drivers/xen/blkback/blkback.c

   1 /******************************************************************************
   2  *
   3  * Back-end of the driver for virtual block devices. This portion of the
   4  * driver exports a 'unified' block-device interface that can be accessed
   5  * by any operating system that implements a compatible front end. A
   6  * reference front-end implementation can be found in:
   7  *  drivers/block/xen-blkfront.c
   8  *
   9  * Copyright (c) 2003-2004, Keir Fraser & Steve Hand
  10  * Copyright (c) 2005, Christopher Clark
  11  *
  12  * This program is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU General Public License version 2
  14  * as published by the Free Software Foundation; or, when distributed
  15  * separately from the Linux kernel or incorporated into other
  16  * software packages, subject to the following license:
  17  *
  18  * Permission is hereby granted, free of charge, to any person obtaining a copy
  19  * of this source file (the "Software"), to deal in the Software without
  20  * restriction, including without limitation the rights to use, copy, modify,
  21  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  22  * and to permit persons to whom the Software is furnished to do so, subject to
  23  * the following conditions:
  24  *
  25  * The above copyright notice and this permission notice shall be included in
  26  * all copies or substantial portions of the Software.
  27  *
  28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  29  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  30  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  31  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  32  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  33  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  34  * IN THE SOFTWARE.
  35  */
  36
  37 #include <linux/spinlock.h>
  38 #include <linux/kthread.h>
  39 #include <linux/list.h>
  40 #include <linux/delay.h>
  41 #include <linux/freezer.h>
  42
  43 #include <xen/events.h>
  44 #include <xen/page.h>
  45 #include <asm/xen/hypervisor.h>
  46 #include <asm/xen/hypercall.h>
  47 #include "common.h"
  48
  49 #define WRITE_BARRIER   (REQ_WRITE | REQ_FLUSH | REQ_FUA)
  50
  51 /*
  52  * These are rather arbitrary. They are fairly large because adjacent requests
  53  * pulled from a communication ring are quite likely to end up being part of
  54  * the same scatter/gather request at the disc.
  55  *
  56  * ** TRY INCREASING 'blkif_reqs' IF WRITE SPEEDS SEEM TOO LOW **
  57  *
  58  * This will increase the chances of being able to write whole tracks.
  59  * 64 should be enough to keep us competitive with Linux.
  60  */
  61 static int blkif_reqs = 64;
  62 module_param_named(reqs, blkif_reqs, int, 0);
  63 MODULE_PARM_DESC(reqs, "Number of blkback requests to allocate");
  64
  65 /* Run-time switchable: /sys/module/blkback/parameters/ */
  66 static unsigned int log_stats;
  67 static unsigned int debug_lvl;
  68 module_param(log_stats, int, 0644);
  69 module_param(debug_lvl, int, 0644);
  70
  71 /*
  72  * Each outstanding request that we've passed to the lower device layers has a
  73  * 'pending_req' allocated to it. Each buffer_head that completes decrements
  74  * the pendcnt towards zero. When it hits zero, the specified domain has a
  75  * response queued for it, with the saved 'id' passed back.
  76  */
  77 struct pending_req {
  78         struct blkif_st       *blkif;
  79         u64            id;
  80         int            nr_pages;
  81         atomic_t       pendcnt;
  82         unsigned short operation;
  83         int            status;
  84         struct list_head free_list;
  85 };
  86
  87 #define BLKBACK_INVALID_HANDLE (~0)
  88
  89 struct xen_blkbk {
  90         struct pending_req      *pending_reqs;
  91         /* List of all 'pending_req' available */
  92         struct list_head        pending_free;
  93         /* And its spinlock. */
  94         spinlock_t              pending_free_lock;
  95         wait_queue_head_t       pending_free_wq;
  96         /* The list of all pages that are available. */
  97         struct page             **pending_pages;
  98         /* And the grant handles that are available. */
  99         grant_handle_t          *pending_grant_handles;
 100 };
 101
 102 static struct xen_blkbk *blkbk;
 103
 104 /*
 105  * Little helpful macro to figure out the index and virtual address of the
 106  * pending_pages[..]. For each 'pending_req' we have have up to
 107  * BLKIF_MAX_SEGMENTS_PER_REQUEST (11) pages. The seg would be from 0 through
 108  * 10 and would index in the pending_pages[..]. */
 109 static inline int vaddr_pagenr(struct pending_req *req, int seg)
