drivers/xen/blkback/blkback.c

   1 /******************************************************************************
   2  * arch/xen/drivers/blkif/backend/main.c
   3  *
   4  * Back-end of the driver for virtual block devices. This portion of the
   5  * driver exports a 'unified' block-device interface that can be accessed
   6  * by any operating system that implements a compatible front end. A
   7  * reference front-end implementation can be found in:
   8  *  arch/xen/drivers/blkif/frontend
   9  *
  10  * Copyright (c) 2003-2004, Keir Fraser & Steve Hand
  11  * Copyright (c) 2005, Christopher Clark
  12  *
  13  * This program is free software; you can redistribute it and/or
  14  * modify it under the terms of the GNU General Public License version 2
  15  * as published by the Free Software Foundation; or, when distributed
  16  * separately from the Linux kernel or incorporated into other
  17  * software packages, subject to the following license:
  18  *
  19  * Permission is hereby granted, free of charge, to any person obtaining a copy
  20  * of this source file (the "Software"), to deal in the Software without
  21  * restriction, including without limitation the rights to use, copy, modify,
  22  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  23  * and to permit persons to whom the Software is furnished to do so, subject to
  24  * the following conditions:
  25  *
  26  * The above copyright notice and this permission notice shall be included in
  27  * all copies or substantial portions of the Software.
  28  *
  29  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  30  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  31  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  32  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  33  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  34  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  35  * IN THE SOFTWARE.
  36  */
  37
  38 #include <linux/spinlock.h>
  39 #include <linux/kthread.h>
  40 #include <linux/freezer.h>
  41 #include <linux/list.h>
  42 #include <linux/module.h>
  43 #include <linux/delay.h>
  44 #include <xen/balloon.h>
  45 #include <xen/evtchn.h>
  46 #include <xen/gnttab.h>
  47 #include <asm/hypervisor.h>
  48 #include "common.h"
  49
  50 /*
  51  * These are rather arbitrary. They are fairly large because adjacent requests
  52  * pulled from a communication ring are quite likely to end up being part of
  53  * the same scatter/gather request at the disc.
  54  *
  55  * ** TRY INCREASING 'blkif_reqs' IF WRITE SPEEDS SEEM TOO LOW **
  56  *
  57  * This will increase the chances of being able to write whole tracks.
  58  * 64 should be enough to keep us competitive with Linux.
  59  */
  60 static int blkif_reqs = 64;
  61 module_param_named(reqs, blkif_reqs, int, 0);
  62 MODULE_PARM_DESC(reqs, "Number of blkback requests to allocate");
  63
  64 /* Run-time switchable: /sys/module/blkback/parameters/ */
  65 static unsigned int log_stats = 0;
  66 static unsigned int debug_lvl = 0;
  67 module_param(log_stats, int, 0644);
  68 module_param(debug_lvl, int, 0644);
  69
  70 /*
  71  * Each outstanding request that we've passed to the lower device layers has a
  72  * 'pending_req' allocated to it. Each buffer_head that completes decrements
  73  * the pendcnt towards zero. When it hits zero, the specified domain has a
  74  * response queued for it, with the saved 'id' passed back.
