drivers/scsi/vmw_pvscsi.c

   1 /*
   2  * Linux driver for VMware's para-virtualized SCSI HBA.
   3  *
   4  * Copyright (C) 2008-2014, VMware, Inc. All Rights Reserved.
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License as published by the
   8  * Free Software Foundation; version 2 of the License and no later version.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  13  * NON INFRINGEMENT.  See the GNU General Public License for more
  14  * details.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * along with this program; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Maintained by: Jim Gill <jgill@vmware.com>
  21  *
  22  */
  23
  24 #include <linux/kernel.h>
  25 #include <linux/module.h>
  26 #include <linux/interrupt.h>
  27 #include <linux/slab.h>
  28 #include <linux/workqueue.h>
  29 #include <linux/pci.h>
  30
  31 #include <scsi/scsi.h>
  32 #include <scsi/scsi_host.h>
  33 #include <scsi/scsi_cmnd.h>
  34 #include <scsi/scsi_device.h>
  35 #include <scsi/scsi_tcq.h>
  36
  37 #include "vmw_pvscsi.h"
  38
  39 #define PVSCSI_LINUX_DRIVER_DESC "VMware PVSCSI driver"
  40
  41 MODULE_DESCRIPTION(PVSCSI_LINUX_DRIVER_DESC);
  42 MODULE_AUTHOR("VMware, Inc.");
  43 MODULE_LICENSE("GPL");
  44 MODULE_VERSION(PVSCSI_DRIVER_VERSION_STRING);
  45
  46 #define PVSCSI_DEFAULT_NUM_PAGES_PER_RING       8
  47 #define PVSCSI_DEFAULT_NUM_PAGES_MSG_RING       1
  48 #define PVSCSI_DEFAULT_QUEUE_DEPTH              254
  49 #define SGL_SIZE                                PAGE_SIZE
  50
  51 struct pvscsi_sg_list {
  52         struct PVSCSISGElement sge[PVSCSI_MAX_NUM_SG_ENTRIES_PER_SEGMENT];
  53 };
  54
  55 struct pvscsi_ctx {
  56         /*
  57          * The index of the context in cmd_map serves as the context ID for a
  58          * 1-to-1 mapping completions back to requests.
  59          */
  60         struct scsi_cmnd        *cmd;
  61         struct pvscsi_sg_list   *sgl;
  62         struct list_head        list;
  63         dma_addr_t              dataPA;
  64         dma_addr_t              sensePA;
  65         dma_addr_t              sglPA;
  66         struct completion       *abort_cmp;
  67 };
  68
  69 struct pvscsi_adapter {
  70         char                            *mmioBase;
  71         unsigned int                    irq;
  72         u8                              rev;
  73         bool                            use_msi;
  74         bool                            use_msix;
  75         bool                            use_msg;
  76         bool                            use_req_threshold;
  77
  78         spinlock_t                      hw_lock;
  79
  80         struct workqueue_struct         *workqueue;
  81         struct work_struct              work;
  82
  83         struct PVSCSIRingReqDesc        *req_ring;
  84         unsigned                        req_pages;
  85         unsigned                        req_depth;
  86         dma_addr_t                      reqRingPA;
  87
  88         struct PVSCSIRingCmpDesc        *cmp_ring;
  89         unsigned                        cmp_pages;
  90         dma_addr_t                      cmpRingPA;
  91
  92         struct PVSCSIRingMsgDesc        *msg_ring;
  93         unsigned                        msg_pages;
  94         dma_addr_t                      msgRingPA;
  95
  96         struct PVSCSIRingsState         *rings_state;
  97         dma_addr_t                      ringStatePA;
  98
  99         struct pci_dev                  *dev;
 100         struct Scsi_Host                *host;
 101
 102         struct list_head                cmd_pool;
 103         struct pvscsi_ctx               *cmd_map;
 104 };
 105
 106
 107 /* Command line parameters */
 108 static int pvscsi_ring_pages;
 109 static int pvscsi_msg_ring_pages = PVSCSI_DEFAULT_NUM_PAGES_MSG_RING;
 110 static int pvscsi_cmd_per_lun    = PVSCSI_DEFAULT_QUEUE_DEPTH;
 111 static bool pvscsi_disable_msi;
 112 static bool pvscsi_disable_msix;
 113 static bool pvscsi_use_msg       = true;
 114 static bool pvscsi_use_req_threshold = true;
 115
 116 #define PVSCSI_RW (S_IRUSR | S_IWUSR)
 117
 118 module_param_named(ring_pages, pvscsi_ring_pages, int, PVSCSI_RW);
 119 MODULE_PARM_DESC(ring_pages, "Number of pages per req/cmp ring - (default="
 120                  __stringify(PVSCSI_DEFAULT_NUM_PAGES_PER_RING)
 121                  "[up to 16 targets],"
 122                  __stringify(PVSCSI_SETUP_RINGS_MAX_NUM_PAGES)
 123                  "[for 16+ targets])");
 124
 125 module_param_named(msg_ring_pages, pvscsi_msg_ring_pages, int, PVSCSI_RW);
 126 MODULE_PARM_DESC(msg_ring_pages, "Number of pages for the msg ring - (default="
 127                  __stringify(PVSCSI_DEFAULT_NUM_PAGES_MSG_RING) ")");
 128
 129 module_param_named(cmd_per_lun, pvscsi_cmd_per_lun, int, PVSCSI_RW);
 130 MODULE_PARM_DESC(cmd_per_lun, "Maximum commands per lun - (default="
 131                  __stringify(PVSCSI_DEFAULT_QUEUE_DEPTH) ")");
 132
 133 module_param_named(disable_msi, pvscsi_disable_msi, bool, PVSCSI_RW);
 134 MODULE_PARM_DESC(disable_msi, "Disable MSI use in driver - (default=0)");
 135
 136 module_param_named(disable_msix, pvscsi_disable_msix, bool, PVSCSI_RW);
 137 MODULE_PARM_DESC(disable_msix, "Disable MSI-X use in driver - (default=0)");
 138
 139 module_param_named(use_msg, pvscsi_use_msg, bool, PVSCSI_RW);
 140 MODULE_PARM_DESC(use_msg, "Use msg ring when available - (default=1)");
 141
 142 module_param_named(use_req_threshold, pvscsi_use_req_threshold,
 143                    bool, PVSCSI_RW);
 144 MODULE_PARM_DESC(use_req_threshold, "Use driver-based request coalescing if configured - (default=1)");
 145
 146 static const struct pci_device_id pvscsi_pci_tbl[] = {
 147         { PCI_VDEVICE(VMWARE, PCI_DEVICE_ID_VMWARE_PVSCSI) },
 148         { 0 }
 149 };
 150
 151 MODULE_DEVICE_TABLE(pci, pvscsi_pci_tbl);
 152
 153 static struct device *
 154 pvscsi_dev(const struct pvscsi_adapter *adapter)
 155 {
 156         return &(adapter->dev->dev);
 157 }
 158
 159 static struct pvscsi_ctx *
 160 pvscsi_find_context(const struct pvscsi_adapter *adapter, struct scsi_cmnd *cmd)
 161 {
 162         struct pvscsi_ctx *ctx, *end;
 163
 164         end = &adapter->cmd_map[adapter->req_depth];
 165         for (ctx = adapter->cmd_map; ctx < end; ctx++)
 166                 if (ctx->cmd == cmd)
 167                         return ctx;
 168
 169         return NULL;
 170 }
 171
 172 static struct pvscsi_ctx *
 173 pvscsi_acquire_context(struct pvscsi_adapter *adapter, struct scsi_cmnd *cmd)
 174 {
 175         struct pvscsi_ctx *ctx;
 176
 177         if (list_empty(&adapter->cmd_pool))
 178                 return NULL;
 179
 180         ctx = list_first_entry(&adapter->cmd_pool, struct pvscsi_ctx, list);
 181         ctx->cmd = cmd;
 182         list_del(&ctx->list);
 183
 184         return ctx;
 185 }
 186
 187 static void pvscsi_release_context(struct pvscsi_adapter *adapter,
 188                                    struct pvscsi_ctx *ctx)
 189 {
 190         ctx->cmd = NULL;
 191         ctx->abort_cmp = NULL;
 192         list_add(&ctx->list, &adapter->cmd_pool);
 193 }
 194
 195 /*
 196  * Map a pvscsi_ctx struct to a context ID field value; we map to a simple
 197  * non-zero integer. ctx always points to an entry in cmd_map array, hence
 198  * the return value is always >=1.
