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_DEBUG, 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_DEBUG, 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 /*
 406  * The device incorrectly doesn't clear the first byte of the sense
 407  * buffer in some cases. We have to do it ourselves.
 408  * Otherwise we run into trouble when SWIOTLB is forced.
 409  */
 410 static void pvscsi_patch_sense(struct scsi_cmnd *cmd)
 411 {
 412         if (cmd->sense_buffer)
 413                 cmd->sense_buffer[0] = 0;
 414 }
 415
 416 static void pvscsi_unmap_buffers(const struct pvscsi_adapter *adapter,
 417                                  struct pvscsi_ctx *ctx)
 418 {
 419         struct scsi_cmnd *cmd;
 420         unsigned bufflen;
 421
 422         cmd = ctx->cmd;
 423         bufflen = scsi_bufflen(cmd);
 424
 425         if (bufflen != 0) {
 426                 unsigned count = scsi_sg_count(cmd);
 427
 428                 if (count != 0) {
 429                         scsi_dma_unmap(cmd);
 430                         if (ctx->sglPA) {
 431                                 dma_unmap_single(&adapter->dev->dev, ctx->sglPA,
 432                                                  SGL_SIZE, DMA_TO_DEVICE);
 433                                 ctx->sglPA = 0;
 434                         }
 435                 } else
 436                         dma_unmap_single(&adapter->dev->dev, ctx->dataPA,
 437                                          bufflen, cmd->sc_data_direction);
 438         }
 439         if (cmd->sense_buffer)
 440                 dma_unmap_single(&adapter->dev->dev, ctx->sensePA,
 441                                  SCSI_SENSE_BUFFERSIZE, DMA_FROM_DEVICE);
 442 }
 443
 444 static int pvscsi_allocate_rings(struct pvscsi_adapter *adapter)
 445 {
 446         adapter->rings_state = dma_alloc_coherent(&adapter->dev->dev, PAGE_SIZE,
 447                         &adapter->ringStatePA, GFP_KERNEL);
 448         if (!adapter->rings_state)
 449                 return -ENOMEM;
 450
 451         adapter->req_pages = min(PVSCSI_MAX_NUM_PAGES_REQ_RING,
 452                                  pvscsi_ring_pages);
 453         adapter->req_depth = adapter->req_pages
 454                                         * PVSCSI_MAX_NUM_REQ_ENTRIES_PER_PAGE;
 455         adapter->req_ring = dma_alloc_coherent(&adapter->dev->dev,
 456                         adapter->req_pages * PAGE_SIZE, &adapter->reqRingPA,
 457                         GFP_KERNEL);
 458         if (!adapter->req_ring)
 459                 return -ENOMEM;
 460
 461         adapter->cmp_pages = min(PVSCSI_MAX_NUM_PAGES_CMP_RING,
 462                                  pvscsi_ring_pages);
 463         adapter->cmp_ring = dma_alloc_coherent(&adapter->dev->dev,
 464                         adapter->cmp_pages * PAGE_SIZE, &adapter->cmpRingPA,
 465                         GFP_KERNEL);
 466         if (!adapter->cmp_ring)
 467                 return -ENOMEM;
 468
 469         BUG_ON(!IS_ALIGNED(adapter->ringStatePA, PAGE_SIZE));
 470         BUG_ON(!IS_ALIGNED(adapter->reqRingPA, PAGE_SIZE));
 471         BUG_ON(!IS_ALIGNED(adapter->cmpRingPA, PAGE_SIZE));
 472
 473         if (!adapter->use_msg)
 474                 return 0;
 475
 476         adapter->msg_pages = min(PVSCSI_MAX_NUM_PAGES_MSG_RING,
 477                                  pvscsi_msg_ring_pages);
 478         adapter->msg_ring = dma_alloc_coherent(&adapter->dev->dev,
 479                         adapter->msg_pages * PAGE_SIZE, &adapter->msgRingPA,
 480                         GFP_KERNEL);
 481         if (!adapter->msg_ring)
 482                 return -ENOMEM;
 483         BUG_ON(!IS_ALIGNED(adapter->msgRingPA, PAGE_SIZE));
 484
 485         return 0;
 486 }
 487
 488 static void pvscsi_setup_all_rings(const struct pvscsi_adapter *adapter)
 489 {
 490         struct PVSCSICmdDescSetupRings cmd = { 0 };
 491         dma_addr_t base;
 492         unsigned i;
 493
 494         cmd.ringsStatePPN   = adapter->ringStatePA >> PAGE_SHIFT;
 495         cmd.reqRingNumPages = adapter->req_pages;
 496         cmd.cmpRingNumPages = adapter->cmp_pages;
 497
 498         base = adapter->reqRingPA;
 499         for (i = 0; i < adapter->req_pages; i++) {
 500                 cmd.reqRingPPNs[i] = base >> PAGE_SHIFT;
 501                 base += PAGE_SIZE;
 502         }
 503
 504         base = adapter->cmpRingPA;
 505         for (i = 0; i < adapter->cmp_pages; i++) {
 506                 cmd.cmpRingPPNs[i] = base >> PAGE_SHIFT;
 507                 base += PAGE_SIZE;
 508         }
 509
 510         memset(adapter->rings_state, 0, PAGE_SIZE);
 511         memset(adapter->req_ring, 0, adapter->req_pages * PAGE_SIZE);
 512         memset(adapter->cmp_ring, 0, adapter->cmp_pages * PAGE_SIZE);
 513
 514         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_SETUP_RINGS,
 515                               &cmd, sizeof(cmd));
 516
 517         if (adapter->use_msg) {
 518                 struct PVSCSICmdDescSetupMsgRing cmd_msg = { 0 };
 519
 520                 cmd_msg.numPages = adapter->msg_pages;
 521
 522                 base = adapter->msgRingPA;
 523                 for (i = 0; i < adapter->msg_pages; i++) {
 524                         cmd_msg.ringPPNs[i] = base >> PAGE_SHIFT;
 525                         base += PAGE_SIZE;
 526                 }
 527                 memset(adapter->msg_ring, 0, adapter->msg_pages * PAGE_SIZE);
 528
 529                 pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_SETUP_MSG_RING,
 530                                       &cmd_msg, sizeof(cmd_msg));
 531         }
 532 }
 533
 534 static int pvscsi_change_queue_depth(struct scsi_device *sdev, int qdepth)
 535 {
 536         if (!sdev->tagged_supported)
 537                 qdepth = 1;
 538         return scsi_change_queue_depth(sdev, qdepth);
 539 }
 540
 541 /*
 542  * Pull a completion descriptor off and pass the completion back
 543  * to the SCSI mid layer.
