1 /******************************************************************************
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
9 * Copyright (c) 2003-2004, Keir Fraser & Steve Hand
10 * Copyright (c) 2005, Christopher Clark
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:
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:
25 * The above copyright notice and this permission notice shall be included in
26 * all copies or substantial portions of the Software.
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
37 #define pr_fmt(fmt) "xen-blkback: " fmt
39 #include <linux/spinlock.h>
40 #include <linux/kthread.h>
41 #include <linux/list.h>
42 #include <linux/delay.h>
43 #include <linux/freezer.h>
44 #include <linux/bitmap.h>
46 #include <xen/events.h>
49 #include <asm/xen/hypervisor.h>
50 #include <asm/xen/hypercall.h>
51 #include <xen/balloon.h>
52 #include <xen/grant_table.h>
56 * Maximum number of unused free pages to keep in the internal buffer.
57 * Setting this to a value too low will reduce memory used in each backend,
58 * but can have a performance penalty.
60 * A sane value is xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST, but can
61 * be set to a lower value that might degrade performance on some intensive
65 static int xen_blkif_max_buffer_pages = 1024;
66 module_param_named(max_buffer_pages, xen_blkif_max_buffer_pages, int, 0644);
67 MODULE_PARM_DESC(max_buffer_pages,
68 "Maximum number of free pages to keep in each block backend buffer");
71 * Maximum number of grants to map persistently in blkback. For maximum
72 * performance this should be the total numbers of grants that can be used
73 * to fill the ring, but since this might become too high, specially with
74 * the use of indirect descriptors, we set it to a value that provides good
75 * performance without using too much memory.
77 * When the list of persistent grants is full we clean it up using a LRU
81 static int xen_blkif_max_pgrants = 1056;
82 module_param_named(max_persistent_grants, xen_blkif_max_pgrants, int, 0644);
83 MODULE_PARM_DESC(max_persistent_grants,
84 "Maximum number of grants to map persistently");
87 * Maximum number of rings/queues blkback supports, allow as many queues as there
88 * are CPUs if user has not specified a value.
90 unsigned int xenblk_max_queues;
91 module_param_named(max_queues, xenblk_max_queues, uint, 0644);
92 MODULE_PARM_DESC(max_queues,
93 "Maximum number of hardware queues per virtual disk." \
94 "By default it is the number of online CPUs.");
97 * Maximum order of pages to be used for the shared ring between front and
98 * backend, 4KB page granularity is used.
100 unsigned int xen_blkif_max_ring_order = XENBUS_MAX_RING_GRANT_ORDER;
101 module_param_named(max_ring_page_order, xen_blkif_max_ring_order, int, S_IRUGO);
102 MODULE_PARM_DESC(max_ring_page_order, "Maximum order of pages to be used for the shared ring");
104 * The LRU mechanism to clean the lists of persistent grants needs to
105 * be executed periodically. The time interval between consecutive executions
106 * of the purge mechanism is set in ms.
108 #define LRU_INTERVAL 100
111 * When the persistent grants list is full we will remove unused grants
112 * from the list. The percent number of grants to be removed at each LRU
115 #define LRU_PERCENT_CLEAN 5
117 /* Run-time switchable: /sys/module/blkback/parameters/ */
118 static unsigned int log_stats;
119 module_param(log_stats, int, 0644);
121 #define BLKBACK_INVALID_HANDLE (~0)
123 /* Number of free pages to remove on each call to gnttab_free_pages */
124 #define NUM_BATCH_FREE_PAGES 10
126 static inline int get_free_page(struct xen_blkif_ring *ring, struct page **page)
130 spin_lock_irqsave(&ring->free_pages_lock, flags);
131 if (list_empty(&ring->free_pages)) {
132 BUG_ON(ring->free_pages_num != 0);
133 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
134 return gnttab_alloc_pages(1, page);
136 BUG_ON(ring->free_pages_num == 0);
137 page[0] = list_first_entry(&ring->free_pages, struct page, lru);
138 list_del(&page[0]->lru);
139 ring->free_pages_num--;
140 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
145 static inline void put_free_pages(struct xen_blkif_ring *ring, struct page **page,
151 spin_lock_irqsave(&ring->free_pages_lock, flags);
152 for (i = 0; i < num; i++)
153 list_add(&page[i]->lru, &ring->free_pages);
154 ring->free_pages_num += num;
155 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
158 static inline void shrink_free_pagepool(struct xen_blkif_ring *ring, int num)
160 /* Remove requested pages in batches of NUM_BATCH_FREE_PAGES */
161 struct page *page[NUM_BATCH_FREE_PAGES];
162 unsigned int num_pages = 0;
165 spin_lock_irqsave(&ring->free_pages_lock, flags);
166 while (ring->free_pages_num > num) {
167 BUG_ON(list_empty(&ring->free_pages));
168 page[num_pages] = list_first_entry(&ring->free_pages,
170 list_del(&page[num_pages]->lru);
171 ring->free_pages_num--;
172 if (++num_pages == NUM_BATCH_FREE_PAGES) {
173 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
174 gnttab_free_pages(num_pages, page);
175 spin_lock_irqsave(&ring->free_pages_lock, flags);
179 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
181 gnttab_free_pages(num_pages, page);
184 #define vaddr(page) ((unsigned long)pfn_to_kaddr(page_to_pfn(page)))
186 static int do_block_io_op(struct xen_blkif_ring *ring);
187 static int dispatch_rw_block_io(struct xen_blkif_ring *ring,
188 struct blkif_request *req,
189 struct pending_req *pending_req);
190 static void make_response(struct xen_blkif_ring *ring, u64 id,
191 unsigned short op, int st);
193 #define foreach_grant_safe(pos, n, rbtree, node) \
194 for ((pos) = container_of(rb_first((rbtree)), typeof(*(pos)), node), \
195 (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL; \
196 &(pos)->node != NULL; \
197 (pos) = container_of(n, typeof(*(pos)), node), \
198 (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL)
202 * We don't need locking around the persistent grant helpers
203 * because blkback uses a single-thread for each backend, so we
204 * can be sure that this functions will never be called recursively.
