]> git.karo-electronics.de Git - linux-beck.git/commitdiff
[SCSI] hpsa: add ioaccell mode 1 RAID offload support.
authorStephen M. Cameron <scameron@beardog.cce.hp.com>
Tue, 18 Feb 2014 19:55:33 +0000 (13:55 -0600)
committerJames Bottomley <JBottomley@Parallels.com>
Sat, 15 Mar 2014 17:19:03 +0000 (10:19 -0700)
This enables sending i/o's destined for RAID logical drives
which can be serviced by a single physical disk down a different,
faster i/o path directly to physical drives for certain logical
volumes on SSDs bypassing the Smart Array RAID stack for a
performance improvement.

Signed-off-by: Matt Gates <matthew.gates@hp.com>
Signed-off-by: Stephen M. Cameron <scameron@beardog.cce.hp.com>
Signed-off-by: Scott Teel <scott.teel@hp.com>
Signed-off-by: Mike Miller <michael.miller@canonical.com>
Signed-off-by: Don Brace <brace@beardog.cce.hp.com>
Signed-off-by: Joe Handzik <joseph.t.handzik@hp.com>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
drivers/scsi/hpsa.c
drivers/scsi/hpsa.h
drivers/scsi/hpsa_cmd.h

index 4b6db4c39b24f0429f73a29083516ee58b7cd2bf..e8489a964ea41784306fe57a8f7ec68b584cabcc 100644 (file)
@@ -49,6 +49,7 @@
 #include <linux/atomic.h>
 #include <linux/kthread.h>
 #include <linux/jiffies.h>
+#include <asm/div64.h>
 #include "hpsa_cmd.h"
 #include "hpsa.h"
 
@@ -216,6 +217,7 @@ static int hpsa_lookup_board_id(struct pci_dev *pdev, u32 *board_id);
 static int hpsa_wait_for_board_state(struct pci_dev *pdev, void __iomem *vaddr,
                                     int wait_for_ready);
 static inline void finish_cmd(struct CommandList *c);
+static void hpsa_wait_for_mode_change_ack(struct ctlr_info *h);
 #define BOARD_NOT_READY 0
 #define BOARD_READY 1
 
@@ -1195,6 +1197,7 @@ static void complete_scsi_command(struct CommandList *cp)
        struct scsi_cmnd *cmd;
        struct ctlr_info *h;
        struct ErrorInfo *ei;
+       struct hpsa_scsi_dev_t *dev;
 
        unsigned char sense_key;
        unsigned char asc;      /* additional sense code */
@@ -1204,6 +1207,7 @@ static void complete_scsi_command(struct CommandList *cp)
        ei = cp->err_info;
        cmd = (struct scsi_cmnd *) cp->scsi_cmd;
        h = cp->h;
+       dev = cmd->device->hostdata;
 
        scsi_dma_unmap(cmd); /* undo the DMA mappings */
        if ((cp->cmd_type == CMD_SCSI) &&
@@ -1242,6 +1246,19 @@ static void complete_scsi_command(struct CommandList *cp)
                cp->Header.Tag.upper = c->Tag.upper;
                memcpy(cp->Header.LUN.LunAddrBytes, c->CISS_LUN, 8);
                memcpy(cp->Request.CDB, c->CDB, cp->Request.CDBLen);
+
+               /* Any RAID offload error results in retry which will use
+                * the normal I/O path so the controller can handle whatever's
+                * wrong.
+                */
+               if (is_logical_dev_addr_mode(dev->scsi3addr)) {
+                       if (ei->CommandStatus == CMD_IOACCEL_DISABLED)
+                               dev->offload_enabled = 0;
+                       cmd->result = DID_SOFT_ERROR << 16;
+                       cmd_free(h, cp);
+                       cmd->scsi_done(cmd);
+                       return;
+               }
        }
 
        /* an error has occurred */
@@ -1406,6 +1423,14 @@ static void complete_scsi_command(struct CommandList *cp)
                cmd->result = DID_ERROR << 16;
                dev_warn(&h->pdev->dev, "Command unabortable\n");
                break;
+       case CMD_IOACCEL_DISABLED:
+               /* This only handles the direct pass-through case since RAID
+                * offload is handled above.  Just attempt a retry.
+                */
+               cmd->result = DID_SOFT_ERROR << 16;
+               dev_warn(&h->pdev->dev,
+                               "cp %p had HP SSD Smart Path error\n", cp);
+               break;
        default:
                cmd->result = DID_ERROR << 16;
                dev_warn(&h->pdev->dev, "cp %p returned unknown status %x\n",
@@ -1650,6 +1675,147 @@ static void hpsa_get_raid_level(struct ctlr_info *h,
        return;
 }
 
