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++) {
 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 = pci_map_single(adapter->dev, ctx->sgl,
 376                                                     SGL_SIZE, PCI_DMA_TODEVICE);
 377                         if (pci_dma_mapping_error(adapter->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 = pci_map_single(adapter->dev, sg, bufflen,
 393                                              cmd->sc_data_direction);
 394                 if (pci_dma_mapping_error(adapter->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                                 pci_unmap_single(adapter->dev, ctx->sglPA,
 421                                                  SGL_SIZE, PCI_DMA_TODEVICE);
 422                                 ctx->sglPA = 0;
 423                         }
 424                 } else
 425                         pci_unmap_single(adapter->dev, ctx->dataPA, bufflen,
 426                                          cmd->sc_data_direction);
 427         }
 428         if (cmd->sense_buffer)
 429                 pci_unmap_single(adapter->dev, ctx->sensePA,
 430                                  SCSI_SENSE_BUFFERSIZE, PCI_DMA_FROMDEVICE);
 431 }
 432
 433 static int pvscsi_allocate_rings(struct pvscsi_adapter *adapter)
 434 {
 435         adapter->rings_state = pci_alloc_consistent(adapter->dev, PAGE_SIZE,
 436                                                     &adapter->ringStatePA);
 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 = pci_alloc_consistent(adapter->dev,
 445                                                  adapter->req_pages * PAGE_SIZE,
 446                                                  &adapter->reqRingPA);
 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 = pci_alloc_consistent(adapter->dev,
 453                                                  adapter->cmp_pages * PAGE_SIZE,
 454                                                  &adapter->cmpRingPA);
 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 = pci_alloc_consistent(adapter->dev,
 468                                                  adapter->msg_pages * PAGE_SIZE,
 469                                                  &adapter->msgRingPA);
 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                         /*
 578                          * Commands like INQUIRY may transfer less data than
 579                          * requested by the initiator via bufflen. Set residual
 580                          * count to make upper layer aware of the actual amount
 581                          * of data returned.
 582                          */
 583                         scsi_set_resid(cmd, scsi_bufflen(cmd) - e->dataLen);
 584                         cmd->result = (DID_OK << 16);
 585                         break;
 586
 587                 case BTSTAT_DATARUN:
 588                 case BTSTAT_DATA_UNDERRUN:
 589                         /* Report residual data in underruns */
 590                         scsi_set_resid(cmd, scsi_bufflen(cmd) - e->dataLen);
 591                         cmd->result = (DID_ERROR << 16);
 592                         break;
 593
 594                 case BTSTAT_SELTIMEO:
 595                         /* Our emulation returns this for non-connected devs */
 596                         cmd->result = (DID_BAD_TARGET << 16);
 597                         break;
 598
 599                 case BTSTAT_LUNMISMATCH:
 600                 case BTSTAT_TAGREJECT:
 601                 case BTSTAT_BADMSG:
 602                         cmd->result = (DRIVER_INVALID << 24);
 603                         /* fall through */
 604
 605                 case BTSTAT_HAHARDWARE:
 606                 case BTSTAT_INVPHASE:
 607                 case BTSTAT_HATIMEOUT:
 608                 case BTSTAT_NORESPONSE:
 609                 case BTSTAT_DISCONNECT:
 610                 case BTSTAT_HASOFTWARE:
 611                 case BTSTAT_BUSFREE:
 612                 case BTSTAT_SENSFAILED:
 613                         cmd->result |= (DID_ERROR << 16);
 614                         break;
 615
 616                 case BTSTAT_SENTRST:
 617                 case BTSTAT_RECVRST:
 618                 case BTSTAT_BUSRESET:
 619                         cmd->result = (DID_RESET << 16);
 620                         break;
 621
 622                 case BTSTAT_ABORTQUEUE:
 623                         cmd->result = (DID_BUS_BUSY << 16);
 624                         break;
 625
 626                 case BTSTAT_SCSIPARITY:
 627                         cmd->result = (DID_PARITY << 16);
 628                         break;
 629
 630                 default:
 631                         cmd->result = (DID_ERROR << 16);
 632                         scmd_printk(KERN_DEBUG, cmd,
 633                                     "Unknown completion status: 0x%x\n",
 634                                     btstat);
 635         }
 636
 637         dev_dbg(&cmd->device->sdev_gendev,
 638                 "cmd=%p %x ctx=%p result=0x%x status=0x%x,%x\n",
 639                 cmd, cmd->cmnd[0], ctx, cmd->result, btstat, sdstat);
 640
 641         cmd->scsi_done(cmd);
 642 }
 643
 644 /*
 645  * barrier usage : Since the PVSCSI device is emulated, there could be cases
 646  * where we may want to serialize some accesses between the driver and the
 647  * emulation layer. We use compiler barriers instead of the more expensive
 648  * memory barriers because PVSCSI is only supported on X86 which has strong
 649  * memory access ordering.
 650  */
 651 static void pvscsi_process_completion_ring(struct pvscsi_adapter *adapter)
 652 {
 653         struct PVSCSIRingsState *s = adapter->rings_state;
 654         struct PVSCSIRingCmpDesc *ring = adapter->cmp_ring;
 655         u32 cmp_entries = s->cmpNumEntriesLog2;
 656
 657         while (s->cmpConsIdx != s->cmpProdIdx) {
 658                 struct PVSCSIRingCmpDesc *e = ring + (s->cmpConsIdx &
 659                                                       MASK(cmp_entries));
 660                 /*
 661                  * This barrier() ensures that *e is not dereferenced while
 662                  * the device emulation still writes data into the slot.
