xen-blkback: don't leak stack data via response ring
[pandora-kernel.git] / drivers / block / 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 #include <linux/loop.h>
43 #include <linux/falloc.h>
44 #include <linux/fs.h>
45
46 #include <xen/events.h>
47 #include <xen/page.h>
48 #include <asm/xen/hypervisor.h>
49 #include <asm/xen/hypercall.h>
50 #include "common.h"
51
52 /*
53  * These are rather arbitrary. They are fairly large because adjacent requests
54  * pulled from a communication ring are quite likely to end up being part of
55  * the same scatter/gather request at the disc.
56  *
57  * ** TRY INCREASING 'xen_blkif_reqs' IF WRITE SPEEDS SEEM TOO LOW **
58  *
59  * This will increase the chances of being able to write whole tracks.
60  * 64 should be enough to keep us competitive with Linux.
61  */
62 static int xen_blkif_reqs = 64;
63 module_param_named(reqs, xen_blkif_reqs, int, 0);
64 MODULE_PARM_DESC(reqs, "Number of blkback requests to allocate");
65
66 /* Run-time switchable: /sys/module/blkback/parameters/ */
67 static unsigned int log_stats;
68 module_param(log_stats, 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 struct pending_req {
77         struct xen_blkif        *blkif;
78         u64                     id;
79         int                     nr_pages;
80         atomic_t                pendcnt;
81         unsigned short          operation;
82         int                     status;
83         struct list_head        free_list;
84 };
85
86 #define BLKBACK_INVALID_HANDLE (~0)
87
88 struct xen_blkbk {
89         struct pending_req      *pending_reqs;
90         /* List of all 'pending_req' available */
91         struct list_head        pending_free;
92         /* And its spinlock. */
93         spinlock_t              pending_free_lock;
94         wait_queue_head_t       pending_free_wq;
95         /* The list of all pages that are available. */
96         struct page             **pending_pages;
97         /* And the grant handles that are available. */
98         grant_handle_t          *pending_grant_handles;
99 };
100
101 static struct xen_blkbk *blkbk;
102
103 /*
104  * Little helpful macro to figure out the index and virtual address of the
105  * pending_pages[..]. For each 'pending_req' we have have up to
106  * BLKIF_MAX_SEGMENTS_PER_REQUEST (11) pages. The seg would be from 0 through
107  * 10 and would index in the pending_pages[..].
108  */
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 xen_blkif *blkif);
128 static int dispatch_rw_block_io(struct xen_blkif *blkif,
129                                 struct blkif_request *req,
130                                 struct pending_req *pending_req);
131 static void make_response(struct xen_blkif *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  * Routines for managing virtual block devices (vbds).
171  */
172 static int xen_vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
173                              int operation)
174 {
175         struct xen_vbd *vbd = &blkif->vbd;
176         int rc = -EACCES;
177
178         if ((operation != READ) && vbd->readonly)
179                 goto out;
180
181         if (likely(req->nr_sects)) {
182                 blkif_sector_t end = req->sector_number + req->nr_sects;
183
184                 if (unlikely(end < req->sector_number))
185                         goto out;
186                 if (unlikely(end > vbd_sz(vbd)))
187                         goto out;
188         }
189
190         req->dev  = vbd->pdevice;
191         req->bdev = vbd->bdev;
192         rc = 0;
193
194  out:
195         return rc;
196 }
197
198 static void xen_vbd_resize(struct xen_blkif *blkif)
199 {
200         struct xen_vbd *vbd = &blkif->vbd;
201         struct xenbus_transaction xbt;
202         int err;
203         struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
204         unsigned long long new_size = vbd_sz(vbd);
205
206         pr_info(DRV_PFX "VBD Resize: Domid: %d, Device: (%d, %d)\n",
207                 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
208         pr_info(DRV_PFX "VBD Resize: new size %llu\n", new_size);
209         vbd->size = new_size;
210 again:
211         err = xenbus_transaction_start(&xbt);
212         if (err) {
213                 pr_warn(DRV_PFX "Error starting transaction");
214                 return;
215         }
216         err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
217                             (unsigned long long)vbd_sz(vbd));
218         if (err) {
219                 pr_warn(DRV_PFX "Error writing new size");
220                 goto abort;
221         }
222         /*
223          * Write the current state; we will use this to synchronize
224          * the front-end. If the current state is "connected" the
225          * front-end will get the new size information online.