+#define HPSA_MAP_DEBUG
+#ifdef HPSA_MAP_DEBUG
+static void hpsa_debug_map_buff(struct ctlr_info *h, int rc,
+                               struct raid_map_data *map_buff)
+{
+       struct raid_map_disk_data *dd = &map_buff->data[0];
+       int map, row, col;
+       u16 map_cnt, row_cnt, disks_per_row;
+
+       if (rc != 0)
+               return;
+
+       dev_info(&h->pdev->dev, "structure_size = %u\n",
+                               le32_to_cpu(map_buff->structure_size));
+       dev_info(&h->pdev->dev, "volume_blk_size = %u\n",
+                       le32_to_cpu(map_buff->volume_blk_size));
+       dev_info(&h->pdev->dev, "volume_blk_cnt = 0x%llx\n",
+                       le64_to_cpu(map_buff->volume_blk_cnt));
+       dev_info(&h->pdev->dev, "physicalBlockShift = %u\n",
+                       map_buff->phys_blk_shift);
+       dev_info(&h->pdev->dev, "parity_rotation_shift = %u\n",
+                       map_buff->parity_rotation_shift);
+       dev_info(&h->pdev->dev, "strip_size = %u\n",
+                       le16_to_cpu(map_buff->strip_size));
+       dev_info(&h->pdev->dev, "disk_starting_blk = 0x%llx\n",
+                       le64_to_cpu(map_buff->disk_starting_blk));
+       dev_info(&h->pdev->dev, "disk_blk_cnt = 0x%llx\n",
+                       le64_to_cpu(map_buff->disk_blk_cnt));
+       dev_info(&h->pdev->dev, "data_disks_per_row = %u\n",
+                       le16_to_cpu(map_buff->data_disks_per_row));
+       dev_info(&h->pdev->dev, "metadata_disks_per_row = %u\n",
+                       le16_to_cpu(map_buff->metadata_disks_per_row));
+       dev_info(&h->pdev->dev, "row_cnt = %u\n",
+                       le16_to_cpu(map_buff->row_cnt));
+       dev_info(&h->pdev->dev, "layout_map_count = %u\n",
+                       le16_to_cpu(map_buff->layout_map_count));
+
+       map_cnt = le16_to_cpu(map_buff->layout_map_count);
+       for (map = 0; map < map_cnt; map++) {
+               dev_info(&h->pdev->dev, "Map%u:\n", map);
+               row_cnt = le16_to_cpu(map_buff->row_cnt);
+               for (row = 0; row < row_cnt; row++) {
+                       dev_info(&h->pdev->dev, "  Row%u:\n", row);
+                       disks_per_row =
+                               le16_to_cpu(map_buff->data_disks_per_row);
+                       for (col = 0; col < disks_per_row; col++, dd++)
+                               dev_info(&h->pdev->dev,
+                                       "    D%02u: h=0x%04x xor=%u,%u\n",
+                                       col, dd->ioaccel_handle,
+                                       dd->xor_mult[0], dd->xor_mult[1]);
+                       disks_per_row =
+                               le16_to_cpu(map_buff->metadata_disks_per_row);
+                       for (col = 0; col < disks_per_row; col++, dd++)
+                               dev_info(&h->pdev->dev,
+                                       "    M%02u: h=0x%04x xor=%u,%u\n",
+                                       col, dd->ioaccel_handle,
+                                       dd->xor_mult[0], dd->xor_mult[1]);
+               }
+       }
+}
+#else
+static void hpsa_debug_map_buff(__attribute__((unused)) struct ctlr_info *h,
+                       __attribute__((unused)) int rc,
+                       __attribute__((unused)) struct raid_map_data *map_buff)
+{
+}
+#endif
+
+static int hpsa_get_raid_map(struct ctlr_info *h,
+       unsigned char *scsi3addr, struct hpsa_scsi_dev_t *this_device)
+{
+       int rc = 0;
+       struct CommandList *c;
+       struct ErrorInfo *ei;
+
+       c = cmd_special_alloc(h);
+       if (c == NULL) {
+               dev_warn(&h->pdev->dev, "cmd_special_alloc returned NULL!\n");
+               return -ENOMEM;
+       }
+       if (fill_cmd(c, HPSA_GET_RAID_MAP, h, &this_device->raid_map,
+                       sizeof(this_device->raid_map), 0,
+                       scsi3addr, TYPE_CMD)) {
+               dev_warn(&h->pdev->dev, "Out of memory in hpsa_get_raid_map()\n");
+               cmd_special_free(h, c);
+               return -ENOMEM;
+       }
+       hpsa_scsi_do_simple_cmd_with_retry(h, c, PCI_DMA_FROMDEVICE);
+       ei = c->err_info;
+       if (ei->CommandStatus != 0 && ei->CommandStatus != CMD_DATA_UNDERRUN) {
+               hpsa_scsi_interpret_error(c);
+               cmd_special_free(h, c);
+               return -1;
+       }
+       cmd_special_free(h, c);
+
+       /* @todo in the future, dynamically allocate RAID map memory */
+       if (le32_to_cpu(this_device->raid_map.structure_size) >
+                               sizeof(this_device->raid_map)) {
+               dev_warn(&h->pdev->dev, "RAID map size is too large!\n");
+               rc = -1;
+       }
+       hpsa_debug_map_buff(h, rc, &this_device->raid_map);
+       return rc;
+}
+
+static void hpsa_get_ioaccel_status(struct ctlr_info *h,
+       unsigned char *scsi3addr, struct hpsa_scsi_dev_t *this_device)
+{
+       int rc;
+       unsigned char *buf;
+       u8 ioaccel_status;
+
+       this_device->offload_config = 0;
+       this_device->offload_enabled = 0;
+
+       buf = kzalloc(64, GFP_KERNEL);
+       if (!buf)
+               return;
+       rc = hpsa_scsi_do_inquiry(h, scsi3addr,
+                       HPSA_VPD_LV_IOACCEL_STATUS, buf, 64);
+       if (rc != 0)
+               goto out;
+
+#define IOACCEL_STATUS_BYTE 4
+#define OFFLOAD_CONFIGURED_BIT 0x01
+#define OFFLOAD_ENABLED_BIT 0x02
+       ioaccel_status = buf[IOACCEL_STATUS_BYTE];
+       this_device->offload_config =
+               !!(ioaccel_status & OFFLOAD_CONFIGURED_BIT);
+       if (this_device->offload_config) {
+               this_device->offload_enabled =
+                       !!(ioaccel_status & OFFLOAD_ENABLED_BIT);
+               if (hpsa_get_raid_map(h, scsi3addr, this_device))
+                       this_device->offload_enabled = 0;
+       }
+out:
+       kfree(buf);
+       return;
+}
+
 /* Get the device id from inquiry page 0x83 */
 static int hpsa_get_device_id(struct ctlr_info *h, unsigned char *scsi3addr,
        unsigned char *device_id, int buflen)
@@ -1698,6 +1864,14 @@ static int hpsa_scsi_do_report_luns(struct ctlr_info *h, int logical,
            ei->CommandStatus != CMD_DATA_UNDERRUN) {
                hpsa_scsi_interpret_error(c);
                rc = -1;
+       } else {
+               if (buf->extended_response_flag != extended_response) {
+                       dev_err(&h->pdev->dev,
+                               "report luns requested format %u, got %u\n",
+                               extended_response,
+                               buf->extended_response_flag);
+                       rc = -1;
+               }
        }
 out:
        cmd_special_free(h, c);
@@ -1763,10 +1937,15 @@ static int hpsa_update_device_info(struct ctlr_info *h,
                sizeof(this_device->device_id));
 
