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                 if (sdstat == SAM_STAT_COMMAND_TERMINATED) {
 568                         cmd->result = (DID_RESET << 16);
 569                 } else {
 570                         cmd->result = (DID_OK << 16) | sdstat;
 571                         if (sdstat == SAM_STAT_CHECK_CONDITION &&
 572                             cmd->sense_buffer)
 573                                 cmd->result |= (DRIVER_SENSE << 24);
 574                 }
 575         } else
 576                 switch (btstat) {
 577                 case BTSTAT_SUCCESS:
 578                 case BTSTAT_LINKED_COMMAND_COMPLETED:
 579                 case BTSTAT_LINKED_COMMAND_COMPLETED_WITH_FLAG:
 580                         /*
 581                          * Commands like INQUIRY may transfer less data than
 582                          * requested by the initiator via bufflen. Set residual
 583                          * count to make upper layer aware of the actual amount
 584                          * of data returned. There are cases when controller
 585                          * returns zero dataLen with non zero data - do not set
 586                          * residual count in that case.
 587                          */
 588                         if (e->dataLen && (e->dataLen < scsi_bufflen(cmd)))
 589                                 scsi_set_resid(cmd, scsi_bufflen(cmd) - e->dataLen);
 590                         cmd->result = (DID_OK << 16);
 591                         break;
 592
 593                 case BTSTAT_DATARUN:
 594                 case BTSTAT_DATA_UNDERRUN:
 595                         /* Report residual data in underruns */
 596                         scsi_set_resid(cmd, scsi_bufflen(cmd) - e->dataLen);
 597                         cmd->result = (DID_ERROR << 16);
 598                         break;
 599
 600                 case BTSTAT_SELTIMEO:
 601                         /* Our emulation returns this for non-connected devs */
 602                         cmd->result = (DID_BAD_TARGET << 16);
 603                         break;
 604
 605                 case BTSTAT_LUNMISMATCH:
 606                 case BTSTAT_TAGREJECT:
 607                 case BTSTAT_BADMSG:
 608                         cmd->result = (DRIVER_INVALID << 24);
 609                         /* fall through */
 610
 611                 case BTSTAT_HAHARDWARE:
 612                 case BTSTAT_INVPHASE:
 613                 case BTSTAT_HATIMEOUT:
 614                 case BTSTAT_NORESPONSE:
 615                 case BTSTAT_DISCONNECT:
 616                 case BTSTAT_HASOFTWARE:
 617                 case BTSTAT_BUSFREE:
 618                 case BTSTAT_SENSFAILED:
 619                         cmd->result |= (DID_ERROR << 16);
 620                         break;
 621
 622                 case BTSTAT_SENTRST:
 623                 case BTSTAT_RECVRST:
 624                 case BTSTAT_BUSRESET:
 625                         cmd->result = (DID_RESET << 16);
 626                         break;
 627
 628                 case BTSTAT_ABORTQUEUE:
 629                         cmd->result = (DID_ABORT << 16);
 630                         break;
 631
 632                 case BTSTAT_SCSIPARITY:
 633                         cmd->result = (DID_PARITY << 16);
 634                         break;
 635
 636                 default:
 637                         cmd->result = (DID_ERROR << 16);
 638                         scmd_printk(KERN_DEBUG, cmd,
 639                                     "Unknown completion status: 0x%x\n",
 640                                     btstat);
 641         }
 642
 643         dev_dbg(&cmd->device->sdev_gendev,
 644                 "cmd=%p %x ctx=%p result=0x%x status=0x%x,%x\n",
 645                 cmd, cmd->cmnd[0], ctx, cmd->result, btstat, sdstat);
 646
 647         cmd->scsi_done(cmd);
 648 }
 649
 650 /*
 651  * barrier usage : Since the PVSCSI device is emulated, there could be cases
 652  * where we may want to serialize some accesses between the driver and the
 653  * emulation layer. We use compiler barriers instead of the more expensive
 654  * memory barriers because PVSCSI is only supported on X86 which has strong
 655  * memory access ordering.
 656  */
 657 static void pvscsi_process_completion_ring(struct pvscsi_adapter *adapter)
 658 {
 659         struct PVSCSIRingsState *s = adapter->rings_state;
 660         struct PVSCSIRingCmpDesc *ring = adapter->cmp_ring;
 661         u32 cmp_entries = s->cmpNumEntriesLog2;
 662
 663         while (s->cmpConsIdx != s->cmpProdIdx) {
 664                 struct PVSCSIRingCmpDesc *e = ring + (s->cmpConsIdx &
 665                                                       MASK(cmp_entries));
 666                 /*
 667                  * This barrier() ensures that *e is not dereferenced while
 668                  * the device emulation still writes data into the slot.
 669                  * Since the device emulation advances s->cmpProdIdx only after
 670                  * updating the slot we want to check it first.
 671                  */
 672                 barrier();
 673                 pvscsi_complete_request(adapter, e);
 674                 /*
 675                  * This barrier() ensures that compiler doesn't reorder write
 676                  * to s->cmpConsIdx before the read of (*e) inside
 677                  * pvscsi_complete_request. Otherwise, device emulation may
 678                  * overwrite *e before we had a chance to read it.
 679                  */
 680                 barrier();
 681                 s->cmpConsIdx++;
 682         }
 683 }
 684
 685 /*
 686  * Translate a Linux SCSI request into a request ring entry.