226          */
227         err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
228         if (err) {
229                 pr_warn(DRV_PFX "Error writing the state");
230                 goto abort;
231         }
232
233         err = xenbus_transaction_end(xbt, 0);
234         if (err == -EAGAIN)
235                 goto again;
236         if (err)
237                 pr_warn(DRV_PFX "Error ending transaction");
238         return;
239 abort:
240         xenbus_transaction_end(xbt, 1);
241 }
242
243 /*
244  * Notification from the guest OS.
245  */
246 static void blkif_notify_work(struct xen_blkif *blkif)
247 {
248         blkif->waiting_reqs = 1;
249         wake_up(&blkif->wq);
250 }
251
252 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
253 {
254         blkif_notify_work(dev_id);
255         return IRQ_HANDLED;
256 }
257
258 /*
259  * SCHEDULER FUNCTIONS
260  */
261
262 static void print_stats(struct xen_blkif *blkif)
263 {
264         pr_info("xen-blkback (%s): oo %3d  |  rd %4d  |  wr %4d  |  f %4d"
265                  "  |  ds %4d\n",
266                  current->comm, blkif->st_oo_req,
267                  blkif->st_rd_req, blkif->st_wr_req,
268                  blkif->st_f_req, blkif->st_ds_req);
269         blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
270         blkif->st_rd_req = 0;
271         blkif->st_wr_req = 0;
272         blkif->st_oo_req = 0;
273         blkif->st_ds_req = 0;
274 }
275
276 int xen_blkif_schedule(void *arg)
277 {
278         struct xen_blkif *blkif = arg;
279         struct xen_vbd *vbd = &blkif->vbd;
280         int ret;
281
282         xen_blkif_get(blkif);
283
284         while (!kthread_should_stop()) {
285                 if (try_to_freeze())
286                         continue;
287                 if (unlikely(vbd->size != vbd_sz(vbd)))
288                         xen_vbd_resize(blkif);
289
290                 wait_event_interruptible(
291                         blkif->wq,
292                         blkif->waiting_reqs || kthread_should_stop());
293                 wait_event_interruptible(
294                         blkbk->pending_free_wq,
295                         !list_empty(&blkbk->pending_free) ||
296                         kthread_should_stop());
297
298                 blkif->waiting_reqs = 0;
299                 smp_mb(); /* clear flag *before* checking for work */
300
301                 ret = do_block_io_op(blkif);
302                 if (ret > 0)
303                         blkif->waiting_reqs = 1;
304                 if (ret == -EACCES)
305                         wait_event_interruptible(blkif->shutdown_wq,
306                                                  kthread_should_stop());
307
308                 if (log_stats && time_after(jiffies, blkif->st_print))
309                         print_stats(blkif);
310         }
311
312         if (log_stats)
313                 print_stats(blkif);
314
315         blkif->xenblkd = NULL;
316         xen_blkif_put(blkif);
317
318         return 0;
319 }
320
321 struct seg_buf {
322         unsigned long buf;
323         unsigned int nsec;
324 };
325 /*
326  * Unmap the grant references, and also remove the M2P over-rides
327  * used in the 'pending_req'.
328  */
329 static void xen_blkbk_unmap(struct pending_req *req)
330 {
331         struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
332         unsigned int i, invcount = 0;
333         grant_handle_t handle;
334         int ret;
335
336         for (i = 0; i < req->nr_pages; i++) {
337                 handle = pending_handle(req, i);
338                 if (handle == BLKBACK_INVALID_HANDLE)
339                         continue;
340                 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
341                                     GNTMAP_host_map, handle);
342                 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
343                 invcount++;
344         }
345
346         ret = HYPERVISOR_grant_table_op(
347                 GNTTABOP_unmap_grant_ref, unmap, invcount);
348         BUG_ON(ret);
349         /*
350          * Note, we use invcount, so nr->pages, so we can't index
351          * using vaddr(req, i).