 544  */
 545 static void pvscsi_complete_request(struct pvscsi_adapter *adapter,
 546                                     const struct PVSCSIRingCmpDesc *e)
 547 {
 548         struct pvscsi_ctx *ctx;
 549         struct scsi_cmnd *cmd;
 550         struct completion *abort_cmp;
 551         u32 btstat = e->hostStatus;
 552         u32 sdstat = e->scsiStatus;
 553
 554         ctx = pvscsi_get_context(adapter, e->context);
 555         cmd = ctx->cmd;
 556         abort_cmp = ctx->abort_cmp;
 557         pvscsi_unmap_buffers(adapter, ctx);
 558         if (sdstat != SAM_STAT_CHECK_CONDITION)
 559                 pvscsi_patch_sense(cmd);
 560         pvscsi_release_context(adapter, ctx);
 561         if (abort_cmp) {
 562                 /*
 563                  * The command was requested to be aborted. Just signal that
 564                  * the request completed and swallow the actual cmd completion
 565                  * here. The abort handler will post a completion for this
 566                  * command indicating that it got successfully aborted.
 567                  */
 568                 complete(abort_cmp);
 569                 return;
 570         }
 571
 572         cmd->result = 0;
 573         if (sdstat != SAM_STAT_GOOD &&
 574             (btstat == BTSTAT_SUCCESS ||
 575              btstat == BTSTAT_LINKED_COMMAND_COMPLETED ||
 576              btstat == BTSTAT_LINKED_COMMAND_COMPLETED_WITH_FLAG)) {
 577                 if (sdstat == SAM_STAT_COMMAND_TERMINATED) {
 578                         cmd->result = (DID_RESET << 16);
 579                 } else {
 580                         cmd->result = (DID_OK << 16) | sdstat;
 581                         if (sdstat == SAM_STAT_CHECK_CONDITION &&
 582                             cmd->sense_buffer)
 583                                 cmd->result |= (DRIVER_SENSE << 24);
 584                 }
 585         } else
 586                 switch (btstat) {
 587                 case BTSTAT_SUCCESS:
 588                 case BTSTAT_LINKED_COMMAND_COMPLETED:
 589                 case BTSTAT_LINKED_COMMAND_COMPLETED_WITH_FLAG:
 590                         /* If everything went fine, let's move on..  */
 591                         cmd->result = (DID_OK << 16);
 592                         break;
 593
 594                 case BTSTAT_DATARUN:
 595                 case BTSTAT_DATA_UNDERRUN:
 596                         /* Report residual data in underruns */
 597                         scsi_set_resid(cmd, scsi_bufflen(cmd) - e->dataLen);
 598                         cmd->result = (DID_ERROR << 16);
 599                         break;
 600
 601                 case BTSTAT_SELTIMEO:
 602                         /* Our emulation returns this for non-connected devs */
 603                         cmd->result = (DID_BAD_TARGET << 16);
 604                         break;
 605
 606                 case BTSTAT_LUNMISMATCH:
 607                 case BTSTAT_TAGREJECT:
 608                 case BTSTAT_BADMSG:
 609                         cmd->result = (DRIVER_INVALID << 24);
 610                         fallthrough;
 611
 612                 case BTSTAT_HAHARDWARE:
 613                 case BTSTAT_INVPHASE:
 614                 case BTSTAT_HATIMEOUT:
 615                 case BTSTAT_NORESPONSE:
 616                 case BTSTAT_DISCONNECT:
 617                 case BTSTAT_HASOFTWARE:
 618                 case BTSTAT_BUSFREE:
 619                 case BTSTAT_SENSFAILED:
 620                         cmd->result |= (DID_ERROR << 16);
 621                         break;
 622
 623                 case BTSTAT_SENTRST:
 624                 case BTSTAT_RECVRST:
 625                 case BTSTAT_BUSRESET:
 626                         cmd->result = (DID_RESET << 16);
 627                         break;
 628
 629                 case BTSTAT_ABORTQUEUE:
 630                         cmd->result = (DID_BUS_BUSY << 16);
 631                         break;
 632
 633                 case BTSTAT_SCSIPARITY:
 634                         cmd->result = (DID_PARITY << 16);
 635                         break;
 636
 637                 default:
 638                         cmd->result = (DID_ERROR << 16);
 639                         scmd_printk(KERN_DEBUG, cmd,
 640                                     "Unknown completion status: 0x%x\n",
 641                                     btstat);
 642         }
 643
 644         dev_dbg(&cmd->device->sdev_gendev,
 645                 "cmd=%p %x ctx=%p result=0x%x status=0x%x,%x\n",
 646                 cmd, cmd->cmnd[0], ctx, cmd->result, btstat, sdstat);
 647
 648         cmd->scsi_done(cmd);
 649 }
 650
 651 /*
 652  * barrier usage : Since the PVSCSI device is emulated, there could be cases
 653  * where we may want to serialize some accesses between the driver and the
 654  * emulation layer. We use compiler barriers instead of the more expensive
 655  * memory barriers because PVSCSI is only supported on X86 which has strong
 656  * memory access ordering.
 657  */
 658 static void pvscsi_process_completion_ring(struct pvscsi_adapter *adapter)
 659 {
 660         struct PVSCSIRingsState *s = adapter->rings_state;
 661         struct PVSCSIRingCmpDesc *ring = adapter->cmp_ring;
 662         u32 cmp_entries = s->cmpNumEntriesLog2;
 663
 664         while (s->cmpConsIdx != s->cmpProdIdx) {
 665                 struct PVSCSIRingCmpDesc *e = ring + (s->cmpConsIdx &
 666                                                       MASK(cmp_entries));
 667                 /*
 668                  * This barrier() ensures that *e is not dereferenced while
 669                  * the device emulation still writes data into the slot.
 670                  * Since the device emulation advances s->cmpProdIdx only after
 671                  * updating the slot we want to check it first.
 672                  */
 673                 barrier();
 674                 pvscsi_complete_request(adapter, e);
 675                 /*
 676                  * This barrier() ensures that compiler doesn't reorder write
 677                  * to s->cmpConsIdx before the read of (*e) inside
 678                  * pvscsi_complete_request. Otherwise, device emulation may
 679                  * overwrite *e before we had a chance to read it.