 663                  * Since the device emulation advances s->cmpProdIdx only after
 664                  * updating the slot we want to check it first.
 665                  */
 666                 barrier();
 667                 pvscsi_complete_request(adapter, e);
 668                 /*
 669                  * This barrier() ensures that compiler doesn't reorder write
 670                  * to s->cmpConsIdx before the read of (*e) inside
 671                  * pvscsi_complete_request. Otherwise, device emulation may
 672                  * overwrite *e before we had a chance to read it.
 673                  */
 674                 barrier();
 675                 s->cmpConsIdx++;
 676         }
 677 }
 678
 679 /*
 680  * Translate a Linux SCSI request into a request ring entry.
 681  */
 682 static int pvscsi_queue_ring(struct pvscsi_adapter *adapter,
 683                              struct pvscsi_ctx *ctx, struct scsi_cmnd *cmd)
 684 {
 685         struct PVSCSIRingsState *s;
 686         struct PVSCSIRingReqDesc *e;
 687         struct scsi_device *sdev;
 688         u32 req_entries;
 689
 690         s = adapter->rings_state;
 691         sdev = cmd->device;
 692         req_entries = s->reqNumEntriesLog2;
 693
 694         /*
 695          * If this condition holds, we might have room on the request ring, but
 696          * we might not have room on the completion ring for the response.
 697          * However, we have already ruled out this possibility - we would not
 698          * have successfully allocated a context if it were true, since we only
 699          * have one context per request entry.  Check for it anyway, since it
 700          * would be a serious bug.
 701          */
 702         if (s->reqProdIdx - s->cmpConsIdx >= 1 << req_entries) {
 703                 scmd_printk(KERN_ERR, cmd, "vmw_pvscsi: "
 704                             "ring full: reqProdIdx=%d cmpConsIdx=%d\n",
 705                             s->reqProdIdx, s->cmpConsIdx);
 706                 return -1;
 707         }
 708
 709         e = adapter->req_ring + (s->reqProdIdx & MASK(req_entries));
 710
 711         e->bus    = sdev->channel;
 712         e->target = sdev->id;
 713         memset(e->lun, 0, sizeof(e->lun));
 714         e->lun[1] = sdev->lun;
 715
 716         if (cmd->sense_buffer) {
 717                 ctx->sensePA = pci_map_single(adapter->dev, cmd->sense_buffer,
 718                                               SCSI_SENSE_BUFFERSIZE,
 719                                               PCI_DMA_FROMDEVICE);
 720                 if (pci_dma_mapping_error(adapter->dev, ctx->sensePA)) {
 721                         scmd_printk(KERN_ERR, cmd,
 722                                     "vmw_pvscsi: Failed to map sense buffer for DMA.\n");
 723                         ctx->sensePA = 0;
 724                         return -ENOMEM;
 725                 }
 726                 e->senseAddr = ctx->sensePA;
 727                 e->senseLen = SCSI_SENSE_BUFFERSIZE;
 728         } else {
 729                 e->senseLen  = 0;
 730                 e->senseAddr = 0;
 731         }
 732         e->cdbLen   = cmd->cmd_len;
 733         e->vcpuHint = smp_processor_id();
 734         memcpy(e->cdb, cmd->cmnd, e->cdbLen);
 735
 736         e->tag = SIMPLE_QUEUE_TAG;
 737
 738         if (cmd->sc_data_direction == DMA_FROM_DEVICE)
 739                 e->flags = PVSCSI_FLAG_CMD_DIR_TOHOST;
 740         else if (cmd->sc_data_direction == DMA_TO_DEVICE)
 741                 e->flags = PVSCSI_FLAG_CMD_DIR_TODEVICE;
 742         else if (cmd->sc_data_direction == DMA_NONE)
 743                 e->flags = PVSCSI_FLAG_CMD_DIR_NONE;
 744         else
 745                 e->flags = 0;
 746
 747         if (pvscsi_map_buffers(adapter, ctx, cmd, e) != 0) {
 748                 if (cmd->sense_buffer) {
 749                         pci_unmap_single(adapter->dev, ctx->sensePA,
 750                                          SCSI_SENSE_BUFFERSIZE,
 751                                          PCI_DMA_FROMDEVICE);
 752                         ctx->sensePA = 0;
 753                 }
 754                 return -ENOMEM;
 755         }
 756
 757         e->context = pvscsi_map_context(adapter, ctx);
 758
 759         barrier();
 760
 761         s->reqProdIdx++;
 762
 763         return 0;
 764 }
 765
 766 static int pvscsi_queue_lck(struct scsi_cmnd *cmd, void (*done)(struct scsi_cmnd *))
 767 {
 768         struct Scsi_Host *host = cmd->device->host;
 769         struct pvscsi_adapter *adapter = shost_priv(host);
 770         struct pvscsi_ctx *ctx;
 771         unsigned long flags;
 772         unsigned char op;
 773
 774         spin_lock_irqsave(&adapter->hw_lock, flags);
 775
 776         ctx = pvscsi_acquire_context(adapter, cmd);
 777         if (!ctx || pvscsi_queue_ring(adapter, ctx, cmd) != 0) {
 778                 if (ctx)
 779                         pvscsi_release_context(adapter, ctx);
 780                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 781                 return SCSI_MLQUEUE_HOST_BUSY;
 782         }
 783
 784         cmd->scsi_done = done;
 785         op = cmd->cmnd[0];
 786
 787         dev_dbg(&cmd->device->sdev_gendev,
 788                 "queued cmd %p, ctx %p, op=%x\n", cmd, ctx, op);
 789
 790         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 791
 792         pvscsi_kick_io(adapter, op);
 793
 794         return 0;
 795 }
 796
 797 static DEF_SCSI_QCMD(pvscsi_queue)
 798
 799 static int pvscsi_abort(struct scsi_cmnd *cmd)
 800 {
 801         struct pvscsi_adapter *adapter = shost_priv(cmd->device->host);
 802         struct pvscsi_ctx *ctx;
 803         unsigned long flags;
 804         int result = SUCCESS;
 805         DECLARE_COMPLETION_ONSTACK(abort_cmp);
 806         int done;
 807
 808         scmd_printk(KERN_DEBUG, cmd, "task abort on host %u, %p\n",
 809                     adapter->host->host_no, cmd);
 810
 811         spin_lock_irqsave(&adapter->hw_lock, flags);
 812
 813         /*
 814          * Poll the completion ring first - we might be trying to abort
 815          * a command that is waiting to be dispatched in the completion ring.