        if (this_device->devtype == TYPE_DISK &&
-               is_logical_dev_addr_mode(scsi3addr))
+               is_logical_dev_addr_mode(scsi3addr)) {
                hpsa_get_raid_level(h, scsi3addr, &this_device->raid_level);
-       else
+               if (h->fw_support & MISC_FW_RAID_OFFLOAD_BASIC)
+                       hpsa_get_ioaccel_status(h, scsi3addr, this_device);
+       } else {
                this_device->raid_level = RAID_UNKNOWN;
+               this_device->offload_config = 0;
+               this_device->offload_enabled = 0;
+       }
 
        if (is_OBDR_device) {
                /* See if this is a One-Button-Disaster-Recovery device
@@ -1903,15 +2082,25 @@ static int add_ext_target_dev(struct ctlr_info *h,
  */
 static int hpsa_gather_lun_info(struct ctlr_info *h,
        int reportlunsize,
-       struct ReportLUNdata *physdev, u32 *nphysicals,
+       struct ReportLUNdata *physdev, u32 *nphysicals, int *physical_mode,
        struct ReportLUNdata *logdev, u32 *nlogicals)
 {
+       int physical_entry_size = 8;
+
+       *physical_mode = 0;
+
+       /* For I/O accelerator mode we need to read physical device handles */
+       if (h->transMethod & CFGTBL_Trans_io_accel1) {
+               *physical_mode = HPSA_REPORT_PHYS_EXTENDED;
+               physical_entry_size = 24;
+       }
        if (hpsa_scsi_do_report_phys_luns(h, physdev, reportlunsize,
-                                       HPSA_REPORT_PHYS_EXTENDED)) {
+                                                       *physical_mode)) {
                dev_err(&h->pdev->dev, "report physical LUNs failed.\n");
                return -1;
        }
-       *nphysicals = be32_to_cpu(*((__be32 *)physdev->LUNListLength)) / 24;
+       *nphysicals = be32_to_cpu(*((__be32 *)physdev->LUNListLength)) /
+                                                       physical_entry_size;
        if (*nphysicals > HPSA_MAX_PHYS_LUN) {
                dev_warn(&h->pdev->dev, "maximum physical LUNs (%d) exceeded."
                        "  %d LUNs ignored.\n", HPSA_MAX_PHYS_LUN,
@@ -1983,10 +2172,11 @@ static void hpsa_update_scsi_devices(struct ctlr_info *h, int hostno)
        struct ReportLUNdata *logdev_list = NULL;
        u32 nphysicals = 0;
        u32 nlogicals = 0;
+       int physical_mode = 0;
        u32 ndev_allocated = 0;
        struct hpsa_scsi_dev_t **currentsd, *this_device, *tmpdevice;
        int ncurrent = 0;
-       int reportlunsize = sizeof(*physdev_list) + HPSA_MAX_PHYS_LUN * 8;
+       int reportlunsize = sizeof(*physdev_list) + HPSA_MAX_PHYS_LUN * 24;
        int i, n_ext_target_devs, ndevs_to_allocate;
        int raid_ctlr_position;
        DECLARE_BITMAP(lunzerobits, MAX_EXT_TARGETS);
@@ -2004,7 +2194,7 @@ static void hpsa_update_scsi_devices(struct ctlr_info *h, int hostno)
 
        if (hpsa_gather_lun_info(h, reportlunsize,
                        (struct ReportLUNdata *) physdev_list, &nphysicals,
-                       logdev_list, &nlogicals))
+                       &physical_mode, logdev_list, &nlogicals))
                goto out;
 
        /* We might see up to the maximum number of logical and physical disks
@@ -2085,12 +2275,16 @@ static void hpsa_update_scsi_devices(struct ctlr_info *h, int hostno)
                                ncurrent++;
                        break;
                case TYPE_DISK:
-                       if (i < nphysicals)
+                       if (i >= nphysicals) {
+                               ncurrent++;
                                break;
-                       memcpy(&this_device->ioaccel_handle,
-                               &lunaddrbytes[20],
-                               sizeof(this_device->ioaccel_handle));
-                       ncurrent++;
+                       }
+                       if (physical_mode == HPSA_REPORT_PHYS_EXTENDED) {
+                               memcpy(&this_device->ioaccel_handle,
+                                       &lunaddrbytes[20],
+                                       sizeof(this_device->ioaccel_handle));
+                               ncurrent++;
+                       }
                        break;
                case TYPE_TAPE:
                case TYPE_MEDIUM_CHANGER:
@@ -2184,15 +2378,62 @@ sglist_finished:
        return 0;
 }
 
