drivers/block/nvme-scsi.c

   1 /*
   2  * NVM Express device driver
   3  * Copyright (c) 2011-2014, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14
  15 /*
  16  * Refer to the SCSI-NVMe Translation spec for details on how
  17  * each command is translated.
  18  */
  19
  20 #include <linux/nvme.h>
  21 #include <linux/bio.h>
  22 #include <linux/bitops.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/compat.h>
  25 #include <linux/delay.h>
  26 #include <linux/errno.h>
  27 #include <linux/fs.h>
  28 #include <linux/genhd.h>
  29 #include <linux/idr.h>
  30 #include <linux/init.h>
  31 #include <linux/interrupt.h>
  32 #include <linux/io.h>
  33 #include <linux/kdev_t.h>
  34 #include <linux/kthread.h>
  35 #include <linux/kernel.h>
  36 #include <linux/mm.h>
  37 #include <linux/module.h>
  38 #include <linux/moduleparam.h>
  39 #include <linux/pci.h>
  40 #include <linux/poison.h>
  41 #include <linux/sched.h>
  42 #include <linux/slab.h>
  43 #include <linux/types.h>
  44 #include <asm/unaligned.h>
  45 #include <scsi/sg.h>
  46 #include <scsi/scsi.h>
  47
  48
  49 static int sg_version_num = 30534;      /* 2 digits for each component */
  50
  51 /* VPD Page Codes */
  52 #define VPD_SUPPORTED_PAGES                             0x00
  53 #define VPD_SERIAL_NUMBER                               0x80
  54 #define VPD_DEVICE_IDENTIFIERS                          0x83
  55 #define VPD_EXTENDED_INQUIRY                            0x86
  56 #define VPD_BLOCK_LIMITS                                0xB0
  57 #define VPD_BLOCK_DEV_CHARACTERISTICS                   0xB1
  58
  59 /* format unit paramter list offsets */
  60 #define FORMAT_UNIT_SHORT_PARM_LIST_LEN                 4
  61 #define FORMAT_UNIT_LONG_PARM_LIST_LEN                  8
  62 #define FORMAT_UNIT_PROT_INT_OFFSET                     3
  63 #define FORMAT_UNIT_PROT_FIELD_USAGE_OFFSET             0
  64 #define FORMAT_UNIT_PROT_FIELD_USAGE_MASK               0x07
  65
  66 /* Misc. defines */
  67 #define FIXED_SENSE_DATA                                0x70
  68 #define DESC_FORMAT_SENSE_DATA                          0x72
  69 #define FIXED_SENSE_DATA_ADD_LENGTH                     10
  70 #define LUN_ENTRY_SIZE                                  8
  71 #define LUN_DATA_HEADER_SIZE                            8
  72 #define ALL_LUNS_RETURNED                               0x02
  73 #define ALL_WELL_KNOWN_LUNS_RETURNED                    0x01
  74 #define RESTRICTED_LUNS_RETURNED                        0x00
  75 #define NVME_POWER_STATE_START_VALID                    0x00
  76 #define NVME_POWER_STATE_ACTIVE                         0x01
  77 #define NVME_POWER_STATE_IDLE                           0x02
  78 #define NVME_POWER_STATE_STANDBY                        0x03
  79 #define NVME_POWER_STATE_LU_CONTROL                     0x07
  80 #define POWER_STATE_0                                   0
  81 #define POWER_STATE_1                                   1
  82 #define POWER_STATE_2                                   2
  83 #define POWER_STATE_3                                   3
  84 #define DOWNLOAD_SAVE_ACTIVATE                          0x05
  85 #define DOWNLOAD_SAVE_DEFER_ACTIVATE                    0x0E
  86 #define ACTIVATE_DEFERRED_MICROCODE                     0x0F
  87 #define FORMAT_UNIT_IMMED_MASK                          0x2
  88 #define FORMAT_UNIT_IMMED_OFFSET                        1
  89 #define KELVIN_TEMP_FACTOR                              273
  90 #define FIXED_FMT_SENSE_DATA_SIZE                       18
  91 #define DESC_FMT_SENSE_DATA_SIZE                        8
  92
  93 /* SCSI/NVMe defines and bit masks */
  94 #define INQ_STANDARD_INQUIRY_PAGE                       0x00
  95 #define INQ_SUPPORTED_VPD_PAGES_PAGE                    0x00
  96 #define INQ_UNIT_SERIAL_NUMBER_PAGE                     0x80
  97 #define INQ_DEVICE_IDENTIFICATION_PAGE                  0x83
  98 #define INQ_EXTENDED_INQUIRY_DATA_PAGE                  0x86
  99 #define INQ_BDEV_LIMITS_PAGE                            0xB0
 100 #define INQ_BDEV_CHARACTERISTICS_PAGE                   0xB1
 101 #define INQ_SERIAL_NUMBER_LENGTH                        0x14
 102 #define INQ_NUM_SUPPORTED_VPD_PAGES                     6
 103 #define VERSION_SPC_4                                   0x06
 104 #define ACA_UNSUPPORTED                                 0
 105 #define STANDARD_INQUIRY_LENGTH                         36
 106 #define ADDITIONAL_STD_INQ_LENGTH                       31
 107 #define EXTENDED_INQUIRY_DATA_PAGE_LENGTH               0x3C
 108 #define RESERVED_FIELD                                  0
 109
 110 /* Mode Sense/Select defines */
 111 #define MODE_PAGE_INFO_EXCEP                            0x1C
 112 #define MODE_PAGE_CACHING                               0x08
 113 #define MODE_PAGE_CONTROL                               0x0A
 114 #define MODE_PAGE_POWER_CONDITION                       0x1A
 115 #define MODE_PAGE_RETURN_ALL                            0x3F
 116 #define MODE_PAGE_BLK_DES_LEN                           0x08
 117 #define MODE_PAGE_LLBAA_BLK_DES_LEN                     0x10
 118 #define MODE_PAGE_CACHING_LEN                           0x14
 119 #define MODE_PAGE_CONTROL_LEN                           0x0C
 120 #define MODE_PAGE_POW_CND_LEN                           0x28
 121 #define MODE_PAGE_INF_EXC_LEN                           0x0C
 122 #define MODE_PAGE_ALL_LEN                               0x54
 123 #define MODE_SENSE6_MPH_SIZE                            4
 124 #define MODE_SENSE_PAGE_CONTROL_MASK                    0xC0
 125 #define MODE_SENSE_PAGE_CODE_OFFSET                     2
 126 #define MODE_SENSE_PAGE_CODE_MASK                       0x3F
 127 #define MODE_SENSE_LLBAA_MASK                           0x10
 128 #define MODE_SENSE_LLBAA_SHIFT                          4
 129 #define MODE_SENSE_DBD_MASK                             8
 130 #define MODE_SENSE_DBD_SHIFT                            3
 131 #define MODE_SENSE10_MPH_SIZE                           8
 132 #define MODE_SELECT_CDB_PAGE_FORMAT_MASK                0x10
 133 #define MODE_SELECT_CDB_SAVE_PAGES_MASK                 0x1
 134 #define MODE_SELECT_6_BD_OFFSET                         3
 135 #define MODE_SELECT_10_BD_OFFSET                        6
 136 #define MODE_SELECT_10_LLBAA_OFFSET                     4
 137 #define MODE_SELECT_10_LLBAA_MASK                       1
 138 #define MODE_SELECT_6_MPH_SIZE                          4
 139 #define MODE_SELECT_10_MPH_SIZE                         8
 140 #define CACHING_MODE_PAGE_WCE_MASK                      0x04
 141 #define MODE_SENSE_BLK_DESC_ENABLED                     0
 142 #define MODE_SENSE_BLK_DESC_COUNT                       1
 143 #define MODE_SELECT_PAGE_CODE_MASK                      0x3F
 144 #define SHORT_DESC_BLOCK                                8
 145 #define LONG_DESC_BLOCK                                 16
 146 #define MODE_PAGE_POW_CND_LEN_FIELD                     0x26
 147 #define MODE_PAGE_INF_EXC_LEN_FIELD                     0x0A
 148 #define MODE_PAGE_CACHING_LEN_FIELD                     0x12
 149 #define MODE_PAGE_CONTROL_LEN_FIELD                     0x0A
 150 #define MODE_SENSE_PC_CURRENT_VALUES                    0
 151
 152 /* Log Sense defines */
 153 #define LOG_PAGE_SUPPORTED_LOG_PAGES_PAGE               0x00
 154 #define LOG_PAGE_SUPPORTED_LOG_PAGES_LENGTH             0x07
 155 #define LOG_PAGE_INFORMATIONAL_EXCEPTIONS_PAGE          0x2F
 156 #define LOG_PAGE_TEMPERATURE_PAGE                       0x0D
 157 #define LOG_SENSE_CDB_SP_NOT_ENABLED                    0
 158 #define LOG_SENSE_CDB_PC_MASK                           0xC0
 159 #define LOG_SENSE_CDB_PC_SHIFT                          6
 160 #define LOG_SENSE_CDB_PC_CUMULATIVE_VALUES              1
 161 #define LOG_SENSE_CDB_PAGE_CODE_MASK                    0x3F
 162 #define REMAINING_INFO_EXCP_PAGE_LENGTH                 0x8
 163 #define LOG_INFO_EXCP_PAGE_LENGTH                       0xC
 164 #define REMAINING_TEMP_PAGE_LENGTH                      0xC
 165 #define LOG_TEMP_PAGE_LENGTH                            0x10
 166 #define LOG_TEMP_UNKNOWN                                0xFF
 167 #define SUPPORTED_LOG_PAGES_PAGE_LENGTH                 0x3
 168
 169 /* Read Capacity defines */
 170 #define READ_CAP_10_RESP_SIZE                           8
 171 #define READ_CAP_16_RESP_SIZE                           32
 172
 173 /* NVMe Namespace and Command Defines */
 174 #define BYTES_TO_DWORDS                                 4
 175 #define NVME_MAX_FIRMWARE_SLOT                          7
 176
 177 /* Report LUNs defines */
 178 #define REPORT_LUNS_FIRST_LUN_OFFSET                    8
 179
 180 /* SCSI ADDITIONAL SENSE Codes */
 181
 182 #define SCSI_ASC_NO_SENSE                               0x00
 183 #define SCSI_ASC_PERIPHERAL_DEV_WRITE_FAULT             0x03
 184 #define SCSI_ASC_LUN_NOT_READY                          0x04
 185 #define SCSI_ASC_WARNING                                0x0B
 186 #define SCSI_ASC_LOG_BLOCK_GUARD_CHECK_FAILED           0x10
 187 #define SCSI_ASC_LOG_BLOCK_APPTAG_CHECK_FAILED          0x10
 188 #define SCSI_ASC_LOG_BLOCK_REFTAG_CHECK_FAILED          0x10
 189 #define SCSI_ASC_UNRECOVERED_READ_ERROR                 0x11
 190 #define SCSI_ASC_MISCOMPARE_DURING_VERIFY               0x1D
 191 #define SCSI_ASC_ACCESS_DENIED_INVALID_LUN_ID           0x20
 192 #define SCSI_ASC_ILLEGAL_COMMAND                        0x20
 193 #define SCSI_ASC_ILLEGAL_BLOCK                          0x21
 194 #define SCSI_ASC_INVALID_CDB                            0x24
 195 #define SCSI_ASC_INVALID_LUN                            0x25
 196 #define SCSI_ASC_INVALID_PARAMETER                      0x26
 197 #define SCSI_ASC_FORMAT_COMMAND_FAILED                  0x31
 198 #define SCSI_ASC_INTERNAL_TARGET_FAILURE                0x44
 199
 200 /* SCSI ADDITIONAL SENSE Code Qualifiers */
 201
 202 #define SCSI_ASCQ_CAUSE_NOT_REPORTABLE                  0x00
 203 #define SCSI_ASCQ_FORMAT_COMMAND_FAILED                 0x01
 204 #define SCSI_ASCQ_LOG_BLOCK_GUARD_CHECK_FAILED          0x01
 205 #define SCSI_ASCQ_LOG_BLOCK_APPTAG_CHECK_FAILED         0x02
 206 #define SCSI_ASCQ_LOG_BLOCK_REFTAG_CHECK_FAILED         0x03
 207 #define SCSI_ASCQ_FORMAT_IN_PROGRESS                    0x04
 208 #define SCSI_ASCQ_POWER_LOSS_EXPECTED                   0x08
 209 #define SCSI_ASCQ_INVALID_LUN_ID                        0x09
 210
 211 /* copied from drivers/usb/gadget/function/storage_common.h */
 212 static inline u32 get_unaligned_be24(u8 *buf)
 213 {
 214         return 0xffffff & (u32) get_unaligned_be32(buf - 1);
 215 }
 216
 217 /* Struct to gather data that needs to be extracted from a SCSI CDB.
