 0
-#define SCSI_PAD_64 0
+#define SCSI_PAD_32 8
+#define SCSI_PAD_64 8
 
 struct cciss_scsi_cmd_stack_elem_t {
 	CommandList_struct cmd;
@@ -213,6 +213,8 @@ scsi_cmd_stack_setup(int ctlr, struct cciss_scsi_adapter_data_t *sa)
 
 	/* Check alignment, see cciss_cmd.h near CommandList_struct def. */
 	BUILD_BUG_ON((sizeof(*stk->pool) % COMMANDLIST_ALIGNMENT) != 0);
+	/* printk(KERN_WARNING "cciss_scsi.c: 0x%08x 0x%08x 0x%08x\n",
+			0xdeadbeef, sizeof(*stk->pool), 0xbeefdead); */
 	/* pci_alloc_consistent guarantees 32-bit DMA address will be used */
 	stk->pool = (struct cciss_scsi_cmd_stack_elem_t *)
 		pci_alloc_consistent(hba[ctlr]->pdev, size, &stk->cmd_pool_handle);