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 max_buffer_pages = 1024;
66 module_param_named(max_buffer_pages, 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 max_pgrants = 1056;
82 module_param_named(max_persistent_grants, max_pgrants, int, 0644);
83 MODULE_PARM_DESC(max_persistent_grants,
84 "Maximum number of grants to map persistently");
87 * How long a persistent grant is allowed to remain allocated without being in
88 * use. The time is in seconds, 0 means indefinitely long.
91 static unsigned int pgrant_timeout = 60;
92 module_param_named(persistent_grant_unused_seconds, pgrant_timeout,
94 MODULE_PARM_DESC(persistent_grant_unused_seconds,
95 "Time in seconds an unused persistent grant is allowed to "
96 "remain allocated. Default is 60, 0 means unlimited.");
99 * Maximum number of rings/queues blkback supports, allow as many queues as there
100 * are CPUs if user has not specified a value.
102 unsigned int xenblk_max_queues;
103 module_param_named(max_queues, xenblk_max_queues, uint, 0644);
104 MODULE_PARM_DESC(max_queues,
105 "Maximum number of hardware queues per virtual disk." \
106 "By default it is the number of online CPUs.");
109 * Maximum order of pages to be used for the shared ring between front and
110 * backend, 4KB page granularity is used.
112 unsigned int xen_blkif_max_ring_order = XENBUS_MAX_RING_GRANT_ORDER;
113 module_param_named(max_ring_page_order, xen_blkif_max_ring_order, int, 0444);
114 MODULE_PARM_DESC(max_ring_page_order, "Maximum order of pages to be used for the shared ring");
116 * The LRU mechanism to clean the lists of persistent grants needs to
117 * be executed periodically. The time interval between consecutive executions
118 * of the purge mechanism is set in ms.
120 #define LRU_INTERVAL 100
123 * When the persistent grants list is full we will remove unused grants
124 * from the list. The percent number of grants to be removed at each LRU
127 #define LRU_PERCENT_CLEAN 5
129 /* Run-time switchable: /sys/module/blkback/parameters/ */
130 static unsigned int log_stats;
131 module_param(log_stats, int, 0644);
133 #define BLKBACK_INVALID_HANDLE (~0)
135 /* Number of free pages to remove on each call to gnttab_free_pages */
136 #define NUM_BATCH_FREE_PAGES 10
138 static inline bool persistent_gnt_timeout(struct persistent_gnt *persistent_gnt)
140 return pgrant_timeout && (jiffies - persistent_gnt->last_used >=
141 HZ * pgrant_timeout);
144 static inline int get_free_page(struct xen_blkif_ring *ring, struct page **page)
148 spin_lock_irqsave(&ring->free_pages_lock, flags);
149 if (list_empty(&ring->free_pages)) {
150 BUG_ON(ring->free_pages_num != 0);
151 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
152 return gnttab_alloc_pages(1, page);
154 BUG_ON(ring->free_pages_num == 0);
155 page[0] = list_first_entry(&ring->free_pages, struct page, lru);
156 list_del(&page[0]->lru);
157 ring->free_pages_num--;
158 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
163 static inline void put_free_pages(struct xen_blkif_ring *ring, struct page **page,
169 spin_lock_irqsave(&ring->free_pages_lock, flags);
170 for (i = 0; i < num; i++)
171 list_add(&page[i]->lru, &ring->free_pages);
172 ring->free_pages_num += num;
173 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
176 static inline void shrink_free_pagepool(struct xen_blkif_ring *ring, int num)
178 /* Remove requested pages in batches of NUM_BATCH_FREE_PAGES */
179 struct page *page[NUM_BATCH_FREE_PAGES];
180 unsigned int num_pages = 0;
183 spin_lock_irqsave(&ring->free_pages_lock, flags);
184 while (ring->free_pages_num > num) {
185 BUG_ON(list_empty(&ring->free_pages));
186 page[num_pages] = list_first_entry(&ring->free_pages,
188 list_del(&page[num_pages]->lru);
189 ring->free_pages_num--;
190 if (++num_pages == NUM_BATCH_FREE_PAGES) {
191 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
192 gnttab_free_pages(num_pages, page);
193 spin_lock_irqsave(&ring->free_pages_lock, flags);
197 spin_unlock_irqrestore(&ring->free_pages_lock, flags);
199 gnttab_free_pages(num_pages, page);
202 #define vaddr(page) ((unsigned long)pfn_to_kaddr(page_to_pfn(page)))
204 static int do_block_io_op(struct xen_blkif_ring *ring);
205 static int dispatch_rw_block_io(struct xen_blkif_ring *ring,
206 struct blkif_request *req,
207 struct pending_req *pending_req);
208 static void make_response(struct xen_blkif_ring *ring, u64 id,
209 unsigned short op, int st);
211 #define foreach_grant_safe(pos, n, rbtree, node) \
212 for ((pos) = container_of(rb_first((rbtree)), typeof(*(pos)), node), \
213 (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL; \
214 &(pos)->node != NULL; \
215 (pos) = container_of(n, typeof(*(pos)), node), \
216 (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL)
220 * We don't need locking around the persistent grant helpers
221 * because blkback uses a single-thread for each backend, so we
222 * can be sure that this functions will never be called recursively.
