sys/cam/ctl/ctl_backend_block.c

   1 /*-
   2  * Copyright (c) 2003 Silicon Graphics International Corp.
   3  * Copyright (c) 2009-2011 Spectra Logic Corporation
   4  * Copyright (c) 2012 The FreeBSD Foundation
   5  * All rights reserved.
   6  *
   7  * Portions of this software were developed by Edward Tomasz Napierala
   8  * under sponsorship from the FreeBSD Foundation.
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions
  12  * are met:
  13  * 1. Redistributions of source code must retain the above copyright
  14  *    notice, this list of conditions, and the following disclaimer,
  15  *    without modification.
  16  * 2. Redistributions in binary form must reproduce at minimum a disclaimer
  17  *    substantially similar to the "NO WARRANTY" disclaimer below
  18  *    ("Disclaimer") and any redistribution must be conditioned upon
  19  *    including a substantially similar Disclaimer requirement for further
  20  *    binary redistribution.
  21  *
  22  * NO WARRANTY
  23  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  24  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  25  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTIBILITY AND FITNESS FOR
  26  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  27  * HOLDERS OR CONTRIBUTORS BE LIABLE FOR SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  31  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
  32  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
  33  * POSSIBILITY OF SUCH DAMAGES.
  34  *
  35  * $Id: //depot/users/kenm/FreeBSD-test2/sys/cam/ctl/ctl_backend_block.c#5 $
  36  */
  37 /*
  38  * CAM Target Layer driver backend for block devices.
  39  *
  40  * Author: Ken Merry <ken@FreeBSD.org>
  41  */
  42 #include <sys/cdefs.h>
  43 __FBSDID("$FreeBSD$");
  44
  45 #include <opt_kdtrace.h>
  46
  47 #include <sys/param.h>
  48 #include <sys/systm.h>
  49 #include <sys/kernel.h>
  50 #include <sys/types.h>
  51 #include <sys/kthread.h>
  52 #include <sys/bio.h>
  53 #include <sys/fcntl.h>
  54 #include <sys/lock.h>
  55 #include <sys/mutex.h>
  56 #include <sys/condvar.h>
  57 #include <sys/malloc.h>
  58 #include <sys/conf.h>
  59 #include <sys/ioccom.h>
  60 #include <sys/queue.h>
  61 #include <sys/sbuf.h>
  62 #include <sys/endian.h>
  63 #include <sys/uio.h>
  64 #include <sys/buf.h>
  65 #include <sys/taskqueue.h>
  66 #include <sys/vnode.h>
  67 #include <sys/namei.h>
  68 #include <sys/mount.h>
  69 #include <sys/disk.h>
  70 #include <sys/fcntl.h>
  71 #include <sys/filedesc.h>
  72 #include <sys/proc.h>
  73 #include <sys/pcpu.h>
  74 #include <sys/module.h>
  75 #include <sys/sdt.h>
  76 #include <sys/devicestat.h>
  77 #include <sys/sysctl.h>
  78
  79 #include <geom/geom.h>
  80
  81 #include <cam/cam.h>
  82 #include <cam/scsi/scsi_all.h>
  83 #include <cam/scsi/scsi_da.h>
  84 #include <cam/ctl/ctl_io.h>
  85 #include <cam/ctl/ctl.h>
  86 #include <cam/ctl/ctl_backend.h>
  87 #include <cam/ctl/ctl_frontend_internal.h>
  88 #include <cam/ctl/ctl_ioctl.h>
  89 #include <cam/ctl/ctl_scsi_all.h>
  90 #include <cam/ctl/ctl_error.h>
  91
  92 /*
  93  * The idea here is that we'll allocate enough S/G space to hold a 16MB
  94  * I/O.  If we get an I/O larger than that, we'll reject it.
  95  */
  96 #define CTLBLK_MAX_IO_SIZE      (16 * 1024 * 1024)
  97 #define CTLBLK_MAX_SEGS         (CTLBLK_MAX_IO_SIZE / MAXPHYS) + 1
  98
  99 #ifdef CTLBLK_DEBUG
 100 #define DPRINTF(fmt, args...) \
 101     printf("cbb(%s:%d): " fmt, __FUNCTION__, __LINE__, ##args)
 102 #else
 103 #define DPRINTF(fmt, args...) do {} while(0)
 104 #endif
 105
 106 SDT_PROVIDER_DEFINE(cbb);
 107
 108 typedef enum {
 109         CTL_BE_BLOCK_LUN_UNCONFIGURED   = 0x01,
 110         CTL_BE_BLOCK_LUN_CONFIG_ERR     = 0x02,
 111         CTL_BE_BLOCK_LUN_WAITING        = 0x04,
 112         CTL_BE_BLOCK_LUN_MULTI_THREAD   = 0x08
 113 } ctl_be_block_lun_flags;
 114
 115 typedef enum {
 116         CTL_BE_BLOCK_NONE,
 117         CTL_BE_BLOCK_DEV,
 118         CTL_BE_BLOCK_FILE
 119 } ctl_be_block_type;
 120
 121 struct ctl_be_block_devdata {
 122         struct cdev *cdev;
 123         struct cdevsw *csw;
 124         int dev_ref;
 125 };
 126
 127 struct ctl_be_block_filedata {
 128         struct ucred *cred;
 129 };
 130
 131 union ctl_be_block_bedata {
 132         struct ctl_be_block_devdata dev;
 133         struct ctl_be_block_filedata file;
 134 };
 135
 136 struct ctl_be_block_io;
 137 struct ctl_be_block_lun;
 138
 139 typedef void (*cbb_dispatch_t)(struct ctl_be_block_lun *be_lun,
 140                                struct ctl_be_block_io *beio);
 141
 142 /*
 143  * Backend LUN structure.  There is a 1:1 mapping between a block device
 144  * and a backend block LUN, and between a backend block LUN and a CTL LUN.
 145  */
 146 struct ctl_be_block_lun {
 147         struct ctl_block_disk *disk;
 148         char lunname[32];
 149         char *dev_path;
 150         ctl_be_block_type dev_type;
 151         struct vnode *vn;
 152         union ctl_be_block_bedata backend;
 153         cbb_dispatch_t dispatch;
 154         cbb_dispatch_t lun_flush;
 155         struct mtx lock;
 156         uma_zone_t lun_zone;
 157         uint64_t size_blocks;
 158         uint64_t size_bytes;
 159         uint32_t blocksize;
 160         int blocksize_shift;
 161         struct ctl_be_block_softc *softc;
 162         struct devstat *disk_stats;
 163         ctl_be_block_lun_flags flags;
 164         STAILQ_ENTRY(ctl_be_block_lun) links;
 165         struct ctl_be_lun ctl_be_lun;
 166         struct taskqueue *io_taskqueue;
 167         struct task io_task;
 168         int num_threads;
 169         STAILQ_HEAD(, ctl_io_hdr) input_queue;
 170         STAILQ_HEAD(, ctl_io_hdr) config_write_queue;
 171         STAILQ_HEAD(, ctl_io_hdr) datamove_queue;
 172 };
 173
 174 /*
 175  * Overall softc structure for the block backend module.
 176  */
 177 struct ctl_be_block_softc {
 178         STAILQ_HEAD(, ctl_be_block_io)   beio_free_queue;
 179         struct mtx                       lock;
 180         int                              prealloc_beio;
 181         int                              num_disks;
 182         STAILQ_HEAD(, ctl_block_disk)    disk_list;
 183         int                              num_luns;
 184         STAILQ_HEAD(, ctl_be_block_lun)  lun_list;
 185 };
 186
 187 static struct ctl_be_block_softc backend_block_softc;
 188
 189 /*
 190  * Per-I/O information.
 191  */
 192 struct ctl_be_block_io {
 193         union ctl_io                    *io;
 194         struct ctl_sg_entry             sg_segs[CTLBLK_MAX_SEGS];
 195         struct iovec                    xiovecs[CTLBLK_MAX_SEGS];
 196         int                             bio_cmd;
 197         int                             bio_flags;
 198         int                             num_segs;
 199         int                             num_bios_sent;
 200         int                             num_bios_done;
 201         int                             send_complete;
 202         int                             num_errors;
 203         struct bintime                  ds_t0;
 204         devstat_tag_type                ds_tag_type;
 205         devstat_trans_flags             ds_trans_type;
 206         uint64_t                        io_len;
 207         uint64_t                        io_offset;
 208         struct ctl_be_block_softc       *softc;
 209         struct ctl_be_block_lun         *lun;
 210         STAILQ_ENTRY(ctl_be_block_io)   links;
 211 };
 212
 213 static int cbb_num_threads = 14;
 214 TUNABLE_INT("kern.cam.ctl.block.num_threads", &cbb_num_threads);
 215 SYSCTL_NODE(_kern_cam_ctl, OID_AUTO, block, CTLFLAG_RD, 0,
 216             "CAM Target Layer Block Backend");
 217 SYSCTL_INT(_kern_cam_ctl_block, OID_AUTO, num_threads, CTLFLAG_RW,
 218            &cbb_num_threads, 0, "Number of threads per backing file");
 219
 220 static struct ctl_be_block_io *ctl_alloc_beio(struct ctl_be_block_softc *softc);
 221 static void ctl_free_beio(struct ctl_be_block_io *beio);
 222 static int ctl_grow_beio(struct ctl_be_block_softc *softc, int count);
 223 #if 0
 224 static void ctl_shrink_beio(struct ctl_be_block_softc *softc);
 225 #endif
 226 static void ctl_complete_beio(struct ctl_be_block_io *beio);
 227 static int ctl_be_block_move_done(union ctl_io *io);
 228 static void ctl_be_block_biodone(struct bio *bio);
 229 static void ctl_be_block_flush_file(struct ctl_be_block_lun *be_lun,
 230                                     struct ctl_be_block_io *beio);
 231 static void ctl_be_block_dispatch_file(struct ctl_be_block_lun *be_lun,
 232                                        struct ctl_be_block_io *beio);
 233 static void ctl_be_block_flush_dev(struct ctl_be_block_lun *be_lun,
 234                                    struct ctl_be_block_io *beio);
 235 static void ctl_be_block_dispatch_dev(struct ctl_be_block_lun *be_lun,
 236                                       struct ctl_be_block_io *beio);
 237 static void ctl_be_block_cw_dispatch(struct ctl_be_block_lun *be_lun,
 238                                     union ctl_io *io);
 239 static void ctl_be_block_dispatch(struct ctl_be_block_lun *be_lun,
 240                                   union ctl_io *io);
 241 static void ctl_be_block_worker(void *context, int pending);
 242 static int ctl_be_block_submit(union ctl_io *io);
 243 static int ctl_be_block_ioctl(struct cdev *dev, u_long cmd, caddr_t addr,
 244                                    int flag, struct thread *td);
 245 static int ctl_be_block_open_file(struct ctl_be_block_lun *be_lun,
 246                                   struct ctl_lun_req *req);
 247 static int ctl_be_block_open_dev(struct ctl_be_block_lun *be_lun,
 248                                  struct ctl_lun_req *req);
 249 static int ctl_be_block_close(struct ctl_be_block_lun *be_lun);
 250 static int ctl_be_block_open(struct ctl_be_block_softc *softc,
 251                              struct ctl_be_block_lun *be_lun,
 252                              struct ctl_lun_req *req);
 253 static int ctl_be_block_create(struct ctl_be_block_softc *softc,
 254                                struct ctl_lun_req *req);
 255 static int ctl_be_block_rm(struct ctl_be_block_softc *softc,
 256                            struct ctl_lun_req *req);
 257 static int ctl_be_block_modify_file(struct ctl_be_block_lun *be_lun,
 258                                   struct ctl_lun_req *req);
 259 static int ctl_be_block_modify_dev(struct ctl_be_block_lun *be_lun,
 260                                  struct ctl_lun_req *req);
 261 static int ctl_be_block_modify(struct ctl_be_block_softc *softc,
 262                            struct ctl_lun_req *req);
 263 static void ctl_be_block_lun_shutdown(void *be_lun);
 264 static void ctl_be_block_lun_config_status(void *be_lun,
 265                                            ctl_lun_config_status status);
 266 static int ctl_be_block_config_write(union ctl_io *io);
 267 static int ctl_be_block_config_read(union ctl_io *io);
 268 static int ctl_be_block_lun_info(void *be_lun, struct sbuf *sb);
 269 int ctl_be_block_init(void);
 270
 271 static struct ctl_backend_driver ctl_be_block_driver =
 272 {
 273         .name = "block",
 274         .flags = CTL_BE_FLAG_HAS_CONFIG,
 275         .init = ctl_be_block_init,
 276         .data_submit = ctl_be_block_submit,
 277         .data_move_done = ctl_be_block_move_done,
 278         .config_read = ctl_be_block_config_read,
 279         .config_write = ctl_be_block_config_write,
 280         .ioctl = ctl_be_block_ioctl,
 281         .lun_info = ctl_be_block_lun_info
 282 };
 283
 284 MALLOC_DEFINE(M_CTLBLK, "ctlblk", "Memory used for CTL block backend");
 285 CTL_BACKEND_DECLARE(cbb, ctl_be_block_driver);
 286
 287 static struct ctl_be_block_io *
 288 ctl_alloc_beio(struct ctl_be_block_softc *softc)
 289 {
 290         struct ctl_be_block_io *beio;
 291         int count;
 292
 293         mtx_lock(&softc->lock);
 294
 295         beio = STAILQ_FIRST(&softc->beio_free_queue);
 296         if (beio != NULL) {
 297                 STAILQ_REMOVE(&softc->beio_free_queue, beio,
 298                               ctl_be_block_io, links);
 299         }
 300         mtx_unlock(&softc->lock);
 301
 302         if (beio != NULL) {
 303                 bzero(beio, sizeof(*beio));
 304                 beio->softc = softc;
 305                 return (beio);
 306         }
 307
 308         for (;;) {
 309
 310                 count = ctl_grow_beio(softc, /*count*/ 10);
 311
 312                 /*
 313                  * This shouldn't be possible, since ctl_grow_beio() uses a
 314                  * blocking malloc.