 110 {
 111         return (req - blkbk->pending_reqs) *
 112                 BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
 113 }
 114
 115 #define pending_page(req, seg) pending_pages[vaddr_pagenr(req, seg)]
 116
 117 static inline unsigned long vaddr(struct pending_req *req, int seg)
 118 {
 119         unsigned long pfn = page_to_pfn(blkbk->pending_page(req, seg));
 120         return (unsigned long)pfn_to_kaddr(pfn);
 121 }
 122
 123 #define pending_handle(_req, _seg) \
 124         (blkbk->pending_grant_handles[vaddr_pagenr(_req, _seg)])
 125
 126
 127 static int do_block_io_op(struct blkif_st *blkif);
 128 static void dispatch_rw_block_io(struct blkif_st *blkif,
 129                                  struct blkif_request *req,
 130                                  struct pending_req *pending_req);
 131 static void make_response(struct blkif_st *blkif, u64 id,
 132                           unsigned short op, int st);
 133
 134 /*
 135  * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
 136  */
 137 static struct pending_req *alloc_req(void)
 138 {
 139         struct pending_req *req = NULL;
 140         unsigned long flags;
 141
 142         spin_lock_irqsave(&blkbk->pending_free_lock, flags);
 143         if (!list_empty(&blkbk->pending_free)) {
 144                 req = list_entry(blkbk->pending_free.next, struct pending_req,
 145                                  free_list);
 146                 list_del(&req->free_list);
 147         }
 148         spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
 149         return req;
 150 }
 151
 152 /*
 153  * Return the 'pending_req' structure back to the freepool. We also
 154  * wake up the thread if it was waiting for a free page.
 155  */
 156 static void free_req(struct pending_req *req)
 157 {
 158         unsigned long flags;
 159         int was_empty;
 160
 161         spin_lock_irqsave(&blkbk->pending_free_lock, flags);
 162         was_empty = list_empty(&blkbk->pending_free);
 163         list_add(&req->free_list, &blkbk->pending_free);
 164         spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
 165         if (was_empty)
 166                 wake_up(&blkbk->pending_free_wq);
 167 }
 168
 169 /*
 170  * Notification from the guest OS.
 171  */
 172 static void blkif_notify_work(struct blkif_st *blkif)
 173 {
 174         blkif->waiting_reqs = 1;
 175         wake_up(&blkif->wq);
 176 }
 177
 178 irqreturn_t blkif_be_int(int irq, void *dev_id)
 179 {
 180         blkif_notify_work(dev_id);
 181         return IRQ_HANDLED;
 182 }
 183
 184 /*
 185  * SCHEDULER FUNCTIONS
 186  */
 187
 188 static void print_stats(struct blkif_st *blkif)
 189 {
 190         printk(KERN_DEBUG "%s: oo %3d  |  rd %4d  |  wr %4d  |  br %4d\n",
 191                current->comm, blkif->st_oo_req,
 192                blkif->st_rd_req, blkif->st_wr_req, blkif->st_br_req);
 193         blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
 194         blkif->st_rd_req = 0;
 195         blkif->st_wr_req = 0;
 196         blkif->st_oo_req = 0;
 197 }
 198
 199 int blkif_schedule(void *arg)
 200 {
 201         struct blkif_st *blkif = arg;
 202         struct vbd *vbd = &blkif->vbd;
 203
 204         blkif_get(blkif);
 205
 206         if (debug_lvl)
 207                 printk(KERN_DEBUG "%s: started\n", current->comm);
 208
 209         while (!kthread_should_stop()) {
 210                 if (try_to_freeze())
 211                         continue;
 212                 if (unlikely(vbd->size != vbd_size(vbd)))
 213                         vbd_resize(blkif);
 214
 215                 wait_event_interruptible(
 216                         blkif->wq,
 217                         blkif->waiting_reqs || kthread_should_stop());
 218                 wait_event_interruptible(
 219                         blkbk->pending_free_wq,
 220                         !list_empty(&blkbk->pending_free) ||
 221                         kthread_should_stop());
 222
 223                 blkif->waiting_reqs = 0;
 224                 smp_mb(); /* clear flag *before* checking for work */
 225
 226                 if (do_block_io_op(blkif))
 227                         blkif->waiting_reqs = 1;
 228
 229                 if (log_stats && time_after(jiffies, blkif->st_print))
 230                         print_stats(blkif);
 231         }
 232
 233         if (log_stats)
 234                 print_stats(blkif);
 235         if (debug_lvl)
 236                 printk(KERN_DEBUG "%s: exiting\n", current->comm);
 237
 238         blkif->xenblkd = NULL;
 239         blkif_put(blkif);
 240
 241         return 0;
 242 }
 243
 244 /*
 245  * Unmap the grant references, and also remove the M2P over-rides
 246  * used in the 'pending_req'.