 687  */
 688 static int pvscsi_queue_ring(struct pvscsi_adapter *adapter,
 689                              struct pvscsi_ctx *ctx, struct scsi_cmnd *cmd)
 690 {
 691         struct PVSCSIRingsState *s;
 692         struct PVSCSIRingReqDesc *e;
 693         struct scsi_device *sdev;
 694         u32 req_entries;
 695
 696         s = adapter->rings_state;
 697         sdev = cmd->device;
 698         req_entries = s->reqNumEntriesLog2;
 699
 700         /*
 701          * If this condition holds, we might have room on the request ring, but
 702          * we might not have room on the completion ring for the response.
 703          * However, we have already ruled out this possibility - we would not
 704          * have successfully allocated a context if it were true, since we only
 705          * have one context per request entry.  Check for it anyway, since it
 706          * would be a serious bug.
 707          */
 708         if (s->reqProdIdx - s->cmpConsIdx >= 1 << req_entries) {
 709                 scmd_printk(KERN_ERR, cmd, "vmw_pvscsi: "
 710                             "ring full: reqProdIdx=%d cmpConsIdx=%d\n",
 711                             s->reqProdIdx, s->cmpConsIdx);
 712                 return -1;
 713         }
 714
 715         e = adapter->req_ring + (s->reqProdIdx & MASK(req_entries));
 716
 717         e->bus    = sdev->channel;
 718         e->target = sdev->id;
 719         memset(e->lun, 0, sizeof(e->lun));
 720         e->lun[1] = sdev->lun;
 721
 722         if (cmd->sense_buffer) {
 723                 ctx->sensePA = pci_map_single(adapter->dev, cmd->sense_buffer,
 724                                               SCSI_SENSE_BUFFERSIZE,
 725                                               PCI_DMA_FROMDEVICE);
 726                 if (pci_dma_mapping_error(adapter->dev, ctx->sensePA)) {
 727                         scmd_printk(KERN_ERR, cmd,
 728                                     "vmw_pvscsi: Failed to map sense buffer for DMA.\n");
 729                         ctx->sensePA = 0;
 730                         return -ENOMEM;
 731                 }
 732                 e->senseAddr = ctx->sensePA;
 733                 e->senseLen = SCSI_SENSE_BUFFERSIZE;
 734         } else {
 735                 e->senseLen  = 0;
 736                 e->senseAddr = 0;
 737         }
 738         e->cdbLen   = cmd->cmd_len;
 739         e->vcpuHint = smp_processor_id();
 740         memcpy(e->cdb, cmd->cmnd, e->cdbLen);
 741
 742         e->tag = SIMPLE_QUEUE_TAG;
 743
 744         if (cmd->sc_data_direction == DMA_FROM_DEVICE)
 745                 e->flags = PVSCSI_FLAG_CMD_DIR_TOHOST;
 746         else if (cmd->sc_data_direction == DMA_TO_DEVICE)
 747                 e->flags = PVSCSI_FLAG_CMD_DIR_TODEVICE;
 748         else if (cmd->sc_data_direction == DMA_NONE)
 749                 e->flags = PVSCSI_FLAG_CMD_DIR_NONE;
 750         else
 751                 e->flags = 0;
 752
 753         if (pvscsi_map_buffers(adapter, ctx, cmd, e) != 0) {
 754                 if (cmd->sense_buffer) {
 755                         pci_unmap_single(adapter->dev, ctx->sensePA,
 756                                          SCSI_SENSE_BUFFERSIZE,
 757                                          PCI_DMA_FROMDEVICE);
 758                         ctx->sensePA = 0;
 759                 }
 760                 return -ENOMEM;
 761         }
 762
 763         e->context = pvscsi_map_context(adapter, ctx);
 764
 765         barrier();
 766
 767         s->reqProdIdx++;
 768
 769         return 0;
 770 }
 771
 772 static int pvscsi_queue_lck(struct scsi_cmnd *cmd, void (*done)(struct scsi_cmnd *))
 773 {
 774         struct Scsi_Host *host = cmd->device->host;
 775         struct pvscsi_adapter *adapter = shost_priv(host);
 776         struct pvscsi_ctx *ctx;
 777         unsigned long flags;
 778         unsigned char op;
 779
 780         spin_lock_irqsave(&adapter->hw_lock, flags);
 781
 782         ctx = pvscsi_acquire_context(adapter, cmd);
 783         if (!ctx || pvscsi_queue_ring(adapter, ctx, cmd) != 0) {
 784                 if (ctx)
 785                         pvscsi_release_context(adapter, ctx);
 786                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 787                 return SCSI_MLQUEUE_HOST_BUSY;
 788         }
 789
 790         cmd->scsi_done = done;
 791         op = cmd->cmnd[0];
 792
 793         dev_dbg(&cmd->device->sdev_gendev,
 794                 "queued cmd %p, ctx %p, op=%x\n", cmd, ctx, op);
 795
 796         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 797
 798         pvscsi_kick_io(adapter, op);
 799
 800         return 0;
 801 }
 802
 803 static DEF_SCSI_QCMD(pvscsi_queue)
 804
 805 static int pvscsi_abort(struct scsi_cmnd *cmd)
 806 {
 807         struct pvscsi_adapter *adapter = shost_priv(cmd->device->host);
 808         struct pvscsi_ctx *ctx;
 809         unsigned long flags;
 810         int result = SUCCESS;
 811         DECLARE_COMPLETION_ONSTACK(abort_cmp);
 812         int done;
 813
 814         scmd_printk(KERN_DEBUG, cmd, "task abort on host %u, %p\n",
 815                     adapter->host->host_no, cmd);
 816
 817         spin_lock_irqsave(&adapter->hw_lock, flags);
 818
 819         /*
 820          * Poll the completion ring first - we might be trying to abort
 821          * a command that is waiting to be dispatched in the completion ring.