 315                  */
 316                 if (count == 0)
 317                         return (NULL);
 318
 319                 /*
 320                  * Since we have to drop the lock when we're allocating beio
 321                  * structures, it's possible someone else can come along and
 322                  * allocate the beio's we've just allocated.
 323                  */
 324                 mtx_lock(&softc->lock);
 325                 beio = STAILQ_FIRST(&softc->beio_free_queue);
 326                 if (beio != NULL) {
 327                         STAILQ_REMOVE(&softc->beio_free_queue, beio,
 328                                       ctl_be_block_io, links);
 329                 }
 330                 mtx_unlock(&softc->lock);
 331
 332                 if (beio != NULL) {
 333                         bzero(beio, sizeof(*beio));
 334                         beio->softc = softc;
 335                         break;
 336                 }
 337         }
 338         return (beio);
 339 }
 340
 341 static void
 342 ctl_free_beio(struct ctl_be_block_io *beio)
 343 {
 344         struct ctl_be_block_softc *softc;
 345         int duplicate_free;
 346         int i;
 347
 348         softc = beio->softc;
 349         duplicate_free = 0;
 350
 351         for (i = 0; i < beio->num_segs; i++) {
 352                 if (beio->sg_segs[i].addr == NULL)
 353                         duplicate_free++;
 354
 355                 uma_zfree(beio->lun->lun_zone, beio->sg_segs[i].addr);
 356                 beio->sg_segs[i].addr = NULL;
 357         }
 358
 359         if (duplicate_free > 0) {
 360                 printf("%s: %d duplicate frees out of %d segments\n", __func__,
 361                        duplicate_free, beio->num_segs);
 362         }
 363         mtx_lock(&softc->lock);
 364         STAILQ_INSERT_TAIL(&softc->beio_free_queue, beio, links);
 365         mtx_unlock(&softc->lock);
 366 }
 367
 368 static int
 369 ctl_grow_beio(struct ctl_be_block_softc *softc, int count)
 370 {
 371         int i;
 372
 373         for (i = 0; i < count; i++) {
 374                 struct ctl_be_block_io *beio;
 375
 376                 beio = (struct ctl_be_block_io *)malloc(sizeof(*beio),
 377                                                            M_CTLBLK,
 378                                                            M_WAITOK | M_ZERO);
 379                 beio->softc = softc;
 380                 mtx_lock(&softc->lock);
 381                 STAILQ_INSERT_TAIL(&softc->beio_free_queue, beio, links);
 382                 mtx_unlock(&softc->lock);
 383         }
 384
 385         return (i);
 386 }
 387
 388 #if 0
 389 static void
 390 ctl_shrink_beio(struct ctl_be_block_softc *softc)
 391 {
 392         struct ctl_be_block_io *beio, *beio_tmp;
 393
 394         mtx_lock(&softc->lock);
 395         STAILQ_FOREACH_SAFE(beio, &softc->beio_free_queue, links, beio_tmp) {
 396                 STAILQ_REMOVE(&softc->beio_free_queue, beio,
 397                               ctl_be_block_io, links);
 398                 free(beio, M_CTLBLK);
 399         }
 400         mtx_unlock(&softc->lock);
 401 }
 402 #endif
 403
 404 static void
 405 ctl_complete_beio(struct ctl_be_block_io *beio)
 406 {
 407         union ctl_io *io;
 408         int io_len;
 409
 410         io = beio->io;
 411
 412         if ((io->io_hdr.status & CTL_STATUS_MASK) == CTL_SUCCESS)
 413                 io_len = beio->io_len;
 414         else
 415                 io_len = 0;
 416
 417         devstat_end_transaction(beio->lun->disk_stats,
 418                                 /*bytes*/ io_len,
 419                                 beio->ds_tag_type,
 420                                 beio->ds_trans_type,
 421                                 /*now*/ NULL,
 422                                 /*then*/&beio->ds_t0);
 423
 424         ctl_free_beio(beio);
 425         ctl_done(io);
 426 }
 427
 428 static int
 429 ctl_be_block_move_done(union ctl_io *io)
 430 {
 431         struct ctl_be_block_io *beio;
 432         struct ctl_be_block_lun *be_lun;
 433 #ifdef CTL_TIME_IO
 434         struct bintime cur_bt;
 435 #endif
 436
 437         beio = (struct ctl_be_block_io *)
 438                 io->io_hdr.ctl_private[CTL_PRIV_BACKEND].ptr;
 439
 440         be_lun = beio->lun;
 441
 442         DPRINTF("entered\n");
 443
 444 #ifdef CTL_TIME_IO
 445         getbintime(&cur_bt);
 446         bintime_sub(&cur_bt, &io->io_hdr.dma_start_bt);
 447         bintime_add(&io->io_hdr.dma_bt, &cur_bt);
 448         io->io_hdr.num_dmas++;
 449 #endif
 450
 451         /*
 452          * We set status at this point for read commands, and write
 453          * commands with errors.
 454          */
 455         if ((beio->bio_cmd == BIO_READ)
 456          && (io->io_hdr.port_status == 0)
 457          && ((io->io_hdr.flags & CTL_FLAG_ABORT) == 0)
 458          && ((io->io_hdr.status & CTL_STATUS_MASK) == CTL_STATUS_NONE))
 459                 ctl_set_success(&io->scsiio);
 460         else if ((io->io_hdr.port_status != 0)
 461               && ((io->io_hdr.flags & CTL_FLAG_ABORT) == 0)
 462               && ((io->io_hdr.status & CTL_STATUS_MASK) == CTL_STATUS_NONE)) {
 463                 /*
 464                  * For hardware error sense keys, the sense key
 465                  * specific value is defined to be a retry count,
 466                  * but we use it to pass back an internal FETD
 467                  * error code.  XXX KDM  Hopefully the FETD is only
 468                  * using 16 bits for an error code, since that's
 469                  * all the space we have in the sks field.
 470                  */
 471                 ctl_set_internal_failure(&io->scsiio,
 472                                          /*sks_valid*/ 1,
 473                                          /*retry_count*/
 474                                          io->io_hdr.port_status);
 475         }
 476
 477         /*
 478          * If this is a read, or a write with errors, it is done.
 479          */
 480         if ((beio->bio_cmd == BIO_READ)
 481          || ((io->io_hdr.flags & CTL_FLAG_ABORT) != 0)
 482          || ((io->io_hdr.status & CTL_STATUS_MASK) != CTL_STATUS_NONE)) {
 483                 ctl_complete_beio(beio);
 484                 return (0);
 485         }
 486
 487         /*
 488          * At this point, we have a write and the DMA completed
 489          * successfully.  We now have to queue it to the task queue to
 490          * execute the backend I/O.  That is because we do blocking
 491          * memory allocations, and in the file backing case, blocking I/O.
 492          * This move done routine is generally called in the SIM's
 493          * interrupt context, and therefore we cannot block.
 494          */
 495         mtx_lock(&be_lun->lock);
 496         /*
 497          * XXX KDM make sure that links is okay to use at this point.
 498          * Otherwise, we either need to add another field to ctl_io_hdr,
 499          * or deal with resource allocation here.
 500          */
 501         STAILQ_INSERT_TAIL(&be_lun->datamove_queue, &io->io_hdr, links);
 502         mtx_unlock(&be_lun->lock);
 503
 504         taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
 505
 506         return (0);
 507 }
 508
 509 static void
 510 ctl_be_block_biodone(struct bio *bio)
 511 {
 512         struct ctl_be_block_io *beio;
 513         struct ctl_be_block_lun *be_lun;
 514         union ctl_io *io;
 515
 516         beio = bio->bio_caller1;
 517         be_lun = beio->lun;
 518         io = beio->io;
 519
 520         DPRINTF("entered\n");
 521
 522         mtx_lock(&be_lun->lock);
 523         if (bio->bio_error != 0)
 524                 beio->num_errors++;
 525
 526         beio->num_bios_done++;
 527
 528         /*
 529          * XXX KDM will this cause WITNESS to complain?  Holding a lock
 530          * during the free might cause it to complain.
 531          */
 532         g_destroy_bio(bio);
 533
 534         /*
 535          * If the send complete bit isn't set, or we aren't the last I/O to
 536          * complete, then we're done.
 537          */
 538         if ((beio->send_complete == 0)
 539          || (beio->num_bios_done < beio->num_bios_sent)) {
 540                 mtx_unlock(&be_lun->lock);
 541                 return;
 542         }
 543
 544         /*
 545          * At this point, we've verified that we are the last I/O to
 546          * complete, so it's safe to drop the lock.
 547          */
 548         mtx_unlock(&be_lun->lock);
 549
 550         /*
 551          * If there are any errors from the backing device, we fail the
 552          * entire I/O with a medium error.
 553          */
 554         if (beio->num_errors > 0) {
 555                 if (beio->bio_cmd == BIO_FLUSH) {
 556                         /* XXX KDM is there is a better error here? */
 557                         ctl_set_internal_failure(&io->scsiio,
 558                                                  /*sks_valid*/ 1,
 559                                                  /*retry_count*/ 0xbad2);
 560                 } else
 561                         ctl_set_medium_error(&io->scsiio);
 562                 ctl_complete_beio(beio);
 563                 return;
 564         }
 565
 566         /*
 567          * If this is a write or a flush, we're all done.
 568          * If this is a read, we can now send the data to the user.