352          */
353         for (i = 0; i < invcount; i++) {
354                 ret = m2p_remove_override(
355                         virt_to_page(unmap[i].host_addr), false);
356                 if (ret) {
357                         pr_alert(DRV_PFX "Failed to remove M2P override for %lx\n",
358                                  (unsigned long)unmap[i].host_addr);
359                         continue;
360                 }
361         }
362 }
363
364 static int xen_blkbk_map(struct blkif_request *req,
365                          struct pending_req *pending_req,
366                          struct seg_buf seg[])
367 {
368         struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
369         int i;
370         int nseg = req->nr_segments;
371         int ret = 0;
372
373         /*
374          * Fill out preq.nr_sects with proper amount of sectors, and setup
375          * assign map[..] with the PFN of the page in our domain with the
376          * corresponding grant reference for each page.
377          */
378         for (i = 0; i < nseg; i++) {
379                 uint32_t flags;
380
381                 flags = GNTMAP_host_map;
382                 if (pending_req->operation != BLKIF_OP_READ)
383                         flags |= GNTMAP_readonly;
384                 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
385                                   req->u.rw.seg[i].gref,
386                                   pending_req->blkif->domid);
387         }
388
389         ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
390         BUG_ON(ret);
391
392         /*
393          * Now swizzle the MFN in our domain with the MFN from the other domain
394          * so that when we access vaddr(pending_req,i) it has the contents of
395          * the page from the other domain.
396          */
397         for (i = 0; i < nseg; i++) {
398                 if (unlikely(map[i].status != 0)) {
399                         pr_debug(DRV_PFX "invalid buffer -- could not remap it\n");
400                         map[i].handle = BLKBACK_INVALID_HANDLE;
401                         ret |= 1;
402                 }
403
404                 pending_handle(pending_req, i) = map[i].handle;
405
406                 if (ret)
407                         continue;
408
409                 ret = m2p_add_override(PFN_DOWN(map[i].dev_bus_addr),
410                         blkbk->pending_page(pending_req, i), NULL);
411                 if (ret) {
412                         pr_alert(DRV_PFX "Failed to install M2P override for %lx (ret: %d)\n",
413                                  (unsigned long)map[i].dev_bus_addr, ret);
414                         /* We could switch over to GNTTABOP_copy */
415                         continue;
416                 }
417
418                 seg[i].buf  = map[i].dev_bus_addr |
419                         (req->u.rw.seg[i].first_sect << 9);
420         }
421         return ret;
422 }
423
424 static void xen_blk_discard(struct xen_blkif *blkif, struct blkif_request *req)
425 {
426         int err = 0;
427         int status = BLKIF_RSP_OKAY;
428         struct block_device *bdev = blkif->vbd.bdev;
429
430         if (blkif->blk_backend_type == BLKIF_BACKEND_PHY)
431                 /* just forward the discard request */
432                 err = blkdev_issue_discard(bdev,
433                                 req->u.discard.sector_number,
434                                 req->u.discard.nr_sectors,
435                                 GFP_KERNEL, 0);
436         else if (blkif->blk_backend_type == BLKIF_BACKEND_FILE) {
437                 /* punch a hole in the backing file */
438                 struct loop_device *lo = bdev->bd_disk->private_data;
439                 struct file *file = lo->lo_backing_file;
440
441                 if (file->f_op->fallocate)
442                         err = file->f_op->fallocate(file,
443                                 FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE,
444                                 req->u.discard.sector_number << 9,
445                                 req->u.discard.nr_sectors << 9);
446                 else
447                         err = -EOPNOTSUPP;
448         } else
449                 err = -EOPNOTSUPP;
450
451         if (err == -EOPNOTSUPP) {
452                 pr_debug(DRV_PFX "discard op failed, not supported\n");
453                 status = BLKIF_RSP_EOPNOTSUPP;
454         } else if (err)
455                 status = BLKIF_RSP_ERROR;
456
457         make_response(blkif, req->id, req->operation, status);
458 }
459
460 static void xen_blk_drain_io(struct xen_blkif *blkif)
461 {
462         atomic_set(&blkif->drain, 1);
463         do {