  75  */
  76 typedef struct {
  77         blkif_t       *blkif;
  78         u64            id;
  79         atomic_t       pendcnt;
  80         unsigned short nr_pages;
  81         unsigned short operation;
  82         struct list_head free_list;
  83 } pending_req_t;
  84
  85 static pending_req_t *pending_reqs;
  86 static struct list_head pending_free;
  87 static DEFINE_SPINLOCK(pending_free_lock);
  88 static DECLARE_WAIT_QUEUE_HEAD(pending_free_wq);
  89
  90 #define BLKBACK_INVALID_HANDLE (~0)
  91
  92 static struct page **pending_pages;
  93 static grant_handle_t *pending_grant_handles;
  94
  95 static inline int vaddr_pagenr(pending_req_t *req, int seg)
  96 {
  97         return (req - pending_reqs) * BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
  98 }
  99
 100 #define pending_page(req, seg) pending_pages[vaddr_pagenr(req, seg)]
 101
 102 static inline unsigned long vaddr(pending_req_t *req, int seg)
 103 {
 104         unsigned long pfn = page_to_pfn(pending_page(req, seg));
 105         return (unsigned long)pfn_to_kaddr(pfn);
 106 }
 107
 108 #define pending_handle(_req, _seg) \
 109         (pending_grant_handles[vaddr_pagenr(_req, _seg)])
 110
 111
 112 static int do_block_io_op(blkif_t *blkif);
 113 static void dispatch_rw_block_io(blkif_t *blkif,
 114                                  blkif_request_t *req,
 115                                  pending_req_t *pending_req);
 116 static void make_response(blkif_t *blkif, u64 id,
 117                           unsigned short op, int st);
 118
 119 /******************************************************************
 120  * misc small helpers
 121  */
 122 static pending_req_t* alloc_req(void)
 123 {
 124         pending_req_t *req = NULL;
 125         unsigned long flags;
 126
 127         spin_lock_irqsave(&pending_free_lock, flags);
 128         if (!list_empty(&pending_free)) {
 129                 req = list_entry(pending_free.next, pending_req_t, free_list);
 130                 list_del(&req->free_list);
 131         }
 132         spin_unlock_irqrestore(&pending_free_lock, flags);
 133         return req;
 134 }
 135
 136 static void free_req(pending_req_t *req)
 137 {
 138         unsigned long flags;
 139         int was_empty;
 140
 141         spin_lock_irqsave(&pending_free_lock, flags);
 142         was_empty = list_empty(&pending_free);
 143         list_add(&req->free_list, &pending_free);
 144         spin_unlock_irqrestore(&pending_free_lock, flags);
 145         if (was_empty)
 146                 wake_up(&pending_free_wq);
 147 }
 148
 149 static void unplug_queue(blkif_t *blkif)
 150 {
 151         if (blkif->plug == NULL)
 152                 return;
 153         kobject_put(&blkif->plug->kobj);
 154         blkif->plug = NULL;
 155 }
 156
 157 static void plug_queue(blkif_t *blkif, struct block_device *bdev)
 158 {
 159         struct request_queue *q = bdev_get_queue(bdev);
 160
 161         if (q == blkif->plug)
 162                 return;
 163         unplug_queue(blkif);
 164         WARN_ON(test_bit(QUEUE_FLAG_DEAD, &q->queue_flags));
 165         kobject_get(&q->kobj);
 166         blkif->plug = q;
 167 }
 168
 169 static void fast_flush_area(pending_req_t *req)
 170 {
 171         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 172         unsigned int i, invcount = 0;
 173         grant_handle_t handle;
 174         int ret;
 175
 176         for (i = 0; i < req->nr_pages; i++) {
 177                 handle = pending_handle(req, i);
 178                 if (handle == BLKBACK_INVALID_HANDLE)
 179                         continue;
 180                 blkback_pagemap_clear(pending_page(req, i));
 181                 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
 182                                     GNTMAP_host_map, handle);
 183                 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
 184                 invcount++;
 185         }
 186
 187         ret = HYPERVISOR_grant_table_op(
 188                 GNTTABOP_unmap_grant_ref, unmap, invcount);
 189         BUG_ON(ret);
 190 }
 191
 192 /******************************************************************
 193  * SCHEDULER FUNCTIONS
 194  */
 195
 196 static void print_stats(blkif_t *blkif)
 197 {
 198         printk(KERN_DEBUG "%s: oo %3d  |  rd %4d  |  wr %4d  |  br %4d"
 199                "  |  fl %4d  |  ds %4d  |  pk %4d\n",
 200                current->comm, blkif->st_oo_req,
 201                blkif->st_rd_req, blkif->st_wr_req,
 202                blkif->st_br_req, blkif->st_fl_req,
 203                blkif->st_ds_req, blkif->st_pk_req);