 822          */
 823         pvscsi_process_completion_ring(adapter);
 824
 825         /*
 826          * If there is no context for the command, it either already succeeded
 827          * or else was never properly issued.  Not our problem.
 828          */
 829         ctx = pvscsi_find_context(adapter, cmd);
 830         if (!ctx) {
 831                 scmd_printk(KERN_DEBUG, cmd, "Failed to abort cmd %p\n", cmd);
 832                 goto out;
 833         }
 834
 835         /*
 836          * Mark that the command has been requested to be aborted and issue
 837          * the abort.
 838          */
 839         ctx->abort_cmp = &abort_cmp;
 840
 841         pvscsi_abort_cmd(adapter, ctx);
 842         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 843         /* Wait for 2 secs for the completion. */
 844         done = wait_for_completion_timeout(&abort_cmp, msecs_to_jiffies(2000));
 845         spin_lock_irqsave(&adapter->hw_lock, flags);
 846
 847         if (!done) {
 848                 /*
 849                  * Failed to abort the command, unmark the fact that it
 850                  * was requested to be aborted.
 851                  */
 852                 ctx->abort_cmp = NULL;
 853                 result = FAILED;
 854                 scmd_printk(KERN_DEBUG, cmd,
 855                             "Failed to get completion for aborted cmd %p\n",
 856                             cmd);
 857                 goto out;
 858         }
 859
 860         /*
 861          * Successfully aborted the command.
 862          */
 863         cmd->result = (DID_ABORT << 16);
 864         cmd->scsi_done(cmd);
 865
 866 out:
 867         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 868         return result;
 869 }
 870
 871 /*
 872  * Abort all outstanding requests.  This is only safe to use if the completion
 873  * ring will never be walked again or the device has been reset, because it
 874  * destroys the 1-1 mapping between context field passed to emulation and our
 875  * request structure.
 876  */
 877 static void pvscsi_reset_all(struct pvscsi_adapter *adapter)
 878 {
 879         unsigned i;
 880
 881         for (i = 0; i < adapter->req_depth; i++) {
 882                 struct pvscsi_ctx *ctx = &adapter->cmd_map[i];
 883                 struct scsi_cmnd *cmd = ctx->cmd;
 884                 if (cmd) {
 885                         scmd_printk(KERN_ERR, cmd,
 886                                     "Forced reset on cmd %p\n", cmd);
 887                         pvscsi_unmap_buffers(adapter, ctx);
 888                         pvscsi_release_context(adapter, ctx);
 889                         cmd->result = (DID_RESET << 16);
 890                         cmd->scsi_done(cmd);
 891                 }
 892         }
 893 }
 894
 895 static int pvscsi_host_reset(struct scsi_cmnd *cmd)
 896 {
 897         struct Scsi_Host *host = cmd->device->host;
 898         struct pvscsi_adapter *adapter = shost_priv(host);
 899         unsigned long flags;
 900         bool use_msg;
 901
 902         scmd_printk(KERN_INFO, cmd, "SCSI Host reset\n");
 903
 904         spin_lock_irqsave(&adapter->hw_lock, flags);
 905
 906         use_msg = adapter->use_msg;
 907
 908         if (use_msg) {
 909                 adapter->use_msg = 0;
 910                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 911
 912                 /*
 913                  * Now that we know that the ISR won't add more work on the
 914                  * workqueue we can safely flush any outstanding work.
 915                  */
 916                 flush_workqueue(adapter->workqueue);
 917                 spin_lock_irqsave(&adapter->hw_lock, flags);
 918         }
 919
 920         /*
 921          * We're going to tear down the entire ring structure and set it back
 922          * up, so stalling new requests until all completions are flushed and
 923          * the rings are back in place.
 924          */
 925
 926         pvscsi_process_request_ring(adapter);
 927
 928         ll_adapter_reset(adapter);
 929
 930         /*
 931          * Now process any completions.  Note we do this AFTER adapter reset,
 932          * which is strange, but stops races where completions get posted
 933          * between processing the ring and issuing the reset.  The backend will
 934          * not touch the ring memory after reset, so the immediately pre-reset
 935          * completion ring state is still valid.
 936          */
 937         pvscsi_process_completion_ring(adapter);
 938
 939         pvscsi_reset_all(adapter);
 940         adapter->use_msg = use_msg;
 941         pvscsi_setup_all_rings(adapter);
 942         pvscsi_unmask_intr(adapter);
 943
 944         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 945
 946         return SUCCESS;
 947 }
 948
 949 static int pvscsi_bus_reset(struct scsi_cmnd *cmd)
 950 {
 951         struct Scsi_Host *host = cmd->device->host;
 952         struct pvscsi_adapter *adapter = shost_priv(host);
 953         unsigned long flags;
 954
 955         scmd_printk(KERN_INFO, cmd, "SCSI Bus reset\n");
 956
 957         /*
 958          * We don't want to queue new requests for this bus after
 959          * flushing all pending requests to emulation, since new
 960          * requests could then sneak in during this bus reset phase,
 961          * so take the lock now.
