Quelle  vmw_pvscsi.c   Sprache: C

/*
 * Linux driver for VMware's para-virtualized SCSI HBA.
 *
 * Copyright (C) 2008-2014, VMware, Inc. All Rights Reserved.
 *
 * This program is free software; you can redistribute it and/or modify it
 * under the terms of the GNU General Public License as published by the
 * Free Software Foundation; version 2 of the License and no later version.
 *
 * This program is distributed in the hope that it will be useful, but
 * WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
 * NON INFRINGEMENT.  See the GNU General Public License for more
 * details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 *
 */

#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/interrupt.h>
#include <linux/slab.h>
#include <linux/workqueue.h>
#include <linux/pci.h>

#include <scsi/scsi.h>
#include <scsi/scsi_host.h>
#include <scsi/scsi_cmnd.h>
#include <scsi/scsi_device.h>
#include <scsi/scsi_tcq.h>

#include "vmw_pvscsi.h"

#define PVSCSI_LINUX_DRIVER_DESC "VMware PVSCSI driver"

MODULE_DESCRIPTION(PVSCSI_LINUX_DRIVER_DESC);
MODULE_AUTHOR("VMware, Inc.");
MODULE_LICENSE("GPL");
MODULE_VERSION(PVSCSI_DRIVER_VERSION_STRING);

#define PVSCSI_DEFAULT_NUM_PAGES_PER_RING 8
#define PVSCSI_DEFAULT_NUM_PAGES_MSG_RING 1
#define PVSCSI_DEFAULT_QUEUE_DEPTH  254
#define SGL_SIZE    PAGE_SIZE

struct pvscsi_sg_list {
 struct PVSCSISGElement sge[PVSCSI_MAX_NUM_SG_ENTRIES_PER_SEGMENT];
};

struct pvscsi_ctx {
 /*
 * The index of the context in cmd_map serves as the context ID for a
 * 1-to-1 mapping completions back to requests.
 */
 struct scsi_cmnd *cmd;
 struct pvscsi_sg_list *sgl;
 struct list_head list;
 dma_addr_t  dataPA;
 dma_addr_t  sensePA;
 dma_addr_t  sglPA;
 struct completion *abort_cmp;
};

struct pvscsi_adapter {
 char    *mmioBase;
 u8    rev;
 bool    use_msg;
 bool    use_req_threshold;

 spinlock_t   hw_lock;

 struct workqueue_struct  *workqueue;
 struct work_struct  work;

 struct PVSCSIRingReqDesc *req_ring;
 unsigned   req_pages;
 unsigned   req_depth;
 dma_addr_t   reqRingPA;

 struct PVSCSIRingCmpDesc *cmp_ring;
 unsigned   cmp_pages;
 dma_addr_t   cmpRingPA;

 struct PVSCSIRingMsgDesc *msg_ring;
 unsigned   msg_pages;
 dma_addr_t   msgRingPA;

 struct PVSCSIRingsState  *rings_state;
 dma_addr_t   ringStatePA;

 struct pci_dev   *dev;
 struct Scsi_Host  *host;

 struct list_head  cmd_pool;
 struct pvscsi_ctx  *cmd_map;
};


/* Command line parameters */
static int pvscsi_ring_pages;
static int pvscsi_msg_ring_pages = PVSCSI_DEFAULT_NUM_PAGES_MSG_RING;
static int pvscsi_cmd_per_lun    = PVSCSI_DEFAULT_QUEUE_DEPTH;
static bool pvscsi_disable_msi;
static bool pvscsi_disable_msix;
static bool pvscsi_use_msg       = true;
static bool pvscsi_use_req_threshold = true;

#define PVSCSI_RW (S_IRUSR | S_IWUSR)

module_param_named(ring_pages, pvscsi_ring_pages, int, PVSCSI_RW);
MODULE_PARM_DESC(ring_pages, "Number of pages per req/cmp ring - (default="
   __stringify(PVSCSI_DEFAULT_NUM_PAGES_PER_RING)
   "[up to 16 targets],"
   __stringify(PVSCSI_SETUP_RINGS_MAX_NUM_PAGES)
   "[for 16+ targets])");

module_param_named(msg_ring_pages, pvscsi_msg_ring_pages, int, PVSCSI_RW);
MODULE_PARM_DESC(msg_ring_pages, "Number of pages for the msg ring - (default="
   __stringify(PVSCSI_DEFAULT_NUM_PAGES_MSG_RING) ")");

module_param_named(cmd_per_lun, pvscsi_cmd_per_lun, int, PVSCSI_RW);
MODULE_PARM_DESC(cmd_per_lun, "Maximum commands per lun - (default="
   __stringify(PVSCSI_DEFAULT_QUEUE_DEPTH) ")");

module_param_named(disable_msi, pvscsi_disable_msi, bool, PVSCSI_RW);
MODULE_PARM_DESC(disable_msi, "Disable MSI use in driver - (default=0)");

module_param_named(disable_msix, pvscsi_disable_msix, bool, PVSCSI_RW);
MODULE_PARM_DESC(disable_msix, "Disable MSI-X use in driver - (default=0)");

module_param_named(use_msg, pvscsi_use_msg, bool, PVSCSI_RW);
MODULE_PARM_DESC(use_msg, "Use msg ring when available - (default=1)");

module_param_named(use_req_threshold, pvscsi_use_req_threshold,
     bool, PVSCSI_RW);
MODULE_PARM_DESC(use_req_threshold, "Use driver-based request coalescing if configured - (default=1)");

static const struct pci_device_id pvscsi_pci_tbl[] = {
 { PCI_VDEVICE(VMWARE, PCI_DEVICE_ID_VMWARE_PVSCSI) },
 { 0 }
};

MODULE_DEVICE_TABLE(pci, pvscsi_pci_tbl);

static struct device *
pvscsi_dev(const struct pvscsi_adapter *adapter)
{
 return &(adapter->dev->dev);
}

static struct pvscsi_ctx *
pvscsi_find_context(const struct pvscsi_adapter *adapter, struct scsi_cmnd *cmd)
{
 struct pvscsi_ctx *ctx, *end;

 end = &adapter->cmd_map[adapter->req_depth];
 for (ctx = adapter->cmd_map; ctx < end; ctx++)
  if (ctx->cmd == cmd)
   return ctx;

 return NULL;
}

static struct pvscsi_ctx *
pvscsi_acquire_context(struct pvscsi_adapter *adapter, struct scsi_cmnd *cmd)
{
 struct pvscsi_ctx *ctx;

 if (list_empty(&adapter->cmd_pool))
  return NULL;

 ctx = list_first_entry(&adapter->cmd_pool, struct pvscsi_ctx, list);
 ctx->cmd = cmd;
 list_del(&ctx->list);

 return ctx;
}

static void pvscsi_release_context(struct pvscsi_adapter *adapter,
       struct pvscsi_ctx *ctx)
{
 ctx->cmd = NULL;
 ctx->abort_cmp = NULL;
 list_add(&ctx->list, &adapter->cmd_pool);
}