 204         blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
 205         blkif->st_rd_req = 0;
 206         blkif->st_wr_req = 0;
 207         blkif->st_oo_req = 0;
 208         blkif->st_br_req = 0;
 209         blkif->st_fl_req = 0;
 210         blkif->st_ds_req = 0;
 211         blkif->st_pk_req = 0;
 212 }
 213
 214 int blkif_schedule(void *arg)
 215 {
 216         blkif_t *blkif = arg;
 217         struct vbd *vbd = &blkif->vbd;
 218
 219         blkif_get(blkif);
 220
 221         if (debug_lvl)
 222                 printk(KERN_DEBUG "%s: started\n", current->comm);
 223
 224         while (!kthread_should_stop()) {
 225                 if (try_to_freeze())
 226                         continue;
 227                 if (unlikely(vbd->size != vbd_size(vbd)))
 228                         vbd_resize(blkif);
 229
 230                 wait_event_interruptible(
 231                         blkif->wq,
 232                         blkif->waiting_reqs || kthread_should_stop());
 233                 wait_event_interruptible(
 234                         pending_free_wq,
 235                         !list_empty(&pending_free) || kthread_should_stop());
 236
 237                 blkif->waiting_reqs = 0;
 238                 smp_mb(); /* clear flag *before* checking for work */
 239
 240                 if (do_block_io_op(blkif))
 241                         blkif->waiting_reqs = 1;
 242                 unplug_queue(blkif);
 243
 244                 if (log_stats && time_after(jiffies, blkif->st_print))
 245                         print_stats(blkif);
 246         }
 247
 248         if (log_stats)
 249                 print_stats(blkif);
 250         if (debug_lvl)
 251                 printk(KERN_DEBUG "%s: exiting\n", current->comm);
 252
 253         blkif->xenblkd = NULL;
 254         blkif_put(blkif);
 255
 256         return 0;
 257 }
 258
 259 static void drain_io(blkif_t *blkif)
 260 {
 261         atomic_set(&blkif->drain, 1);
 262         do {
 263                 /* The initial value is one, and one refcnt taken at the
 264                  * start of the blkif_schedule thread. */
 265                 if (atomic_read(&blkif->refcnt) <= 2)
 266                         break;
 267
 268                 wait_for_completion_interruptible_timeout(
 269                                 &blkif->drain_complete, HZ);
 270
 271                 if (!atomic_read(&blkif->drain))
 272                         break;
 273         } while (!kthread_should_stop());
 274         atomic_set(&blkif->drain, 0);
 275 }
 276
 277 /******************************************************************
 278  * COMPLETION CALLBACK -- Called as bh->b_end_io()
 279  */
 280
 281 static void __end_block_io_op(pending_req_t *pending_req, int error)
 282 {
 283         blkif_t *blkif = pending_req->blkif;
 284         int status = BLKIF_RSP_OKAY;
 285
 286         /* An error fails the entire request. */
 287         if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
 288             (error == -EOPNOTSUPP)) {
 289                 DPRINTK("blkback: write barrier op failed, not supported\n");
 290                 blkback_barrier(XBT_NIL, blkif->be, 0);
 291                 status = BLKIF_RSP_EOPNOTSUPP;
 292         } else if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
 293                    (error == -EOPNOTSUPP)) {
 294                 DPRINTK("blkback: flush diskcache op failed, not supported\n");
 295                 blkback_flush_diskcache(XBT_NIL, blkif->be, 0);
 296                 status = BLKIF_RSP_EOPNOTSUPP;
 297         } else if (error) {
 298                 DPRINTK("Buffer not up-to-date at end of operation, "
 299                         "error=%d\n", error);
 300                 status = BLKIF_RSP_ERROR;
 301         }
 302
 303         if (atomic_dec_and_test(&pending_req->pendcnt)) {
 304                 fast_flush_area(pending_req);
 305                 make_response(blkif, pending_req->id,
 306                               pending_req->operation, status);
 307                 free_req(pending_req);
 308                 if (atomic_read(&blkif->drain)
 309                     && atomic_read(&blkif->refcnt) <= 2)
 310                         complete(&blkif->drain_complete);
 311                 blkif_put(blkif);
 312         }
 313 }
 314
 315 static void end_block_io_op(struct bio *bio, int error)
 316 {
 317         __end_block_io_op(bio->bi_private, error);
 318         bio_put(bio);
 319 }
 320
 321
 322 /******************************************************************************
 323  * NOTIFICATION FROM GUEST OS.