 816          */
 817         pvscsi_process_completion_ring(adapter);
 818
 819         /*
 820          * If there is no context for the command, it either already succeeded
 821          * or else was never properly issued.  Not our problem.
 822          */
 823         ctx = pvscsi_find_context(adapter, cmd);
 824         if (!ctx) {
 825                 scmd_printk(KERN_DEBUG, cmd, "Failed to abort cmd %p\n", cmd);
 826                 goto out;
 827         }
 828
 829         /*
 830          * Mark that the command has been requested to be aborted and issue
 831          * the abort.
 832          */
 833         ctx->abort_cmp = &abort_cmp;
 834
 835         pvscsi_abort_cmd(adapter, ctx);
 836         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 837         /* Wait for 2 secs for the completion. */
 838         done = wait_for_completion_timeout(&abort_cmp, msecs_to_jiffies(2000));
 839         spin_lock_irqsave(&adapter->hw_lock, flags);
 840
 841         if (!done) {
 842                 /*
 843                  * Failed to abort the command, unmark the fact that it
 844                  * was requested to be aborted.
 845                  */
 846                 ctx->abort_cmp = NULL;
 847                 result = FAILED;
 848                 scmd_printk(KERN_DEBUG, cmd,
 849                             "Failed to get completion for aborted cmd %p\n",
 850                             cmd);
 851                 goto out;
 852         }
 853
 854         /*
 855          * Successfully aborted the command.
 856          */
 857         cmd->result = (DID_ABORT << 16);
 858         cmd->scsi_done(cmd);
 859
 860 out:
 861         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 862         return result;
 863 }
 864
 865 /*
 866  * Abort all outstanding requests.  This is only safe to use if the completion
 867  * ring will never be walked again or the device has been reset, because it
 868  * destroys the 1-1 mapping between context field passed to emulation and our
 869  * request structure.
 870  */
 871 static void pvscsi_reset_all(struct pvscsi_adapter *adapter)
 872 {
 873         unsigned i;
 874
 875         for (i = 0; i < adapter->req_depth; i++) {
 876                 struct pvscsi_ctx *ctx = &adapter->cmd_map[i];
 877                 struct scsi_cmnd *cmd = ctx->cmd;
 878                 if (cmd) {
 879                         scmd_printk(KERN_ERR, cmd,
 880                                     "Forced reset on cmd %p\n", cmd);
 881                         pvscsi_unmap_buffers(adapter, ctx);
 882                         pvscsi_release_context(adapter, ctx);
 883                         cmd->result = (DID_RESET << 16);
 884                         cmd->scsi_done(cmd);
 885                 }
 886         }
 887 }
 888
 889 static int pvscsi_host_reset(struct scsi_cmnd *cmd)
 890 {
 891         struct Scsi_Host *host = cmd->device->host;
 892         struct pvscsi_adapter *adapter = shost_priv(host);
 893         unsigned long flags;
 894         bool use_msg;
 895
 896         scmd_printk(KERN_INFO, cmd, "SCSI Host reset\n");
 897
 898         spin_lock_irqsave(&adapter->hw_lock, flags);
 899
 900         use_msg = adapter->use_msg;
 901
 902         if (use_msg) {
 903                 adapter->use_msg = 0;
 904                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 905
 906                 /*
 907                  * Now that we know that the ISR won't add more work on the
 908                  * workqueue we can safely flush any outstanding work.
 909                  */
 910                 flush_workqueue(adapter->workqueue);
 911                 spin_lock_irqsave(&adapter->hw_lock, flags);
 912         }
 913
 914         /*
 915          * We're going to tear down the entire ring structure and set it back
 916          * up, so stalling new requests until all completions are flushed and
 917          * the rings are back in place.
 918          */
 919
 920         pvscsi_process_request_ring(adapter);
 921
 922         ll_adapter_reset(adapter);
 923
 924         /*
 925          * Now process any completions.  Note we do this AFTER adapter reset,
 926          * which is strange, but stops races where completions get posted
 927          * between processing the ring and issuing the reset.  The backend will
 928          * not touch the ring memory after reset, so the immediately pre-reset
 929          * completion ring state is still valid.
 930          */
 931         pvscsi_process_completion_ring(adapter);
 932
 933         pvscsi_reset_all(adapter);
 934         adapter->use_msg = use_msg;
 935         pvscsi_setup_all_rings(adapter);
 936         pvscsi_unmask_intr(adapter);
 937
 938         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 939
 940         return SUCCESS;
 941 }
 942
 943 static int pvscsi_bus_reset(struct scsi_cmnd *cmd)
 944 {
 945         struct Scsi_Host *host = cmd->device->host;
 946         struct pvscsi_adapter *adapter = shost_priv(host);
 947         unsigned long flags;
 948
 949         scmd_printk(KERN_INFO, cmd, "SCSI Bus reset\n");
 950
 951         /*
 952          * We don't want to queue new requests for this bus after
 953          * flushing all pending requests to emulation, since new
 954          * requests could then sneak in during this bus reset phase,
 955          * so take the lock now.
