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  * Give back a reference count on the underlaying storage.
 171  * It is OK to make multiple calls in this function as it
 172  * resets the plug to NULL when it is done on the first call.
 173  */
 174 static void unplug_queue(struct blkif_st *blkif)
 175 {
 176         if (blkif->plug == NULL)
 177                 return;
 178         if (blkif->plug->unplug_fn)
 179                 blkif->plug->unplug_fn(blkif->plug);
 180         blk_put_queue(blkif->plug);
 181         blkif->plug = NULL;
 182 }
 183
 184 /*
 185  * Take a reference count on the underlaying storage.
 186  * It is OK to call this multiple times as we check to make sure
 187  * not to double reference. We also give back a reference count
 188  * if it corresponds to another queue.
 189  */
 190 static void plug_queue(struct blkif_st *blkif, struct block_device *bdev)
 191 {
 192         struct request_queue *q = bdev_get_queue(bdev);
 193
 194         if (q == blkif->plug)
 195                 return;
 196         unplug_queue(blkif);
 197         blk_get_queue(q);
 198         blkif->plug = q;
 199 }
 200
 201 /*
 202  * Unmap the grant references, and also remove the M2P over-rides
 203  * used in the 'pending_req'.
 204 */
 205 static void fast_flush_area(struct pending_req *req)
 206 {
 207         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 208         unsigned int i, invcount = 0;
 209         grant_handle_t handle;
 210         int ret;
 211
 212         for (i = 0; i < req->nr_pages; i++) {
 213                 handle = pending_handle(req, i);
 214                 if (handle == BLKBACK_INVALID_HANDLE)
 215                         continue;
 216                 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
 217                                     GNTMAP_host_map, handle);
 218                 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
 219                 invcount++;
 220         }
 221
 222         ret = HYPERVISOR_grant_table_op(
 223                 GNTTABOP_unmap_grant_ref, unmap, invcount);
 224         BUG_ON(ret);
 225         /* Note, we use invcount, so nr->pages, so we can't index
 226          * using vaddr(req, i).
 227          */
 228         for (i = 0; i < invcount; i++) {
 229                 ret = m2p_remove_override(
 230                         virt_to_page(unmap[i].host_addr), false);
 231                 if (ret) {
 232                         printk(KERN_ALERT "Failed to remove M2P override for " \
 233                                 "%lx\n", (unsigned long)unmap[i].host_addr);
 234                         continue;
 235                 }
 236         }
 237 }
 238
 239 /*
 240  * SCHEDULER FUNCTIONS
 241  */
 242
 243 static void print_stats(struct blkif_st *blkif)
 244 {
 245         printk(KERN_DEBUG "%s: oo %3d  |  rd %4d  |  wr %4d  |  br %4d\n",
 246                current->comm, blkif->st_oo_req,
 247                blkif->st_rd_req, blkif->st_wr_req, blkif->st_br_req);
 248         blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
 249         blkif->st_rd_req = 0;
 250         blkif->st_wr_req = 0;
 251         blkif->st_oo_req = 0;
 252 }
 253
 254 int blkif_schedule(void *arg)
 255 {
 256         struct blkif_st *blkif = arg;
 257         struct vbd *vbd = &blkif->vbd;
 258
 259         blkif_get(blkif);
 260
 261         if (debug_lvl)
 262                 printk(KERN_DEBUG "%s: started\n", current->comm);
 263
 264         while (!kthread_should_stop()) {
 265                 if (try_to_freeze())
 266                         continue;
 267                 if (unlikely(vbd->size != vbd_size(vbd)))
 268                         vbd_resize(blkif);
 269
 270                 wait_event_interruptible(
 271                         blkif->wq,
 272                         blkif->waiting_reqs || kthread_should_stop());
 273                 wait_event_interruptible(
 274                         blkbk->pending_free_wq,
 275                         !list_empty(&blkbk->pending_free) ||
 276                         kthread_should_stop());
 277
 278                 blkif->waiting_reqs = 0;
 279                 smp_mb(); /* clear flag *before* checking for work */
 280
 281                 if (do_block_io_op(blkif))
 282                         blkif->waiting_reqs = 1;
 283                 unplug_queue(blkif);
 284
 285                 if (log_stats && time_after(jiffies, blkif->st_print))
 286                         print_stats(blkif);
 287         }
 288
 289         if (log_stats)
 290                 print_stats(blkif);
 291         if (debug_lvl)
 292                 printk(KERN_DEBUG "%s: exiting\n", current->comm);
 293
 294         blkif->xenblkd = NULL;
 295         blkif_put(blkif);
 296
 297         return 0;
 298 }
 299
 300 /*
 301  * Completion callback on the bio's. Called as bh->b_end_io()
 302  */
 303
 304 static void __end_block_io_op(struct pending_req *pending_req, int error)
 305 {
 306         /* An error fails the entire request. */
 307         if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
 308             (error == -EOPNOTSUPP)) {
 309                 DPRINTK("blkback: write barrier op failed, not supported\n");
 310                 blkback_barrier(XBT_NIL, pending_req->blkif->be, 0);
 311                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
 312         } else if (error) {
 313                 DPRINTK("Buffer not up-to-date at end of operation, "
 314                         "error=%d\n", error);
 315                 pending_req->status = BLKIF_RSP_ERROR;
 316         }
 317
 318         /* If all of the bio's have completed it is time to unmap
 319          * the grant references associated with 'request' and provide
 320          * the proper response on the ring.