 569          */
 570         if ((beio->bio_cmd == BIO_WRITE)
 571          || (beio->bio_cmd == BIO_FLUSH)) {
 572                 ctl_set_success(&io->scsiio);
 573                 ctl_complete_beio(beio);
 574         } else {
 575                 io->scsiio.be_move_done = ctl_be_block_move_done;
 576                 io->scsiio.kern_data_ptr = (uint8_t *)beio->sg_segs;
 577                 io->scsiio.kern_data_len = beio->io_len;
 578                 io->scsiio.kern_total_len = beio->io_len;
 579                 io->scsiio.kern_rel_offset = 0;
 580                 io->scsiio.kern_data_resid = 0;
 581                 io->scsiio.kern_sg_entries = beio->num_segs;
 582                 io->io_hdr.flags |= CTL_FLAG_ALLOCATED | CTL_FLAG_KDPTR_SGLIST;
 583 #ifdef CTL_TIME_IO
 584                 getbintime(&io->io_hdr.dma_start_bt);
 585 #endif
 586                 ctl_datamove(io);
 587         }
 588 }
 589
 590 static void
 591 ctl_be_block_flush_file(struct ctl_be_block_lun *be_lun,
 592                         struct ctl_be_block_io *beio)
 593 {
 594         union ctl_io *io;
 595         struct mount *mountpoint;
 596         int vfs_is_locked, error, lock_flags;
 597
 598         DPRINTF("entered\n");
 599
 600         io = beio->io;
 601
 602         vfs_is_locked = VFS_LOCK_GIANT(be_lun->vn->v_mount);
 603
 604         (void) vn_start_write(be_lun->vn, &mountpoint, V_WAIT);
 605
 606         if (MNT_SHARED_WRITES(mountpoint)
 607          || ((mountpoint == NULL)
 608           && MNT_SHARED_WRITES(be_lun->vn->v_mount)))
 609                 lock_flags = LK_SHARED;
 610         else
 611                 lock_flags = LK_EXCLUSIVE;
 612
 613         vn_lock(be_lun->vn, lock_flags | LK_RETRY);
 614
 615         binuptime(&beio->ds_t0);
 616         devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
 617
 618         error = VOP_FSYNC(be_lun->vn, MNT_WAIT, curthread);
 619         VOP_UNLOCK(be_lun->vn, 0);
 620
 621         vn_finished_write(mountpoint);
 622
 623         VFS_UNLOCK_GIANT(vfs_is_locked);
 624
 625         if (error == 0)
 626                 ctl_set_success(&io->scsiio);
 627         else {
 628                 /* XXX KDM is there is a better error here? */
 629                 ctl_set_internal_failure(&io->scsiio,
 630                                          /*sks_valid*/ 1,
 631                                          /*retry_count*/ 0xbad1);
 632         }
 633
 634         ctl_complete_beio(beio);
 635 }
 636
 637 SDT_PROBE_DEFINE1(cbb, kernel, read, file_start, file_start, "uint64_t");
 638 SDT_PROBE_DEFINE1(cbb, kernel, write, file_start, file_start, "uint64_t");
 639 SDT_PROBE_DEFINE1(cbb, kernel, read, file_done, file_done,"uint64_t");
 640 SDT_PROBE_DEFINE1(cbb, kernel, write, file_done, file_done, "uint64_t");
 641
 642 static void
 643 ctl_be_block_dispatch_file(struct ctl_be_block_lun *be_lun,
 644                            struct ctl_be_block_io *beio)
 645 {
 646         struct ctl_be_block_filedata *file_data;
 647         union ctl_io *io;
 648         struct uio xuio;
 649         struct iovec *xiovec;
 650         int vfs_is_locked, flags;
 651         int error, i;
 652
 653         DPRINTF("entered\n");
 654
 655         file_data = &be_lun->backend.file;
 656         io = beio->io;
 657         flags = beio->bio_flags;
 658
 659         if (beio->bio_cmd == BIO_READ) {
 660                 SDT_PROBE(cbb, kernel, read, file_start, 0, 0, 0, 0, 0);
 661         } else {
 662                 SDT_PROBE(cbb, kernel, write, file_start, 0, 0, 0, 0, 0);
 663         }
 664
 665         bzero(&xuio, sizeof(xuio));
 666         if (beio->bio_cmd == BIO_READ)
 667                 xuio.uio_rw = UIO_READ;
 668         else
 669                 xuio.uio_rw = UIO_WRITE;
 670
 671         xuio.uio_offset = beio->io_offset;
 672         xuio.uio_resid = beio->io_len;
 673         xuio.uio_segflg = UIO_SYSSPACE;
 674         xuio.uio_iov = beio->xiovecs;
 675         xuio.uio_iovcnt = beio->num_segs;
 676         xuio.uio_td = curthread;
 677
 678         for (i = 0, xiovec = xuio.uio_iov; i < xuio.uio_iovcnt; i++, xiovec++) {
 679                 xiovec->iov_base = beio->sg_segs[i].addr;
 680                 xiovec->iov_len = beio->sg_segs[i].len;
 681         }
 682
 683         vfs_is_locked = VFS_LOCK_GIANT(be_lun->vn->v_mount);
 684         if (beio->bio_cmd == BIO_READ) {
 685                 vn_lock(be_lun->vn, LK_SHARED | LK_RETRY);
 686
 687                 binuptime(&beio->ds_t0);
 688                 devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
 689
 690                 /*
 691                  * UFS pays attention to IO_DIRECT for reads.  If the
 692                  * DIRECTIO option is configured into the kernel, it calls
 693                  * ffs_rawread().  But that only works for single-segment
 694                  * uios with user space addresses.  In our case, with a
 695                  * kernel uio, it still reads into the buffer cache, but it
 696                  * will just try to release the buffer from the cache later
 697                  * on in ffs_read().
 698                  *
 699                  * ZFS does not pay attention to IO_DIRECT for reads.
 700                  *
 701                  * UFS does not pay attention to IO_SYNC for reads.
 702                  *
 703                  * ZFS pays attention to IO_SYNC (which translates into the
 704                  * Solaris define FRSYNC for zfs_read()) for reads.  It
 705                  * attempts to sync the file before reading.
 706                  *
 707                  * So, to attempt to provide some barrier semantics in the
 708                  * BIO_ORDERED case, set both IO_DIRECT and IO_SYNC.
 709                  */
 710                 error = VOP_READ(be_lun->vn, &xuio, (flags & BIO_ORDERED) ?
 711                                  (IO_DIRECT|IO_SYNC) : 0, file_data->cred);
 712
 713                 VOP_UNLOCK(be_lun->vn, 0);
 714         } else {
 715                 struct mount *mountpoint;
 716                 int lock_flags;
 717
 718                 (void)vn_start_write(be_lun->vn, &mountpoint, V_WAIT);
 719
 720                 if (MNT_SHARED_WRITES(mountpoint)
 721                  || ((mountpoint == NULL)
 722                   && MNT_SHARED_WRITES(be_lun->vn->v_mount)))
 723                         lock_flags = LK_SHARED;
 724                 else
 725                         lock_flags = LK_EXCLUSIVE;
 726
 727                 vn_lock(be_lun->vn, lock_flags | LK_RETRY);
 728
 729                 binuptime(&beio->ds_t0);
 730                 devstat_start_transaction(beio->lun->disk_stats, &beio->ds_t0);
 731
 732                 /*
 733                  * UFS pays attention to IO_DIRECT for writes.  The write
 734                  * is done asynchronously.  (Normally the write would just
 735                  * get put into cache.
 736                  *
 737                  * UFS pays attention to IO_SYNC for writes.  It will
 738                  * attempt to write the buffer out synchronously if that
 739                  * flag is set.
 740                  *
 741                  * ZFS does not pay attention to IO_DIRECT for writes.
 742                  *
 743                  * ZFS pays attention to IO_SYNC (a.k.a. FSYNC or FRSYNC)
 744                  * for writes.  It will flush the transaction from the
 745                  * cache before returning.
 746                  *
 747                  * So if we've got the BIO_ORDERED flag set, we want
 748                  * IO_SYNC in either the UFS or ZFS case.
 749                  */
 750                 error = VOP_WRITE(be_lun->vn, &xuio, (flags & BIO_ORDERED) ?
 751                                   IO_SYNC : 0, file_data->cred);
 752                 VOP_UNLOCK(be_lun->vn, 0);
 753
 754                 vn_finished_write(mountpoint);
 755         }
 756         VFS_UNLOCK_GIANT(vfs_is_locked);
 757
 758         /*
 759          * If we got an error, set the sense data to "MEDIUM ERROR" and
 760          * return the I/O to the user.
 761          */
 762         if (error != 0) {
 763                 char path_str[32];
 764
 765                 ctl_scsi_path_string(io, path_str, sizeof(path_str));
 766                 /*
 767                  * XXX KDM ZFS returns ENOSPC when the underlying
 768                  * filesystem fills up.  What kind of SCSI error should we
 769                  * return for that?
 770                  */
 771                 printf("%s%s command returned errno %d\n", path_str,
 772                        (beio->bio_cmd == BIO_READ) ? "READ" : "WRITE", error);
 773                 ctl_set_medium_error(&io->scsiio);
 774                 ctl_complete_beio(beio);
 775                 return;
 776         }
 777
 778         /*
 779          * If this is a write, we're all done.
 780          * If this is a read, we can now send the data to the user.
 781          */
 782         if (beio->bio_cmd == BIO_WRITE) {
 783                 ctl_set_success(&io->scsiio);
 784                 SDT_PROBE(cbb, kernel, write, file_done, 0, 0, 0, 0, 0);
 785                 ctl_complete_beio(beio);
 786         } else {
 787                 SDT_PROBE(cbb, kernel, read, file_done, 0, 0, 0, 0, 0);
 788                 io->scsiio.be_move_done = ctl_be_block_move_done;
 789                 io->scsiio.kern_data_ptr = (uint8_t *)beio->sg_segs;
 790                 io->scsiio.kern_data_len = beio->io_len;
 791                 io->scsiio.kern_total_len = beio->io_len;
 792                 io->scsiio.kern_rel_offset = 0;
 793                 io->scsiio.kern_data_resid = 0;
 794                 io->scsiio.kern_sg_entries = beio->num_segs;
 795                 io->io_hdr.flags |= CTL_FLAG_ALLOCATED | CTL_FLAG_KDPTR_SGLIST;
 796 #ifdef CTL_TIME_IO
 797                 getbintime(&io->io_hdr.dma_start_bt);
 798 #endif
 799                 ctl_datamove(io);
 800         }
 801 }
 802
 803 static void
 804 ctl_be_block_flush_dev(struct ctl_be_block_lun *be_lun,
 805                        struct ctl_be_block_io *beio)
 806 {
 807         struct bio *bio;
 808         union ctl_io *io;
 809         struct ctl_be_block_devdata *dev_data;
 810
 811         dev_data = &be_lun->backend.dev;
 812         io = beio->io;
 813
 814         DPRINTF("entered\n");
 815
 816         /* This can't fail, it's a blocking allocation. */
 817         bio = g_alloc_bio();
 818
 819         bio->bio_cmd        = BIO_FLUSH;
 820         bio->bio_flags     |= BIO_ORDERED;
 821         bio->bio_dev        = dev_data->cdev;
 822         bio->bio_offset     = 0;
 823         bio->bio_data       = 0;
 824         bio->bio_done       = ctl_be_block_biodone;
 825         bio->bio_caller1    = beio;
 826         bio->bio_pblkno     = 0;
 827
 828         /*
 829          * We don't need to acquire the LUN lock here, because we are only
 830          * sending one bio, and so there is no other context to synchronize
 831          * with.
 832          */
 833         beio->num_bios_sent = 1;
 834         beio->send_complete = 1;
 835
 836         binuptime(&beio->ds_t0);
 837         devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
 838
 839         (*dev_data->csw->d_strategy)(bio);
 840 }
 841
 842 static void
 843 ctl_be_block_dispatch_dev(struct ctl_be_block_lun *be_lun,
 844                           struct ctl_be_block_io *beio)
 845 {
 846         int i;
 847         struct bio *bio;
 848         struct ctl_be_block_devdata *dev_data;
 849         off_t cur_offset;
 850         int max_iosize;
 851
 852         DPRINTF("entered\n");
 853
 854         dev_data = &be_lun->backend.dev;
 855
 856         /*
 857          * We have to limit our I/O size to the maximum supported by the
 858          * backend device.  Hopefully it is MAXPHYS.  If the driver doesn't
 859          * set it properly, use DFLTPHYS.