/*
 * Map a pvscsi_ctx struct to a context ID field value; we map to a simple
 * non-zero integer. ctx always points to an entry in cmd_map array, hence
 * the return value is always >=1.
 */
static u64 pvscsi_map_context(const struct pvscsi_adapter *adapter,
         const struct pvscsi_ctx *ctx)
{
 return ctx - adapter->cmd_map + 1;
}

static struct pvscsi_ctx *
pvscsi_get_context(const struct pvscsi_adapter *adapter, u64 context)
{
 return &adapter->cmd_map[context - 1];
}

static void pvscsi_reg_write(const struct pvscsi_adapter *adapter,
        u32 offset, u32 val)
{
 writel(val, adapter->mmioBase + offset);
}

static u32 pvscsi_reg_read(const struct pvscsi_adapter *adapter, u32 offset)
{
 return readl(adapter->mmioBase + offset);
}

static u32 pvscsi_read_intr_status(const struct pvscsi_adapter *adapter)
{
 return pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_INTR_STATUS);
}

static void pvscsi_write_intr_status(const struct pvscsi_adapter *adapter,
         u32 val)
{
 pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_INTR_STATUS, val);
}

static void pvscsi_unmask_intr(const struct pvscsi_adapter *adapter)
{
 u32 intr_bits;

 intr_bits = PVSCSI_INTR_CMPL_MASK;
 if (adapter->use_msg)
  intr_bits |= PVSCSI_INTR_MSG_MASK;

 pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_INTR_MASK, intr_bits);
}

static void pvscsi_mask_intr(const struct pvscsi_adapter *adapter)
{
 pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_INTR_MASK, 0);
}

static void pvscsi_write_cmd_desc(const struct pvscsi_adapter *adapter,
      u32 cmd, const void *desc, size_t len)
{
 const u32 *ptr = desc;
 size_t i;

 len /= sizeof(*ptr);
 pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND, cmd);
 for (i = 0; i < len; i++)
  pvscsi_reg_write(adapter,
     PVSCSI_REG_OFFSET_COMMAND_DATA, ptr[i]);
}

static void pvscsi_abort_cmd(const struct pvscsi_adapter *adapter,
        const struct pvscsi_ctx *ctx)
{
 struct PVSCSICmdDescAbortCmd cmd = { 0 };

 cmd.target = ctx->cmd->device->id;
 cmd.context = pvscsi_map_context(adapter, ctx);

 pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_ABORT_CMD, &cmd, sizeof(cmd));
}

static void pvscsi_kick_rw_io(const struct pvscsi_adapter *adapter)
{
 pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_KICK_RW_IO, 0);
}

static void pvscsi_process_request_ring(const struct pvscsi_adapter *adapter)
{
 pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_KICK_NON_RW_IO, 0);
}

static int scsi_is_rw(unsigned char op)
{
 return op == READ_6  || op == WRITE_6 ||
        op == READ_10 || op == WRITE_10 ||
        op == READ_12 || op == WRITE_12 ||
        op == READ_16 || op == WRITE_16;
}

static void pvscsi_kick_io(const struct pvscsi_adapter *adapter,
      unsigned char op)
{
 if (scsi_is_rw(op)) {
  struct PVSCSIRingsState *s = adapter->rings_state;

  if (!adapter->use_req_threshold ||
      s->reqProdIdx - s->reqConsIdx >= s->reqCallThreshold)
   pvscsi_kick_rw_io(adapter);
 } else {
  pvscsi_process_request_ring(adapter);
 }
}

static void ll_adapter_reset(const struct pvscsi_adapter *adapter)
{
 dev_dbg(pvscsi_dev(adapter), "Adapter Reset on %p\n", adapter);

 pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_ADAPTER_RESET, NULL, 0);
}

static void ll_bus_reset(const struct pvscsi_adapter *adapter)
{
 dev_dbg(pvscsi_dev(adapter), "Resetting bus on %p\n", adapter);

 pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_RESET_BUS, NULL, 0);
}

static void ll_device_reset(const struct pvscsi_adapter *adapter, u32 target)
{
 struct PVSCSICmdDescResetDevice cmd = { 0 };

 dev_dbg(pvscsi_dev(adapter), "Resetting device: target=%u\n", target);

 cmd.target = target;

 pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_RESET_DEVICE,
         &cmd, sizeof(cmd));
}

static void pvscsi_create_sg(struct pvscsi_ctx *ctx,
        struct scatterlist *sg, unsigned count)
{
 unsigned i;
 struct PVSCSISGElement *sge;

 BUG_ON(count > PVSCSI_MAX_NUM_SG_ENTRIES_PER_SEGMENT);

 sge = &ctx->sgl->sge[0];
 for (i = 0; i < count; i++, sg = sg_next(sg)) {
  sge[i].addr   = sg_dma_address(sg);
  sge[i].length = sg_dma_len(sg);
  sge[i].flags  = 0;
 }
}

/*
 * Map all data buffers for a command into PCI space and
 * setup the scatter/gather list if needed.
 */
static int pvscsi_map_buffers(struct pvscsi_adapter *adapter,
         struct pvscsi_ctx *ctx, struct scsi_cmnd *cmd,
         struct PVSCSIRingReqDesc *e)
{
 unsigned count;
 unsigned bufflen = scsi_bufflen(cmd);
 struct scatterlist *sg;

 e->dataLen = bufflen;
 e->dataAddr = 0;
 if (bufflen == 0)
  return 0;

 sg = scsi_sglist(cmd);
 count = scsi_sg_count(cmd);
 if (count != 0) {
  int segs = scsi_dma_map(cmd);

  if (segs == -ENOMEM) {
   scmd_printk(KERN_DEBUG, cmd,
        "vmw_pvscsi: Failed to map cmd sglist for DMA.\n");
   return -ENOMEM;
  } else if (segs > 1) {
   pvscsi_create_sg(ctx, sg, segs);

   e->flags |= PVSCSI_FLAG_CMD_WITH_SG_LIST;
   ctx->sglPA = dma_map_single(&adapter->dev->dev,
     ctx->sgl, SGL_SIZE, DMA_TO_DEVICE);
   if (dma_mapping_error(&adapter->dev->dev, ctx->sglPA)) {
    scmd_printk(KERN_ERR, cmd,
         "vmw_pvscsi: Failed to map ctx sglist for DMA.\n");
    scsi_dma_unmap(cmd);
    ctx->sglPA = 0;
    return -ENOMEM;
   }
   e->dataAddr = ctx->sglPA;
  } else
   e->dataAddr = sg_dma_address(sg);
 } else {
  /*
 * In case there is no S/G list, scsi_sglist points
 * directly to the buffer.
 */
  ctx->dataPA = dma_map_single(&adapter->dev->dev, sg, bufflen,
          cmd->sc_data_direction);
  if (dma_mapping_error(&adapter->dev->dev, ctx->dataPA)) {
   scmd_printk(KERN_DEBUG, cmd,
        "vmw_pvscsi: Failed to map direct data buffer for DMA.\n");
   return -ENOMEM;
  }
  e->dataAddr = ctx->dataPA;
 }

 return 0;
}

/*
 * The device incorrectly doesn't clear the first byte of the sense
 * buffer in some cases. We have to do it ourselves.
 * Otherwise we run into trouble when SWIOTLB is forced.
 */
static void pvscsi_patch_sense(struct scsi_cmnd *cmd)
{
 if (cmd->sense_buffer)
  cmd->sense_buffer[0] = 0;
}