 956          */
 957         spin_lock_irqsave(&adapter->hw_lock, flags);
 958
 959         pvscsi_process_request_ring(adapter);
 960         ll_bus_reset(adapter);
 961         pvscsi_process_completion_ring(adapter);
 962
 963         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 964
 965         return SUCCESS;
 966 }
 967
 968 static int pvscsi_device_reset(struct scsi_cmnd *cmd)
 969 {
 970         struct Scsi_Host *host = cmd->device->host;
 971         struct pvscsi_adapter *adapter = shost_priv(host);
 972         unsigned long flags;
 973
 974         scmd_printk(KERN_INFO, cmd, "SCSI device reset on scsi%u:%u\n",
 975                     host->host_no, cmd->device->id);
 976
 977         /*
 978          * We don't want to queue new requests for this device after flushing
 979          * all pending requests to emulation, since new requests could then
 980          * sneak in during this device reset phase, so take the lock now.
 981          */
 982         spin_lock_irqsave(&adapter->hw_lock, flags);
 983
 984         pvscsi_process_request_ring(adapter);
 985         ll_device_reset(adapter, cmd->device->id);
 986         pvscsi_process_completion_ring(adapter);
 987
 988         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 989
 990         return SUCCESS;
 991 }
 992
 993 static struct scsi_host_template pvscsi_template;
 994
 995 static const char *pvscsi_info(struct Scsi_Host *host)
 996 {
 997         struct pvscsi_adapter *adapter = shost_priv(host);
 998         static char buf[256];
 999
1000         sprintf(buf, "VMware PVSCSI storage adapter rev %d, req/cmp/msg rings: "
1001                 "%u/%u/%u pages, cmd_per_lun=%u", adapter->rev,
1002                 adapter->req_pages, adapter->cmp_pages, adapter->msg_pages,
1003                 pvscsi_template.cmd_per_lun);
1004
1005         return buf;
1006 }
1007
1008 static struct scsi_host_template pvscsi_template = {
1009         .module                         = THIS_MODULE,
1010         .name                           = "VMware PVSCSI Host Adapter",
1011         .proc_name                      = "vmw_pvscsi",
1012         .info                           = pvscsi_info,
1013         .queuecommand                   = pvscsi_queue,
1014         .this_id                        = -1,
1015         .sg_tablesize                   = PVSCSI_MAX_NUM_SG_ENTRIES_PER_SEGMENT,
1016         .dma_boundary                   = UINT_MAX,
1017         .max_sectors                    = 0xffff,
1018         .use_clustering                 = ENABLE_CLUSTERING,
1019         .change_queue_depth             = pvscsi_change_queue_depth,
1020         .eh_abort_handler               = pvscsi_abort,
1021         .eh_device_reset_handler        = pvscsi_device_reset,
1022         .eh_bus_reset_handler           = pvscsi_bus_reset,
1023         .eh_host_reset_handler          = pvscsi_host_reset,
1024 };
1025
1026 static void pvscsi_process_msg(const struct pvscsi_adapter *adapter,
1027                                const struct PVSCSIRingMsgDesc *e)
1028 {
1029         struct PVSCSIRingsState *s = adapter->rings_state;
1030         struct Scsi_Host *host = adapter->host;
1031         struct scsi_device *sdev;
1032
1033         printk(KERN_INFO "vmw_pvscsi: msg type: 0x%x - MSG RING: %u/%u (%u) \n",
1034                e->type, s->msgProdIdx, s->msgConsIdx, s->msgNumEntriesLog2);
1035
1036         BUILD_BUG_ON(PVSCSI_MSG_LAST != 2);
1037
1038         if (e->type == PVSCSI_MSG_DEV_ADDED) {
1039                 struct PVSCSIMsgDescDevStatusChanged *desc;
1040                 desc = (struct PVSCSIMsgDescDevStatusChanged *)e;
1041
1042                 printk(KERN_INFO
1043                        "vmw_pvscsi: msg: device added at scsi%u:%u:%u\n",
1044                        desc->bus, desc->target, desc->lun[1]);
1045
1046                 if (!scsi_host_get(host))
1047                         return;
1048
1049                 sdev = scsi_device_lookup(host, desc->bus, desc->target,
1050                                           desc->lun[1]);
1051                 if (sdev) {
1052                         printk(KERN_INFO "vmw_pvscsi: device already exists\n");
1053                         scsi_device_put(sdev);
1054                 } else
1055                         scsi_add_device(adapter->host, desc->bus,
1056                                         desc->target, desc->lun[1]);
1057
1058                 scsi_host_put(host);
1059         } else if (e->type == PVSCSI_MSG_DEV_REMOVED) {
1060                 struct PVSCSIMsgDescDevStatusChanged *desc;
1061                 desc = (struct PVSCSIMsgDescDevStatusChanged *)e;
1062
1063                 printk(KERN_INFO
1064                        "vmw_pvscsi: msg: device removed at scsi%u:%u:%u\n",
1065                        desc->bus, desc->target, desc->lun[1]);
1066
1067                 if (!scsi_host_get(host))
1068                         return;
1069
1070                 sdev = scsi_device_lookup(host, desc->bus, desc->target,
1071                                           desc->lun[1]);
1072                 if (sdev) {
1073                         scsi_remove_device(sdev);
1074                         scsi_device_put(sdev);
1075                 } else
1076                         printk(KERN_INFO
1077                                "vmw_pvscsi: failed to lookup scsi%u:%u:%u\n",
1078                                desc->bus, desc->target, desc->lun[1]);
1079
1080                 scsi_host_put(host);
1081         }
1082 }
1083
1084 static int pvscsi_msg_pending(const struct pvscsi_adapter *adapter)
1085 {
1086         struct PVSCSIRingsState *s = adapter->rings_state;
1087
1088         return s->msgProdIdx != s->msgConsIdx;
1089 }
1090
1091 static void pvscsi_process_msg_ring(const struct pvscsi_adapter *adapter)
1092 {
1093         struct PVSCSIRingsState *s = adapter->rings_state;