 962          */
 963         spin_lock_irqsave(&adapter->hw_lock, flags);
 964
 965         pvscsi_process_request_ring(adapter);
 966         ll_bus_reset(adapter);
 967         pvscsi_process_completion_ring(adapter);
 968
 969         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 970
 971         return SUCCESS;
 972 }
 973
 974 static int pvscsi_device_reset(struct scsi_cmnd *cmd)
 975 {
 976         struct Scsi_Host *host = cmd->device->host;
 977         struct pvscsi_adapter *adapter = shost_priv(host);
 978         unsigned long flags;
 979
 980         scmd_printk(KERN_INFO, cmd, "SCSI device reset on scsi%u:%u\n",
 981                     host->host_no, cmd->device->id);
 982
 983         /*
 984          * We don't want to queue new requests for this device after flushing
 985          * all pending requests to emulation, since new requests could then
 986          * sneak in during this device reset phase, so take the lock now.
 987          */
 988         spin_lock_irqsave(&adapter->hw_lock, flags);
 989
 990         pvscsi_process_request_ring(adapter);
 991         ll_device_reset(adapter, cmd->device->id);
 992         pvscsi_process_completion_ring(adapter);
 993
 994         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 995
 996         return SUCCESS;
 997 }
 998
 999 static struct scsi_host_template pvscsi_template;
1000
1001 static const char *pvscsi_info(struct Scsi_Host *host)
1002 {
1003         struct pvscsi_adapter *adapter = shost_priv(host);
1004         static char buf[256];
1005
1006         sprintf(buf, "VMware PVSCSI storage adapter rev %d, req/cmp/msg rings: "
1007                 "%u/%u/%u pages, cmd_per_lun=%u", adapter->rev,
1008                 adapter->req_pages, adapter->cmp_pages, adapter->msg_pages,
1009                 pvscsi_template.cmd_per_lun);
1010
1011         return buf;
1012 }
1013
1014 static struct scsi_host_template pvscsi_template = {
1015         .module                         = THIS_MODULE,
1016         .name                           = "VMware PVSCSI Host Adapter",
1017         .proc_name                      = "vmw_pvscsi",
1018         .info                           = pvscsi_info,
1019         .queuecommand                   = pvscsi_queue,
1020         .this_id                        = -1,
1021         .sg_tablesize                   = PVSCSI_MAX_NUM_SG_ENTRIES_PER_SEGMENT,
1022         .dma_boundary                   = UINT_MAX,
1023         .max_sectors                    = 0xffff,
1024         .use_clustering                 = ENABLE_CLUSTERING,
1025         .change_queue_depth             = pvscsi_change_queue_depth,
1026         .eh_abort_handler               = pvscsi_abort,
1027         .eh_device_reset_handler        = pvscsi_device_reset,
1028         .eh_bus_reset_handler           = pvscsi_bus_reset,
1029         .eh_host_reset_handler          = pvscsi_host_reset,
1030 };
1031
1032 static void pvscsi_process_msg(const struct pvscsi_adapter *adapter,
1033                                const struct PVSCSIRingMsgDesc *e)
1034 {
1035         struct PVSCSIRingsState *s = adapter->rings_state;
1036         struct Scsi_Host *host = adapter->host;
1037         struct scsi_device *sdev;
1038
1039         printk(KERN_INFO "vmw_pvscsi: msg type: 0x%x - MSG RING: %u/%u (%u) \n",
1040                e->type, s->msgProdIdx, s->msgConsIdx, s->msgNumEntriesLog2);
1041
1042         BUILD_BUG_ON(PVSCSI_MSG_LAST != 2);
1043
1044         if (e->type == PVSCSI_MSG_DEV_ADDED) {
1045                 struct PVSCSIMsgDescDevStatusChanged *desc;
1046                 desc = (struct PVSCSIMsgDescDevStatusChanged *)e;
1047
1048                 printk(KERN_INFO
1049                        "vmw_pvscsi: msg: device added at scsi%u:%u:%u\n",
1050                        desc->bus, desc->target, desc->lun[1]);
1051
1052                 if (!scsi_host_get(host))
1053                         return;
1054
1055                 sdev = scsi_device_lookup(host, desc->bus, desc->target,
1056                                           desc->lun[1]);
1057                 if (sdev) {
1058                         printk(KERN_INFO "vmw_pvscsi: device already exists\n");
1059                         scsi_device_put(sdev);
1060                 } else
1061                         scsi_add_device(adapter->host, desc->bus,
1062                                         desc->target, desc->lun[1]);
1063
1064                 scsi_host_put(host);
1065         } else if (e->type == PVSCSI_MSG_DEV_REMOVED) {
1066                 struct PVSCSIMsgDescDevStatusChanged *desc;
1067                 desc = (struct PVSCSIMsgDescDevStatusChanged *)e;
1068
1069                 printk(KERN_INFO
1070                        "vmw_pvscsi: msg: device removed at scsi%u:%u:%u\n",
1071                        desc->bus, desc->target, desc->lun[1]);
1072
1073                 if (!scsi_host_get(host))
1074                         return;
1075
1076                 sdev = scsi_device_lookup(host, desc->bus, desc->target,
1077                                           desc->lun[1]);
1078                 if (sdev) {
1079                         scsi_remove_device(sdev);
1080                         scsi_device_put(sdev);
1081                 } else
1082                         printk(KERN_INFO
1083                                "vmw_pvscsi: failed to lookup scsi%u:%u:%u\n",
1084                                desc->bus, desc->target, desc->lun[1]);
1085
1086                 scsi_host_put(host);
1087         }
1088 }
1089
1090 static int pvscsi_msg_pending(const struct pvscsi_adapter *adapter)
1091 {
1092         struct PVSCSIRingsState *s = adapter->rings_state;
1093
1094         return s->msgProdIdx != s->msgConsIdx;
1095 }
1096
1097 static void pvscsi_process_msg_ring(const struct pvscsi_adapter *adapter)
1098 {
1099         struct PVSCSIRingsState *s = adapter->rings_state;
1100         struct PVSCSIRingMsgDesc *ring = adapter->msg_ring;
1101         u32 msg_entries = s->msgNumEntriesLog2;
1102
1103         while (pvscsi_msg_pending(adapter)) {
1104                 struct PVSCSIRingMsgDesc *e = ring + (s->msgConsIdx &
1105                                                       MASK(msg_entries));
1106
1107                 barrier();
1108                 pvscsi_process_msg(adapter, e);
