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: Arvind Kumar <arvindkumar@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 void 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;
 364
 365         sg = scsi_sglist(cmd);
 366         count = scsi_sg_count(cmd);
 367         if (count != 0) {
 368                 int segs = scsi_dma_map(cmd);
 369                 if (segs > 1) {
 370                         pvscsi_create_sg(ctx, sg, segs);
 371
 372                         e->flags |= PVSCSI_FLAG_CMD_WITH_SG_LIST;
 373                         ctx->sglPA = pci_map_single(adapter->dev, ctx->sgl,
 374                                                     SGL_SIZE, PCI_DMA_TODEVICE);
 375                         e->dataAddr = ctx->sglPA;
 376                 } else
 377                         e->dataAddr = sg_dma_address(sg);
 378         } else {
 379                 /*
 380                  * In case there is no S/G list, scsi_sglist points
 381                  * directly to the buffer.
 382                  */
 383                 ctx->dataPA = pci_map_single(adapter->dev, sg, bufflen,
 384                                              cmd->sc_data_direction);
 385                 e->dataAddr = ctx->dataPA;
 386         }
 387 }
 388
 389 static void pvscsi_unmap_buffers(const struct pvscsi_adapter *adapter,
 390                                  struct pvscsi_ctx *ctx)
 391 {
 392         struct scsi_cmnd *cmd;
 393         unsigned bufflen;
 394
 395         cmd = ctx->cmd;
 396         bufflen = scsi_bufflen(cmd);
 397
 398         if (bufflen != 0) {
 399                 unsigned count = scsi_sg_count(cmd);
 400
 401                 if (count != 0) {
 402                         scsi_dma_unmap(cmd);
 403                         if (ctx->sglPA) {
 404                                 pci_unmap_single(adapter->dev, ctx->sglPA,
 405                                                  SGL_SIZE, PCI_DMA_TODEVICE);
 406                                 ctx->sglPA = 0;
 407                         }
 408                 } else
 409                         pci_unmap_single(adapter->dev, ctx->dataPA, bufflen,
 410                                          cmd->sc_data_direction);
 411         }
 412         if (cmd->sense_buffer)
 413                 pci_unmap_single(adapter->dev, ctx->sensePA,
 414                                  SCSI_SENSE_BUFFERSIZE, PCI_DMA_FROMDEVICE);
 415 }
 416
 417 static int pvscsi_allocate_rings(struct pvscsi_adapter *adapter)
 418 {
 419         adapter->rings_state = pci_alloc_consistent(adapter->dev, PAGE_SIZE,
 420                                                     &adapter->ringStatePA);
 421         if (!adapter->rings_state)
 422                 return -ENOMEM;
 423
 424         adapter->req_pages = min(PVSCSI_MAX_NUM_PAGES_REQ_RING,
 425                                  pvscsi_ring_pages);
 426         adapter->req_depth = adapter->req_pages
 427                                         * PVSCSI_MAX_NUM_REQ_ENTRIES_PER_PAGE;
 428         adapter->req_ring = pci_alloc_consistent(adapter->dev,
 429                                                  adapter->req_pages * PAGE_SIZE,
 430                                                  &adapter->reqRingPA);
 431         if (!adapter->req_ring)
 432                 return -ENOMEM;
 433
 434         adapter->cmp_pages = min(PVSCSI_MAX_NUM_PAGES_CMP_RING,
 435                                  pvscsi_ring_pages);
 436         adapter->cmp_ring = pci_alloc_consistent(adapter->dev,
 437                                                  adapter->cmp_pages * PAGE_SIZE,
 438                                                  &adapter->cmpRingPA);
 439         if (!adapter->cmp_ring)
 440                 return -ENOMEM;
 441
 442         BUG_ON(!IS_ALIGNED(adapter->ringStatePA, PAGE_SIZE));
 443         BUG_ON(!IS_ALIGNED(adapter->reqRingPA, PAGE_SIZE));
 444         BUG_ON(!IS_ALIGNED(adapter->cmpRingPA, PAGE_SIZE));
 445
 446         if (!adapter->use_msg)
 447                 return 0;
 448
 449         adapter->msg_pages = min(PVSCSI_MAX_NUM_PAGES_MSG_RING,
 450                                  pvscsi_msg_ring_pages);
 451         adapter->msg_ring = pci_alloc_consistent(adapter->dev,
 452                                                  adapter->msg_pages * PAGE_SIZE,
 453                                                  &adapter->msgRingPA);
 454         if (!adapter->msg_ring)
 455                 return -ENOMEM;
 456         BUG_ON(!IS_ALIGNED(adapter->msgRingPA, PAGE_SIZE));
 457
 458         return 0;
 459 }
 460
 461 static void pvscsi_setup_all_rings(const struct pvscsi_adapter *adapter)
 462 {
 463         struct PVSCSICmdDescSetupRings cmd = { 0 };
 464         dma_addr_t base;
 465         unsigned i;
 466
 467         cmd.ringsStatePPN   = adapter->ringStatePA >> PAGE_SHIFT;
 468         cmd.reqRingNumPages = adapter->req_pages;
 469         cmd.cmpRingNumPages = adapter->cmp_pages;
 470
 471         base = adapter->reqRingPA;
 472         for (i = 0; i < adapter->req_pages; i++) {
 473                 cmd.reqRingPPNs[i] = base >> PAGE_SHIFT;
 474                 base += PAGE_SIZE;
 475         }
 476
 477         base = adapter->cmpRingPA;
 478         for (i = 0; i < adapter->cmp_pages; i++) {
 479                 cmd.cmpRingPPNs[i] = base >> PAGE_SHIFT;
 480                 base += PAGE_SIZE;
 481         }
 482
 483         memset(adapter->rings_state, 0, PAGE_SIZE);
 484         memset(adapter->req_ring, 0, adapter->req_pages * PAGE_SIZE);
 485         memset(adapter->cmp_ring, 0, adapter->cmp_pages * PAGE_SIZE);
 486
 487         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_SETUP_RINGS,
 488                               &cmd, sizeof(cmd));
 489
 490         if (adapter->use_msg) {
 491                 struct PVSCSICmdDescSetupMsgRing cmd_msg = { 0 };
 492
 493                 cmd_msg.numPages = adapter->msg_pages;
 494
 495                 base = adapter->msgRingPA;
 496                 for (i = 0; i < adapter->msg_pages; i++) {
 497                         cmd_msg.ringPPNs[i] = base >> PAGE_SHIFT;
 498                         base += PAGE_SIZE;
 499                 }
 500                 memset(adapter->msg_ring, 0, adapter->msg_pages * PAGE_SIZE);
 501
 502                 pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_SETUP_MSG_RING,
 503                                       &cmd_msg, sizeof(cmd_msg));
 504         }
 505 }
 506
 507 static int pvscsi_change_queue_depth(struct scsi_device *sdev, int qdepth)
 508 {
 509         if (!sdev->tagged_supported)
 510                 qdepth = 1;
 511         scsi_change_queue_depth(sdev, qdepth);
 512
 513         if (sdev->inquiry_len > 7)
 514                 sdev_printk(KERN_INFO, sdev,
 515                             "qdepth(%d), tagged(%d), simple(%d), scsi_level(%d), cmd_que(%d)\n",
 516                             sdev->queue_depth, sdev->tagged_supported,
 517                             sdev->simple_tags,
 518                             sdev->scsi_level, (sdev->inquiry[7] & 2) >> 1);
 519         return sdev->queue_depth;
 520 }
 521
 522 /*
 523  * Pull a completion descriptor off and pass the completion back
 524  * to the SCSI mid layer.