 247 */
 248 static void fast_flush_area(struct pending_req *req)
 249 {
 250         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 251         unsigned int i, invcount = 0;
 252         grant_handle_t handle;
 253         int ret;
 254
 255         for (i = 0; i < req->nr_pages; i++) {
 256                 handle = pending_handle(req, i);
 257                 if (handle == BLKBACK_INVALID_HANDLE)
 258                         continue;
 259                 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
 260                                     GNTMAP_host_map, handle);
 261                 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
 262                 invcount++;
 263         }
 264
 265         ret = HYPERVISOR_grant_table_op(
 266                 GNTTABOP_unmap_grant_ref, unmap, invcount);
 267         BUG_ON(ret);
 268         /* Note, we use invcount, so nr->pages, so we can't index
 269          * using vaddr(req, i).
 270          */
 271         for (i = 0; i < invcount; i++) {
 272                 ret = m2p_remove_override(
 273                         virt_to_page(unmap[i].host_addr), false);
 274                 if (ret) {
 275                         printk(KERN_ALERT "Failed to remove M2P override for " \
 276                                 "%lx\n", (unsigned long)unmap[i].host_addr);
 277                         continue;
 278                 }
 279         }
 280 }
 281 /*
 282  * Completion callback on the bio's. Called as bh->b_end_io()
 283  */
 284
 285 static void __end_block_io_op(struct pending_req *pending_req, int error)
 286 {
 287         /* An error fails the entire request. */
 288         if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
 289             (error == -EOPNOTSUPP)) {
 290                 DPRINTK("blkback: write barrier op failed, not supported\n");
 291                 blkback_barrier(XBT_NIL, pending_req->blkif->be, 0);
 292                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
 293         } else if (error) {
 294                 DPRINTK("Buffer not up-to-date at end of operation, "
 295                         "error=%d\n", error);
 296                 pending_req->status = BLKIF_RSP_ERROR;
 297         }
 298
 299         /* If all of the bio's have completed it is time to unmap
 300          * the grant references associated with 'request' and provide
 301          * the proper response on the ring.
 302          */
 303         if (atomic_dec_and_test(&pending_req->pendcnt)) {
 304                 fast_flush_area(pending_req);
 305                 make_response(pending_req->blkif, pending_req->id,
 306                               pending_req->operation, pending_req->status);
 307                 blkif_put(pending_req->blkif);
 308                 free_req(pending_req);
 309         }
 310 }
 311
 312 /*
 313  * bio callback.
 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 /*
 324  * Function to copy the from the ring buffer the 'struct blkif_request'
 325  * (which has the sectors we want, number of them, grant references, etc),
 326  * and transmute  it to the block API to hand it over to the proper block disk.
 327  */
 328 static int do_block_io_op(struct blkif_st *blkif)
 329 {
 330         union blkif_back_rings *blk_rings = &blkif->blk_rings;
 331         struct blkif_request req;
 332         struct pending_req *pending_req;
 333         RING_IDX rc, rp;
 334         int more_to_do = 0;
 335
 336         rc = blk_rings->common.req_cons;
 337         rp = blk_rings->common.sring->req_prod;
 338         rmb(); /* Ensure we see queued requests up to 'rp'. */
 339
 340         while (rc != rp) {
 341
 342                 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
 343                         break;
 344
 345                 if (kthread_should_stop()) {
 346                         more_to_do = 1;
 347                         break;
 348                 }
 349
 350                 pending_req = alloc_req();
 351                 if (NULL == pending_req) {
 352                         blkif->st_oo_req++;
 353                         more_to_do = 1;
 354                         break;
 355                 }
 356
 357                 switch (blkif->blk_protocol) {
 358                 case BLKIF_PROTOCOL_NATIVE:
 359                         memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
 360                         break;
 361                 case BLKIF_PROTOCOL_X86_32:
 362                         blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
 363                         break;
 364                 case BLKIF_PROTOCOL_X86_64:
 365                         blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
 366                         break;
 367                 default:
 368                         BUG();
 369                 }
 370                 blk_rings->common.req_cons = ++rc; /* before make_response() */
 371
 372                 /* Apply all sanity checks to /private copy/ of request. */
 373                 barrier();
 374
 375                 switch (req.operation) {
 376                 case BLKIF_OP_READ:
 377                         blkif->st_rd_req++;
 378                         dispatch_rw_block_io(blkif, &req, pending_req);
 379                         break;
 380                 case BLKIF_OP_WRITE_BARRIER:
 381                         blkif->st_br_req++;
 382                         /* fall through */
 383                 case BLKIF_OP_WRITE:
 384                         blkif->st_wr_req++;
 385                         dispatch_rw_block_io(blkif, &req, pending_req);
 386                         break;
 387                 default:
 388                         /* A good sign something is wrong: sleep for a while to
 389                          * avoid excessive CPU consumption by a bad guest. */
 390                         msleep(1);
 391                         DPRINTK("error: unknown block io operation [%d]\n",
 392                                 req.operation);
 393                         make_response(blkif, req.id, req.operation,
 394                                       BLKIF_RSP_ERROR);
 395                         free_req(pending_req);
 396                         break;
 397                 }
 398
 399                 /* Yield point for this unbounded loop. */
 400                 cond_resched();
 401         }
 402
 403         return more_to_do;
 404 }
 405
 406 /*
 407  * Transumation of the 'struct blkif_request' to a proper 'struct bio'
 408  * and call the 'submit_bio' to pass it to the underlaying storage.