1109                 barrier();
1110                 s->msgConsIdx++;
1111         }
1112 }
1113
1114 static void pvscsi_msg_workqueue_handler(struct work_struct *data)
1115 {
1116         struct pvscsi_adapter *adapter;
1117
1118         adapter = container_of(data, struct pvscsi_adapter, work);
1119
1120         pvscsi_process_msg_ring(adapter);
1121 }
1122
1123 static int pvscsi_setup_msg_workqueue(struct pvscsi_adapter *adapter)
1124 {
1125         char name[32];
1126
1127         if (!pvscsi_use_msg)
1128                 return 0;
1129
1130         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
1131                          PVSCSI_CMD_SETUP_MSG_RING);
1132
1133         if (pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS) == -1)
1134                 return 0;
1135
1136         snprintf(name, sizeof(name),
1137                  "vmw_pvscsi_wq_%u", adapter->host->host_no);
1138
1139         adapter->workqueue = create_singlethread_workqueue(name);
1140         if (!adapter->workqueue) {
1141                 printk(KERN_ERR "vmw_pvscsi: failed to create work queue\n");
1142                 return 0;
1143         }
1144         INIT_WORK(&adapter->work, pvscsi_msg_workqueue_handler);
1145
1146         return 1;
1147 }
1148
1149 static bool pvscsi_setup_req_threshold(struct pvscsi_adapter *adapter,
1150                                       bool enable)
1151 {
1152         u32 val;
1153
1154         if (!pvscsi_use_req_threshold)
1155                 return false;
1156
1157         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
1158                          PVSCSI_CMD_SETUP_REQCALLTHRESHOLD);
1159         val = pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS);
1160         if (val == -1) {
1161                 printk(KERN_INFO "vmw_pvscsi: device does not support req_threshold\n");
1162                 return false;
1163         } else {
1164                 struct PVSCSICmdDescSetupReqCall cmd_msg = { 0 };
1165                 cmd_msg.enable = enable;
1166                 printk(KERN_INFO
1167                        "vmw_pvscsi: %sabling reqCallThreshold\n",
1168                         enable ? "en" : "dis");
1169                 pvscsi_write_cmd_desc(adapter,
1170                                       PVSCSI_CMD_SETUP_REQCALLTHRESHOLD,
1171                                       &cmd_msg, sizeof(cmd_msg));
1172                 return pvscsi_reg_read(adapter,
1173                                        PVSCSI_REG_OFFSET_COMMAND_STATUS) != 0;
1174         }
1175 }
1176
1177 static irqreturn_t pvscsi_isr(int irq, void *devp)
1178 {
1179         struct pvscsi_adapter *adapter = devp;
1180         int handled;
1181
1182         if (adapter->use_msi || adapter->use_msix)
1183                 handled = true;
1184         else {
1185                 u32 val = pvscsi_read_intr_status(adapter);
1186                 handled = (val & PVSCSI_INTR_ALL_SUPPORTED) != 0;
1187                 if (handled)
1188                         pvscsi_write_intr_status(devp, val);
1189         }
1190
1191         if (handled) {
1192                 unsigned long flags;
1193
1194                 spin_lock_irqsave(&adapter->hw_lock, flags);
1195
1196                 pvscsi_process_completion_ring(adapter);
1197                 if (adapter->use_msg && pvscsi_msg_pending(adapter))
1198                         queue_work(adapter->workqueue, &adapter->work);
1199
1200                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
1201         }
1202
1203         return IRQ_RETVAL(handled);
1204 }
1205
1206 static void pvscsi_free_sgls(const struct pvscsi_adapter *adapter)
1207 {
1208         struct pvscsi_ctx *ctx = adapter->cmd_map;
1209         unsigned i;
1210
1211         for (i = 0; i < adapter->req_depth; ++i, ++ctx)
1212                 free_pages((unsigned long)ctx->sgl, get_order(SGL_SIZE));
1213 }
1214
1215 static int pvscsi_setup_msix(const struct pvscsi_adapter *adapter,
1216                              unsigned int *irq)
1217 {
1218         struct msix_entry entry = { 0, PVSCSI_VECTOR_COMPLETION };
1219         int ret;
1220
1221         ret = pci_enable_msix_exact(adapter->dev, &entry, 1);
1222         if (ret)
1223                 return ret;
1224
1225         *irq = entry.vector;
1226
1227         return 0;
1228 }
1229
1230 static void pvscsi_shutdown_intr(struct pvscsi_adapter *adapter)
1231 {
1232         if (adapter->irq) {
1233                 free_irq(adapter->irq, adapter);
1234                 adapter->irq = 0;
1235         }
1236         if (adapter->use_msi) {
1237                 pci_disable_msi(adapter->dev);
1238                 adapter->use_msi = 0;
1239         } else if (adapter->use_msix) {
1240                 pci_disable_msix(adapter->dev);
1241                 adapter->use_msix = 0;
1242         }
1243 }
1244
1245 static void pvscsi_release_resources(struct pvscsi_adapter *adapter)
1246 {
1247         if (adapter->workqueue)
1248                 destroy_workqueue(adapter->workqueue);
1249
1250         if (adapter->mmioBase)
1251                 pci_iounmap(adapter->dev, adapter->mmioBase);
1252
1253         pci_release_regions(adapter->dev);
1254
1255         if (adapter->cmd_map) {
1256                 pvscsi_free_sgls(adapter);
1257                 kfree(adapter->cmd_map);
1258         }
1259
1260         if (adapter->rings_state)
1261                 pci_free_consistent(adapter->dev, PAGE_SIZE,
1262                                     adapter->rings_state, adapter->ringStatePA);
1263
1264         if (adapter->req_ring)
1265                 pci_free_consistent(adapter->dev,
1266                                     adapter->req_pages * PAGE_SIZE,
1267                                     adapter->req_ring, adapter->reqRingPA);
1268
1269         if (adapter->cmp_ring)
1270                 pci_free_consistent(adapter->dev,
1271                                     adapter->cmp_pages * PAGE_SIZE,
1272                                     adapter->cmp_ring, adapter->cmpRingPA);
1273
1274         if (adapter->msg_ring)
1275                 pci_free_consistent(adapter->dev,
1276                                     adapter->msg_pages * PAGE_SIZE,
1277                                     adapter->msg_ring, adapter->msgRingPA);
1278 }
1279
1280 /*
1281  * Allocate scatter gather lists.