 680                  */
 681                 barrier();
 682                 s->cmpConsIdx++;
 683         }
 684 }
 685
 686 /*
 687  * Translate a Linux SCSI request into a request ring entry.
 688  */
 689 static int pvscsi_queue_ring(struct pvscsi_adapter *adapter,
 690                              struct pvscsi_ctx *ctx, struct scsi_cmnd *cmd)
 691 {
 692         struct PVSCSIRingsState *s;
 693         struct PVSCSIRingReqDesc *e;
 694         struct scsi_device *sdev;
 695         u32 req_entries;
 696
 697         s = adapter->rings_state;
 698         sdev = cmd->device;
 699         req_entries = s->reqNumEntriesLog2;
 700
 701         /*
 702          * If this condition holds, we might have room on the request ring, but
 703          * we might not have room on the completion ring for the response.
 704          * However, we have already ruled out this possibility - we would not
 705          * have successfully allocated a context if it were true, since we only
 706          * have one context per request entry.  Check for it anyway, since it
 707          * would be a serious bug.
 708          */
 709         if (s->reqProdIdx - s->cmpConsIdx >= 1 << req_entries) {
 710                 scmd_printk(KERN_ERR, cmd, "vmw_pvscsi: "
 711                             "ring full: reqProdIdx=%d cmpConsIdx=%d\n",
 712                             s->reqProdIdx, s->cmpConsIdx);
 713                 return -1;
 714         }
 715
 716         e = adapter->req_ring + (s->reqProdIdx & MASK(req_entries));
 717
 718         e->bus    = sdev->channel;
 719         e->target = sdev->id;
 720         memset(e->lun, 0, sizeof(e->lun));
 721         e->lun[1] = sdev->lun;
 722
 723         if (cmd->sense_buffer) {
 724                 ctx->sensePA = dma_map_single(&adapter->dev->dev,
 725                                 cmd->sense_buffer, SCSI_SENSE_BUFFERSIZE,
 726                                 DMA_FROM_DEVICE);
 727                 if (dma_mapping_error(&adapter->dev->dev, ctx->sensePA)) {
 728                         scmd_printk(KERN_DEBUG, cmd,
 729                                     "vmw_pvscsi: Failed to map sense buffer for DMA.\n");
 730                         ctx->sensePA = 0;
 731                         return -ENOMEM;
 732                 }
 733                 e->senseAddr = ctx->sensePA;
 734                 e->senseLen = SCSI_SENSE_BUFFERSIZE;
 735         } else {
 736                 e->senseLen  = 0;
 737                 e->senseAddr = 0;
 738         }
 739         e->cdbLen   = cmd->cmd_len;
 740         e->vcpuHint = smp_processor_id();
 741         memcpy(e->cdb, cmd->cmnd, e->cdbLen);
 742
 743         e->tag = SIMPLE_QUEUE_TAG;
 744
 745         if (cmd->sc_data_direction == DMA_FROM_DEVICE)
 746                 e->flags = PVSCSI_FLAG_CMD_DIR_TOHOST;
 747         else if (cmd->sc_data_direction == DMA_TO_DEVICE)
 748                 e->flags = PVSCSI_FLAG_CMD_DIR_TODEVICE;
 749         else if (cmd->sc_data_direction == DMA_NONE)
 750                 e->flags = PVSCSI_FLAG_CMD_DIR_NONE;
 751         else
 752                 e->flags = 0;
 753
 754         if (pvscsi_map_buffers(adapter, ctx, cmd, e) != 0) {
 755                 if (cmd->sense_buffer) {
 756                         dma_unmap_single(&adapter->dev->dev, ctx->sensePA,
 757                                          SCSI_SENSE_BUFFERSIZE,
 758                                          DMA_FROM_DEVICE);
 759                         ctx->sensePA = 0;
 760                 }
 761                 return -ENOMEM;
 762         }
 763
 764         e->context = pvscsi_map_context(adapter, ctx);
 765
 766         barrier();
 767
 768         s->reqProdIdx++;
 769
 770         return 0;
 771 }
 772
 773 static int pvscsi_queue_lck(struct scsi_cmnd *cmd, void (*done)(struct scsi_cmnd *))
 774 {
 775         struct Scsi_Host *host = cmd->device->host;
 776         struct pvscsi_adapter *adapter = shost_priv(host);
 777         struct pvscsi_ctx *ctx;
 778         unsigned long flags;
 779         unsigned char op;
 780
 781         spin_lock_irqsave(&adapter->hw_lock, flags);
 782
 783         ctx = pvscsi_acquire_context(adapter, cmd);
 784         if (!ctx || pvscsi_queue_ring(adapter, ctx, cmd) != 0) {
 785                 if (ctx)
 786                         pvscsi_release_context(adapter, ctx);
 787                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 788                 return SCSI_MLQUEUE_HOST_BUSY;
 789         }
 790
 791         cmd->scsi_done = done;
 792         op = cmd->cmnd[0];
 793
 794         dev_dbg(&cmd->device->sdev_gendev,
 795                 "queued cmd %p, ctx %p, op=%x\n", cmd, ctx, op);
 796
 797         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 798
 799         pvscsi_kick_io(adapter, op);
 800
 801         return 0;
 802 }
 803
 804 static DEF_SCSI_QCMD(pvscsi_queue)
 805
 806 static int pvscsi_abort(struct scsi_cmnd *cmd)
 807 {
 808         struct pvscsi_adapter *adapter = shost_priv(cmd->device->host);
 809         struct pvscsi_ctx *ctx;
 810         unsigned long flags;
 811         int result = SUCCESS;
 812         DECLARE_COMPLETION_ONSTACK(abort_cmp);
 813         int done;
 814
 815         scmd_printk(KERN_DEBUG, cmd, "task abort on host %u, %p\n",
 816                     adapter->host->host_no, cmd);
 817
 818         spin_lock_irqsave(&adapter->hw_lock, flags);
 819
 820         /*
 821          * Poll the completion ring first - we might be trying to abort
 822          * a command that is waiting to be dispatched in the completion ring.
 823          */
 824         pvscsi_process_completion_ring(adapter);
 825
 826         /*
 827          * If there is no context for the command, it either already succeeded
 828          * or else was never properly issued.  Not our problem.
 829          */
 830         ctx = pvscsi_find_context(adapter, cmd);
 831         if (!ctx) {
 832                 scmd_printk(KERN_DEBUG, cmd, "Failed to abort cmd %p\n", cmd);
 833                 goto out;
 834         }
 835
 836         /*
 837          * Mark that the command has been requested to be aborted and issue
 838          * the abort.