206 * The only exception to that is put_persistent_grant, that can be called
207 * from interrupt context (by xen_blkbk_unmap), so we have to use atomic
208 * bit operations to modify the flags of a persistent grant and to count
209 * the number of used grants.
211 static int add_persistent_gnt(struct xen_blkif_ring *ring,
212 struct persistent_gnt *persistent_gnt)
214 struct rb_node **new = NULL, *parent = NULL;
215 struct persistent_gnt *this;
216 struct xen_blkif *blkif = ring->blkif;
218 if (ring->persistent_gnt_c >= xen_blkif_max_pgrants) {
219 if (!blkif->vbd.overflow_max_grants)
220 blkif->vbd.overflow_max_grants = 1;
223 /* Figure out where to put new node */
224 new = &ring->persistent_gnts.rb_node;
226 this = container_of(*new, struct persistent_gnt, node);
229 if (persistent_gnt->gnt < this->gnt)
230 new = &((*new)->rb_left);
231 else if (persistent_gnt->gnt > this->gnt)
232 new = &((*new)->rb_right);
234 pr_alert_ratelimited("trying to add a gref that's already in the tree\n");
239 bitmap_zero(persistent_gnt->flags, PERSISTENT_GNT_FLAGS_SIZE);
240 set_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags);
241 /* Add new node and rebalance tree. */
242 rb_link_node(&(persistent_gnt->node), parent, new);
243 rb_insert_color(&(persistent_gnt->node), &ring->persistent_gnts);
244 ring->persistent_gnt_c++;
245 atomic_inc(&ring->persistent_gnt_in_use);
249 static struct persistent_gnt *get_persistent_gnt(struct xen_blkif_ring *ring,
252 struct persistent_gnt *data;
253 struct rb_node *node = NULL;
255 node = ring->persistent_gnts.rb_node;
257 data = container_of(node, struct persistent_gnt, node);
259 if (gref < data->gnt)
260 node = node->rb_left;
261 else if (gref > data->gnt)
262 node = node->rb_right;
264 if(test_bit(PERSISTENT_GNT_ACTIVE, data->flags)) {
265 pr_alert_ratelimited("requesting a grant already in use\n");
268 set_bit(PERSISTENT_GNT_ACTIVE, data->flags);
269 atomic_inc(&ring->persistent_gnt_in_use);
276 static void put_persistent_gnt(struct xen_blkif_ring *ring,
277 struct persistent_gnt *persistent_gnt)
279 if(!test_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags))
280 pr_alert_ratelimited("freeing a grant already unused\n");
281 set_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags);
282 clear_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags);
283 atomic_dec(&ring->persistent_gnt_in_use);
286 static void free_persistent_gnts(struct xen_blkif_ring *ring, struct rb_root *root,
289 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
290 struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
291 struct persistent_gnt *persistent_gnt;
293 int segs_to_unmap = 0;
294 struct gntab_unmap_queue_data unmap_data;
296 unmap_data.pages = pages;
297 unmap_data.unmap_ops = unmap;
298 unmap_data.kunmap_ops = NULL;
300 foreach_grant_safe(persistent_gnt, n, root, node) {
301 BUG_ON(persistent_gnt->handle ==
302 BLKBACK_INVALID_HANDLE);
303 gnttab_set_unmap_op(&unmap[segs_to_unmap],
304 (unsigned long) pfn_to_kaddr(page_to_pfn(
305 persistent_gnt->page)),
307 persistent_gnt->handle);
309 pages[segs_to_unmap] = persistent_gnt->page;
311 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST ||
312 !rb_next(&persistent_gnt->node)) {
314 unmap_data.count = segs_to_unmap;
315 BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
317 put_free_pages(ring, pages, segs_to_unmap);
321 rb_erase(&persistent_gnt->node, root);
322 kfree(persistent_gnt);
328 void xen_blkbk_unmap_purged_grants(struct work_struct *work)
330 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
331 struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
332 struct persistent_gnt *persistent_gnt;
333 int segs_to_unmap = 0;
334 struct xen_blkif_ring *ring = container_of(work, typeof(*ring), persistent_purge_work);
335 struct gntab_unmap_queue_data unmap_data;
337 unmap_data.pages = pages;
338 unmap_data.unmap_ops = unmap;
339 unmap_data.kunmap_ops = NULL;
341 while(!list_empty(&ring->persistent_purge_list)) {
342 persistent_gnt = list_first_entry(&ring->persistent_purge_list,
343 struct persistent_gnt,
345 list_del(&persistent_gnt->remove_node);
347 gnttab_set_unmap_op(&unmap[segs_to_unmap],
348 vaddr(persistent_gnt->page),
350 persistent_gnt->handle);
352 pages[segs_to_unmap] = persistent_gnt->page;
354 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
355 unmap_data.count = segs_to_unmap;
356 BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
357 put_free_pages(ring, pages, segs_to_unmap);
360 kfree(persistent_gnt);
362 if (segs_to_unmap > 0) {
363 unmap_data.count = segs_to_unmap;
364 BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
365 put_free_pages(ring, pages, segs_to_unmap);
369 static void purge_persistent_gnt(struct xen_blkif_ring *ring)