 218    Not conforming to any particular CDB variant, but compatible with all. */
 219
 220 struct nvme_trans_io_cdb {
 221         u8 fua;
 222         u8 prot_info;
 223         u64 lba;
 224         u32 xfer_len;
 225 };
 226
 227
 228 /* Internal Helper Functions */
 229
 230
 231 /* Copy data to userspace memory */
 232
 233 static int nvme_trans_copy_to_user(struct sg_io_hdr *hdr, void *from,
 234                                                                 unsigned long n)
 235 {
 236         int i;
 237         void *index = from;
 238         size_t remaining = n;
 239         size_t xfer_len;
 240
 241         if (hdr->iovec_count > 0) {
 242                 struct sg_iovec sgl;
 243
 244                 for (i = 0; i < hdr->iovec_count; i++) {
 245                         if (copy_from_user(&sgl, hdr->dxferp +
 246                                                 i * sizeof(struct sg_iovec),
 247                                                 sizeof(struct sg_iovec)))
 248                                 return -EFAULT;
 249                         xfer_len = min(remaining, sgl.iov_len);
 250                         if (copy_to_user(sgl.iov_base, index, xfer_len))
 251                                 return -EFAULT;
 252
 253                         index += xfer_len;
 254                         remaining -= xfer_len;
 255                         if (remaining == 0)
 256                                 break;
 257                 }
 258                 return 0;
 259         }
 260
 261         if (copy_to_user(hdr->dxferp, from, n))
 262                 return -EFAULT;
 263         return 0;
 264 }
 265
 266 /* Copy data from userspace memory */
 267
 268 static int nvme_trans_copy_from_user(struct sg_io_hdr *hdr, void *to,
 269                                                                 unsigned long n)
 270 {
 271         int i;
 272         void *index = to;
 273         size_t remaining = n;
 274         size_t xfer_len;
 275
 276         if (hdr->iovec_count > 0) {
 277                 struct sg_iovec sgl;
 278
 279                 for (i = 0; i < hdr->iovec_count; i++) {
 280                         if (copy_from_user(&sgl, hdr->dxferp +
 281                                                 i * sizeof(struct sg_iovec),
 282                                                 sizeof(struct sg_iovec)))
 283                                 return -EFAULT;
 284                         xfer_len = min(remaining, sgl.iov_len);
 285                         if (copy_from_user(index, sgl.iov_base, xfer_len))
 286                                 return -EFAULT;
 287                         index += xfer_len;
 288                         remaining -= xfer_len;
 289                         if (remaining == 0)
 290                                 break;
 291                 }
 292                 return 0;
 293         }
 294
 295         if (copy_from_user(to, hdr->dxferp, n))
 296                 return -EFAULT;
 297         return 0;
 298 }
 299
 300 /* Status/Sense Buffer Writeback */
 301
 302 static int nvme_trans_completion(struct sg_io_hdr *hdr, u8 status, u8 sense_key,
 303                                  u8 asc, u8 ascq)
 304 {
 305         u8 xfer_len;
 306         u8 resp[DESC_FMT_SENSE_DATA_SIZE];
 307
 308         if (scsi_status_is_good(status)) {
 309                 hdr->status = SAM_STAT_GOOD;
 310                 hdr->masked_status = GOOD;
 311                 hdr->host_status = DID_OK;
 312                 hdr->driver_status = DRIVER_OK;
 313                 hdr->sb_len_wr = 0;
 314         } else {
 315                 hdr->status = status;
 316                 hdr->masked_status = status >> 1;
 317                 hdr->host_status = DID_OK;
 318                 hdr->driver_status = DRIVER_OK;
 319
 320                 memset(resp, 0, DESC_FMT_SENSE_DATA_SIZE);
 321                 resp[0] = DESC_FORMAT_SENSE_DATA;
 322                 resp[1] = sense_key;
 323                 resp[2] = asc;
 324                 resp[3] = ascq;
 325
 326                 xfer_len = min_t(u8, hdr->mx_sb_len, DESC_FMT_SENSE_DATA_SIZE);
 327                 hdr->sb_len_wr = xfer_len;
 328                 if (copy_to_user(hdr->sbp, resp, xfer_len) > 0)
 329                         return -EFAULT;
 330         }
 331
 332         return 0;
 333 }
 334
 335 /*
 336  * Take a status code from a lowlevel routine, and if it was a positive NVMe
 337  * error code update the sense data based on it.  In either case the passed
 338  * in value is returned again, unless an -EFAULT from copy_to_user overrides
 339  * it.
 340  */
 341 static int nvme_trans_status_code(struct sg_io_hdr *hdr, int nvme_sc)
 342 {
 343         u8 status, sense_key, asc, ascq;
 344         int res;
 345
 346         /* For non-nvme (Linux) errors, simply return the error code */
 347         if (nvme_sc < 0)
 348                 return nvme_sc;
 349
 350         /* Mask DNR, More, and reserved fields */
 351         switch (nvme_sc & 0x7FF) {
 352         /* Generic Command Status */
 353         case NVME_SC_SUCCESS:
 354                 status = SAM_STAT_GOOD;
 355                 sense_key = NO_SENSE;
 356                 asc = SCSI_ASC_NO_SENSE;
 357                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 358                 break;
 359         case NVME_SC_INVALID_OPCODE:
 360                 status = SAM_STAT_CHECK_CONDITION;
 361                 sense_key = ILLEGAL_REQUEST;
 362                 asc = SCSI_ASC_ILLEGAL_COMMAND;
 363                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 364                 break;
 365         case NVME_SC_INVALID_FIELD:
 366                 status = SAM_STAT_CHECK_CONDITION;
 367                 sense_key = ILLEGAL_REQUEST;
 368                 asc = SCSI_ASC_INVALID_CDB;
 369                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 370                 break;
 371         case NVME_SC_DATA_XFER_ERROR:
 372                 status = SAM_STAT_CHECK_CONDITION;
 373                 sense_key = MEDIUM_ERROR;
 374                 asc = SCSI_ASC_NO_SENSE;
 375                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 376                 break;
 377         case NVME_SC_POWER_LOSS:
 378                 status = SAM_STAT_TASK_ABORTED;
 379                 sense_key = ABORTED_COMMAND;
 380                 asc = SCSI_ASC_WARNING;
 381                 ascq = SCSI_ASCQ_POWER_LOSS_EXPECTED;
 382                 break;
 383         case NVME_SC_INTERNAL:
 384                 status = SAM_STAT_CHECK_CONDITION;
 385                 sense_key = HARDWARE_ERROR;
 386                 asc = SCSI_ASC_INTERNAL_TARGET_FAILURE;
 387                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 388                 break;
 389         case NVME_SC_ABORT_REQ:
 390                 status = SAM_STAT_TASK_ABORTED;
 391                 sense_key = ABORTED_COMMAND;
 392                 asc = SCSI_ASC_NO_SENSE;
 393                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 394                 break;
 395         case NVME_SC_ABORT_QUEUE:
 396                 status = SAM_STAT_TASK_ABORTED;
 397                 sense_key = ABORTED_COMMAND;
 398                 asc = SCSI_ASC_NO_SENSE;
 399                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 400                 break;
 401         case NVME_SC_FUSED_FAIL:
 402                 status = SAM_STAT_TASK_ABORTED;
 403                 sense_key = ABORTED_COMMAND;
 404                 asc = SCSI_ASC_NO_SENSE;
 405                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 406                 break;
 407         case NVME_SC_FUSED_MISSING:
 408                 status = SAM_STAT_TASK_ABORTED;
 409                 sense_key = ABORTED_COMMAND;
 410                 asc = SCSI_ASC_NO_SENSE;
 411                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 412                 break;
 413         case NVME_SC_INVALID_NS:
 414                 status = SAM_STAT_CHECK_CONDITION;
 415                 sense_key = ILLEGAL_REQUEST;
 416                 asc = SCSI_ASC_ACCESS_DENIED_INVALID_LUN_ID;
 417                 ascq = SCSI_ASCQ_INVALID_LUN_ID;
 418                 break;
 419         case NVME_SC_LBA_RANGE:
 420                 status = SAM_STAT_CHECK_CONDITION;
 421                 sense_key = ILLEGAL_REQUEST;
 422                 asc = SCSI_ASC_ILLEGAL_BLOCK;
 423                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 424                 break;
 425         case NVME_SC_CAP_EXCEEDED:
 426                 status = SAM_STAT_CHECK_CONDITION;
 427                 sense_key = MEDIUM_ERROR;
 428                 asc = SCSI_ASC_NO_SENSE;
 429                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 430                 break;
 431         case NVME_SC_NS_NOT_READY:
 432                 status = SAM_STAT_CHECK_CONDITION;
 433                 sense_key = NOT_READY;
 434                 asc = SCSI_ASC_LUN_NOT_READY;
 435                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 436                 break;
 437
 438         /* Command Specific Status */
 439         case NVME_SC_INVALID_FORMAT:
 440                 status = SAM_STAT_CHECK_CONDITION;
 441                 sense_key = ILLEGAL_REQUEST;
 442                 asc = SCSI_ASC_FORMAT_COMMAND_FAILED;
 443                 ascq = SCSI_ASCQ_FORMAT_COMMAND_FAILED;
 444                 break;
 445         case NVME_SC_BAD_ATTRIBUTES:
 446                 status = SAM_STAT_CHECK_CONDITION;
 447                 sense_key = ILLEGAL_REQUEST;
 448                 asc = SCSI_ASC_INVALID_CDB;
 449                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 450                 break;
 451
 452         /* Media Errors */
 453         case NVME_SC_WRITE_FAULT:
 454                 status = SAM_STAT_CHECK_CONDITION;
 455                 sense_key = MEDIUM_ERROR;
 456                 asc = SCSI_ASC_PERIPHERAL_DEV_WRITE_FAULT;
 457                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 458                 break;
 459         case NVME_SC_READ_ERROR:
 460                 status = SAM_STAT_CHECK_CONDITION;
 461                 sense_key = MEDIUM_ERROR;
 462                 asc = SCSI_ASC_UNRECOVERED_READ_ERROR;
 463                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 464                 break;
 465         case NVME_SC_GUARD_CHECK:
 466                 status = SAM_STAT_CHECK_CONDITION;
 467                 sense_key = MEDIUM_ERROR;
 468                 asc = SCSI_ASC_LOG_BLOCK_GUARD_CHECK_FAILED;
 469                 ascq = SCSI_ASCQ_LOG_BLOCK_GUARD_CHECK_FAILED;
 470                 break;
 471         case NVME_SC_APPTAG_CHECK:
 472                 status = SAM_STAT_CHECK_CONDITION;
 473                 sense_key = MEDIUM_ERROR;
 474                 asc = SCSI_ASC_LOG_BLOCK_APPTAG_CHECK_FAILED;
 475                 ascq = SCSI_ASCQ_LOG_BLOCK_APPTAG_CHECK_FAILED;
 476                 break;
 477         case NVME_SC_REFTAG_CHECK:
 478                 status = SAM_STAT_CHECK_CONDITION;
 479                 sense_key = MEDIUM_ERROR;
 480                 asc = SCSI_ASC_LOG_BLOCK_REFTAG_CHECK_FAILED;
 481                 ascq = SCSI_ASCQ_LOG_BLOCK_REFTAG_CHECK_FAILED;
 482                 break;
 483         case NVME_SC_COMPARE_FAILED:
 484                 status = SAM_STAT_CHECK_CONDITION;
 485                 sense_key = MISCOMPARE;
 486                 asc = SCSI_ASC_MISCOMPARE_DURING_VERIFY;
 487                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 488                 break;
 489         case NVME_SC_ACCESS_DENIED:
 490                 status = SAM_STAT_CHECK_CONDITION;
 491                 sense_key = ILLEGAL_REQUEST;
 492                 asc = SCSI_ASC_ACCESS_DENIED_INVALID_LUN_ID;
 493                 ascq = SCSI_ASCQ_INVALID_LUN_ID;
 494                 break;
 495
 496         /* Unspecified/Default */
 497         case NVME_SC_CMDID_CONFLICT:
 498         case NVME_SC_CMD_SEQ_ERROR:
 499         case NVME_SC_CQ_INVALID:
 500         case NVME_SC_QID_INVALID:
 501         case NVME_SC_QUEUE_SIZE:
 502         case NVME_SC_ABORT_LIMIT:
 503         case NVME_SC_ABORT_MISSING:
 504         case NVME_SC_ASYNC_LIMIT:
 505         case NVME_SC_FIRMWARE_SLOT:
 506         case NVME_SC_FIRMWARE_IMAGE:
 507         case NVME_SC_INVALID_VECTOR:
 508         case NVME_SC_INVALID_LOG_PAGE:
 509         default:
 510                 status = SAM_STAT_CHECK_CONDITION;
 511                 sense_key = ILLEGAL_REQUEST;
 512                 asc = SCSI_ASC_NO_SENSE;
 513                 ascq = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
 514                 break;
 515         }
 516
 517         res = nvme_trans_completion(hdr, status, sense_key, asc, ascq);
 518         return res ? res : nvme_sc;
 519 }
 520
 521 /* INQUIRY Helper Functions */
 522
 523 static int nvme_trans_standard_inquiry_page(struct nvme_ns *ns,
 524                                         struct sg_io_hdr *hdr, u8 *inq_response,
 525                                         int alloc_len)
 526 {
 527         struct nvme_dev *dev = ns->dev;
 528         struct nvme_id_ns *id_ns;
 529         int res;
 530         int nvme_sc;
 531         int xfer_len;
 532         u8 resp_data_format = 0x02;
 533         u8 protect;
 534         u8 cmdque = 0x01 << 1;
 535         u8 fw_offset = sizeof(dev->firmware_rev);
 536
 537         /* nvme ns identify - use DPS value for PROTECT field */