 839          */
 840         ctx->abort_cmp = &abort_cmp;
 841
 842         pvscsi_abort_cmd(adapter, ctx);
 843         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 844         /* Wait for 2 secs for the completion. */
 845         done = wait_for_completion_timeout(&abort_cmp, msecs_to_jiffies(2000));
 846         spin_lock_irqsave(&adapter->hw_lock, flags);
 847
 848         if (!done) {
 849                 /*
 850                  * Failed to abort the command, unmark the fact that it
 851                  * was requested to be aborted.
 852                  */
 853                 ctx->abort_cmp = NULL;
 854                 result = FAILED;
 855                 scmd_printk(KERN_DEBUG, cmd,
 856                             "Failed to get completion for aborted cmd %p\n",
 857                             cmd);
 858                 goto out;
 859         }
 860
 861         /*
 862          * Successfully aborted the command.
 863          */
 864         cmd->result = (DID_ABORT << 16);
 865         cmd->scsi_done(cmd);
 866
 867 out:
 868         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 869         return result;
 870 }
 871
 872 /*
 873  * Abort all outstanding requests.  This is only safe to use if the completion
 874  * ring will never be walked again or the device has been reset, because it
 875  * destroys the 1-1 mapping between context field passed to emulation and our
 876  * request structure.
 877  */
 878 static void pvscsi_reset_all(struct pvscsi_adapter *adapter)
 879 {
 880         unsigned i;
 881
 882         for (i = 0; i < adapter->req_depth; i++) {
 883                 struct pvscsi_ctx *ctx = &adapter->cmd_map[i];
 884                 struct scsi_cmnd *cmd = ctx->cmd;
 885                 if (cmd) {
 886                         scmd_printk(KERN_ERR, cmd,
 887                                     "Forced reset on cmd %p\n", cmd);
 888                         pvscsi_unmap_buffers(adapter, ctx);
 889                         pvscsi_patch_sense(cmd);
 890                         pvscsi_release_context(adapter, ctx);
 891                         cmd->result = (DID_RESET << 16);
 892                         cmd->scsi_done(cmd);
 893                 }
 894         }
 895 }
 896
 897 static int pvscsi_host_reset(struct scsi_cmnd *cmd)
 898 {
 899         struct Scsi_Host *host = cmd->device->host;
 900         struct pvscsi_adapter *adapter = shost_priv(host);
 901         unsigned long flags;
 902         bool use_msg;
 903
 904         scmd_printk(KERN_INFO, cmd, "SCSI Host reset\n");
 905
 906         spin_lock_irqsave(&adapter->hw_lock, flags);
 907
 908         use_msg = adapter->use_msg;
 909
 910         if (use_msg) {
 911                 adapter->use_msg = false;
 912                 spin_unlock_irqrestore(&adapter->hw_lock, flags);
 913
 914                 /*
 915                  * Now that we know that the ISR won't add more work on the
 916                  * workqueue we can safely flush any outstanding work.
 917                  */
 918                 flush_workqueue(adapter->workqueue);
 919                 spin_lock_irqsave(&adapter->hw_lock, flags);
 920         }
 921
 922         /*
 923          * We're going to tear down the entire ring structure and set it back
 924          * up, so stalling new requests until all completions are flushed and
 925          * the rings are back in place.
 926          */
 927
 928         pvscsi_process_request_ring(adapter);
 929
 930         ll_adapter_reset(adapter);
 931
 932         /*
 933          * Now process any completions.  Note we do this AFTER adapter reset,
 934          * which is strange, but stops races where completions get posted
 935          * between processing the ring and issuing the reset.  The backend will
 936          * not touch the ring memory after reset, so the immediately pre-reset
 937          * completion ring state is still valid.
 938          */
 939         pvscsi_process_completion_ring(adapter);
 940
 941         pvscsi_reset_all(adapter);
 942         adapter->use_msg = use_msg;
 943         pvscsi_setup_all_rings(adapter);
 944         pvscsi_unmask_intr(adapter);
 945
 946         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 947
 948         return SUCCESS;
 949 }
 950
 951 static int pvscsi_bus_reset(struct scsi_cmnd *cmd)
 952 {
 953         struct Scsi_Host *host = cmd->device->host;
 954         struct pvscsi_adapter *adapter = shost_priv(host);
 955         unsigned long flags;
 956
 957         scmd_printk(KERN_INFO, cmd, "SCSI Bus reset\n");
 958
 959         /*
 960          * We don't want to queue new requests for this bus after
 961          * flushing all pending requests to emulation, since new
 962          * requests could then sneak in during this bus reset phase,
 963          * so take the lock now.
 964          */
 965         spin_lock_irqsave(&adapter->hw_lock, flags);
 966
 967         pvscsi_process_request_ring(adapter);
 968         ll_bus_reset(adapter);
 969         pvscsi_process_completion_ring(adapter);
 970
 971         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 972
 973         return SUCCESS;
 974 }
 975
 976 static int pvscsi_device_reset(struct scsi_cmnd *cmd)
 977 {
 978         struct Scsi_Host *host = cmd->device->host;
 979         struct pvscsi_adapter *adapter = shost_priv(host);
 980         unsigned long flags;
 981
 982         scmd_printk(KERN_INFO, cmd, "SCSI device reset on scsi%u:%u\n",
 983                     host->host_no, cmd->device->id);
 984
 985         /*
 986          * We don't want to queue new requests for this device after flushing
 987          * all pending requests to emulation, since new requests could then
 988          * sneak in during this device reset phase, so take the lock now.