 860          */
 861         max_iosize = dev_data->cdev->si_iosize_max;
 862         if (max_iosize < PAGE_SIZE)
 863                 max_iosize = DFLTPHYS;
 864
 865         cur_offset = beio->io_offset;
 866
 867         /*
 868          * XXX KDM need to accurately reflect the number of I/Os outstanding
 869          * to a device.
 870          */
 871         binuptime(&beio->ds_t0);
 872         devstat_start_transaction(be_lun->disk_stats, &beio->ds_t0);
 873
 874         for (i = 0; i < beio->num_segs; i++) {
 875                 size_t cur_size;
 876                 uint8_t *cur_ptr;
 877
 878                 cur_size = beio->sg_segs[i].len;
 879                 cur_ptr = beio->sg_segs[i].addr;
 880
 881                 while (cur_size > 0) {
 882                         /* This can't fail, it's a blocking allocation. */
 883                         bio = g_alloc_bio();
 884
 885                         KASSERT(bio != NULL, ("g_alloc_bio() failed!\n"));
 886
 887                         bio->bio_cmd = beio->bio_cmd;
 888                         bio->bio_flags |= beio->bio_flags;
 889                         bio->bio_dev = dev_data->cdev;
 890                         bio->bio_caller1 = beio;
 891                         bio->bio_length = min(cur_size, max_iosize);
 892                         bio->bio_offset = cur_offset;
 893                         bio->bio_data = cur_ptr;
 894                         bio->bio_done = ctl_be_block_biodone;
 895                         bio->bio_pblkno = cur_offset / be_lun->blocksize;
 896
 897                         cur_offset += bio->bio_length;
 898                         cur_ptr += bio->bio_length;
 899                         cur_size -= bio->bio_length;
 900
 901                         /*
 902                          * Make sure we set the complete bit just before we
 903                          * issue the last bio so we don't wind up with a
 904                          * race.
 905                          *
 906                          * Use the LUN mutex here instead of a combination
 907                          * of atomic variables for simplicity.
 908                          *
 909                          * XXX KDM we could have a per-IO lock, but that
 910                          * would cause additional per-IO setup and teardown
 911                          * overhead.  Hopefully there won't be too much
 912                          * contention on the LUN lock.
 913                          */
 914                         mtx_lock(&be_lun->lock);
 915
 916                         beio->num_bios_sent++;
 917
 918                         if ((i == beio->num_segs - 1)
 919                          && (cur_size == 0))
 920                                 beio->send_complete = 1;
 921
 922                         mtx_unlock(&be_lun->lock);
 923
 924                         (*dev_data->csw->d_strategy)(bio);
 925                 }
 926         }
 927 }
 928
 929 static void
 930 ctl_be_block_cw_dispatch(struct ctl_be_block_lun *be_lun,
 931                          union ctl_io *io)
 932 {
 933         struct ctl_be_block_io *beio;
 934         struct ctl_be_block_softc *softc;
 935
 936         DPRINTF("entered\n");
 937
 938         softc = be_lun->softc;
 939         beio = ctl_alloc_beio(softc);
 940         if (beio == NULL) {
 941                 /*
 942                  * This should not happen.  ctl_alloc_beio() will call
 943                  * ctl_grow_beio() with a blocking malloc as needed.
 944                  * A malloc with M_WAITOK should not fail.
 945                  */
 946                 ctl_set_busy(&io->scsiio);
 947                 ctl_done(io);
 948                 return;
 949         }
 950
 951         beio->io = io;
 952         beio->softc = softc;
 953         beio->lun = be_lun;
 954         io->io_hdr.ctl_private[CTL_PRIV_BACKEND].ptr = beio;
 955
 956         switch (io->scsiio.cdb[0]) {
 957         case SYNCHRONIZE_CACHE:
 958         case SYNCHRONIZE_CACHE_16:
 959                 beio->ds_trans_type = DEVSTAT_NO_DATA;
 960                 beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
 961                 beio->io_len = 0;
 962                 be_lun->lun_flush(be_lun, beio);
 963                 break;
 964         default:
 965                 panic("Unhandled CDB type %#x", io->scsiio.cdb[0]);
 966                 break;
 967         }
 968 }
 969
 970 SDT_PROBE_DEFINE1(cbb, kernel, read, start, start, "uint64_t");
 971 SDT_PROBE_DEFINE1(cbb, kernel, write, start, start, "uint64_t");
 972 SDT_PROBE_DEFINE1(cbb, kernel, read, alloc_done, alloc_done, "uint64_t");
 973 SDT_PROBE_DEFINE1(cbb, kernel, write, alloc_done, alloc_done, "uint64_t");
 974
 975 static void
 976 ctl_be_block_dispatch(struct ctl_be_block_lun *be_lun,
 977                            union ctl_io *io)
 978 {
 979         struct ctl_be_block_io *beio;
 980         struct ctl_be_block_softc *softc;
 981         struct ctl_lba_len lbalen;
 982         uint64_t len_left, io_size_bytes;
 983         int i;
 984
 985         softc = be_lun->softc;
 986
 987         DPRINTF("entered\n");
 988
 989         if ((io->io_hdr.flags & CTL_FLAG_DATA_MASK) == CTL_FLAG_DATA_IN) {
 990                 SDT_PROBE(cbb, kernel, read, start, 0, 0, 0, 0, 0);
 991         } else {
 992                 SDT_PROBE(cbb, kernel, write, start, 0, 0, 0, 0, 0);
 993         }
 994
 995         memcpy(&lbalen, io->io_hdr.ctl_private[CTL_PRIV_LBA_LEN].bytes,
 996                sizeof(lbalen));
 997
 998         io_size_bytes = lbalen.len * be_lun->blocksize;
 999
1000         /*
1001          * XXX KDM this is temporary, until we implement chaining of beio
1002          * structures and multiple datamove calls to move all the data in
1003          * or out.
1004          */
1005         if (io_size_bytes > CTLBLK_MAX_IO_SIZE) {
1006                 printf("%s: IO length %ju > max io size %u\n", __func__,
1007                        io_size_bytes, CTLBLK_MAX_IO_SIZE);
1008                 ctl_set_invalid_field(&io->scsiio,
1009                                       /*sks_valid*/ 0,
1010                                       /*command*/ 1,
1011                                       /*field*/ 0,
1012                                       /*bit_valid*/ 0,
1013                                       /*bit*/ 0);
1014                 ctl_done(io);
1015                 return;
1016         }
1017
1018         beio = ctl_alloc_beio(softc);
1019         if (beio == NULL) {
1020                 /*
1021                  * This should not happen.  ctl_alloc_beio() will call
1022                  * ctl_grow_beio() with a blocking malloc as needed.
1023                  * A malloc with M_WAITOK should not fail.
1024                  */
1025                 ctl_set_busy(&io->scsiio);
1026                 ctl_done(io);
1027                 return;
1028         }
1029
1030         beio->io = io;
1031         beio->softc = softc;
1032         beio->lun = be_lun;
1033         io->io_hdr.ctl_private[CTL_PRIV_BACKEND].ptr = beio;
1034
1035         /*
1036          * If the I/O came down with an ordered or head of queue tag, set
1037          * the BIO_ORDERED attribute.  For head of queue tags, that's
1038          * pretty much the best we can do.
1039          *
1040          * XXX KDM we don't have a great way to easily know about the FUA
1041          * bit right now (it is decoded in ctl_read_write(), but we don't
1042          * pass that knowledge to the backend), and in any case we would
1043          * need to determine how to handle it.
1044          */
1045         if ((io->scsiio.tag_type == CTL_TAG_ORDERED)
1046          || (io->scsiio.tag_type == CTL_TAG_HEAD_OF_QUEUE))
1047                 beio->bio_flags = BIO_ORDERED;
1048
1049         switch (io->scsiio.tag_type) {
1050         case CTL_TAG_ORDERED:
1051                 beio->ds_tag_type = DEVSTAT_TAG_ORDERED;
1052                 break;
1053         case CTL_TAG_HEAD_OF_QUEUE:
1054                 beio->ds_tag_type = DEVSTAT_TAG_HEAD;
1055                 break;
1056         case CTL_TAG_UNTAGGED:
1057         case CTL_TAG_SIMPLE:
1058         case CTL_TAG_ACA:
1059         default:
1060                 beio->ds_tag_type = DEVSTAT_TAG_SIMPLE;
1061                 break;
1062         }
1063
1064         /*
1065          * This path handles read and write only.  The config write path
1066          * handles flush operations.
1067          */
1068         if ((io->io_hdr.flags & CTL_FLAG_DATA_MASK) == CTL_FLAG_DATA_IN) {
1069                 beio->bio_cmd = BIO_READ;
1070                 beio->ds_trans_type = DEVSTAT_READ;
1071         } else {
1072                 beio->bio_cmd = BIO_WRITE;
1073                 beio->ds_trans_type = DEVSTAT_WRITE;
1074         }
1075
1076         beio->io_len = lbalen.len * be_lun->blocksize;
1077         beio->io_offset = lbalen.lba * be_lun->blocksize;
1078
1079         DPRINTF("%s at LBA %jx len %u\n",
1080                (beio->bio_cmd == BIO_READ) ? "READ" : "WRITE",
1081                (uintmax_t)lbalen.lba, lbalen.len);
1082
1083         for (i = 0, len_left = io_size_bytes; i < CTLBLK_MAX_SEGS &&
1084              len_left > 0; i++) {
1085
1086                 /*
1087                  * Setup the S/G entry for this chunk.
1088                  */
1089                 beio->sg_segs[i].len = min(MAXPHYS, len_left);
1090                 beio->sg_segs[i].addr = uma_zalloc(be_lun->lun_zone, M_WAITOK);
1091
1092                 DPRINTF("segment %d addr %p len %zd\n", i,
1093                         beio->sg_segs[i].addr, beio->sg_segs[i].len);
1094
1095                 beio->num_segs++;
1096                 len_left -= beio->sg_segs[i].len;
1097         }
1098
1099         /*
1100          * For the read case, we need to read the data into our buffers and
1101          * then we can send it back to the user.  For the write case, we
1102          * need to get the data from the user first.