+#define IO_ACCEL_INELIGIBLE (1)
+static int fixup_ioaccel_cdb(u8 *cdb, int *cdb_len)
+{
+       int is_write = 0;
+       u32 block;
+       u32 block_cnt;
+
+       /* Perform some CDB fixups if needed using 10 byte reads/writes only */
+       switch (cdb[0]) {
+       case WRITE_6:
+       case WRITE_12:
+               is_write = 1;
+       case READ_6:
+       case READ_12:
+               if (*cdb_len == 6) {
+                       block = (((u32) cdb[2]) << 8) | cdb[3];
+                       block_cnt = cdb[4];
+               } else {
+                       BUG_ON(*cdb_len != 12);
+                       block = (((u32) cdb[2]) << 24) |
+                               (((u32) cdb[3]) << 16) |
+                               (((u32) cdb[4]) << 8) |
+                               cdb[5];
+                       block_cnt =
+                               (((u32) cdb[6]) << 24) |
+                               (((u32) cdb[7]) << 16) |
+                               (((u32) cdb[8]) << 8) |
+                               cdb[9];
+               }
+               if (block_cnt > 0xffff)
+                       return IO_ACCEL_INELIGIBLE;
+
+               cdb[0] = is_write ? WRITE_10 : READ_10;
+               cdb[1] = 0;
+               cdb[2] = (u8) (block >> 24);
+               cdb[3] = (u8) (block >> 16);
+               cdb[4] = (u8) (block >> 8);
+               cdb[5] = (u8) (block);
+               cdb[6] = 0;
+               cdb[7] = (u8) (block_cnt >> 8);
+               cdb[8] = (u8) (block_cnt);
+               cdb[9] = 0;
+               *cdb_len = 10;
+               break;
+       }
+       return 0;
+}
+
 /*
  * Queue a command to the I/O accelerator path.
- * This method does not currently support S/G chaining.
  */
 static int hpsa_scsi_ioaccel_queue_command(struct ctlr_info *h,
-                                       struct CommandList *c)
+       struct CommandList *c, u32 ioaccel_handle, u8 *cdb, int cdb_len,
+       u8 *scsi3addr)
 {
        struct scsi_cmnd *cmd = c->scsi_cmd;
-       struct hpsa_scsi_dev_t *dev = cmd->device->hostdata;
        struct io_accel1_cmd *cp = &h->ioaccel_cmd_pool[c->cmdindex];
        unsigned int len;
        unsigned int total_len = 0;
@@ -2202,8 +2443,15 @@ static int hpsa_scsi_ioaccel_queue_command(struct ctlr_info *h,
        struct SGDescriptor *curr_sg;
        u32 control = IOACCEL1_CONTROL_SIMPLEQUEUE;
 
+       /* TODO: implement chaining support */
+       if (scsi_sg_count(cmd) > h->ioaccel_maxsg)
+               return IO_ACCEL_INELIGIBLE;
+
        BUG_ON(cmd->cmd_len > IOACCEL1_IOFLAGS_CDBLEN_MAX);
 
+       if (fixup_ioaccel_cdb(cdb, &cdb_len))
+               return IO_ACCEL_INELIGIBLE;
+
        c->cmd_type = CMD_IOACCEL1;
 
        /* Adjust the DMA address to point to the accelerated command buffer */
@@ -2254,13 +2502,13 @@ static int hpsa_scsi_ioaccel_queue_command(struct ctlr_info *h,
        }
 
        /* Fill out the command structure to submit */
-       cp->dev_handle = dev->ioaccel_handle;
+       cp->dev_handle = ioaccel_handle & 0xFFFF;
        cp->transfer_len = total_len;
        cp->io_flags = IOACCEL1_IOFLAGS_IO_REQ |
-                       (cmd->cmd_len & IOACCEL1_IOFLAGS_CDBLEN_MASK);
+                       (cdb_len & IOACCEL1_IOFLAGS_CDBLEN_MASK);
        cp->control = control;
-       memcpy(cp->CDB, cmd->cmnd, cmd->cmd_len);
-       memcpy(cp->CISS_LUN, dev->scsi3addr, 8);
+       memcpy(cp->CDB, cdb, cdb_len);
+       memcpy(cp->CISS_LUN, scsi3addr, 8);
 
        /* Tell the controller to post the reply to the queue for this
         * processor.  This seems to give the best I/O throughput.
@@ -2274,15 +2522,214 @@ static int hpsa_scsi_ioaccel_queue_command(struct ctlr_info *h,
         */
        c->busaddr |= 1 | (h->ioaccel1_blockFetchTable[use_sg] << 1) |
                                        IOACCEL1_BUSADDR_CMDTYPE;
-
-       /* execute command (bypassing cmd queue if possible) */
-       if (unlikely(h->access.fifo_full(h)))
-               enqueue_cmd_and_start_io(h, c);
-       else
-               h->access.submit_command(h, c);
+       enqueue_cmd_and_start_io(h, c);
        return 0;
 }
 
