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