1103          */
1104         if (beio->bio_cmd == BIO_READ) {
1105                 SDT_PROBE(cbb, kernel, read, alloc_done, 0, 0, 0, 0, 0);
1106                 be_lun->dispatch(be_lun, beio);
1107         } else {
1108                 SDT_PROBE(cbb, kernel, write, alloc_done, 0, 0, 0, 0, 0);
1109                 io->scsiio.be_move_done = ctl_be_block_move_done;
1110                 io->scsiio.kern_data_ptr = (uint8_t *)beio->sg_segs;
1111                 io->scsiio.kern_data_len = beio->io_len;
1112                 io->scsiio.kern_total_len = beio->io_len;
1113                 io->scsiio.kern_rel_offset = 0;
1114                 io->scsiio.kern_data_resid = 0;
1115                 io->scsiio.kern_sg_entries = beio->num_segs;
1116                 io->io_hdr.flags |= CTL_FLAG_ALLOCATED | CTL_FLAG_KDPTR_SGLIST;
1117 #ifdef CTL_TIME_IO
1118                 getbintime(&io->io_hdr.dma_start_bt);
1119 #endif
1120                 ctl_datamove(io);
1121         }
1122 }
1123
1124 static void
1125 ctl_be_block_worker(void *context, int pending)
1126 {
1127         struct ctl_be_block_lun *be_lun;
1128         struct ctl_be_block_softc *softc;
1129         union ctl_io *io;
1130
1131         be_lun = (struct ctl_be_block_lun *)context;
1132         softc = be_lun->softc;
1133
1134         DPRINTF("entered\n");
1135
1136         mtx_lock(&be_lun->lock);
1137         for (;;) {
1138                 io = (union ctl_io *)STAILQ_FIRST(&be_lun->datamove_queue);
1139                 if (io != NULL) {
1140                         struct ctl_be_block_io *beio;
1141
1142                         DPRINTF("datamove queue\n");
1143
1144                         STAILQ_REMOVE(&be_lun->datamove_queue, &io->io_hdr,
1145                                       ctl_io_hdr, links);
1146
1147                         mtx_unlock(&be_lun->lock);
1148
1149                         beio = (struct ctl_be_block_io *)
1150                             io->io_hdr.ctl_private[CTL_PRIV_BACKEND].ptr;
1151
1152                         be_lun->dispatch(be_lun, beio);
1153
1154                         mtx_lock(&be_lun->lock);
1155                         continue;
1156                 }
1157                 io = (union ctl_io *)STAILQ_FIRST(&be_lun->config_write_queue);
1158                 if (io != NULL) {
1159
1160                         DPRINTF("config write queue\n");
1161
1162                         STAILQ_REMOVE(&be_lun->config_write_queue, &io->io_hdr,
1163                                       ctl_io_hdr, links);
1164
1165                         mtx_unlock(&be_lun->lock);
1166
1167                         ctl_be_block_cw_dispatch(be_lun, io);
1168
1169                         mtx_lock(&be_lun->lock);
1170                         continue;
1171                 }
1172                 io = (union ctl_io *)STAILQ_FIRST(&be_lun->input_queue);
1173                 if (io != NULL) {
1174                         DPRINTF("input queue\n");
1175
1176                         STAILQ_REMOVE(&be_lun->input_queue, &io->io_hdr,
1177                                       ctl_io_hdr, links);
1178                         mtx_unlock(&be_lun->lock);
1179
1180                         /*
1181                          * We must drop the lock, since this routine and
1182                          * its children may sleep.
1183                          */
1184                         ctl_be_block_dispatch(be_lun, io);
1185
1186                         mtx_lock(&be_lun->lock);
1187                         continue;
1188                 }
1189
1190                 /*
1191                  * If we get here, there is no work left in the queues, so
1192                  * just break out and let the task queue go to sleep.
1193                  */
1194                 break;
1195         }
1196         mtx_unlock(&be_lun->lock);
1197 }
1198
1199 /*
1200  * Entry point from CTL to the backend for I/O.  We queue everything to a
1201  * work thread, so this just puts the I/O on a queue and wakes up the
1202  * thread.
1203  */
1204 static int
1205 ctl_be_block_submit(union ctl_io *io)
1206 {
1207         struct ctl_be_block_lun *be_lun;
1208         struct ctl_be_lun *ctl_be_lun;
1209         int retval;
1210
1211         DPRINTF("entered\n");
1212
1213         retval = CTL_RETVAL_COMPLETE;
1214
1215         ctl_be_lun = (struct ctl_be_lun *)io->io_hdr.ctl_private[
1216                 CTL_PRIV_BACKEND_LUN].ptr;
1217         be_lun = (struct ctl_be_block_lun *)ctl_be_lun->be_lun;
1218
1219         /*
1220          * Make sure we only get SCSI I/O.
1221          */
1222         KASSERT(io->io_hdr.io_type == CTL_IO_SCSI, ("Non-SCSI I/O (type "
1223                 "%#x) encountered", io->io_hdr.io_type));
1224
1225         mtx_lock(&be_lun->lock);
1226         /*
1227          * XXX KDM make sure that links is okay to use at this point.
1228          * Otherwise, we either need to add another field to ctl_io_hdr,
1229          * or deal with resource allocation here.
1230          */
1231         STAILQ_INSERT_TAIL(&be_lun->input_queue, &io->io_hdr, links);
1232         mtx_unlock(&be_lun->lock);
1233
1234         taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
1235
1236         return (retval);
1237 }
1238
1239 static int
1240 ctl_be_block_ioctl(struct cdev *dev, u_long cmd, caddr_t addr,
1241                         int flag, struct thread *td)
1242 {
1243         struct ctl_be_block_softc *softc;
1244         int error;
1245
1246         softc = &backend_block_softc;
1247
1248         error = 0;
1249
1250         switch (cmd) {
1251         case CTL_LUN_REQ: {
1252                 struct ctl_lun_req *lun_req;
1253
1254                 lun_req = (struct ctl_lun_req *)addr;
1255
1256                 switch (lun_req->reqtype) {
1257                 case CTL_LUNREQ_CREATE:
1258                         error = ctl_be_block_create(softc, lun_req);
1259                         break;
1260                 case CTL_LUNREQ_RM:
1261                         error = ctl_be_block_rm(softc, lun_req);
1262                         break;
1263                 case CTL_LUNREQ_MODIFY:
1264                         error = ctl_be_block_modify(softc, lun_req);
1265                         break;
1266                 default:
1267                         lun_req->status = CTL_LUN_ERROR;
1268                         snprintf(lun_req->error_str, sizeof(lun_req->error_str),
1269                                  "%s: invalid LUN request type %d", __func__,
1270                                  lun_req->reqtype);
1271                         break;
1272                 }
1273                 break;
1274         }
1275         default:
1276                 error = ENOTTY;
1277                 break;
1278         }
1279
1280         return (error);
1281 }
1282
1283 static int
1284 ctl_be_block_open_file(struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1285 {
1286         struct ctl_be_block_filedata *file_data;
1287         struct ctl_lun_create_params *params;
1288         struct vattr                  vattr;
1289         int                           error;
1290
1291         error = 0;
1292         file_data = &be_lun->backend.file;
1293         params = &req->reqdata.create;
1294
1295         be_lun->dev_type = CTL_BE_BLOCK_FILE;
1296         be_lun->dispatch = ctl_be_block_dispatch_file;
1297         be_lun->lun_flush = ctl_be_block_flush_file;
1298
1299         error = VOP_GETATTR(be_lun->vn, &vattr, curthread->td_ucred);
1300         if (error != 0) {
1301                 snprintf(req->error_str, sizeof(req->error_str),
1302                          "error calling VOP_GETATTR() for file %s",
1303                          be_lun->dev_path);
1304                 return (error);
1305         }
1306
1307         /*
1308          * Verify that we have the ability to upgrade to exclusive
1309          * access on this file so we can trap errors at open instead
1310          * of reporting them during first access.
1311          */
1312         if (VOP_ISLOCKED(be_lun->vn) != LK_EXCLUSIVE) {
1313                 vn_lock(be_lun->vn, LK_UPGRADE | LK_RETRY);
1314                 if (be_lun->vn->v_iflag & VI_DOOMED) {
1315                         error = EBADF;
1316                         snprintf(req->error_str, sizeof(req->error_str),
1317                                  "error locking file %s", be_lun->dev_path);
1318                         return (error);
1319                 }
1320         }
1321
1322
1323         file_data->cred = crhold(curthread->td_ucred);
1324         if (params->lun_size_bytes != 0)
1325                 be_lun->size_bytes = params->lun_size_bytes;
1326         else
1327                 be_lun->size_bytes = vattr.va_size;
1328         /*
1329          * We set the multi thread flag for file operations because all
1330          * filesystems (in theory) are capable of allowing multiple readers
1331          * of a file at once.  So we want to get the maximum possible
1332          * concurrency.
1333          */
1334         be_lun->flags |= CTL_BE_BLOCK_LUN_MULTI_THREAD;
1335
1336         /*
1337          * XXX KDM vattr.va_blocksize may be larger than 512 bytes here.
1338          * With ZFS, it is 131072 bytes.  Block sizes that large don't work
1339          * with disklabel and UFS on FreeBSD at least.  Large block sizes
1340          * may not work with other OSes as well.  So just export a sector
1341          * size of 512 bytes, which should work with any OS or
1342          * application.  Since our backing is a file, any block size will
1343          * work fine for the backing store.
1344          */
1345 #if 0
1346         be_lun->blocksize= vattr.va_blocksize;
1347 #endif
1348         if (params->blocksize_bytes != 0)
1349                 be_lun->blocksize = params->blocksize_bytes;
1350         else
1351                 be_lun->blocksize = 512;
1352
1353         /*
1354          * Sanity check.  The media size has to be at least one
1355          * sector long.
1356          */
1357         if (be_lun->size_bytes < be_lun->blocksize) {
1358                 error = EINVAL;
1359                 snprintf(req->error_str, sizeof(req->error_str),
1360                          "file %s size %ju < block size %u", be_lun->dev_path,
1361                          (uintmax_t)be_lun->size_bytes, be_lun->blocksize);
1362         }
1363         return (error);
1364 }
1365
1366 static int
1367 ctl_be_block_open_dev(struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1368 {
1369         struct ctl_lun_create_params *params;
1370         struct vattr                  vattr;
1371         struct cdev                  *dev;
1372         struct cdevsw                *devsw;
1373         int                           error;
1374
1375         params = &req->reqdata.create;
1376
1377         be_lun->dev_type = CTL_BE_BLOCK_DEV;
1378         be_lun->dispatch = ctl_be_block_dispatch_dev;
1379         be_lun->lun_flush = ctl_be_block_flush_dev;
1380         be_lun->backend.dev.cdev = be_lun->vn->v_rdev;
1381         be_lun->backend.dev.csw = dev_refthread(be_lun->backend.dev.cdev,
1382                                              &be_lun->backend.dev.dev_ref);
1383         if (be_lun->backend.dev.csw == NULL)
1384                 panic("Unable to retrieve device switch");
1385
1386         error = VOP_GETATTR(be_lun->vn, &vattr, NOCRED);
1387         if (error) {
1388                 snprintf(req->error_str, sizeof(req->error_str),
1389                          "%s: error getting vnode attributes for device %s",
1390                          __func__, be_lun->dev_path);
1391                 return (error);
1392         }
1393
1394         dev = be_lun->vn->v_rdev;
1395         devsw = dev->si_devsw;
1396         if (!devsw->d_ioctl) {
1397                 snprintf(req->error_str, sizeof(req->error_str),
1398                          "%s: no d_ioctl for device %s!", __func__,
1399                          be_lun->dev_path);
1400                 return (ENODEV);
1401         }
1402
1403         error = devsw->d_ioctl(dev, DIOCGSECTORSIZE,
1404                                (caddr_t)&be_lun->blocksize, FREAD,
1405                                curthread);
1406         if (error) {
1407                 snprintf(req->error_str, sizeof(req->error_str),
1408                          "%s: error %d returned for DIOCGSECTORSIZE ioctl "
1409                          "on %s!", __func__, error, be_lun->dev_path);
1410                 return (error);
1411         }
1412
1413         /*
1414          * If the user has asked for a blocksize that is greater than the
1415          * backing device's blocksize, we can do it only if the blocksize
1416          * the user is asking for is an even multiple of the underlying
1417          * device's blocksize.