static void pvscsi_unmap_buffers(const struct pvscsi_adapter *adapter,
     struct pvscsi_ctx *ctx)
{
 struct scsi_cmnd *cmd;
 unsigned bufflen;

 cmd = ctx->cmd;
 bufflen = scsi_bufflen(cmd);

 if (bufflen != 0) {
  unsigned count = scsi_sg_count(cmd);

  if (count != 0) {
   scsi_dma_unmap(cmd);
   if (ctx->sglPA) {
    dma_unmap_single(&adapter->dev->dev, ctx->sglPA,
       SGL_SIZE, DMA_TO_DEVICE);
    ctx->sglPA = 0;
   }
  } else
   dma_unmap_single(&adapter->dev->dev, ctx->dataPA,
      bufflen, cmd->sc_data_direction);
 }
 if (cmd->sense_buffer)
  dma_unmap_single(&adapter->dev->dev, ctx->sensePA,
     SCSI_SENSE_BUFFERSIZE, DMA_FROM_DEVICE);
}

static int pvscsi_allocate_rings(struct pvscsi_adapter *adapter)
{
 adapter->rings_state = dma_alloc_coherent(&adapter->dev->dev, PAGE_SIZE,
   &adapter->ringStatePA, GFP_KERNEL);
 if (!adapter->rings_state)
  return -ENOMEM;

 adapter->req_pages = min(PVSCSI_MAX_NUM_PAGES_REQ_RING,
     pvscsi_ring_pages);
 adapter->req_depth = adapter->req_pages
     * PVSCSI_MAX_NUM_REQ_ENTRIES_PER_PAGE;
 adapter->req_ring = dma_alloc_coherent(&adapter->dev->dev,
   adapter->req_pages * PAGE_SIZE, &adapter->reqRingPA,
   GFP_KERNEL);
 if (!adapter->req_ring)
  return -ENOMEM;

 adapter->cmp_pages = min(PVSCSI_MAX_NUM_PAGES_CMP_RING,
     pvscsi_ring_pages);
 adapter->cmp_ring = dma_alloc_coherent(&adapter->dev->dev,
   adapter->cmp_pages * PAGE_SIZE, &adapter->cmpRingPA,
   GFP_KERNEL);
 if (!adapter->cmp_ring)
  return -ENOMEM;

 BUG_ON(!IS_ALIGNED(adapter->ringStatePA, PAGE_SIZE));
 BUG_ON(!IS_ALIGNED(adapter->reqRingPA, PAGE_SIZE));
 BUG_ON(!IS_ALIGNED(adapter->cmpRingPA, PAGE_SIZE));

 if (!adapter->use_msg)
  return 0;

 adapter->msg_pages = min(PVSCSI_MAX_NUM_PAGES_MSG_RING,
     pvscsi_msg_ring_pages);
 adapter->msg_ring = dma_alloc_coherent(&adapter->dev->dev,
   adapter->msg_pages * PAGE_SIZE, &adapter->msgRingPA,
   GFP_KERNEL);
 if (!adapter->msg_ring)
  return -ENOMEM;
 BUG_ON(!IS_ALIGNED(adapter->msgRingPA, PAGE_SIZE));

 return 0;
}

static void pvscsi_setup_all_rings(const struct pvscsi_adapter *adapter)
{
 struct PVSCSICmdDescSetupRings cmd = { 0 };
 dma_addr_t base;
 unsigned i;

 cmd.ringsStatePPN   = adapter->ringStatePA >> PAGE_SHIFT;
 cmd.reqRingNumPages = adapter->req_pages;
 cmd.cmpRingNumPages = adapter->cmp_pages;

 base = adapter->reqRingPA;
 for (i = 0; i < adapter->req_pages; i++) {
  cmd.reqRingPPNs[i] = base >> PAGE_SHIFT;
  base += PAGE_SIZE;
 }

 base = adapter->cmpRingPA;
 for (i = 0; i < adapter->cmp_pages; i++) {
  cmd.cmpRingPPNs[i] = base >> PAGE_SHIFT;
  base += PAGE_SIZE;
 }

 memset(adapter->rings_state, 0, PAGE_SIZE);
 memset(adapter->req_ring, 0, adapter->req_pages * PAGE_SIZE);
 memset(adapter->cmp_ring, 0, adapter->cmp_pages * PAGE_SIZE);

 pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_SETUP_RINGS,
         &cmd, sizeof(cmd));

 if (adapter->use_msg) {
  struct PVSCSICmdDescSetupMsgRing cmd_msg = { 0 };

  cmd_msg.numPages = adapter->msg_pages;

  base = adapter->msgRingPA;
  for (i = 0; i < adapter->msg_pages; i++) {
   cmd_msg.ringPPNs[i] = base >> PAGE_SHIFT;
   base += PAGE_SIZE;
  }
  memset(adapter->msg_ring, 0, adapter->msg_pages * PAGE_SIZE);

  pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_SETUP_MSG_RING,
          &cmd_msg, sizeof(cmd_msg));
 }
}

static int pvscsi_change_queue_depth(struct scsi_device *sdev, int qdepth)
{
 if (!sdev->tagged_supported)
  qdepth = 1;
 return scsi_change_queue_depth(sdev, qdepth);
}

/*
 * Pull a completion descriptor off and pass the completion back
 * to the SCSI mid layer.
 */
static void pvscsi_complete_request(struct pvscsi_adapter *adapter,
        const struct PVSCSIRingCmpDesc *e)
{
 struct pvscsi_ctx *ctx;
 struct scsi_cmnd *cmd;
 struct completion *abort_cmp;
 u32 btstat = e->hostStatus;
 u32 sdstat = e->scsiStatus;

 ctx = pvscsi_get_context(adapter, e->context);
 cmd = ctx->cmd;
 abort_cmp = ctx->abort_cmp;
 pvscsi_unmap_buffers(adapter, ctx);
 if (sdstat != SAM_STAT_CHECK_CONDITION)
  pvscsi_patch_sense(cmd);
 pvscsi_release_context(adapter, ctx);
 if (abort_cmp) {
  /*
 * The command was requested to be aborted. Just signal that
 * the request completed and swallow the actual cmd completion
 * here. The abort handler will post a completion for this
 * command indicating that it got successfully aborted.
 */
  complete(abort_cmp);
  return;
 }

 cmd->result = 0;
 if (sdstat != SAM_STAT_GOOD &&
     (btstat == BTSTAT_SUCCESS ||
      btstat == BTSTAT_LINKED_COMMAND_COMPLETED ||
      btstat == BTSTAT_LINKED_COMMAND_COMPLETED_WITH_FLAG)) {
  if (sdstat == SAM_STAT_COMMAND_TERMINATED) {
   cmd->result = (DID_RESET << 16);
  } else {
   cmd->result = (DID_OK << 16) | sdstat;
  }
 } else
  switch (btstat) {
  case BTSTAT_SUCCESS:
  case BTSTAT_LINKED_COMMAND_COMPLETED:
  case BTSTAT_LINKED_COMMAND_COMPLETED_WITH_FLAG:
   /*
 * Commands like INQUIRY may transfer less data than
 * requested by the initiator via bufflen. Set residual
 * count to make upper layer aware of the actual amount
 * of data returned. There are cases when controller
 * returns zero dataLen with non zero data - do not set
 * residual count in that case.
 */
   if (e->dataLen && (e->dataLen < scsi_bufflen(cmd)))
    scsi_set_resid(cmd, scsi_bufflen(cmd) - e->dataLen);
   cmd->result = (DID_OK << 16);
   break;