371 struct persistent_gnt *persistent_gnt;
373 unsigned int num_clean, total;
374 bool scan_used = false, clean_used = false;
375 struct rb_root *root;
377 if (ring->persistent_gnt_c < xen_blkif_max_pgrants ||
378 (ring->persistent_gnt_c == xen_blkif_max_pgrants &&
379 !ring->blkif->vbd.overflow_max_grants)) {
383 if (work_busy(&ring->persistent_purge_work)) {
384 pr_alert_ratelimited("Scheduled work from previous purge is still busy, cannot purge list\n");
388 num_clean = (xen_blkif_max_pgrants / 100) * LRU_PERCENT_CLEAN;
389 num_clean = ring->persistent_gnt_c - xen_blkif_max_pgrants + num_clean;
390 num_clean = min(ring->persistent_gnt_c, num_clean);
391 if ((num_clean == 0) ||
392 (num_clean > (ring->persistent_gnt_c - atomic_read(&ring->persistent_gnt_in_use))))
396 * At this point, we can assure that there will be no calls
397 * to get_persistent_grant (because we are executing this code from
398 * xen_blkif_schedule), there can only be calls to put_persistent_gnt,
399 * which means that the number of currently used grants will go down,
400 * but never up, so we will always be able to remove the requested
406 pr_debug("Going to purge %u persistent grants\n", num_clean);
408 BUG_ON(!list_empty(&ring->persistent_purge_list));
409 root = &ring->persistent_gnts;
411 foreach_grant_safe(persistent_gnt, n, root, node) {
412 BUG_ON(persistent_gnt->handle ==
413 BLKBACK_INVALID_HANDLE);
416 clear_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags);
420 if (test_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags))
423 (test_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags)))
426 rb_erase(&persistent_gnt->node, root);
427 list_add(&persistent_gnt->remove_node,
428 &ring->persistent_purge_list);
429 if (--num_clean == 0)
433 * If we get here it means we also need to start cleaning
434 * grants that were used since last purge in order to cope
435 * with the requested num
437 if (!scan_used && !clean_used) {
438 pr_debug("Still missing %u purged frames\n", num_clean);
444 pr_debug("Finished scanning for grants to clean, removing used flag\n");
449 ring->persistent_gnt_c -= (total - num_clean);
450 ring->blkif->vbd.overflow_max_grants = 0;
452 /* We can defer this work */
453 schedule_work(&ring->persistent_purge_work);
454 pr_debug("Purged %u/%u\n", (total - num_clean), total);
461 * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
463 static struct pending_req *alloc_req(struct xen_blkif_ring *ring)
465 struct pending_req *req = NULL;
468 spin_lock_irqsave(&ring->pending_free_lock, flags);
469 if (!list_empty(&ring->pending_free)) {
470 req = list_entry(ring->pending_free.next, struct pending_req,
472 list_del(&req->free_list);
474 spin_unlock_irqrestore(&ring->pending_free_lock, flags);
479 * Return the 'pending_req' structure back to the freepool. We also
480 * wake up the thread if it was waiting for a free page.
482 static void free_req(struct xen_blkif_ring *ring, struct pending_req *req)
487 spin_lock_irqsave(&ring->pending_free_lock, flags);
488 was_empty = list_empty(&ring->pending_free);
489 list_add(&req->free_list, &ring->pending_free);
490 spin_unlock_irqrestore(&ring->pending_free_lock, flags);
492 wake_up(&ring->pending_free_wq);
496 * Routines for managing virtual block devices (vbds).
498 static int xen_vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
501 struct xen_vbd *vbd = &blkif->vbd;
504 if ((operation != READ) && vbd->readonly)
507 if (likely(req->nr_sects)) {
508 blkif_sector_t end = req->sector_number + req->nr_sects;
510 if (unlikely(end < req->sector_number))
512 if (unlikely(end > vbd_sz(vbd)))
516 req->dev = vbd->pdevice;
517 req->bdev = vbd->bdev;
524 static void xen_vbd_resize(struct xen_blkif *blkif)
526 struct xen_vbd *vbd = &blkif->vbd;
527 struct xenbus_transaction xbt;
529 struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
530 unsigned long long new_size = vbd_sz(vbd);
532 pr_info("VBD Resize: Domid: %d, Device: (%d, %d)\n",
533 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
534 pr_info("VBD Resize: new size %llu\n", new_size);
535 vbd->size = new_size;
537 err = xenbus_transaction_start(&xbt);
539 pr_warn("Error starting transaction\n");
542 err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
543 (unsigned long long)vbd_sz(vbd));
545 pr_warn("Error writing new size\n");
549 * Write the current state; we will use this to synchronize
550 * the front-end. If the current state is "connected" the
551 * front-end will get the new size information online.
553 err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
555 pr_warn("Error writing the state\n");
559 err = xenbus_transaction_end(xbt, 0);
563 pr_warn("Error ending transaction\n");
566 xenbus_transaction_end(xbt, 1);
570 * Notification from the guest OS.