1418          */
1419         if ((params->blocksize_bytes != 0)
1420          && (params->blocksize_bytes > be_lun->blocksize)) {
1421                 uint32_t bs_multiple, tmp_blocksize;
1422
1423                 bs_multiple = params->blocksize_bytes / be_lun->blocksize;
1424
1425                 tmp_blocksize = bs_multiple * be_lun->blocksize;
1426
1427                 if (tmp_blocksize == params->blocksize_bytes) {
1428                         be_lun->blocksize = params->blocksize_bytes;
1429                 } else {
1430                         snprintf(req->error_str, sizeof(req->error_str),
1431                                  "%s: requested blocksize %u is not an even "
1432                                  "multiple of backing device blocksize %u",
1433                                  __func__, params->blocksize_bytes,
1434                                  be_lun->blocksize);
1435                         return (EINVAL);
1436
1437                 }
1438         } else if ((params->blocksize_bytes != 0)
1439                 && (params->blocksize_bytes != be_lun->blocksize)) {
1440                 snprintf(req->error_str, sizeof(req->error_str),
1441                          "%s: requested blocksize %u < backing device "
1442                          "blocksize %u", __func__, params->blocksize_bytes,
1443                          be_lun->blocksize);
1444                 return (EINVAL);
1445         }
1446
1447         error = devsw->d_ioctl(dev, DIOCGMEDIASIZE,
1448                                (caddr_t)&be_lun->size_bytes, FREAD,
1449                                curthread);
1450         if (error) {
1451                 snprintf(req->error_str, sizeof(req->error_str),
1452                          "%s: error %d returned for DIOCGMEDIASIZE "
1453                          " ioctl on %s!", __func__, error,
1454                          be_lun->dev_path);
1455                 return (error);
1456         }
1457
1458         if (params->lun_size_bytes != 0) {
1459                 if (params->lun_size_bytes > be_lun->size_bytes) {
1460                         snprintf(req->error_str, sizeof(req->error_str),
1461                                  "%s: requested LUN size %ju > backing device "
1462                                  "size %ju", __func__,
1463                                  (uintmax_t)params->lun_size_bytes,
1464                                  (uintmax_t)be_lun->size_bytes);
1465                         return (EINVAL);
1466                 }
1467
1468                 be_lun->size_bytes = params->lun_size_bytes;
1469         }
1470
1471         return (0);
1472 }
1473
1474 static int
1475 ctl_be_block_close(struct ctl_be_block_lun *be_lun)
1476 {
1477         DROP_GIANT();
1478         if (be_lun->vn) {
1479                 int flags = FREAD | FWRITE;
1480                 int vfs_is_locked = 0;
1481
1482                 switch (be_lun->dev_type) {
1483                 case CTL_BE_BLOCK_DEV:
1484                         if (be_lun->backend.dev.csw) {
1485                                 dev_relthread(be_lun->backend.dev.cdev,
1486                                               be_lun->backend.dev.dev_ref);
1487                                 be_lun->backend.dev.csw  = NULL;
1488                                 be_lun->backend.dev.cdev = NULL;
1489                         }
1490                         break;
1491                 case CTL_BE_BLOCK_FILE:
1492                         vfs_is_locked = VFS_LOCK_GIANT(be_lun->vn->v_mount);
1493                         break;
1494                 case CTL_BE_BLOCK_NONE:
1495                 default:
1496                         panic("Unexpected backend type.");
1497                         break;
1498                 }
1499
1500                 (void)vn_close(be_lun->vn, flags, NOCRED, curthread);
1501                 be_lun->vn = NULL;
1502
1503                 switch (be_lun->dev_type) {
1504                 case CTL_BE_BLOCK_DEV:
1505                         break;
1506                 case CTL_BE_BLOCK_FILE:
1507                         VFS_UNLOCK_GIANT(vfs_is_locked);
1508                         if (be_lun->backend.file.cred != NULL) {
1509                                 crfree(be_lun->backend.file.cred);
1510                                 be_lun->backend.file.cred = NULL;
1511                         }
1512                         break;
1513                 case CTL_BE_BLOCK_NONE:
1514                 default:
1515                         panic("Unexpected backend type.");
1516                         break;
1517                 }
1518         }
1519         PICKUP_GIANT();
1520
1521         return (0);
1522 }
1523
1524 static int
1525 ctl_be_block_open(struct ctl_be_block_softc *softc,
1526                        struct ctl_be_block_lun *be_lun, struct ctl_lun_req *req)
1527 {
1528         struct nameidata nd;
1529         int              flags;
1530         int              error;
1531         int              vfs_is_locked;
1532
1533         /*
1534          * XXX KDM allow a read-only option?
1535          */
1536         flags = FREAD | FWRITE;
1537         error = 0;
1538
1539         if (rootvnode == NULL) {
1540                 snprintf(req->error_str, sizeof(req->error_str),
1541                          "%s: Root filesystem is not mounted", __func__);
1542                 return (1);
1543         }
1544
1545         if (!curthread->td_proc->p_fd->fd_cdir) {
1546                 curthread->td_proc->p_fd->fd_cdir = rootvnode;
1547                 VREF(rootvnode);
1548         }
1549         if (!curthread->td_proc->p_fd->fd_rdir) {
1550                 curthread->td_proc->p_fd->fd_rdir = rootvnode;
1551                 VREF(rootvnode);
1552         }
1553         if (!curthread->td_proc->p_fd->fd_jdir) {
1554                 curthread->td_proc->p_fd->fd_jdir = rootvnode;
1555                 VREF(rootvnode);
1556         }
1557
1558  again:
1559         NDINIT(&nd, LOOKUP, FOLLOW, UIO_SYSSPACE, be_lun->dev_path, curthread);
1560         error = vn_open(&nd, &flags, 0, NULL);
1561         if (error) {
1562                 /*
1563                  * This is the only reasonable guess we can make as far as
1564                  * path if the user doesn't give us a fully qualified path.
1565                  * If they want to specify a file, they need to specify the
1566                  * full path.
1567                  */
1568                 if (be_lun->dev_path[0] != '/') {
1569                         char *dev_path = "/dev/";
1570                         char *dev_name;
1571
1572                         /* Try adding device path at beginning of name */
1573                         dev_name = malloc(strlen(be_lun->dev_path)
1574                                         + strlen(dev_path) + 1,
1575                                           M_CTLBLK, M_WAITOK);
1576                         if (dev_name) {
1577                                 sprintf(dev_name, "%s%s", dev_path,
1578                                         be_lun->dev_path);
1579                                 free(be_lun->dev_path, M_CTLBLK);
1580                                 be_lun->dev_path = dev_name;
1581                                 goto again;
1582                         }
1583                 }
1584                 snprintf(req->error_str, sizeof(req->error_str),
1585                          "%s: error opening %s", __func__, be_lun->dev_path);
1586                 return (error);
1587         }
1588
1589         vfs_is_locked = NDHASGIANT(&nd);
1590
1591         NDFREE(&nd, NDF_ONLY_PNBUF);
1592
1593         be_lun->vn = nd.ni_vp;
1594
1595         /* We only support disks and files. */
1596         if (vn_isdisk(be_lun->vn, &error)) {
1597                 error = ctl_be_block_open_dev(be_lun, req);
1598         } else if (be_lun->vn->v_type == VREG) {
1599                 error = ctl_be_block_open_file(be_lun, req);
1600         } else {
1601                 error = EINVAL;
1602                 snprintf(req->error_str, sizeof(req->error_str),
1603                          "%s is not a disk or file", be_lun->dev_path);
1604         }
1605         VOP_UNLOCK(be_lun->vn, 0);
1606         VFS_UNLOCK_GIANT(vfs_is_locked);
1607
1608         if (error != 0) {
1609                 ctl_be_block_close(be_lun);
1610                 return (error);
1611         }
1612
1613         be_lun->blocksize_shift = fls(be_lun->blocksize) - 1;
1614         be_lun->size_blocks = be_lun->size_bytes >> be_lun->blocksize_shift;
1615
1616         return (0);
1617 }
1618
1619 static int
1620 ctl_be_block_mem_ctor(void *mem, int size, void *arg, int flags)
1621 {
1622         return (0);
1623 }
1624
1625 static void
1626 ctl_be_block_mem_dtor(void *mem, int size, void *arg)
1627 {
1628         bzero(mem, size);
1629 }
1630
1631 static int
1632 ctl_be_block_create(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
1633 {
1634         struct ctl_be_block_lun *be_lun;
1635         struct ctl_lun_create_params *params;
1636         struct ctl_be_arg *file_arg;
1637         char tmpstr[32];
1638         int retval, num_threads;
1639         int i;
1640
1641         params = &req->reqdata.create;
1642         retval = 0;
1643
1644         num_threads = cbb_num_threads;
1645
1646         file_arg = NULL;
1647
1648         be_lun = malloc(sizeof(*be_lun), M_CTLBLK, M_ZERO | M_WAITOK);
1649
1650         be_lun->softc = softc;
1651         STAILQ_INIT(&be_lun->input_queue);
1652         STAILQ_INIT(&be_lun->config_write_queue);
1653         STAILQ_INIT(&be_lun->datamove_queue);
1654         sprintf(be_lun->lunname, "cblk%d", softc->num_luns);
1655         mtx_init(&be_lun->lock, be_lun->lunname, NULL, MTX_DEF);
1656
1657         be_lun->lun_zone = uma_zcreate(be_lun->lunname, MAXPHYS,
1658             ctl_be_block_mem_ctor, ctl_be_block_mem_dtor, NULL, NULL,
1659             /*align*/ 0, /*flags*/0);
1660
1661         if (be_lun->lun_zone == NULL) {
1662                 snprintf(req->error_str, sizeof(req->error_str),
1663                          "%s: error allocating UMA zone", __func__);
1664                 goto bailout_error;
1665         }
1666
1667         if (params->flags & CTL_LUN_FLAG_DEV_TYPE)
1668                 be_lun->ctl_be_lun.lun_type = params->device_type;
1669         else
1670                 be_lun->ctl_be_lun.lun_type = T_DIRECT;
1671
1672         if (be_lun->ctl_be_lun.lun_type == T_DIRECT) {
1673                 for (i = 0; i < req->num_be_args; i++) {
1674                         if (strcmp(req->kern_be_args[i].kname, "file") == 0) {
1675                                 file_arg = &req->kern_be_args[i];
1676                                 break;
1677                         }
1678                 }
1679
1680                 if (file_arg == NULL) {
1681                         snprintf(req->error_str, sizeof(req->error_str),
1682                                  "%s: no file argument specified", __func__);
1683                         goto bailout_error;
1684                 }
1685
1686                 be_lun->dev_path = malloc(file_arg->vallen, M_CTLBLK,
1687                                           M_WAITOK | M_ZERO);
1688
1689                 strlcpy(be_lun->dev_path, (char *)file_arg->kvalue,
1690                         file_arg->vallen);
1691
1692                 retval = ctl_be_block_open(softc, be_lun, req);
1693                 if (retval != 0) {
1694                         retval = 0;
1695                         goto bailout_error;
1696                 }
1697
1698                 /*
1699                  * Tell the user the size of the file/device.
1700                  */
1701                 params->lun_size_bytes = be_lun->size_bytes;
1702
1703                 /*
1704                  * The maximum LBA is the size - 1.
1705                  */
1706                 be_lun->ctl_be_lun.maxlba = be_lun->size_blocks - 1;
1707         } else {
1708                 /*
1709                  * For processor devices, we don't have any size.
1710                  */
1711                 be_lun->blocksize = 0;
1712                 be_lun->size_blocks = 0;
1713                 be_lun->size_bytes = 0;
1714                 be_lun->ctl_be_lun.maxlba = 0;
1715                 params->lun_size_bytes = 0;
1716
1717                 /*
1718                  * Default to just 1 thread for processor devices.
1719                  */
1720                 num_threads = 1;
1721         }
1722
1723         /*
1724          * XXX This searching loop might be refactored to be combined with
1725          * the loop above,
1726          */
1727         for (i = 0; i < req->num_be_args; i++) {
1728                 if (strcmp(req->kern_be_args[i].kname, "num_threads") == 0) {
1729                         struct ctl_be_arg *thread_arg;
1730                         char num_thread_str[16];
1731                         int tmp_num_threads;
1732
1733
1734                         thread_arg = &req->kern_be_args[i];
1735
1736                         strlcpy(num_thread_str, (char *)thread_arg->kvalue,
1737                                 min(thread_arg->vallen,
1738                                 sizeof(num_thread_str)));
1739
1740                         tmp_num_threads = strtol(num_thread_str, NULL, 0);
1741
1742                         /*
1743                          * We don't let the user specify less than one
1744                          * thread, but hope he's clueful enough not to
1745                          * specify 1000 threads.