 538         nvme_sc = nvme_identify_ns(dev, ns->ns_id, &id_ns);
 539         res = nvme_trans_status_code(hdr, nvme_sc);
 540         if (res)
 541                 return res;
 542
 543         if (id_ns->dps)
 544                 protect = 0x01;
 545         else
 546                 protect = 0;
 547         kfree(id_ns);
 548
 549         memset(inq_response, 0, STANDARD_INQUIRY_LENGTH);
 550         inq_response[2] = VERSION_SPC_4;
 551         inq_response[3] = resp_data_format;     /*normaca=0 | hisup=0 */
 552         inq_response[4] = ADDITIONAL_STD_INQ_LENGTH;
 553         inq_response[5] = protect;      /* sccs=0 | acc=0 | tpgs=0 | pc3=0 */
 554         inq_response[7] = cmdque;       /* wbus16=0 | sync=0 | vs=0 */
 555         strncpy(&inq_response[8], "NVMe    ", 8);
 556         strncpy(&inq_response[16], dev->model, 16);
 557
 558         while (dev->firmware_rev[fw_offset - 1] == ' ' && fw_offset > 4)
 559                 fw_offset--;
 560         fw_offset -= 4;
 561         strncpy(&inq_response[32], dev->firmware_rev + fw_offset, 4);
 562
 563         xfer_len = min(alloc_len, STANDARD_INQUIRY_LENGTH);
 564         return nvme_trans_copy_to_user(hdr, inq_response, xfer_len);
 565 }
 566
 567 static int nvme_trans_supported_vpd_pages(struct nvme_ns *ns,
 568                                         struct sg_io_hdr *hdr, u8 *inq_response,
 569                                         int alloc_len)
 570 {
 571         int xfer_len;
 572
 573         memset(inq_response, 0, STANDARD_INQUIRY_LENGTH);
 574         inq_response[1] = INQ_SUPPORTED_VPD_PAGES_PAGE;   /* Page Code */
 575         inq_response[3] = INQ_NUM_SUPPORTED_VPD_PAGES;    /* Page Length */
 576         inq_response[4] = INQ_SUPPORTED_VPD_PAGES_PAGE;
 577         inq_response[5] = INQ_UNIT_SERIAL_NUMBER_PAGE;
 578         inq_response[6] = INQ_DEVICE_IDENTIFICATION_PAGE;
 579         inq_response[7] = INQ_EXTENDED_INQUIRY_DATA_PAGE;
 580         inq_response[8] = INQ_BDEV_CHARACTERISTICS_PAGE;
 581         inq_response[9] = INQ_BDEV_LIMITS_PAGE;
 582
 583         xfer_len = min(alloc_len, STANDARD_INQUIRY_LENGTH);
 584         return nvme_trans_copy_to_user(hdr, inq_response, xfer_len);
 585 }
 586
 587 static int nvme_trans_unit_serial_page(struct nvme_ns *ns,
 588                                         struct sg_io_hdr *hdr, u8 *inq_response,
 589                                         int alloc_len)
 590 {
 591         struct nvme_dev *dev = ns->dev;
 592         int xfer_len;
 593
 594         memset(inq_response, 0, STANDARD_INQUIRY_LENGTH);
 595         inq_response[1] = INQ_UNIT_SERIAL_NUMBER_PAGE; /* Page Code */
 596         inq_response[3] = INQ_SERIAL_NUMBER_LENGTH;    /* Page Length */
 597         strncpy(&inq_response[4], dev->serial, INQ_SERIAL_NUMBER_LENGTH);
 598
 599         xfer_len = min(alloc_len, STANDARD_INQUIRY_LENGTH);
 600         return nvme_trans_copy_to_user(hdr, inq_response, xfer_len);
 601 }
 602
 603 static int nvme_trans_device_id_page(struct nvme_ns *ns, struct sg_io_hdr *hdr,
 604                                         u8 *inq_response, int alloc_len)
 605 {
 606         struct nvme_dev *dev = ns->dev;
 607         int res;
 608         int nvme_sc;
 609         int xfer_len;
 610         __be32 tmp_id = cpu_to_be32(ns->ns_id);
 611
 612         memset(inq_response, 0, alloc_len);
 613         inq_response[1] = INQ_DEVICE_IDENTIFICATION_PAGE;    /* Page Code */
 614         if (readl(&dev->bar->vs) >= NVME_VS(1, 1)) {
 615                 struct nvme_id_ns *id_ns;
 616                 void *eui;
 617                 int len;
 618
 619                 nvme_sc = nvme_identify_ns(dev, ns->ns_id, &id_ns);
 620                 res = nvme_trans_status_code(hdr, nvme_sc);
 621                 if (res)
 622                         return res;
 623
 624                 eui = id_ns->eui64;
 625                 len = sizeof(id_ns->eui64);
 626                 if (readl(&dev->bar->vs) >= NVME_VS(1, 2)) {
 627                         if (bitmap_empty(eui, len * 8)) {
 628                                 eui = id_ns->nguid;
 629                                 len = sizeof(id_ns->nguid);
 630                         }
 631                 }
 632                 if (bitmap_empty(eui, len * 8)) {
 633                         kfree(id_ns);
 634                         goto scsi_string;
 635                 }
 636
 637                 inq_response[3] = 4 + len; /* Page Length */
 638                 /* Designation Descriptor start */
 639                 inq_response[4] = 0x01;    /* Proto ID=0h | Code set=1h */
 640                 inq_response[5] = 0x02;    /* PIV=0b | Asso=00b | Designator Type=2h */
 641                 inq_response[6] = 0x00;    /* Rsvd */
 642                 inq_response[7] = len;     /* Designator Length */
 643                 memcpy(&inq_response[8], eui, len);
 644                 kfree(id_ns);
 645         } else {
 646  scsi_string:
 647                 if (alloc_len < 72) {
 648                         return nvme_trans_completion(hdr,
 649                                         SAM_STAT_CHECK_CONDITION,
 650                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
 651                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
 652                 }
 653                 inq_response[3] = 0x48;    /* Page Length */
 654                 /* Designation Descriptor start */
 655                 inq_response[4] = 0x03;    /* Proto ID=0h | Code set=3h */
 656                 inq_response[5] = 0x08;    /* PIV=0b | Asso=00b | Designator Type=8h */
 657                 inq_response[6] = 0x00;    /* Rsvd */
 658                 inq_response[7] = 0x44;    /* Designator Length */
 659
 660                 sprintf(&inq_response[8], "%04x", to_pci_dev(dev->dev)->vendor);
 661                 memcpy(&inq_response[12], dev->model, sizeof(dev->model));
 662                 sprintf(&inq_response[52], "%04x", tmp_id);
 663                 memcpy(&inq_response[56], dev->serial, sizeof(dev->serial));
 664         }
 665         xfer_len = alloc_len;
 666         return nvme_trans_copy_to_user(hdr, inq_response, xfer_len);
 667 }
 668
 669 static int nvme_trans_ext_inq_page(struct nvme_ns *ns, struct sg_io_hdr *hdr,
 670                                         int alloc_len)
 671 {
 672         u8 *inq_response;
 673         int res;
 674         int nvme_sc;
 675         struct nvme_dev *dev = ns->dev;
 676         struct nvme_id_ctrl *id_ctrl;
 677         struct nvme_id_ns *id_ns;
 678         int xfer_len;
 679         u8 microcode = 0x80;
 680         u8 spt;
 681         u8 spt_lut[8] = {0, 0, 2, 1, 4, 6, 5, 7};
 682         u8 grd_chk, app_chk, ref_chk, protect;
 683         u8 uask_sup = 0x20;
 684         u8 v_sup;
 685         u8 luiclr = 0x01;
 686
 687         inq_response = kmalloc(EXTENDED_INQUIRY_DATA_PAGE_LENGTH, GFP_KERNEL);
 688         if (inq_response == NULL)
 689                 return -ENOMEM;
 690
 691         nvme_sc = nvme_identify_ns(dev, ns->ns_id, &id_ns);
 692         res = nvme_trans_status_code(hdr, nvme_sc);
 693         if (res)
 694                 goto out_free_inq;
 695
 696         spt = spt_lut[id_ns->dpc & 0x07] << 3;
 697         if (id_ns->dps)
 698                 protect = 0x01;
 699         else
 700                 protect = 0;
 701         kfree(id_ns);
 702
 703         grd_chk = protect << 2;
 704         app_chk = protect << 1;
 705         ref_chk = protect;
 706
 707         nvme_sc = nvme_identify_ctrl(dev, &id_ctrl);
 708         res = nvme_trans_status_code(hdr, nvme_sc);
 709         if (res)
 710                 goto out_free_inq;
 711
 712         v_sup = id_ctrl->vwc;
 713         kfree(id_ctrl);
 714
 715         memset(inq_response, 0, EXTENDED_INQUIRY_DATA_PAGE_LENGTH);
 716         inq_response[1] = INQ_EXTENDED_INQUIRY_DATA_PAGE;    /* Page Code */
 717         inq_response[2] = 0x00;    /* Page Length MSB */
 718         inq_response[3] = 0x3C;    /* Page Length LSB */
 719         inq_response[4] = microcode | spt | grd_chk | app_chk | ref_chk;
 720         inq_response[5] = uask_sup;
 721         inq_response[6] = v_sup;
 722         inq_response[7] = luiclr;
 723         inq_response[8] = 0;
 724         inq_response[9] = 0;
 725
 726         xfer_len = min(alloc_len, EXTENDED_INQUIRY_DATA_PAGE_LENGTH);
 727         res = nvme_trans_copy_to_user(hdr, inq_response, xfer_len);
 728
 729  out_free_inq:
 730         kfree(inq_response);
 731         return res;
 732 }
 733
 734 static int nvme_trans_bdev_limits_page(struct nvme_ns *ns, struct sg_io_hdr *hdr,
 735                                         u8 *inq_response, int alloc_len)
 736 {
 737         __be32 max_sectors = cpu_to_be32(queue_max_hw_sectors(ns->queue));
 738         __be32 max_discard = cpu_to_be32(ns->queue->limits.max_discard_sectors);
 739         __be32 discard_desc_count = cpu_to_be32(0x100);
 740
 741         memset(inq_response, 0, STANDARD_INQUIRY_LENGTH);
 742         inq_response[1] = VPD_BLOCK_LIMITS;
 743         inq_response[3] = 0x3c; /* Page Length */
 744         memcpy(&inq_response[8], &max_sectors, sizeof(u32));
 745         memcpy(&inq_response[20], &max_discard, sizeof(u32));
 746
 747         if (max_discard)
 748                 memcpy(&inq_response[24], &discard_desc_count, sizeof(u32));
 749
 750         return nvme_trans_copy_to_user(hdr, inq_response, 0x3c);
 751 }
 752
 753 static int nvme_trans_bdev_char_page(struct nvme_ns *ns, struct sg_io_hdr *hdr,
 754                                         int alloc_len)
 755 {
 756         u8 *inq_response;
 757         int res;
 758         int xfer_len;
 759
 760         inq_response = kzalloc(EXTENDED_INQUIRY_DATA_PAGE_LENGTH, GFP_KERNEL);
 761         if (inq_response == NULL) {
 762                 res = -ENOMEM;
 763                 goto out_mem;
 764         }
 765
 766         inq_response[1] = INQ_BDEV_CHARACTERISTICS_PAGE;    /* Page Code */
 767         inq_response[2] = 0x00;    /* Page Length MSB */
 768         inq_response[3] = 0x3C;    /* Page Length LSB */
 769         inq_response[4] = 0x00;    /* Medium Rotation Rate MSB */
 770         inq_response[5] = 0x01;    /* Medium Rotation Rate LSB */
 771         inq_response[6] = 0x00;    /* Form Factor */
 772
 773         xfer_len = min(alloc_len, EXTENDED_INQUIRY_DATA_PAGE_LENGTH);
 774         res = nvme_trans_copy_to_user(hdr, inq_response, xfer_len);
 775
 776         kfree(inq_response);
 777  out_mem:
 778         return res;
 779 }
 780
 781 /* LOG SENSE Helper Functions */
 782
 783 static int nvme_trans_log_supp_pages(struct nvme_ns *ns, struct sg_io_hdr *hdr,
 784                                         int alloc_len)
 785 {
 786         int res;
 787         int xfer_len;
 788         u8 *log_response;
 789
 790         log_response = kzalloc(LOG_PAGE_SUPPORTED_LOG_PAGES_LENGTH, GFP_KERNEL);
 791         if (log_response == NULL) {
 792                 res = -ENOMEM;
 793                 goto out_mem;
 794         }
 795
 796         log_response[0] = LOG_PAGE_SUPPORTED_LOG_PAGES_PAGE;
 797         /* Subpage=0x00, Page Length MSB=0 */
 798         log_response[3] = SUPPORTED_LOG_PAGES_PAGE_LENGTH;
 799         log_response[4] = LOG_PAGE_SUPPORTED_LOG_PAGES_PAGE;
 800         log_response[5] = LOG_PAGE_INFORMATIONAL_EXCEPTIONS_PAGE;
 801         log_response[6] = LOG_PAGE_TEMPERATURE_PAGE;
 802
 803         xfer_len = min(alloc_len, LOG_PAGE_SUPPORTED_LOG_PAGES_LENGTH);
 804         res = nvme_trans_copy_to_user(hdr, log_response, xfer_len);
 805
 806         kfree(log_response);
 807  out_mem:
 808         return res;
 809 }
 810
 811 static int nvme_trans_log_info_exceptions(struct nvme_ns *ns,
 812                                         struct sg_io_hdr *hdr, int alloc_len)
 813 {
 814         int res;
 815         int xfer_len;
 816         u8 *log_response;
 817         struct nvme_dev *dev = ns->dev;
 818         struct nvme_smart_log *smart_log;
 819         u8 temp_c;
 820         u16 temp_k;
 821
 822         log_response = kzalloc(LOG_INFO_EXCP_PAGE_LENGTH, GFP_KERNEL);
 823         if (log_response == NULL)
 824                 return -ENOMEM;
 825
 826         res = nvme_get_log_page(dev, &smart_log);
 827         if (res < 0)
 828                 goto out_free_response;
 829
 830         if (res != NVME_SC_SUCCESS) {
 831                 temp_c = LOG_TEMP_UNKNOWN;
 832         } else {
 833                 temp_k = (smart_log->temperature[1] << 8) +
 834                                 (smart_log->temperature[0]);
 835                 temp_c = temp_k - KELVIN_TEMP_FACTOR;
 836         }
 837         kfree(smart_log);
 838
 839         log_response[0] = LOG_PAGE_INFORMATIONAL_EXCEPTIONS_PAGE;
 840         /* Subpage=0x00, Page Length MSB=0 */
 841         log_response[3] = REMAINING_INFO_EXCP_PAGE_LENGTH;
 842         /* Informational Exceptions Log Parameter 1 Start */
 843         /* Parameter Code=0x0000 bytes 4,5 */
 844         log_response[6] = 0x23; /* DU=0, TSD=1, ETC=0, TMC=0, FMT_AND_LNK=11b */
 845         log_response[7] = 0x04; /* PARAMETER LENGTH */
 846         /* Add sense Code and qualifier = 0x00 each */
 847         /* Use Temperature from NVMe Get Log Page, convert to C from K */
 848         log_response[10] = temp_c;
 849
 850         xfer_len = min(alloc_len, LOG_INFO_EXCP_PAGE_LENGTH);
 851         res = nvme_trans_copy_to_user(hdr, log_response, xfer_len);
 852
 853  out_free_response:
 854         kfree(log_response);
 855         return res;
 856 }
 857