 199  */
 200 static u64 pvscsi_map_context(const struct pvscsi_adapter *adapter,
 201                               const struct pvscsi_ctx *ctx)
 202 {
 203         return ctx - adapter->cmd_map + 1;
 204 }
 205
 206 static struct pvscsi_ctx *
 207 pvscsi_get_context(const struct pvscsi_adapter *adapter, u64 context)
 208 {
 209         return &adapter->cmd_map[context - 1];
 210 }
 211
 212 static void pvscsi_reg_write(const struct pvscsi_adapter *adapter,
 213                              u32 offset, u32 val)
 214 {
 215         writel(val, adapter->mmioBase + offset);
 216 }
 217
 218 static u32 pvscsi_reg_read(const struct pvscsi_adapter *adapter, u32 offset)
 219 {
 220         return readl(adapter->mmioBase + offset);
 221 }
 222
 223 static u32 pvscsi_read_intr_status(const struct pvscsi_adapter *adapter)
 224 {
 225         return pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_INTR_STATUS);
 226 }
 227
 228 static void pvscsi_write_intr_status(const struct pvscsi_adapter *adapter,
 229                                      u32 val)
 230 {
 231         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_INTR_STATUS, val);
 232 }
 233
 234 static void pvscsi_unmask_intr(const struct pvscsi_adapter *adapter)
 235 {
 236         u32 intr_bits;
 237
 238         intr_bits = PVSCSI_INTR_CMPL_MASK;
 239         if (adapter->use_msg)
 240                 intr_bits |= PVSCSI_INTR_MSG_MASK;
 241
 242         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_INTR_MASK, intr_bits);
 243 }
 244
 245 static void pvscsi_mask_intr(const struct pvscsi_adapter *adapter)
 246 {
 247         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_INTR_MASK, 0);
 248 }
 249
 250 static void pvscsi_write_cmd_desc(const struct pvscsi_adapter *adapter,
 251                                   u32 cmd, const void *desc, size_t len)
 252 {
 253         const u32 *ptr = desc;
 254         size_t i;
 255
 256         len /= sizeof(*ptr);
 257         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND, cmd);
 258         for (i = 0; i < len; i++)
 259                 pvscsi_reg_write(adapter,
 260                                  PVSCSI_REG_OFFSET_COMMAND_DATA, ptr[i]);
 261 }
 262
 263 static void pvscsi_abort_cmd(const struct pvscsi_adapter *adapter,
 264                              const struct pvscsi_ctx *ctx)
 265 {
 266         struct PVSCSICmdDescAbortCmd cmd = { 0 };
 267
 268         cmd.target = ctx->cmd->device->id;
 269         cmd.context = pvscsi_map_context(adapter, ctx);
 270
 271         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_ABORT_CMD, &cmd, sizeof(cmd));
 272 }
 273
 274 static void pvscsi_kick_rw_io(const struct pvscsi_adapter *adapter)
 275 {
 276         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_KICK_RW_IO, 0);
 277 }
 278
 279 static void pvscsi_process_request_ring(const struct pvscsi_adapter *adapter)
 280 {
 281         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_KICK_NON_RW_IO, 0);
 282 }
 283
 284 static int scsi_is_rw(unsigned char op)
 285 {
 286         return op == READ_6  || op == WRITE_6 ||
 287                op == READ_10 || op == WRITE_10 ||
 288                op == READ_12 || op == WRITE_12 ||
 289                op == READ_16 || op == WRITE_16;
 290 }
 291
 292 static void pvscsi_kick_io(const struct pvscsi_adapter *adapter,
 293                            unsigned char op)
 294 {
 295         if (scsi_is_rw(op)) {
 296                 struct PVSCSIRingsState *s = adapter->rings_state;
 297
 298                 if (!adapter->use_req_threshold ||
 299                     s->reqProdIdx - s->reqConsIdx >= s->reqCallThreshold)
 300                         pvscsi_kick_rw_io(adapter);
 301         } else {
 302                 pvscsi_process_request_ring(adapter);
 303         }
 304 }
 305
 306 static void ll_adapter_reset(const struct pvscsi_adapter *adapter)
 307 {
 308         dev_dbg(pvscsi_dev(adapter), "Adapter Reset on %p\n", adapter);
 309
 310         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_ADAPTER_RESET, NULL, 0);
 311 }
 312
 313 static void ll_bus_reset(const struct pvscsi_adapter *adapter)
 314 {
 315         dev_dbg(pvscsi_dev(adapter), "Resetting bus on %p\n", adapter);
 316
 317         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_RESET_BUS, NULL, 0);
 318 }
 319
 320 static void ll_device_reset(const struct pvscsi_adapter *adapter, u32 target)
 321 {
 322         struct PVSCSICmdDescResetDevice cmd = { 0 };
 323
 324         dev_dbg(pvscsi_dev(adapter), "Resetting device: target=%u\n", target);
 325
 326         cmd.target = target;
 327
 328         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_RESET_DEVICE,
 329                               &cmd, sizeof(cmd));
 330 }
 331
 332 static void pvscsi_create_sg(struct pvscsi_ctx *ctx,
 333                              struct scatterlist *sg, unsigned count)
 334 {
 335         unsigned i;
 336         struct PVSCSISGElement *sge;
 337
 338         BUG_ON(count > PVSCSI_MAX_NUM_SG_ENTRIES_PER_SEGMENT);
 339
 340         sge = &ctx->sgl->sge[0];
 341         for (i = 0; i < count; i++, sg++) {
 342                 sge[i].addr   = sg_dma_address(sg);
 343                 sge[i].length = sg_dma_len(sg);
 344                 sge[i].flags  = 0;
 345         }
 346 }
 347
 348 /*
 349  * Map all data buffers for a command into PCI space and
 350  * setup the scatter/gather list if needed.
 351  */
 352 static int pvscsi_map_buffers(struct pvscsi_adapter *adapter,
 353                               struct pvscsi_ctx *ctx, struct scsi_cmnd *cmd,
 354                               struct PVSCSIRingReqDesc *e)
 355 {
 356         unsigned count;
 357         unsigned bufflen = scsi_bufflen(cmd);
 358         struct scatterlist *sg;
 359
 360         e->dataLen = bufflen;
 361         e->dataAddr = 0;
 362         if (bufflen == 0)
 363                 return 0;
 364
 365         sg = scsi_sglist(cmd);
 366         count = scsi_sg_count(cmd);
 367         if (count != 0) {
 368                 int segs = scsi_dma_map(cmd);
 369
 370                 if (segs == -ENOMEM) {
 371                         scmd_printk(KERN_ERR, cmd,
 372                                     "vmw_pvscsi: Failed to map cmd sglist for DMA.\n");
 373                         return -ENOMEM;
 374                 } else if (segs > 1) {
 375                         pvscsi_create_sg(ctx, sg, segs);
 376
 377                         e->flags |= PVSCSI_FLAG_CMD_WITH_SG_LIST;
 378                         ctx->sglPA = pci_map_single(adapter->dev, ctx->sgl,
 379                                                     SGL_SIZE, PCI_DMA_TODEVICE);
 380                         if (pci_dma_mapping_error(adapter->dev, ctx->sglPA)) {
 381                                 scmd_printk(KERN_ERR, cmd,
 382                                             "vmw_pvscsi: Failed to map ctx sglist for DMA.\n");
 383                                 scsi_dma_unmap(cmd);
 384                                 ctx->sglPA = 0;
 385                                 return -ENOMEM;
 386                         }
 387                         e->dataAddr = ctx->sglPA;
 388                 } else
 389                         e->dataAddr = sg_dma_address(sg);
 390         } else {
 391                 /*
 392                  * In case there is no S/G list, scsi_sglist points
 393                  * directly to the buffer.