1094         struct PVSCSIRingMsgDesc *ring = adapter->msg_ring;
1095         u32 msg_entries = s->msgNumEntriesLog2;
1096
1097         while (pvscsi_msg_pending(adapter)) {
1098                 struct PVSCSIRingMsgDesc *e = ring + (s->msgConsIdx &
1099                                                       MASK(msg_entries));
1100
1101                 barrier();
1102                 pvscsi_process_msg(adapter, e);
1103                 barrier();
1104                 s->msgConsIdx++;
1105         }
1106 }
1107
1108 static void pvscsi_msg_workqueue_handler(struct work_struct *data)
1109 {
1110         struct pvscsi_adapter *adapter;
1111
1112         adapter = container_of(data, struct pvscsi_adapter, work);
1113
1114         pvscsi_process_msg_ring(adapter);
1115 }
1116
1117 static int pvscsi_setup_msg_workqueue(struct pvscsi_adapter *adapter)
1118 {
1119         char name[32];
1120
1121         if (!pvscsi_use_msg)
1122                 return 0;
1123
1124         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
1125                          PVSCSI_CMD_SETUP_MSG_RING);
1126
1127         if (pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS) == -1)
1128                 return 0;
1129
1130         snprintf(name, sizeof(name),
1131                  "vmw_pvscsi_wq_%u", adapter->host->host_no);
1132
1133         adapter->workqueue = create_singlethread_workqueue(name);
1134         if (!adapter->workqueue) {
1135                 printk(KERN_ERR "vmw_pvscsi: failed to create work queue\n");
1136                 return 0;
1137         }
1138         INIT_WORK(&adapter->work, pvscsi_msg_workqueue_handler);
1139
1140         return 1;
1141 }
1142
1143 static bool pvscsi_setup_req_threshold(struct pvscsi_adapter *adapter,
1144                                       bool enable)
1145 {
1146         u32 val;
1147
1148         if (!pvscsi_use_req_threshold)
1149                 return false;
1150
1151         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
1152                          PVSCSI_CMD_SETUP_REQCALLTHRESHOLD);
1153         val = pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS);
1154         if (val == -1) {
1155                 printk(KERN_INFO "vmw_pvscsi: device does not support req_threshold\n");
1156                 return false;
1157         } else {
1158                 struct PVSCSICmdDescSetupReqCall cmd_msg = { 0 };
1159                 cmd_msg.enable = enable;
1160                 printk(KERN_INFO
1161                        "vmw_pvscsi: %sabling reqCallThreshold\n",
1162                         enable ? "en" : "dis");
1163                 pvscsi_write_cmd_desc(adapter,
1164                                       PVSCSI_CMD_SETUP_REQCALLTHRESHOLD,
1165                                       &cmd_msg, sizeof(cmd_msg));
1166                 return pvscsi_reg_read(adapter,
1167                                        PVSCSI_REG_OFFSET_COMMAND_STATUS) != 0;
1168         }
1169 }
1170
1171 static irqreturn_t pvscsi_isr(int irq, void *devp)
1172 {
1173         struct pvscsi_adapter *adapter = devp;
1174         unsigned long flags;
1175
1176         spin_lock_irqsave(&adapter->hw_lock, flags);
1177         pvscsi_process_completion_ring(adapter);
1178         if (adapter->use_msg && pvscsi_msg_pending(adapter))
1179                 queue_work(adapter->workqueue, &adapter->work);
1180         spin_unlock_irqrestore(&adapter->hw_lock, flags);
1181
1182         return IRQ_HANDLED;
1183 }
1184
1185 static irqreturn_t pvscsi_shared_isr(int irq, void *devp)
1186 {
1187         struct pvscsi_adapter *adapter = devp;
1188         u32 val = pvscsi_read_intr_status(adapter);
1189
1190         if (!(val & PVSCSI_INTR_ALL_SUPPORTED))
1191                 return IRQ_NONE;
1192         pvscsi_write_intr_status(devp, val);
1193         return pvscsi_isr(irq, devp);
1194 }
1195
1196 static void pvscsi_free_sgls(const struct pvscsi_adapter *adapter)
1197 {
1198         struct pvscsi_ctx *ctx = adapter->cmd_map;
1199         unsigned i;
1200
1201         for (i = 0; i < adapter->req_depth; ++i, ++ctx)
1202                 free_pages((unsigned long)ctx->sgl, get_order(SGL_SIZE));
1203 }
1204
1205 static void pvscsi_shutdown_intr(struct pvscsi_adapter *adapter)
1206 {
1207         free_irq(pci_irq_vector(adapter->dev, 0), adapter);
1208         pci_free_irq_vectors(adapter->dev);
1209 }
1210
1211 static void pvscsi_release_resources(struct pvscsi_adapter *adapter)
1212 {
1213         if (adapter->workqueue)
1214                 destroy_workqueue(adapter->workqueue);
1215
1216         if (adapter->mmioBase)
1217                 pci_iounmap(adapter->dev, adapter->mmioBase);
1218
1219         pci_release_regions(adapter->dev);
1220
1221         if (adapter->cmd_map) {
1222                 pvscsi_free_sgls(adapter);
1223                 kfree(adapter->cmd_map);
1224         }
1225
1226         if (adapter->rings_state)
1227                 pci_free_consistent(adapter->dev, PAGE_SIZE,
1228                                     adapter->rings_state, adapter->ringStatePA);
1229
1230         if (adapter->req_ring)
1231                 pci_free_consistent(adapter->dev,
1232                                     adapter->req_pages * PAGE_SIZE,
1233                                     adapter->req_ring, adapter->reqRingPA);
1234
1235         if (adapter->cmp_ring)
1236                 pci_free_consistent(adapter->dev,
1237                                     adapter->cmp_pages * PAGE_SIZE,
1238                                     adapter->cmp_ring, adapter->cmpRingPA);
1239
1240         if (adapter->msg_ring)
1241                 pci_free_consistent(adapter->dev,
1242                                     adapter->msg_pages * PAGE_SIZE,
1243                                     adapter->msg_ring, adapter->msgRingPA);
1244 }
1245
1246 /*
1247  * Allocate scatter gather lists.