572 static void blkif_notify_work(struct xen_blkif_ring *ring)
574 ring->waiting_reqs = 1;
578 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
580 blkif_notify_work(dev_id);
585 * SCHEDULER FUNCTIONS
588 static void print_stats(struct xen_blkif_ring *ring)
590 struct xen_blkif *blkif = ring->blkif;
592 pr_info("(%s): oo %3llu | rd %4llu | wr %4llu | f %4llu"
593 " | ds %4llu | pg: %4u/%4d\n",
594 current->comm, blkif->st_oo_req,
595 blkif->st_rd_req, blkif->st_wr_req,
596 blkif->st_f_req, blkif->st_ds_req,
597 ring->persistent_gnt_c,
598 xen_blkif_max_pgrants);
599 blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
600 blkif->st_rd_req = 0;
601 blkif->st_wr_req = 0;
602 blkif->st_oo_req = 0;
603 blkif->st_ds_req = 0;
606 int xen_blkif_schedule(void *arg)
608 struct xen_blkif_ring *ring = arg;
609 struct xen_blkif *blkif = ring->blkif;
610 struct xen_vbd *vbd = &blkif->vbd;
611 unsigned long timeout;
614 xen_blkif_get(blkif);
617 while (!kthread_should_stop()) {
620 if (unlikely(vbd->size != vbd_sz(vbd)))
621 xen_vbd_resize(blkif);
623 timeout = msecs_to_jiffies(LRU_INTERVAL);
625 timeout = wait_event_interruptible_timeout(
627 ring->waiting_reqs || kthread_should_stop(),
631 timeout = wait_event_interruptible_timeout(
632 ring->pending_free_wq,
633 !list_empty(&ring->pending_free) ||
634 kthread_should_stop(),
639 ring->waiting_reqs = 0;
640 smp_mb(); /* clear flag *before* checking for work */
642 ret = do_block_io_op(ring);
644 ring->waiting_reqs = 1;
646 wait_event_interruptible(ring->shutdown_wq,
647 kthread_should_stop());
650 if (blkif->vbd.feature_gnt_persistent &&
651 time_after(jiffies, ring->next_lru)) {
652 purge_persistent_gnt(ring);
653 ring->next_lru = jiffies + msecs_to_jiffies(LRU_INTERVAL);
656 /* Shrink if we have more than xen_blkif_max_buffer_pages */
657 shrink_free_pagepool(ring, xen_blkif_max_buffer_pages);
659 if (log_stats && time_after(jiffies, ring->blkif->st_print))
663 /* Drain pending purge work */
664 flush_work(&ring->persistent_purge_work);
669 ring->xenblkd = NULL;
670 xen_blkif_put(blkif);
676 * Remove persistent grants and empty the pool of free pages
678 void xen_blkbk_free_caches(struct xen_blkif_ring *ring)
680 /* Free all persistent grant pages */
681 if (!RB_EMPTY_ROOT(&ring->persistent_gnts))
682 free_persistent_gnts(ring, &ring->persistent_gnts,
683 ring->persistent_gnt_c);
685 BUG_ON(!RB_EMPTY_ROOT(&ring->persistent_gnts));
686 ring->persistent_gnt_c = 0;
688 /* Since we are shutting down remove all pages from the buffer */
689 shrink_free_pagepool(ring, 0 /* All */);
692 static unsigned int xen_blkbk_unmap_prepare(
693 struct xen_blkif_ring *ring,
694 struct grant_page **pages,
696 struct gnttab_unmap_grant_ref *unmap_ops,
697 struct page **unmap_pages)
699 unsigned int i, invcount = 0;
701 for (i = 0; i < num; i++) {
702 if (pages[i]->persistent_gnt != NULL) {
703 put_persistent_gnt(ring, pages[i]->persistent_gnt);
706 if (pages[i]->handle == BLKBACK_INVALID_HANDLE)
708 unmap_pages[invcount] = pages[i]->page;
709 gnttab_set_unmap_op(&unmap_ops[invcount], vaddr(pages[i]->page),
710 GNTMAP_host_map, pages[i]->handle);
711 pages[i]->handle = BLKBACK_INVALID_HANDLE;
718 static void xen_blkbk_unmap_and_respond_callback(int result, struct gntab_unmap_queue_data *data)
720 struct pending_req *pending_req = (struct pending_req *)(data->data);
721 struct xen_blkif_ring *ring = pending_req->ring;
722 struct xen_blkif *blkif = ring->blkif;
724 /* BUG_ON used to reproduce existing behaviour,
725 but is this the best way to deal with this? */
728 put_free_pages(ring, data->pages, data->count);
729 make_response(ring, pending_req->id,
730 pending_req->operation, pending_req->status);
731 free_req(ring, pending_req);
733 * Make sure the request is freed before releasing blkif,
734 * or there could be a race between free_req and the
735 * cleanup done in xen_blkif_free during shutdown.
737 * NB: The fact that we might try to wake up pending_free_wq
738 * before drain_complete (in case there's a drain going on)
739 * it's not a problem with our current implementation
740 * because we can assure there's no thread waiting on
741 * pending_free_wq if there's a drain going on, but it has
742 * to be taken into account if the current model is changed.
744 if (atomic_dec_and_test(&ring->inflight) && atomic_read(&blkif->drain)) {
745 complete(&blkif->drain_complete);
747 xen_blkif_put(blkif);
750 static void xen_blkbk_unmap_and_respond(struct pending_req *req)
752 struct gntab_unmap_queue_data* work = &req->gnttab_unmap_data;
753 struct xen_blkif_ring *ring = req->ring;
754 struct grant_page **pages = req->segments;
755 unsigned int invcount;
757 invcount = xen_blkbk_unmap_prepare(ring, pages, req->nr_segs,
758 req->unmap, req->unmap_pages);
761 work->done = xen_blkbk_unmap_and_respond_callback;
762 work->unmap_ops = req->unmap;
763 work->kunmap_ops = NULL;
764 work->pages = req->unmap_pages;
765 work->count = invcount;
767 gnttab_unmap_refs_async(&req->gnttab_unmap_data);
772 * Unmap the grant references.