 989          */
 990         spin_lock_irqsave(&adapter->hw_lock, flags);
 991
 992         pvscsi_process_request_ring(adapter);
 993         ll_device_reset(adapter, cmd->device->id);
 994         pvscsi_process_completion_ring(adapter);
 995
 996         spin_unlock_irqrestore(&adapter->hw_lock, flags);
 997
 998         return SUCCESS;
 999 }
1000
1001 static struct scsi_host_template pvscsi_template;
1002
1003 static const char *pvscsi_info(struct Scsi_Host *host)
1004 {
1005         struct pvscsi_adapter *adapter = shost_priv(host);
1006         static char buf[256];
1007
1008         sprintf(buf, "VMware PVSCSI storage adapter rev %d, req/cmp/msg rings: "
1009                 "%u/%u/%u pages, cmd_per_lun=%u", adapter->rev,
1010                 adapter->req_pages, adapter->cmp_pages, adapter->msg_pages,
1011                 pvscsi_template.cmd_per_lun);
1012
1013         return buf;
1014 }
1015
1016 static struct scsi_host_template pvscsi_template = {
1017         .module                         = THIS_MODULE,
1018         .name                           = "VMware PVSCSI Host Adapter",
1019         .proc_name                      = "vmw_pvscsi",
1020         .info                           = pvscsi_info,
1021         .queuecommand                   = pvscsi_queue,
1022         .this_id                        = -1,
1023         .sg_tablesize                   = PVSCSI_MAX_NUM_SG_ENTRIES_PER_SEGMENT,
1024         .dma_boundary                   = UINT_MAX,
1025         .max_sectors                    = 0xffff,
1026         .change_queue_depth             = pvscsi_change_queue_depth,
1027         .eh_abort_handler               = pvscsi_abort,
1028         .eh_device_reset_handler        = pvscsi_device_reset,
1029         .eh_bus_reset_handler           = pvscsi_bus_reset,
1030         .eh_host_reset_handler          = pvscsi_host_reset,
1031 };
1032
1033 static void pvscsi_process_msg(const struct pvscsi_adapter *adapter,
1034                                const struct PVSCSIRingMsgDesc *e)
1035 {
1036         struct PVSCSIRingsState *s = adapter->rings_state;
1037         struct Scsi_Host *host = adapter->host;
1038         struct scsi_device *sdev;
1039
1040         printk(KERN_INFO "vmw_pvscsi: msg type: 0x%x - MSG RING: %u/%u (%u) \n",
1041                e->type, s->msgProdIdx, s->msgConsIdx, s->msgNumEntriesLog2);
1042
1043         BUILD_BUG_ON(PVSCSI_MSG_LAST != 2);
1044
1045         if (e->type == PVSCSI_MSG_DEV_ADDED) {
1046                 struct PVSCSIMsgDescDevStatusChanged *desc;
1047                 desc = (struct PVSCSIMsgDescDevStatusChanged *)e;
1048
1049                 printk(KERN_INFO
1050                        "vmw_pvscsi: msg: device added at scsi%u:%u:%u\n",
1051                        desc->bus, desc->target, desc->lun[1]);
1052
1053                 if (!scsi_host_get(host))
1054                         return;
1055
1056                 sdev = scsi_device_lookup(host, desc->bus, desc->target,
1057                                           desc->lun[1]);
1058                 if (sdev) {
1059                         printk(KERN_INFO "vmw_pvscsi: device already exists\n");
1060                         scsi_device_put(sdev);
1061                 } else
1062                         scsi_add_device(adapter->host, desc->bus,
1063                                         desc->target, desc->lun[1]);
1064
1065                 scsi_host_put(host);
1066         } else if (e->type == PVSCSI_MSG_DEV_REMOVED) {
1067                 struct PVSCSIMsgDescDevStatusChanged *desc;
1068                 desc = (struct PVSCSIMsgDescDevStatusChanged *)e;
1069
1070                 printk(KERN_INFO
1071                        "vmw_pvscsi: msg: device removed at scsi%u:%u:%u\n",
1072                        desc->bus, desc->target, desc->lun[1]);
1073
1074                 if (!scsi_host_get(host))
1075                         return;
1076
1077                 sdev = scsi_device_lookup(host, desc->bus, desc->target,
1078                                           desc->lun[1]);
1079                 if (sdev) {
1080                         scsi_remove_device(sdev);
1081                         scsi_device_put(sdev);
1082                 } else
1083                         printk(KERN_INFO
1084                                "vmw_pvscsi: failed to lookup scsi%u:%u:%u\n",
1085                                desc->bus, desc->target, desc->lun[1]);
1086
1087                 scsi_host_put(host);
1088         }
1089 }
1090
1091 static int pvscsi_msg_pending(const struct pvscsi_adapter *adapter)
1092 {
1093         struct PVSCSIRingsState *s = adapter->rings_state;
1094
1095         return s->msgProdIdx != s->msgConsIdx;
1096 }
1097
1098 static void pvscsi_process_msg_ring(const struct pvscsi_adapter *adapter)
1099 {
1100         struct PVSCSIRingsState *s = adapter->rings_state;
1101         struct PVSCSIRingMsgDesc *ring = adapter->msg_ring;
1102         u32 msg_entries = s->msgNumEntriesLog2;
1103
1104         while (pvscsi_msg_pending(adapter)) {
1105                 struct PVSCSIRingMsgDesc *e = ring + (s->msgConsIdx &
1106                                                       MASK(msg_entries));
1107
1108                 barrier();
1109                 pvscsi_process_msg(adapter, e);
1110                 barrier();
1111                 s->msgConsIdx++;
1112         }
1113 }
1114
1115 static void pvscsi_msg_workqueue_handler(struct work_struct *data)
1116 {
1117         struct pvscsi_adapter *adapter;
1118
1119         adapter = container_of(data, struct pvscsi_adapter, work);
1120
1121         pvscsi_process_msg_ring(adapter);
1122 }
1123
1124 static int pvscsi_setup_msg_workqueue(struct pvscsi_adapter *adapter)
1125 {
1126         char name[32];
1127
1128         if (!pvscsi_use_msg)
1129                 return 0;
1130
1131         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
1132                          PVSCSI_CMD_SETUP_MSG_RING);
1133
1134         if (pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS) == -1)
1135                 return 0;
1136
1137         snprintf(name, sizeof(name),
1138                  "vmw_pvscsi_wq_%u", adapter->host->host_no);
1139
1140         adapter->workqueue = create_singlethread_workqueue(name);