224 * The only exception to that is put_persistent_grant, that can be called
225 * from interrupt context (by xen_blkbk_unmap), so we have to use atomic
226 * bit operations to modify the flags of a persistent grant and to count
227 * the number of used grants.
229 static int add_persistent_gnt(struct xen_blkif_ring *ring,
230 struct persistent_gnt *persistent_gnt)
232 struct rb_node **new = NULL, *parent = NULL;
233 struct persistent_gnt *this;
234 struct xen_blkif *blkif = ring->blkif;
236 if (ring->persistent_gnt_c >= max_pgrants) {
237 if (!blkif->vbd.overflow_max_grants)
238 blkif->vbd.overflow_max_grants = 1;
241 /* Figure out where to put new node */
242 new = &ring->persistent_gnts.rb_node;
244 this = container_of(*new, struct persistent_gnt, node);
247 if (persistent_gnt->gnt < this->gnt)
248 new = &((*new)->rb_left);
249 else if (persistent_gnt->gnt > this->gnt)
250 new = &((*new)->rb_right);
252 pr_alert_ratelimited("trying to add a gref that's already in the tree\n");
257 persistent_gnt->active = true;
258 /* Add new node and rebalance tree. */
259 rb_link_node(&(persistent_gnt->node), parent, new);
260 rb_insert_color(&(persistent_gnt->node), &ring->persistent_gnts);
261 ring->persistent_gnt_c++;
262 atomic_inc(&ring->persistent_gnt_in_use);
266 static struct persistent_gnt *get_persistent_gnt(struct xen_blkif_ring *ring,
269 struct persistent_gnt *data;
270 struct rb_node *node = NULL;
272 node = ring->persistent_gnts.rb_node;
274 data = container_of(node, struct persistent_gnt, node);
276 if (gref < data->gnt)
277 node = node->rb_left;
278 else if (gref > data->gnt)
279 node = node->rb_right;
282 pr_alert_ratelimited("requesting a grant already in use\n");
286 atomic_inc(&ring->persistent_gnt_in_use);
293 static void put_persistent_gnt(struct xen_blkif_ring *ring,
294 struct persistent_gnt *persistent_gnt)
296 if (!persistent_gnt->active)
297 pr_alert_ratelimited("freeing a grant already unused\n");
298 persistent_gnt->last_used = jiffies;
299 persistent_gnt->active = false;
300 atomic_dec(&ring->persistent_gnt_in_use);
303 static void free_persistent_gnts(struct xen_blkif_ring *ring, struct rb_root *root,
306 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
307 struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
308 struct persistent_gnt *persistent_gnt;
310 int segs_to_unmap = 0;
311 struct gntab_unmap_queue_data unmap_data;
313 unmap_data.pages = pages;
314 unmap_data.unmap_ops = unmap;
315 unmap_data.kunmap_ops = NULL;
317 foreach_grant_safe(persistent_gnt, n, root, node) {
318 BUG_ON(persistent_gnt->handle ==
319 BLKBACK_INVALID_HANDLE);
320 gnttab_set_unmap_op(&unmap[segs_to_unmap],
321 (unsigned long) pfn_to_kaddr(page_to_pfn(
322 persistent_gnt->page)),
324 persistent_gnt->handle);
326 pages[segs_to_unmap] = persistent_gnt->page;
328 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST ||
329 !rb_next(&persistent_gnt->node)) {
331 unmap_data.count = segs_to_unmap;
332 BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
334 put_free_pages(ring, pages, segs_to_unmap);
338 rb_erase(&persistent_gnt->node, root);
339 kfree(persistent_gnt);
345 void xen_blkbk_unmap_purged_grants(struct work_struct *work)
347 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
348 struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
349 struct persistent_gnt *persistent_gnt;
350 int segs_to_unmap = 0;
351 struct xen_blkif_ring *ring = container_of(work, typeof(*ring), persistent_purge_work);