 321          */
 322         if (atomic_dec_and_test(&pending_req->pendcnt)) {
 323                 fast_flush_area(pending_req);
 324                 make_response(pending_req->blkif, pending_req->id,
 325                               pending_req->operation, pending_req->status);
 326                 blkif_put(pending_req->blkif);
 327                 free_req(pending_req);
 328         }
 329 }
 330
 331 /*
 332  * bio callback.
 333  */
 334 static void end_block_io_op(struct bio *bio, int error)
 335 {
 336         __end_block_io_op(bio->bi_private, error);
 337         bio_put(bio);
 338 }
 339
 340
 341 /*
 342  * Notification from the guest OS.
 343  */
 344
 345 static void blkif_notify_work(struct blkif_st *blkif)
 346 {
 347         blkif->waiting_reqs = 1;
 348         wake_up(&blkif->wq);
 349 }
 350
 351 irqreturn_t blkif_be_int(int irq, void *dev_id)
 352 {
 353         blkif_notify_work(dev_id);
 354         return IRQ_HANDLED;
 355 }
 356
 357
 358
 359 /*
 360  * Function to copy the from the ring buffer the 'struct blkif_request'
 361  * (which has the sectors we want, number of them, grant references, etc),
 362  * and transmute  it to the block API to hand it over to the proper block disk.
 363  */
 364 static int do_block_io_op(struct blkif_st *blkif)
 365 {
 366         union blkif_back_rings *blk_rings = &blkif->blk_rings;
 367         struct blkif_request req;
 368         struct pending_req *pending_req;
 369         RING_IDX rc, rp;
 370         int more_to_do = 0;
 371
 372         rc = blk_rings->common.req_cons;
 373         rp = blk_rings->common.sring->req_prod;
 374         rmb(); /* Ensure we see queued requests up to 'rp'. */
 375
 376         while (rc != rp) {
 377
 378                 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
 379                         break;
 380
 381                 if (kthread_should_stop()) {
 382                         more_to_do = 1;
 383                         break;
 384                 }
 385
 386                 pending_req = alloc_req();
 387                 if (NULL == pending_req) {
 388                         blkif->st_oo_req++;
 389                         more_to_do = 1;
 390                         break;
 391                 }
 392
 393                 switch (blkif->blk_protocol) {
 394                 case BLKIF_PROTOCOL_NATIVE:
 395                         memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
 396                         break;
 397                 case BLKIF_PROTOCOL_X86_32:
 398                         blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
 399                         break;
 400                 case BLKIF_PROTOCOL_X86_64:
 401                         blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
 402                         break;
 403                 default:
 404                         BUG();
 405                 }
 406                 blk_rings->common.req_cons = ++rc; /* before make_response() */
 407
 408                 /* Apply all sanity checks to /private copy/ of request. */
 409                 barrier();
 410
 411                 switch (req.operation) {
 412                 case BLKIF_OP_READ:
 413                         blkif->st_rd_req++;
 414                         dispatch_rw_block_io(blkif, &req, pending_req);
 415                         break;
 416                 case BLKIF_OP_WRITE_BARRIER:
 417                         blkif->st_br_req++;
 418                         /* fall through */
 419                 case BLKIF_OP_WRITE:
 420                         blkif->st_wr_req++;
 421                         dispatch_rw_block_io(blkif, &req, pending_req);
 422                         break;
 423                 default:
 424                         /* A good sign something is wrong: sleep for a while to
 425                          * avoid excessive CPU consumption by a bad guest. */
 426                         msleep(1);
 427                         DPRINTK("error: unknown block io operation [%d]\n",
 428                                 req.operation);
 429                         make_response(blkif, req.id, req.operation,
 430                                       BLKIF_RSP_ERROR);
 431                         free_req(pending_req);
 432                         break;
 433                 }
 434
 435                 /* Yield point for this unbounded loop. */
 436                 cond_resched();
 437         }
 438
 439         return more_to_do;
 440 }
 441
 442 /*
 443  * Transumation of the 'struct blkif_request' to a proper 'struct bio'
 444  * and call the 'submit_bio' to pass it to the underlaying storage.