  case BTSTAT_DATARUN:
  case BTSTAT_DATA_UNDERRUN:
   /* Report residual data in underruns */
   scsi_set_resid(cmd, scsi_bufflen(cmd) - e->dataLen);
   cmd->result = (DID_ERROR << 16);
   break;

  case BTSTAT_SELTIMEO:
   /* Our emulation returns this for non-connected devs */
   cmd->result = (DID_BAD_TARGET << 16);
   break;

  case BTSTAT_LUNMISMATCH:
  case BTSTAT_TAGREJECT:
  case BTSTAT_BADMSG:
  case BTSTAT_HAHARDWARE:
  case BTSTAT_INVPHASE:
  case BTSTAT_HATIMEOUT:
  case BTSTAT_NORESPONSE:
  case BTSTAT_DISCONNECT:
  case BTSTAT_HASOFTWARE:
  case BTSTAT_BUSFREE:
  case BTSTAT_SENSFAILED:
   cmd->result |= (DID_ERROR << 16);
   break;

  case BTSTAT_SENTRST:
  case BTSTAT_RECVRST:
  case BTSTAT_BUSRESET:
   cmd->result = (DID_RESET << 16);
   break;

  case BTSTAT_ABORTQUEUE:
   cmd->result = (DID_BUS_BUSY << 16);
   break;

  case BTSTAT_SCSIPARITY:
   cmd->result = (DID_PARITY << 16);
   break;

  default:
   cmd->result = (DID_ERROR << 16);
   scmd_printk(KERN_DEBUG, cmd,
        "Unknown completion status: 0x%x\n",
        btstat);
 }

 dev_dbg(&cmd->device->sdev_gendev,
  "cmd=%p %x ctx=%p result=0x%x status=0x%x,%x\n",
  cmd, cmd->cmnd[0], ctx, cmd->result, btstat, sdstat);

 scsi_done(cmd);
}

/*
 * barrier usage : Since the PVSCSI device is emulated, there could be cases
 * where we may want to serialize some accesses between the driver and the
 * emulation layer. We use compiler barriers instead of the more expensive
 * memory barriers because PVSCSI is only supported on X86 which has strong
 * memory access ordering.
 */
static void pvscsi_process_completion_ring(struct pvscsi_adapter *adapter)
{
 struct PVSCSIRingsState *s = adapter->rings_state;
 struct PVSCSIRingCmpDesc *ring = adapter->cmp_ring;
 u32 cmp_entries = s->cmpNumEntriesLog2;

 while (s->cmpConsIdx != s->cmpProdIdx) {
  struct PVSCSIRingCmpDesc *e = ring + (s->cmpConsIdx &
            MASK(cmp_entries));
  /*
 * This barrier() ensures that *e is not dereferenced while
 * the device emulation still writes data into the slot.
 * Since the device emulation advances s->cmpProdIdx only after
 * updating the slot we want to check it first.
 */
  barrier();
  pvscsi_complete_request(adapter, e);
  /*
 * This barrier() ensures that compiler doesn't reorder write
 * to s->cmpConsIdx before the read of (*e) inside
 * pvscsi_complete_request. Otherwise, device emulation may
 * overwrite *e before we had a chance to read it.
 */
  barrier();
  s->cmpConsIdx++;
 }
}

/*
 * Translate a Linux SCSI request into a request ring entry.
 */
static int pvscsi_queue_ring(struct pvscsi_adapter *adapter,
        struct pvscsi_ctx *ctx, struct scsi_cmnd *cmd)
{
 struct PVSCSIRingsState *s;
 struct PVSCSIRingReqDesc *e;
 struct scsi_device *sdev;
 u32 req_entries;

 s = adapter->rings_state;
 sdev = cmd->device;
 req_entries = s->reqNumEntriesLog2;

 /*
 * If this condition holds, we might have room on the request ring, but
 * we might not have room on the completion ring for the response.
 * However, we have already ruled out this possibility - we would not
 * have successfully allocated a context if it were true, since we only
 * have one context per request entry.  Check for it anyway, since it
 * would be a serious bug.
 */
 if (s->reqProdIdx - s->cmpConsIdx >= 1 << req_entries) {
  scmd_printk(KERN_ERR, cmd, "vmw_pvscsi: "
       "ring full: reqProdIdx=%d cmpConsIdx=%d\n",
       s->reqProdIdx, s->cmpConsIdx);
  return -1;
 }

 e = adapter->req_ring + (s->reqProdIdx & MASK(req_entries));

 e->bus    = sdev->channel;
 e->target = sdev->id;
 memset(e->lun, 0, sizeof(e->lun));
 e->lun[1] = sdev->lun;

 if (cmd->sense_buffer) {
  ctx->sensePA = dma_map_single(&adapter->dev->dev,
    cmd->sense_buffer, SCSI_SENSE_BUFFERSIZE,
    DMA_FROM_DEVICE);
  if (dma_mapping_error(&adapter->dev->dev, ctx->sensePA)) {
   scmd_printk(KERN_DEBUG, cmd,
        "vmw_pvscsi: Failed to map sense buffer for DMA.\n");
   ctx->sensePA = 0;
   return -ENOMEM;
  }
  e->senseAddr = ctx->sensePA;
  e->senseLen = SCSI_SENSE_BUFFERSIZE;
 } else {
  e->senseLen  = 0;
  e->senseAddr = 0;
 }
 e->cdbLen   = cmd->cmd_len;
 e->vcpuHint = smp_processor_id();
 memcpy(e->cdb, cmd->cmnd, e->cdbLen);

 e->tag = SIMPLE_QUEUE_TAG;

 if (cmd->sc_data_direction == DMA_FROM_DEVICE)
  e->flags = PVSCSI_FLAG_CMD_DIR_TOHOST;
 else if (cmd->sc_data_direction == DMA_TO_DEVICE)
  e->flags = PVSCSI_FLAG_CMD_DIR_TODEVICE;
 else if (cmd->sc_data_direction == DMA_NONE)
  e->flags = PVSCSI_FLAG_CMD_DIR_NONE;
 else
  e->flags = 0;

 if (pvscsi_map_buffers(adapter, ctx, cmd, e) != 0) {
  if (cmd->sense_buffer) {
   dma_unmap_single(&adapter->dev->dev, ctx->sensePA,
      SCSI_SENSE_BUFFERSIZE,
      DMA_FROM_DEVICE);
   ctx->sensePA = 0;
  }
  return -ENOMEM;
 }

 e->context = pvscsi_map_context(adapter, ctx);

 barrier();

 s->reqProdIdx++;

 return 0;
}

static int pvscsi_queue_lck(struct scsi_cmnd *cmd)
{
 struct Scsi_Host *host = cmd->device->host;
 struct pvscsi_adapter *adapter = shost_priv(host);
 struct pvscsi_ctx *ctx;
 unsigned long flags;
 unsigned char op;