1141         if (!adapter->workqueue) {
1142                 printk(KERN_ERR "vmw_pvscsi: failed to create work queue\n");
1143                 return 0;
1144         }
1145         INIT_WORK(&adapter->work, pvscsi_msg_workqueue_handler);
1146
1147         return 1;
1148 }
1149
1150 static bool pvscsi_setup_req_threshold(struct pvscsi_adapter *adapter,
1151                                       bool enable)
1152 {
1153         u32 val;
1154
1155         if (!pvscsi_use_req_threshold)
1156                 return false;
1157
1158         pvscsi_reg_write(adapter, PVSCSI_REG_OFFSET_COMMAND,
1159                          PVSCSI_CMD_SETUP_REQCALLTHRESHOLD);
1160         val = pvscsi_reg_read(adapter, PVSCSI_REG_OFFSET_COMMAND_STATUS);
1161         if (val == -1) {
1162                 printk(KERN_INFO "vmw_pvscsi: device does not support req_threshold\n");
1163                 return false;
1164         } else {
1165                 struct PVSCSICmdDescSetupReqCall cmd_msg = { 0 };
1166                 cmd_msg.enable = enable;
1167                 printk(KERN_INFO
1168                        "vmw_pvscsi: %sabling reqCallThreshold\n",
1169                         enable ? "en" : "dis");
1170                 pvscsi_write_cmd_desc(adapter,
1171                                       PVSCSI_CMD_SETUP_REQCALLTHRESHOLD,
1172                                       &cmd_msg, sizeof(cmd_msg));
1173                 return pvscsi_reg_read(adapter,
1174                                        PVSCSI_REG_OFFSET_COMMAND_STATUS) != 0;
1175         }
1176 }
1177
1178 static irqreturn_t pvscsi_isr(int irq, void *devp)
1179 {
1180         struct pvscsi_adapter *adapter = devp;
1181         unsigned long flags;
1182
1183         spin_lock_irqsave(&adapter->hw_lock, flags);
1184         pvscsi_process_completion_ring(adapter);
1185         if (adapter->use_msg && pvscsi_msg_pending(adapter))
1186                 queue_work(adapter->workqueue, &adapter->work);
1187         spin_unlock_irqrestore(&adapter->hw_lock, flags);
1188
1189         return IRQ_HANDLED;
1190 }
1191
1192 static irqreturn_t pvscsi_shared_isr(int irq, void *devp)
1193 {
1194         struct pvscsi_adapter *adapter = devp;
1195         u32 val = pvscsi_read_intr_status(adapter);
1196
1197         if (!(val & PVSCSI_INTR_ALL_SUPPORTED))
1198                 return IRQ_NONE;
1199         pvscsi_write_intr_status(devp, val);
1200         return pvscsi_isr(irq, devp);
1201 }
1202
1203 static void pvscsi_free_sgls(const struct pvscsi_adapter *adapter)
1204 {
1205         struct pvscsi_ctx *ctx = adapter->cmd_map;
1206         unsigned i;
1207
1208         for (i = 0; i < adapter->req_depth; ++i, ++ctx)
1209                 free_pages((unsigned long)ctx->sgl, get_order(SGL_SIZE));
1210 }
1211
1212 static void pvscsi_shutdown_intr(struct pvscsi_adapter *adapter)
1213 {
1214         free_irq(pci_irq_vector(adapter->dev, 0), adapter);
1215         pci_free_irq_vectors(adapter->dev);
1216 }
1217
1218 static void pvscsi_release_resources(struct pvscsi_adapter *adapter)
1219 {
1220         if (adapter->workqueue)
1221                 destroy_workqueue(adapter->workqueue);
1222
1223         if (adapter->mmioBase)
1224                 pci_iounmap(adapter->dev, adapter->mmioBase);
1225
1226         pci_release_regions(adapter->dev);
1227
1228         if (adapter->cmd_map) {
1229                 pvscsi_free_sgls(adapter);
1230                 kfree(adapter->cmd_map);
1231         }
1232
1233         if (adapter->rings_state)
1234                 dma_free_coherent(&adapter->dev->dev, PAGE_SIZE,
1235                                     adapter->rings_state, adapter->ringStatePA);
1236
1237         if (adapter->req_ring)
1238                 dma_free_coherent(&adapter->dev->dev,
1239                                     adapter->req_pages * PAGE_SIZE,
1240                                     adapter->req_ring, adapter->reqRingPA);
1241
1242         if (adapter->cmp_ring)
1243                 dma_free_coherent(&adapter->dev->dev,
1244                                     adapter->cmp_pages * PAGE_SIZE,
1245                                     adapter->cmp_ring, adapter->cmpRingPA);
1246
1247         if (adapter->msg_ring)
1248                 dma_free_coherent(&adapter->dev->dev,
1249                                     adapter->msg_pages * PAGE_SIZE,
1250                                     adapter->msg_ring, adapter->msgRingPA);
1251 }
1252
1253 /*
1254  * Allocate scatter gather lists.
1255  *
1256  * These are statically allocated.  Trying to be clever was not worth it.
1257  *
1258  * Dynamic allocation can fail, and we can't go deep into the memory
1259  * allocator, since we're a SCSI driver, and trying too hard to allocate
1260  * memory might generate disk I/O.  We also don't want to fail disk I/O
1261  * in that case because we can't get an allocation - the I/O could be
1262  * trying to swap out data to free memory.  Since that is pathological,
1263  * just use a statically allocated scatter list.
1264  *
1265  */
1266 static int pvscsi_allocate_sg(struct pvscsi_adapter *adapter)
1267 {
1268         struct pvscsi_ctx *ctx;
1269         int i;
1270
1271         ctx = adapter->cmd_map;
1272         BUILD_BUG_ON(sizeof(struct pvscsi_sg_list) > SGL_SIZE);
1273
1274         for (i = 0; i < adapter->req_depth; ++i, ++ctx) {
1275                 ctx->sgl = (void *)__get_free_pages(GFP_KERNEL,
1276                                                     get_order(SGL_SIZE));
1277                 ctx->sglPA = 0;
1278                 BUG_ON(!IS_ALIGNED(((unsigned long)ctx->sgl), PAGE_SIZE));
1279                 if (!ctx->sgl) {
1280                         for (; i >= 0; --i, --ctx) {
1281                                 free_pages((unsigned long)ctx->sgl,
1282                                            get_order(SGL_SIZE));
1283                                 ctx->sgl = NULL;
1284                         }
1285                         return -ENOMEM;
1286                 }
1287         }
1288
1289         return 0;
1290 }
1291
1292 /*
1293  * Query the device, fetch the config info and return the
1294  * maximum number of targets on the adapter. In case of
1295  * failure due to any reason return default i.e. 16.