1248  *
1249  * These are statically allocated.  Trying to be clever was not worth it.
1250  *
1251  * Dynamic allocation can fail, and we can't go deep into the memory
1252  * allocator, since we're a SCSI driver, and trying too hard to allocate
1253  * memory might generate disk I/O.  We also don't want to fail disk I/O
1254  * in that case because we can't get an allocation - the I/O could be
1255  * trying to swap out data to free memory.  Since that is pathological,
1256  * just use a statically allocated scatter list.
1257  *
1258  */
1259 static int pvscsi_allocate_sg(struct pvscsi_adapter *adapter)
1260 {
1261         struct pvscsi_ctx *ctx;
1262         int i;
1263
1264         ctx = adapter->cmd_map;
1265         BUILD_BUG_ON(sizeof(struct pvscsi_sg_list) > SGL_SIZE);
1266
1267         for (i = 0; i < adapter->req_depth; ++i, ++ctx) {
1268                 ctx->sgl = (void *)__get_free_pages(GFP_KERNEL,
1269                                                     get_order(SGL_SIZE));
1270                 ctx->sglPA = 0;
1271                 BUG_ON(!IS_ALIGNED(((unsigned long)ctx->sgl), PAGE_SIZE));
1272                 if (!ctx->sgl) {
1273                         for (; i >= 0; --i, --ctx) {
1274                                 free_pages((unsigned long)ctx->sgl,
1275                                            get_order(SGL_SIZE));
1276                                 ctx->sgl = NULL;
1277                         }
1278                         return -ENOMEM;
1279                 }
1280         }
1281
1282         return 0;
1283 }
1284
1285 /*
1286  * Query the device, fetch the config info and return the
1287  * maximum number of targets on the adapter. In case of
1288  * failure due to any reason return default i.e. 16.
1289  */
1290 static u32 pvscsi_get_max_targets(struct pvscsi_adapter *adapter)
1291 {
1292         struct PVSCSICmdDescConfigCmd cmd;
1293         struct PVSCSIConfigPageHeader *header;
1294         struct device *dev;
1295         dma_addr_t configPagePA;
1296         void *config_page;
1297         u32 numPhys = 16;
1298
1299         dev = pvscsi_dev(adapter);
1300         config_page = pci_alloc_consistent(adapter->dev, PAGE_SIZE,
1301                                            &configPagePA);
1302         if (!config_page) {
1303                 dev_warn(dev, "vmw_pvscsi: failed to allocate memory for config page\n");
1304                 goto exit;
1305         }
1306         BUG_ON(configPagePA & ~PAGE_MASK);
1307
1308         /* Fetch config info from the device. */
1309         cmd.configPageAddress = ((u64)PVSCSI_CONFIG_CONTROLLER_ADDRESS) << 32;
1310         cmd.configPageNum = PVSCSI_CONFIG_PAGE_CONTROLLER;
1311         cmd.cmpAddr = configPagePA;
1312         cmd._pad = 0;
1313
1314         /*
1315          * Mark the completion page header with error values. If the device
1316          * completes the command successfully, it sets the status values to
1317          * indicate success.
1318          */
1319         header = config_page;
1320         memset(header, 0, sizeof *header);
1321         header->hostStatus = BTSTAT_INVPARAM;
1322         header->scsiStatus = SDSTAT_CHECK;
1323
1324         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_CONFIG, &cmd, sizeof cmd);
1325
1326         if (header->hostStatus == BTSTAT_SUCCESS &&
1327             header->scsiStatus == SDSTAT_GOOD) {
1328                 struct PVSCSIConfigPageController *config;
1329
1330                 config = config_page;
1331                 numPhys = config->numPhys;
1332         } else
1333                 dev_warn(dev, "vmw_pvscsi: PVSCSI_CMD_CONFIG failed. hostStatus = 0x%x, scsiStatus = 0x%x\n",
1334                          header->hostStatus, header->scsiStatus);
1335         pci_free_consistent(adapter->dev, PAGE_SIZE, config_page, configPagePA);
1336 exit:
1337         return numPhys;
1338 }
1339
1340 static int pvscsi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1341 {
1342         unsigned int irq_flag = PCI_IRQ_MSIX | PCI_IRQ_MSI | PCI_IRQ_LEGACY;
1343         struct pvscsi_adapter *adapter;
1344         struct pvscsi_adapter adapter_temp;
1345         struct Scsi_Host *host = NULL;
1346         unsigned int i;
1347         int error;
1348         u32 max_id;
1349
1350         error = -ENODEV;
1351
1352         if (pci_enable_device(pdev))
1353                 return error;
1354
1355         if (pci_set_dma_mask(pdev, DMA_BIT_MASK(64)) == 0 &&
1356             pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64)) == 0) {
1357                 printk(KERN_INFO "vmw_pvscsi: using 64bit dma\n");
1358         } else if (pci_set_dma_mask(pdev, DMA_BIT_MASK(32)) == 0 &&
1359                    pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32)) == 0) {
1360                 printk(KERN_INFO "vmw_pvscsi: using 32bit dma\n");
1361         } else {
1362                 printk(KERN_ERR "vmw_pvscsi: failed to set DMA mask\n");
1363                 goto out_disable_device;
1364         }
1365
1366         /*
1367          * Let's use a temp pvscsi_adapter struct until we find the number of
1368          * targets on the adapter, after that we will switch to the real
1369          * allocated struct.