1282  *
1283  * These are statically allocated.  Trying to be clever was not worth it.
1284  *
1285  * Dynamic allocation can fail, and we can't go deep into the memory
1286  * allocator, since we're a SCSI driver, and trying too hard to allocate
1287  * memory might generate disk I/O.  We also don't want to fail disk I/O
1288  * in that case because we can't get an allocation - the I/O could be
1289  * trying to swap out data to free memory.  Since that is pathological,
1290  * just use a statically allocated scatter list.
1291  *
1292  */
1293 static int pvscsi_allocate_sg(struct pvscsi_adapter *adapter)
1294 {
1295         struct pvscsi_ctx *ctx;
1296         int i;
1297
1298         ctx = adapter->cmd_map;
1299         BUILD_BUG_ON(sizeof(struct pvscsi_sg_list) > SGL_SIZE);
1300
1301         for (i = 0; i < adapter->req_depth; ++i, ++ctx) {
1302                 ctx->sgl = (void *)__get_free_pages(GFP_KERNEL,
1303                                                     get_order(SGL_SIZE));
1304                 ctx->sglPA = 0;
1305                 BUG_ON(!IS_ALIGNED(((unsigned long)ctx->sgl), PAGE_SIZE));
1306                 if (!ctx->sgl) {
1307                         for (; i >= 0; --i, --ctx) {
1308                                 free_pages((unsigned long)ctx->sgl,
1309                                            get_order(SGL_SIZE));
1310                                 ctx->sgl = NULL;
1311                         }
1312                         return -ENOMEM;
1313                 }
1314         }
1315
1316         return 0;
1317 }
1318
1319 /*
1320  * Query the device, fetch the config info and return the
1321  * maximum number of targets on the adapter. In case of
1322  * failure due to any reason return default i.e. 16.
1323  */
1324 static u32 pvscsi_get_max_targets(struct pvscsi_adapter *adapter)
1325 {
1326         struct PVSCSICmdDescConfigCmd cmd;
1327         struct PVSCSIConfigPageHeader *header;
1328         struct device *dev;
1329         dma_addr_t configPagePA;
1330         void *config_page;
1331         u32 numPhys = 16;
1332
1333         dev = pvscsi_dev(adapter);
1334         config_page = pci_alloc_consistent(adapter->dev, PAGE_SIZE,
1335                                            &configPagePA);
1336         if (!config_page) {
1337                 dev_warn(dev, "vmw_pvscsi: failed to allocate memory for config page\n");
1338                 goto exit;
1339         }
1340         BUG_ON(configPagePA & ~PAGE_MASK);
1341
1342         /* Fetch config info from the device. */
1343         cmd.configPageAddress = ((u64)PVSCSI_CONFIG_CONTROLLER_ADDRESS) << 32;
1344         cmd.configPageNum = PVSCSI_CONFIG_PAGE_CONTROLLER;
1345         cmd.cmpAddr = configPagePA;
1346         cmd._pad = 0;
1347
1348         /*
1349          * Mark the completion page header with error values. If the device
1350          * completes the command successfully, it sets the status values to
1351          * indicate success.