 858 static int nvme_trans_log_temperature(struct nvme_ns *ns, struct sg_io_hdr *hdr,
 859                                         int alloc_len)
 860 {
 861         int res;
 862         int xfer_len;
 863         u8 *log_response;
 864         struct nvme_dev *dev = ns->dev;
 865         struct nvme_smart_log *smart_log;
 866         u32 feature_resp;
 867         u8 temp_c_cur, temp_c_thresh;
 868         u16 temp_k;
 869
 870         log_response = kzalloc(LOG_TEMP_PAGE_LENGTH, GFP_KERNEL);
 871         if (log_response == NULL)
 872                 return -ENOMEM;
 873
 874         res = nvme_get_log_page(dev, &smart_log);
 875         if (res < 0)
 876                 goto out_free_response;
 877
 878         if (res != NVME_SC_SUCCESS) {
 879                 temp_c_cur = LOG_TEMP_UNKNOWN;
 880         } else {
 881                 temp_k = (smart_log->temperature[1] << 8) +
 882                                 (smart_log->temperature[0]);
 883                 temp_c_cur = temp_k - KELVIN_TEMP_FACTOR;
 884         }
 885         kfree(smart_log);
 886
 887         /* Get Features for Temp Threshold */
 888         res = nvme_get_features(dev, NVME_FEAT_TEMP_THRESH, 0, 0,
 889                                                                 &feature_resp);
 890         if (res != NVME_SC_SUCCESS)
 891                 temp_c_thresh = LOG_TEMP_UNKNOWN;
 892         else
 893                 temp_c_thresh = (feature_resp & 0xFFFF) - KELVIN_TEMP_FACTOR;
 894
 895         log_response[0] = LOG_PAGE_TEMPERATURE_PAGE;
 896         /* Subpage=0x00, Page Length MSB=0 */
 897         log_response[3] = REMAINING_TEMP_PAGE_LENGTH;
 898         /* Temperature Log Parameter 1 (Temperature) Start */
 899         /* Parameter Code = 0x0000 */
 900         log_response[6] = 0x01;         /* Format and Linking = 01b */
 901         log_response[7] = 0x02;         /* Parameter Length */
 902         /* Use Temperature from NVMe Get Log Page, convert to C from K */
 903         log_response[9] = temp_c_cur;
 904         /* Temperature Log Parameter 2 (Reference Temperature) Start */
 905         log_response[11] = 0x01;        /* Parameter Code = 0x0001 */
 906         log_response[12] = 0x01;        /* Format and Linking = 01b */
 907         log_response[13] = 0x02;        /* Parameter Length */
 908         /* Use Temperature Thresh from NVMe Get Log Page, convert to C from K */
 909         log_response[15] = temp_c_thresh;
 910
 911         xfer_len = min(alloc_len, LOG_TEMP_PAGE_LENGTH);
 912         res = nvme_trans_copy_to_user(hdr, log_response, xfer_len);
 913
 914  out_free_response:
 915         kfree(log_response);
 916         return res;
 917 }
 918
 919 /* MODE SENSE Helper Functions */
 920
 921 static int nvme_trans_fill_mode_parm_hdr(u8 *resp, int len, u8 cdb10, u8 llbaa,
 922                                         u16 mode_data_length, u16 blk_desc_len)
 923 {
 924         /* Quick check to make sure I don't stomp on my own memory... */
 925         if ((cdb10 && len < 8) || (!cdb10 && len < 4))
 926                 return -EINVAL;
 927
 928         if (cdb10) {
 929                 resp[0] = (mode_data_length & 0xFF00) >> 8;
 930                 resp[1] = (mode_data_length & 0x00FF);
 931                 resp[3] = 0x10 /* DPOFUA */;
 932                 resp[4] = llbaa;
 933                 resp[5] = RESERVED_FIELD;
 934                 resp[6] = (blk_desc_len & 0xFF00) >> 8;
 935                 resp[7] = (blk_desc_len & 0x00FF);
 936         } else {
 937                 resp[0] = (mode_data_length & 0x00FF);
 938                 resp[2] = 0x10 /* DPOFUA */;
 939                 resp[3] = (blk_desc_len & 0x00FF);
 940         }
 941
 942         return 0;
 943 }
 944
 945 static int nvme_trans_fill_blk_desc(struct nvme_ns *ns, struct sg_io_hdr *hdr,
 946                                     u8 *resp, int len, u8 llbaa)
 947 {
 948         int res;
 949         int nvme_sc;
 950         struct nvme_dev *dev = ns->dev;
 951         struct nvme_id_ns *id_ns;
 952         u8 flbas;
 953         u32 lba_length;
 954
 955         if (llbaa == 0 && len < MODE_PAGE_BLK_DES_LEN)
 956                 return -EINVAL;
 957         else if (llbaa > 0 && len < MODE_PAGE_LLBAA_BLK_DES_LEN)
 958                 return -EINVAL;
 959
 960         nvme_sc = nvme_identify_ns(dev, ns->ns_id, &id_ns);
 961         res = nvme_trans_status_code(hdr, nvme_sc);
 962         if (res)
 963                 return res;
 964
 965         flbas = (id_ns->flbas) & 0x0F;
 966         lba_length = (1 << (id_ns->lbaf[flbas].ds));
 967
 968         if (llbaa == 0) {
 969                 __be32 tmp_cap = cpu_to_be32(le64_to_cpu(id_ns->ncap));
 970                 /* Byte 4 is reserved */
 971                 __be32 tmp_len = cpu_to_be32(lba_length & 0x00FFFFFF);
 972
 973                 memcpy(resp, &tmp_cap, sizeof(u32));
 974                 memcpy(&resp[4], &tmp_len, sizeof(u32));
 975         } else {
 976                 __be64 tmp_cap = cpu_to_be64(le64_to_cpu(id_ns->ncap));
 977                 __be32 tmp_len = cpu_to_be32(lba_length);
 978
 979                 memcpy(resp, &tmp_cap, sizeof(u64));
 980                 /* Bytes 8, 9, 10, 11 are reserved */
 981                 memcpy(&resp[12], &tmp_len, sizeof(u32));
 982         }
 983
 984         kfree(id_ns);
 985         return res;
 986 }
 987
 988 static int nvme_trans_fill_control_page(struct nvme_ns *ns,
 989                                         struct sg_io_hdr *hdr, u8 *resp,
 990                                         int len)
 991 {
 992         if (len < MODE_PAGE_CONTROL_LEN)
 993                 return -EINVAL;
 994
 995         resp[0] = MODE_PAGE_CONTROL;
 996         resp[1] = MODE_PAGE_CONTROL_LEN_FIELD;
 997         resp[2] = 0x0E;         /* TST=000b, TMF_ONLY=0, DPICZ=1,
 998                                  * D_SENSE=1, GLTSD=1, RLEC=0 */
 999         resp[3] = 0x12;         /* Q_ALGO_MODIFIER=1h, NUAR=0, QERR=01b */
1000         /* Byte 4:  VS=0, RAC=0, UA_INT=0, SWP=0 */
1001         resp[5] = 0x40;         /* ATO=0, TAS=1, ATMPE=0, RWWP=0, AUTOLOAD=0 */
1002         /* resp[6] and [7] are obsolete, thus zero */
1003         resp[8] = 0xFF;         /* Busy timeout period = 0xffff */
1004         resp[9] = 0xFF;
1005         /* Bytes 10,11: Extended selftest completion time = 0x0000 */
1006
1007         return 0;
1008 }
1009
1010 static int nvme_trans_fill_caching_page(struct nvme_ns *ns,
1011                                         struct sg_io_hdr *hdr,
1012                                         u8 *resp, int len)
1013 {
1014         int res = 0;
1015         int nvme_sc;
1016         struct nvme_dev *dev = ns->dev;
1017         u32 feature_resp;
1018         u8 vwc;
1019
1020         if (len < MODE_PAGE_CACHING_LEN)
1021                 return -EINVAL;
1022
1023         nvme_sc = nvme_get_features(dev, NVME_FEAT_VOLATILE_WC, 0, 0,
1024                                                                 &feature_resp);
1025         res = nvme_trans_status_code(hdr, nvme_sc);
1026         if (res)
1027                 return res;
1028
1029         vwc = feature_resp & 0x00000001;
1030
1031         resp[0] = MODE_PAGE_CACHING;
1032         resp[1] = MODE_PAGE_CACHING_LEN_FIELD;
1033         resp[2] = vwc << 2;
1034         return 0;
1035 }
1036
1037 static int nvme_trans_fill_pow_cnd_page(struct nvme_ns *ns,
1038                                         struct sg_io_hdr *hdr, u8 *resp,
1039                                         int len)
1040 {
1041         if (len < MODE_PAGE_POW_CND_LEN)
1042                 return -EINVAL;
1043
1044         resp[0] = MODE_PAGE_POWER_CONDITION;
1045         resp[1] = MODE_PAGE_POW_CND_LEN_FIELD;
1046         /* All other bytes are zero */
1047
1048         return 0;
1049 }
1050
1051 static int nvme_trans_fill_inf_exc_page(struct nvme_ns *ns,
1052                                         struct sg_io_hdr *hdr, u8 *resp,
1053                                         int len)
1054 {
1055         if (len < MODE_PAGE_INF_EXC_LEN)
1056                 return -EINVAL;
1057
1058         resp[0] = MODE_PAGE_INFO_EXCEP;
1059         resp[1] = MODE_PAGE_INF_EXC_LEN_FIELD;
1060         resp[2] = 0x88;
1061         /* All other bytes are zero */
1062
1063         return 0;
1064 }
1065
1066 static int nvme_trans_fill_all_pages(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1067                                      u8 *resp, int len)
1068 {
1069         int res;
1070         u16 mode_pages_offset_1 = 0;
1071         u16 mode_pages_offset_2, mode_pages_offset_3, mode_pages_offset_4;
1072
1073         mode_pages_offset_2 = mode_pages_offset_1 + MODE_PAGE_CACHING_LEN;
1074         mode_pages_offset_3 = mode_pages_offset_2 + MODE_PAGE_CONTROL_LEN;
1075         mode_pages_offset_4 = mode_pages_offset_3 + MODE_PAGE_POW_CND_LEN;
1076
1077         res = nvme_trans_fill_caching_page(ns, hdr, &resp[mode_pages_offset_1],
1078                                         MODE_PAGE_CACHING_LEN);
1079         if (res)
1080                 return res;
1081         res = nvme_trans_fill_control_page(ns, hdr, &resp[mode_pages_offset_2],
1082                                         MODE_PAGE_CONTROL_LEN);
1083         if (res)
1084                 return res;
1085         res = nvme_trans_fill_pow_cnd_page(ns, hdr, &resp[mode_pages_offset_3],
1086                                         MODE_PAGE_POW_CND_LEN);
1087         if (res)
1088                 return res;
1089         return nvme_trans_fill_inf_exc_page(ns, hdr, &resp[mode_pages_offset_4],
1090                                         MODE_PAGE_INF_EXC_LEN);
1091 }
1092
1093 static inline int nvme_trans_get_blk_desc_len(u8 dbd, u8 llbaa)
1094 {
1095         if (dbd == MODE_SENSE_BLK_DESC_ENABLED) {
1096                 /* SPC-4: len = 8 x Num_of_descriptors if llbaa = 0, 16x if 1 */
1097                 return 8 * (llbaa + 1) * MODE_SENSE_BLK_DESC_COUNT;
1098         } else {
1099                 return 0;
1100         }
1101 }
1102
1103 static int nvme_trans_mode_page_create(struct nvme_ns *ns,
1104                                         struct sg_io_hdr *hdr, u8 *cmd,
1105                                         u16 alloc_len, u8 cdb10,
1106                                         int (*mode_page_fill_func)
1107                                         (struct nvme_ns *,
1108                                         struct sg_io_hdr *hdr, u8 *, int),
1109                                         u16 mode_pages_tot_len)
1110 {
1111         int res;
1112         int xfer_len;
1113         u8 *response;
1114         u8 dbd, llbaa;
1115         u16 resp_size;
1116         int mph_size;
1117         u16 mode_pages_offset_1;
1118         u16 blk_desc_len, blk_desc_offset, mode_data_length;
1119
1120         dbd = (cmd[1] & MODE_SENSE_DBD_MASK) >> MODE_SENSE_DBD_SHIFT;
1121         llbaa = (cmd[1] & MODE_SENSE_LLBAA_MASK) >> MODE_SENSE_LLBAA_SHIFT;
1122         mph_size = cdb10 ? MODE_SENSE10_MPH_SIZE : MODE_SENSE6_MPH_SIZE;
1123
1124         blk_desc_len = nvme_trans_get_blk_desc_len(dbd, llbaa);
1125
1126         resp_size = mph_size + blk_desc_len + mode_pages_tot_len;
1127         /* Refer spc4r34 Table 440 for calculation of Mode data Length field */
1128         mode_data_length = 3 + (3 * cdb10) + blk_desc_len + mode_pages_tot_len;
1129
1130         blk_desc_offset = mph_size;
1131         mode_pages_offset_1 = blk_desc_offset + blk_desc_len;
1132
1133         response = kzalloc(resp_size, GFP_KERNEL);
1134         if (response == NULL) {
1135                 res = -ENOMEM;
1136                 goto out_mem;
1137         }
1138
1139         res = nvme_trans_fill_mode_parm_hdr(&response[0], mph_size, cdb10,
1140                                         llbaa, mode_data_length, blk_desc_len);
1141         if (res)
1142                 goto out_free;
1143         if (blk_desc_len > 0) {
1144                 res = nvme_trans_fill_blk_desc(ns, hdr,
1145                                                &response[blk_desc_offset],
1146                                                blk_desc_len, llbaa);
1147                 if (res)
1148                         goto out_free;
1149         }
1150         res = mode_page_fill_func(ns, hdr, &response[mode_pages_offset_1],
1151                                         mode_pages_tot_len);
1152         if (res)
1153                 goto out_free;
1154
1155         xfer_len = min(alloc_len, resp_size);
1156         res = nvme_trans_copy_to_user(hdr, response, xfer_len);
1157
1158  out_free:
1159         kfree(response);
1160  out_mem:
1161         return res;
1162 }
1163
1164 /* Read Capacity Helper Functions */
1165
1166 static void nvme_trans_fill_read_cap(u8 *response, struct nvme_id_ns *id_ns,
1167                                                                 u8 cdb16)
1168 {
1169         u8 flbas;
1170         u32 lba_length;
1171         u64 rlba;
1172         u8 prot_en;
1173         u8 p_type_lut[4] = {0, 0, 1, 2};
1174         __be64 tmp_rlba;
1175         __be32 tmp_rlba_32;
1176         __be32 tmp_len;
1177
1178         flbas = (id_ns->flbas) & 0x0F;
1179         lba_length = (1 << (id_ns->lbaf[flbas].ds));
1180         rlba = le64_to_cpup(&id_ns->nsze) - 1;
1181         (id_ns->dps) ? (prot_en = 0x01) : (prot_en = 0);
1182
1183         if (!cdb16) {
1184                 if (rlba > 0xFFFFFFFF)
1185                         rlba = 0xFFFFFFFF;
1186                 tmp_rlba_32 = cpu_to_be32(rlba);
1187                 tmp_len = cpu_to_be32(lba_length);