+/*
+ * Queue a command directly to a device behind the controller using the
+ * I/O accelerator path.
+ */
+static int hpsa_scsi_ioaccel_direct_map(struct ctlr_info *h,
+       struct CommandList *c)
+{
+       struct scsi_cmnd *cmd = c->scsi_cmd;
+       struct hpsa_scsi_dev_t *dev = cmd->device->hostdata;
+
+       return hpsa_scsi_ioaccel_queue_command(h, c, dev->ioaccel_handle,
+               cmd->cmnd, cmd->cmd_len, dev->scsi3addr);
+}
+
+/*
+ * Attempt to perform offload RAID mapping for a logical volume I/O.
+ */
+static int hpsa_scsi_ioaccel_raid_map(struct ctlr_info *h,
+       struct CommandList *c)
+{
+       struct scsi_cmnd *cmd = c->scsi_cmd;
+       struct hpsa_scsi_dev_t *dev = cmd->device->hostdata;
+       struct raid_map_data *map = &dev->raid_map;
+       struct raid_map_disk_data *dd = &map->data[0];
+       int is_write = 0;
+       u32 map_index;
+       u64 first_block, last_block;
+       u32 block_cnt;
+       u32 blocks_per_row;
+       u64 first_row, last_row;
+       u32 first_row_offset, last_row_offset;
+       u32 first_column, last_column;
+       u32 map_row;
+       u32 disk_handle;
+       u64 disk_block;
+       u32 disk_block_cnt;
+       u8 cdb[16];
+       u8 cdb_len;
+#if BITS_PER_LONG == 32
+       u64 tmpdiv;
+#endif
+
+       BUG_ON(!(dev->offload_config && dev->offload_enabled));
+
+       /* check for valid opcode, get LBA and block count */
+       switch (cmd->cmnd[0]) {
+       case WRITE_6:
+               is_write = 1;
+       case READ_6:
+               first_block =
+                       (((u64) cmd->cmnd[2]) << 8) |
+                       cmd->cmnd[3];
+               block_cnt = cmd->cmnd[4];
+               break;
+       case WRITE_10:
+               is_write = 1;
+       case READ_10:
+               first_block =
+                       (((u64) cmd->cmnd[2]) << 24) |
+                       (((u64) cmd->cmnd[3]) << 16) |
+                       (((u64) cmd->cmnd[4]) << 8) |
+                       cmd->cmnd[5];
+               block_cnt =
+                       (((u32) cmd->cmnd[7]) << 8) |
+                       cmd->cmnd[8];
+               break;
+       case WRITE_12:
+               is_write = 1;
+       case READ_12:
+               first_block =
+                       (((u64) cmd->cmnd[2]) << 24) |
+                       (((u64) cmd->cmnd[3]) << 16) |
+                       (((u64) cmd->cmnd[4]) << 8) |
+                       cmd->cmnd[5];
+               block_cnt =
+                       (((u32) cmd->cmnd[6]) << 24) |
+                       (((u32) cmd->cmnd[7]) << 16) |
+                       (((u32) cmd->cmnd[8]) << 8) |
+               cmd->cmnd[9];
+               break;
+       case WRITE_16:
+               is_write = 1;
+       case READ_16:
+               first_block =
+                       (((u64) cmd->cmnd[2]) << 56) |
+                       (((u64) cmd->cmnd[3]) << 48) |
+                       (((u64) cmd->cmnd[4]) << 40) |
+                       (((u64) cmd->cmnd[5]) << 32) |
+                       (((u64) cmd->cmnd[6]) << 24) |
+                       (((u64) cmd->cmnd[7]) << 16) |
+                       (((u64) cmd->cmnd[8]) << 8) |
+                       cmd->cmnd[9];
+               block_cnt =
+                       (((u32) cmd->cmnd[10]) << 24) |
+                       (((u32) cmd->cmnd[11]) << 16) |
+                       (((u32) cmd->cmnd[12]) << 8) |
+                       cmd->cmnd[13];
+               break;
+       default:
+               return IO_ACCEL_INELIGIBLE; /* process via normal I/O path */
+       }
+       BUG_ON(block_cnt == 0);
+       last_block = first_block + block_cnt - 1;
+
+       /* check for write to non-RAID-0 */
+       if (is_write && dev->raid_level != 0)
+               return IO_ACCEL_INELIGIBLE;
+
+       /* check for invalid block or wraparound */
+       if (last_block >= map->volume_blk_cnt || last_block < first_block)
+               return IO_ACCEL_INELIGIBLE;
+
+       /* calculate stripe information for the request */
+       blocks_per_row = map->data_disks_per_row * map->strip_size;
+#if BITS_PER_LONG == 32
+       tmpdiv = first_block;
+       (void) do_div(tmpdiv, blocks_per_row);
+       first_row = tmpdiv;
+       tmpdiv = last_block;
+       (void) do_div(tmpdiv, blocks_per_row);
+       last_row = tmpdiv;
+       first_row_offset = (u32) (first_block - (first_row * blocks_per_row));
+       last_row_offset = (u32) (last_block - (last_row * blocks_per_row));
+       tmpdiv = first_row_offset;
+       (void) do_div(tmpdiv,  map->strip_size);
+       first_column = tmpdiv;
+       tmpdiv = last_row_offset;
+       (void) do_div(tmpdiv, map->strip_size);
+       last_column = tmpdiv;
+#else
+       first_row = first_block / blocks_per_row;
+       last_row = last_block / blocks_per_row;
+       first_row_offset = (u32) (first_block - (first_row * blocks_per_row));
+       last_row_offset = (u32) (last_block - (last_row * blocks_per_row));
+       first_column = first_row_offset / map->strip_size;
+       last_column = last_row_offset / map->strip_size;
+#endif
+
+       /* if this isn't a single row/column then give to the controller */
+       if ((first_row != last_row) || (first_column != last_column))
+               return IO_ACCEL_INELIGIBLE;
+
+       /* proceeding with driver mapping */
+       map_row = ((u32)(first_row >> map->parity_rotation_shift)) %
+                               map->row_cnt;
+       map_index = (map_row * (map->data_disks_per_row +
+                               map->metadata_disks_per_row)) + first_column;
+       if (dev->raid_level == 2) {
+               /* simple round-robin balancing of RAID 1+0 reads across
+                * primary and mirror members.  this is appropriate for SSD
+                * but not optimal for HDD.
+                */
+               if (dev->offload_to_mirror)
+                       map_index += map->data_disks_per_row;
+               dev->offload_to_mirror = !dev->offload_to_mirror;
+       }
+       disk_handle = dd[map_index].ioaccel_handle;
+       disk_block = map->disk_starting_blk + (first_row * map->strip_size) +
+                       (first_row_offset - (first_column * map->strip_size));
+       disk_block_cnt = block_cnt;
+
+       /* handle differing logical/physical block sizes */
+       if (map->phys_blk_shift) {
+               disk_block <<= map->phys_blk_shift;
+               disk_block_cnt <<= map->phys_blk_shift;
+       }
+       BUG_ON(disk_block_cnt > 0xffff);
+
+       /* build the new CDB for the physical disk I/O */
+       if (disk_block > 0xffffffff) {
+               cdb[0] = is_write ? WRITE_16 : READ_16;
+               cdb[1] = 0;
+               cdb[2] = (u8) (disk_block >> 56);
+               cdb[3] = (u8) (disk_block >> 48);
+               cdb[4] = (u8) (disk_block >> 40);
+               cdb[5] = (u8) (disk_block >> 32);
+               cdb[6] = (u8) (disk_block >> 24);
+               cdb[7] = (u8) (disk_block >> 16);
+               cdb[8] = (u8) (disk_block >> 8);
+               cdb[9] = (u8) (disk_block);
+               cdb[10] = (u8) (disk_block_cnt >> 24);
+               cdb[11] = (u8) (disk_block_cnt >> 16);
+               cdb[12] = (u8) (disk_block_cnt >> 8);
+               cdb[13] = (u8) (disk_block_cnt);
+               cdb[14] = 0;
+               cdb[15] = 0;
+               cdb_len = 16;
+       } else {
+               cdb[0] = is_write ? WRITE_10 : READ_10;
+               cdb[1] = 0;
+               cdb[2] = (u8) (disk_block >> 24);
+               cdb[3] = (u8) (disk_block >> 16);
+               cdb[4] = (u8) (disk_block >> 8);
+               cdb[5] = (u8) (disk_block);
+               cdb[6] = 0;
+               cdb[7] = (u8) (disk_block_cnt >> 8);
+               cdb[8] = (u8) (disk_block_cnt);
+               cdb[9] = 0;
+               cdb_len = 10;
+       }
+       return hpsa_scsi_ioaccel_queue_command(h, c, disk_handle, cdb, cdb_len,
+                                               dev->scsi3addr);
+}
+
 static int hpsa_scsi_queue_command_lck(struct scsi_cmnd *cmd,
        void (*done)(struct scsi_cmnd *))
 {
@@ -2291,6 +2738,7 @@ static int hpsa_scsi_queue_command_lck(struct scsi_cmnd *cmd,
        unsigned char scsi3addr[8];
        struct CommandList *c;
        unsigned long flags;
+       int rc = 0;
 