352 struct gntab_unmap_queue_data unmap_data;
354 unmap_data.pages = pages;
355 unmap_data.unmap_ops = unmap;
356 unmap_data.kunmap_ops = NULL;
358 while(!list_empty(&ring->persistent_purge_list)) {
359 persistent_gnt = list_first_entry(&ring->persistent_purge_list,
360 struct persistent_gnt,
362 list_del(&persistent_gnt->remove_node);
364 gnttab_set_unmap_op(&unmap[segs_to_unmap],
365 vaddr(persistent_gnt->page),
367 persistent_gnt->handle);
369 pages[segs_to_unmap] = persistent_gnt->page;
371 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
372 unmap_data.count = segs_to_unmap;
373 BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
374 put_free_pages(ring, pages, segs_to_unmap);
377 kfree(persistent_gnt);
379 if (segs_to_unmap > 0) {
380 unmap_data.count = segs_to_unmap;
381 BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
382 put_free_pages(ring, pages, segs_to_unmap);
386 static void purge_persistent_gnt(struct xen_blkif_ring *ring)
388 struct persistent_gnt *persistent_gnt;
390 unsigned int num_clean, total;
391 bool scan_used = false;
392 struct rb_root *root;
394 if (work_busy(&ring->persistent_purge_work)) {
395 pr_alert_ratelimited("Scheduled work from previous purge is still busy, cannot purge list\n");
399 if (ring->persistent_gnt_c < max_pgrants ||
400 (ring->persistent_gnt_c == max_pgrants &&
401 !ring->blkif->vbd.overflow_max_grants)) {
404 num_clean = (max_pgrants / 100) * LRU_PERCENT_CLEAN;
405 num_clean = ring->persistent_gnt_c - max_pgrants + num_clean;
406 num_clean = min(ring->persistent_gnt_c, num_clean);
407 pr_debug("Going to purge at least %u persistent grants\n",
412 * At this point, we can assure that there will be no calls
413 * to get_persistent_grant (because we are executing this code from
414 * xen_blkif_schedule), there can only be calls to put_persistent_gnt,
415 * which means that the number of currently used grants will go down,
416 * but never up, so we will always be able to remove the requested
422 BUG_ON(!list_empty(&ring->persistent_purge_list));
423 root = &ring->persistent_gnts;
425 foreach_grant_safe(persistent_gnt, n, root, node) {
426 BUG_ON(persistent_gnt->handle ==
427 BLKBACK_INVALID_HANDLE);
429 if (persistent_gnt->active)
431 if (!scan_used && !persistent_gnt_timeout(persistent_gnt))
433 if (scan_used && total >= num_clean)
436 rb_erase(&persistent_gnt->node, root);
437 list_add(&persistent_gnt->remove_node,
438 &ring->persistent_purge_list);
442 * Check whether we also need to start cleaning
443 * grants that were used since last purge in order to cope
444 * with the requested num
446 if (!scan_used && total < num_clean) {
447 pr_debug("Still missing %u purged frames\n", num_clean - total);
453 ring->persistent_gnt_c -= total;
454 ring->blkif->vbd.overflow_max_grants = 0;
456 /* We can defer this work */
457 schedule_work(&ring->persistent_purge_work);
458 pr_debug("Purged %u/%u\n", num_clean, total);
466 * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
468 static struct pending_req *alloc_req(struct xen_blkif_ring *ring)
470 struct pending_req *req = NULL;
473 spin_lock_irqsave(&ring->pending_free_lock, flags);
474 if (!list_empty(&ring->pending_free)) {
475 req = list_entry(ring->pending_free.next, struct pending_req,
477 list_del(&req->free_list);
479 spin_unlock_irqrestore(&ring->pending_free_lock, flags);
484 * Return the 'pending_req' structure back to the freepool. We also
485 * wake up the thread if it was waiting for a free page.