1188                 memcpy(response, &tmp_rlba_32, sizeof(u32));
1189                 memcpy(&response[4], &tmp_len, sizeof(u32));
1190         } else {
1191                 tmp_rlba = cpu_to_be64(rlba);
1192                 tmp_len = cpu_to_be32(lba_length);
1193                 memcpy(response, &tmp_rlba, sizeof(u64));
1194                 memcpy(&response[8], &tmp_len, sizeof(u32));
1195                 response[12] = (p_type_lut[id_ns->dps & 0x3] << 1) | prot_en;
1196                 /* P_I_Exponent = 0x0 | LBPPBE = 0x0 */
1197                 /* LBPME = 0 | LBPRZ = 0 | LALBA = 0x00 */
1198                 /* Bytes 16-31 - Reserved */
1199         }
1200 }
1201
1202 /* Start Stop Unit Helper Functions */
1203
1204 static int nvme_trans_power_state(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1205                                                 u8 pc, u8 pcmod, u8 start)
1206 {
1207         int res;
1208         int nvme_sc;
1209         struct nvme_dev *dev = ns->dev;
1210         struct nvme_id_ctrl *id_ctrl;
1211         int lowest_pow_st;      /* max npss = lowest power consumption */
1212         unsigned ps_desired = 0;
1213
1214         nvme_sc = nvme_identify_ctrl(dev, &id_ctrl);
1215         res = nvme_trans_status_code(hdr, nvme_sc);
1216         if (res)
1217                 return res;
1218
1219         lowest_pow_st = max(POWER_STATE_0, (int)(id_ctrl->npss - 1));
1220         kfree(id_ctrl);
1221
1222         switch (pc) {
1223         case NVME_POWER_STATE_START_VALID:
1224                 /* Action unspecified if POWER CONDITION MODIFIER != 0 */
1225                 if (pcmod == 0 && start == 0x1)
1226                         ps_desired = POWER_STATE_0;
1227                 if (pcmod == 0 && start == 0x0)
1228                         ps_desired = lowest_pow_st;
1229                 break;
1230         case NVME_POWER_STATE_ACTIVE:
1231                 /* Action unspecified if POWER CONDITION MODIFIER != 0 */
1232                 if (pcmod == 0)
1233                         ps_desired = POWER_STATE_0;
1234                 break;
1235         case NVME_POWER_STATE_IDLE:
1236                 /* Action unspecified if POWER CONDITION MODIFIER != [0,1,2] */
1237                 if (pcmod == 0x0)
1238                         ps_desired = POWER_STATE_1;
1239                 else if (pcmod == 0x1)
1240                         ps_desired = POWER_STATE_2;
1241                 else if (pcmod == 0x2)
1242                         ps_desired = POWER_STATE_3;
1243                 break;
1244         case NVME_POWER_STATE_STANDBY:
1245                 /* Action unspecified if POWER CONDITION MODIFIER != [0,1] */
1246                 if (pcmod == 0x0)
1247                         ps_desired = max(POWER_STATE_0, (lowest_pow_st - 2));
1248                 else if (pcmod == 0x1)
1249                         ps_desired = max(POWER_STATE_0, (lowest_pow_st - 1));
1250                 break;
1251         case NVME_POWER_STATE_LU_CONTROL:
1252         default:
1253                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1254                                 ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
1255                                 SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1256                 break;
1257         }
1258         nvme_sc = nvme_set_features(dev, NVME_FEAT_POWER_MGMT, ps_desired, 0,
1259                                     NULL);
1260         return nvme_trans_status_code(hdr, nvme_sc);
1261 }
1262
1263 static int nvme_trans_send_activate_fw_cmd(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1264                                         u8 buffer_id)
1265 {
1266         struct nvme_command c;
1267         int nvme_sc;
1268
1269         memset(&c, 0, sizeof(c));
1270         c.common.opcode = nvme_admin_activate_fw;
1271         c.common.cdw10[0] = cpu_to_le32(buffer_id | NVME_FWACT_REPL_ACTV);
1272
1273         nvme_sc = nvme_submit_sync_cmd(ns->queue, &c, NULL, 0);
1274         return nvme_trans_status_code(hdr, nvme_sc);
1275 }
1276
1277 static int nvme_trans_send_download_fw_cmd(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1278                                         u8 opcode, u32 tot_len, u32 offset,
1279                                         u8 buffer_id)
1280 {
1281         int nvme_sc;
1282         struct nvme_dev *dev = ns->dev;
1283         struct nvme_command c;
1284
1285         if (hdr->iovec_count > 0) {
1286                 /* Assuming SGL is not allowed for this command */
1287                 return nvme_trans_completion(hdr,
1288                                         SAM_STAT_CHECK_CONDITION,
1289                                         ILLEGAL_REQUEST,
1290                                         SCSI_ASC_INVALID_CDB,
1291                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1292         }
1293
1294         memset(&c, 0, sizeof(c));
1295         c.common.opcode = nvme_admin_download_fw;
1296         c.dlfw.numd = cpu_to_le32((tot_len/BYTES_TO_DWORDS) - 1);
1297         c.dlfw.offset = cpu_to_le32(offset/BYTES_TO_DWORDS);
1298
1299         nvme_sc = __nvme_submit_sync_cmd(dev->admin_q, &c, NULL,
1300                         hdr->dxferp, tot_len, NULL, 0);
1301         return nvme_trans_status_code(hdr, nvme_sc);
1302 }
1303
1304 /* Mode Select Helper Functions */
1305
1306 static inline void nvme_trans_modesel_get_bd_len(u8 *parm_list, u8 cdb10,
1307                                                 u16 *bd_len, u8 *llbaa)
1308 {
1309         if (cdb10) {
1310                 /* 10 Byte CDB */
1311                 *bd_len = (parm_list[MODE_SELECT_10_BD_OFFSET] << 8) +
1312                         parm_list[MODE_SELECT_10_BD_OFFSET + 1];
1313                 *llbaa = parm_list[MODE_SELECT_10_LLBAA_OFFSET] &
1314                                 MODE_SELECT_10_LLBAA_MASK;
1315         } else {
1316                 /* 6 Byte CDB */
1317                 *bd_len = parm_list[MODE_SELECT_6_BD_OFFSET];
1318         }
1319 }
1320
1321 static void nvme_trans_modesel_save_bd(struct nvme_ns *ns, u8 *parm_list,
1322                                         u16 idx, u16 bd_len, u8 llbaa)
1323 {
1324         u16 bd_num;
1325
1326         bd_num = bd_len / ((llbaa == 0) ?
1327                         SHORT_DESC_BLOCK : LONG_DESC_BLOCK);
1328         /* Store block descriptor info if a FORMAT UNIT comes later */
1329         /* TODO Saving 1st BD info; what to do if multiple BD received? */
1330         if (llbaa == 0) {
1331                 /* Standard Block Descriptor - spc4r34 7.5.5.1 */
1332                 ns->mode_select_num_blocks =
1333                                 (parm_list[idx + 1] << 16) +
1334                                 (parm_list[idx + 2] << 8) +
1335                                 (parm_list[idx + 3]);
1336
1337                 ns->mode_select_block_len =
1338                                 (parm_list[idx + 5] << 16) +
1339                                 (parm_list[idx + 6] << 8) +
1340                                 (parm_list[idx + 7]);
1341         } else {
1342                 /* Long LBA Block Descriptor - sbc3r27 6.4.2.3 */
1343                 ns->mode_select_num_blocks =
1344                                 (((u64)parm_list[idx + 0]) << 56) +
1345                                 (((u64)parm_list[idx + 1]) << 48) +
1346                                 (((u64)parm_list[idx + 2]) << 40) +
1347                                 (((u64)parm_list[idx + 3]) << 32) +
1348                                 (((u64)parm_list[idx + 4]) << 24) +
1349                                 (((u64)parm_list[idx + 5]) << 16) +
1350                                 (((u64)parm_list[idx + 6]) << 8) +
1351                                 ((u64)parm_list[idx + 7]);
1352
1353                 ns->mode_select_block_len =
1354                                 (parm_list[idx + 12] << 24) +
1355                                 (parm_list[idx + 13] << 16) +
1356                                 (parm_list[idx + 14] << 8) +
1357                                 (parm_list[idx + 15]);
1358         }
1359 }
1360
1361 static int nvme_trans_modesel_get_mp(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1362                                         u8 *mode_page, u8 page_code)
1363 {
1364         int res = 0;
1365         int nvme_sc;
1366         struct nvme_dev *dev = ns->dev;
1367         unsigned dword11;
1368
1369         switch (page_code) {
1370         case MODE_PAGE_CACHING:
1371                 dword11 = ((mode_page[2] & CACHING_MODE_PAGE_WCE_MASK) ? 1 : 0);
1372                 nvme_sc = nvme_set_features(dev, NVME_FEAT_VOLATILE_WC, dword11,
1373                                             0, NULL);
1374                 res = nvme_trans_status_code(hdr, nvme_sc);
1375                 break;
1376         case MODE_PAGE_CONTROL:
1377                 break;
1378         case MODE_PAGE_POWER_CONDITION:
1379                 /* Verify the OS is not trying to set timers */
1380                 if ((mode_page[2] & 0x01) != 0 || (mode_page[3] & 0x0F) != 0) {
1381                         res = nvme_trans_completion(hdr,
1382                                                 SAM_STAT_CHECK_CONDITION,
1383                                                 ILLEGAL_REQUEST,
1384                                                 SCSI_ASC_INVALID_PARAMETER,
1385                                                 SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1386                         break;
1387                 }
1388                 break;
1389         default:
1390                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1391                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
1392                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1393                 break;
1394         }
1395
1396         return res;
1397 }
1398
1399 static int nvme_trans_modesel_data(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1400                                         u8 *cmd, u16 parm_list_len, u8 pf,
1401                                         u8 sp, u8 cdb10)
1402 {
1403         int res;
1404         u8 *parm_list;
1405         u16 bd_len;
1406         u8 llbaa = 0;
1407         u16 index, saved_index;
1408         u8 page_code;
1409         u16 mp_size;
1410
1411         /* Get parm list from data-in/out buffer */
1412         parm_list = kmalloc(parm_list_len, GFP_KERNEL);
1413         if (parm_list == NULL) {
1414                 res = -ENOMEM;
1415                 goto out;
1416         }
1417
1418         res = nvme_trans_copy_from_user(hdr, parm_list, parm_list_len);
1419         if (res)
1420                 goto out_mem;
1421
1422         nvme_trans_modesel_get_bd_len(parm_list, cdb10, &bd_len, &llbaa);
1423         index = (cdb10) ? (MODE_SELECT_10_MPH_SIZE) : (MODE_SELECT_6_MPH_SIZE);
1424
1425         if (bd_len != 0) {
1426                 /* Block Descriptors present, parse */
1427                 nvme_trans_modesel_save_bd(ns, parm_list, index, bd_len, llbaa);
1428                 index += bd_len;
1429         }
1430         saved_index = index;
1431
1432         /* Multiple mode pages may be present; iterate through all */
1433         /* In 1st Iteration, don't do NVME Command, only check for CDB errors */
1434         do {
1435                 page_code = parm_list[index] & MODE_SELECT_PAGE_CODE_MASK;
1436                 mp_size = parm_list[index + 1] + 2;
1437                 if ((page_code != MODE_PAGE_CACHING) &&
1438                     (page_code != MODE_PAGE_CONTROL) &&
1439                     (page_code != MODE_PAGE_POWER_CONDITION)) {
1440                         res = nvme_trans_completion(hdr,
1441                                                 SAM_STAT_CHECK_CONDITION,
1442                                                 ILLEGAL_REQUEST,
1443                                                 SCSI_ASC_INVALID_CDB,
1444                                                 SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1445                         goto out_mem;
1446                 }
1447                 index += mp_size;
1448         } while (index < parm_list_len);
1449
1450         /* In 2nd Iteration, do the NVME Commands */
1451         index = saved_index;
1452         do {
1453                 page_code = parm_list[index] & MODE_SELECT_PAGE_CODE_MASK;
1454                 mp_size = parm_list[index + 1] + 2;
1455                 res = nvme_trans_modesel_get_mp(ns, hdr, &parm_list[index],
1456                                                                 page_code);
1457                 if (res)
1458                         break;
1459                 index += mp_size;
1460         } while (index < parm_list_len);
1461
1462  out_mem:
1463         kfree(parm_list);
1464  out:
1465         return res;
1466 }
1467
1468 /* Format Unit Helper Functions */
1469
1470 static int nvme_trans_fmt_set_blk_size_count(struct nvme_ns *ns,
1471                                              struct sg_io_hdr *hdr)
1472 {
1473         int res = 0;
1474         int nvme_sc;
1475         struct nvme_dev *dev = ns->dev;
1476         u8 flbas;
1477
1478         /*
1479          * SCSI Expects a MODE SELECT would have been issued prior to
1480          * a FORMAT UNIT, and the block size and number would be used
1481          * from the block descriptor in it. If a MODE SELECT had not
1482          * been issued, FORMAT shall use the current values for both.