        /* Get the ptr to our adapter structure out of cmd->host. */
        h = sdev_to_hba(cmd->device);
@@ -2326,13 +2774,29 @@ static int hpsa_scsi_queue_command_lck(struct scsi_cmnd *cmd,
        c->cmd_type = CMD_SCSI;
        c->scsi_cmd = cmd;
 
-       /* Call alternate submit routine for I/O accelerated commands */
-       if ((likely(h->transMethod & CFGTBL_Trans_io_accel1)) &&
-               (dev->ioaccel_handle) &&
-               ((cmd->cmnd[0] == READ_10) || (cmd->cmnd[0] == WRITE_10)) &&
-               (scsi_sg_count(cmd) <= IOACCEL1_MAXSGENTRIES) &&
-               likely(cmd->request->cmd_type == REQ_TYPE_FS))
-               return hpsa_scsi_ioaccel_queue_command(h, c);
+       /* Call alternate submit routine for I/O accelerated commands.
+        * Retries always go down the normal I/O path.
+        */
+       if (likely(cmd->retries == 0 &&
+               cmd->request->cmd_type == REQ_TYPE_FS)) {
+               if (dev->offload_enabled) {
+                       rc = hpsa_scsi_ioaccel_raid_map(h, c);
+                       if (rc == 0)
+                               return 0; /* Sent on ioaccel path */
+                       if (rc < 0) {   /* scsi_dma_map failed. */
+                               cmd_free(h, c);
+                               return SCSI_MLQUEUE_HOST_BUSY;
+                       }
+               } else if (dev->ioaccel_handle) {
+                       rc = hpsa_scsi_ioaccel_direct_map(h, c);
+                       if (rc == 0)
+                               return 0; /* Sent on direct map path */
+                       if (rc < 0) {   /* scsi_dma_map failed. */
+                               cmd_free(h, c);
+                               return SCSI_MLQUEUE_HOST_BUSY;
+                       }
+               }
+       }
 
        c->Header.ReplyQueue = 0;  /* unused in simple mode */
        memcpy(&c->Header.LUN.LunAddrBytes[0], &scsi3addr[0], 8);
@@ -3515,6 +3979,18 @@ static int fill_cmd(struct CommandList *c, u8 cmd, struct ctlr_info *h,
                        c->Request.Type.Direction = XFER_NONE;
                        c->Request.Timeout = 0;
                        break;
+               case HPSA_GET_RAID_MAP:
+                       c->Request.CDBLen = 12;
+                       c->Request.Type.Attribute = ATTR_SIMPLE;
+                       c->Request.Type.Direction = XFER_READ;
+                       c->Request.Timeout = 0;
+                       c->Request.CDB[0] = HPSA_CISS_READ;
+                       c->Request.CDB[1] = cmd;
+                       c->Request.CDB[6] = (size >> 24) & 0xFF; /* MSB */
+                       c->Request.CDB[7] = (size >> 16) & 0xFF;
+                       c->Request.CDB[8] = (size >> 8) & 0xFF;
+                       c->Request.CDB[9] = size & 0xFF;
+                       break;
                default:
                        dev_warn(&h->pdev->dev, "unknown command 0x%c\n", cmd);
                        BUG();
@@ -4485,6 +4961,7 @@ static void hpsa_find_board_params(struct ctlr_info *h)
        hpsa_get_max_perf_mode_cmds(h);
        h->nr_cmds = h->max_commands - 4; /* Allow room for some ioctls */
        h->maxsgentries = readl(&(h->cfgtable->MaxScatterGatherElements));
+       h->fw_support = readl(&(h->cfgtable->misc_fw_support));
        /*
         * Limit in-command s/g elements to 32 save dma'able memory.
         * Howvever spec says if 0, use 31
@@ -4569,18 +5046,19 @@ static int hpsa_enter_simple_mode(struct ctlr_info *h)
                return -ENOTSUPP;
 