 525  */
 526 static void pvscsi_complete_request(struct pvscsi_adapter *adapter,
 527                                     const struct PVSCSIRingCmpDesc *e)
 528 {
 529         struct pvscsi_ctx *ctx;
 530         struct scsi_cmnd *cmd;
 531         struct completion *abort_cmp;
 532         u32 btstat = e->hostStatus;
 533         u32 sdstat = e->scsiStatus;
 534
 535         ctx = pvscsi_get_context(adapter, e->context);
 536         cmd = ctx->cmd;
 537         abort_cmp = ctx->abort_cmp;
 538         pvscsi_unmap_buffers(adapter, ctx);
 539         pvscsi_release_context(adapter, ctx);
 540         if (abort_cmp) {
 541                 /*
 542                  * The command was requested to be aborted. Just signal that
 543                  * the request completed and swallow the actual cmd completion
 544                  * here. The abort handler will post a completion for this
 545                  * command indicating that it got successfully aborted.
 546                  */
 547                 complete(abort_cmp);
 548                 return;
 549         }
 550
 551         cmd->result = 0;
 552         if (sdstat != SAM_STAT_GOOD &&
 553             (btstat == BTSTAT_SUCCESS ||
 554              btstat == BTSTAT_LINKED_COMMAND_COMPLETED ||
 555              btstat == BTSTAT_LINKED_COMMAND_COMPLETED_WITH_FLAG)) {
 556                 cmd->result = (DID_OK << 16) | sdstat;
 557                 if (sdstat == SAM_STAT_CHECK_CONDITION && cmd->sense_buffer)
 558                         cmd->result |= (DRIVER_SENSE << 24);
 559         } else
 560                 switch (btstat) {
 561                 case BTSTAT_SUCCESS:
 562                 case BTSTAT_LINKED_COMMAND_COMPLETED:
 563                 case BTSTAT_LINKED_COMMAND_COMPLETED_WITH_FLAG:
 564                         /* If everything went fine, let's move on..  */
 565                         cmd->result = (DID_OK << 16);
 566                         break;
 567
 568                 case BTSTAT_DATARUN:
 569                 case BTSTAT_DATA_UNDERRUN:
 570                         /* Report residual data in underruns */
 571                         scsi_set_resid(cmd, scsi_bufflen(cmd) - e->dataLen);
 572                         cmd->result = (DID_ERROR << 16);
 573                         break;
 574
 575                 case BTSTAT_SELTIMEO:
 576                         /* Our emulation returns this for non-connected devs */
 577                         cmd->result = (DID_BAD_TARGET << 16);
 578                         break;
 579
 580                 case BTSTAT_LUNMISMATCH:
 581                 case BTSTAT_TAGREJECT:
 582                 case BTSTAT_BADMSG:
 583                         cmd->result = (DRIVER_INVALID << 24);
 584                         /* fall through */
 585
 586                 case BTSTAT_HAHARDWARE:
 587                 case BTSTAT_INVPHASE:
 588                 case BTSTAT_HATIMEOUT:
 589                 case BTSTAT_NORESPONSE:
 590                 case BTSTAT_DISCONNECT:
 591                 case BTSTAT_HASOFTWARE:
 592                 case BTSTAT_BUSFREE:
 593                 case BTSTAT_SENSFAILED:
 594                         cmd->result |= (DID_ERROR << 16);
 595                         break;
 596
 597                 case BTSTAT_SENTRST:
 598                 case BTSTAT_RECVRST:
 599                 case BTSTAT_BUSRESET:
 600                         cmd->result = (DID_RESET << 16);
 601                         break;
 602
 603                 case BTSTAT_ABORTQUEUE:
 604                         cmd->result = (DID_ABORT << 16);
 605                         break;
 606
 607                 case BTSTAT_SCSIPARITY:
 608                         cmd->result = (DID_PARITY << 16);
 609                         break;
 610
 611                 default:
 612                         cmd->result = (DID_ERROR << 16);
 613                         scmd_printk(KERN_DEBUG, cmd,
 614                                     "Unknown completion status: 0x%x\n",
 615                                     btstat);
 616         }
 617
 618         dev_dbg(&cmd->device->sdev_gendev,
 619                 "cmd=%p %x ctx=%p result=0x%x status=0x%x,%x\n",
 620                 cmd, cmd->cmnd[0], ctx, cmd->result, btstat, sdstat);
 621
 622         cmd->scsi_done(cmd);
 623 }
 624
 625 /*
 626  * barrier usage : Since the PVSCSI device is emulated, there could be cases
 627  * where we may want to serialize some accesses between the driver and the
 628  * emulation layer. We use compiler barriers instead of the more expensive
 629  * memory barriers because PVSCSI is only supported on X86 which has strong
 630  * memory access ordering.
 631  */
 632 static void pvscsi_process_completion_ring(struct pvscsi_adapter *adapter)
 633 {
 634         struct PVSCSIRingsState *s = adapter->rings_state;
 635         struct PVSCSIRingCmpDesc *ring = adapter->cmp_ring;
 636         u32 cmp_entries = s->cmpNumEntriesLog2;
 637
 638         while (s->cmpConsIdx != s->cmpProdIdx) {
 639                 struct PVSCSIRingCmpDesc *e = ring + (s->cmpConsIdx &
 640                                                       MASK(cmp_entries));
 641                 /*
 642                  * This barrier() ensures that *e is not dereferenced while
 643                  * the device emulation still writes data into the slot.