 445  */
 446 static void dispatch_rw_block_io(struct blkif_st *blkif,
 447                                  struct blkif_request *req,
 448                                  struct pending_req *pending_req)
 449 {
 450         struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 451         struct phys_req preq;
 452         struct {
 453                 unsigned long buf; unsigned int nsec;
 454         } seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 455         unsigned int nseg;
 456         struct bio *bio = NULL;
 457         int ret, i;
 458         int operation;
 459
 460         switch (req->operation) {
 461         case BLKIF_OP_READ:
 462                 operation = READ;
 463                 break;
 464         case BLKIF_OP_WRITE:
 465                 operation = WRITE;
 466                 break;
 467         case BLKIF_OP_WRITE_BARRIER:
 468                 operation = WRITE_BARRIER;
 469                 break;
 470         default:
 471                 operation = 0; /* make gcc happy */
 472                 BUG();
 473         }
 474
 475         /* Check that the number of segments is sane. */
 476         nseg = req->nr_segments;
 477         if (unlikely(nseg == 0 && operation != WRITE_BARRIER) ||
 478             unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
 479                 DPRINTK("Bad number of segments in request (%d)\n", nseg);
 480                 goto fail_response;
 481         }
 482
 483         preq.dev           = req->handle;
 484         preq.sector_number = req->u.rw.sector_number;
 485         preq.nr_sects      = 0;
 486
 487         pending_req->blkif     = blkif;
 488         pending_req->id        = req->id;
 489         pending_req->operation = req->operation;
 490         pending_req->status    = BLKIF_RSP_OKAY;
 491         pending_req->nr_pages  = nseg;
 492
 493         /* Fill out preq.nr_sects with proper amount of sectors, and setup
 494          * assign map[..] with the PFN of the page in our domain with the
 495          * corresponding grant reference for each page.
 496          */
 497         for (i = 0; i < nseg; i++) {
 498                 uint32_t flags;
 499
 500                 seg[i].nsec = req->u.rw.seg[i].last_sect -
 501                         req->u.rw.seg[i].first_sect + 1;
 502                 if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
 503                     (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
 504                         goto fail_response;
 505                 preq.nr_sects += seg[i].nsec;
 506
 507                 flags = GNTMAP_host_map;
 508                 if (operation != READ)
 509                         flags |= GNTMAP_readonly;
 510                 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
 511                                   req->u.rw.seg[i].gref, blkif->domid);
 512         }
 513
 514         ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
 515         BUG_ON(ret);
 516
 517         /* Now swizzel the MFN in our domain with the MFN from the other domain
 518          * so that when we access vaddr(pending_req,i) it has the contents of
 519          * the page from the other domain.
 520          */
 521         for (i = 0; i < nseg; i++) {
 522                 if (unlikely(map[i].status != 0)) {
 523                         DPRINTK("invalid buffer -- could not remap it\n");
 524                         map[i].handle = BLKBACK_INVALID_HANDLE;
 525                         ret |= 1;
 526                 }
 527
 528                 pending_handle(pending_req, i) = map[i].handle;
 529
 530                 if (ret)
 531                         continue;
 532
 533                 ret = m2p_add_override(PFN_DOWN(map[i].dev_bus_addr),
 534                         blkbk->pending_page(pending_req, i), false);
 535                 if (ret) {
 536                         printk(KERN_ALERT "Failed to install M2P override for"\
 537                                 " %lx (ret: %d)\n", (unsigned long)
 538                                 map[i].dev_bus_addr, ret);
 539                         /* We could switch over to GNTTABOP_copy */
 540                         continue;
 541                 }
 542
 543                 seg[i].buf  = map[i].dev_bus_addr |
 544                         (req->u.rw.seg[i].first_sect << 9);
 545         }
 546
 547         /* If we have failed at this point, we need to undo the M2P override,
 548          * set gnttab_set_unmap_op on all of the grant references and perform
 549          * the hypercall to unmap the grants - that is all done in
 550          * fast_flush_area.