 394                  */
 395                 ctx->dataPA = pci_map_single(adapter->dev, sg, bufflen,
 396                                              cmd->sc_data_direction);
 397                 if (pci_dma_mapping_error(adapter->dev, ctx->dataPA)) {
 398                         scmd_printk(KERN_ERR, cmd,
 399                                     "vmw_pvscsi: Failed to map direct data buffer for DMA.\n");
 400                         return -ENOMEM;
 401                 }
 402                 e->dataAddr = ctx->dataPA;
 403         }
 404
 405         return 0;
 406 }
 407
 408 static void pvscsi_unmap_buffers(const struct pvscsi_adapter *adapter,
 409                                  struct pvscsi_ctx *ctx)
 410 {
 411         struct scsi_cmnd *cmd;
 412         unsigned bufflen;
 413
 414         cmd = ctx->cmd;
 415         bufflen = scsi_bufflen(cmd);
 416
 417         if (bufflen != 0) {
 418                 unsigned count = scsi_sg_count(cmd);
 419
 420                 if (count != 0) {
 421                         scsi_dma_unmap(cmd);
 422                         if (ctx->sglPA) {
 423                                 pci_unmap_single(adapter->dev, ctx->sglPA,
 424                                                  SGL_SIZE, PCI_DMA_TODEVICE);
 425                                 ctx->sglPA = 0;
 426                         }
 427                 } else
 428                         pci_unmap_single(adapter->dev, ctx->dataPA, bufflen,
 429                                          cmd->sc_data_direction);
 430         }
 431         if (cmd->sense_buffer)
 432                 pci_unmap_single(adapter->dev, ctx->sensePA,
 433                                  SCSI_SENSE_BUFFERSIZE, PCI_DMA_FROMDEVICE);
 434 }
 435
 436 static int pvscsi_allocate_rings(struct pvscsi_adapter *adapter)
 437 {
 438         adapter->rings_state = pci_alloc_consistent(adapter->dev, PAGE_SIZE,
 439                                                     &adapter->ringStatePA);
 440         if (!adapter->rings_state)
 441                 return -ENOMEM;
 442
 443         adapter->req_pages = min(PVSCSI_MAX_NUM_PAGES_REQ_RING,
 444                                  pvscsi_ring_pages);
 445         adapter->req_depth = adapter->req_pages
 446                                         * PVSCSI_MAX_NUM_REQ_ENTRIES_PER_PAGE;
 447         adapter->req_ring = pci_alloc_consistent(adapter->dev,
 448                                                  adapter->req_pages * PAGE_SIZE,
 449                                                  &adapter->reqRingPA);
 450         if (!adapter->req_ring)
 451                 return -ENOMEM;
 452
 453         adapter->cmp_pages = min(PVSCSI_MAX_NUM_PAGES_CMP_RING,
 454                                  pvscsi_ring_pages);
 455         adapter->cmp_ring = pci_alloc_consistent(adapter->dev,
 456                                                  adapter->cmp_pages * PAGE_SIZE,
 457                                                  &adapter->cmpRingPA);
 458         if (!adapter->cmp_ring)
 459                 return -ENOMEM;
 460
 461         BUG_ON(!IS_ALIGNED(adapter->ringStatePA, PAGE_SIZE));
 462         BUG_ON(!IS_ALIGNED(adapter->reqRingPA, PAGE_SIZE));
 463         BUG_ON(!IS_ALIGNED(adapter->cmpRingPA, PAGE_SIZE));
 464
 465         if (!adapter->use_msg)
 466                 return 0;
 467
 468         adapter->msg_pages = min(PVSCSI_MAX_NUM_PAGES_MSG_RING,
 469                                  pvscsi_msg_ring_pages);
 470         adapter->msg_ring = pci_alloc_consistent(adapter->dev,
 471                                                  adapter->msg_pages * PAGE_SIZE,
 472                                                  &adapter->msgRingPA);
 473         if (!adapter->msg_ring)
 474                 return -ENOMEM;
 475         BUG_ON(!IS_ALIGNED(adapter->msgRingPA, PAGE_SIZE));
 476
 477         return 0;
 478 }
 479
 480 static void pvscsi_setup_all_rings(const struct pvscsi_adapter *adapter)
 481 {
 482         struct PVSCSICmdDescSetupRings cmd = { 0 };
 483         dma_addr_t base;
 484         unsigned i;
 485
 486         cmd.ringsStatePPN   = adapter->ringStatePA >> PAGE_SHIFT;
 487         cmd.reqRingNumPages = adapter->req_pages;
 488         cmd.cmpRingNumPages = adapter->cmp_pages;
 489
 490         base = adapter->reqRingPA;
 491         for (i = 0; i < adapter->req_pages; i++) {
 492                 cmd.reqRingPPNs[i] = base >> PAGE_SHIFT;
 493                 base += PAGE_SIZE;
 494         }
 495
 496         base = adapter->cmpRingPA;
 497         for (i = 0; i < adapter->cmp_pages; i++) {
 498                 cmd.cmpRingPPNs[i] = base >> PAGE_SHIFT;
 499                 base += PAGE_SIZE;
 500         }
 501
 502         memset(adapter->rings_state, 0, PAGE_SIZE);
 503         memset(adapter->req_ring, 0, adapter->req_pages * PAGE_SIZE);
 504         memset(adapter->cmp_ring, 0, adapter->cmp_pages * PAGE_SIZE);
 505
 506         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_SETUP_RINGS,
 507                               &cmd, sizeof(cmd));
 508
 509         if (adapter->use_msg) {
 510                 struct PVSCSICmdDescSetupMsgRing cmd_msg = { 0 };
 511
 512                 cmd_msg.numPages = adapter->msg_pages;
 513
 514                 base = adapter->msgRingPA;
 515                 for (i = 0; i < adapter->msg_pages; i++) {
 516                         cmd_msg.ringPPNs[i] = base >> PAGE_SHIFT;
 517                         base += PAGE_SIZE;
 518                 }
 519                 memset(adapter->msg_ring, 0, adapter->msg_pages * PAGE_SIZE);
 520
 521                 pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_SETUP_MSG_RING,
 522                                       &cmd_msg, sizeof(cmd_msg));
 523         }
 524 }
 525
 526 static int pvscsi_change_queue_depth(struct scsi_device *sdev, int qdepth)
 527 {
 528         if (!sdev->tagged_supported)
 529                 qdepth = 1;
 530         return scsi_change_queue_depth(sdev, qdepth);
 531 }
 532
 533 /*
 534  * Pull a completion descriptor off and pass the completion back
 535  * to the SCSI mid layer.
 536  */
 537 static void pvscsi_complete_request(struct pvscsi_adapter *adapter,
 538                                     const struct PVSCSIRingCmpDesc *e)
 539 {
 540         struct pvscsi_ctx *ctx;
 541         struct scsi_cmnd *cmd;
 542         struct completion *abort_cmp;
 543         u32 btstat = e->hostStatus;
 544         u32 sdstat = e->scsiStatus;
 545
 546         ctx = pvscsi_get_context(adapter, e->context);
 547         cmd = ctx->cmd;
 548         abort_cmp = ctx->abort_cmp;
 549         pvscsi_unmap_buffers(adapter, ctx);
 550         pvscsi_release_context(adapter, ctx);
 551         if (abort_cmp) {
 552                 /*
 553                  * The command was requested to be aborted. Just signal that
 554                  * the request completed and swallow the actual cmd completion
 555                  * here. The abort handler will post a completion for this
 556                  * command indicating that it got successfully aborted.