        h->max_commands = readl(&(h->cfgtable->CmdsOutMax));
+
        /* Update the field, and then ring the doorbell */
        writel(CFGTBL_Trans_Simple, &(h->cfgtable->HostWrite.TransportRequest));
        writel(CFGTBL_ChangeReq, h->vaddr + SA5_DOORBELL);
        hpsa_wait_for_mode_change_ack(h);
        print_cfg_table(&h->pdev->dev, h->cfgtable);
-       if (!(readl(&(h->cfgtable->TransportActive)) & CFGTBL_Trans_Simple)) {
-               dev_warn(&h->pdev->dev,
-                       "unable to get board into simple mode\n");
-               return -ENODEV;
-       }
+       if (!(readl(&(h->cfgtable->TransportActive)) & CFGTBL_Trans_Simple))
+               goto error;
        h->transMethod = CFGTBL_Trans_Simple;
        return 0;
+error:
+       dev_warn(&h->pdev->dev, "unable to get board into simple mode\n");
+       return -ENODEV;
 }
 
 static int hpsa_pci_init(struct ctlr_info *h)
@@ -4961,7 +5439,7 @@ reinit_after_soft_reset:
         * the 5 lower bits of the address are used by the hardware. and by
         * the driver.  See comments in hpsa.h for more info.
         */
-#define COMMANDLIST_ALIGNMENT 32
+#define COMMANDLIST_ALIGNMENT 128
        BUILD_BUG_ON(sizeof(struct CommandList) % COMMANDLIST_ALIGNMENT);
        h = kzalloc(sizeof(*h), GFP_KERNEL);
        if (!h)
@@ -5338,8 +5816,8 @@ static void hpsa_enter_performant_mode(struct ctlr_info *h, u32 trans_support)
                h->reply_queue[i].current_entry =
                        readl(h->vaddr + IOACCEL_MODE1_PRODUCER_INDEX);
        }
-       bft[7] = IOACCEL1_MAXSGENTRIES + 8;
-       calc_bucket_map(bft, ARRAY_SIZE(bft), IOACCEL1_MAXSGENTRIES, 8,
+       bft[7] = h->ioaccel_maxsg + 8;
+       calc_bucket_map(bft, ARRAY_SIZE(bft), h->ioaccel_maxsg, 8,
                        h->ioaccel1_blockFetchTable);
 
        /* initialize all reply queue entries to unused */
@@ -5370,6 +5848,11 @@ static void hpsa_enter_performant_mode(struct ctlr_info *h, u32 trans_support)
 
 static int hpsa_alloc_ioaccel_cmd_and_bft(struct ctlr_info *h)
 {
+       h->ioaccel_maxsg =
+               readl(&(h->cfgtable->io_accel_max_embedded_sg_count));
+       if (h->ioaccel_maxsg > IOACCEL1_MAXSGENTRIES)
+               h->ioaccel_maxsg = IOACCEL1_MAXSGENTRIES;
+
        /* Command structures must be aligned on a 128-byte boundary
         * because the 7 lower bits of the address are used by the
         * hardware.
@@ -5383,7 +5866,7 @@ static int hpsa_alloc_ioaccel_cmd_and_bft(struct ctlr_info *h)
                        &(h->ioaccel_cmd_pool_dhandle));
 
        h->ioaccel1_blockFetchTable =
-               kmalloc(((IOACCEL1_MAXSGENTRIES + 1) *
+               kmalloc(((h->ioaccel_maxsg + 1) *
                                sizeof(u32)), GFP_KERNEL);
 
        if ((h->ioaccel_cmd_pool == NULL) ||
index c7865f30ffd1f453022cf1ce1c2411012a9aa579..ae08f1c46272567e8d432c9deba245abbb1cc28b 100644 (file)
@@ -47,6 +47,13 @@ struct hpsa_scsi_dev_t {
        unsigned char model[16];        /* bytes 16-31 of inquiry data */
        unsigned char raid_level;       /* from inquiry page 0xC1 */
        u32 ioaccel_handle;
+       int offload_config;             /* I/O accel RAID offload configured */
+       int offload_enabled;            /* I/O accel RAID offload enabled */
+       int offload_to_mirror;          /* Send next I/O accelerator RAID
+                                        * offload request to mirror drive
+                                        */
+       struct raid_map_data raid_map;  /* I/O accelerator RAID map */
+
 };
 
 struct reply_pool {
@@ -133,6 +140,10 @@ struct ctlr_info {
        u32 *blockFetchTable;
        u32 *ioaccel1_blockFetchTable;
        unsigned char *hba_inquiry_data;
+       u32 driver_support;
+       u32 fw_support;
+       int ioaccel_support;
+       int ioaccel_maxsg;
        u64 last_intr_timestamp;
        u32 last_heartbeat;
        u64 last_heartbeat_timestamp;
@@ -406,8 +417,7 @@ static bool SA5_ioaccel_mode1_intr_pending(struct ctlr_info *h)
 #define IOACCEL_MODE1_CONSUMER_INDEX     0x1BC
 #define IOACCEL_MODE1_REPLY_UNUSED       0xFFFFFFFFFFFFFFFFULL
 
-static unsigned long SA5_ioaccel_mode1_completed(struct ctlr_info *h,
-                                                       u8 q)
+static unsigned long SA5_ioaccel_mode1_completed(struct ctlr_info *h, u8 q)
 {
        u64 register_value;
        struct reply_pool *rq = &h->reply_queue[q];
@@ -420,12 +430,18 @@ static unsigned long SA5_ioaccel_mode1_completed(struct ctlr_info *h,
                rq->head[rq->current_entry] = IOACCEL_MODE1_REPLY_UNUSED;
                if (++rq->current_entry == rq->size)
                        rq->current_entry = 0;
+               /*
+                * @todo
+                *
+                * Don't really need to write the new index after each command,
+                * but with current driver design this is easiest.
+                */
+               wmb();
+               writel((q << 24) | rq->current_entry, h->vaddr +
+                               IOACCEL_MODE1_CONSUMER_INDEX);
                spin_lock_irqsave(&h->lock, flags);
                h->commands_outstanding--;
                spin_unlock_irqrestore(&h->lock, flags);
-       } else {
-               writel((q << 24) | rq->current_entry,
-                       h->vaddr + IOACCEL_MODE1_CONSUMER_INDEX);
        }
        return (unsigned long) register_value;
 }
index e682d2e6b387f70633b931becc00fe74ff67f1b6..c1ae8d2a6bf2ab7889a25650f474f4a38b124a9d 100644 (file)
@@ -42,6 +42,8 @@
 #define CMD_UNSOLICITED_ABORT   0x000A
 #define CMD_TIMEOUT             0x000B
 #define CMD_UNABORTABLE                0x000C
+#define CMD_IOACCEL_DISABLED   0x000E
+
 