464                 /* The initial value is one, and one refcnt taken at the
465                  * start of the xen_blkif_schedule thread. */
466                 if (atomic_read(&blkif->refcnt) <= 2)
467                         break;
468                 wait_for_completion_interruptible_timeout(
469                                 &blkif->drain_complete, HZ);
470
471                 if (!atomic_read(&blkif->drain))
472                         break;
473         } while (!kthread_should_stop());
474         atomic_set(&blkif->drain, 0);
475 }
476
477 /*
478  * Completion callback on the bio's. Called as bh->b_end_io()
479  */
480
481 static void __end_block_io_op(struct pending_req *pending_req, int error)
482 {
483         /* An error fails the entire request. */
484         if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
485             (error == -EOPNOTSUPP)) {
486                 pr_debug(DRV_PFX "flush diskcache op failed, not supported\n");
487                 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->blkif->be, 0);
488                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
489         } else if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
490                     (error == -EOPNOTSUPP)) {
491                 pr_debug(DRV_PFX "write barrier op failed, not supported\n");
492                 xen_blkbk_barrier(XBT_NIL, pending_req->blkif->be, 0);
493                 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
494         } else if (error) {
495                 pr_debug(DRV_PFX "Buffer not up-to-date at end of operation,"
496                          " error=%d\n", error);
497                 pending_req->status = BLKIF_RSP_ERROR;
498         }
499
500         /*
501          * If all of the bio's have completed it is time to unmap
502          * the grant references associated with 'request' and provide
503          * the proper response on the ring.
504          */
505         if (atomic_dec_and_test(&pending_req->pendcnt)) {
506                 xen_blkbk_unmap(pending_req);
507                 make_response(pending_req->blkif, pending_req->id,
508                               pending_req->operation, pending_req->status);
509                 xen_blkif_put(pending_req->blkif);
510                 if (atomic_read(&pending_req->blkif->refcnt) <= 2) {
511                         if (atomic_read(&pending_req->blkif->drain))
512                                 complete(&pending_req->blkif->drain_complete);
513                 }
514                 free_req(pending_req);
515         }
516 }
517
518 /*
519  * bio callback.
520  */
521 static void end_block_io_op(struct bio *bio, int error)
522 {
523         __end_block_io_op(bio->bi_private, error);
524         bio_put(bio);
525 }
526
527
528
529 /*
530  * Function to copy the from the ring buffer the 'struct blkif_request'
531  * (which has the sectors we want, number of them, grant references, etc),
532  * and transmute  it to the block API to hand it over to the proper block disk.
533  */
534 static int
535 __do_block_io_op(struct xen_blkif *blkif)
536 {
537         union blkif_back_rings *blk_rings = &blkif->blk_rings;
538         struct blkif_request req;
539         struct pending_req *pending_req;
540         RING_IDX rc, rp;
541         int more_to_do = 0;
542
543         rc = blk_rings->common.req_cons;
544         rp = blk_rings->common.sring->req_prod;
545         rmb(); /* Ensure we see queued requests up to 'rp'. */
546
547         if (RING_REQUEST_PROD_OVERFLOW(&blk_rings->common, rp)) {
548                 rc = blk_rings->common.rsp_prod_pvt;
549                 pr_warn(DRV_PFX "Frontend provided bogus ring requests (%d - %d = %d). Halting ring processing on dev=%04x\n",
550                         rp, rc, rp - rc, blkif->vbd.pdevice);
551                 return -EACCES;
552         }
553         while (rc != rp) {
554
555                 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
556                         break;
557
558                 if (kthread_should_stop()) {
559                         more_to_do = 1;
560                         break;
561                 }
562
563                 pending_req = alloc_req();
564                 if (NULL == pending_req) {
565                         blkif->st_oo_req++;
566                         more_to_do = 1;
567                         break;
568                 }
569
570                 switch (blkif->blk_protocol) {
571                 case BLKIF_PROTOCOL_NATIVE:
572                         memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
573                         break;
574                 case BLKIF_PROTOCOL_X86_32:
575                         blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