 557                  */
 558                 complete(abort_cmp);
 559                 return;
 560         }
 561
 562         cmd->result = 0;
 563         if (sdstat != SAM_STAT_GOOD &&
 564             (btstat == BTSTAT_SUCCESS ||
 565              btstat == BTSTAT_LINKED_COMMAND_COMPLETED ||
 566              btstat == BTSTAT_LINKED_COMMAND_COMPLETED_WITH_FLAG)) {
 567                 cmd->result = (DID_OK << 16) | sdstat;
 568                 if (sdstat == SAM_STAT_CHECK_CONDITION && cmd->sense_buffer)
 569                         cmd->result |= (DRIVER_SENSE << 24);
 570         } else
 571                 switch (btstat) {
 572                 case BTSTAT_SUCCESS:
 573                 case BTSTAT_LINKED_COMMAND_COMPLETED:
 574                 case BTSTAT_LINKED_COMMAND_COMPLETED_WITH_FLAG:
 575                         /* If everything went fine, let's move on..  */
 576                         cmd->result = (DID_OK << 16);
 577                         break;
 578
 579                 case BTSTAT_DATARUN:
 580                 case BTSTAT_DATA_UNDERRUN:
 581                         /* Report residual data in underruns */
 582                         scsi_set_resid(cmd, scsi_bufflen(cmd) - e->dataLen);
 583                         cmd->result = (DID_ERROR << 16);
 584                         break;
 585
 586                 case BTSTAT_SELTIMEO:
 587                         /* Our emulation returns this for non-connected devs */
 588                         cmd->result = (DID_BAD_TARGET << 16);
 589                         break;
 590
 591                 case BTSTAT_LUNMISMATCH:
 592                 case BTSTAT_TAGREJECT:
 593                 case BTSTAT_BADMSG:
 594                         cmd->result = (DRIVER_INVALID << 24);
 595                         /* fall through */
 596
 597                 case BTSTAT_HAHARDWARE:
 598                 case BTSTAT_INVPHASE:
 599                 case BTSTAT_HATIMEOUT:
 600                 case BTSTAT_NORESPONSE:
 601                 case BTSTAT_DISCONNECT:
 602                 case BTSTAT_HASOFTWARE:
 603                 case BTSTAT_BUSFREE:
 604                 case BTSTAT_SENSFAILED:
 605                         cmd->result |= (DID_ERROR << 16);
 606                         break;
 607
 608                 case BTSTAT_SENTRST:
 609                 case BTSTAT_RECVRST:
 610                 case BTSTAT_BUSRESET:
 611                         cmd->result = (DID_RESET << 16);
 612                         break;
 613
 614                 case BTSTAT_ABORTQUEUE:
 615                         cmd->result = (DID_ABORT << 16);
 616                         break;
 617
 618                 case BTSTAT_SCSIPARITY:
 619                         cmd->result = (DID_PARITY << 16);
 620                         break;
 621
 622                 default:
 623                         cmd->result = (DID_ERROR << 16);
 624                         scmd_printk(KERN_DEBUG, cmd,
 625                                     "Unknown completion status: 0x%x\n",
 626                                     btstat);
 627         }
 628
 629         dev_dbg(&cmd->device->sdev_gendev,
 630                 "cmd=%p %x ctx=%p result=0x%x status=0x%x,%x\n",
 631                 cmd, cmd->cmnd[0], ctx, cmd->result, btstat, sdstat);
 632
 633         cmd->scsi_done(cmd);
 634 }
 635
 636 /*
 637  * barrier usage : Since the PVSCSI device is emulated, there could be cases
 638  * where we may want to serialize some accesses between the driver and the
 639  * emulation layer. We use compiler barriers instead of the more expensive
 640  * memory barriers because PVSCSI is only supported on X86 which has strong
 641  * memory access ordering.
 642  */
 643 static void pvscsi_process_completion_ring(struct pvscsi_adapter *adapter)
 644 {
 645         struct PVSCSIRingsState *s = adapter->rings_state;
 646         struct PVSCSIRingCmpDesc *ring = adapter->cmp_ring;
 647         u32 cmp_entries = s->cmpNumEntriesLog2;
 648
 649         while (s->cmpConsIdx != s->cmpProdIdx) {
 650                 struct PVSCSIRingCmpDesc *e = ring + (s->cmpConsIdx &
 651                                                       MASK(cmp_entries));
 652                 /*
 653                  * This barrier() ensures that *e is not dereferenced while
 654                  * the device emulation still writes data into the slot.
 655                  * Since the device emulation advances s->cmpProdIdx only after
 656                  * updating the slot we want to check it first.
 657                  */
 658                 barrier();
 659                 pvscsi_complete_request(adapter, e);
 660                 /*
 661                  * This barrier() ensures that compiler doesn't reorder write
 662                  * to s->cmpConsIdx before the read of (*e) inside
 663                  * pvscsi_complete_request. Otherwise, device emulation may
 664                  * overwrite *e before we had a chance to read it.
 665                  */
 666                 barrier();
 667                 s->cmpConsIdx++;
 668         }
 669 }
 670
 671 /*
 672  * Translate a Linux SCSI request into a request ring entry.
 673  */
 674 static int pvscsi_queue_ring(struct pvscsi_adapter *adapter,
 675                              struct pvscsi_ctx *ctx, struct scsi_cmnd *cmd)
 676 {
 677         struct PVSCSIRingsState *s;
 678         struct PVSCSIRingReqDesc *e;
 679         struct scsi_device *sdev;
 680         u32 req_entries;
 681
 682         s = adapter->rings_state;
 683         sdev = cmd->device;
 684         req_entries = s->reqNumEntriesLog2;
 685
 686         /*
 687          * If this condition holds, we might have room on the request ring, but
 688          * we might not have room on the completion ring for the response.
 689          * However, we have already ruled out this possibility - we would not
 690          * have successfully allocated a context if it were true, since we only
 691          * have one context per request entry.  Check for it anyway, since it
 692          * would be a serious bug.
 693          */
 694         if (s->reqProdIdx - s->cmpConsIdx >= 1 << req_entries) {
 695                 scmd_printk(KERN_ERR, cmd, "vmw_pvscsi: "
 696                             "ring full: reqProdIdx=%d cmpConsIdx=%d\n",
 697                             s->reqProdIdx, s->cmpConsIdx);
 698                 return -1;
 699         }
 700
 701         e = adapter->req_ring + (s->reqProdIdx & MASK(req_entries));
 702
 703         e->bus    = sdev->channel;
 704         e->target = sdev->id;
 705         memset(e->lun, 0, sizeof(e->lun));
 706         e->lun[1] = sdev->lun;
 707
 708         if (cmd->sense_buffer) {
 709                 ctx->sensePA = pci_map_single(adapter->dev, cmd->sense_buffer,
 710                                               SCSI_SENSE_BUFFERSIZE,
 711                                               PCI_DMA_FROMDEVICE);
 712                 if (pci_dma_mapping_error(adapter->dev, ctx->sensePA)) {
 713                         scmd_printk(KERN_ERR, cmd,
 714                                     "vmw_pvscsi: Failed to map sense buffer for DMA.\n");
 715                         ctx->sensePA = 0;
 716                         return -ENOMEM;
 717                 }
 718                 e->senseAddr = ctx->sensePA;
 719                 e->senseLen = SCSI_SENSE_BUFFERSIZE;
 720         } else {
 721                 e->senseLen  = 0;
 722                 e->senseAddr = 0;
 723         }
 724         e->cdbLen   = cmd->cmd_len;
 725         e->vcpuHint = smp_processor_id();
 726         memcpy(e->cdb, cmd->cmnd, e->cdbLen);
 727
 728         e->tag = SIMPLE_QUEUE_TAG;
 729
 730         if (cmd->sc_data_direction == DMA_FROM_DEVICE)
 731                 e->flags = PVSCSI_FLAG_CMD_DIR_TOHOST;
 732         else if (cmd->sc_data_direction == DMA_TO_DEVICE)
 733                 e->flags = PVSCSI_FLAG_CMD_DIR_TODEVICE;
 734         else if (cmd->sc_data_direction == DMA_NONE)
 735                 e->flags = PVSCSI_FLAG_CMD_DIR_NONE;
 736         else
 737                 e->flags = 0;
 738
 739         if (pvscsi_map_buffers(adapter, ctx, cmd, e) != 0) {
 740                 if (cmd->sense_buffer) {
 741                         pci_unmap_single(adapter->dev, ctx->sensePA,
 742                                          SCSI_SENSE_BUFFERSIZE,
 743                                          PCI_DMA_FROMDEVICE);
 744                         ctx->sensePA = 0;
 745                 }
 746                 return -ENOMEM;
 747         }
 748
 749         e->context = pvscsi_map_context(adapter, ctx);
 750
 751         barrier();
 752
 753         s->reqProdIdx++;
 754
 755         return 0;
 756 }
 757
 758 static int pvscsi_queue_lck(struct scsi_cmnd *cmd, void (*done)(struct scsi_cmnd *))
 759 {
 760         struct Scsi_Host *host = cmd->device->host;
 761         struct pvscsi_adapter *adapter = shost_priv(host);
 762         struct pvscsi_ctx *ctx;
 763         unsigned long flags;
 764
 765         spin_lock_irqsave(&adapter->hw_lock, flags);
 766
 767         ctx = pvscsi_acquire_context(adapter, cmd);
 768         if (!ctx || pvscsi_queue_ring(adapter, ctx, cmd) != 0) {
 769                 if (ctx)
 770                         pvscsi_release_context(adapter, ctx);
 771                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 772                 return SCSI_MLQUEUE_HOST_BUSY;
 773         }
 774
 775         cmd->scsi_done = done;
 776
 777         dev_dbg(&cmd->device->sdev_gendev,
 778                 "queued cmd %p, ctx %p, op=%x\n", cmd, ctx, cmd->cmnd[0]);
 779
 780         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 781
 782         pvscsi_kick_io(adapter, cmd->cmnd[0]);
 783
 784         return 0;
 785 }
 786
 787 static DEF_SCSI_QCMD(pvscsi_queue)
 788
 789 static int pvscsi_abort(struct scsi_cmnd *cmd)
 790 {
 791         struct pvscsi_adapter *adapter = shost_priv(cmd->device->host);
 792         struct pvscsi_ctx *ctx;
 793         unsigned long flags;
 794         int result = SUCCESS;
 795         DECLARE_COMPLETION_ONSTACK(abort_cmp);
 796         int done;
 797
 798         scmd_printk(KERN_DEBUG, cmd, "task abort on host %u, %p\n",
 799                     adapter->host->host_no, cmd);
 800
 801         spin_lock_irqsave(&adapter->hw_lock, flags);
 802
 803         /*
 804          * Poll the completion ring first - we might be trying to abort
 805          * a command that is waiting to be dispatched in the completion ring.