774 * This could accumulate ops up to the batch size to reduce the number
775 * of hypercalls, but since this is only used in error paths there's
778 static void xen_blkbk_unmap(struct xen_blkif_ring *ring,
779 struct grant_page *pages[],
782 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
783 struct page *unmap_pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
784 unsigned int invcount = 0;
788 unsigned int batch = min(num, BLKIF_MAX_SEGMENTS_PER_REQUEST);
790 invcount = xen_blkbk_unmap_prepare(ring, pages, batch,
793 ret = gnttab_unmap_refs(unmap, NULL, unmap_pages, invcount);
795 put_free_pages(ring, unmap_pages, invcount);
802 static int xen_blkbk_map(struct xen_blkif_ring *ring,
803 struct grant_page *pages[],
806 struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
807 struct page *pages_to_gnt[BLKIF_MAX_SEGMENTS_PER_REQUEST];
808 struct persistent_gnt *persistent_gnt = NULL;
809 phys_addr_t addr = 0;
810 int i, seg_idx, new_map_idx;
813 int last_map = 0, map_until = 0;
814 int use_persistent_gnts;
815 struct xen_blkif *blkif = ring->blkif;
817 use_persistent_gnts = (blkif->vbd.feature_gnt_persistent);
820 * Fill out preq.nr_sects with proper amount of sectors, and setup
821 * assign map[..] with the PFN of the page in our domain with the
822 * corresponding grant reference for each page.
825 for (i = map_until; i < num; i++) {
828 if (use_persistent_gnts) {
829 persistent_gnt = get_persistent_gnt(
834 if (persistent_gnt) {
836 * We are using persistent grants and
837 * the grant is already mapped
839 pages[i]->page = persistent_gnt->page;
840 pages[i]->persistent_gnt = persistent_gnt;
842 if (get_free_page(ring, &pages[i]->page))
844 addr = vaddr(pages[i]->page);
845 pages_to_gnt[segs_to_map] = pages[i]->page;
846 pages[i]->persistent_gnt = NULL;
847 flags = GNTMAP_host_map;
848 if (!use_persistent_gnts && ro)
849 flags |= GNTMAP_readonly;
850 gnttab_set_map_op(&map[segs_to_map++], addr,
851 flags, pages[i]->gref,
855 if (segs_to_map == BLKIF_MAX_SEGMENTS_PER_REQUEST)
860 ret = gnttab_map_refs(map, NULL, pages_to_gnt, segs_to_map);
865 * Now swizzle the MFN in our domain with the MFN from the other domain
866 * so that when we access vaddr(pending_req,i) it has the contents of
867 * the page from the other domain.
869 for (seg_idx = last_map, new_map_idx = 0; seg_idx < map_until; seg_idx++) {
870 if (!pages[seg_idx]->persistent_gnt) {
871 /* This is a newly mapped grant */
872 BUG_ON(new_map_idx >= segs_to_map);
873 if (unlikely(map[new_map_idx].status != 0)) {
874 pr_debug("invalid buffer -- could not remap it\n");
875 put_free_pages(ring, &pages[seg_idx]->page, 1);
876 pages[seg_idx]->handle = BLKBACK_INVALID_HANDLE;
880 pages[seg_idx]->handle = map[new_map_idx].handle;
884 if (use_persistent_gnts &&
885 ring->persistent_gnt_c < xen_blkif_max_pgrants) {
887 * We are using persistent grants, the grant is
888 * not mapped but we might have room for it.
890 persistent_gnt = kmalloc(sizeof(struct persistent_gnt),
892 if (!persistent_gnt) {
894 * If we don't have enough memory to
895 * allocate the persistent_gnt struct
896 * map this grant non-persistenly
900 persistent_gnt->gnt = map[new_map_idx].ref;
901 persistent_gnt->handle = map[new_map_idx].handle;
902 persistent_gnt->page = pages[seg_idx]->page;
903 if (add_persistent_gnt(ring,
905 kfree(persistent_gnt);
906 persistent_gnt = NULL;
909 pages[seg_idx]->persistent_gnt = persistent_gnt;
910 pr_debug("grant %u added to the tree of persistent grants, using %u/%u\n",
911 persistent_gnt->gnt, ring->persistent_gnt_c,
912 xen_blkif_max_pgrants);
915 if (use_persistent_gnts && !blkif->vbd.overflow_max_grants) {
916 blkif->vbd.overflow_max_grants = 1;
917 pr_debug("domain %u, device %#x is using maximum number of persistent grants\n",
918 blkif->domid, blkif->vbd.handle);
921 * We could not map this grant persistently, so use it as
922 * a non-persistent grant.