576                         break;
577                 case BLKIF_PROTOCOL_X86_64:
578                         blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
579                         break;
580                 default:
581                         BUG();
582                 }
583                 blk_rings->common.req_cons = ++rc; /* before make_response() */
584
585                 /* Apply all sanity checks to /private copy/ of request. */
586                 barrier();
587
588                 if (dispatch_rw_block_io(blkif, &req, pending_req))
589                         break;
590
591                 /* Yield point for this unbounded loop. */
592                 cond_resched();
593         }
594
595         return more_to_do;
596 }
597
598 static int
599 do_block_io_op(struct xen_blkif *blkif)
600 {
601         union blkif_back_rings *blk_rings = &blkif->blk_rings;
602         int more_to_do;
603
604         do {
605                 more_to_do = __do_block_io_op(blkif);
606                 if (more_to_do)
607                         break;
608
609                 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
610         } while (more_to_do);
611
612         return more_to_do;
613 }
614 /*
615  * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
616  * and call the 'submit_bio' to pass it to the underlying storage.
617  */
618 static int dispatch_rw_block_io(struct xen_blkif *blkif,
619                                 struct blkif_request *req,
620                                 struct pending_req *pending_req)
621 {
622         struct phys_req preq;
623         struct seg_buf seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
624         unsigned int nseg;
625         struct bio *bio = NULL;
626         struct bio *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
627         int i, nbio = 0;
628         int operation;
629         struct blk_plug plug;
630         bool drain = false;
631
632         switch (req->operation) {
633         case BLKIF_OP_READ:
634                 blkif->st_rd_req++;
635                 operation = READ;
636                 break;
637         case BLKIF_OP_WRITE:
638                 blkif->st_wr_req++;
639                 operation = WRITE_ODIRECT;
640                 break;
641         case BLKIF_OP_WRITE_BARRIER:
642                 drain = true;
643         case BLKIF_OP_FLUSH_DISKCACHE:
644                 blkif->st_f_req++;
645                 operation = WRITE_FLUSH;
646                 break;
647         case BLKIF_OP_DISCARD:
648                 blkif->st_ds_req++;
649                 operation = REQ_DISCARD;
650                 break;
651         default:
652                 operation = 0; /* make gcc happy */
653                 goto fail_response;
654                 break;
655         }
656
657         /* Check that the number of segments is sane. */
658         nseg = req->nr_segments;
659         if (unlikely(nseg == 0 && operation != WRITE_FLUSH &&
660                                 operation != REQ_DISCARD) ||
661             unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
662                 pr_debug(DRV_PFX "Bad number of segments in request (%d)\n",
663                          nseg);
664                 /* Haven't submitted any bio's yet. */
665                 goto fail_response;
666         }
667
668         preq.dev           = req->handle;
669         preq.sector_number = req->u.rw.sector_number;
670         preq.nr_sects      = 0;
671
672         pending_req->blkif     = blkif;
673         pending_req->id        = req->id;
674         pending_req->operation = req->operation;
675         pending_req->status    = BLKIF_RSP_OKAY;
676         pending_req->nr_pages  = nseg;
677
678         for (i = 0; i < nseg; i++) {
679                 seg[i].nsec = req->u.rw.seg[i].last_sect -
680                         req->u.rw.seg[i].first_sect + 1;
681                 if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
682                     (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
683                         goto fail_response;
684                 preq.nr_sects += seg[i].nsec;
685
686         }
687
688         if (xen_vbd_translate(&preq, blkif, operation) != 0) {
689                 pr_debug(DRV_PFX "access denied: %s of [%llu,%llu] on dev=%04x\n",
690                          operation == READ ? "read" : "write",
691                          preq.sector_number,
692                          preq.sector_number + preq.nr_sects, preq.dev);
693                 goto fail_response;
694         }
695
696         /*
697          * This check _MUST_ be done after xen_vbd_translate as the preq.bdev
698          * is set there.