 644                  * Since the device emulation advances s->cmpProdIdx only after
 645                  * updating the slot we want to check it first.
 646                  */
 647                 barrier();
 648                 pvscsi_complete_request(adapter, e);
 649                 /*
 650                  * This barrier() ensures that compiler doesn't reorder write
 651                  * to s->cmpConsIdx before the read of (*e) inside
 652                  * pvscsi_complete_request. Otherwise, device emulation may
 653                  * overwrite *e before we had a chance to read it.
 654                  */
 655                 barrier();
 656                 s->cmpConsIdx++;
 657         }
 658 }
 659
 660 /*
 661  * Translate a Linux SCSI request into a request ring entry.
 662  */
 663 static int pvscsi_queue_ring(struct pvscsi_adapter *adapter,
 664                              struct pvscsi_ctx *ctx, struct scsi_cmnd *cmd)
 665 {
 666         struct PVSCSIRingsState *s;
 667         struct PVSCSIRingReqDesc *e;
 668         struct scsi_device *sdev;
 669         u32 req_entries;
 670
 671         s = adapter->rings_state;
 672         sdev = cmd->device;
 673         req_entries = s->reqNumEntriesLog2;
 674
 675         /*
 676          * If this condition holds, we might have room on the request ring, but
 677          * we might not have room on the completion ring for the response.
 678          * However, we have already ruled out this possibility - we would not
 679          * have successfully allocated a context if it were true, since we only
 680          * have one context per request entry.  Check for it anyway, since it
 681          * would be a serious bug.
 682          */
 683         if (s->reqProdIdx - s->cmpConsIdx >= 1 << req_entries) {
 684                 scmd_printk(KERN_ERR, cmd, "vmw_pvscsi: "
 685                             "ring full: reqProdIdx=%d cmpConsIdx=%d\n",
 686                             s->reqProdIdx, s->cmpConsIdx);
 687                 return -1;
 688         }
 689
 690         e = adapter->req_ring + (s->reqProdIdx & MASK(req_entries));
 691
 692         e->bus    = sdev->channel;
 693         e->target = sdev->id;
 694         memset(e->lun, 0, sizeof(e->lun));
 695         e->lun[1] = sdev->lun;
 696
 697         if (cmd->sense_buffer) {
 698                 ctx->sensePA = pci_map_single(adapter->dev, cmd->sense_buffer,
 699                                               SCSI_SENSE_BUFFERSIZE,
 700                                               PCI_DMA_FROMDEVICE);
 701                 e->senseAddr = ctx->sensePA;
 702                 e->senseLen = SCSI_SENSE_BUFFERSIZE;
 703         } else {
 704                 e->senseLen  = 0;
 705                 e->senseAddr = 0;
 706         }
 707         e->cdbLen   = cmd->cmd_len;
 708         e->vcpuHint = smp_processor_id();
 709         memcpy(e->cdb, cmd->cmnd, e->cdbLen);
 710
 711         e->tag = SIMPLE_QUEUE_TAG;
 712
 713         if (cmd->sc_data_direction == DMA_FROM_DEVICE)
 714                 e->flags = PVSCSI_FLAG_CMD_DIR_TOHOST;
 715         else if (cmd->sc_data_direction == DMA_TO_DEVICE)
 716                 e->flags = PVSCSI_FLAG_CMD_DIR_TODEVICE;
 717         else if (cmd->sc_data_direction == DMA_NONE)
 718                 e->flags = PVSCSI_FLAG_CMD_DIR_NONE;
 719         else
 720                 e->flags = 0;
 721
 722         pvscsi_map_buffers(adapter, ctx, cmd, e);
 723
 724         e->context = pvscsi_map_context(adapter, ctx);
 725
 726         barrier();
 727
 728         s->reqProdIdx++;
 729
 730         return 0;
 731 }
 732
 733 static int pvscsi_queue_lck(struct scsi_cmnd *cmd, void (*done)(struct scsi_cmnd *))
 734 {
 735         struct Scsi_Host *host = cmd->device->host;
 736         struct pvscsi_adapter *adapter = shost_priv(host);
 737         struct pvscsi_ctx *ctx;
 738         unsigned long flags;
 739
 740         spin_lock_irqsave(&adapter->hw_lock, flags);
 741
 742         ctx = pvscsi_acquire_context(adapter, cmd);
 743         if (!ctx || pvscsi_queue_ring(adapter, ctx, cmd) != 0) {
 744                 if (ctx)
 745                         pvscsi_release_context(adapter, ctx);
 746                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 747                 return SCSI_MLQUEUE_HOST_BUSY;
 748         }
 749
 750         cmd->scsi_done = done;
 751
 752         dev_dbg(&cmd->device->sdev_gendev,
 753                 "queued cmd %p, ctx %p, op=%x\n", cmd, ctx, cmd->cmnd[0]);
 754
 755         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 756
 757         pvscsi_kick_io(adapter, cmd->cmnd[0]);
 758
 759         return 0;
 760 }
 761
 762 static DEF_SCSI_QCMD(pvscsi_queue)
 763
 764 static int pvscsi_abort(struct scsi_cmnd *cmd)
 765 {
 766         struct pvscsi_adapter *adapter = shost_priv(cmd->device->host);
 767         struct pvscsi_ctx *ctx;
 768         unsigned long flags;
 769         int result = SUCCESS;
 770         DECLARE_COMPLETION_ONSTACK(abort_cmp);
 771
 772         scmd_printk(KERN_DEBUG, cmd, "task abort on host %u, %p\n",
 773                     adapter->host->host_no, cmd);
 774
 775         spin_lock_irqsave(&adapter->hw_lock, flags);
 776
 777         /*
 778          * Poll the completion ring first - we might be trying to abort
 779          * a command that is waiting to be dispatched in the completion ring.
 780          */
 781         pvscsi_process_completion_ring(adapter);
 782
 783         /*
 784          * If there is no context for the command, it either already succeeded
 785          * or else was never properly issued.  Not our problem.
 786          */
 787         ctx = pvscsi_find_context(adapter, cmd);
 788         if (!ctx) {
 789                 scmd_printk(KERN_DEBUG, cmd, "Failed to abort cmd %p\n", cmd);
 790                 goto out;
 791         }
 792
 793         /*
 794          * Mark that the command has been requested to be aborted and issue
 795          * the abort.
