drivers/scsi/vmw_pvscsi.c

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