 /* Unit Attentions ASC's as defined for the MSA2012sa */
 #define POWER_OR_RESET                 0x29
 #define CFGTBL_BusType_Ultra3   0x00000002l
 #define CFGTBL_BusType_Fibre1G  0x00000100l
 #define CFGTBL_BusType_Fibre2G  0x00000200l
+
+/* VPD Inquiry types */
+#define HPSA_VPD_LV_DEVICE_GEOMETRY     0xC1
+#define HPSA_VPD_LV_IOACCEL_STATUS      0xC2
+
 struct vals32 {
        u32   lower;
        u32   upper;
@@ -165,9 +172,46 @@ struct InquiryData {
 #define HPSA_REPORT_LOG 0xc2    /* Report Logical LUNs */
 #define HPSA_REPORT_PHYS 0xc3   /* Report Physical LUNs */
 #define HPSA_REPORT_PHYS_EXTENDED 0x02
+#define HPSA_CISS_READ 0xc0    /* CISS Read */
+#define HPSA_GET_RAID_MAP 0xc8 /* CISS Get RAID Layout Map */
+
+#define RAID_MAP_MAX_ENTRIES   256
+
+struct raid_map_disk_data {
+       u32   ioaccel_handle;         /**< Handle to access this disk via the
+                                       *  I/O accelerator */
+       u8    xor_mult[2];            /**< XOR multipliers for this position,
+                                       *  valid for data disks only */
+       u8    reserved[2];
+};
+
+struct raid_map_data {
+       u32   structure_size;           /* Size of entire structure in bytes */
+       u32   volume_blk_size;          /* bytes / block in the volume */
+       u64   volume_blk_cnt;           /* logical blocks on the volume */
+       u8    phys_blk_shift;           /* Shift factor to convert between
+                                        * units of logical blocks and physical
+                                        * disk blocks */
+       u8    parity_rotation_shift;    /* Shift factor to convert between units
+                                        * of logical stripes and physical
+                                        * stripes */
+       u16   strip_size;               /* blocks used on each disk / stripe */
+       u64   disk_starting_blk;        /* First disk block used in volume */
+       u64   disk_blk_cnt;             /* disk blocks used by volume / disk */
+       u16   data_disks_per_row;       /* data disk entries / row in the map */
+       u16   metadata_disks_per_row;   /* mirror/parity disk entries / row
+                                        * in the map */
+       u16   row_cnt;                  /* rows in each layout map */
+       u16   layout_map_count;         /* layout maps (1 map per mirror/parity
+                                        * group) */
+       u8    reserved[20];
+       struct raid_map_disk_data data[RAID_MAP_MAX_ENTRIES];
+};
+
 struct ReportLUNdata {
        u8 LUNListLength[4];
-       u32 reserved;
+       u8 extended_response_flag;
+       u8 reserved[3];
        u8 LUN[HPSA_MAX_LUN][8];
 };
 
@@ -331,7 +375,7 @@ struct CommandList {
  */
 #define IS_32_BIT ((8 - sizeof(long))/4)
 #define IS_64_BIT (!IS_32_BIT)
-#define PAD_32 (4)
+#define PAD_32 (36)
 #define PAD_64 (4)
 #define COMMANDLIST_PAD (IS_32_BIT * PAD_32 + IS_64_BIT * PAD_64)
        u8 pad[COMMANDLIST_PAD];
@@ -371,6 +415,11 @@ struct io_accel1_cmd {
        struct vals32 host_addr;        /* 0x70 - 0x77 */
        u8  CISS_LUN[8];                /* 0x78 - 0x7F */
        struct SGDescriptor SG[IOACCEL1_MAXSGENTRIES];
+#define IOACCEL1_PAD_64 0
+#define IOACCEL1_PAD_32 0
+#define IOACCEL1_PAD (IS_32_BIT * IOACCEL1_PAD_32 + \
+                       IS_64_BIT * IOACCEL1_PAD_64)
+       u8 pad[IOACCEL1_PAD];
 };
 
 #define IOACCEL1_FUNCTION_SCSIIO        0x00
@@ -407,6 +456,8 @@ struct HostWrite {
 #define MEMQ_MODE       0x08
 #define IOACCEL_MODE_1  0x80
 
+#define DRIVER_SUPPORT_UA_ENABLE        0x00000001
+
 struct CfgTable {
        u8            Signature[4];
        u32             SpecValence;
@@ -435,8 +486,16 @@ struct CfgTable {
        u32             misc_fw_support; /* offset 0x78 */
 #define                        MISC_FW_DOORBELL_RESET (0x02)
 #define                        MISC_FW_DOORBELL_RESET2 (0x010)
+#define                        MISC_FW_RAID_OFFLOAD_BASIC (0x020)
+#define                        MISC_FW_EVENT_NOTIFY (0x080)
        u8              driver_version[32];
-
+       u32             max_cached_write_size;
+       u8              driver_scratchpad[16];
+       u32             max_error_info_length;
+       u32             io_accel_max_embedded_sg_count;
+       u32             io_accel_request_size_offset;
+       u32             event_notify;
+       u32             clear_event_notify;
 };
 
 #define NUM_BLOCKFETCH_ENTRIES 8