699          */
700         for (i = 0; i < nseg; i++) {
701                 if (((int)preq.sector_number|(int)seg[i].nsec) &
702                     ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
703                         pr_debug(DRV_PFX "Misaligned I/O request from domain %d",
704                                  blkif->domid);
705                         goto fail_response;
706                 }
707         }
708
709         /* Wait on all outstanding I/O's and once that has been completed
710          * issue the WRITE_FLUSH.
711          */
712         if (drain)
713                 xen_blk_drain_io(pending_req->blkif);
714
715         /*
716          * If we have failed at this point, we need to undo the M2P override,
717          * set gnttab_set_unmap_op on all of the grant references and perform
718          * the hypercall to unmap the grants - that is all done in
719          * xen_blkbk_unmap.
720          */
721         if (operation != REQ_DISCARD && xen_blkbk_map(req, pending_req, seg))
722                 goto fail_flush;
723
724         /*
725          * This corresponding xen_blkif_put is done in __end_block_io_op, or
726          * below (in "!bio") if we are handling a BLKIF_OP_DISCARD.
727          */
728         xen_blkif_get(blkif);
729
730         for (i = 0; i < nseg; i++) {
731                 while ((bio == NULL) ||
732                        (bio_add_page(bio,
733                                      blkbk->pending_page(pending_req, i),
734                                      seg[i].nsec << 9,
735                                      seg[i].buf & ~PAGE_MASK) == 0)) {
736
737                         bio = bio_alloc(GFP_KERNEL, nseg-i);
738                         if (unlikely(bio == NULL))
739                                 goto fail_put_bio;
740
741                         biolist[nbio++] = bio;
742                         bio->bi_bdev    = preq.bdev;
743                         bio->bi_private = pending_req;
744                         bio->bi_end_io  = end_block_io_op;
745                         bio->bi_sector  = preq.sector_number;
746                 }
747
748                 preq.sector_number += seg[i].nsec;
749         }
750
751         /* This will be hit if the operation was a flush or discard. */
752         if (!bio) {
753                 BUG_ON(operation != WRITE_FLUSH && operation != REQ_DISCARD);
754
755                 if (operation == WRITE_FLUSH) {
756                         bio = bio_alloc(GFP_KERNEL, 0);
757                         if (unlikely(bio == NULL))
758                                 goto fail_put_bio;
759
760                         biolist[nbio++] = bio;
761                         bio->bi_bdev    = preq.bdev;
762                         bio->bi_private = pending_req;
763                         bio->bi_end_io  = end_block_io_op;
764                 } else if (operation == REQ_DISCARD) {
765                         xen_blk_discard(blkif, req);
766                         xen_blkif_put(blkif);
767                         free_req(pending_req);
768                         return 0;
769                 }
770         }
771
772         atomic_set(&pending_req->pendcnt, nbio);
773         blk_start_plug(&plug);
774
775         for (i = 0; i < nbio; i++)
776                 submit_bio(operation, biolist[i]);
777
778         /* Let the I/Os go.. */
779         blk_finish_plug(&plug);
780
781         if (operation == READ)
782                 blkif->st_rd_sect += preq.nr_sects;
783         else if (operation & WRITE)
784                 blkif->st_wr_sect += preq.nr_sects;
785
786         return 0;
787
788  fail_flush:
789         xen_blkbk_unmap(pending_req);
790  fail_response:
791         /* Haven't submitted any bio's yet. */
792         make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
793         free_req(pending_req);
794         msleep(1); /* back off a bit */
795         return -EIO;
796
797  fail_put_bio:
798         for (i = 0; i < nbio; i++)
799                 bio_put(biolist[i]);
800         atomic_set(&pending_req->pendcnt, 1);
801         __end_block_io_op(pending_req, -EINVAL);
802         msleep(1); /* back off a bit */
803         return -EIO;
804 }
805
806
807
808 /*
809  * Put a response on the ring on how the operation fared.