 806          */
 807         pvscsi_process_completion_ring(adapter);
 808
 809         /*
 810          * If there is no context for the command, it either already succeeded
 811          * or else was never properly issued.  Not our problem.
 812          */
 813         ctx = pvscsi_find_context(adapter, cmd);
 814         if (!ctx) {
 815                 scmd_printk(KERN_DEBUG, cmd, "Failed to abort cmd %p\n", cmd);
 816                 goto out;
 817         }
 818
 819         /*
 820          * Mark that the command has been requested to be aborted and issue
 821          * the abort.
 822          */
 823         ctx->abort_cmp = &abort_cmp;
 824
 825         pvscsi_abort_cmd(adapter, ctx);
 826         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 827         /* Wait for 2 secs for the completion. */
 828         done = wait_for_completion_timeout(&abort_cmp, msecs_to_jiffies(2000));
 829         spin_lock_irqsave(&adapter->hw_lock, flags);
 830
 831         if (!done) {
 832                 /*
 833                  * Failed to abort the command, unmark the fact that it
 834                  * was requested to be aborted.
 835                  */
 836                 ctx->abort_cmp = NULL;
 837                 result = FAILED;
 838                 scmd_printk(KERN_DEBUG, cmd,
 839                             "Failed to get completion for aborted cmd %p\n",
 840                             cmd);
 841                 goto out;
 842         }
 843
 844         /*
 845          * Successfully aborted the command.
 846          */
 847         cmd->result = (DID_ABORT << 16);
 848         cmd->scsi_done(cmd);
 849
 850 out:
 851         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 852         return result;
 853 }
 854
 855 /*
 856  * Abort all outstanding requests.  This is only safe to use if the completion
 857  * ring will never be walked again or the device has been reset, because it
 858  * destroys the 1-1 mapping between context field passed to emulation and our
 859  * request structure.
 860  */
 861 static void pvscsi_reset_all(struct pvscsi_adapter *adapter)
 862 {
 863         unsigned i;
 864
 865         for (i = 0; i < adapter->req_depth; i++) {
 866                 struct pvscsi_ctx *ctx = &adapter->cmd_map[i];
 867                 struct scsi_cmnd *cmd = ctx->cmd;
 868                 if (cmd) {
 869                         scmd_printk(KERN_ERR, cmd,
 870                                     "Forced reset on cmd %p\n", cmd);
 871                         pvscsi_unmap_buffers(adapter, ctx);
 872                         pvscsi_release_context(adapter, ctx);
 873                         cmd->result = (DID_RESET << 16);
 874                         cmd->scsi_done(cmd);
 875                 }
 876         }
 877 }
 878
 879 static int pvscsi_host_reset(struct scsi_cmnd *cmd)
 880 {
 881         struct Scsi_Host *host = cmd->device->host;
 882         struct pvscsi_adapter *adapter = shost_priv(host);
 883         unsigned long flags;
 884         bool use_msg;
 885
 886         scmd_printk(KERN_INFO, cmd, "SCSI Host reset\n");
 887
 888         spin_lock_irqsave(&adapter->hw_lock, flags);
 889
 890         use_msg = adapter->use_msg;
 891
 892         if (use_msg) {
 893                 adapter->use_msg = 0;
 894                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 895
 896                 /*
 897                  * Now that we know that the ISR won't add more work on the
 898                  * workqueue we can safely flush any outstanding work.
 899                  */
 900                 flush_workqueue(adapter->workqueue);
 901                 spin_lock_irqsave(&adapter->hw_lock, flags);
 902         }
 903
 904         /*
 905          * We're going to tear down the entire ring structure and set it back
 906          * up, so stalling new requests until all completions are flushed and
 907          * the rings are back in place.
 908          */
 909
 910         pvscsi_process_request_ring(adapter);
 911
 912         ll_adapter_reset(adapter);
 913
 914         /*
 915          * Now process any completions.  Note we do this AFTER adapter reset,
 916          * which is strange, but stops races where completions get posted
 917          * between processing the ring and issuing the reset.  The backend will
 918          * not touch the ring memory after reset, so the immediately pre-reset
 919          * completion ring state is still valid.
 920          */
 921         pvscsi_process_completion_ring(adapter);
 922
 923         pvscsi_reset_all(adapter);
 924         adapter->use_msg = use_msg;
 925         pvscsi_setup_all_rings(adapter);
 926         pvscsi_unmask_intr(adapter);
 927
 928         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 929
 930         return SUCCESS;
 931 }
 932
 933 static int pvscsi_bus_reset(struct scsi_cmnd *cmd)
 934 {
 935         struct Scsi_Host *host = cmd->device->host;
 936         struct pvscsi_adapter *adapter = shost_priv(host);
 937         unsigned long flags;
 938
 939         scmd_printk(KERN_INFO, cmd, "SCSI Bus reset\n");
 940
 941         /*
 942          * We don't want to queue new requests for this bus after
 943          * flushing all pending requests to emulation, since new
 944          * requests could then sneak in during this bus reset phase,
 945          * so take the lock now.
 946          */
 947         spin_lock_irqsave(&adapter->hw_lock, flags);
 948
 949         pvscsi_process_request_ring(adapter);
 950         ll_bus_reset(adapter);
 951         pvscsi_process_completion_ring(adapter);
 952
 953         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 954
 955         return SUCCESS;
 956 }
 957
 958 static int pvscsi_device_reset(struct scsi_cmnd *cmd)
 959 {
 960         struct Scsi_Host *host = cmd->device->host;
 961         struct pvscsi_adapter *adapter = shost_priv(host);
 962         unsigned long flags;
 963
 964         scmd_printk(KERN_INFO, cmd, "SCSI device reset on scsi%u:%u\n",
 965                     host->host_no, cmd->device->id);
 966
 967         /*
 968          * We don't want to queue new requests for this device after flushing
 969          * all pending requests to emulation, since new requests could then
 970          * sneak in during this device reset phase, so take the lock now.