 551          */
 552         if (ret)
 553                 goto fail_flush;
 554
 555         if (vbd_translate(&preq, blkif, operation) != 0) {
 556                 DPRINTK("access denied: %s of [%llu,%llu] on dev=%04x\n",
 557                         operation == READ ? "read" : "write",
 558                         preq.sector_number,
 559                         preq.sector_number + preq.nr_sects, preq.dev);
 560                 goto fail_flush;
 561         }
 562
 563         /* Get a reference count for the disk queue and start sending I/O */
 564         plug_queue(blkif, preq.bdev);
 565
 566         /* We set it one so that the last submit_bio does not have to call
 567          * atomic_inc.
 568          */
 569         atomic_set(&pending_req->pendcnt, 1);
 570         blkif_get(blkif);
 571
 572         for (i = 0; i < nseg; i++) {
 573                 if (((int)preq.sector_number|(int)seg[i].nsec) &
 574                     ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
 575                         DPRINTK("Misaligned I/O request from domain %d",
 576                                 blkif->domid);
 577                         goto fail_put_bio;
 578                 }
 579
 580                 while ((bio == NULL) ||
 581                        (bio_add_page(bio,
 582                                      blkbk->pending_page(pending_req, i),
 583                                      seg[i].nsec << 9,
 584                                      seg[i].buf & ~PAGE_MASK) == 0)) {
 585                         if (bio) {
 586                                 atomic_inc(&pending_req->pendcnt);
 587                                 submit_bio(operation, bio);
 588                         }
 589
 590                         bio = bio_alloc(GFP_KERNEL, nseg-i);
 591                         if (unlikely(bio == NULL))
 592                                 goto fail_put_bio;
 593
 594                         bio->bi_bdev    = preq.bdev;
 595                         bio->bi_private = pending_req;
 596                         bio->bi_end_io  = end_block_io_op;
 597                         bio->bi_sector  = preq.sector_number;
 598                 }
 599
 600                 preq.sector_number += seg[i].nsec;
 601         }
 602
 603         /* This will be hit if the operation was a barrier. */
 604         if (!bio) {
 605                 BUG_ON(operation != WRITE_BARRIER);
 606                 bio = bio_alloc(GFP_KERNEL, 0);
 607                 if (unlikely(bio == NULL))
 608                         goto fail_put_bio;
 609
 610                 bio->bi_bdev    = preq.bdev;
 611                 bio->bi_private = pending_req;
 612                 bio->bi_end_io  = end_block_io_op;
 613                 bio->bi_sector  = -1;
 614         }
 615
 616         submit_bio(operation, bio);
 617
 618         if (operation == READ)
 619                 blkif->st_rd_sect += preq.nr_sects;
 620         else if (operation == WRITE || operation == WRITE_BARRIER)
 621                 blkif->st_wr_sect += preq.nr_sects;
 622
 623         return;
 624
 625  fail_flush:
 626         fast_flush_area(pending_req);
 627  fail_response:
 628         make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
 629         free_req(pending_req);
 630         msleep(1); /* back off a bit */
 631         return;
 632
 633  fail_put_bio:
 634         __end_block_io_op(pending_req, -EINVAL);
 635         if (bio)
 636                 bio_put(bio);
 637         unplug_queue(blkif);
 638         msleep(1); /* back off a bit */
 639         return;
 640 }
 641
 642
 643
 644 /*
 645  * Put a response on the ring on how the operation fared.