 796          */
 797         ctx->abort_cmp = &abort_cmp;
 798
 799         pvscsi_abort_cmd(adapter, ctx);
 800         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 801         /* Wait for 2 secs for the completion. */
 802         wait_for_completion_timeout(&abort_cmp, msecs_to_jiffies(2000));
 803         spin_lock_irqsave(&adapter->hw_lock, flags);
 804
 805         if (!completion_done(&abort_cmp)) {
 806                 /*
 807                  * Failed to abort the command, unmark the fact that it
 808                  * was requested to be aborted.
 809                  */
 810                 ctx->abort_cmp = NULL;
 811                 result = FAILED;
 812                 scmd_printk(KERN_DEBUG, cmd,
 813                             "Failed to get completion for aborted cmd %p\n",
 814                             cmd);
 815                 goto out;
 816         }
 817
 818         /*
 819          * Successfully aborted the command.
 820          */
 821         cmd->result = (DID_ABORT << 16);
 822         cmd->scsi_done(cmd);
 823
 824 out:
 825         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 826         return result;
 827 }
 828
 829 /*
 830  * Abort all outstanding requests.  This is only safe to use if the completion
 831  * ring will never be walked again or the device has been reset, because it
 832  * destroys the 1-1 mapping between context field passed to emulation and our
 833  * request structure.
 834  */
 835 static void pvscsi_reset_all(struct pvscsi_adapter *adapter)
 836 {
 837         unsigned i;
 838
 839         for (i = 0; i < adapter->req_depth; i++) {
 840                 struct pvscsi_ctx *ctx = &adapter->cmd_map[i];
 841                 struct scsi_cmnd *cmd = ctx->cmd;
 842                 if (cmd) {
 843                         scmd_printk(KERN_ERR, cmd,
 844                                     "Forced reset on cmd %p\n", cmd);
 845                         pvscsi_unmap_buffers(adapter, ctx);
 846                         pvscsi_release_context(adapter, ctx);
 847                         cmd->result = (DID_RESET << 16);
 848                         cmd->scsi_done(cmd);
 849                 }
 850         }
 851 }
 852
 853 static int pvscsi_host_reset(struct scsi_cmnd *cmd)
 854 {
 855         struct Scsi_Host *host = cmd->device->host;
 856         struct pvscsi_adapter *adapter = shost_priv(host);
 857         unsigned long flags;
 858         bool use_msg;
 859
 860         scmd_printk(KERN_INFO, cmd, "SCSI Host reset\n");
 861
 862         spin_lock_irqsave(&adapter->hw_lock, flags);
 863
 864         use_msg = adapter->use_msg;
 865
 866         if (use_msg) {
 867                 adapter->use_msg = 0;
 868                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 869
 870                 /*
 871                  * Now that we know that the ISR won't add more work on the
 872                  * workqueue we can safely flush any outstanding work.
 873                  */
 874                 flush_workqueue(adapter->workqueue);
 875                 spin_lock_irqsave(&adapter->hw_lock, flags);
 876         }
 877
 878         /*
 879          * We're going to tear down the entire ring structure and set it back
 880          * up, so stalling new requests until all completions are flushed and
 881          * the rings are back in place.
 882          */
 883
 884         pvscsi_process_request_ring(adapter);
 885
 886         ll_adapter_reset(adapter);
 887
 888         /*
 889          * Now process any completions.  Note we do this AFTER adapter reset,
 890          * which is strange, but stops races where completions get posted
 891          * between processing the ring and issuing the reset.  The backend will
 892          * not touch the ring memory after reset, so the immediately pre-reset
 893          * completion ring state is still valid.
 894          */
 895         pvscsi_process_completion_ring(adapter);
 896
 897         pvscsi_reset_all(adapter);
 898         adapter->use_msg = use_msg;
 899         pvscsi_setup_all_rings(adapter);
 900         pvscsi_unmask_intr(adapter);
 901
 902         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 903
 904         return SUCCESS;
 905 }
 906
 907 static int pvscsi_bus_reset(struct scsi_cmnd *cmd)
 908 {
 909         struct Scsi_Host *host = cmd->device->host;
 910         struct pvscsi_adapter *adapter = shost_priv(host);
 911         unsigned long flags;
 912
 913         scmd_printk(KERN_INFO, cmd, "SCSI Bus reset\n");
 914
 915         /*
 916          * We don't want to queue new requests for this bus after
 917          * flushing all pending requests to emulation, since new
 918          * requests could then sneak in during this bus reset phase,
 919          * so take the lock now.
 920          */
 921         spin_lock_irqsave(&adapter->hw_lock, flags);
 922
 923         pvscsi_process_request_ring(adapter);
 924         ll_bus_reset(adapter);
 925         pvscsi_process_completion_ring(adapter);
 926
 927         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 928
 929         return SUCCESS;
 930 }
 931
 932 static int pvscsi_device_reset(struct scsi_cmnd *cmd)
 933 {
 934         struct Scsi_Host *host = cmd->device->host;
 935         struct pvscsi_adapter *adapter = shost_priv(host);
 936         unsigned long flags;
 937
 938         scmd_printk(KERN_INFO, cmd, "SCSI device reset on scsi%u:%u\n",
 939                     host->host_no, cmd->device->id);
 940
 941         /*
 942          * We don't want to queue new requests for this device after flushing
 943          * all pending requests to emulation, since new requests could then
 944          * sneak in during this device reset phase, so take the lock now.