487 static void free_req(struct xen_blkif_ring *ring, struct pending_req *req)
492 spin_lock_irqsave(&ring->pending_free_lock, flags);
493 was_empty = list_empty(&ring->pending_free);
494 list_add(&req->free_list, &ring->pending_free);
495 spin_unlock_irqrestore(&ring->pending_free_lock, flags);
497 wake_up(&ring->pending_free_wq);
501 * Routines for managing virtual block devices (vbds).
503 static int xen_vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
506 struct xen_vbd *vbd = &blkif->vbd;
509 if ((operation != REQ_OP_READ) && vbd->readonly)
512 if (likely(req->nr_sects)) {
513 blkif_sector_t end = req->sector_number + req->nr_sects;
515 if (unlikely(end < req->sector_number))
517 if (unlikely(end > vbd_sz(vbd)))
521 req->dev = vbd->pdevice;
522 req->bdev = vbd->bdev;
529 static void xen_vbd_resize(struct xen_blkif *blkif)
531 struct xen_vbd *vbd = &blkif->vbd;
532 struct xenbus_transaction xbt;
534 struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
535 unsigned long long new_size = vbd_sz(vbd);
537 pr_info("VBD Resize: Domid: %d, Device: (%d, %d)\n",
538 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
539 pr_info("VBD Resize: new size %llu\n", new_size);
540 vbd->size = new_size;
542 err = xenbus_transaction_start(&xbt);
544 pr_warn("Error starting transaction\n");
547 err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
548 (unsigned long long)vbd_sz(vbd));
550 pr_warn("Error writing new size\n");
554 * Write the current state; we will use this to synchronize
555 * the front-end. If the current state is "connected" the
556 * front-end will get the new size information online.
558 err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
560 pr_warn("Error writing the state\n");
564 err = xenbus_transaction_end(xbt, 0);
568 pr_warn("Error ending transaction\n");
571 xenbus_transaction_end(xbt, 1);
575 * Notification from the guest OS.
577 static void blkif_notify_work(struct xen_blkif_ring *ring)
579 ring->waiting_reqs = 1;
583 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
585 blkif_notify_work(dev_id);
590 * SCHEDULER FUNCTIONS
593 static void print_stats(struct xen_blkif_ring *ring)
595 pr_info("(%s): oo %3llu | rd %4llu | wr %4llu | f %4llu"
596 " | ds %4llu | pg: %4u/%4d\n",
597 current->comm, ring->st_oo_req,
598 ring->st_rd_req, ring->st_wr_req,
599 ring->st_f_req, ring->st_ds_req,
600 ring->persistent_gnt_c, max_pgrants);
601 ring->st_print = jiffies + msecs_to_jiffies(10 * 1000);
608 int xen_blkif_schedule(void *arg)
610 struct xen_blkif_ring *ring = arg;
611 struct xen_blkif *blkif = ring->blkif;
612 struct xen_vbd *vbd = &blkif->vbd;
613 unsigned long timeout;
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 the free pages pool if it is too large. */
657 if (time_before(jiffies, blkif->buffer_squeeze_end))
658 shrink_free_pagepool(ring, 0);
660 shrink_free_pagepool(ring, max_buffer_pages);
662 if (log_stats && time_after(jiffies, ring->st_print))
666 /* Drain pending purge work */
667 flush_work(&ring->persistent_purge_work);