1352          */
1353         header = config_page;
1354         memset(header, 0, sizeof *header);
1355         header->hostStatus = BTSTAT_INVPARAM;
1356         header->scsiStatus = SDSTAT_CHECK;
1357
1358         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_CONFIG, &cmd, sizeof cmd);
1359
1360         if (header->hostStatus == BTSTAT_SUCCESS &&
1361             header->scsiStatus == SDSTAT_GOOD) {
1362                 struct PVSCSIConfigPageController *config;
1363
1364                 config = config_page;
1365                 numPhys = config->numPhys;
1366         } else
1367                 dev_warn(dev, "vmw_pvscsi: PVSCSI_CMD_CONFIG failed. hostStatus = 0x%x, scsiStatus = 0x%x\n",
1368                          header->hostStatus, header->scsiStatus);
1369         pci_free_consistent(adapter->dev, PAGE_SIZE, config_page, configPagePA);
1370 exit:
1371         return numPhys;
1372 }
1373
1374 static int pvscsi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1375 {
1376         struct pvscsi_adapter *adapter;
1377         struct pvscsi_adapter adapter_temp;
1378         struct Scsi_Host *host = NULL;
1379         unsigned int i;
1380         unsigned long flags = 0;
1381         int error;
1382         u32 max_id;
1383
1384         error = -ENODEV;
1385
1386         if (pci_enable_device(pdev))
1387                 return error;
1388
1389         if (pci_set_dma_mask(pdev, DMA_BIT_MASK(64)) == 0 &&
1390             pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64)) == 0) {
1391                 printk(KERN_INFO "vmw_pvscsi: using 64bit dma\n");
1392         } else if (pci_set_dma_mask(pdev, DMA_BIT_MASK(32)) == 0 &&
1393                    pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32)) == 0) {
1394                 printk(KERN_INFO "vmw_pvscsi: using 32bit dma\n");
1395         } else {
1396                 printk(KERN_ERR "vmw_pvscsi: failed to set DMA mask\n");
1397                 goto out_disable_device;
1398         }
1399
1400         /*
1401          * Let's use a temp pvscsi_adapter struct until we find the number of
1402          * targets on the adapter, after that we will switch to the real
1403          * allocated struct.
1404          */
1405         adapter = &adapter_temp;
1406         memset(adapter, 0, sizeof(*adapter));
1407         adapter->dev  = pdev;
1408         adapter->rev = pdev->revision;
1409
1410         if (pci_request_regions(pdev, "vmw_pvscsi")) {
1411                 printk(KERN_ERR "vmw_pvscsi: pci memory selection failed\n");
1412                 goto out_disable_device;
1413         }
1414
1415         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
1416                 if ((pci_resource_flags(pdev, i) & PCI_BASE_ADDRESS_SPACE_IO))
1417                         continue;
1418
1419                 if (pci_resource_len(pdev, i) < PVSCSI_MEM_SPACE_SIZE)
1420                         continue;
1421
1422                 break;
1423         }
1424
1425         if (i == DEVICE_COUNT_RESOURCE) {
1426                 printk(KERN_ERR
1427                        "vmw_pvscsi: adapter has no suitable MMIO region\n");
1428                 goto out_release_resources_and_disable;
1429         }
1430
1431         adapter->mmioBase = pci_iomap(pdev, i, PVSCSI_MEM_SPACE_SIZE);
1432
1433         if (!adapter->mmioBase) {
1434                 printk(KERN_ERR
1435                        "vmw_pvscsi: can't iomap for BAR %d memsize %lu\n",
1436                        i, PVSCSI_MEM_SPACE_SIZE);
1437                 goto out_release_resources_and_disable;
1438         }
1439
1440         pci_set_master(pdev);
1441
1442         /*
1443          * Ask the device for max number of targets before deciding the
1444          * default pvscsi_ring_pages value.
1445          */
1446         max_id = pvscsi_get_max_targets(adapter);
1447         printk(KERN_INFO "vmw_pvscsi: max_id: %u\n", max_id);
1448
1449         if (pvscsi_ring_pages == 0)
1450                 /*
1451                  * Set the right default value. Up to 16 it is 8, above it is
1452                  * max.
1453                  */
1454                 pvscsi_ring_pages = (max_id > 16) ?
1455                         PVSCSI_SETUP_RINGS_MAX_NUM_PAGES :
1456                         PVSCSI_DEFAULT_NUM_PAGES_PER_RING;
1457         printk(KERN_INFO
1458                "vmw_pvscsi: setting ring_pages to %d\n",
1459                pvscsi_ring_pages);
1460
1461         pvscsi_template.can_queue =
1462                 min(PVSCSI_MAX_NUM_PAGES_REQ_RING, pvscsi_ring_pages) *
1463                 PVSCSI_MAX_NUM_REQ_ENTRIES_PER_PAGE;
1464         pvscsi_template.cmd_per_lun =
1465                 min(pvscsi_template.can_queue, pvscsi_cmd_per_lun);
1466         host = scsi_host_alloc(&pvscsi_template, sizeof(struct pvscsi_adapter));
1467         if (!host) {
1468                 printk(KERN_ERR "vmw_pvscsi: failed to allocate host\n");
1469                 goto out_release_resources_and_disable;
1470         }
1471
1472         /*
1473          * Let's use the real pvscsi_adapter struct here onwards.
1474          */
1475         adapter = shost_priv(host);
1476         memset(adapter, 0, sizeof(*adapter));
1477         adapter->dev  = pdev;
1478         adapter->host = host;
1479         /*
1480          * Copy back what we already have to the allocated adapter struct.
1481          */
1482         adapter->rev = adapter_temp.rev;
1483         adapter->mmioBase = adapter_temp.mmioBase;
1484
1485         spin_lock_init(&adapter->hw_lock);
1486         host->max_channel = 0;
1487         host->max_lun     = 1;
1488         host->max_cmd_len = 16;
1489         host->max_id      = max_id;
1490
1491         pci_set_drvdata(pdev, host);
1492
1493         ll_adapter_reset(adapter);
1494
1495         adapter->use_msg = pvscsi_setup_msg_workqueue(adapter);
1496
1497         error = pvscsi_allocate_rings(adapter);
1498         if (error) {
1499                 printk(KERN_ERR "vmw_pvscsi: unable to allocate ring memory\n");
1500                 goto out_release_resources;
1501         }
1502
1503         /*
1504          * From this point on we should reset the adapter if anything goes
1505          * wrong.