1483          */
1484
1485         if (ns->mode_select_num_blocks == 0 || ns->mode_select_block_len == 0) {
1486                 struct nvme_id_ns *id_ns;
1487
1488                 nvme_sc = nvme_identify_ns(dev, ns->ns_id, &id_ns);
1489                 res = nvme_trans_status_code(hdr, nvme_sc);
1490                 if (res)
1491                         return res;
1492
1493                 if (ns->mode_select_num_blocks == 0)
1494                         ns->mode_select_num_blocks = le64_to_cpu(id_ns->ncap);
1495                 if (ns->mode_select_block_len == 0) {
1496                         flbas = (id_ns->flbas) & 0x0F;
1497                         ns->mode_select_block_len =
1498                                                 (1 << (id_ns->lbaf[flbas].ds));
1499                 }
1500
1501                 kfree(id_ns);
1502         }
1503
1504         return 0;
1505 }
1506
1507 static int nvme_trans_fmt_get_parm_header(struct sg_io_hdr *hdr, u8 len,
1508                                         u8 format_prot_info, u8 *nvme_pf_code)
1509 {
1510         int res;
1511         u8 *parm_list;
1512         u8 pf_usage, pf_code;
1513
1514         parm_list = kmalloc(len, GFP_KERNEL);
1515         if (parm_list == NULL) {
1516                 res = -ENOMEM;
1517                 goto out;
1518         }
1519         res = nvme_trans_copy_from_user(hdr, parm_list, len);
1520         if (res)
1521                 goto out_mem;
1522
1523         if ((parm_list[FORMAT_UNIT_IMMED_OFFSET] &
1524                                 FORMAT_UNIT_IMMED_MASK) != 0) {
1525                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1526                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
1527                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1528                 goto out_mem;
1529         }
1530
1531         if (len == FORMAT_UNIT_LONG_PARM_LIST_LEN &&
1532             (parm_list[FORMAT_UNIT_PROT_INT_OFFSET] & 0x0F) != 0) {
1533                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1534                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
1535                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1536                 goto out_mem;
1537         }
1538         pf_usage = parm_list[FORMAT_UNIT_PROT_FIELD_USAGE_OFFSET] &
1539                         FORMAT_UNIT_PROT_FIELD_USAGE_MASK;
1540         pf_code = (pf_usage << 2) | format_prot_info;
1541         switch (pf_code) {
1542         case 0:
1543                 *nvme_pf_code = 0;
1544                 break;
1545         case 2:
1546                 *nvme_pf_code = 1;
1547                 break;
1548         case 3:
1549                 *nvme_pf_code = 2;
1550                 break;
1551         case 7:
1552                 *nvme_pf_code = 3;
1553                 break;
1554         default:
1555                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1556                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
1557                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1558                 break;
1559         }
1560
1561  out_mem:
1562         kfree(parm_list);
1563  out:
1564         return res;
1565 }
1566
1567 static int nvme_trans_fmt_send_cmd(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1568                                    u8 prot_info)
1569 {
1570         int res;
1571         int nvme_sc;
1572         struct nvme_dev *dev = ns->dev;
1573         struct nvme_id_ns *id_ns;
1574         u8 i;
1575         u8 flbas, nlbaf;
1576         u8 selected_lbaf = 0xFF;
1577         u32 cdw10 = 0;
1578         struct nvme_command c;
1579
1580         /* Loop thru LBAF's in id_ns to match reqd lbaf, put in cdw10 */
1581         nvme_sc = nvme_identify_ns(dev, ns->ns_id, &id_ns);
1582         res = nvme_trans_status_code(hdr, nvme_sc);
1583         if (res)
1584                 return res;
1585
1586         flbas = (id_ns->flbas) & 0x0F;
1587         nlbaf = id_ns->nlbaf;
1588
1589         for (i = 0; i < nlbaf; i++) {
1590                 if (ns->mode_select_block_len == (1 << (id_ns->lbaf[i].ds))) {
1591                         selected_lbaf = i;
1592                         break;
1593                 }
1594         }
1595         if (selected_lbaf > 0x0F) {
1596                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1597                                 ILLEGAL_REQUEST, SCSI_ASC_INVALID_PARAMETER,
1598                                 SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1599         }
1600         if (ns->mode_select_num_blocks != le64_to_cpu(id_ns->ncap)) {
1601                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1602                                 ILLEGAL_REQUEST, SCSI_ASC_INVALID_PARAMETER,
1603                                 SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1604         }
1605
1606         cdw10 |= prot_info << 5;
1607         cdw10 |= selected_lbaf & 0x0F;
1608         memset(&c, 0, sizeof(c));
1609         c.format.opcode = nvme_admin_format_nvm;
1610         c.format.nsid = cpu_to_le32(ns->ns_id);
1611         c.format.cdw10 = cpu_to_le32(cdw10);
1612
1613         nvme_sc = nvme_submit_sync_cmd(dev->admin_q, &c, NULL, 0);
1614         res = nvme_trans_status_code(hdr, nvme_sc);
1615
1616         kfree(id_ns);
1617         return res;
1618 }
1619
1620 static inline u32 nvme_trans_io_get_num_cmds(struct sg_io_hdr *hdr,
1621                                         struct nvme_trans_io_cdb *cdb_info,
1622                                         u32 max_blocks)
1623 {
1624         /* If using iovecs, send one nvme command per vector */
1625         if (hdr->iovec_count > 0)
1626                 return hdr->iovec_count;
1627         else if (cdb_info->xfer_len > max_blocks)
1628                 return ((cdb_info->xfer_len - 1) / max_blocks) + 1;
1629         else
1630                 return 1;
1631 }
1632
1633 static u16 nvme_trans_io_get_control(struct nvme_ns *ns,
1634                                         struct nvme_trans_io_cdb *cdb_info)
1635 {
1636         u16 control = 0;
1637
1638         /* When Protection information support is added, implement here */
1639
1640         if (cdb_info->fua > 0)
1641                 control |= NVME_RW_FUA;
1642
1643         return control;
1644 }
1645
1646 static int nvme_trans_do_nvme_io(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1647                                 struct nvme_trans_io_cdb *cdb_info, u8 is_write)
1648 {
1649         int nvme_sc = NVME_SC_SUCCESS;
1650         u32 num_cmds;
1651         u64 unit_len;
1652         u64 unit_num_blocks;    /* Number of blocks to xfer in each nvme cmd */
1653         u32 retcode;
1654         u32 i = 0;
1655         u64 nvme_offset = 0;
1656         void __user *next_mapping_addr;
1657         struct nvme_command c;
1658         u8 opcode = (is_write ? nvme_cmd_write : nvme_cmd_read);
1659         u16 control;
1660         u32 max_blocks = queue_max_hw_sectors(ns->queue);
1661
1662         num_cmds = nvme_trans_io_get_num_cmds(hdr, cdb_info, max_blocks);
1663
1664         /*
1665          * This loop handles two cases.
1666          * First, when an SGL is used in the form of an iovec list:
1667          *   - Use iov_base as the next mapping address for the nvme command_id
1668          *   - Use iov_len as the data transfer length for the command.
1669          * Second, when we have a single buffer
1670          *   - If larger than max_blocks, split into chunks, offset
1671          *        each nvme command accordingly.
1672          */
1673         for (i = 0; i < num_cmds; i++) {
1674                 memset(&c, 0, sizeof(c));
1675                 if (hdr->iovec_count > 0) {
1676                         struct sg_iovec sgl;
1677
1678                         retcode = copy_from_user(&sgl, hdr->dxferp +
1679                                         i * sizeof(struct sg_iovec),
1680                                         sizeof(struct sg_iovec));
1681                         if (retcode)
1682                                 return -EFAULT;
1683                         unit_len = sgl.iov_len;
1684                         unit_num_blocks = unit_len >> ns->lba_shift;
1685                         next_mapping_addr = sgl.iov_base;
1686                 } else {
1687                         unit_num_blocks = min((u64)max_blocks,
1688                                         (cdb_info->xfer_len - nvme_offset));
1689                         unit_len = unit_num_blocks << ns->lba_shift;
1690                         next_mapping_addr = hdr->dxferp +
1691                                         ((1 << ns->lba_shift) * nvme_offset);
1692                 }
1693
1694                 c.rw.opcode = opcode;
1695                 c.rw.nsid = cpu_to_le32(ns->ns_id);
1696                 c.rw.slba = cpu_to_le64(cdb_info->lba + nvme_offset);
1697                 c.rw.length = cpu_to_le16(unit_num_blocks - 1);
1698                 control = nvme_trans_io_get_control(ns, cdb_info);
1699                 c.rw.control = cpu_to_le16(control);
1700
1701                 if (get_capacity(ns->disk) - unit_num_blocks <
1702                                 cdb_info->lba + nvme_offset) {
1703                         nvme_sc = NVME_SC_LBA_RANGE;
1704                         break;
1705                 }
1706                 nvme_sc = __nvme_submit_sync_cmd(ns->queue, &c, NULL,
1707                                 next_mapping_addr, unit_len, NULL, 0);
1708                 if (nvme_sc)
1709                         break;
1710
1711                 nvme_offset += unit_num_blocks;
1712         }
1713
1714         return nvme_trans_status_code(hdr, nvme_sc);
1715 }
1716
1717
1718 /* SCSI Command Translation Functions */
1719
1720 static int nvme_trans_io(struct nvme_ns *ns, struct sg_io_hdr *hdr, u8 is_write,
1721                                                         u8 *cmd)
1722 {
1723         int res = 0;
1724         struct nvme_trans_io_cdb cdb_info = { 0, };
1725         u8 opcode = cmd[0];
1726         u64 xfer_bytes;
1727         u64 sum_iov_len = 0;
1728         struct sg_iovec sgl;
1729         int i;
1730         size_t not_copied;
1731
1732         /*
1733          * The FUA and WPROTECT fields are not supported in 6-byte CDBs,
1734          * but always in the same place for all others.