 971          */
 972         spin_lock_irqsave(&adapter->hw_lock, flags);
 973
 974         pvscsi_process_request_ring(adapter);
 975         ll_device_reset(adapter, cmd->device->id);
 976         pvscsi_process_completion_ring(adapter);
 977
 978         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 979
 980         return SUCCESS;
 981 }
 982
 983 static struct scsi_host_template pvscsi_template;
 984
 985 static const char *pvscsi_info(struct Scsi_Host *host)
 986 {
 987         struct pvscsi_adapter *adapter = shost_priv(host);
 988         static char buf[256];
 989
 990         sprintf(buf, "VMware PVSCSI storage adapter rev %d, req/cmp/msg rings: "
 991                 "%u/%u/%u pages, cmd_per_lun=%u", adapter->rev,
 992                 adapter->req_pages, adapter->cmp_pages, adapter->msg_pages,
 993                 pvscsi_template.cmd_per_lun);
 994
 995         return buf;
 996 }
 997
 998 static struct scsi_host_template pvscsi_template = {
 999         .module                         = THIS_MODULE,
1000         .name                           = "VMware PVSCSI Host Adapter",
1001         .proc_name                      = "vmw_pvscsi",
1002         .info                           = pvscsi_info,
1003         .queuecommand                   = pvscsi_queue,
1004         .this_id                        = -1,
1005         .sg_tablesize                   = PVSCSI_MAX_NUM_SG_ENTRIES_PER_SEGMENT,
1006         .dma_boundary                   = UINT_MAX,
1007         .max_sectors                    = 0xffff,
1008         .use_clustering                 = ENABLE_CLUSTERING,
1009         .change_queue_depth             = pvscsi_change_queue_depth,
1010         .eh_abort_handler               = pvscsi_abort,
1011         .eh_device_reset_handler        = pvscsi_device_reset,
1012         .eh_bus_reset_handler           = pvscsi_bus_reset,
1013         .eh_host_reset_handler          = pvscsi_host_reset,
1014 };
1015
1016 static void pvscsi_process_msg(const struct pvscsi_adapter *adapter,
1017                                const struct PVSCSIRingMsgDesc *e)
1018 {
1019         struct PVSCSIRingsState *s = adapter->rings_state;
1020         struct Scsi_Host *host = adapter->host;
1021         struct scsi_device *sdev;
1022
1023         printk(KERN_INFO "vmw_pvscsi: msg type: 0x%x - MSG RING: %u/%u (%u) \n",
1024                e->type, s->msgProdIdx, s->msgConsIdx, s->msgNumEntriesLog2);
1025
1026         BUILD_BUG_ON(PVSCSI_MSG_LAST != 2);
1027
1028         if (e->type == PVSCSI_MSG_DEV_ADDED) {
1029                 struct PVSCSIMsgDescDevStatusChanged *desc;
1030                 desc = (struct PVSCSIMsgDescDevStatusChanged *)e;
1031
1032                 printk(KERN_INFO
1033                        "vmw_pvscsi: msg: device added at scsi%u:%u:%u\n",
1034                        desc->bus, desc->target, desc->lun[1]);
1035
1036                 if (!scsi_host_get(host))
1037                         return;
1038
1039                 sdev = scsi_device_lookup(host, desc->bus, desc->target,
1040                                           desc->lun[1]);
1041                 if (sdev) {
1042                         printk(KERN_INFO "vmw_pvscsi: device already exists\n");
1043                         scsi_device_put(sdev);
1044                 } else
1045                         scsi_add_device(adapter->host, desc->bus,
1046                                         desc->target, desc->lun[1]);
1047
1048                 scsi_host_put(host);
1049         } else if (e->type == PVSCSI_MSG_DEV_REMOVED) {
1050                 struct PVSCSIMsgDescDevStatusChanged *desc;
1051                 desc = (struct PVSCSIMsgDescDevStatusChanged *)e;
1052
1053                 printk(KERN_INFO
1054                        "vmw_pvscsi: msg: device removed at scsi%u:%u:%u\n",
1055                        desc->bus, desc->target, desc->lun[1]);
1056
1057                 if (!scsi_host_get(host))
1058                         return;
1059
1060                 sdev = scsi_device_lookup(host, desc->bus, desc->target,
1061                                           desc->lun[1]);
1062                 if (sdev) {
1063                         scsi_remove_device(sdev);
1064                         scsi_device_put(sdev);
1065                 } else
1066                         printk(KERN_INFO
1067                                "vmw_pvscsi: failed to lookup scsi%u:%u:%u\n",
1068                                desc->bus, desc->target, desc->lun[1]);
1069
1070                 scsi_host_put(host);
1071         }
1072 }
1073
1074 static int pvscsi_msg_pending(const struct pvscsi_adapter *adapter)
1075 {
1076         struct PVSCSIRingsState *s = adapter->rings_state;
1077
1078         return s->msgProdIdx != s->msgConsIdx;
1079 }
1080
1081 static void pvscsi_process_msg_ring(const struct pvscsi_adapter *adapter)
1082 {
1083         struct PVSCSIRingsState *s = adapter->rings_state;
1084         struct PVSCSIRingMsgDesc *ring = adapter->msg_ring;
1085         u32 msg_entries = s->msgNumEntriesLog2;
1086
1087         while (pvscsi_msg_pending(adapter)) {
1088                 struct PVSCSIRingMsgDesc *e = ring + (s->msgConsIdx &
1089                                                       MASK(msg_entries));
1090
1091                 barrier();
1092                 pvscsi_process_msg(adapter, e);
1093                 barrier();
1094                 s->msgConsIdx++;
1095         }
1096 }
1097
1098 static void pvscsi_msg_workqueue_handler(struct work_struct *data)
1099 {
1100         struct pvscsi_adapter *adapter;
1101
1102         adapter = container_of(data, struct pvscsi_adapter, work);
1103
1104         pvscsi_process_msg_ring(adapter);
1105 }
1106
1107 static int pvscsi_setup_msg_workqueue(struct pvscsi_adapter *adapter)
1108 {
1109         char name[32];
1110
1111         if (!pvscsi_use_msg)
1112                 return 0;
1113
1114         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
1115                          PVSCSI_CMD_SETUP_MSG_RING);
1116
1117         if (pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS) == -1)
1118                 return 0;
1119
1120         snprintf(name, sizeof(name),
1121                  "vmw_pvscsi_wq_%u", adapter->host->host_no);
1122
1123         adapter->workqueue = create_singlethread_workqueue(name);
1124         if (!adapter->workqueue) {
1125                 printk(KERN_ERR "vmw_pvscsi: failed to create work queue\n");
1126                 return 0;
1127         }
1128         INIT_WORK(&adapter->work, pvscsi_msg_workqueue_handler);
1129
1130         return 1;
1131 }
1132
1133 static bool pvscsi_setup_req_threshold(struct pvscsi_adapter *adapter,
1134                                       bool enable)
1135 {
1136         u32 val;
1137
1138         if (!pvscsi_use_req_threshold)
1139                 return false;
1140
1141         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
1142                          PVSCSI_CMD_SETUP_REQCALLTHRESHOLD);
1143         val = pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS);
1144         if (val == -1) {
1145                 printk(KERN_INFO "vmw_pvscsi: device does not support req_threshold\n");
1146                 return false;
1147         } else {
1148                 struct PVSCSICmdDescSetupReqCall cmd_msg = { 0 };
1149                 cmd_msg.enable = enable;
1150                 printk(KERN_INFO
1151                        "vmw_pvscsi: %sabling reqCallThreshold\n",
1152                         enable ? "en" : "dis");
1153                 pvscsi_write_cmd_desc(adapter,
1154                                       PVSCSI_CMD_SETUP_REQCALLTHRESHOLD,
1155                                       &cmd_msg, sizeof(cmd_msg));
1156                 return pvscsi_reg_read(adapter,
1157                                        PVSCSI_REG_OFFSET_COMMAND_STATUS) != 0;
1158         }
1159 }
1160
1161 static irqreturn_t pvscsi_isr(int irq, void *devp)
1162 {
1163         struct pvscsi_adapter *adapter = devp;
1164         int handled;
1165
1166         if (adapter->use_msi || adapter->use_msix)
1167                 handled = true;
1168         else {
1169                 u32 val = pvscsi_read_intr_status(adapter);
1170                 handled = (val & PVSCSI_INTR_ALL_SUPPORTED) != 0;
1171                 if (handled)
1172                         pvscsi_write_intr_status(devp, val);
1173         }
1174
1175         if (handled) {
1176                 unsigned long flags;
1177
1178                 spin_lock_irqsave(&adapter->hw_lock, flags);
1179
1180                 pvscsi_process_completion_ring(adapter);
1181                 if (adapter->use_msg && pvscsi_msg_pending(adapter))
1182                         queue_work(adapter->workqueue, &adapter->work);
1183
1184                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
1185         }
1186
1187         return IRQ_RETVAL(handled);
1188 }
1189
1190 static void pvscsi_free_sgls(const struct pvscsi_adapter *adapter)
1191 {
1192         struct pvscsi_ctx *ctx = adapter->cmd_map;
1193         unsigned i;
1194
1195         for (i = 0; i < adapter->req_depth; ++i, ++ctx)
1196                 free_pages((unsigned long)ctx->sgl, get_order(SGL_SIZE));
1197 }
1198
1199 static int pvscsi_setup_msix(const struct pvscsi_adapter *adapter,
1200                              unsigned int *irq)
1201 {
1202         struct msix_entry entry = { 0, PVSCSI_VECTOR_COMPLETION };
1203         int ret;
1204
1205         ret = pci_enable_msix_exact(adapter->dev, &entry, 1);
1206         if (ret)
1207                 return ret;
1208
1209         *irq = entry.vector;
1210
1211         return 0;
1212 }
1213
1214 static void pvscsi_shutdown_intr(struct pvscsi_adapter *adapter)
1215 {
1216         if (adapter->irq) {
1217                 free_irq(adapter->irq, adapter);
1218                 adapter->irq = 0;
1219         }
1220         if (adapter->use_msi) {
1221                 pci_disable_msi(adapter->dev);
1222                 adapter->use_msi = 0;
1223         } else if (adapter->use_msix) {
1224                 pci_disable_msix(adapter->dev);
1225                 adapter->use_msix = 0;
1226         }
1227 }
1228
1229 static void pvscsi_release_resources(struct pvscsi_adapter *adapter)
1230 {
1231         pvscsi_shutdown_intr(adapter);
1232
1233         if (adapter->workqueue)
1234                 destroy_workqueue(adapter->workqueue);
1235
1236         if (adapter->mmioBase)
1237                 pci_iounmap(adapter->dev, adapter->mmioBase);
1238
1239         pci_release_regions(adapter->dev);
1240
1241         if (adapter->cmd_map) {
1242                 pvscsi_free_sgls(adapter);
1243                 kfree(adapter->cmd_map);
1244         }
1245
1246         if (adapter->rings_state)
1247                 pci_free_consistent(adapter->dev, PAGE_SIZE,
1248                                     adapter->rings_state, adapter->ringStatePA);
1249
1250         if (adapter->req_ring)
1251                 pci_free_consistent(adapter->dev,
1252                                     adapter->req_pages * PAGE_SIZE,
1253                                     adapter->req_ring, adapter->reqRingPA);
1254
1255         if (adapter->cmp_ring)
1256                 pci_free_consistent(adapter->dev,
1257                                     adapter->cmp_pages * PAGE_SIZE,
1258                                     adapter->cmp_ring, adapter->cmpRingPA);
1259
1260         if (adapter->msg_ring)
1261                 pci_free_consistent(adapter->dev,
1262                                     adapter->msg_pages * PAGE_SIZE,
1263                                     adapter->msg_ring, adapter->msgRingPA);
1264 }
1265
1266 /*
1267  * Allocate scatter gather lists.