 945          */
 946         spin_lock_irqsave(&adapter->hw_lock, flags);
 947
 948         pvscsi_process_request_ring(adapter);
 949         ll_device_reset(adapter, cmd->device->id);
 950         pvscsi_process_completion_ring(adapter);
 951
 952         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 953
 954         return SUCCESS;
 955 }
 956
 957 static struct scsi_host_template pvscsi_template;
 958
 959 static const char *pvscsi_info(struct Scsi_Host *host)
 960 {
 961         struct pvscsi_adapter *adapter = shost_priv(host);
 962         static char buf[256];
 963
 964         sprintf(buf, "VMware PVSCSI storage adapter rev %d, req/cmp/msg rings: "
 965                 "%u/%u/%u pages, cmd_per_lun=%u", adapter->rev,
 966                 adapter->req_pages, adapter->cmp_pages, adapter->msg_pages,
 967                 pvscsi_template.cmd_per_lun);
 968
 969         return buf;
 970 }
 971
 972 static struct scsi_host_template pvscsi_template = {
 973         .module                         = THIS_MODULE,
 974         .name                           = "VMware PVSCSI Host Adapter",
 975         .proc_name                      = "vmw_pvscsi",
 976         .info                           = pvscsi_info,
 977         .queuecommand                   = pvscsi_queue,
 978         .this_id                        = -1,
 979         .sg_tablesize                   = PVSCSI_MAX_NUM_SG_ENTRIES_PER_SEGMENT,
 980         .dma_boundary                   = UINT_MAX,
 981         .max_sectors                    = 0xffff,
 982         .use_clustering                 = ENABLE_CLUSTERING,
 983         .change_queue_depth             = pvscsi_change_queue_depth,
 984         .eh_abort_handler               = pvscsi_abort,
 985         .eh_device_reset_handler        = pvscsi_device_reset,
 986         .eh_bus_reset_handler           = pvscsi_bus_reset,
 987         .eh_host_reset_handler          = pvscsi_host_reset,
 988 };
 989
 990 static void pvscsi_process_msg(const struct pvscsi_adapter *adapter,
 991                                const struct PVSCSIRingMsgDesc *e)
 992 {
 993         struct PVSCSIRingsState *s = adapter->rings_state;
 994         struct Scsi_Host *host = adapter->host;
 995         struct scsi_device *sdev;
 996
 997         printk(KERN_INFO "vmw_pvscsi: msg type: 0x%x - MSG RING: %u/%u (%u) \n",
 998                e->type, s->msgProdIdx, s->msgConsIdx, s->msgNumEntriesLog2);
 999
1000         BUILD_BUG_ON(PVSCSI_MSG_LAST != 2);
1001
1002         if (e->type == PVSCSI_MSG_DEV_ADDED) {
1003                 struct PVSCSIMsgDescDevStatusChanged *desc;
1004                 desc = (struct PVSCSIMsgDescDevStatusChanged *)e;
1005
1006                 printk(KERN_INFO
1007                        "vmw_pvscsi: msg: device added at scsi%u:%u:%u\n",
1008                        desc->bus, desc->target, desc->lun[1]);
1009
1010                 if (!scsi_host_get(host))
1011                         return;
1012
1013                 sdev = scsi_device_lookup(host, desc->bus, desc->target,
1014                                           desc->lun[1]);
1015                 if (sdev) {
1016                         printk(KERN_INFO "vmw_pvscsi: device already exists\n");
1017                         scsi_device_put(sdev);
1018                 } else
1019                         scsi_add_device(adapter->host, desc->bus,
1020                                         desc->target, desc->lun[1]);
1021
1022                 scsi_host_put(host);
1023         } else if (e->type == PVSCSI_MSG_DEV_REMOVED) {
1024                 struct PVSCSIMsgDescDevStatusChanged *desc;
1025                 desc = (struct PVSCSIMsgDescDevStatusChanged *)e;
1026
1027                 printk(KERN_INFO
1028                        "vmw_pvscsi: msg: device removed at scsi%u:%u:%u\n",
1029                        desc->bus, desc->target, desc->lun[1]);
1030
1031                 if (!scsi_host_get(host))
1032                         return;
1033
1034                 sdev = scsi_device_lookup(host, desc->bus, desc->target,
1035                                           desc->lun[1]);
1036                 if (sdev) {
1037                         scsi_remove_device(sdev);
1038                         scsi_device_put(sdev);
1039                 } else
1040                         printk(KERN_INFO
1041                                "vmw_pvscsi: failed to lookup scsi%u:%u:%u\n",
1042                                desc->bus, desc->target, desc->lun[1]);
1043
1044                 scsi_host_put(host);
1045         }
1046 }
1047
1048 static int pvscsi_msg_pending(const struct pvscsi_adapter *adapter)
1049 {
1050         struct PVSCSIRingsState *s = adapter->rings_state;
1051
1052         return s->msgProdIdx != s->msgConsIdx;
1053 }
1054
1055 static void pvscsi_process_msg_ring(const struct pvscsi_adapter *adapter)
1056 {
1057         struct PVSCSIRingsState *s = adapter->rings_state;
1058         struct PVSCSIRingMsgDesc *ring = adapter->msg_ring;
1059         u32 msg_entries = s->msgNumEntriesLog2;
1060
1061         while (pvscsi_msg_pending(adapter)) {
1062                 struct PVSCSIRingMsgDesc *e = ring + (s->msgConsIdx &
1063                                                       MASK(msg_entries));
1064
1065                 barrier();
1066                 pvscsi_process_msg(adapter, e);
1067                 barrier();
1068                 s->msgConsIdx++;
1069         }
1070 }
1071
1072 static void pvscsi_msg_workqueue_handler(struct work_struct *data)
1073 {
1074         struct pvscsi_adapter *adapter;
1075
1076         adapter = container_of(data, struct pvscsi_adapter, work);
1077
1078         pvscsi_process_msg_ring(adapter);
1079 }
1080
1081 static int pvscsi_setup_msg_workqueue(struct pvscsi_adapter *adapter)
1082 {
1083         char name[32];
1084
1085         if (!pvscsi_use_msg)
1086                 return 0;
1087
1088         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
1089                          PVSCSI_CMD_SETUP_MSG_RING);
1090
1091         if (pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS) == -1)
1092                 return 0;
1093
1094         snprintf(name, sizeof(name),
1095                  "vmw_pvscsi_wq_%u", adapter->host->host_no);
1096
1097         adapter->workqueue = create_singlethread_workqueue(name);
1098         if (!adapter->workqueue) {
1099                 printk(KERN_ERR "vmw_pvscsi: failed to create work queue\n");
1100                 return 0;
1101         }
1102         INIT_WORK(&adapter->work, pvscsi_msg_workqueue_handler);
1103
1104         return 1;
1105 }
1106
1107 static bool pvscsi_setup_req_threshold(struct pvscsi_adapter *adapter,
1108                                       bool enable)
1109 {
