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