 409  */
 410 static void dispatch_rw_block_io(struct blkif_st *blkif,
 411                                  struct blkif_request *req,
 412                                  struct pending_req *pending_req)
 413 {
 414         struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 415         struct phys_req preq;
 416         struct {
 417                 unsigned long buf; unsigned int nsec;
 418         } seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 419         unsigned int nseg;
 420         struct bio *bio = NULL;
 421         struct bio *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 422         int ret, i, nbio = 0;
 423         int operation;
 424         struct blk_plug plug;
 425         struct request_queue *q;
 426
 427         switch (req->operation) {
 428         case BLKIF_OP_READ:
 429                 operation = READ;
 430                 break;
 431         case BLKIF_OP_WRITE:
 432                 operation = WRITE;
 433                 break;
 434         case BLKIF_OP_WRITE_BARRIER:
 435                 operation = WRITE_BARRIER;
 436                 break;
 437         default:
 438                 operation = 0; /* make gcc happy */
 439                 BUG();
 440         }
 441
 442         /* Check that the number of segments is sane. */
 443         nseg = req->nr_segments;
 444         if (unlikely(nseg == 0 && operation != WRITE_BARRIER) ||
 445             unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
 446                 DPRINTK("Bad number of segments in request (%d)\n", nseg);
 447                 goto fail_response;
 448         }
 449
 450         preq.dev           = req->handle;
 451         preq.sector_number = req->u.rw.sector_number;
 452         preq.nr_sects      = 0;
 453
 454         pending_req->blkif     = blkif;
 455         pending_req->id        = req->id;
 456         pending_req->operation = req->operation;
 457         pending_req->status    = BLKIF_RSP_OKAY;
 458         pending_req->nr_pages  = nseg;
 459
 460         /* Fill out preq.nr_sects with proper amount of sectors, and setup
 461          * assign map[..] with the PFN of the page in our domain with the
 462          * corresponding grant reference for each page.
 463          */
 464         for (i = 0; i < nseg; i++) {
 465                 uint32_t flags;
 466
 467                 seg[i].nsec = req->u.rw.seg[i].last_sect -
 468                         req->u.rw.seg[i].first_sect + 1;
 469                 if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
 470                     (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
 471                         goto fail_response;
 472                 preq.nr_sects += seg[i].nsec;
 473
 474                 flags = GNTMAP_host_map;
 475                 if (operation != READ)
 476                         flags |= GNTMAP_readonly;
 477                 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
 478                                   req->u.rw.seg[i].gref, blkif->domid);
 479         }
 480
 481         ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
 482         BUG_ON(ret);
 483
 484         /* Now swizzel the MFN in our domain with the MFN from the other domain
 485          * so that when we access vaddr(pending_req,i) it has the contents of
 486          * the page from the other domain.