1746                          */
1747                         if (tmp_num_threads < 1) {
1748                                 snprintf(req->error_str, sizeof(req->error_str),
1749                                          "%s: invalid number of threads %s",
1750                                          __func__, num_thread_str);
1751                                 goto bailout_error;
1752                         }
1753
1754                         num_threads = tmp_num_threads;
1755                 }
1756         }
1757
1758         be_lun->flags = CTL_BE_BLOCK_LUN_UNCONFIGURED;
1759         be_lun->ctl_be_lun.flags = CTL_LUN_FLAG_PRIMARY;
1760         be_lun->ctl_be_lun.be_lun = be_lun;
1761         be_lun->ctl_be_lun.blocksize = be_lun->blocksize;
1762         /* Tell the user the blocksize we ended up using */
1763         params->blocksize_bytes = be_lun->blocksize;
1764         if (params->flags & CTL_LUN_FLAG_ID_REQ) {
1765                 be_lun->ctl_be_lun.req_lun_id = params->req_lun_id;
1766                 be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_ID_REQ;
1767         } else
1768                 be_lun->ctl_be_lun.req_lun_id = 0;
1769
1770         be_lun->ctl_be_lun.lun_shutdown = ctl_be_block_lun_shutdown;
1771         be_lun->ctl_be_lun.lun_config_status =
1772                 ctl_be_block_lun_config_status;
1773         be_lun->ctl_be_lun.be = &ctl_be_block_driver;
1774
1775         if ((params->flags & CTL_LUN_FLAG_SERIAL_NUM) == 0) {
1776                 snprintf(tmpstr, sizeof(tmpstr), "MYSERIAL%4d",
1777                          softc->num_luns);
1778                 strncpy((char *)be_lun->ctl_be_lun.serial_num, tmpstr,
1779                         ctl_min(sizeof(be_lun->ctl_be_lun.serial_num),
1780                         sizeof(tmpstr)));
1781
1782                 /* Tell the user what we used for a serial number */
1783                 strncpy((char *)params->serial_num, tmpstr,
1784                         ctl_min(sizeof(params->serial_num), sizeof(tmpstr)));
1785         } else {
1786                 strncpy((char *)be_lun->ctl_be_lun.serial_num,
1787                         params->serial_num,
1788                         ctl_min(sizeof(be_lun->ctl_be_lun.serial_num),
1789                         sizeof(params->serial_num)));
1790         }
1791         if ((params->flags & CTL_LUN_FLAG_DEVID) == 0) {
1792                 snprintf(tmpstr, sizeof(tmpstr), "MYDEVID%4d", softc->num_luns);
1793                 strncpy((char *)be_lun->ctl_be_lun.device_id, tmpstr,
1794                         ctl_min(sizeof(be_lun->ctl_be_lun.device_id),
1795                         sizeof(tmpstr)));
1796
1797                 /* Tell the user what we used for a device ID */
1798                 strncpy((char *)params->device_id, tmpstr,
1799                         ctl_min(sizeof(params->device_id), sizeof(tmpstr)));
1800         } else {
1801                 strncpy((char *)be_lun->ctl_be_lun.device_id,
1802                         params->device_id,
1803                         ctl_min(sizeof(be_lun->ctl_be_lun.device_id),
1804                                 sizeof(params->device_id)));
1805         }
1806
1807         TASK_INIT(&be_lun->io_task, /*priority*/0, ctl_be_block_worker, be_lun);
1808
1809         be_lun->io_taskqueue = taskqueue_create(be_lun->lunname, M_WAITOK,
1810             taskqueue_thread_enqueue, /*context*/&be_lun->io_taskqueue);
1811
1812         if (be_lun->io_taskqueue == NULL) {
1813                 snprintf(req->error_str, sizeof(req->error_str),
1814                          "%s: Unable to create taskqueue", __func__);
1815                 goto bailout_error;
1816         }
1817
1818         /*
1819          * Note that we start the same number of threads by default for
1820          * both the file case and the block device case.  For the file
1821          * case, we need multiple threads to allow concurrency, because the
1822          * vnode interface is designed to be a blocking interface.  For the
1823          * block device case, ZFS zvols at least will block the caller's
1824          * context in many instances, and so we need multiple threads to
1825          * overcome that problem.  Other block devices don't need as many
1826          * threads, but they shouldn't cause too many problems.
1827          *
1828          * If the user wants to just have a single thread for a block
1829          * device, he can specify that when the LUN is created, or change
1830          * the tunable/sysctl to alter the default number of threads.
1831          */
1832         retval = taskqueue_start_threads(&be_lun->io_taskqueue,
1833                                          /*num threads*/num_threads,
1834                                          /*priority*/PWAIT,
1835                                          /*thread name*/
1836                                          "%s taskq", be_lun->lunname);
1837
1838         if (retval != 0)
1839                 goto bailout_error;
1840
1841         be_lun->num_threads = num_threads;
1842
1843         mtx_lock(&softc->lock);
1844         softc->num_luns++;
1845         STAILQ_INSERT_TAIL(&softc->lun_list, be_lun, links);
1846
1847         mtx_unlock(&softc->lock);
1848
1849         retval = ctl_add_lun(&be_lun->ctl_be_lun);
1850         if (retval != 0) {
1851                 mtx_lock(&softc->lock);
1852                 STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun,
1853                               links);
1854                 softc->num_luns--;
1855                 mtx_unlock(&softc->lock);
1856                 snprintf(req->error_str, sizeof(req->error_str),
1857                          "%s: ctl_add_lun() returned error %d, see dmesg for "
1858                         "details", __func__, retval);
1859                 retval = 0;
1860                 goto bailout_error;
1861         }
1862
1863         mtx_lock(&softc->lock);
1864
1865         /*
1866          * Tell the config_status routine that we're waiting so it won't
1867          * clean up the LUN in the event of an error.
1868          */
1869         be_lun->flags |= CTL_BE_BLOCK_LUN_WAITING;
1870
1871         while (be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) {
1872                 retval = msleep(be_lun, &softc->lock, PCATCH, "ctlblk", 0);
1873                 if (retval == EINTR)
1874                         break;
1875         }
1876         be_lun->flags &= ~CTL_BE_BLOCK_LUN_WAITING;
1877
1878         if (be_lun->flags & CTL_BE_BLOCK_LUN_CONFIG_ERR) {
1879                 snprintf(req->error_str, sizeof(req->error_str),
1880                          "%s: LUN configuration error, see dmesg for details",
1881                          __func__);
1882                 STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun,
1883                               links);
1884                 softc->num_luns--;
1885                 mtx_unlock(&softc->lock);
1886                 goto bailout_error;
1887         } else {
1888                 params->req_lun_id = be_lun->ctl_be_lun.lun_id;
1889         }
1890
1891         mtx_unlock(&softc->lock);
1892
1893         be_lun->disk_stats = devstat_new_entry("cbb", params->req_lun_id,
1894                                                be_lun->blocksize,
1895                                                DEVSTAT_ALL_SUPPORTED,
1896                                                be_lun->ctl_be_lun.lun_type
1897                                                | DEVSTAT_TYPE_IF_OTHER,
1898                                                DEVSTAT_PRIORITY_OTHER);
1899
1900
1901         req->status = CTL_LUN_OK;
1902
1903         return (retval);
1904
1905 bailout_error:
1906         req->status = CTL_LUN_ERROR;
1907
1908         ctl_be_block_close(be_lun);
1909
1910         free(be_lun->dev_path, M_CTLBLK);
1911         free(be_lun, M_CTLBLK);
1912
1913         return (retval);
1914 }
1915
1916 static int
1917 ctl_be_block_rm(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
1918 {
1919         struct ctl_lun_rm_params *params;
1920         struct ctl_be_block_lun *be_lun;
1921         int retval;
1922
1923         params = &req->reqdata.rm;
1924
1925         mtx_lock(&softc->lock);
1926
1927         be_lun = NULL;
1928
1929         STAILQ_FOREACH(be_lun, &softc->lun_list, links) {
1930                 if (be_lun->ctl_be_lun.lun_id == params->lun_id)
1931                         break;
1932         }
1933         mtx_unlock(&softc->lock);
1934
1935         if (be_lun == NULL) {
1936                 snprintf(req->error_str, sizeof(req->error_str),
1937                          "%s: LUN %u is not managed by the block backend",
1938                          __func__, params->lun_id);
1939                 goto bailout_error;
1940         }
1941
1942         retval = ctl_disable_lun(&be_lun->ctl_be_lun);
1943
1944         if (retval != 0) {
1945                 snprintf(req->error_str, sizeof(req->error_str),
1946                          "%s: error %d returned from ctl_disable_lun() for "
1947                          "LUN %d", __func__, retval, params->lun_id);
1948                 goto bailout_error;
1949
1950         }
1951
1952         retval = ctl_invalidate_lun(&be_lun->ctl_be_lun);
1953         if (retval != 0) {
1954                 snprintf(req->error_str, sizeof(req->error_str),
1955                          "%s: error %d returned from ctl_invalidate_lun() for "
1956                          "LUN %d", __func__, retval, params->lun_id);
1957                 goto bailout_error;
1958         }
1959
1960         mtx_lock(&softc->lock);
1961
1962         be_lun->flags |= CTL_BE_BLOCK_LUN_WAITING;
1963
1964         while ((be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) == 0) {
1965                 retval = msleep(be_lun, &softc->lock, PCATCH, "ctlblk", 0);
1966                 if (retval == EINTR)
1967                         break;
1968         }
1969
1970         be_lun->flags &= ~CTL_BE_BLOCK_LUN_WAITING;
1971
1972         if ((be_lun->flags & CTL_BE_BLOCK_LUN_UNCONFIGURED) == 0) {
1973                 snprintf(req->error_str, sizeof(req->error_str),
1974                          "%s: interrupted waiting for LUN to be freed",
1975                          __func__);
1976                 mtx_unlock(&softc->lock);
1977                 goto bailout_error;
1978         }
1979
1980         STAILQ_REMOVE(&softc->lun_list, be_lun, ctl_be_block_lun, links);
1981
1982         softc->num_luns--;
1983         mtx_unlock(&softc->lock);
1984
1985         taskqueue_drain(be_lun->io_taskqueue, &be_lun->io_task);
1986
1987         taskqueue_free(be_lun->io_taskqueue);
1988
1989         ctl_be_block_close(be_lun);
1990
1991         if (be_lun->disk_stats != NULL)
1992                 devstat_remove_entry(be_lun->disk_stats);
1993
1994         uma_zdestroy(be_lun->lun_zone);
1995
1996         free(be_lun->dev_path, M_CTLBLK);
1997
1998         free(be_lun, M_CTLBLK);
1999
2000         req->status = CTL_LUN_OK;
2001
2002         return (0);
2003
2004 bailout_error:
2005
2006         req->status = CTL_LUN_ERROR;
2007
2008         return (0);
2009 }
2010
2011 static int
2012 ctl_be_block_modify_file(struct ctl_be_block_lun *be_lun,
2013                          struct ctl_lun_req *req)
2014 {
2015         struct vattr vattr;
2016         int error;
2017         struct ctl_lun_modify_params *params;
2018
2019         params = &req->reqdata.modify;
2020
2021         if (params->lun_size_bytes != 0) {
2022                 be_lun->size_bytes = params->lun_size_bytes;
2023         } else  {
2024                 error = VOP_GETATTR(be_lun->vn, &vattr, curthread->td_ucred);
2025                 if (error != 0) {
2026                         snprintf(req->error_str, sizeof(req->error_str),