1268  *
1269  * These are statically allocated.  Trying to be clever was not worth it.
1270  *
1271  * Dynamic allocation can fail, and we can't go deep into the memory
1272  * allocator, since we're a SCSI driver, and trying too hard to allocate
1273  * memory might generate disk I/O.  We also don't want to fail disk I/O
1274  * in that case because we can't get an allocation - the I/O could be
1275  * trying to swap out data to free memory.  Since that is pathological,
1276  * just use a statically allocated scatter list.
1277  *
1278  */
1279 static int pvscsi_allocate_sg(struct pvscsi_adapter *adapter)
1280 {
1281         struct pvscsi_ctx *ctx;
1282         int i;
1283
1284         ctx = adapter->cmd_map;
1285         BUILD_BUG_ON(sizeof(struct pvscsi_sg_list) > SGL_SIZE);
1286
1287         for (i = 0; i < adapter->req_depth; ++i, ++ctx) {
1288                 ctx->sgl = (void *)__get_free_pages(GFP_KERNEL,
1289                                                     get_order(SGL_SIZE));
1290                 ctx->sglPA = 0;
1291                 BUG_ON(!IS_ALIGNED(((unsigned long)ctx->sgl), PAGE_SIZE));
1292                 if (!ctx->sgl) {
1293                         for (; i >= 0; --i, --ctx) {
1294                                 free_pages((unsigned long)ctx->sgl,
1295                                            get_order(SGL_SIZE));
1296                                 ctx->sgl = NULL;
1297                         }
1298                         return -ENOMEM;
1299                 }
1300         }
1301
1302         return 0;
1303 }
1304
1305 /*
1306  * Query the device, fetch the config info and return the
1307  * maximum number of targets on the adapter. In case of
1308  * failure due to any reason return default i.e. 16.
1309  */
1310 static u32 pvscsi_get_max_targets(struct pvscsi_adapter *adapter)
1311 {
1312         struct PVSCSICmdDescConfigCmd cmd;
1313         struct PVSCSIConfigPageHeader *header;
1314         struct device *dev;
1315         dma_addr_t configPagePA;
1316         void *config_page;
1317         u32 numPhys = 16;
1318
1319         dev = pvscsi_dev(adapter);
1320         config_page = pci_alloc_consistent(adapter->dev, PAGE_SIZE,
1321                                            &configPagePA);
1322         if (!config_page) {
1323                 dev_warn(dev, "vmw_pvscsi: failed to allocate memory for config page\n");
1324                 goto exit;
1325         }
1326         BUG_ON(configPagePA & ~PAGE_MASK);
1327
1328         /* Fetch config info from the device. */
1329         cmd.configPageAddress = ((u64)PVSCSI_CONFIG_CONTROLLER_ADDRESS) << 32;
1330         cmd.configPageNum = PVSCSI_CONFIG_PAGE_CONTROLLER;
1331         cmd.cmpAddr = configPagePA;
1332         cmd._pad = 0;
1333
1334         /*
1335          * Mark the completion page header with error values. If the device
1336          * completes the command successfully, it sets the status values to
1337          * indicate success.
1338          */
1339         header = config_page;
1340         memset(header, 0, sizeof *header);
1341         header->hostStatus = BTSTAT_INVPARAM;
1342         header->scsiStatus = SDSTAT_CHECK;
1343
1344         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_CONFIG, &cmd, sizeof cmd);
1345
1346         if (header->hostStatus == BTSTAT_SUCCESS &&
1347             header->scsiStatus == SDSTAT_GOOD) {
1348                 struct PVSCSIConfigPageController *config;
1349
1350                 config = config_page;
1351                 numPhys = config->numPhys;
1352         } else
1353                 dev_warn(dev, "vmw_pvscsi: PVSCSI_CMD_CONFIG failed. hostStatus = 0x%x, scsiStatus = 0x%x\n",
1354                          header->hostStatus, header->scsiStatus);
1355         pci_free_consistent(adapter->dev, PAGE_SIZE, config_page, configPagePA);
1356 exit:
1357         return numPhys;
1358 }
1359
1360 static int pvscsi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1361 {
1362         struct pvscsi_adapter *adapter;
1363         struct pvscsi_adapter adapter_temp;
1364         struct Scsi_Host *host = NULL;
1365         unsigned int i;
1366         unsigned long flags = 0;
1367         int error;
1368         u32 max_id;
1369
1370         error = -ENODEV;
1371
1372         if (pci_enable_device(pdev))
1373                 return error;
1374
1375         if (pci_set_dma_mask(pdev, DMA_BIT_MASK(64)) == 0 &&
1376             pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64)) == 0) {
1377                 printk(KERN_INFO "vmw_pvscsi: using 64bit dma\n");
1378         } else if (pci_set_dma_mask(pdev, DMA_BIT_MASK(32)) == 0 &&
1379                    pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32)) == 0) {
1380                 printk(KERN_INFO "vmw_pvscsi: using 32bit dma\n");
1381         } else {
1382                 printk(KERN_ERR "vmw_pvscsi: failed to set DMA mask\n");
1383                 goto out_disable_device;
1384         }
1385
1386         /*
1387          * Let's use a temp pvscsi_adapter struct until we find the number of
1388          * targets on the adapter, after that we will switch to the real
1389          * allocated struct.
1390          */
1391         adapter = &adapter_temp;
1392         memset(adapter, 0, sizeof(*adapter));
1393         adapter->dev  = pdev;
1394         adapter->rev = pdev->revision;
1395
1396         if (pci_request_regions(pdev, "vmw_pvscsi")) {
1397                 printk(KERN_ERR "vmw_pvscsi: pci memory selection failed\n");
1398                 goto out_disable_device;
1399         }
1400
1401         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
1402                 if ((pci_resource_flags(pdev, i) & PCI_BASE_ADDRESS_SPACE_IO))
1403                         continue;
1404
1405                 if (pci_resource_len(pdev, i) < PVSCSI_MEM_SPACE_SIZE)
1406                         continue;
1407
1408                 break;
1409         }
1410
1411         if (i == DEVICE_COUNT_RESOURCE) {
1412                 printk(KERN_ERR
1413                        "vmw_pvscsi: adapter has no suitable MMIO region\n");
1414                 goto out_release_resources_and_disable;
1415         }
1416
1417         adapter->mmioBase = pci_iomap(pdev, i, PVSCSI_MEM_SPACE_SIZE);
1418
1419         if (!adapter->mmioBase) {
1420                 printk(KERN_ERR
1421                        "vmw_pvscsi: can't iomap for BAR %d memsize %lu\n",
1422                        i, PVSCSI_MEM_SPACE_SIZE);
1423                 goto out_release_resources_and_disable;
1424         }
1425
1426         pci_set_master(pdev);
1427
1428         /*
1429          * Ask the device for max number of targets before deciding the
1430          * default pvscsi_ring_pages value.