672 ring->xenblkd = NULL;
678 * Remove persistent grants and empty the pool of free pages
680 void xen_blkbk_free_caches(struct xen_blkif_ring *ring)
682 /* Free all persistent grant pages */
683 if (!RB_EMPTY_ROOT(&ring->persistent_gnts))
684 free_persistent_gnts(ring, &ring->persistent_gnts,
685 ring->persistent_gnt_c);
687 BUG_ON(!RB_EMPTY_ROOT(&ring->persistent_gnts));
688 ring->persistent_gnt_c = 0;
690 /* Since we are shutting down remove all pages from the buffer */
691 shrink_free_pagepool(ring, 0 /* All */);
694 static unsigned int xen_blkbk_unmap_prepare(
695 struct xen_blkif_ring *ring,
696 struct grant_page **pages,
698 struct gnttab_unmap_grant_ref *unmap_ops,
699 struct page **unmap_pages)
701 unsigned int i, invcount = 0;
703 for (i = 0; i < num; i++) {
704 if (pages[i]->persistent_gnt != NULL) {
705 put_persistent_gnt(ring, pages[i]->persistent_gnt);
708 if (pages[i]->handle == BLKBACK_INVALID_HANDLE)
710 unmap_pages[invcount] = pages[i]->page;
711 gnttab_set_unmap_op(&unmap_ops[invcount], vaddr(pages[i]->page),
712 GNTMAP_host_map, pages[i]->handle);
713 pages[i]->handle = BLKBACK_INVALID_HANDLE;
720 static void xen_blkbk_unmap_and_respond_callback(int result, struct gntab_unmap_queue_data *data)
722 struct pending_req *pending_req = (struct pending_req *)(data->data);
723 struct xen_blkif_ring *ring = pending_req->ring;
724 struct xen_blkif *blkif = ring->blkif;
726 /* BUG_ON used to reproduce existing behaviour,
727 but is this the best way to deal with this? */
730 put_free_pages(ring, data->pages, data->count);
731 make_response(ring, pending_req->id,
732 pending_req->operation, pending_req->status);
733 free_req(ring, pending_req);
735 * Make sure the request is freed before releasing blkif,
736 * or there could be a race between free_req and the
737 * cleanup done in xen_blkif_free during shutdown.
739 * NB: The fact that we might try to wake up pending_free_wq
740 * before drain_complete (in case there's a drain going on)
741 * it's not a problem with our current implementation
742 * because we can assure there's no thread waiting on
743 * pending_free_wq if there's a drain going on, but it has
744 * to be taken into account if the current model is changed.
746 if (atomic_dec_and_test(&ring->inflight) && atomic_read(&blkif->drain)) {
747 complete(&blkif->drain_complete);
749 xen_blkif_put(blkif);
752 static void xen_blkbk_unmap_and_respond(struct pending_req *req)
754 struct gntab_unmap_queue_data* work = &req->gnttab_unmap_data;
755 struct xen_blkif_ring *ring = req->ring;
756 struct grant_page **pages = req->segments;
757 unsigned int invcount;
759 invcount = xen_blkbk_unmap_prepare(ring, pages, req->nr_segs,
760 req->unmap, req->unmap_pages);
763 work->done = xen_blkbk_unmap_and_respond_callback;
764 work->unmap_ops = req->unmap;
765 work->kunmap_ops = NULL;
766 work->pages = req->unmap_pages;
767 work->count = invcount;
769 gnttab_unmap_refs_async(&req->gnttab_unmap_data);
774 * Unmap the grant references.