 spin_lock_irqsave(&adapter->hw_lock, flags);

 ctx = pvscsi_acquire_context(adapter, cmd);
 if (!ctx || pvscsi_queue_ring(adapter, ctx, cmd) != 0) {
  if (ctx)
   pvscsi_release_context(adapter, ctx);
  spin_unlock_irqrestore(&adapter->hw_lock, flags);
  return SCSI_MLQUEUE_HOST_BUSY;
 }

 op = cmd->cmnd[0];

 dev_dbg(&cmd->device->sdev_gendev,
  "queued cmd %p, ctx %p, op=%x\n", cmd, ctx, op);

 spin_unlock_irqrestore(&adapter->hw_lock, flags);

 pvscsi_kick_io(adapter, op);

 return 0;
}

static DEF_SCSI_QCMD(pvscsi_queue)

static int pvscsi_abort(struct scsi_cmnd *cmd)
{
 struct pvscsi_adapter *adapter = shost_priv(cmd->device->host);
 struct pvscsi_ctx *ctx;
 unsigned long flags;
 int result = SUCCESS;
 DECLARE_COMPLETION_ONSTACK(abort_cmp);
 int done;

 scmd_printk(KERN_DEBUG, cmd, "task abort on host %u, %p\n",
      adapter->host->host_no, cmd);

 spin_lock_irqsave(&adapter->hw_lock, flags);

 /*
 * Poll the completion ring first - we might be trying to abort
 * a command that is waiting to be dispatched in the completion ring.
 */
 pvscsi_process_completion_ring(adapter);

 /*
 * If there is no context for the command, it either already succeeded
 * or else was never properly issued.  Not our problem.
 */
 ctx = pvscsi_find_context(adapter, cmd);
 if (!ctx) {
  scmd_printk(KERN_DEBUG, cmd, "Failed to abort cmd %p\n", cmd);
  goto out;
 }

 /*
 * Mark that the command has been requested to be aborted and issue
 * the abort.
 */
 ctx->abort_cmp = &abort_cmp;

 pvscsi_abort_cmd(adapter, ctx);
 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 /* Wait for 2 secs for the completion. */
 done = wait_for_completion_timeout(&abort_cmp, msecs_to_jiffies(2000));
 spin_lock_irqsave(&adapter->hw_lock, flags);

 if (!done) {
  /*
 * Failed to abort the command, unmark the fact that it
 * was requested to be aborted.
 */
  ctx->abort_cmp = NULL;
  result = FAILED;
  scmd_printk(KERN_DEBUG, cmd,
       "Failed to get completion for aborted cmd %p\n",
       cmd);
  goto out;
 }

 /*
 * Successfully aborted the command.
 */
 cmd->result = (DID_ABORT << 16);
 scsi_done(cmd);

out:
 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 return result;
}

/*
 * Abort all outstanding requests.  This is only safe to use if the completion
 * ring will never be walked again or the device has been reset, because it
 * destroys the 1-1 mapping between context field passed to emulation and our
 * request structure.
 */
static void pvscsi_reset_all(struct pvscsi_adapter *adapter)
{
 unsigned i;

 for (i = 0; i < adapter->req_depth; i++) {
  struct pvscsi_ctx *ctx = &adapter->cmd_map[i];
  struct scsi_cmnd *cmd = ctx->cmd;
  if (cmd) {
   scmd_printk(KERN_ERR, cmd,
        "Forced reset on cmd %p\n", cmd);
   pvscsi_unmap_buffers(adapter, ctx);
   pvscsi_patch_sense(cmd);
   pvscsi_release_context(adapter, ctx);
   cmd->result = (DID_RESET << 16);
   scsi_done(cmd);
  }
 }
}

static int pvscsi_host_reset(struct scsi_cmnd *cmd)
{
 struct Scsi_Host *host = cmd->device->host;
 struct pvscsi_adapter *adapter = shost_priv(host);
 unsigned long flags;
 bool use_msg;

 scmd_printk(KERN_INFO, cmd, "SCSI Host reset\n");

 spin_lock_irqsave(&adapter->hw_lock, flags);

 use_msg = adapter->use_msg;

 if (use_msg) {
  adapter->use_msg = false;
  spin_unlock_irqrestore(&adapter->hw_lock, flags);

  /*
 * Now that we know that the ISR won't add more work on the
 * workqueue we can safely flush any outstanding work.
 */
  flush_workqueue(adapter->workqueue);
  spin_lock_irqsave(&adapter->hw_lock, flags);
 }

 /*
 * We're going to tear down the entire ring structure and set it back
 * up, so stalling new requests until all completions are flushed and
 * the rings are back in place.
 */

 pvscsi_process_request_ring(adapter);

 ll_adapter_reset(adapter);

 /*
 * Now process any completions.  Note we do this AFTER adapter reset,
 * which is strange, but stops races where completions get posted
 * between processing the ring and issuing the reset.  The backend will
 * not touch the ring memory after reset, so the immediately pre-reset
 * completion ring state is still valid.
 */
 pvscsi_process_completion_ring(adapter);

 pvscsi_reset_all(adapter);
 adapter->use_msg = use_msg;
 pvscsi_setup_all_rings(adapter);
 pvscsi_unmask_intr(adapter);

 spin_unlock_irqrestore(&adapter->hw_lock, flags);

 return SUCCESS;
}

static int pvscsi_bus_reset(struct scsi_cmnd *cmd)
{
 struct Scsi_Host *host = cmd->device->host;
 struct pvscsi_adapter *adapter = shost_priv(host);
 unsigned long flags;

 scmd_printk(KERN_INFO, cmd, "SCSI Bus reset\n");

 /*
 * We don't want to queue new requests for this bus after
 * flushing all pending requests to emulation, since new
 * requests could then sneak in during this bus reset phase,
 * so take the lock now.
 */
 spin_lock_irqsave(&adapter->hw_lock, flags);

 pvscsi_process_request_ring(adapter);
 ll_bus_reset(adapter);
 pvscsi_process_completion_ring(adapter);

 spin_unlock_irqrestore(&adapter->hw_lock, flags);

 return SUCCESS;
}

static int pvscsi_device_reset(struct scsi_cmnd *cmd)
{
 struct Scsi_Host *host = cmd->device->host;
 struct pvscsi_adapter *adapter = shost_priv(host);
 unsigned long flags;

 scmd_printk(KERN_INFO, cmd, "SCSI device reset on scsi%u:%u\n",
      host->host_no, cmd->device->id);