810  */
811 static void make_response(struct xen_blkif *blkif, u64 id,
812                           unsigned short op, int st)
813 {
814         struct blkif_response *resp;
815         unsigned long     flags;
816         union blkif_back_rings *blk_rings = &blkif->blk_rings;
817         int notify;
818
819         spin_lock_irqsave(&blkif->blk_ring_lock, flags);
820         /* Place on the response ring for the relevant domain. */
821         switch (blkif->blk_protocol) {
822         case BLKIF_PROTOCOL_NATIVE:
823                 resp = RING_GET_RESPONSE(&blk_rings->native,
824                                          blk_rings->native.rsp_prod_pvt);
825                 break;
826         case BLKIF_PROTOCOL_X86_32:
827                 resp = RING_GET_RESPONSE(&blk_rings->x86_32,
828                                          blk_rings->x86_32.rsp_prod_pvt);
829                 break;
830         case BLKIF_PROTOCOL_X86_64:
831                 resp = RING_GET_RESPONSE(&blk_rings->x86_64,
832                                          blk_rings->x86_64.rsp_prod_pvt);
833                 break;
834         default:
835                 BUG();
836         }
837
838         resp->id        = id;
839         resp->operation = op;
840         resp->status    = st;
841
842         blk_rings->common.rsp_prod_pvt++;
843         RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
844         spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
845         if (notify)
846                 notify_remote_via_irq(blkif->irq);
847 }
848
849 static int __init xen_blkif_init(void)
850 {
851         int i, mmap_pages;
852         int rc = 0;
853
854         if (!xen_pv_domain())
855                 return -ENODEV;
856
857         blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
858         if (!blkbk) {
859                 pr_alert(DRV_PFX "%s: out of memory!\n", __func__);
860                 return -ENOMEM;
861         }
862
863         mmap_pages = xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
864
865         blkbk->pending_reqs          = kzalloc(sizeof(blkbk->pending_reqs[0]) *
866                                         xen_blkif_reqs, GFP_KERNEL);
867         blkbk->pending_grant_handles = kmalloc(sizeof(blkbk->pending_grant_handles[0]) *
868                                         mmap_pages, GFP_KERNEL);
869         blkbk->pending_pages         = kzalloc(sizeof(blkbk->pending_pages[0]) *
870                                         mmap_pages, GFP_KERNEL);
871
872         if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
873             !blkbk->pending_pages) {
874                 rc = -ENOMEM;
875                 goto out_of_memory;
876         }
877
878         for (i = 0; i < mmap_pages; i++) {
879                 blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
880                 blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
881                 if (blkbk->pending_pages[i] == NULL) {
882                         rc = -ENOMEM;
883                         goto out_of_memory;
884                 }
885         }
886         rc = xen_blkif_interface_init();
887         if (rc)
888                 goto failed_init;
889
890         INIT_LIST_HEAD(&blkbk->pending_free);
891         spin_lock_init(&blkbk->pending_free_lock);
892         init_waitqueue_head(&blkbk->pending_free_wq);
893
894         for (i = 0; i < xen_blkif_reqs; i++)
895                 list_add_tail(&blkbk->pending_reqs[i].free_list,
896                               &blkbk->pending_free);
897
898         rc = xen_blkif_xenbus_init();
899         if (rc)
900                 goto failed_init;
901
902         return 0;
903
904  out_of_memory:
905         pr_alert(DRV_PFX "%s: out of memory\n", __func__);
906  failed_init:
907         kfree(blkbk->pending_reqs);
908         kfree(blkbk->pending_grant_handles);
909         if (blkbk->pending_pages) {
910                 for (i = 0; i < mmap_pages; i++) {
911                         if (blkbk->pending_pages[i])
912                                 __free_page(blkbk->pending_pages[i]);
913                 }
914                 kfree(blkbk->pending_pages);
915         }
916         kfree(blkbk);
917         blkbk = NULL;
918         return rc;
919 }
920
921 module_init(xen_blkif_init);
922
923 MODULE_LICENSE("Dual BSD/GPL");
924 MODULE_ALIAS("xen-backend:vbd");