 646  */
 647 static void make_response(struct blkif_st *blkif, u64 id,
 648                           unsigned short op, int st)
 649 {
 650         struct blkif_response  resp;
 651         unsigned long     flags;
 652         union blkif_back_rings *blk_rings = &blkif->blk_rings;
 653         int more_to_do = 0;
 654         int notify;
 655
 656         resp.id        = id;
 657         resp.operation = op;
 658         resp.status    = st;
 659
 660         spin_lock_irqsave(&blkif->blk_ring_lock, flags);
 661         /* Place on the response ring for the relevant domain. */
 662         switch (blkif->blk_protocol) {
 663         case BLKIF_PROTOCOL_NATIVE:
 664                 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
 665                        &resp, sizeof(resp));
 666                 break;
 667         case BLKIF_PROTOCOL_X86_32:
 668                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
 669                        &resp, sizeof(resp));
 670                 break;
 671         case BLKIF_PROTOCOL_X86_64:
 672                 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
 673                        &resp, sizeof(resp));
 674                 break;
 675         default:
 676                 BUG();
 677         }
 678         blk_rings->common.rsp_prod_pvt++;
 679         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
 680         if (blk_rings->common.rsp_prod_pvt == blk_rings->common.req_cons) {
 681                 /*
 682                  * Tail check for pending requests. Allows frontend to avoid
 683                  * notifications if requests are already in flight (lower
 684                  * overheads and promotes batching).
 685                  */
 686                 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
 687
 688         } else if (RING_HAS_UNCONSUMED_REQUESTS(&blk_rings->common)) {
 689                 more_to_do = 1;
 690         }
 691
 692         spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
 693
 694         if (more_to_do)
 695                 blkif_notify_work(blkif);
 696         if (notify)
 697                 notify_remote_via_irq(blkif->irq);
 698 }
 699
 700 static int __init blkif_init(void)
 701 {
 702         int i, mmap_pages;
 703         int rc = 0;
 704
 705         if (!xen_pv_domain())
 706                 return -ENODEV;
 707
 708         blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
 709         if (!blkbk) {
 710                 printk(KERN_ALERT "%s: out of memory!\n", __func__);
 711                 return -ENOMEM;
 712         }
 713
 714         mmap_pages = blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
 715
 716         blkbk->pending_reqs          = kmalloc(sizeof(blkbk->pending_reqs[0]) *
 717                                         blkif_reqs, GFP_KERNEL);
 718         blkbk->pending_grant_handles = kzalloc(sizeof(blkbk->pending_grant_handles[0]) *
 719                                         mmap_pages, GFP_KERNEL);
 720         blkbk->pending_pages         = kzalloc(sizeof(blkbk->pending_pages[0]) *
 721                                         mmap_pages, GFP_KERNEL);
 722
 723         if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
 724             !blkbk->pending_pages) {
 725                 rc = -ENOMEM;
 726                 goto out_of_memory;
 727         }
 728
 729         for (i = 0; i < mmap_pages; i++) {
 730                 blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
 731                 blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
 732                 if (blkbk->pending_pages[i] == NULL) {
 733                         rc = -ENOMEM;
 734                         goto out_of_memory;
 735                 }
 736         }
 737         rc = blkif_interface_init();
 738         if (rc)
 739                 goto failed_init;
 740
 741         memset(blkbk->pending_reqs, 0, sizeof(blkbk->pending_reqs));
 742
 743         INIT_LIST_HEAD(&blkbk->pending_free);
 744         spin_lock_init(&blkbk->pending_free_lock);
 745         init_waitqueue_head(&blkbk->pending_free_wq);
 746
 747         for (i = 0; i < blkif_reqs; i++)
 748                 list_add_tail(&blkbk->pending_reqs[i].free_list,
 749                               &blkbk->pending_free);
 750
 751         rc = blkif_xenbus_init();
 752         if (rc)
 753                 goto failed_init;
 754
 755         return 0;
 756
 757  out_of_memory:
 758         printk(KERN_ERR "%s: out of memory\n", __func__);
 759  failed_init:
 760         kfree(blkbk->pending_reqs);
 761         kfree(blkbk->pending_grant_handles);
 762         for (i = 0; i < mmap_pages; i++) {
 763                 if (blkbk->pending_pages[i])
 764                         __free_page(blkbk->pending_pages[i]);
 765         }
 766         kfree(blkbk->pending_pages);
 767         kfree(blkbk);
 768         blkbk = NULL;
 769         return rc;
 770 }
 771
 772 module_init(blkif_init);
 773
 774 MODULE_LICENSE("Dual BSD/GPL");