 /*
 * We don't want to queue new requests for this device after flushing
 * all pending requests to emulation, since new requests could then
 * sneak in during this device reset phase, so take the lock now.
 */
 spin_lock_irqsave(&adapter->hw_lock, flags);

 pvscsi_process_request_ring(adapter);
 ll_device_reset(adapter, cmd->device->id);
 pvscsi_process_completion_ring(adapter);

 spin_unlock_irqrestore(&adapter->hw_lock, flags);

 return SUCCESS;
}

static struct scsi_host_template pvscsi_template;

static const char *pvscsi_info(struct Scsi_Host *host)
{
 struct pvscsi_adapter *adapter = shost_priv(host);
 static char buf[256];

 sprintf(buf, "VMware PVSCSI storage adapter rev %d, req/cmp/msg rings: "
  "%u/%u/%u pages, cmd_per_lun=%u", adapter->rev,
  adapter->req_pages, adapter->cmp_pages, adapter->msg_pages,
  pvscsi_template.cmd_per_lun);

 return buf;
}

static struct scsi_host_template pvscsi_template = {
 .module    = THIS_MODULE,
 .name    = "VMware PVSCSI Host Adapter",
 .proc_name   = "vmw_pvscsi",
 .info    = pvscsi_info,
 .queuecommand   = pvscsi_queue,
 .this_id   = -1,
 .sg_tablesize   = PVSCSI_MAX_NUM_SG_ENTRIES_PER_SEGMENT,
 .dma_boundary   = UINT_MAX,
 .max_sectors   = 0xffff,
 .change_queue_depth  = pvscsi_change_queue_depth,
 .eh_abort_handler  = pvscsi_abort,
 .eh_device_reset_handler = pvscsi_device_reset,
 .eh_bus_reset_handler  = pvscsi_bus_reset,
 .eh_host_reset_handler  = pvscsi_host_reset,
};

static void pvscsi_process_msg(const struct pvscsi_adapter *adapter,
          const struct PVSCSIRingMsgDesc *e)
{
 struct PVSCSIRingsState *s = adapter->rings_state;
 struct Scsi_Host *host = adapter->host;
 struct scsi_device *sdev;

 printk(KERN_INFO "vmw_pvscsi: msg type: 0x%x - MSG RING: %u/%u (%u) \n",
        e->type, s->msgProdIdx, s->msgConsIdx, s->msgNumEntriesLog2);

 BUILD_BUG_ON(PVSCSI_MSG_LAST != 2);

 if (e->type == PVSCSI_MSG_DEV_ADDED) {
  struct PVSCSIMsgDescDevStatusChanged *desc;
  desc = (struct PVSCSIMsgDescDevStatusChanged *)e;

  printk(KERN_INFO
         "vmw_pvscsi: msg: device added at scsi%u:%u:%u\n",
         desc->bus, desc->target, desc->lun[1]);

  if (!scsi_host_get(host))
   return;

  sdev = scsi_device_lookup(host, desc->bus, desc->target,
       desc->lun[1]);
  if (sdev) {
   printk(KERN_INFO "vmw_pvscsi: device already exists\n");
   scsi_device_put(sdev);
  } else
   scsi_add_device(adapter->host, desc->bus,
     desc->target, desc->lun[1]);

  scsi_host_put(host);
 } else if (e->type == PVSCSI_MSG_DEV_REMOVED) {
  struct PVSCSIMsgDescDevStatusChanged *desc;
  desc = (struct PVSCSIMsgDescDevStatusChanged *)e;

  printk(KERN_INFO
         "vmw_pvscsi: msg: device removed at scsi%u:%u:%u\n",
         desc->bus, desc->target, desc->lun[1]);

  if (!scsi_host_get(host))
   return;

  sdev = scsi_device_lookup(host, desc->bus, desc->target,
       desc->lun[1]);
  if (sdev) {
   scsi_remove_device(sdev);
   scsi_device_put(sdev);
  } else
   printk(KERN_INFO
          "vmw_pvscsi: failed to lookup scsi%u:%u:%u\n",
          desc->bus, desc->target, desc->lun[1]);

  scsi_host_put(host);
 }
}

static int pvscsi_msg_pending(const struct pvscsi_adapter *adapter)
{
 struct PVSCSIRingsState *s = adapter->rings_state;

 return s->msgProdIdx != s->msgConsIdx;
}

static void pvscsi_process_msg_ring(const struct pvscsi_adapter *adapter)
{
 struct PVSCSIRingsState *s = adapter->rings_state;
 struct PVSCSIRingMsgDesc *ring = adapter->msg_ring;
 u32 msg_entries = s->msgNumEntriesLog2;

 while (pvscsi_msg_pending(adapter)) {
  struct PVSCSIRingMsgDesc *e = ring + (s->msgConsIdx &
            MASK(msg_entries));

  barrier();
  pvscsi_process_msg(adapter, e);
  barrier();
  s->msgConsIdx++;
 }
}

static void pvscsi_msg_workqueue_handler(struct work_struct *data)
{
 struct pvscsi_adapter *adapter;

 adapter = container_of(data, struct pvscsi_adapter, work);

 pvscsi_process_msg_ring(adapter);
}

static int pvscsi_setup_msg_workqueue(struct pvscsi_adapter *adapter)
{
 char name[32];

 if (!pvscsi_use_msg)
  return 0;

 pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
    PVSCSI_CMD_SETUP_MSG_RING);

 if (pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS) == -1)
  return 0;

 snprintf(name, sizeof(name),
   "vmw_pvscsi_wq_%u", adapter->host->host_no);

 adapter->workqueue =
  alloc_ordered_workqueue("%s", WQ_MEM_RECLAIM, name);
 if (!adapter->workqueue) {
  printk(KERN_ERR "vmw_pvscsi: failed to create work queue\n");
  return 0;
 }
 INIT_WORK(&adapter->work, pvscsi_msg_workqueue_handler);

 return 1;
}

static bool pvscsi_setup_req_threshold(struct pvscsi_adapter *adapter,
          bool enable)
{
 u32 val;

 if (!pvscsi_use_req_threshold)
  return false;

 pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
    PVSCSI_CMD_SETUP_REQCALLTHRESHOLD);
 val = pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS);
 if (val == -1) {
  printk(KERN_INFO "vmw_pvscsi: device does not support req_threshold\n");
  return false;
 } else {
  struct PVSCSICmdDescSetupReqCall cmd_msg = { 0 };
  cmd_msg.enable = enable;
  printk(KERN_INFO
         "vmw_pvscsi: %sabling reqCallThreshold\n",
   enable ? "en" : "dis");
  pvscsi_write_cmd_desc(adapter,
          PVSCSI_CMD_SETUP_REQCALLTHRESHOLD,
          &cmd_msg, sizeof(cmd_msg));
  return pvscsi_reg_read(adapter,
           PVSCSI_REG_OFFSET_COMMAND_STATUS) != 0;
 }
}

static irqreturn_t pvscsi_isr(int irq, void *devp)
{
 struct pvscsi_adapter *adapter = devp;
 unsigned long flags;

 spin_lock_irqsave(&adapter->hw_lock, flags);
 pvscsi_process_completion_ring(adapter);
 if (adapter->use_msg && pvscsi_msg_pending(adapter))
  queue_work(adapter->workqueue, &adapter->work);
 spin_unlock_irqrestore(&adapter->hw_lock, flags);

 return IRQ_HANDLED;
}

static irqreturn_t pvscsi_shared_isr(int irq, void *devp)
{
 struct pvscsi_adapter *adapter = devp;
 u32 val = pvscsi_read_intr_status(adapter);

 if (!(val & PVSCSI_INTR_ALL_SUPPORTED))
  return IRQ_NONE;
 pvscsi_write_intr_status(devp, val);
 return pvscsi_isr(irq, devp);
}

static void pvscsi_free_sgls(const struct pvscsi_adapter *adapter)
{
 struct pvscsi_ctx *ctx = adapter->cmd_map;
 unsigned i;