928 last_map = map_until;
929 if (map_until != num)
935 pr_alert("%s: out of memory\n", __func__);
936 put_free_pages(ring, pages_to_gnt, segs_to_map);
940 static int xen_blkbk_map_seg(struct pending_req *pending_req)
944 rc = xen_blkbk_map(pending_req->ring, pending_req->segments,
945 pending_req->nr_segs,
946 (pending_req->operation != BLKIF_OP_READ));
951 static int xen_blkbk_parse_indirect(struct blkif_request *req,
952 struct pending_req *pending_req,
953 struct seg_buf seg[],
954 struct phys_req *preq)
956 struct grant_page **pages = pending_req->indirect_pages;
957 struct xen_blkif_ring *ring = pending_req->ring;
958 int indirect_grefs, rc, n, nseg, i;
959 struct blkif_request_segment *segments = NULL;
961 nseg = pending_req->nr_segs;
962 indirect_grefs = INDIRECT_PAGES(nseg);
963 BUG_ON(indirect_grefs > BLKIF_MAX_INDIRECT_PAGES_PER_REQUEST);
965 for (i = 0; i < indirect_grefs; i++)
966 pages[i]->gref = req->u.indirect.indirect_grefs[i];
968 rc = xen_blkbk_map(ring, pages, indirect_grefs, true);
972 for (n = 0, i = 0; n < nseg; n++) {
973 if ((n % SEGS_PER_INDIRECT_FRAME) == 0) {
974 /* Map indirect segments */
976 kunmap_atomic(segments);
977 segments = kmap_atomic(pages[n/SEGS_PER_INDIRECT_FRAME]->page);
979 i = n % SEGS_PER_INDIRECT_FRAME;
980 pending_req->segments[n]->gref = segments[i].gref;
981 seg[n].nsec = segments[i].last_sect -
982 segments[i].first_sect + 1;
983 seg[n].offset = (segments[i].first_sect << 9);
984 if ((segments[i].last_sect >= (XEN_PAGE_SIZE >> 9)) ||
985 (segments[i].last_sect < segments[i].first_sect)) {
989 preq->nr_sects += seg[n].nsec;
994 kunmap_atomic(segments);
995 xen_blkbk_unmap(ring, pages, indirect_grefs);
999 static int dispatch_discard_io(struct xen_blkif_ring *ring,
1000 struct blkif_request *req)
1003 int status = BLKIF_RSP_OKAY;
1004 struct xen_blkif *blkif = ring->blkif;
1005 struct block_device *bdev = blkif->vbd.bdev;
1006 unsigned long secure;
1007 struct phys_req preq;
1009 xen_blkif_get(blkif);
1011 preq.sector_number = req->u.discard.sector_number;
1012 preq.nr_sects = req->u.discard.nr_sectors;
1014 err = xen_vbd_translate(&preq, blkif, WRITE);
1016 pr_warn("access denied: DISCARD [%llu->%llu] on dev=%04x\n",
1018 preq.sector_number + preq.nr_sects, blkif->vbd.pdevice);
1023 secure = (blkif->vbd.discard_secure &&
1024 (req->u.discard.flag & BLKIF_DISCARD_SECURE)) ?
1025 BLKDEV_DISCARD_SECURE : 0;
1027 err = blkdev_issue_discard(bdev, req->u.discard.sector_number,
1028 req->u.discard.nr_sectors,
1029 GFP_KERNEL, secure);
1031 if (err == -EOPNOTSUPP) {
1032 pr_debug("discard op failed, not supported\n");
1033 status = BLKIF_RSP_EOPNOTSUPP;
1035 status = BLKIF_RSP_ERROR;
1037 make_response(ring, req->u.discard.id, req->operation, status);
1038 xen_blkif_put(blkif);
1042 static int dispatch_other_io(struct xen_blkif_ring *ring,
1043 struct blkif_request *req,
1044 struct pending_req *pending_req)
1046 free_req(ring, pending_req);
1047 make_response(ring, req->u.other.id, req->operation,
1048 BLKIF_RSP_EOPNOTSUPP);
1052 static void xen_blk_drain_io(struct xen_blkif_ring *ring)
1054 struct xen_blkif *blkif = ring->blkif;
1056 atomic_set(&blkif->drain, 1);
1058 if (atomic_read(&ring->inflight) == 0)
1060 wait_for_completion_interruptible_timeout(
1061 &blkif->drain_complete, HZ);
1063 if (!atomic_read(&blkif->drain))
1065 } while (!kthread_should_stop());
1066 atomic_set(&blkif->drain, 0);
1070 * Completion callback on the bio's. Called as bh->b_end_io()
1073 static void __end_block_io_op(struct pending_req *pending_req, int error)
1075 /* An error fails the entire request. */
1076 if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
1077 (error == -EOPNOTSUPP)) {
1078 pr_debug("flush diskcache op failed, not supported\n");
1079 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->ring->blkif->be, 0);
1080 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
1081 } else if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
1082 (error == -EOPNOTSUPP)) {
1083 pr_debug("write barrier op failed, not supported\n");
1084 xen_blkbk_barrier(XBT_NIL, pending_req->ring->blkif->be, 0);
1085 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
1087 pr_debug("Buffer not up-to-date at end of operation,"
1088 " error=%d\n", error);
1089 pending_req->status = BLKIF_RSP_ERROR;
1093 * If all of the bio's have completed it is time to unmap
1094 * the grant references associated with 'request' and provide
1095 * the proper response on the ring.
1097 if (atomic_dec_and_test(&pending_req->pendcnt))
1098 xen_blkbk_unmap_and_respond(pending_req);
1104 static void end_block_io_op(struct bio *bio)
1106 __end_block_io_op(bio->bi_private, bio->bi_error);
1113 * Function to copy the from the ring buffer the 'struct blkif_request'
1114 * (which has the sectors we want, number of them, grant references, etc),
1115 * and transmute it to the block API to hand it over to the proper block disk.