 324  */
 325
 326 static void blkif_notify_work(blkif_t *blkif)
 327 {
 328         blkif->waiting_reqs = 1;
 329         wake_up(&blkif->wq);
 330 }
 331
 332 irqreturn_t blkif_be_int(int irq, void *dev_id)
 333 {
 334         blkif_notify_work(dev_id);
 335         return IRQ_HANDLED;
 336 }
 337
 338
 339
 340 /******************************************************************
 341  * DOWNWARD CALLS -- These interface with the block-device layer proper.
 342  */
 343
 344 static void dispatch_discard(blkif_t *blkif, struct blkif_request_discard *req)
 345 {
 346         unsigned long secure = (blkif->vbd.discard_secure &&
 347                                 (req->flag & BLKIF_DISCARD_SECURE)) ?
 348                                BLKDEV_DISCARD_SECURE : 0;
 349         struct phys_req preq;
 350         int status;
 351
 352         blkif->st_ds_req++;
 353
 354         preq.dev           = req->handle;
 355         preq.sector_number = req->sector_number;
 356         preq.nr_sects      = req->nr_sectors;
 357
 358         if (vbd_translate(&preq, blkif, REQ_DISCARD) != 0) {
 359                 DPRINTK("access denied: discard of [%Lu,%Lu) on dev=%04x\n",
 360                         preq.sector_number,
 361                         preq.sector_number + preq.nr_sects, preq.dev);
 362                 make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
 363                 msleep(1); /* back off a bit */
 364                 return;
 365         }
 366
 367         plug_queue(blkif, preq.bdev);
 368
 369         switch (blkdev_issue_discard(preq.bdev, preq.sector_number,
 370                                      preq.nr_sects, GFP_KERNEL, secure)) {
 371         case 0:
 372                 status = BLKIF_RSP_OKAY;
 373                 break;
 374         case -EOPNOTSUPP:
 375                 DPRINTK("discard op failed, not supported\n");
 376                 status = BLKIF_RSP_EOPNOTSUPP;
 377                 break;
 378         default:
 379                 status = BLKIF_RSP_ERROR;
 380                 break;
 381         }
 382
 383         make_response(blkif, req->id, req->operation, status);
 384 }
 385
 386 static int _do_block_io_op(blkif_t *blkif)
 387 {
 388         blkif_back_rings_t *blk_rings = &blkif->blk_rings;
 389         blkif_request_t req;
 390         pending_req_t *pending_req;
 391         RING_IDX rc, rp;
 392
 393         rc = blk_rings->common.req_cons;
 394         rp = blk_rings->common.sring->req_prod;
 395         rmb(); /* Ensure we see queued requests up to 'rp'. */
 396
 397         while (rc != rp) {
 398                 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
 399                         break;
 400
 401                 if (kthread_should_stop())
 402                         return 1;
 403
 404                 switch (blkif->blk_protocol) {
 405                 case BLKIF_PROTOCOL_NATIVE:
 406                         req = *RING_GET_REQUEST(&blk_rings->native, rc);
 407                         break;
 408                 case BLKIF_PROTOCOL_X86_32:
 409                         blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
 410                         break;
 411                 case BLKIF_PROTOCOL_X86_64:
 412                         blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
 413                         break;
 414                 default:
 415                         BUG();
 416                         return 0; /* make compiler happy */
 417                 }
 418
 419                 ++rc;
 420
 421                 switch (req.operation) {
 422                 case BLKIF_OP_READ:
 423                 case BLKIF_OP_WRITE:
 424                 case BLKIF_OP_WRITE_BARRIER:
 425                 case BLKIF_OP_FLUSH_DISKCACHE:
 426                         pending_req = alloc_req();
 427                         if (!pending_req) {
 428                                 blkif->st_oo_req++;
 429                                 return 1;
 430                         }
 431
 432                         /* before make_response() */
 433                         blk_rings->common.req_cons = rc;
 434
 435                         /* Apply all sanity checks to /private copy/ of request. */
 436                         barrier();
 437
 438                         dispatch_rw_block_io(blkif, &req, pending_req);