776 * This could accumulate ops up to the batch size to reduce the number
777 * of hypercalls, but since this is only used in error paths there's
780 static void xen_blkbk_unmap(struct xen_blkif_ring *ring,
781 struct grant_page *pages[],
784 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
785 struct page *unmap_pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
786 unsigned int invcount = 0;
790 unsigned int batch = min(num, BLKIF_MAX_SEGMENTS_PER_REQUEST);
792 invcount = xen_blkbk_unmap_prepare(ring, pages, batch,
795 ret = gnttab_unmap_refs(unmap, NULL, unmap_pages, invcount);
797 put_free_pages(ring, unmap_pages, invcount);
804 static int xen_blkbk_map(struct xen_blkif_ring *ring,
805 struct grant_page *pages[],
808 struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
809 struct page *pages_to_gnt[BLKIF_MAX_SEGMENTS_PER_REQUEST];
810 struct persistent_gnt *persistent_gnt = NULL;
811 phys_addr_t addr = 0;
812 int i, seg_idx, new_map_idx;
815 int last_map = 0, map_until = 0;
816 int use_persistent_gnts;
817 struct xen_blkif *blkif = ring->blkif;
819 use_persistent_gnts = (blkif->vbd.feature_gnt_persistent);
822 * Fill out preq.nr_sects with proper amount of sectors, and setup
823 * assign map[..] with the PFN of the page in our domain with the
824 * corresponding grant reference for each page.
827 for (i = map_until; i < num; i++) {
830 if (use_persistent_gnts) {
831 persistent_gnt = get_persistent_gnt(
836 if (persistent_gnt) {
838 * We are using persistent grants and
839 * the grant is already mapped
841 pages[i]->page = persistent_gnt->page;
842 pages[i]->persistent_gnt = persistent_gnt;
844 if (get_free_page(ring, &pages[i]->page))
846 addr = vaddr(pages[i]->page);
847 pages_to_gnt[segs_to_map] = pages[i]->page;
848 pages[i]->persistent_gnt = NULL;
849 flags = GNTMAP_host_map;
850 if (!use_persistent_gnts && ro)
851 flags |= GNTMAP_readonly;
852 gnttab_set_map_op(&map[segs_to_map++], addr,
853 flags, pages[i]->gref,
857 if (segs_to_map == BLKIF_MAX_SEGMENTS_PER_REQUEST)
862 ret = gnttab_map_refs(map, NULL, pages_to_gnt, segs_to_map);
867 * Now swizzle the MFN in our domain with the MFN from the other domain
868 * so that when we access vaddr(pending_req,i) it has the contents of
869 * the page from the other domain.
871 for (seg_idx = last_map, new_map_idx = 0; seg_idx < map_until; seg_idx++) {
872 if (!pages[seg_idx]->persistent_gnt) {
873 /* This is a newly mapped grant */
874 BUG_ON(new_map_idx >= segs_to_map);
875 if (unlikely(map[new_map_idx].status != 0)) {
876 pr_debug("invalid buffer -- could not remap it\n");
877 put_free_pages(ring, &pages[seg_idx]->page, 1);
878 pages[seg_idx]->handle = BLKBACK_INVALID_HANDLE;
882 pages[seg_idx]->handle = map[new_map_idx].handle;
886 if (use_persistent_gnts &&
887 ring->persistent_gnt_c < max_pgrants) {
889 * We are using persistent grants, the grant is
890 * not mapped but we might have room for it.
892 persistent_gnt = kmalloc(sizeof(struct persistent_gnt),
894 if (!persistent_gnt) {
896 * If we don't have enough memory to
897 * allocate the persistent_gnt struct
898 * map this grant non-persistenly
902 persistent_gnt->gnt = map[new_map_idx].ref;
903 persistent_gnt->handle = map[new_map_idx].handle;
904 persistent_gnt->page = pages[seg_idx]->page;
905 if (add_persistent_gnt(ring,
907 kfree(persistent_gnt);
908 persistent_gnt = NULL;
911 pages[seg_idx]->persistent_gnt = persistent_gnt;
912 pr_debug("grant %u added to the tree of persistent grants, using %u/%u\n",
913 persistent_gnt->gnt, ring->persistent_gnt_c,
917 if (use_persistent_gnts && !blkif->vbd.overflow_max_grants) {
918 blkif->vbd.overflow_max_grants = 1;
919 pr_debug("domain %u, device %#x is using maximum number of persistent grants\n",
920 blkif->domid, blkif->vbd.handle);
923 * We could not map this grant persistently, so use it as
924 * a non-persistent grant.