1118 __do_block_io_op(struct xen_blkif_ring *ring)
1120 union blkif_back_rings *blk_rings = &ring->blk_rings;
1121 struct blkif_request req;
1122 struct pending_req *pending_req;
1126 rc = blk_rings->common.req_cons;
1127 rp = blk_rings->common.sring->req_prod;
1128 rmb(); /* Ensure we see queued requests up to 'rp'. */
1130 if (RING_REQUEST_PROD_OVERFLOW(&blk_rings->common, rp)) {
1131 rc = blk_rings->common.rsp_prod_pvt;
1132 pr_warn("Frontend provided bogus ring requests (%d - %d = %d). Halting ring processing on dev=%04x\n",
1133 rp, rc, rp - rc, ring->blkif->vbd.pdevice);
1138 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
1141 if (kthread_should_stop()) {
1146 pending_req = alloc_req(ring);
1147 if (NULL == pending_req) {
1148 ring->blkif->st_oo_req++;
1153 switch (ring->blkif->blk_protocol) {
1154 case BLKIF_PROTOCOL_NATIVE:
1155 memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
1157 case BLKIF_PROTOCOL_X86_32:
1158 blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
1160 case BLKIF_PROTOCOL_X86_64:
1161 blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
1166 blk_rings->common.req_cons = ++rc; /* before make_response() */
1168 /* Apply all sanity checks to /private copy/ of request. */
1171 switch (req.operation) {
1173 case BLKIF_OP_WRITE:
1174 case BLKIF_OP_WRITE_BARRIER:
1175 case BLKIF_OP_FLUSH_DISKCACHE:
1176 case BLKIF_OP_INDIRECT:
1177 if (dispatch_rw_block_io(ring, &req, pending_req))
1180 case BLKIF_OP_DISCARD:
1181 free_req(ring, pending_req);
1182 if (dispatch_discard_io(ring, &req))
1186 if (dispatch_other_io(ring, &req, pending_req))
1191 /* Yield point for this unbounded loop. */
1199 do_block_io_op(struct xen_blkif_ring *ring)
1201 union blkif_back_rings *blk_rings = &ring->blk_rings;
1205 more_to_do = __do_block_io_op(ring);
1209 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
1210 } while (more_to_do);
1215 * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
1216 * and call the 'submit_bio' to pass it to the underlying storage.
1218 static int dispatch_rw_block_io(struct xen_blkif_ring *ring,
1219 struct blkif_request *req,
1220 struct pending_req *pending_req)
1222 struct phys_req preq;
1223 struct seg_buf *seg = pending_req->seg;
1225 struct bio *bio = NULL;
1226 struct bio **biolist = pending_req->biolist;
1229 struct blk_plug plug;
1231 struct grant_page **pages = pending_req->segments;
1232 unsigned short req_operation;
1234 req_operation = req->operation == BLKIF_OP_INDIRECT ?
1235 req->u.indirect.indirect_op : req->operation;
1237 if ((req->operation == BLKIF_OP_INDIRECT) &&
1238 (req_operation != BLKIF_OP_READ) &&
1239 (req_operation != BLKIF_OP_WRITE)) {
1240 pr_debug("Invalid indirect operation (%u)\n", req_operation);
1244 switch (req_operation) {
1246 ring->blkif->st_rd_req++;
1249 case BLKIF_OP_WRITE:
1250 ring->blkif->st_wr_req++;
1251 operation = WRITE_ODIRECT;
1253 case BLKIF_OP_WRITE_BARRIER:
1255 case BLKIF_OP_FLUSH_DISKCACHE:
1256 ring->blkif->st_f_req++;
1257 operation = WRITE_FLUSH;
1260 operation = 0; /* make gcc happy */
1265 /* Check that the number of segments is sane. */
1266 nseg = req->operation == BLKIF_OP_INDIRECT ?
1267 req->u.indirect.nr_segments : req->u.rw.nr_segments;
1269 if (unlikely(nseg == 0 && operation != WRITE_FLUSH) ||
1270 unlikely((req->operation != BLKIF_OP_INDIRECT) &&
1271 (nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) ||
1272 unlikely((req->operation == BLKIF_OP_INDIRECT) &&
1273 (nseg > MAX_INDIRECT_SEGMENTS))) {
1274 pr_debug("Bad number of segments in request (%d)\n", nseg);
1275 /* Haven't submitted any bio's yet. */
1281 pending_req->ring = ring;
1282 pending_req->id = req->u.rw.id;
1283 pending_req->operation = req_operation;
1284 pending_req->status = BLKIF_RSP_OKAY;
1285 pending_req->nr_segs = nseg;
1287 if (req->operation != BLKIF_OP_INDIRECT) {
1288 preq.dev = req->u.rw.handle;
1289 preq.sector_number = req->u.rw.sector_number;
1290 for (i = 0; i < nseg; i++) {
1291 pages[i]->gref = req->u.rw.seg[i].gref;
1292 seg[i].nsec = req->u.rw.seg[i].last_sect -
1293 req->u.rw.seg[i].first_sect + 1;
1294 seg[i].offset = (req->u.rw.seg[i].first_sect << 9);
1295 if ((req->u.rw.seg[i].last_sect >= (XEN_PAGE_SIZE >> 9)) ||
1296 (req->u.rw.seg[i].last_sect <
1297 req->u.rw.seg[i].first_sect))
1299 preq.nr_sects += seg[i].nsec;
1302 preq.dev = req->u.indirect.handle;
1303 preq.sector_number = req->u.indirect.sector_number;
1304 if (xen_blkbk_parse_indirect(req, pending_req, seg, &preq))
1308 if (xen_vbd_translate(&preq, ring->blkif, operation) != 0) {
1309 pr_debug("access denied: %s of [%llu,%llu] on dev=%04x\n",
1310 operation == READ ? "read" : "write",
1312 preq.sector_number + preq.nr_sects,
1313 ring->blkif->vbd.pdevice);
1318 * This check _MUST_ be done after xen_vbd_translate as the preq.bdev
1321 for (i = 0; i < nseg; i++) {
1322 if (((int)preq.sector_number|(int)seg[i].nsec) &
1323 ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
1324 pr_debug("Misaligned I/O request from domain %d\n",
1325 ring->blkif->domid);
1330 /* Wait on all outstanding I/O's and once that has been completed
1331 * issue the WRITE_FLUSH.