 439                         break;
 440                 case BLKIF_OP_DISCARD:
 441                         blk_rings->common.req_cons = rc;
 442                         barrier();
 443                         dispatch_discard(blkif, (void *)&req);
 444                         break;
 445                 case BLKIF_OP_PACKET:
 446                         blk_rings->common.req_cons = rc;
 447                         barrier();
 448                         blkif->st_pk_req++;
 449                         DPRINTK("error: block operation BLKIF_OP_PACKET not implemented\n");
 450                         make_response(blkif, req.id, req.operation,
 451                                       BLKIF_RSP_ERROR);
 452                         break;
 453                 default:
 454                         /* A good sign something is wrong: sleep for a while to
 455                          * avoid excessive CPU consumption by a bad guest. */
 456                         msleep(1);
 457                         blk_rings->common.req_cons = rc;
 458                         barrier();
 459                         DPRINTK("error: unknown block io operation [%d]\n",
 460                                 req.operation);
 461                         make_response(blkif, req.id, req.operation,
 462                                       BLKIF_RSP_ERROR);
 463                         break;
 464                 }
 465
 466                 /* Yield point for this unbounded loop. */
 467                 cond_resched();
 468         }
 469
 470         return 0;
 471 }
 472
 473 static int
 474 do_block_io_op(blkif_t *blkif)
 475 {
 476         blkif_back_rings_t *blk_rings = &blkif->blk_rings;
 477         int more_to_do;
 478
 479         do {
 480                 more_to_do = _do_block_io_op(blkif);
 481                 if (more_to_do)
 482                         break;
 483
 484                 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
 485         } while (more_to_do);
 486
 487         return more_to_do;
 488 }
 489
 490 static void dispatch_rw_block_io(blkif_t *blkif,
 491                                  blkif_request_t *req,
 492                                  pending_req_t *pending_req)
 493 {
 494         struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 495         struct phys_req preq;
 496         struct {
 497                 unsigned long buf; unsigned int nsec;
 498         } seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 499         unsigned int nseg;
 500         struct bio *bio = NULL;
 501         uint32_t flags;
 502         int ret, i;
 503         int operation;
 504
 505         switch (req->operation) {
 506         case BLKIF_OP_READ:
 507                 blkif->st_rd_req++;
 508                 operation = READ;
 509                 break;
 510         case BLKIF_OP_WRITE:
 511                 blkif->st_wr_req++;
 512                 operation = WRITE;
 513                 break;
 514         case BLKIF_OP_WRITE_BARRIER:
 515                 blkif->st_br_req++;
 516                 operation = WRITE_FLUSH_FUA;
 517                 break;
 518         case BLKIF_OP_FLUSH_DISKCACHE:
 519                 blkif->st_fl_req++;
 520                 operation = WRITE_FLUSH;
 521                 break;
 522         default:
 523                 operation = 0; /* make gcc happy */
 524                 BUG();
 525         }
 526
 527         /* Check that number of segments is sane. */
 528         nseg = req->nr_segments;
 529         if (unlikely(nseg == 0 && !(operation & REQ_FLUSH)) ||
 530             unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
 531                 DPRINTK("Bad number of segments in request (%d)\n", nseg);
 532                 goto fail_response;
 533         }
 534
 535         preq.dev           = req->handle;
 536         preq.sector_number = req->sector_number;
 537         preq.nr_sects      = 0;
 538
 539         pending_req->blkif     = blkif;
 540         pending_req->id        = req->id;
 541         pending_req->operation = req->operation;
 542         pending_req->nr_pages  = nseg;
 543
 544         flags = GNTMAP_host_map;
 545         if (operation != READ)
 546                 flags |= GNTMAP_readonly;
 547
 548         for (i = 0; i < nseg; i++) {
 549                 seg[i].nsec = req->seg[i].last_sect -