930 last_map = map_until;
931 if (map_until != num)
937 pr_alert("%s: out of memory\n", __func__);
938 put_free_pages(ring, pages_to_gnt, segs_to_map);
939 for (i = last_map; i < num; i++)
940 pages[i]->handle = BLKBACK_INVALID_HANDLE;
944 static int xen_blkbk_map_seg(struct pending_req *pending_req)
948 rc = xen_blkbk_map(pending_req->ring, pending_req->segments,
949 pending_req->nr_segs,
950 (pending_req->operation != BLKIF_OP_READ));
955 static int xen_blkbk_parse_indirect(struct blkif_request *req,
956 struct pending_req *pending_req,
957 struct seg_buf seg[],
958 struct phys_req *preq)
960 struct grant_page **pages = pending_req->indirect_pages;
961 struct xen_blkif_ring *ring = pending_req->ring;
962 int indirect_grefs, rc, n, nseg, i;
963 struct blkif_request_segment *segments = NULL;
965 nseg = pending_req->nr_segs;
966 indirect_grefs = INDIRECT_PAGES(nseg);
967 BUG_ON(indirect_grefs > BLKIF_MAX_INDIRECT_PAGES_PER_REQUEST);
969 for (i = 0; i < indirect_grefs; i++)
970 pages[i]->gref = req->u.indirect.indirect_grefs[i];
972 rc = xen_blkbk_map(ring, pages, indirect_grefs, true);
976 for (n = 0, i = 0; n < nseg; n++) {
977 uint8_t first_sect, last_sect;
979 if ((n % SEGS_PER_INDIRECT_FRAME) == 0) {
980 /* Map indirect segments */
982 kunmap_atomic(segments);
983 segments = kmap_atomic(pages[n/SEGS_PER_INDIRECT_FRAME]->page);
985 i = n % SEGS_PER_INDIRECT_FRAME;
987 pending_req->segments[n]->gref = segments[i].gref;
989 first_sect = READ_ONCE(segments[i].first_sect);
990 last_sect = READ_ONCE(segments[i].last_sect);
991 if (last_sect >= (XEN_PAGE_SIZE >> 9) || last_sect < first_sect) {
996 seg[n].nsec = last_sect - first_sect + 1;
997 seg[n].offset = first_sect << 9;
998 preq->nr_sects += seg[n].nsec;
1003 kunmap_atomic(segments);
1004 xen_blkbk_unmap(ring, pages, indirect_grefs);
1008 static int dispatch_discard_io(struct xen_blkif_ring *ring,
1009 struct blkif_request *req)
1012 int status = BLKIF_RSP_OKAY;
1013 struct xen_blkif *blkif = ring->blkif;
1014 struct block_device *bdev = blkif->vbd.bdev;
1015 unsigned long secure;
1016 struct phys_req preq;
1018 xen_blkif_get(blkif);
1020 preq.sector_number = req->u.discard.sector_number;
1021 preq.nr_sects = req->u.discard.nr_sectors;
1023 err = xen_vbd_translate(&preq, blkif, REQ_OP_WRITE);
1025 pr_warn("access denied: DISCARD [%llu->%llu] on dev=%04x\n",
1027 preq.sector_number + preq.nr_sects, blkif->vbd.pdevice);
1032 secure = (blkif->vbd.discard_secure &&
1033 (req->u.discard.flag & BLKIF_DISCARD_SECURE)) ?