 for (i = 0; i < adapter->req_depth; ++i, ++ctx)
  free_pages((unsigned long)ctx->sgl, get_order(SGL_SIZE));
}

static void pvscsi_shutdown_intr(struct pvscsi_adapter *adapter)
{
 free_irq(pci_irq_vector(adapter->dev, 0), adapter);
 pci_free_irq_vectors(adapter->dev);
}

static void pvscsi_release_resources(struct pvscsi_adapter *adapter)
{
 if (adapter->workqueue)
  destroy_workqueue(adapter->workqueue);

 if (adapter->mmioBase)
  pci_iounmap(adapter->dev, adapter->mmioBase);

 pci_release_regions(adapter->dev);

 if (adapter->cmd_map) {
  pvscsi_free_sgls(adapter);
  kfree(adapter->cmd_map);
 }

 if (adapter->rings_state)
  dma_free_coherent(&adapter->dev->dev, PAGE_SIZE,
        adapter->rings_state, adapter->ringStatePA);

 if (adapter->req_ring)
  dma_free_coherent(&adapter->dev->dev,
        adapter->req_pages * PAGE_SIZE,
        adapter->req_ring, adapter->reqRingPA);

 if (adapter->cmp_ring)
  dma_free_coherent(&adapter->dev->dev,
        adapter->cmp_pages * PAGE_SIZE,
        adapter->cmp_ring, adapter->cmpRingPA);

 if (adapter->msg_ring)
  dma_free_coherent(&adapter->dev->dev,
        adapter->msg_pages * PAGE_SIZE,
        adapter->msg_ring, adapter->msgRingPA);
}

/*
 * Allocate scatter gather lists.
 *
 * These are statically allocated.  Trying to be clever was not worth it.
 *
 * Dynamic allocation can fail, and we can't go deep into the memory
 * allocator, since we're a SCSI driver, and trying too hard to allocate
 * memory might generate disk I/O.  We also don't want to fail disk I/O
 * in that case because we can't get an allocation - the I/O could be
 * trying to swap out data to free memory.  Since that is pathological,
 * just use a statically allocated scatter list.
 *
 */
static int pvscsi_allocate_sg(struct pvscsi_adapter *adapter)
{
 struct pvscsi_ctx *ctx;
 int i;

 ctx = adapter->cmd_map;
 BUILD_BUG_ON(sizeof(struct pvscsi_sg_list) > SGL_SIZE);

 for (i = 0; i < adapter->req_depth; ++i, ++ctx) {
  ctx->sgl = (void *)__get_free_pages(GFP_KERNEL,
          get_order(SGL_SIZE));
  ctx->sglPA = 0;
  BUG_ON(!IS_ALIGNED(((unsigned long)ctx->sgl), PAGE_SIZE));
  if (!ctx->sgl) {
   for (; i >= 0; --i, --ctx) {
    free_pages((unsigned long)ctx->sgl,
        get_order(SGL_SIZE));
    ctx->sgl = NULL;
   }
   return -ENOMEM;
  }
 }

 return 0;
}

/*
 * Query the device, fetch the config info and return the
 * maximum number of targets on the adapter. In case of
 * failure due to any reason return default i.e. 16.
 */
static u32 pvscsi_get_max_targets(struct pvscsi_adapter *adapter)
{
 struct PVSCSICmdDescConfigCmd cmd;
 struct PVSCSIConfigPageHeader *header;
 struct device *dev;
 dma_addr_t configPagePA;
 void *config_page;
 u32 numPhys = 16;

 dev = pvscsi_dev(adapter);
 config_page = dma_alloc_coherent(&adapter->dev->dev, PAGE_SIZE,
   &configPagePA, GFP_KERNEL);
 if (!config_page) {
  dev_warn(dev, "vmw_pvscsi: failed to allocate memory for config page\n");
  goto exit;
 }
 BUG_ON(configPagePA & ~PAGE_MASK);

 /* Fetch config info from the device. */
 cmd.configPageAddress = ((u64)PVSCSI_CONFIG_CONTROLLER_ADDRESS) << 32;
 cmd.configPageNum = PVSCSI_CONFIG_PAGE_CONTROLLER;
 cmd.cmpAddr = configPagePA;
 cmd._pad = 0;

 /*
 * Mark the completion page header with error values. If the device
 * completes the command successfully, it sets the status values to
 * indicate success.
 */
 header = config_page;
 header->hostStatus = BTSTAT_INVPARAM;
 header->scsiStatus = SDSTAT_CHECK;

 pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_CONFIG, &cmd, sizeof cmd);

 if (header->hostStatus == BTSTAT_SUCCESS &&
     header->scsiStatus == SDSTAT_GOOD) {
  struct PVSCSIConfigPageController *config;

  config = config_page;
  numPhys = config->numPhys;
 } else
  dev_warn(dev, "vmw_pvscsi: PVSCSI_CMD_CONFIG failed. hostStatus = 0x%x, scsiStatus = 0x%x\n",
    header->hostStatus, header->scsiStatus);
 dma_free_coherent(&adapter->dev->dev, PAGE_SIZE, config_page,
     configPagePA);
exit:
 return numPhys;
}

static int pvscsi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
{
 unsigned int irq_flag = PCI_IRQ_ALL_TYPES;
 struct pvscsi_adapter *adapter;
 struct pvscsi_adapter adapter_temp;
 struct Scsi_Host *host = NULL;
 unsigned int i;
 int error;
 u32 max_id;

 error = -ENODEV;

 if (pci_enable_device(pdev))
  return error;

 if (!dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64))) {
  printk(KERN_INFO "vmw_pvscsi: using 64bit dma\n");
 } else if (!dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32))) {
  printk(KERN_INFO "vmw_pvscsi: using 32bit dma\n");
 } else {
  printk(KERN_ERR "vmw_pvscsi: failed to set DMA mask\n");
  goto out_disable_device;
 }

 /*
 * Let's use a temp pvscsi_adapter struct until we find the number of
 * targets on the adapter, after that we will switch to the real
 * allocated struct.
 */
 adapter = &adapter_temp;
 memset(adapter, 0, sizeof(*adapter));
 adapter->dev  = pdev;
 adapter->rev = pdev->revision;

 if (pci_request_regions(pdev, "vmw_pvscsi")) {
  printk(KERN_ERR "vmw_pvscsi: pci memory selection failed\n");
  goto out_disable_device;
 }

 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
  if ((pci_resource_flags(pdev, i) & PCI_BASE_ADDRESS_SPACE_IO))
   continue;

  if (pci_resource_len(pdev, i) < PVSCSI_MEM_SPACE_SIZE)
   continue;

  break;
 }

 if (i == DEVICE_COUNT_RESOURCE) {
  printk(KERN_ERR
         "vmw_pvscsi: adapter has no suitable MMIO region\n");
  goto out_release_resources_and_disable;
 }

 adapter->mmioBase = pci_iomap(pdev, i, PVSCSI_MEM_SPACE_SIZE);

 if (!adapter->mmioBase) {
  printk(KERN_ERR
         "vmw_pvscsi: can't iomap for BAR %d memsize %lu\n",
         i, PVSCSI_MEM_SPACE_SIZE);
  goto out_release_resources_and_disable;
 }

 pci_set_master(pdev);