1370          */
1371         adapter = &adapter_temp;
1372         memset(adapter, 0, sizeof(*adapter));
1373         adapter->dev  = pdev;
1374         adapter->rev = pdev->revision;
1375
1376         if (pci_request_regions(pdev, "vmw_pvscsi")) {
1377                 printk(KERN_ERR "vmw_pvscsi: pci memory selection failed\n");
1378                 goto out_disable_device;
1379         }
1380
1381         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
1382                 if ((pci_resource_flags(pdev, i) & PCI_BASE_ADDRESS_SPACE_IO))
1383                         continue;
1384
1385                 if (pci_resource_len(pdev, i) < PVSCSI_MEM_SPACE_SIZE)
1386                         continue;
1387
1388                 break;
1389         }
1390
1391         if (i == DEVICE_COUNT_RESOURCE) {
1392                 printk(KERN_ERR
1393                        "vmw_pvscsi: adapter has no suitable MMIO region\n");
1394                 goto out_release_resources_and_disable;
1395         }
1396
1397         adapter->mmioBase = pci_iomap(pdev, i, PVSCSI_MEM_SPACE_SIZE);
1398
1399         if (!adapter->mmioBase) {
1400                 printk(KERN_ERR
1401                        "vmw_pvscsi: can't iomap for BAR %d memsize %lu\n",
1402                        i, PVSCSI_MEM_SPACE_SIZE);
1403                 goto out_release_resources_and_disable;
1404         }
1405
1406         pci_set_master(pdev);
1407
1408         /*
1409          * Ask the device for max number of targets before deciding the
1410          * default pvscsi_ring_pages value.
1411          */
1412         max_id = pvscsi_get_max_targets(adapter);
1413         printk(KERN_INFO "vmw_pvscsi: max_id: %u\n", max_id);
1414
1415         if (pvscsi_ring_pages == 0)
1416                 /*
1417                  * Set the right default value. Up to 16 it is 8, above it is
1418                  * max.
1419                  */
1420                 pvscsi_ring_pages = (max_id > 16) ?
1421                         PVSCSI_SETUP_RINGS_MAX_NUM_PAGES :
1422                         PVSCSI_DEFAULT_NUM_PAGES_PER_RING;
1423         printk(KERN_INFO
1424                "vmw_pvscsi: setting ring_pages to %d\n",
1425                pvscsi_ring_pages);
1426
1427         pvscsi_template.can_queue =
1428                 min(PVSCSI_MAX_NUM_PAGES_REQ_RING, pvscsi_ring_pages) *
1429                 PVSCSI_MAX_NUM_REQ_ENTRIES_PER_PAGE;
1430         pvscsi_template.cmd_per_lun =
1431                 min(pvscsi_template.can_queue, pvscsi_cmd_per_lun);
1432         host = scsi_host_alloc(&pvscsi_template, sizeof(struct pvscsi_adapter));
1433         if (!host) {
1434                 printk(KERN_ERR "vmw_pvscsi: failed to allocate host\n");
1435                 goto out_release_resources_and_disable;
1436         }
1437
1438         /*
1439          * Let's use the real pvscsi_adapter struct here onwards.
1440          */
1441         adapter = shost_priv(host);
1442         memset(adapter, 0, sizeof(*adapter));
1443         adapter->dev  = pdev;
1444         adapter->host = host;
1445         /*
1446          * Copy back what we already have to the allocated adapter struct.
1447          */
1448         adapter->rev = adapter_temp.rev;
1449         adapter->mmioBase = adapter_temp.mmioBase;
1450
1451         spin_lock_init(&adapter->hw_lock);
1452         host->max_channel = 0;
1453         host->max_lun     = 1;
1454         host->max_cmd_len = 16;
1455         host->max_id      = max_id;
1456
1457         pci_set_drvdata(pdev, host);
1458
1459         ll_adapter_reset(adapter);
1460
1461         adapter->use_msg = pvscsi_setup_msg_workqueue(adapter);
1462
1463         error = pvscsi_allocate_rings(adapter);
1464         if (error) {
1465                 printk(KERN_ERR "vmw_pvscsi: unable to allocate ring memory\n");
1466                 goto out_release_resources;
1467         }
1468
1469         /*
1470          * From this point on we should reset the adapter if anything goes
1471          * wrong.