1735          */
1736         switch (opcode) {
1737         case WRITE_6:
1738         case READ_6:
1739                 break;
1740         default:
1741                 cdb_info.fua = cmd[1] & 0x8;
1742                 cdb_info.prot_info = (cmd[1] & 0xe0) >> 5;
1743                 if (cdb_info.prot_info && !ns->pi_type) {
1744                         return nvme_trans_completion(hdr,
1745                                         SAM_STAT_CHECK_CONDITION,
1746                                         ILLEGAL_REQUEST,
1747                                         SCSI_ASC_INVALID_CDB,
1748                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1749                 }
1750         }
1751
1752         switch (opcode) {
1753         case WRITE_6:
1754         case READ_6:
1755                 cdb_info.lba = get_unaligned_be24(&cmd[1]);
1756                 cdb_info.xfer_len = cmd[4];
1757                 if (cdb_info.xfer_len == 0)
1758                         cdb_info.xfer_len = 256;
1759                 break;
1760         case WRITE_10:
1761         case READ_10:
1762                 cdb_info.lba = get_unaligned_be32(&cmd[2]);
1763                 cdb_info.xfer_len = get_unaligned_be16(&cmd[7]);
1764                 break;
1765         case WRITE_12:
1766         case READ_12:
1767                 cdb_info.lba = get_unaligned_be32(&cmd[2]);
1768                 cdb_info.xfer_len = get_unaligned_be32(&cmd[6]);
1769                 break;
1770         case WRITE_16:
1771         case READ_16:
1772                 cdb_info.lba = get_unaligned_be64(&cmd[2]);
1773                 cdb_info.xfer_len = get_unaligned_be32(&cmd[10]);
1774                 break;
1775         default:
1776                 /* Will never really reach here */
1777                 res = -EIO;
1778                 goto out;
1779         }
1780
1781         /* Calculate total length of transfer (in bytes) */
1782         if (hdr->iovec_count > 0) {
1783                 for (i = 0; i < hdr->iovec_count; i++) {
1784                         not_copied = copy_from_user(&sgl, hdr->dxferp +
1785                                                 i * sizeof(struct sg_iovec),
1786                                                 sizeof(struct sg_iovec));
1787                         if (not_copied)
1788                                 return -EFAULT;
1789                         sum_iov_len += sgl.iov_len;
1790                         /* IO vector sizes should be multiples of block size */
1791                         if (sgl.iov_len % (1 << ns->lba_shift) != 0) {
1792                                 res = nvme_trans_completion(hdr,
1793                                                 SAM_STAT_CHECK_CONDITION,
1794                                                 ILLEGAL_REQUEST,
1795                                                 SCSI_ASC_INVALID_PARAMETER,
1796                                                 SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1797                                 goto out;
1798                         }
1799                 }
1800         } else {
1801                 sum_iov_len = hdr->dxfer_len;
1802         }
1803
1804         /* As Per sg ioctl howto, if the lengths differ, use the lower one */
1805         xfer_bytes = min(((u64)hdr->dxfer_len), sum_iov_len);
1806
1807         /* If block count and actual data buffer size dont match, error out */
1808         if (xfer_bytes != (cdb_info.xfer_len << ns->lba_shift)) {
1809                 res = -EINVAL;
1810                 goto out;
1811         }
1812
1813         /* Check for 0 length transfer - it is not illegal */
1814         if (cdb_info.xfer_len == 0)
1815                 goto out;
1816
1817         /* Send NVMe IO Command(s) */
1818         res = nvme_trans_do_nvme_io(ns, hdr, &cdb_info, is_write);
1819         if (res)
1820                 goto out;
1821
1822  out:
1823         return res;
1824 }
1825
1826 static int nvme_trans_inquiry(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1827                                                         u8 *cmd)
1828 {
1829         int res = 0;
1830         u8 evpd;
1831         u8 page_code;
1832         int alloc_len;
1833         u8 *inq_response;
1834
1835         evpd = cmd[1] & 0x01;
1836         page_code = cmd[2];
1837         alloc_len = get_unaligned_be16(&cmd[3]);
1838
1839         inq_response = kmalloc(alloc_len, GFP_KERNEL);
1840         if (inq_response == NULL) {
1841                 res = -ENOMEM;
1842                 goto out_mem;
1843         }
1844
1845         if (evpd == 0) {
1846                 if (page_code == INQ_STANDARD_INQUIRY_PAGE) {
1847                         res = nvme_trans_standard_inquiry_page(ns, hdr,
1848                                                 inq_response, alloc_len);
1849                 } else {
1850                         res = nvme_trans_completion(hdr,
1851                                                 SAM_STAT_CHECK_CONDITION,
1852                                                 ILLEGAL_REQUEST,
1853                                                 SCSI_ASC_INVALID_CDB,
1854                                                 SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1855                 }
1856         } else {
1857                 switch (page_code) {
1858                 case VPD_SUPPORTED_PAGES:
1859                         res = nvme_trans_supported_vpd_pages(ns, hdr,
1860                                                 inq_response, alloc_len);
1861                         break;
1862                 case VPD_SERIAL_NUMBER:
1863                         res = nvme_trans_unit_serial_page(ns, hdr, inq_response,
1864                                                                 alloc_len);
1865                         break;
1866                 case VPD_DEVICE_IDENTIFIERS:
1867                         res = nvme_trans_device_id_page(ns, hdr, inq_response,
1868                                                                 alloc_len);
1869                         break;
1870                 case VPD_EXTENDED_INQUIRY:
1871                         res = nvme_trans_ext_inq_page(ns, hdr, alloc_len);
1872                         break;
1873                 case VPD_BLOCK_LIMITS:
1874                         res = nvme_trans_bdev_limits_page(ns, hdr, inq_response,
1875                                                                 alloc_len);
1876                         break;
1877                 case VPD_BLOCK_DEV_CHARACTERISTICS:
1878                         res = nvme_trans_bdev_char_page(ns, hdr, alloc_len);
1879                         break;
1880                 default:
1881                         res = nvme_trans_completion(hdr,
1882                                                 SAM_STAT_CHECK_CONDITION,
1883                                                 ILLEGAL_REQUEST,
1884                                                 SCSI_ASC_INVALID_CDB,
1885                                                 SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1886                         break;
1887                 }
1888         }
1889         kfree(inq_response);
1890  out_mem:
1891         return res;
1892 }
1893
1894 static int nvme_trans_log_sense(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1895                                                         u8 *cmd)
1896 {
1897         int res;
1898         u16 alloc_len;
1899         u8 pc;
1900         u8 page_code;
1901
1902         if (cmd[1] != LOG_SENSE_CDB_SP_NOT_ENABLED) {
1903                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1904                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
1905                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1906                 goto out;
1907         }
1908
1909         page_code = cmd[2] & LOG_SENSE_CDB_PAGE_CODE_MASK;
1910         pc = (cmd[2] & LOG_SENSE_CDB_PC_MASK) >> LOG_SENSE_CDB_PC_SHIFT;
1911         if (pc != LOG_SENSE_CDB_PC_CUMULATIVE_VALUES) {
1912                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1913                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
1914                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1915                 goto out;
1916         }
1917         alloc_len = get_unaligned_be16(&cmd[7]);
1918         switch (page_code) {
1919         case LOG_PAGE_SUPPORTED_LOG_PAGES_PAGE:
1920                 res = nvme_trans_log_supp_pages(ns, hdr, alloc_len);
1921                 break;
1922         case LOG_PAGE_INFORMATIONAL_EXCEPTIONS_PAGE:
1923                 res = nvme_trans_log_info_exceptions(ns, hdr, alloc_len);
1924                 break;
1925         case LOG_PAGE_TEMPERATURE_PAGE:
1926                 res = nvme_trans_log_temperature(ns, hdr, alloc_len);
1927                 break;
1928         default:
1929                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1930                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
1931                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1932                 break;
1933         }
1934
1935  out:
1936         return res;
1937 }
1938
1939 static int nvme_trans_mode_select(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1940                                                         u8 *cmd)
1941 {
1942         u8 cdb10 = 0;
1943         u16 parm_list_len;
1944         u8 page_format;
1945         u8 save_pages;
1946
1947         page_format = cmd[1] & MODE_SELECT_CDB_PAGE_FORMAT_MASK;
1948         save_pages = cmd[1] & MODE_SELECT_CDB_SAVE_PAGES_MASK;
1949
1950         if (cmd[0] == MODE_SELECT) {
1951                 parm_list_len = cmd[4];
1952         } else {
1953                 parm_list_len = cmd[7];
1954                 cdb10 = 1;
1955         }
1956
1957         if (parm_list_len != 0) {
1958                 /*
1959                  * According to SPC-4 r24, a paramter list length field of 0
1960                  * shall not be considered an error
1961                  */
1962                 return nvme_trans_modesel_data(ns, hdr, cmd, parm_list_len,
1963                                                 page_format, save_pages, cdb10);
1964         }
1965
1966         return 0;
1967 }
1968
1969 static int nvme_trans_mode_sense(struct nvme_ns *ns, struct sg_io_hdr *hdr,
1970                                                         u8 *cmd)
1971 {
1972         int res = 0;
1973         u16 alloc_len;
1974         u8 cdb10 = 0;
1975
1976         if (cmd[0] == MODE_SENSE) {
1977                 alloc_len = cmd[4];
1978         } else {
1979                 alloc_len = get_unaligned_be16(&cmd[7]);
1980                 cdb10 = 1;
1981         }
1982
1983         if ((cmd[2] & MODE_SENSE_PAGE_CONTROL_MASK) !=
1984                         MODE_SENSE_PC_CURRENT_VALUES) {
1985                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
1986                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
1987                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
1988                 goto out;
1989         }
1990
1991         switch (cmd[2] & MODE_SENSE_PAGE_CODE_MASK) {
1992         case MODE_PAGE_CACHING:
1993                 res = nvme_trans_mode_page_create(ns, hdr, cmd, alloc_len,
1994                                                 cdb10,
1995                                                 &nvme_trans_fill_caching_page,
1996                                                 MODE_PAGE_CACHING_LEN);
1997                 break;
1998         case MODE_PAGE_CONTROL:
1999                 res = nvme_trans_mode_page_create(ns, hdr, cmd, alloc_len,
2000                                                 cdb10,
2001                                                 &nvme_trans_fill_control_page,
2002                                                 MODE_PAGE_CONTROL_LEN);
2003                 break;
2004         case MODE_PAGE_POWER_CONDITION:
2005                 res = nvme_trans_mode_page_create(ns, hdr, cmd, alloc_len,
2006                                                 cdb10,
2007                                                 &nvme_trans_fill_pow_cnd_page,
2008                                                 MODE_PAGE_POW_CND_LEN);
2009                 break;
2010         case MODE_PAGE_INFO_EXCEP:
2011                 res = nvme_trans_mode_page_create(ns, hdr, cmd, alloc_len,
2012                                                 cdb10,
2013                                                 &nvme_trans_fill_inf_exc_page,
2014                                                 MODE_PAGE_INF_EXC_LEN);
2015                 break;
2016         case MODE_PAGE_RETURN_ALL:
2017                 res = nvme_trans_mode_page_create(ns, hdr, cmd, alloc_len,
2018                                                 cdb10,
2019                                                 &nvme_trans_fill_all_pages,
2020                                                 MODE_PAGE_ALL_LEN);
2021                 break;
2022         default:
2023                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
2024                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
2025                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2026                 break;
2027         }
2028
2029  out:
2030         return res;
2031 }
2032
2033 static int nvme_trans_read_capacity(struct nvme_ns *ns, struct sg_io_hdr *hdr,
2034                                                         u8 *cmd, u8 cdb16)
2035 {
2036         int res;
2037         int nvme_sc;
2038         u32 alloc_len;
2039         u32 resp_size;
2040         u32 xfer_len;
2041         struct nvme_dev *dev = ns->dev;
2042         struct nvme_id_ns *id_ns;
2043         u8 *response;
2044
2045         if (cdb16) {
2046                 alloc_len = get_unaligned_be32(&cmd[10]);
2047                 resp_size = READ_CAP_16_RESP_SIZE;
2048         } else {
2049                 alloc_len = READ_CAP_10_RESP_SIZE;
2050                 resp_size = READ_CAP_10_RESP_SIZE;
2051         }
2052
2053         nvme_sc = nvme_identify_ns(dev, ns->ns_id, &id_ns);
2054         res = nvme_trans_status_code(hdr, nvme_sc);
2055         if (res)
2056                 return res;
2057
2058         response = kzalloc(resp_size, GFP_KERNEL);
2059         if (response == NULL) {
2060                 res = -ENOMEM;
2061                 goto out_free_id;
2062         }
2063         nvme_trans_fill_read_cap(response, id_ns, cdb16);
2064
2065         xfer_len = min(alloc_len, resp_size);
2066         res = nvme_trans_copy_to_user(hdr, response, xfer_len);
2067
2068         kfree(response);
2069  out_free_id:
2070         kfree(id_ns);
2071         return res;
2072 }
2073
2074 static int nvme_trans_report_luns(struct nvme_ns *ns, struct sg_io_hdr *hdr,
2075                                                         u8 *cmd)
2076 {
2077         int res;
2078         int nvme_sc;
2079         u32 alloc_len, xfer_len, resp_size;
2080         u8 *response;
2081         struct nvme_dev *dev = ns->dev;
2082         struct nvme_id_ctrl *id_ctrl;
2083         u32 ll_length, lun_id;
2084         u8 lun_id_offset = REPORT_LUNS_FIRST_LUN_OFFSET;
2085         __be32 tmp_len;
2086
2087         switch (cmd[2]) {
2088         default:
2089                 return nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
2090                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
2091                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2092         case ALL_LUNS_RETURNED:
2093         case ALL_WELL_KNOWN_LUNS_RETURNED:
2094         case RESTRICTED_LUNS_RETURNED:
2095                 nvme_sc = nvme_identify_ctrl(dev, &id_ctrl);
2096                 res = nvme_trans_status_code(hdr, nvme_sc);
2097                 if (res)
2098                         return res;
2099
2100                 ll_length = le32_to_cpu(id_ctrl->nn) * LUN_ENTRY_SIZE;
2101                 resp_size = ll_length + LUN_DATA_HEADER_SIZE;
2102
2103                 alloc_len = get_unaligned_be32(&cmd[6]);
2104                 if (alloc_len < resp_size) {
2105                         res = nvme_trans_completion(hdr,
2106                                         SAM_STAT_CHECK_CONDITION,
2107                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
2108                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2109                         goto out_free_id;
2110                 }
2111
2112                 response = kzalloc(resp_size, GFP_KERNEL);
2113                 if (response == NULL) {
2114                         res = -ENOMEM;
2115                         goto out_free_id;
2116                 }
2117
2118                 /* The first LUN ID will always be 0 per the SAM spec */
2119                 for (lun_id = 0; lun_id < le32_to_cpu(id_ctrl->nn); lun_id++) {
2120                         /*
2121                          * Set the LUN Id and then increment to the next LUN
2122                          * location in the parameter data.