1506          */
1507         pvscsi_setup_all_rings(adapter);
1508
1509         adapter->cmd_map = kcalloc(adapter->req_depth,
1510                                    sizeof(struct pvscsi_ctx), GFP_KERNEL);
1511         if (!adapter->cmd_map) {
1512                 printk(KERN_ERR "vmw_pvscsi: failed to allocate memory.\n");
1513                 error = -ENOMEM;
1514                 goto out_reset_adapter;
1515         }
1516
1517         INIT_LIST_HEAD(&adapter->cmd_pool);
1518         for (i = 0; i < adapter->req_depth; i++) {
1519                 struct pvscsi_ctx *ctx = adapter->cmd_map + i;
1520                 list_add(&ctx->list, &adapter->cmd_pool);
1521         }
1522
1523         error = pvscsi_allocate_sg(adapter);
1524         if (error) {
1525                 printk(KERN_ERR "vmw_pvscsi: unable to allocate s/g table\n");
1526                 goto out_reset_adapter;
1527         }
1528
1529         if (!pvscsi_disable_msix &&
1530             pvscsi_setup_msix(adapter, &adapter->irq) == 0) {
1531                 printk(KERN_INFO "vmw_pvscsi: using MSI-X\n");
1532                 adapter->use_msix = 1;
1533         } else if (!pvscsi_disable_msi && pci_enable_msi(pdev) == 0) {
1534                 printk(KERN_INFO "vmw_pvscsi: using MSI\n");
1535                 adapter->use_msi = 1;
1536                 adapter->irq = pdev->irq;
1537         } else {
1538                 printk(KERN_INFO "vmw_pvscsi: using INTx\n");
1539                 adapter->irq = pdev->irq;
1540                 flags = IRQF_SHARED;
1541         }
1542
1543         adapter->use_req_threshold = pvscsi_setup_req_threshold(adapter, true);
1544         printk(KERN_DEBUG "vmw_pvscsi: driver-based request coalescing %sabled\n",
1545                adapter->use_req_threshold ? "en" : "dis");
1546
1547         error = request_irq(adapter->irq, pvscsi_isr, flags,
1548                             "vmw_pvscsi", adapter);
1549         if (error) {
1550                 printk(KERN_ERR
1551                        "vmw_pvscsi: unable to request IRQ: %d\n", error);
1552                 adapter->irq = 0;
1553                 goto out_reset_adapter;
1554         }
1555
1556         error = scsi_add_host(host, &pdev->dev);
1557         if (error) {
1558                 printk(KERN_ERR
1559                        "vmw_pvscsi: scsi_add_host failed: %d\n", error);
1560                 goto out_reset_adapter;
1561         }
1562
1563         dev_info(&pdev->dev, "VMware PVSCSI rev %d host #%u\n",
1564                  adapter->rev, host->host_no);
1565
1566         pvscsi_unmask_intr(adapter);
1567
1568         scsi_scan_host(host);
1569
1570         return 0;
1571
1572 out_reset_adapter:
1573         ll_adapter_reset(adapter);
1574 out_release_resources:
1575         pvscsi_shutdown_intr(adapter);
1576         pvscsi_release_resources(adapter);
1577         scsi_host_put(host);
1578 out_disable_device:
1579         pci_disable_device(pdev);
1580
1581         return error;
1582
1583 out_release_resources_and_disable:
1584         pvscsi_shutdown_intr(adapter);
1585         pvscsi_release_resources(adapter);
1586         goto out_disable_device;
1587 }
1588
1589 static void __pvscsi_shutdown(struct pvscsi_adapter *adapter)
1590 {
1591         pvscsi_mask_intr(adapter);
1592
1593         if (adapter->workqueue)
1594                 flush_workqueue(adapter->workqueue);
1595
1596         pvscsi_shutdown_intr(adapter);
1597
1598         pvscsi_process_request_ring(adapter);
1599         pvscsi_process_completion_ring(adapter);
1600         ll_adapter_reset(adapter);
1601 }
1602
1603 static void pvscsi_shutdown(struct pci_dev *dev)
1604 {
1605         struct Scsi_Host *host = pci_get_drvdata(dev);
1606         struct pvscsi_adapter *adapter = shost_priv(host);
1607
1608         __pvscsi_shutdown(adapter);
1609 }
1610
1611 static void pvscsi_remove(struct pci_dev *pdev)
1612 {
1613         struct Scsi_Host *host = pci_get_drvdata(pdev);
1614         struct pvscsi_adapter *adapter = shost_priv(host);
1615
1616         scsi_remove_host(host);
1617
1618         __pvscsi_shutdown(adapter);
1619         pvscsi_release_resources(adapter);
1620
1621         scsi_host_put(host);
1622
1623         pci_disable_device(pdev);
1624 }
1625
1626 static struct pci_driver pvscsi_pci_driver = {
1627         .name           = "vmw_pvscsi",
1628         .id_table       = pvscsi_pci_tbl,
1629         .probe          = pvscsi_probe,
1630         .remove         = pvscsi_remove,
1631         .shutdown       = pvscsi_shutdown,
1632 };
1633
1634 static int __init pvscsi_init(void)
1635 {
1636         pr_info("%s - version %s\n",
1637                 PVSCSI_LINUX_DRIVER_DESC, PVSCSI_DRIVER_VERSION_STRING);
1638         return pci_register_driver(&pvscsi_pci_driver);
1639 }
1640
1641 static void __exit pvscsi_exit(void)
1642 {
1643         pci_unregister_driver(&pvscsi_pci_driver);
1644 }
1645
1646 module_init(pvscsi_init);
1647 module_exit(pvscsi_exit);