 550                         req->seg[i].first_sect + 1;
 551
 552                 if ((req->seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
 553                     (req->seg[i].last_sect < req->seg[i].first_sect))
 554                         goto fail_response;
 555                 preq.nr_sects += seg[i].nsec;
 556
 557                 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
 558                                   req->seg[i].gref, blkif->domid);
 559         }
 560
 561         ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
 562         BUG_ON(ret);
 563
 564         for (i = 0; i < nseg; i++) {
 565                 if (unlikely(map[i].status == GNTST_eagain))
 566                         gnttab_check_GNTST_eagain_do_while(GNTTABOP_map_grant_ref, &map[i])
 567                 if (unlikely(map[i].status != GNTST_okay)) {
 568                         DPRINTK("invalid buffer -- could not remap it\n");
 569                         map[i].handle = BLKBACK_INVALID_HANDLE;
 570                         ret = 1;
 571                 } else {
 572                         blkback_pagemap_set(vaddr_pagenr(pending_req, i),
 573                                             pending_page(pending_req, i),
 574                                             blkif->domid, req->handle,
 575                                             req->seg[i].gref);
 576                 }
 577
 578                 pending_handle(pending_req, i) = map[i].handle;
 579
 580                 if (ret)
 581                         continue;
 582
 583                 set_phys_to_machine(
 584                         page_to_pfn(pending_page(pending_req, i)),
 585                         FOREIGN_FRAME(map[i].dev_bus_addr >> PAGE_SHIFT));
 586                 seg[i].buf  = map[i].dev_bus_addr |
 587                         (req->seg[i].first_sect << 9);
 588         }
 589
 590         if (ret)
 591                 goto fail_flush;
 592
 593         if (vbd_translate(&preq, blkif, operation) != 0) {
 594                 DPRINTK("access denied: %s of [%llu,%llu] on dev=%04x\n",
 595                         operation == READ ? "read" : "write",
 596                         preq.sector_number,
 597                         preq.sector_number + preq.nr_sects, preq.dev);
 598                 goto fail_flush;
 599         }
 600
 601         /* Wait on all outstanding I/O's and once that has been completed
 602          * issue the WRITE_FLUSH.
 603          */
 604         if (req->operation == BLKIF_OP_WRITE_BARRIER)
 605                 drain_io(blkif);
 606
 607         plug_queue(blkif, preq.bdev);
 608         atomic_set(&pending_req->pendcnt, 1);
 609         blkif_get(blkif);
 610
 611         for (i = 0; i < nseg; i++) {
 612                 if (((int)preq.sector_number|(int)seg[i].nsec) &
 613                     ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
 614                         DPRINTK("Misaligned I/O request from domain %d",
 615                                 blkif->domid);
 616                         goto fail_put_bio;
 617                 }
 618
 619                 while ((bio == NULL) ||
 620                        (bio_add_page(bio,
 621                                      pending_page(pending_req, i),
 622                                      seg[i].nsec << 9,
 623                                      seg[i].buf & ~PAGE_MASK) == 0)) {
 624                         if (bio) {
 625                                 atomic_inc(&pending_req->pendcnt);
 626                                 submit_bio(operation, bio);
 627                         }
 628
 629                         bio = bio_alloc(GFP_KERNEL, nseg-i);
 630                         if (unlikely(bio == NULL))
 631                                 goto fail_put_bio;
 632
 633                         bio->bi_bdev    = preq.bdev;
 634                         bio->bi_private = pending_req;
 635                         bio->bi_end_io  = end_block_io_op;
 636                         bio->bi_sector  = preq.sector_number;
 637                 }
 638
 639                 preq.sector_number += seg[i].nsec;
 640         }
 641
 642         if (!bio) {
 643                 BUG_ON(!(operation & (REQ_FLUSH|REQ_FUA)));
 644                 bio = bio_alloc(GFP_KERNEL, 0);