2027                                  "error calling VOP_GETATTR() for file %s",
2028                                  be_lun->dev_path);
2029                         return (error);
2030                 }
2031
2032                 be_lun->size_bytes = vattr.va_size;
2033         }
2034
2035         return (0);
2036 }
2037
2038 static int
2039 ctl_be_block_modify_dev(struct ctl_be_block_lun *be_lun,
2040                         struct ctl_lun_req *req)
2041 {
2042         struct cdev *dev;
2043         struct cdevsw *devsw;
2044         int error;
2045         struct ctl_lun_modify_params *params;
2046         uint64_t size_bytes;
2047
2048         params = &req->reqdata.modify;
2049
2050         dev = be_lun->vn->v_rdev;
2051         devsw = dev->si_devsw;
2052         if (!devsw->d_ioctl) {
2053                 snprintf(req->error_str, sizeof(req->error_str),
2054                          "%s: no d_ioctl for device %s!", __func__,
2055                          be_lun->dev_path);
2056                 return (ENODEV);
2057         }
2058
2059         error = devsw->d_ioctl(dev, DIOCGMEDIASIZE,
2060                                (caddr_t)&size_bytes, FREAD,
2061                                curthread);
2062         if (error) {
2063                 snprintf(req->error_str, sizeof(req->error_str),
2064                          "%s: error %d returned for DIOCGMEDIASIZE ioctl "
2065                          "on %s!", __func__, error, be_lun->dev_path);
2066                 return (error);
2067         }
2068
2069         if (params->lun_size_bytes != 0) {
2070                 if (params->lun_size_bytes > size_bytes) {
2071                         snprintf(req->error_str, sizeof(req->error_str),
2072                                  "%s: requested LUN size %ju > backing device "
2073                                  "size %ju", __func__,
2074                                  (uintmax_t)params->lun_size_bytes,
2075                                  (uintmax_t)size_bytes);
2076                         return (EINVAL);
2077                 }
2078
2079                 be_lun->size_bytes = params->lun_size_bytes;
2080         } else {
2081                 be_lun->size_bytes = size_bytes;
2082         }
2083
2084         return (0);
2085 }
2086
2087 static int
2088 ctl_be_block_modify(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
2089 {
2090         struct ctl_lun_modify_params *params;
2091         struct ctl_be_block_lun *be_lun;
2092         int vfs_is_locked, error;
2093
2094         params = &req->reqdata.modify;
2095
2096         mtx_lock(&softc->lock);
2097
2098         be_lun = NULL;
2099
2100         STAILQ_FOREACH(be_lun, &softc->lun_list, links) {
2101                 if (be_lun->ctl_be_lun.lun_id == params->lun_id)
2102                         break;
2103         }
2104         mtx_unlock(&softc->lock);
2105
2106         if (be_lun == NULL) {
2107                 snprintf(req->error_str, sizeof(req->error_str),
2108                          "%s: LUN %u is not managed by the block backend",
2109                          __func__, params->lun_id);
2110                 goto bailout_error;
2111         }
2112
2113         if (params->lun_size_bytes != 0) {
2114                 if (params->lun_size_bytes < be_lun->blocksize) {
2115                         snprintf(req->error_str, sizeof(req->error_str),
2116                                 "%s: LUN size %ju < blocksize %u", __func__,
2117                                 params->lun_size_bytes, be_lun->blocksize);
2118                         goto bailout_error;
2119                 }
2120         }
2121
2122         vfs_is_locked = VFS_LOCK_GIANT(be_lun->vn->v_mount);
2123         vn_lock(be_lun->vn, LK_SHARED | LK_RETRY);
2124
2125         if (be_lun->vn->v_type == VREG)
2126                 error = ctl_be_block_modify_file(be_lun, req);
2127         else
2128                 error = ctl_be_block_modify_dev(be_lun, req);
2129
2130         VOP_UNLOCK(be_lun->vn, 0);
2131         VFS_UNLOCK_GIANT(vfs_is_locked);
2132
2133         if (error != 0)
2134                 goto bailout_error;
2135
2136         be_lun->size_blocks = be_lun->size_bytes >> be_lun->blocksize_shift;
2137
2138         /*
2139          * The maximum LBA is the size - 1.
2140          *
2141          * XXX: Note that this field is being updated without locking,
2142          *      which might cause problems on 32-bit architectures.
2143          */
2144         be_lun->ctl_be_lun.maxlba = be_lun->size_blocks - 1;
2145         ctl_lun_capacity_changed(&be_lun->ctl_be_lun);
2146
2147         /* Tell the user the exact size we ended up using */
2148         params->lun_size_bytes = be_lun->size_bytes;
2149
2150         req->status = CTL_LUN_OK;
2151
2152         return (0);
2153
2154 bailout_error:
2155         req->status = CTL_LUN_ERROR;
2156
2157         return (0);
2158 }
2159
2160 static void
2161 ctl_be_block_lun_shutdown(void *be_lun)
2162 {
2163         struct ctl_be_block_lun *lun;
2164         struct ctl_be_block_softc *softc;
2165
2166         lun = (struct ctl_be_block_lun *)be_lun;
2167
2168         softc = lun->softc;
2169
2170         mtx_lock(&softc->lock);
2171         lun->flags |= CTL_BE_BLOCK_LUN_UNCONFIGURED;
2172         if (lun->flags & CTL_BE_BLOCK_LUN_WAITING)
2173                 wakeup(lun);
2174         mtx_unlock(&softc->lock);
2175
2176 }
2177
2178 static void
2179 ctl_be_block_lun_config_status(void *be_lun, ctl_lun_config_status status)
2180 {
2181         struct ctl_be_block_lun *lun;
2182         struct ctl_be_block_softc *softc;
2183
2184         lun = (struct ctl_be_block_lun *)be_lun;
2185         softc = lun->softc;
2186
2187         if (status == CTL_LUN_CONFIG_OK) {
2188                 mtx_lock(&softc->lock);
2189                 lun->flags &= ~CTL_BE_BLOCK_LUN_UNCONFIGURED;
2190                 if (lun->flags & CTL_BE_BLOCK_LUN_WAITING)
2191                         wakeup(lun);
2192                 mtx_unlock(&softc->lock);
2193
2194                 /*
2195                  * We successfully added the LUN, attempt to enable it.
2196                  */
2197                 if (ctl_enable_lun(&lun->ctl_be_lun) != 0) {
2198                         printf("%s: ctl_enable_lun() failed!\n", __func__);
2199                         if (ctl_invalidate_lun(&lun->ctl_be_lun) != 0) {
2200                                 printf("%s: ctl_invalidate_lun() failed!\n",
2201                                        __func__);
2202                         }
2203                 }
2204
2205                 return;
2206         }
2207
2208
2209         mtx_lock(&softc->lock);
2210         lun->flags &= ~CTL_BE_BLOCK_LUN_UNCONFIGURED;
2211         lun->flags |= CTL_BE_BLOCK_LUN_CONFIG_ERR;
2212         wakeup(lun);
2213         mtx_unlock(&softc->lock);
2214 }
2215
2216
2217 static int
2218 ctl_be_block_config_write(union ctl_io *io)
2219 {
2220         struct ctl_be_block_lun *be_lun;
2221         struct ctl_be_lun *ctl_be_lun;
2222         int retval;
2223
2224         retval = 0;
2225
2226         DPRINTF("entered\n");
2227
2228         ctl_be_lun = (struct ctl_be_lun *)io->io_hdr.ctl_private[
2229                 CTL_PRIV_BACKEND_LUN].ptr;
2230         be_lun = (struct ctl_be_block_lun *)ctl_be_lun->be_lun;
2231
2232         switch (io->scsiio.cdb[0]) {
2233         case SYNCHRONIZE_CACHE:
2234         case SYNCHRONIZE_CACHE_16:
2235                 /*
2236                  * The upper level CTL code will filter out any CDBs with
2237                  * the immediate bit set and return the proper error.
2238                  *
2239                  * We don't really need to worry about what LBA range the
2240                  * user asked to be synced out.  When they issue a sync
2241                  * cache command, we'll sync out the whole thing.
2242                  */
2243                 mtx_lock(&be_lun->lock);
2244                 STAILQ_INSERT_TAIL(&be_lun->config_write_queue, &io->io_hdr,
2245                                    links);
2246                 mtx_unlock(&be_lun->lock);
2247                 taskqueue_enqueue(be_lun->io_taskqueue, &be_lun->io_task);
2248                 break;
2249         case START_STOP_UNIT: {
2250                 struct scsi_start_stop_unit *cdb;
2251
2252                 cdb = (struct scsi_start_stop_unit *)io->scsiio.cdb;
2253
2254                 if (cdb->how & SSS_START)
2255                         retval = ctl_start_lun(ctl_be_lun);
2256                 else {
2257                         retval = ctl_stop_lun(ctl_be_lun);
2258                         /*
2259                          * XXX KDM Copan-specific offline behavior.
2260                          * Figure out a reasonable way to port this?
2261                          */
2262 #ifdef NEEDTOPORT
2263                         if ((retval == 0)
2264                          && (cdb->byte2 & SSS_ONOFFLINE))
2265                                 retval = ctl_lun_offline(ctl_be_lun);
2266 #endif
2267                 }
2268
2269                 /*
2270                  * In general, the above routines should not fail.  They
2271                  * just set state for the LUN.  So we've got something
2272                  * pretty wrong here if we can't start or stop the LUN.
2273                  */
2274                 if (retval != 0) {
2275                         ctl_set_internal_failure(&io->scsiio,
2276                                                  /*sks_valid*/ 1,
2277                                                  /*retry_count*/ 0xf051);
2278                         retval = CTL_RETVAL_COMPLETE;
2279                 } else {
2280                         ctl_set_success(&io->scsiio);
2281                 }
2282                 ctl_config_write_done(io);
2283                 break;
2284         }
2285         default:
2286                 ctl_set_invalid_opcode(&io->scsiio);
2287                 ctl_config_write_done(io);
2288                 retval = CTL_RETVAL_COMPLETE;
2289                 break;
2290         }
2291
2292         return (retval);
2293
2294 }
2295
2296 static int
2297 ctl_be_block_config_read(union ctl_io *io)
2298 {
2299         return (0);
2300 }
2301
2302 static int
2303 ctl_be_block_lun_info(void *be_lun, struct sbuf *sb)
2304 {
2305         struct ctl_be_block_lun *lun;
2306         int retval;
2307
2308         lun = (struct ctl_be_block_lun *)be_lun;
2309         retval = 0;
2310
2311         retval = sbuf_printf(sb, "<num_threads>");
2312
2313         if (retval != 0)
2314                 goto bailout;
2315
2316         retval = sbuf_printf(sb, "%d", lun->num_threads);
2317
2318         if (retval != 0)
2319                 goto bailout;
2320
2321         retval = sbuf_printf(sb, "</num_threads>");
2322
2323         /*
2324          * For processor devices, we don't have a path variable.
2325          */
2326         if ((retval != 0)
2327          || (lun->dev_path == NULL))
2328                 goto bailout;
2329
2330         retval = sbuf_printf(sb, "<file>");
2331
2332         if (retval != 0)
2333                 goto bailout;
2334
2335         retval = ctl_sbuf_printf_esc(sb, lun->dev_path);
2336
2337         if (retval != 0)
2338                 goto bailout;
2339
2340         retval = sbuf_printf(sb, "</file>\n");
2341
2342 bailout:
2343
2344         return (retval);
2345 }
2346
2347 int
2348 ctl_be_block_init(void)
2349 {
2350         struct ctl_be_block_softc *softc;
2351         int retval;
2352
2353         softc = &backend_block_softc;
2354         retval = 0;
2355
2356         mtx_init(&softc->lock, "ctlblk", NULL, MTX_DEF);
2357         STAILQ_INIT(&softc->beio_free_queue);
2358         STAILQ_INIT(&softc->disk_list);
2359         STAILQ_INIT(&softc->lun_list);
2360         ctl_grow_beio(softc, 200);
2361
2362         return (retval);
2363 }