 487          */
 488         for (i = 0; i < nseg; i++) {
 489                 if (unlikely(map[i].status != 0)) {
 490                         DPRINTK("invalid buffer -- could not remap it\n");
 491                         map[i].handle = BLKBACK_INVALID_HANDLE;
 492                         ret |= 1;
 493                 }
 494
 495                 pending_handle(pending_req, i) = map[i].handle;
 496
 497                 if (ret)
 498                         continue;
 499
 500                 ret = m2p_add_override(PFN_DOWN(map[i].dev_bus_addr),
 501                         blkbk->pending_page(pending_req, i), false);
 502                 if (ret) {
 503                         printk(KERN_ALERT "Failed to install M2P override for"\
 504                                 " %lx (ret: %d)\n", (unsigned long)
 505                                 map[i].dev_bus_addr, ret);
 506                         /* We could switch over to GNTTABOP_copy */
 507                         continue;
 508                 }
 509
 510                 seg[i].buf  = map[i].dev_bus_addr |
 511                         (req->u.rw.seg[i].first_sect << 9);
 512         }
 513
 514         /* If we have failed at this point, we need to undo the M2P override,
 515          * set gnttab_set_unmap_op on all of the grant references and perform
 516          * the hypercall to unmap the grants - that is all done in
 517          * fast_flush_area.
 518          */
 519         if (ret)
 520                 goto fail_flush;
 521
 522         if (vbd_translate(&preq, blkif, operation) != 0) {
 523                 DPRINTK("access denied: %s of [%llu,%llu] on dev=%04x\n",
 524                         operation == READ ? "read" : "write",
 525                         preq.sector_number,
 526                         preq.sector_number + preq.nr_sects, preq.dev);
 527                 goto fail_flush;
 528         }
 529
 530         /* This corresponding blkif_put is done in __end_block_io_op */
 531         blkif_get(blkif);
 532
 533         for (i = 0; i < nseg; i++) {
 534                 if (((int)preq.sector_number|(int)seg[i].nsec) &
 535                     ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
 536                         DPRINTK("Misaligned I/O request from domain %d",
 537                                 blkif->domid);
 538                         goto fail_put_bio;
 539                 }
 540
 541                 while ((bio == NULL) ||
 542                        (bio_add_page(bio,
 543                                      blkbk->pending_page(pending_req, i),
 544                                      seg[i].nsec << 9,
 545                                      seg[i].buf & ~PAGE_MASK) == 0)) {
 546
 547                         bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, nseg-i);
 548                         if (unlikely(bio == NULL))
 549                                 goto fail_put_bio;
 550
 551                         bio->bi_bdev    = preq.bdev;
 552                         bio->bi_private = pending_req;
 553                         bio->bi_end_io  = end_block_io_op;
 554                         bio->bi_sector  = preq.sector_number;
 555                 }
 556
 557                 preq.sector_number += seg[i].nsec;
 558         }
 559
 560         /* This will be hit if the operation was a barrier. */
 561         if (!bio) {
 562                 BUG_ON(operation != WRITE_BARRIER);
 563                 bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, 0);
 564                 if (unlikely(bio == NULL))
 565                         goto fail_put_bio;
 566
 567                 bio->bi_bdev    = preq.bdev;
 568                 bio->bi_private = pending_req;
 569                 bio->bi_end_io  = end_block_io_op;
 570                 bio->bi_sector  = -1;
 571         }
 572
 573
 574         /* We set it one so that the last submit_bio does not have to call
 575          * atomic_inc.
 576          */
 577         atomic_set(&pending_req->pendcnt, nbio);
 578
 579         /* Get a reference count for the disk queue and start sending I/O */
 580         blk_get_queue(q);
 581         blk_start_plug(&plug);
 582
 583         for (i = 0; i < nbio; i++)
 584                 submit_bio(operation, biolist[i]);
 585
 586         blk_finish_plug(&plug);
 587         /* Let the I/Os go.. */
 588         blk_put_queue(q);
 589
 590         if (operation == READ)
 591                 blkif->st_rd_sect += preq.nr_sects;
 592         else if (operation == WRITE || operation == WRITE_BARRIER)
 593                 blkif->st_wr_sect += preq.nr_sects;
 594
 595         return;
 596
 597  fail_flush:
 598         fast_flush_area(pending_req);
 599  fail_response:
 600         /* Haven't submitted any bio's yet. */
 601         make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
 602         free_req(pending_req);
 603         msleep(1); /* back off a bit */
 604         return;
 605
 606  fail_put_bio:
 607         for (i = 0; i < (nbio-1); i++)
 608                 bio_put(biolist[i]);
 609         __end_block_io_op(pending_req, -EINVAL);
 610         msleep(1); /* back off a bit */
 611         return;
 612 }
 613
 614
 615
 616 /*
 617  * Put a response on the ring on how the operation fared.