1472          */
1473         pvscsi_setup_all_rings(adapter);
1474
1475         adapter->cmd_map = kcalloc(adapter->req_depth,
1476                                    sizeof(struct pvscsi_ctx), GFP_KERNEL);
1477         if (!adapter->cmd_map) {
1478                 printk(KERN_ERR "vmw_pvscsi: failed to allocate memory.\n");
1479                 error = -ENOMEM;
1480                 goto out_reset_adapter;
1481         }
1482
1483         INIT_LIST_HEAD(&adapter->cmd_pool);
1484         for (i = 0; i < adapter->req_depth; i++) {
1485                 struct pvscsi_ctx *ctx = adapter->cmd_map + i;
1486                 list_add(&ctx->list, &adapter->cmd_pool);
1487         }
1488
1489         error = pvscsi_allocate_sg(adapter);
1490         if (error) {
1491                 printk(KERN_ERR "vmw_pvscsi: unable to allocate s/g table\n");
1492                 goto out_reset_adapter;
1493         }
1494
1495         if (pvscsi_disable_msix)
1496                 irq_flag &= ~PCI_IRQ_MSIX;
1497         if (pvscsi_disable_msi)
1498                 irq_flag &= ~PCI_IRQ_MSI;
1499
1500         error = pci_alloc_irq_vectors(adapter->dev, 1, 1, irq_flag);
1501         if (error < 0)
1502                 goto out_reset_adapter;
1503
1504         adapter->use_req_threshold = pvscsi_setup_req_threshold(adapter, true);
1505         printk(KERN_DEBUG "vmw_pvscsi: driver-based request coalescing %sabled\n",
1506                adapter->use_req_threshold ? "en" : "dis");
1507
1508         if (adapter->dev->msix_enabled || adapter->dev->msi_enabled) {
1509                 printk(KERN_INFO "vmw_pvscsi: using MSI%s\n",
1510                         adapter->dev->msix_enabled ? "-X" : "");
1511                 error = request_irq(pci_irq_vector(pdev, 0), pvscsi_isr,
1512                                 0, "vmw_pvscsi", adapter);
1513         } else {
1514                 printk(KERN_INFO "vmw_pvscsi: using INTx\n");
1515                 error = request_irq(pci_irq_vector(pdev, 0), pvscsi_shared_isr,
1516                                 IRQF_SHARED, "vmw_pvscsi", adapter);
1517         }
1518
1519         if (error) {
1520                 printk(KERN_ERR
1521                        "vmw_pvscsi: unable to request IRQ: %d\n", error);
1522                 goto out_reset_adapter;
1523         }
1524
1525         error = scsi_add_host(host, &pdev->dev);
1526         if (error) {
1527                 printk(KERN_ERR
1528                        "vmw_pvscsi: scsi_add_host failed: %d\n", error);
1529                 goto out_reset_adapter;
1530         }
1531
1532         dev_info(&pdev->dev, "VMware PVSCSI rev %d host #%u\n",
1533                  adapter->rev, host->host_no);
1534
1535         pvscsi_unmask_intr(adapter);
1536
1537         scsi_scan_host(host);
1538
1539         return 0;
1540
1541 out_reset_adapter:
1542         ll_adapter_reset(adapter);
1543 out_release_resources:
1544         pvscsi_shutdown_intr(adapter);
1545         pvscsi_release_resources(adapter);
1546         scsi_host_put(host);
1547 out_disable_device:
1548         pci_disable_device(pdev);
1549
1550         return error;
1551
1552 out_release_resources_and_disable:
1553         pvscsi_shutdown_intr(adapter);
1554         pvscsi_release_resources(adapter);
1555         goto out_disable_device;
1556 }
1557
1558 static void __pvscsi_shutdown(struct pvscsi_adapter *adapter)
1559 {
1560         pvscsi_mask_intr(adapter);
1561
1562         if (adapter->workqueue)
1563                 flush_workqueue(adapter->workqueue);
1564
1565         pvscsi_shutdown_intr(adapter);
1566
1567         pvscsi_process_request_ring(adapter);
1568         pvscsi_process_completion_ring(adapter);
1569         ll_adapter_reset(adapter);
1570 }
1571
1572 static void pvscsi_shutdown(struct pci_dev *dev)
1573 {
1574         struct Scsi_Host *host = pci_get_drvdata(dev);
1575         struct pvscsi_adapter *adapter = shost_priv(host);
1576
1577         __pvscsi_shutdown(adapter);
1578 }
1579
1580 static void pvscsi_remove(struct pci_dev *pdev)
1581 {
1582         struct Scsi_Host *host = pci_get_drvdata(pdev);
1583         struct pvscsi_adapter *adapter = shost_priv(host);
1584
1585         scsi_remove_host(host);
1586
1587         __pvscsi_shutdown(adapter);
1588         pvscsi_release_resources(adapter);
1589
1590         scsi_host_put(host);
1591
1592         pci_disable_device(pdev);
1593 }
1594
1595 static struct pci_driver pvscsi_pci_driver = {
1596         .name           = "vmw_pvscsi",
1597         .id_table       = pvscsi_pci_tbl,
1598         .probe          = pvscsi_probe,
1599         .remove         = pvscsi_remove,
1600         .shutdown       = pvscsi_shutdown,
1601 };
1602
1603 static int __init pvscsi_init(void)
1604 {
1605         pr_info("%s - version %s\n",
1606                 PVSCSI_LINUX_DRIVER_DESC, PVSCSI_DRIVER_VERSION_STRING);
1607         return pci_register_driver(&pvscsi_pci_driver);
1608 }
1609
1610 static void __exit pvscsi_exit(void)
1611 {
1612         pci_unregister_driver(&pvscsi_pci_driver);
1613 }
1614
1615 module_init(pvscsi_init);
1616 module_exit(pvscsi_exit);