1110         u32 val;
1111
1112         if (!pvscsi_use_req_threshold)
1113                 return false;
1114
1115         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
1116                          PVSCSI_CMD_SETUP_REQCALLTHRESHOLD);
1117         val = pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS);
1118         if (val == -1) {
1119                 printk(KERN_INFO "vmw_pvscsi: device does not support req_threshold\n");
1120                 return false;
1121         } else {
1122                 struct PVSCSICmdDescSetupReqCall cmd_msg = { 0 };
1123                 cmd_msg.enable = enable;
1124                 printk(KERN_INFO
1125                        "vmw_pvscsi: %sabling reqCallThreshold\n",
1126                         enable ? "en" : "dis");
1127                 pvscsi_write_cmd_desc(adapter,
1128                                       PVSCSI_CMD_SETUP_REQCALLTHRESHOLD,
1129                                       &cmd_msg, sizeof(cmd_msg));
1130                 return pvscsi_reg_read(adapter,
1131                                        PVSCSI_REG_OFFSET_COMMAND_STATUS) != 0;
1132         }
1133 }
1134
1135 static irqreturn_t pvscsi_isr(int irq, void *devp)
1136 {
1137         struct pvscsi_adapter *adapter = devp;
1138         int handled;
1139
1140         if (adapter->use_msi || adapter->use_msix)
1141                 handled = true;
1142         else {
1143                 u32 val = pvscsi_read_intr_status(adapter);
1144                 handled = (val & PVSCSI_INTR_ALL_SUPPORTED) != 0;
1145                 if (handled)
1146                         pvscsi_write_intr_status(devp, val);
1147         }
1148
1149         if (handled) {
1150                 unsigned long flags;
1151
1152                 spin_lock_irqsave(&adapter->hw_lock, flags);
1153
1154                 pvscsi_process_completion_ring(adapter);
1155                 if (adapter->use_msg && pvscsi_msg_pending(adapter))
1156                         queue_work(adapter->workqueue, &adapter->work);
1157
1158                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
1159         }
1160
1161         return IRQ_RETVAL(handled);
1162 }
1163
1164 static void pvscsi_free_sgls(const struct pvscsi_adapter *adapter)
1165 {
1166         struct pvscsi_ctx *ctx = adapter->cmd_map;
1167         unsigned i;
1168
1169         for (i = 0; i < adapter->req_depth; ++i, ++ctx)
1170                 free_pages((unsigned long)ctx->sgl, get_order(SGL_SIZE));
1171 }
1172
1173 static int pvscsi_setup_msix(const struct pvscsi_adapter *adapter,
1174                              unsigned int *irq)
1175 {
1176         struct msix_entry entry = { 0, PVSCSI_VECTOR_COMPLETION };
1177         int ret;
1178
1179         ret = pci_enable_msix_exact(adapter->dev, &entry, 1);
1180         if (ret)
1181                 return ret;
1182
1183         *irq = entry.vector;
1184
1185         return 0;
1186 }
1187
1188 static void pvscsi_shutdown_intr(struct pvscsi_adapter *adapter)
1189 {
1190         if (adapter->irq) {
1191                 free_irq(adapter->irq, adapter);
1192                 adapter->irq = 0;
1193         }
1194         if (adapter->use_msi) {
1195                 pci_disable_msi(adapter->dev);
1196                 adapter->use_msi = 0;
1197         } else if (adapter->use_msix) {
1198                 pci_disable_msix(adapter->dev);
1199                 adapter->use_msix = 0;
1200         }
1201 }
1202
1203 static void pvscsi_release_resources(struct pvscsi_adapter *adapter)
1204 {
1205         pvscsi_shutdown_intr(adapter);
1206
1207         if (adapter->workqueue)
1208                 destroy_workqueue(adapter->workqueue);
1209
1210         if (adapter->mmioBase)
1211                 pci_iounmap(adapter->dev, adapter->mmioBase);
1212
1213         pci_release_regions(adapter->dev);
1214
1215         if (adapter->cmd_map) {
1216                 pvscsi_free_sgls(adapter);
1217                 kfree(adapter->cmd_map);
1218         }
1219
1220         if (adapter->rings_state)
1221                 pci_free_consistent(adapter->dev, PAGE_SIZE,
1222                                     adapter->rings_state, adapter->ringStatePA);
1223
1224         if (adapter->req_ring)
1225                 pci_free_consistent(adapter->dev,
1226                                     adapter->req_pages * PAGE_SIZE,
1227                                     adapter->req_ring, adapter->reqRingPA);
1228
1229         if (adapter->cmp_ring)
1230                 pci_free_consistent(adapter->dev,
1231                                     adapter->cmp_pages * PAGE_SIZE,
1232                                     adapter->cmp_ring, adapter->cmpRingPA);
1233
1234         if (adapter->msg_ring)
1235                 pci_free_consistent(adapter->dev,
1236                                     adapter->msg_pages * PAGE_SIZE,
1237                                     adapter->msg_ring, adapter->msgRingPA);
1238 }
1239
1240 /*
1241  * Allocate scatter gather lists.
1242  *
1243  * These are statically allocated.  Trying to be clever was not worth it.
1244  *
1245  * Dynamic allocation can fail, and we can't go deep into the memory
1246  * allocator, since we're a SCSI driver, and trying too hard to allocate
1247  * memory might generate disk I/O.  We also don't want to fail disk I/O
1248  * in that case because we can't get an allocation - the I/O could be
1249  * trying to swap out data to free memory.  Since that is pathological,
1250  * just use a statically allocated scatter list.
1251  *
1252  */
1253 static int pvscsi_allocate_sg(struct pvscsi_adapter *adapter)
1254 {
1255         struct pvscsi_ctx *ctx;
1256         int i;
1257
1258         ctx = adapter->cmd_map;
1259         BUILD_BUG_ON(sizeof(struct pvscsi_sg_list) > SGL_SIZE);
1260
1261         for (i = 0; i < adapter->req_depth; ++i, ++ctx) {
1262                 ctx->sgl = (void *)__get_free_pages(GFP_KERNEL,
1263                                                     get_order(SGL_SIZE));
1264                 ctx->sglPA = 0;
1265                 BUG_ON(!IS_ALIGNED(((unsigned long)ctx->sgl), PAGE_SIZE));
1266                 if (!ctx->sgl) {
1267                         for (; i >= 0; --i, --ctx) {
1268                                 free_pages((unsigned long)ctx->sgl,
1269                                            get_order(SGL_SIZE));
1270                                 ctx->sgl = NULL;
1271                         }
1272                         return -ENOMEM;
1273                 }
1274         }
1275
1276         return 0;
1277 }
1278
1279 /*
1280  * Query the device, fetch the config info and return the
1281  * maximum number of targets on the adapter. In case of
1282  * failure due to any reason return default i.e. 16.