2123                          */
2124                         __be64 tmp_id = cpu_to_be64(lun_id);
2125                         memcpy(&response[lun_id_offset], &tmp_id, sizeof(u64));
2126                         lun_id_offset += LUN_ENTRY_SIZE;
2127                 }
2128                 tmp_len = cpu_to_be32(ll_length);
2129                 memcpy(response, &tmp_len, sizeof(u32));
2130         }
2131
2132         xfer_len = min(alloc_len, resp_size);
2133         res = nvme_trans_copy_to_user(hdr, response, xfer_len);
2134
2135         kfree(response);
2136  out_free_id:
2137         kfree(id_ctrl);
2138         return res;
2139 }
2140
2141 static int nvme_trans_request_sense(struct nvme_ns *ns, struct sg_io_hdr *hdr,
2142                                                         u8 *cmd)
2143 {
2144         int res;
2145         u8 alloc_len, xfer_len, resp_size;
2146         u8 desc_format;
2147         u8 *response;
2148
2149         desc_format = cmd[1] & 0x01;
2150         alloc_len = cmd[4];
2151
2152         resp_size = ((desc_format) ? (DESC_FMT_SENSE_DATA_SIZE) :
2153                                         (FIXED_FMT_SENSE_DATA_SIZE));
2154         response = kzalloc(resp_size, GFP_KERNEL);
2155         if (response == NULL) {
2156                 res = -ENOMEM;
2157                 goto out;
2158         }
2159
2160         if (desc_format) {
2161                 /* Descriptor Format Sense Data */
2162                 response[0] = DESC_FORMAT_SENSE_DATA;
2163                 response[1] = NO_SENSE;
2164                 /* TODO How is LOW POWER CONDITION ON handled? (byte 2) */
2165                 response[2] = SCSI_ASC_NO_SENSE;
2166                 response[3] = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
2167                 /* SDAT_OVFL = 0 | Additional Sense Length = 0 */
2168         } else {
2169                 /* Fixed Format Sense Data */
2170                 response[0] = FIXED_SENSE_DATA;
2171                 /* Byte 1 = Obsolete */
2172                 response[2] = NO_SENSE; /* FM, EOM, ILI, SDAT_OVFL = 0 */
2173                 /* Bytes 3-6 - Information - set to zero */
2174                 response[7] = FIXED_SENSE_DATA_ADD_LENGTH;
2175                 /* Bytes 8-11 - Cmd Specific Information - set to zero */
2176                 response[12] = SCSI_ASC_NO_SENSE;
2177                 response[13] = SCSI_ASCQ_CAUSE_NOT_REPORTABLE;
2178                 /* Byte 14 = Field Replaceable Unit Code = 0 */
2179                 /* Bytes 15-17 - SKSV=0; Sense Key Specific = 0 */
2180         }
2181
2182         xfer_len = min(alloc_len, resp_size);
2183         res = nvme_trans_copy_to_user(hdr, response, xfer_len);
2184
2185         kfree(response);
2186  out:
2187         return res;
2188 }
2189
2190 static int nvme_trans_security_protocol(struct nvme_ns *ns,
2191                                         struct sg_io_hdr *hdr,
2192                                         u8 *cmd)
2193 {
2194         return nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
2195                                 ILLEGAL_REQUEST, SCSI_ASC_ILLEGAL_COMMAND,
2196                                 SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2197 }
2198
2199 static int nvme_trans_synchronize_cache(struct nvme_ns *ns,
2200                                         struct sg_io_hdr *hdr)
2201 {
2202         int nvme_sc;
2203         struct nvme_command c;
2204
2205         memset(&c, 0, sizeof(c));
2206         c.common.opcode = nvme_cmd_flush;
2207         c.common.nsid = cpu_to_le32(ns->ns_id);
2208
2209         nvme_sc = nvme_submit_sync_cmd(ns->queue, &c, NULL, 0);
2210         return nvme_trans_status_code(hdr, nvme_sc);
2211 }
2212
2213 static int nvme_trans_start_stop(struct nvme_ns *ns, struct sg_io_hdr *hdr,
2214                                                         u8 *cmd)
2215 {
2216         u8 immed, pcmod, pc, no_flush, start;
2217
2218         immed = cmd[1] & 0x01;
2219         pcmod = cmd[3] & 0x0f;
2220         pc = (cmd[4] & 0xf0) >> 4;
2221         no_flush = cmd[4] & 0x04;
2222         start = cmd[4] & 0x01;
2223
2224         if (immed != 0) {
2225                 return nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
2226                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
2227                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2228         } else {
2229                 if (no_flush == 0) {
2230                         /* Issue NVME FLUSH command prior to START STOP UNIT */
2231                         int res = nvme_trans_synchronize_cache(ns, hdr);
2232                         if (res)
2233                                 return res;
2234                 }
2235                 /* Setup the expected power state transition */
2236                 return nvme_trans_power_state(ns, hdr, pc, pcmod, start);
2237         }
2238 }
2239
2240 static int nvme_trans_format_unit(struct nvme_ns *ns, struct sg_io_hdr *hdr,
2241                                                         u8 *cmd)
2242 {
2243         int res;
2244         u8 parm_hdr_len = 0;
2245         u8 nvme_pf_code = 0;
2246         u8 format_prot_info, long_list, format_data;
2247
2248         format_prot_info = (cmd[1] & 0xc0) >> 6;
2249         long_list = cmd[1] & 0x20;
2250         format_data = cmd[1] & 0x10;
2251
2252         if (format_data != 0) {
2253                 if (format_prot_info != 0) {
2254                         if (long_list == 0)
2255                                 parm_hdr_len = FORMAT_UNIT_SHORT_PARM_LIST_LEN;
2256                         else
2257                                 parm_hdr_len = FORMAT_UNIT_LONG_PARM_LIST_LEN;
2258                 }
2259         } else if (format_data == 0 && format_prot_info != 0) {
2260                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
2261                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
2262                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2263                 goto out;
2264         }
2265
2266         /* Get parm header from data-in/out buffer */
2267         /*
2268          * According to the translation spec, the only fields in the parameter
2269          * list we are concerned with are in the header. So allocate only that.
2270          */
2271         if (parm_hdr_len > 0) {
2272                 res = nvme_trans_fmt_get_parm_header(hdr, parm_hdr_len,
2273                                         format_prot_info, &nvme_pf_code);
2274                 if (res)
2275                         goto out;
2276         }
2277
2278         /* Attempt to activate any previously downloaded firmware image */
2279         res = nvme_trans_send_activate_fw_cmd(ns, hdr, 0);
2280
2281         /* Determine Block size and count and send format command */
2282         res = nvme_trans_fmt_set_blk_size_count(ns, hdr);
2283         if (res)
2284                 goto out;
2285
2286         res = nvme_trans_fmt_send_cmd(ns, hdr, nvme_pf_code);
2287
2288  out:
2289         return res;
2290 }
2291
2292 static int nvme_trans_test_unit_ready(struct nvme_ns *ns,
2293                                         struct sg_io_hdr *hdr,
2294                                         u8 *cmd)
2295 {
2296         struct nvme_dev *dev = ns->dev;
2297
2298         if (!(readl(&dev->bar->csts) & NVME_CSTS_RDY))
2299                 return nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
2300                                             NOT_READY, SCSI_ASC_LUN_NOT_READY,
2301                                             SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2302         else
2303                 return nvme_trans_completion(hdr, SAM_STAT_GOOD, NO_SENSE, 0, 0);
2304 }
2305
2306 static int nvme_trans_write_buffer(struct nvme_ns *ns, struct sg_io_hdr *hdr,
2307                                                         u8 *cmd)
2308 {
2309         int res = 0;
2310         u32 buffer_offset, parm_list_length;
2311         u8 buffer_id, mode;
2312
2313         parm_list_length = get_unaligned_be24(&cmd[6]);
2314         if (parm_list_length % BYTES_TO_DWORDS != 0) {
2315                 /* NVMe expects Firmware file to be a whole number of DWORDS */
2316                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
2317                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
2318                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2319                 goto out;
2320         }
2321         buffer_id = cmd[2];
2322         if (buffer_id > NVME_MAX_FIRMWARE_SLOT) {
2323                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
2324                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
2325                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2326                 goto out;
2327         }
2328         mode = cmd[1] & 0x1f;
2329         buffer_offset = get_unaligned_be24(&cmd[3]);
2330
2331         switch (mode) {
2332         case DOWNLOAD_SAVE_ACTIVATE:
2333                 res = nvme_trans_send_download_fw_cmd(ns, hdr, nvme_admin_download_fw,
2334                                                 parm_list_length, buffer_offset,
2335                                                 buffer_id);
2336                 if (res)
2337                         goto out;
2338                 res = nvme_trans_send_activate_fw_cmd(ns, hdr, buffer_id);
2339                 break;
2340         case DOWNLOAD_SAVE_DEFER_ACTIVATE:
2341                 res = nvme_trans_send_download_fw_cmd(ns, hdr, nvme_admin_download_fw,
2342                                                 parm_list_length, buffer_offset,
2343                                                 buffer_id);
2344                 break;
2345         case ACTIVATE_DEFERRED_MICROCODE:
2346                 res = nvme_trans_send_activate_fw_cmd(ns, hdr, buffer_id);
2347                 break;
2348         default:
2349                 res = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
2350                                         ILLEGAL_REQUEST, SCSI_ASC_INVALID_CDB,
2351                                         SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2352                 break;
2353         }
2354
2355  out:
2356         return res;
2357 }
2358
2359 struct scsi_unmap_blk_desc {
2360         __be64  slba;
2361         __be32  nlb;
2362         u32     resv;
2363 };
2364
2365 struct scsi_unmap_parm_list {
2366         __be16  unmap_data_len;
2367         __be16  unmap_blk_desc_data_len;
2368         u32     resv;
2369         struct scsi_unmap_blk_desc desc[0];
2370 };
2371
2372 static int nvme_trans_unmap(struct nvme_ns *ns, struct sg_io_hdr *hdr,
2373                                                         u8 *cmd)
2374 {
2375         struct scsi_unmap_parm_list *plist;
2376         struct nvme_dsm_range *range;
2377         struct nvme_command c;
2378         int i, nvme_sc, res = -ENOMEM;
2379         u16 ndesc, list_len;
2380
2381         list_len = get_unaligned_be16(&cmd[7]);
2382         if (!list_len)
2383                 return -EINVAL;
2384
2385         plist = kmalloc(list_len, GFP_KERNEL);
2386         if (!plist)
2387                 return -ENOMEM;
2388
2389         res = nvme_trans_copy_from_user(hdr, plist, list_len);
2390         if (res)
2391                 goto out;
2392
2393         ndesc = be16_to_cpu(plist->unmap_blk_desc_data_len) >> 4;
2394         if (!ndesc || ndesc > 256) {
2395                 res = -EINVAL;
2396                 goto out;
2397         }
2398
2399         range = kcalloc(ndesc, sizeof(*range), GFP_KERNEL);
2400         if (!range)
2401                 goto out;
2402
2403         for (i = 0; i < ndesc; i++) {
2404                 range[i].nlb = cpu_to_le32(be32_to_cpu(plist->desc[i].nlb));
2405                 range[i].slba = cpu_to_le64(be64_to_cpu(plist->desc[i].slba));
2406                 range[i].cattr = 0;
2407         }
2408
2409         memset(&c, 0, sizeof(c));
2410         c.dsm.opcode = nvme_cmd_dsm;
2411         c.dsm.nsid = cpu_to_le32(ns->ns_id);
2412         c.dsm.nr = cpu_to_le32(ndesc - 1);
2413         c.dsm.attributes = cpu_to_le32(NVME_DSMGMT_AD);
2414
2415         nvme_sc = nvme_submit_sync_cmd(ns->queue, &c, range,
2416                         ndesc * sizeof(*range));
2417         res = nvme_trans_status_code(hdr, nvme_sc);
2418
2419         kfree(range);
2420  out:
2421         kfree(plist);
2422         return res;
2423 }
2424
2425 static int nvme_scsi_translate(struct nvme_ns *ns, struct sg_io_hdr *hdr)
2426 {
2427         u8 cmd[BLK_MAX_CDB];
2428         int retcode;
2429         unsigned int opcode;
2430
2431         if (hdr->cmdp == NULL)
2432                 return -EMSGSIZE;
2433         if (copy_from_user(cmd, hdr->cmdp, hdr->cmd_len))
2434                 return -EFAULT;
2435
2436         /*
2437          * Prime the hdr with good status for scsi commands that don't require
2438          * an nvme command for translation.
2439          */
2440         retcode = nvme_trans_status_code(hdr, NVME_SC_SUCCESS);
2441         if (retcode)
2442                 return retcode;
2443
2444         opcode = cmd[0];
2445
2446         switch (opcode) {
2447         case READ_6:
2448         case READ_10:
2449         case READ_12:
2450         case READ_16:
2451                 retcode = nvme_trans_io(ns, hdr, 0, cmd);
2452                 break;
2453         case WRITE_6:
2454         case WRITE_10:
2455         case WRITE_12:
2456         case WRITE_16:
2457                 retcode = nvme_trans_io(ns, hdr, 1, cmd);
2458                 break;
2459         case INQUIRY:
2460                 retcode = nvme_trans_inquiry(ns, hdr, cmd);
2461                 break;
2462         case LOG_SENSE:
2463                 retcode = nvme_trans_log_sense(ns, hdr, cmd);
2464                 break;
2465         case MODE_SELECT:
2466         case MODE_SELECT_10:
2467                 retcode = nvme_trans_mode_select(ns, hdr, cmd);
2468                 break;
2469         case MODE_SENSE:
2470         case MODE_SENSE_10:
2471                 retcode = nvme_trans_mode_sense(ns, hdr, cmd);
2472                 break;
2473         case READ_CAPACITY:
2474                 retcode = nvme_trans_read_capacity(ns, hdr, cmd, 0);
2475                 break;
2476         case SERVICE_ACTION_IN_16:
2477                 switch (cmd[1]) {
2478                 case SAI_READ_CAPACITY_16:
2479                         retcode = nvme_trans_read_capacity(ns, hdr, cmd, 1);
2480                         break;
2481                 default:
2482                         goto out;
2483                 }
2484                 break;
2485         case REPORT_LUNS:
2486                 retcode = nvme_trans_report_luns(ns, hdr, cmd);
2487                 break;
2488         case REQUEST_SENSE:
2489                 retcode = nvme_trans_request_sense(ns, hdr, cmd);
2490                 break;
2491         case SECURITY_PROTOCOL_IN:
2492         case SECURITY_PROTOCOL_OUT:
2493                 retcode = nvme_trans_security_protocol(ns, hdr, cmd);
2494                 break;
2495         case START_STOP:
2496                 retcode = nvme_trans_start_stop(ns, hdr, cmd);
2497                 break;
2498         case SYNCHRONIZE_CACHE:
2499                 retcode = nvme_trans_synchronize_cache(ns, hdr);
2500                 break;
2501         case FORMAT_UNIT:
2502                 retcode = nvme_trans_format_unit(ns, hdr, cmd);
2503                 break;
2504         case TEST_UNIT_READY:
2505                 retcode = nvme_trans_test_unit_ready(ns, hdr, cmd);
2506                 break;
2507         case WRITE_BUFFER:
2508                 retcode = nvme_trans_write_buffer(ns, hdr, cmd);
2509                 break;
2510         case UNMAP:
2511                 retcode = nvme_trans_unmap(ns, hdr, cmd);
2512                 break;
2513         default:
2514  out:
2515                 retcode = nvme_trans_completion(hdr, SAM_STAT_CHECK_CONDITION,
2516                                 ILLEGAL_REQUEST, SCSI_ASC_ILLEGAL_COMMAND,
2517                                 SCSI_ASCQ_CAUSE_NOT_REPORTABLE);
2518                 break;
2519         }
2520         return retcode;
2521 }
2522
2523 int nvme_sg_io(struct nvme_ns *ns, struct sg_io_hdr __user *u_hdr)
2524 {
2525         struct sg_io_hdr hdr;
2526         int retcode;
2527
2528         if (!capable(CAP_SYS_ADMIN))
2529                 return -EACCES;
2530         if (copy_from_user(&hdr, u_hdr, sizeof(hdr)))
2531                 return -EFAULT;
2532         if (hdr.interface_id != 'S')
2533                 return -EINVAL;
2534         if (hdr.cmd_len > BLK_MAX_CDB)
2535                 return -EINVAL;
2536
2537         /*
2538          * A positive return code means a NVMe status, which has been
2539          * translated to sense data.
2540          */
2541         retcode = nvme_scsi_translate(ns, &hdr);
2542         if (retcode < 0)
2543                 return retcode;
2544         if (copy_to_user(u_hdr, &hdr, sizeof(sg_io_hdr_t)) > 0)
2545                 return -EFAULT;
2546         return 0;
2547 }
2548
2549 int nvme_sg_get_version_num(int __user *ip)
2550 {
2551         return put_user(sg_version_num, ip);
2552 }