 618  */
 619 static void make_response(struct blkif_st *blkif, u64 id,
 620                           unsigned short op, int st)
 621 {
 622         struct blkif_response  resp;
 623         unsigned long     flags;
 624         union blkif_back_rings *blk_rings = &blkif->blk_rings;
 625         int more_to_do = 0;
 626         int notify;
 627
 628         resp.id        = id;
 629         resp.operation = op;
 630         resp.status    = st;
 631
 632         spin_lock_irqsave(&blkif->blk_ring_lock, flags);
 633         /* Place on the response ring for the relevant domain. */
 634         switch (blkif->blk_protocol) {
 635         case BLKIF_PROTOCOL_NATIVE:
 636                 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
 637                        &resp, sizeof(resp));
 638                 break;
 639         case BLKIF_PROTOCOL_X86_32:
 640                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
 641                        &resp, sizeof(resp));
 642                 break;
 643         case BLKIF_PROTOCOL_X86_64:
 644                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
 645                        &resp, sizeof(resp));
 646                 break;
 647         default:
 648                 BUG();
 649         }
 650         blk_rings->common.rsp_prod_pvt++;
 651         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
 652         if (blk_rings->common.rsp_prod_pvt == blk_rings->common.req_cons) {
 653                 /*
 654                  * Tail check for pending requests. Allows frontend to avoid
 655                  * notifications if requests are already in flight (lower
 656                  * overheads and promotes batching).
 657                  */
 658                 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
 659
 660         } else if (RING_HAS_UNCONSUMED_REQUESTS(&blk_rings->common)) {
 661                 more_to_do = 1;
 662         }
 663
 664         spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
 665
 666         if (more_to_do)
 667                 blkif_notify_work(blkif);
 668         if (notify)
 669                 notify_remote_via_irq(blkif->irq);
 670 }
 671
 672 static int __init blkif_init(void)
 673 {
 674         int i, mmap_pages;
 675         int rc = 0;
 676
 677         if (!xen_pv_domain())
 678                 return -ENODEV;
 679
 680         blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
 681         if (!blkbk) {
 682                 printk(KERN_ALERT "%s: out of memory!\n", __func__);
 683                 return -ENOMEM;
 684         }
 685
 686         mmap_pages = blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
 687
 688         blkbk->pending_reqs          = kmalloc(sizeof(blkbk->pending_reqs[0]) *
 689                                         blkif_reqs, GFP_KERNEL);
 690         blkbk->pending_grant_handles = kzalloc(sizeof(blkbk->pending_grant_handles[0]) *
 691                                         mmap_pages, GFP_KERNEL);
 692         blkbk->pending_pages         = kzalloc(sizeof(blkbk->pending_pages[0]) *
 693                                         mmap_pages, GFP_KERNEL);
 694
 695         if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
 696             !blkbk->pending_pages) {
 697                 rc = -ENOMEM;
 698                 goto out_of_memory;
 699         }
 700
 701         for (i = 0; i < mmap_pages; i++) {
 702                 blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
 703                 blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
 704                 if (blkbk->pending_pages[i] == NULL) {
 705                         rc = -ENOMEM;
 706                         goto out_of_memory;
 707                 }
 708         }
 709         rc = blkif_interface_init();
 710         if (rc)
 711                 goto failed_init;
 712
 713         memset(blkbk->pending_reqs, 0, sizeof(blkbk->pending_reqs));
 714
 715         INIT_LIST_HEAD(&blkbk->pending_free);
 716         spin_lock_init(&blkbk->pending_free_lock);
 717         init_waitqueue_head(&blkbk->pending_free_wq);
 718
 719         for (i = 0; i < blkif_reqs; i++)
 720                 list_add_tail(&blkbk->pending_reqs[i].free_list,
 721                               &blkbk->pending_free);
 722
 723         rc = blkif_xenbus_init();
 724         if (rc)
 725                 goto failed_init;
 726
 727         return 0;
 728
 729  out_of_memory:
 730         printk(KERN_ERR "%s: out of memory\n", __func__);
 731  failed_init:
 732         kfree(blkbk->pending_reqs);
 733         kfree(blkbk->pending_grant_handles);
 734         for (i = 0; i < mmap_pages; i++) {
 735                 if (blkbk->pending_pages[i])
 736                         __free_page(blkbk->pending_pages[i]);
 737         }
 738         kfree(blkbk->pending_pages);
 739         kfree(blkbk);
 740         blkbk = NULL;
 741         return rc;
 742 }
 743
 744 module_init(blkif_init);
 745
 746 MODULE_LICENSE("Dual BSD/GPL");