1283  */
1284 static u32 pvscsi_get_max_targets(struct pvscsi_adapter *adapter)
1285 {
1286         struct PVSCSICmdDescConfigCmd cmd;
1287         struct PVSCSIConfigPageHeader *header;
1288         struct device *dev;
1289         dma_addr_t configPagePA;
1290         void *config_page;
1291         u32 numPhys = 16;
1292
1293         dev = pvscsi_dev(adapter);
1294         config_page = pci_alloc_consistent(adapter->dev, PAGE_SIZE,
1295                                            &configPagePA);
1296         if (!config_page) {
1297                 dev_warn(dev, "vmw_pvscsi: failed to allocate memory for config page\n");
1298                 goto exit;
1299         }
1300         BUG_ON(configPagePA & ~PAGE_MASK);
1301
1302         /* Fetch config info from the device. */
1303         cmd.configPageAddress = ((u64)PVSCSI_CONFIG_CONTROLLER_ADDRESS) << 32;
1304         cmd.configPageNum = PVSCSI_CONFIG_PAGE_CONTROLLER;
1305         cmd.cmpAddr = configPagePA;
1306         cmd._pad = 0;
1307
1308         /*
1309          * Mark the completion page header with error values. If the device
1310          * completes the command successfully, it sets the status values to
1311          * indicate success.
1312          */
1313         header = config_page;
1314         memset(header, 0, sizeof *header);
1315         header->hostStatus = BTSTAT_INVPARAM;
1316         header->scsiStatus = SDSTAT_CHECK;
1317
1318         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_CONFIG, &cmd, sizeof cmd);
1319
1320         if (header->hostStatus == BTSTAT_SUCCESS &&
1321             header->scsiStatus == SDSTAT_GOOD) {
1322                 struct PVSCSIConfigPageController *config;
1323
1324                 config = config_page;
1325                 numPhys = config->numPhys;
1326         } else
1327                 dev_warn(dev, "vmw_pvscsi: PVSCSI_CMD_CONFIG failed. hostStatus = 0x%x, scsiStatus = 0x%x\n",
1328                          header->hostStatus, header->scsiStatus);
1329         pci_free_consistent(adapter->dev, PAGE_SIZE, config_page, configPagePA);
1330 exit:
1331         return numPhys;
1332 }
1333
1334 static int pvscsi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1335 {
1336         struct pvscsi_adapter *adapter;
1337         struct pvscsi_adapter adapter_temp;
1338         struct Scsi_Host *host = NULL;
1339         unsigned int i;
1340         unsigned long flags = 0;
1341         int error;
1342         u32 max_id;
1343
1344         error = -ENODEV;
1345
1346         if (pci_enable_device(pdev))
1347                 return error;
1348
1349         if (pci_set_dma_mask(pdev, DMA_BIT_MASK(64)) == 0 &&
1350             pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64)) == 0) {
1351                 printk(KERN_INFO "vmw_pvscsi: using 64bit dma\n");
1352         } else if (pci_set_dma_mask(pdev, DMA_BIT_MASK(32)) == 0 &&
1353                    pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32)) == 0) {
1354                 printk(KERN_INFO "vmw_pvscsi: using 32bit dma\n");
1355         } else {
1356                 printk(KERN_ERR "vmw_pvscsi: failed to set DMA mask\n");
1357                 goto out_disable_device;
1358         }
1359
1360         /*
1361          * Let's use a temp pvscsi_adapter struct until we find the number of
1362          * targets on the adapter, after that we will switch to the real
1363          * allocated struct.
1364          */
1365         adapter = &adapter_temp;
1366         memset(adapter, 0, sizeof(*adapter));
1367         adapter->dev  = pdev;
1368         adapter->rev = pdev->revision;
1369
1370         if (pci_request_regions(pdev, "vmw_pvscsi")) {
1371                 printk(KERN_ERR "vmw_pvscsi: pci memory selection failed\n");
1372                 goto out_disable_device;
1373         }
1374
1375         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
1376                 if ((pci_resource_flags(pdev, i) & PCI_BASE_ADDRESS_SPACE_IO))
1377                         continue;
1378
1379                 if (pci_resource_len(pdev, i) < PVSCSI_MEM_SPACE_SIZE)
1380                         continue;
1381
1382                 break;
1383         }
1384
1385         if (i == DEVICE_COUNT_RESOURCE) {
1386                 printk(KERN_ERR
1387                        "vmw_pvscsi: adapter has no suitable MMIO region\n");
1388                 goto out_release_resources_and_disable;
1389         }
1390
1391         adapter->mmioBase = pci_iomap(pdev, i, PVSCSI_MEM_SPACE_SIZE);
1392
1393         if (!adapter->mmioBase) {
1394                 printk(KERN_ERR
1395                        "vmw_pvscsi: can't iomap for BAR %d memsize %lu\n",
1396                        i, PVSCSI_MEM_SPACE_SIZE);
1397                 goto out_release_resources_and_disable;
1398         }
1399
1400         pci_set_master(pdev);
1401
1402         /*
1403          * Ask the device for max number of targets before deciding the
1404          * default pvscsi_ring_pages value.
1405          */
1406         max_id = pvscsi_get_max_targets(adapter);
1407         printk(KERN_INFO "vmw_pvscsi: max_id: %u\n", max_id);
1408
1409         if (pvscsi_ring_pages == 0)
1410                 /*
1411                  * Set the right default value. Up to 16 it is 8, above it is
1412                  * max.
1413                  */
1414                 pvscsi_ring_pages = (max_id > 16) ?
1415                         PVSCSI_SETUP_RINGS_MAX_NUM_PAGES :
1416                         PVSCSI_DEFAULT_NUM_PAGES_PER_RING;
1417         printk(KERN_INFO
1418                "vmw_pvscsi: setting ring_pages to %d\n",
1419                pvscsi_ring_pages);
1420
1421         pvscsi_template.can_queue =
1422                 min(PVSCSI_MAX_NUM_PAGES_REQ_RING, pvscsi_ring_pages) *
1423                 PVSCSI_MAX_NUM_REQ_ENTRIES_PER_PAGE;
1424         pvscsi_template.cmd_per_lun =
1425                 min(pvscsi_template.can_queue, pvscsi_cmd_per_lun);
1426         host = scsi_host_alloc(&pvscsi_template, sizeof(struct pvscsi_adapter));
1427         if (!host) {
1428                 printk(KERN_ERR "vmw_pvscsi: failed to allocate host\n");
1429                 goto out_release_resources_and_disable;
1430         }
1431
1432         /*
1433          * Let's use the real pvscsi_adapter struct here onwards.
1434          */
1435         adapter = shost_priv(host);
1436         memset(adapter, 0, sizeof(*adapter));
1437         adapter->dev  = pdev;
1438         adapter->host = host;
1439         /*
1440          * Copy back what we already have to the allocated adapter struct.