 645                 if (unlikely(bio == NULL))
 646                         goto fail_put_bio;
 647
 648                 bio->bi_bdev    = preq.bdev;
 649                 bio->bi_private = pending_req;
 650                 bio->bi_end_io  = end_block_io_op;
 651                 bio->bi_sector  = -1;
 652         }
 653
 654         submit_bio(operation, bio);
 655
 656         if (operation == READ)
 657                 blkif->st_rd_sect += preq.nr_sects;
 658         else
 659                 blkif->st_wr_sect += preq.nr_sects;
 660
 661         return;
 662
 663  fail_flush:
 664         fast_flush_area(pending_req);
 665  fail_response:
 666         make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
 667         free_req(pending_req);
 668         msleep(1); /* back off a bit */
 669         return;
 670
 671  fail_put_bio:
 672         __end_block_io_op(pending_req, -EINVAL);
 673         if (bio)
 674                 bio_put(bio);
 675         unplug_queue(blkif);
 676         msleep(1); /* back off a bit */
 677         return;
 678 }
 679
 680
 681
 682 /******************************************************************
 683  * MISCELLANEOUS SETUP / TEARDOWN / DEBUGGING
 684  */
 685
 686
 687 static void make_response(blkif_t *blkif, u64 id,
 688                           unsigned short op, int st)
 689 {
 690         blkif_response_t  resp;
 691         unsigned long     flags;
 692         blkif_back_rings_t *blk_rings = &blkif->blk_rings;
 693         int notify;
 694
 695         resp.id        = id;
 696         resp.operation = op;
 697         resp.status    = st;
 698
 699         spin_lock_irqsave(&blkif->blk_ring_lock, flags);
 700         /* Place on the response ring for the relevant domain. */
 701         switch (blkif->blk_protocol) {
 702         case BLKIF_PROTOCOL_NATIVE:
 703                 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
 704                        &resp, sizeof(resp));
 705                 break;
 706         case BLKIF_PROTOCOL_X86_32:
 707                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
 708                        &resp, sizeof(resp));
 709                 break;
 710         case BLKIF_PROTOCOL_X86_64:
 711                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
 712                        &resp, sizeof(resp));
 713                 break;
 714         default:
 715                 BUG();
 716         }
 717         blk_rings->common.rsp_prod_pvt++;
 718         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
 719         spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
 720
 721         if (notify)
 722                 notify_remote_via_irq(blkif->irq);
 723 }
 724
 725 static int __init blkif_init(void)
 726 {
 727         int i, mmap_pages;
 728
 729         if (!is_running_on_xen())
 730                 return -ENODEV;
 731
 732         mmap_pages = blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
 733
 734         pending_reqs          = kzalloc(sizeof(pending_reqs[0]) *
 735                                         blkif_reqs, GFP_KERNEL);
 736         pending_grant_handles = kmalloc(sizeof(pending_grant_handles[0]) *
 737                                         mmap_pages, GFP_KERNEL);
 738         pending_pages         = alloc_empty_pages_and_pagevec(mmap_pages);
 739
 740         if (blkback_pagemap_init(mmap_pages))
 741                 goto out_of_memory;
 742
 743         if (!pending_reqs || !pending_grant_handles || !pending_pages)
 744                 goto out_of_memory;
 745
 746         for (i = 0; i < mmap_pages; i++)
 747                 pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
 748
 749         blkif_interface_init();
 750
 751         INIT_LIST_HEAD(&pending_free);
 752
 753         for (i = 0; i < blkif_reqs; i++)
 754                 list_add_tail(&pending_reqs[i].free_list, &pending_free);
 755
 756         blkif_xenbus_init();
 757
 758         return 0;
 759
 760  out_of_memory:
 761         kfree(pending_reqs);
 762         kfree(pending_grant_handles);
 763         free_empty_pages_and_pagevec(pending_pages, mmap_pages);
 764         pr_warning("%s: out of memory\n", __FUNCTION__);
 765         return -ENOMEM;
 766 }
 767
 768 module_init(blkif_init);
 769
 770 MODULE_LICENSE("Dual BSD/GPL");
 771 MODULE_ALIAS("xen-backend:vbd");