1296  */
1297 static u32 pvscsi_get_max_targets(struct pvscsi_adapter *adapter)
1298 {
1299         struct PVSCSICmdDescConfigCmd cmd;
1300         struct PVSCSIConfigPageHeader *header;
1301         struct device *dev;
1302         dma_addr_t configPagePA;
1303         void *config_page;
1304         u32 numPhys = 16;
1305
1306         dev = pvscsi_dev(adapter);
1307         config_page = dma_alloc_coherent(&adapter->dev->dev, PAGE_SIZE,
1308                         &configPagePA, GFP_KERNEL);
1309         if (!config_page) {
1310                 dev_warn(dev, "vmw_pvscsi: failed to allocate memory for config page\n");
1311                 goto exit;
1312         }
1313         BUG_ON(configPagePA & ~PAGE_MASK);
1314
1315         /* Fetch config info from the device. */
1316         cmd.configPageAddress = ((u64)PVSCSI_CONFIG_CONTROLLER_ADDRESS) << 32;
1317         cmd.configPageNum = PVSCSI_CONFIG_PAGE_CONTROLLER;
1318         cmd.cmpAddr = configPagePA;
1319         cmd._pad = 0;
1320
1321         /*
1322          * Mark the completion page header with error values. If the device
1323          * completes the command successfully, it sets the status values to
1324          * indicate success.
1325          */
1326         header = config_page;
1327         memset(header, 0, sizeof *header);
1328         header->hostStatus = BTSTAT_INVPARAM;
1329         header->scsiStatus = SDSTAT_CHECK;
1330
1331         pvscsi_write_cmd_desc(adapter, PVSCSI_CMD_CONFIG, &cmd, sizeof cmd);
1332
1333         if (header->hostStatus == BTSTAT_SUCCESS &&
1334             header->scsiStatus == SDSTAT_GOOD) {
1335                 struct PVSCSIConfigPageController *config;
1336
1337                 config = config_page;
1338                 numPhys = config->numPhys;
1339         } else
1340                 dev_warn(dev, "vmw_pvscsi: PVSCSI_CMD_CONFIG failed. hostStatus = 0x%x, scsiStatus = 0x%x\n",
1341                          header->hostStatus, header->scsiStatus);
1342         dma_free_coherent(&adapter->dev->dev, PAGE_SIZE, config_page,
1343                           configPagePA);
1344 exit:
1345         return numPhys;
1346 }
1347
1348 static int pvscsi_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1349 {
1350         unsigned int irq_flag = PCI_IRQ_MSIX | PCI_IRQ_MSI | PCI_IRQ_LEGACY;
1351         struct pvscsi_adapter *adapter;
1352         struct pvscsi_adapter adapter_temp;
1353         struct Scsi_Host *host = NULL;
1354         unsigned int i;
1355         int error;
1356         u32 max_id;
1357
1358         error = -ENODEV;
1359
1360         if (pci_enable_device(pdev))
1361                 return error;
1362
1363         if (!dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64))) {
1364                 printk(KERN_INFO "vmw_pvscsi: using 64bit dma\n");
1365         } else if (!dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(32))) {
1366                 printk(KERN_INFO "vmw_pvscsi: using 32bit dma\n");
1367         } else {
1368                 printk(KERN_ERR "vmw_pvscsi: failed to set DMA mask\n");
1369                 goto out_disable_device;
1370         }
1371
1372         /*
1373          * Let's use a temp pvscsi_adapter struct until we find the number of
1374          * targets on the adapter, after that we will switch to the real
1375          * allocated struct.
1376          */
1377         adapter = &adapter_temp;
1378         memset(adapter, 0, sizeof(*adapter));
1379         adapter->dev  = pdev;
1380         adapter->rev = pdev->revision;
1381
1382         if (pci_request_regions(pdev, "vmw_pvscsi")) {
1383                 printk(KERN_ERR "vmw_pvscsi: pci memory selection failed\n");
1384                 goto out_disable_device;
1385         }
1386
1387         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
1388                 if ((pci_resource_flags(pdev, i) & PCI_BASE_ADDRESS_SPACE_IO))
1389                         continue;
1390
1391                 if (pci_resource_len(pdev, i) < PVSCSI_MEM_SPACE_SIZE)
1392                         continue;
1393
1394                 break;
1395         }
1396
1397         if (i == DEVICE_COUNT_RESOURCE) {
1398                 printk(KERN_ERR
1399                        "vmw_pvscsi: adapter has no suitable MMIO region\n");
1400                 goto out_release_resources_and_disable;
1401         }
1402
1403         adapter->mmioBase = pci_iomap(pdev, i, PVSCSI_MEM_SPACE_SIZE);
1404
1405         if (!adapter->mmioBase) {
1406                 printk(KERN_ERR
1407                        "vmw_pvscsi: can't iomap for BAR %d memsize %lu\n",
1408                        i, PVSCSI_MEM_SPACE_SIZE);
1409                 goto out_release_resources_and_disable;
1410         }
1411
1412         pci_set_master(pdev);
1413
1414         /*
1415          * Ask the device for max number of targets before deciding the
1416          * default pvscsi_ring_pages value.
1417          */
1418         max_id = pvscsi_get_max_targets(adapter);
1419         printk(KERN_INFO "vmw_pvscsi: max_id: %u\n", max_id);
1420
1421         if (pvscsi_ring_pages == 0)
1422                 /*
1423                  * Set the right default value. Up to 16 it is 8, above it is
1424                  * max.
1425                  */
1426                 pvscsi_ring_pages = (max_id > 16) ?
1427                         PVSCSI_SETUP_RINGS_MAX_NUM_PAGES :
1428                         PVSCSI_DEFAULT_NUM_PAGES_PER_RING;
1429         printk(KERN_INFO
1430                "vmw_pvscsi: setting ring_pages to %d\n",
1431                pvscsi_ring_pages);
1432
1433         pvscsi_template.can_queue =
1434                 min(PVSCSI_MAX_NUM_PAGES_REQ_RING, pvscsi_ring_pages) *
1435                 PVSCSI_MAX_NUM_REQ_ENTRIES_PER_PAGE;
1436         pvscsi_template.cmd_per_lun =
1437                 min(pvscsi_template.can_queue, pvscsi_cmd_per_lun);
1438         host = scsi_host_alloc(&pvscsi_template, sizeof(struct pvscsi_adapter));
1439         if (!host) {
1440                 printk(KERN_ERR "vmw_pvscsi: failed to allocate host\n");
1441                 goto out_release_resources_and_disable;
1442         }
1443
1444         /*
1445          * Let's use the real pvscsi_adapter struct here onwards.
1446          */
1447         adapter = shost_priv(host);
1448         memset(adapter, 0, sizeof(*adapter));
1449         adapter->dev  = pdev;
1450         adapter->host = host;
1451         /*
1452          * Copy back what we already have to the allocated adapter struct.