1441          */
1442         adapter->rev = adapter_temp.rev;
1443         adapter->mmioBase = adapter_temp.mmioBase;
1444
1445         spin_lock_init(&adapter->hw_lock);
1446         host->max_channel = 0;
1447         host->max_lun     = 1;
1448         host->max_cmd_len = 16;
1449         host->max_id      = max_id;
1450
1451         pci_set_drvdata(pdev, host);
1452
1453         ll_adapter_reset(adapter);
1454
1455         adapter->use_msg = pvscsi_setup_msg_workqueue(adapter);
1456
1457         error = pvscsi_allocate_rings(adapter);
1458         if (error) {
1459                 printk(KERN_ERR "vmw_pvscsi: unable to allocate ring memory\n");
1460                 goto out_release_resources;
1461         }
1462
1463         /*
1464          * From this point on we should reset the adapter if anything goes
1465          * wrong.
1466          */
1467         pvscsi_setup_all_rings(adapter);
1468
1469         adapter->cmd_map = kcalloc(adapter->req_depth,
1470                                    sizeof(struct pvscsi_ctx), GFP_KERNEL);
1471         if (!adapter->cmd_map) {
1472                 printk(KERN_ERR "vmw_pvscsi: failed to allocate memory.\n");
1473                 error = -ENOMEM;
1474                 goto out_reset_adapter;
1475         }
1476
1477         INIT_LIST_HEAD(&adapter->cmd_pool);
1478         for (i = 0; i < adapter->req_depth; i++) {
1479                 struct pvscsi_ctx *ctx = adapter->cmd_map + i;
1480                 list_add(&ctx->list, &adapter->cmd_pool);
1481         }
1482
1483         error = pvscsi_allocate_sg(adapter);
1484         if (error) {
1485                 printk(KERN_ERR "vmw_pvscsi: unable to allocate s/g table\n");
1486                 goto out_reset_adapter;
1487         }
1488
1489         if (!pvscsi_disable_msix &&
1490             pvscsi_setup_msix(adapter, &adapter->irq) == 0) {
1491                 printk(KERN_INFO "vmw_pvscsi: using MSI-X\n");
1492                 adapter->use_msix = 1;
1493         } else if (!pvscsi_disable_msi && pci_enable_msi(pdev) == 0) {
1494                 printk(KERN_INFO "vmw_pvscsi: using MSI\n");
1495                 adapter->use_msi = 1;
1496                 adapter->irq = pdev->irq;
1497         } else {
1498                 printk(KERN_INFO "vmw_pvscsi: using INTx\n");
1499                 adapter->irq = pdev->irq;
1500                 flags = IRQF_SHARED;
1501         }
1502
1503         adapter->use_req_threshold = pvscsi_setup_req_threshold(adapter, true);
1504         printk(KERN_DEBUG "vmw_pvscsi: driver-based request coalescing %sabled\n",
1505                adapter->use_req_threshold ? "en" : "dis");
1506
1507         error = request_irq(adapter->irq, pvscsi_isr, flags,
1508                             "vmw_pvscsi", adapter);
1509         if (error) {
1510                 printk(KERN_ERR
1511                        "vmw_pvscsi: unable to request IRQ: %d\n", error);
1512                 adapter->irq = 0;
1513                 goto out_reset_adapter;
1514         }
1515
1516         error = scsi_add_host(host, &pdev->dev);
1517         if (error) {
1518                 printk(KERN_ERR
1519                        "vmw_pvscsi: scsi_add_host failed: %d\n", error);
1520                 goto out_reset_adapter;
1521         }
1522
1523         dev_info(&pdev->dev, "VMware PVSCSI rev %d host #%u\n",
1524                  adapter->rev, host->host_no);
1525
1526         pvscsi_unmask_intr(adapter);
1527
1528         scsi_scan_host(host);
1529
1530         return 0;
1531
1532 out_reset_adapter:
1533         ll_adapter_reset(adapter);
1534 out_release_resources:
1535         pvscsi_release_resources(adapter);
1536         scsi_host_put(host);
1537 out_disable_device:
1538         pci_disable_device(pdev);
1539
1540         return error;
1541
1542 out_release_resources_and_disable:
1543         pvscsi_release_resources(adapter);
1544         goto out_disable_device;
1545 }
1546
1547 static void __pvscsi_shutdown(struct pvscsi_adapter *adapter)
1548 {
1549         pvscsi_mask_intr(adapter);
1550
1551         if (adapter->workqueue)
1552                 flush_workqueue(adapter->workqueue);
1553
1554         pvscsi_shutdown_intr(adapter);
1555
1556         pvscsi_process_request_ring(adapter);
1557         pvscsi_process_completion_ring(adapter);
1558         ll_adapter_reset(adapter);
1559 }
1560
1561 static void pvscsi_shutdown(struct pci_dev *dev)
1562 {
1563         struct Scsi_Host *host = pci_get_drvdata(dev);
1564         struct pvscsi_adapter *adapter = shost_priv(host);
1565
1566         __pvscsi_shutdown(adapter);
1567 }
1568
1569 static void pvscsi_remove(struct pci_dev *pdev)
1570 {
1571         struct Scsi_Host *host = pci_get_drvdata(pdev);
1572         struct pvscsi_adapter *adapter = shost_priv(host);
1573
1574         scsi_remove_host(host);
1575
1576         __pvscsi_shutdown(adapter);
1577         pvscsi_release_resources(adapter);
1578
1579         scsi_host_put(host);
1580
1581         pci_disable_device(pdev);
1582 }
1583
1584 static struct pci_driver pvscsi_pci_driver = {
1585         .name           = "vmw_pvscsi",
1586         .id_table       = pvscsi_pci_tbl,
1587         .probe          = pvscsi_probe,
1588         .remove         = pvscsi_remove,
1589         .shutdown       = pvscsi_shutdown,
1590 };
1591
1592 static int __init pvscsi_init(void)
1593 {
1594         pr_info("%s - version %s\n",
1595                 PVSCSI_LINUX_DRIVER_DESC, PVSCSI_DRIVER_VERSION_STRING);
1596         return pci_register_driver(&pvscsi_pci_driver);
1597 }
1598
1599 static void __exit pvscsi_exit(void)
1600 {
1601         pci_unregister_driver(&pvscsi_pci_driver);
1602 }
1603
1604 module_init(pvscsi_init);
1605 module_exit(pvscsi_exit);