 /*
 * Ask the device for max number of targets before deciding the
 * default pvscsi_ring_pages value.
 */
 max_id = pvscsi_get_max_targets(adapter);
 printk(KERN_INFO "vmw_pvscsi: max_id: %u\n", max_id);

 if (pvscsi_ring_pages == 0)
  /*
 * Set the right default value. Up to 16 it is 8, above it is
 * max.
 */
  pvscsi_ring_pages = (max_id > 16) ?
   PVSCSI_SETUP_RINGS_MAX_NUM_PAGES :
   PVSCSI_DEFAULT_NUM_PAGES_PER_RING;
 printk(KERN_INFO
        "vmw_pvscsi: setting ring_pages to %d\n",
        pvscsi_ring_pages);

 pvscsi_template.can_queue =
  min(PVSCSI_MAX_NUM_PAGES_REQ_RING, pvscsi_ring_pages) *
  PVSCSI_MAX_NUM_REQ_ENTRIES_PER_PAGE;
 pvscsi_template.cmd_per_lun =
  min(pvscsi_template.can_queue, pvscsi_cmd_per_lun);
 host = scsi_host_alloc(&pvscsi_template, sizeof(struct pvscsi_adapter));
 if (!host) {
  printk(KERN_ERR "vmw_pvscsi: failed to allocate host\n");
  goto out_release_resources_and_disable;
 }

 /*
 * Let's use the real pvscsi_adapter struct here onwards.
 */
 adapter = shost_priv(host);
 memset(adapter, 0, sizeof(*adapter));
 adapter->dev  = pdev;
 adapter->host = host;
 /*
 * Copy back what we already have to the allocated adapter struct.
 */
 adapter->rev = adapter_temp.rev;
 adapter->mmioBase = adapter_temp.mmioBase;

 spin_lock_init(&adapter->hw_lock);
 host->max_channel = 0;
 host->max_lun     = 1;
 host->max_cmd_len = 16;
 host->max_id      = max_id;

 pci_set_drvdata(pdev, host);

 ll_adapter_reset(adapter);

 adapter->use_msg = pvscsi_setup_msg_workqueue(adapter);

 error = pvscsi_allocate_rings(adapter);
 if (error) {
  printk(KERN_ERR "vmw_pvscsi: unable to allocate ring memory\n");
  goto out_release_resources;
 }

 /*
 * From this point on we should reset the adapter if anything goes
 * wrong.
 */
 pvscsi_setup_all_rings(adapter);

 adapter->cmd_map = kcalloc(adapter->req_depth,
       sizeof(struct pvscsi_ctx), GFP_KERNEL);
 if (!adapter->cmd_map) {
  printk(KERN_ERR "vmw_pvscsi: failed to allocate memory.\n");
  error = -ENOMEM;
  goto out_reset_adapter;
 }

 INIT_LIST_HEAD(&adapter->cmd_pool);
 for (i = 0; i < adapter->req_depth; i++) {
  struct pvscsi_ctx *ctx = adapter->cmd_map + i;
  list_add(&ctx->list, &adapter->cmd_pool);
 }

 error = pvscsi_allocate_sg(adapter);
 if (error) {
  printk(KERN_ERR "vmw_pvscsi: unable to allocate s/g table\n");
  goto out_reset_adapter;
 }

 if (pvscsi_disable_msix)
  irq_flag &= ~PCI_IRQ_MSIX;
 if (pvscsi_disable_msi)
  irq_flag &= ~PCI_IRQ_MSI;

 error = pci_alloc_irq_vectors(adapter->dev, 1, 1, irq_flag);
 if (error < 0)
  goto out_reset_adapter;

 adapter->use_req_threshold = pvscsi_setup_req_threshold(adapter, true);
 printk(KERN_DEBUG "vmw_pvscsi: driver-based request coalescing %sabled\n",
        adapter->use_req_threshold ? "en" : "dis");

 if (adapter->dev->msix_enabled || adapter->dev->msi_enabled) {
  printk(KERN_INFO "vmw_pvscsi: using MSI%s\n",
   adapter->dev->msix_enabled ? "-X" : "");
  error = request_irq(pci_irq_vector(pdev, 0), pvscsi_isr,
    0, "vmw_pvscsi", adapter);
 } else {
  printk(KERN_INFO "vmw_pvscsi: using INTx\n");
  error = request_irq(pci_irq_vector(pdev, 0), pvscsi_shared_isr,
    IRQF_SHARED, "vmw_pvscsi", adapter);
 }

 if (error) {
  printk(KERN_ERR
         "vmw_pvscsi: unable to request IRQ: %d\n", error);
  goto out_reset_adapter;
 }

 error = scsi_add_host(host, &pdev->dev);
 if (error) {
  printk(KERN_ERR
         "vmw_pvscsi: scsi_add_host failed: %d\n", error);
  goto out_reset_adapter;
 }

 dev_info(&pdev->dev, "VMware PVSCSI rev %d host #%u\n",
   adapter->rev, host->host_no);

 pvscsi_unmask_intr(adapter);

 scsi_scan_host(host);

 return 0;

out_reset_adapter:
 ll_adapter_reset(adapter);
out_release_resources:
 pvscsi_shutdown_intr(adapter);
 pvscsi_release_resources(adapter);
 scsi_host_put(host);
out_disable_device:
 pci_disable_device(pdev);

 return error;

out_release_resources_and_disable:
 pvscsi_shutdown_intr(adapter);
 pvscsi_release_resources(adapter);
 goto out_disable_device;
}

static void __pvscsi_shutdown(struct pvscsi_adapter *adapter)
{
 pvscsi_mask_intr(adapter);

 if (adapter->workqueue)
  flush_workqueue(adapter->workqueue);

 pvscsi_shutdown_intr(adapter);

 pvscsi_process_request_ring(adapter);
 pvscsi_process_completion_ring(adapter);
 ll_adapter_reset(adapter);
}

static void pvscsi_shutdown(struct pci_dev *dev)
{
 struct Scsi_Host *host = pci_get_drvdata(dev);
 struct pvscsi_adapter *adapter = shost_priv(host);

 __pvscsi_shutdown(adapter);
}

static void pvscsi_remove(struct pci_dev *pdev)
{
 struct Scsi_Host *host = pci_get_drvdata(pdev);
 struct pvscsi_adapter *adapter = shost_priv(host);

 scsi_remove_host(host);

 __pvscsi_shutdown(adapter);
 pvscsi_release_resources(adapter);

 scsi_host_put(host);

 pci_disable_device(pdev);
}

static struct pci_driver pvscsi_pci_driver = {
 .name  = "vmw_pvscsi",
 .id_table = pvscsi_pci_tbl,
 .probe  = pvscsi_probe,
 .remove  = pvscsi_remove,
 .shutdown       = pvscsi_shutdown,
};

static int __init pvscsi_init(void)
{
 pr_info("%s - version %s\n",
  PVSCSI_LINUX_DRIVER_DESC, PVSCSI_DRIVER_VERSION_STRING);
 return pci_register_driver(&pvscsi_pci_driver);
}

static void __exit pvscsi_exit(void)
{
 pci_unregister_driver(&pvscsi_pci_driver);
}

module_init(pvscsi_init);
module_exit(pvscsi_exit);