1431          */
1432         max_id = pvscsi_get_max_targets(adapter);
1433         printk(KERN_INFO "vmw_pvscsi: max_id: %u\n", max_id);
1434
1435         if (pvscsi_ring_pages == 0)
1436                 /*
1437                  * Set the right default value. Up to 16 it is 8, above it is
1438                  * max.
1439                  */
1440                 pvscsi_ring_pages = (max_id > 16) ?
1441                         PVSCSI_SETUP_RINGS_MAX_NUM_PAGES :
1442                         PVSCSI_DEFAULT_NUM_PAGES_PER_RING;
1443         printk(KERN_INFO
1444                "vmw_pvscsi: setting ring_pages to %d\n",
1445                pvscsi_ring_pages);
1446
1447         pvscsi_template.can_queue =
1448                 min(PVSCSI_MAX_NUM_PAGES_REQ_RING, pvscsi_ring_pages) *
1449                 PVSCSI_MAX_NUM_REQ_ENTRIES_PER_PAGE;
1450         pvscsi_template.cmd_per_lun =
1451                 min(pvscsi_template.can_queue, pvscsi_cmd_per_lun);
1452         host = scsi_host_alloc(&pvscsi_template, sizeof(struct pvscsi_adapter));
1453         if (!host) {
1454                 printk(KERN_ERR "vmw_pvscsi: failed to allocate host\n");
1455                 goto out_release_resources_and_disable;
1456         }
1457
1458         /*
1459          * Let's use the real pvscsi_adapter struct here onwards.
1460          */
1461         adapter = shost_priv(host);
1462         memset(adapter, 0, sizeof(*adapter));
1463         adapter->dev  = pdev;
1464         adapter->host = host;
1465         /*
1466          * Copy back what we already have to the allocated adapter struct.
1467          */
1468         adapter->rev = adapter_temp.rev;
1469         adapter->mmioBase = adapter_temp.mmioBase;
1470
1471         spin_lock_init(&adapter->hw_lock);
1472         host->max_channel = 0;
1473         host->max_lun     = 1;
1474         host->max_cmd_len = 16;
1475         host->max_id      = max_id;
1476
1477         pci_set_drvdata(pdev, host);
1478
1479         ll_adapter_reset(adapter);
1480
1481         adapter->use_msg = pvscsi_setup_msg_workqueue(adapter);
1482
1483         error = pvscsi_allocate_rings(adapter);
1484         if (error) {
1485                 printk(KERN_ERR "vmw_pvscsi: unable to allocate ring memory\n");
1486                 goto out_release_resources;
1487         }
1488
1489         /*
1490          * From this point on we should reset the adapter if anything goes
1491          * wrong.
1492          */
1493         pvscsi_setup_all_rings(adapter);
1494
1495         adapter->cmd_map = kcalloc(adapter->req_depth,
1496                                    sizeof(struct pvscsi_ctx), GFP_KERNEL);
1497         if (!adapter->cmd_map) {
1498                 printk(KERN_ERR "vmw_pvscsi: failed to allocate memory.\n");
1499                 error = -ENOMEM;
1500                 goto out_reset_adapter;
1501         }
1502
1503         INIT_LIST_HEAD(&adapter->cmd_pool);
1504         for (i = 0; i < adapter->req_depth; i++) {
1505                 struct pvscsi_ctx *ctx = adapter->cmd_map + i;
1506                 list_add(&ctx->list, &adapter->cmd_pool);
1507         }
1508
1509         error = pvscsi_allocate_sg(adapter);
1510         if (error) {
1511                 printk(KERN_ERR "vmw_pvscsi: unable to allocate s/g table\n");
1512                 goto out_reset_adapter;
1513         }
1514
1515         if (!pvscsi_disable_msix &&
1516             pvscsi_setup_msix(adapter, &adapter->irq) == 0) {
1517                 printk(KERN_INFO "vmw_pvscsi: using MSI-X\n");
1518                 adapter->use_msix = 1;
1519         } else if (!pvscsi_disable_msi && pci_enable_msi(pdev) == 0) {
1520                 printk(KERN_INFO "vmw_pvscsi: using MSI\n");
1521                 adapter->use_msi = 1;
1522                 adapter->irq = pdev->irq;
1523         } else {
1524                 printk(KERN_INFO "vmw_pvscsi: using INTx\n");
1525                 adapter->irq = pdev->irq;
1526                 flags = IRQF_SHARED;
1527         }
1528
1529         adapter->use_req_threshold = pvscsi_setup_req_threshold(adapter, true);
1530         printk(KERN_DEBUG "vmw_pvscsi: driver-based request coalescing %sabled\n",
1531                adapter->use_req_threshold ? "en" : "dis");
1532
1533         error = request_irq(adapter->irq, pvscsi_isr, flags,
1534                             "vmw_pvscsi", adapter);
1535         if (error) {
1536                 printk(KERN_ERR
1537                        "vmw_pvscsi: unable to request IRQ: %d\n", error);
1538                 adapter->irq = 0;
1539                 goto out_reset_adapter;
1540         }
1541
1542         error = scsi_add_host(host, &pdev->dev);
1543         if (error) {
1544                 printk(KERN_ERR
1545                        "vmw_pvscsi: scsi_add_host failed: %d\n", error);
1546                 goto out_reset_adapter;
1547         }
1548
1549         dev_info(&pdev->dev, "VMware PVSCSI rev %d host #%u\n",
1550                  adapter->rev, host->host_no);
1551
1552         pvscsi_unmask_intr(adapter);
1553
1554         scsi_scan_host(host);
1555
1556         return 0;
1557
1558 out_reset_adapter:
1559         ll_adapter_reset(adapter);
1560 out_release_resources:
1561         pvscsi_release_resources(adapter);
1562         scsi_host_put(host);
1563 out_disable_device:
1564         pci_disable_device(pdev);
1565
1566         return error;
1567
1568 out_release_resources_and_disable:
1569         pvscsi_release_resources(adapter);
1570         goto out_disable_device;
1571 }
1572
1573 static void __pvscsi_shutdown(struct pvscsi_adapter *adapter)
1574 {
1575         pvscsi_mask_intr(adapter);
1576
1577         if (adapter->workqueue)
1578                 flush_workqueue(adapter->workqueue);
1579
1580         pvscsi_shutdown_intr(adapter);
1581
1582         pvscsi_process_request_ring(adapter);
1583         pvscsi_process_completion_ring(adapter);
1584         ll_adapter_reset(adapter);
1585 }
1586
1587 static void pvscsi_shutdown(struct pci_dev *dev)
1588 {
1589         struct Scsi_Host *host = pci_get_drvdata(dev);
1590         struct pvscsi_adapter *adapter = shost_priv(host);
1591
1592         __pvscsi_shutdown(adapter);
1593 }
1594
1595 static void pvscsi_remove(struct pci_dev *pdev)
1596 {
1597         struct Scsi_Host *host = pci_get_drvdata(pdev);
1598         struct pvscsi_adapter *adapter = shost_priv(host);
1599
1600         scsi_remove_host(host);
1601
1602         __pvscsi_shutdown(adapter);
1603         pvscsi_release_resources(adapter);
1604
1605         scsi_host_put(host);
1606
1607         pci_disable_device(pdev);
1608 }
1609
1610 static struct pci_driver pvscsi_pci_driver = {
1611         .name           = "vmw_pvscsi",
1612         .id_table       = pvscsi_pci_tbl,
1613         .probe          = pvscsi_probe,
1614         .remove         = pvscsi_remove,
1615         .shutdown       = pvscsi_shutdown,
1616 };
1617
1618 static int __init pvscsi_init(void)
1619 {
1620         pr_info("%s - version %s\n",
1621                 PVSCSI_LINUX_DRIVER_DESC, PVSCSI_DRIVER_VERSION_STRING);
1622         return pci_register_driver(&pvscsi_pci_driver);
1623 }
1624
1625 static void __exit pvscsi_exit(void)
1626 {
1627         pci_unregister_driver(&pvscsi_pci_driver);
1628 }
1629
1630 module_init(pvscsi_init);
1631 module_exit(pvscsi_exit);