1453          */
1454         adapter->rev = adapter_temp.rev;
1455         adapter->mmioBase = adapter_temp.mmioBase;
1456
1457         spin_lock_init(&adapter->hw_lock);
1458         host->max_channel = 0;
1459         host->max_lun     = 1;
1460         host->max_cmd_len = 16;
1461         host->max_id      = max_id;
1462
1463         pci_set_drvdata(pdev, host);
1464
1465         ll_adapter_reset(adapter);
1466
1467         adapter->use_msg = pvscsi_setup_msg_workqueue(adapter);
1468
1469         error = pvscsi_allocate_rings(adapter);
1470         if (error) {
1471                 printk(KERN_ERR "vmw_pvscsi: unable to allocate ring memory\n");
1472                 goto out_release_resources;
1473         }
1474
1475         /*
1476          * From this point on we should reset the adapter if anything goes
1477          * wrong.
1478          */
1479         pvscsi_setup_all_rings(adapter);
1480
1481         adapter->cmd_map = kcalloc(adapter->req_depth,
1482                                    sizeof(struct pvscsi_ctx), GFP_KERNEL);
1483         if (!adapter->cmd_map) {
1484                 printk(KERN_ERR "vmw_pvscsi: failed to allocate memory.\n");
1485                 error = -ENOMEM;
1486                 goto out_reset_adapter;
1487         }
1488
1489         INIT_LIST_HEAD(&adapter->cmd_pool);
1490         for (i = 0; i < adapter->req_depth; i++) {
1491                 struct pvscsi_ctx *ctx = adapter->cmd_map + i;
1492                 list_add(&ctx->list, &adapter->cmd_pool);
1493         }
1494
1495         error = pvscsi_allocate_sg(adapter);
1496         if (error) {
1497                 printk(KERN_ERR "vmw_pvscsi: unable to allocate s/g table\n");
1498                 goto out_reset_adapter;
1499         }
1500
1501         if (pvscsi_disable_msix)
1502                 irq_flag &= ~PCI_IRQ_MSIX;
1503         if (pvscsi_disable_msi)
1504                 irq_flag &= ~PCI_IRQ_MSI;
1505
1506         error = pci_alloc_irq_vectors(adapter->dev, 1, 1, irq_flag);
1507         if (error < 0)
1508                 goto out_reset_adapter;
1509
1510         adapter->use_req_threshold = pvscsi_setup_req_threshold(adapter, true);
1511         printk(KERN_DEBUG "vmw_pvscsi: driver-based request coalescing %sabled\n",
1512                adapter->use_req_threshold ? "en" : "dis");
1513
1514         if (adapter->dev->msix_enabled || adapter->dev->msi_enabled) {
1515                 printk(KERN_INFO "vmw_pvscsi: using MSI%s\n",
1516                         adapter->dev->msix_enabled ? "-X" : "");
1517                 error = request_irq(pci_irq_vector(pdev, 0), pvscsi_isr,
1518                                 0, "vmw_pvscsi", adapter);
1519         } else {
1520                 printk(KERN_INFO "vmw_pvscsi: using INTx\n");
1521                 error = request_irq(pci_irq_vector(pdev, 0), pvscsi_shared_isr,
1522                                 IRQF_SHARED, "vmw_pvscsi", adapter);
1523         }
1524
1525         if (error) {
1526                 printk(KERN_ERR
1527                        "vmw_pvscsi: unable to request IRQ: %d\n", error);
1528                 goto out_reset_adapter;
1529         }
1530
1531         error = scsi_add_host(host, &pdev->dev);
1532         if (error) {
1533                 printk(KERN_ERR
1534                        "vmw_pvscsi: scsi_add_host failed: %d\n", error);
1535                 goto out_reset_adapter;
1536         }
1537
1538         dev_info(&pdev->dev, "VMware PVSCSI rev %d host #%u\n",
1539                  adapter->rev, host->host_no);
1540
1541         pvscsi_unmask_intr(adapter);
1542
1543         scsi_scan_host(host);
1544
1545         return 0;
1546
1547 out_reset_adapter:
1548         ll_adapter_reset(adapter);
1549 out_release_resources:
1550         pvscsi_shutdown_intr(adapter);
1551         pvscsi_release_resources(adapter);
1552         scsi_host_put(host);
1553 out_disable_device:
1554         pci_disable_device(pdev);
1555
1556         return error;
1557
1558 out_release_resources_and_disable:
1559         pvscsi_shutdown_intr(adapter);
1560         pvscsi_release_resources(adapter);
1561         goto out_disable_device;
1562 }
1563
1564 static void __pvscsi_shutdown(struct pvscsi_adapter *adapter)
1565 {
1566         pvscsi_mask_intr(adapter);
1567
1568         if (adapter->workqueue)
1569                 flush_workqueue(adapter->workqueue);
1570
1571         pvscsi_shutdown_intr(adapter);
1572
1573         pvscsi_process_request_ring(adapter);
1574         pvscsi_process_completion_ring(adapter);
1575         ll_adapter_reset(adapter);
1576 }
1577
1578 static void pvscsi_shutdown(struct pci_dev *dev)
1579 {
1580         struct Scsi_Host *host = pci_get_drvdata(dev);
1581         struct pvscsi_adapter *adapter = shost_priv(host);
1582
1583         __pvscsi_shutdown(adapter);
1584 }
1585
1586 static void pvscsi_remove(struct pci_dev *pdev)
1587 {
1588         struct Scsi_Host *host = pci_get_drvdata(pdev);
1589         struct pvscsi_adapter *adapter = shost_priv(host);
1590
1591         scsi_remove_host(host);
1592
1593         __pvscsi_shutdown(adapter);
1594         pvscsi_release_resources(adapter);
1595
1596         scsi_host_put(host);
1597
1598         pci_disable_device(pdev);
1599 }
1600
1601 static struct pci_driver pvscsi_pci_driver = {
1602         .name           = "vmw_pvscsi",
1603         .id_table       = pvscsi_pci_tbl,
1604         .probe          = pvscsi_probe,
1605         .remove         = pvscsi_remove,
1606         .shutdown       = pvscsi_shutdown,
1607 };
1608
1609 static int __init pvscsi_init(void)
1610 {
1611         pr_info("%s - version %s\n",
1612                 PVSCSI_LINUX_DRIVER_DESC, PVSCSI_DRIVER_VERSION_STRING);
1613         return pci_register_driver(&pvscsi_pci_driver);
1614 }
1615
1616 static void __exit pvscsi_exit(void)
1617 {
1618         pci_unregister_driver(&pvscsi_pci_driver);
1619 }
1620
1621 module_init(pvscsi_init);
1622 module_exit(pvscsi_exit);