1334 xen_blk_drain_io(pending_req->ring);
1337 * If we have failed at this point, we need to undo the M2P override,
1338 * set gnttab_set_unmap_op on all of the grant references and perform
1339 * the hypercall to unmap the grants - that is all done in
1342 if (xen_blkbk_map_seg(pending_req))
1346 * This corresponding xen_blkif_put is done in __end_block_io_op, or
1347 * below (in "!bio") if we are handling a BLKIF_OP_DISCARD.
1349 xen_blkif_get(ring->blkif);
1350 atomic_inc(&ring->inflight);
1352 for (i = 0; i < nseg; i++) {
1353 while ((bio == NULL) ||
1357 seg[i].offset) == 0)) {
1359 int nr_iovecs = min_t(int, (nseg-i), BIO_MAX_PAGES);
1360 bio = bio_alloc(GFP_KERNEL, nr_iovecs);
1361 if (unlikely(bio == NULL))
1364 biolist[nbio++] = bio;
1365 bio->bi_bdev = preq.bdev;
1366 bio->bi_private = pending_req;
1367 bio->bi_end_io = end_block_io_op;
1368 bio->bi_iter.bi_sector = preq.sector_number;
1371 preq.sector_number += seg[i].nsec;
1374 /* This will be hit if the operation was a flush or discard. */
1376 BUG_ON(operation != WRITE_FLUSH);
1378 bio = bio_alloc(GFP_KERNEL, 0);
1379 if (unlikely(bio == NULL))
1382 biolist[nbio++] = bio;
1383 bio->bi_bdev = preq.bdev;
1384 bio->bi_private = pending_req;
1385 bio->bi_end_io = end_block_io_op;
1388 atomic_set(&pending_req->pendcnt, nbio);
1389 blk_start_plug(&plug);
1391 for (i = 0; i < nbio; i++)
1392 submit_bio(operation, biolist[i]);
1394 /* Let the I/Os go.. */
1395 blk_finish_plug(&plug);
1397 if (operation == READ)
1398 ring->blkif->st_rd_sect += preq.nr_sects;
1399 else if (operation & WRITE)
1400 ring->blkif->st_wr_sect += preq.nr_sects;
1405 xen_blkbk_unmap(ring, pending_req->segments,
1406 pending_req->nr_segs);
1408 /* Haven't submitted any bio's yet. */
1409 make_response(ring, req->u.rw.id, req_operation, BLKIF_RSP_ERROR);
1410 free_req(ring, pending_req);
1411 msleep(1); /* back off a bit */
1415 for (i = 0; i < nbio; i++)
1416 bio_put(biolist[i]);
1417 atomic_set(&pending_req->pendcnt, 1);
1418 __end_block_io_op(pending_req, -EINVAL);
1419 msleep(1); /* back off a bit */
1426 * Put a response on the ring on how the operation fared.
1428 static void make_response(struct xen_blkif_ring *ring, u64 id,
1429 unsigned short op, int st)
1431 struct blkif_response resp;
1432 unsigned long flags;
1433 union blkif_back_rings *blk_rings;
1437 resp.operation = op;
1440 spin_lock_irqsave(&ring->blk_ring_lock, flags);
1441 blk_rings = &ring->blk_rings;
1442 /* Place on the response ring for the relevant domain. */
1443 switch (ring->blkif->blk_protocol) {
1444 case BLKIF_PROTOCOL_NATIVE:
1445 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
1446 &resp, sizeof(resp));
1448 case BLKIF_PROTOCOL_X86_32:
1449 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
1450 &resp, sizeof(resp));
1452 case BLKIF_PROTOCOL_X86_64:
1453 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
1454 &resp, sizeof(resp));
1459 blk_rings->common.rsp_prod_pvt++;
1460 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
1461 spin_unlock_irqrestore(&ring->blk_ring_lock, flags);
1463 notify_remote_via_irq(ring->irq);
1466 static int __init xen_blkif_init(void)
1473 if (xen_blkif_max_ring_order > XENBUS_MAX_RING_GRANT_ORDER) {
1474 pr_info("Invalid max_ring_order (%d), will use default max: %d.\n",
1475 xen_blkif_max_ring_order, XENBUS_MAX_RING_GRANT_ORDER);
1476 xen_blkif_max_ring_order = XENBUS_MAX_RING_GRANT_ORDER;
1479 if (xenblk_max_queues == 0)
1480 xenblk_max_queues = num_online_cpus();
1482 rc = xen_blkif_interface_init();
1486 rc = xen_blkif_xenbus_init();
1494 module_init(xen_blkif_init);
1496 MODULE_LICENSE("Dual BSD/GPL");
1497 MODULE_ALIAS("xen-backend:vbd");