1034 BLKDEV_DISCARD_SECURE : 0;
1036 err = blkdev_issue_discard(bdev, req->u.discard.sector_number,
1037 req->u.discard.nr_sectors,
1038 GFP_KERNEL, secure);
1040 if (err == -EOPNOTSUPP) {
1041 pr_debug("discard op failed, not supported\n");
1042 status = BLKIF_RSP_EOPNOTSUPP;
1044 status = BLKIF_RSP_ERROR;
1046 make_response(ring, req->u.discard.id, req->operation, status);
1047 xen_blkif_put(blkif);
1051 static int dispatch_other_io(struct xen_blkif_ring *ring,
1052 struct blkif_request *req,
1053 struct pending_req *pending_req)
1055 free_req(ring, pending_req);
1056 make_response(ring, req->u.other.id, req->operation,
1057 BLKIF_RSP_EOPNOTSUPP);
1061 static void xen_blk_drain_io(struct xen_blkif_ring *ring)
1063 struct xen_blkif *blkif = ring->blkif;
1065 atomic_set(&blkif->drain, 1);
1067 if (atomic_read(&ring->inflight) == 0)
1069 wait_for_completion_interruptible_timeout(
1070 &blkif->drain_complete, HZ);
1072 if (!atomic_read(&blkif->drain))
1074 } while (!kthread_should_stop());
1075 atomic_set(&blkif->drain, 0);
1078 static void __end_block_io_op(struct pending_req *pending_req,
1081 /* An error fails the entire request. */
1082 if (pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE &&
1083 error == BLK_STS_NOTSUPP) {
1084 pr_debug("flush diskcache op failed, not supported\n");
1085 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->ring->blkif->be, 0);
1086 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
1087 } else if (pending_req->operation == BLKIF_OP_WRITE_BARRIER &&
1088 error == BLK_STS_NOTSUPP) {
1089 pr_debug("write barrier op failed, not supported\n");
1090 xen_blkbk_barrier(XBT_NIL, pending_req->ring->blkif->be, 0);
1091 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
1093 pr_debug("Buffer not up-to-date at end of operation,"
1094 " error=%d\n", error);
1095 pending_req->status = BLKIF_RSP_ERROR;
1099 * If all of the bio's have completed it is time to unmap
1100 * the grant references associated with 'request' and provide
1101 * the proper response on the ring.
1103 if (atomic_dec_and_test(&pending_req->pendcnt))
1104 xen_blkbk_unmap_and_respond(pending_req);
1110 static void end_block_io_op(struct bio *bio)
1112 __end_block_io_op(bio->bi_private, bio->bi_status);
1119 * Function to copy the from the ring buffer the 'struct blkif_request'
1120 * (which has the sectors we want, number of them, grant references, etc),
1121 * and transmute it to the block API to hand it over to the proper block disk.
1124 __do_block_io_op(struct xen_blkif_ring *ring)
1126 union blkif_back_rings *blk_rings = &ring->blk_rings;
1127 struct blkif_request req;
1128 struct pending_req *pending_req;
1132 rc = blk_rings->common.req_cons;
1133 rp = blk_rings->common.sring->req_prod;
1134 rmb(); /* Ensure we see queued requests up to 'rp'. */
1136 if (RING_REQUEST_PROD_OVERFLOW(&blk_rings->common, rp)) {
1137 rc = blk_rings->common.rsp_prod_pvt;
1138 pr_warn("Frontend provided bogus ring requests (%d - %d = %d). Halting ring processing on dev=%04x\n",
1139 rp, rc, rp - rc, ring->blkif->vbd.pdevice);
1144 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
1147 if (kthread_should_stop()) {
1152 pending_req = alloc_req(ring);
1153 if (NULL == pending_req) {
1159 switch (ring->blkif->blk_protocol) {
1160 case BLKIF_PROTOCOL_NATIVE:
1161 memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
1163 case BLKIF_PROTOCOL_X86_32:
1164 blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
1166 case BLKIF_PROTOCOL_X86_64:
1167 blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
1172 blk_rings->common.req_cons = ++rc; /* before make_response() */
1174 /* Apply all sanity checks to /private copy/ of request. */
1177 switch (req.operation) {
1179 case BLKIF_OP_WRITE:
1180 case BLKIF_OP_WRITE_BARRIER:
1181 case BLKIF_OP_FLUSH_DISKCACHE:
1182 case BLKIF_OP_INDIRECT:
1183 if (dispatch_rw_block_io(ring, &req, pending_req))
1186 case BLKIF_OP_DISCARD:
1187 free_req(ring, pending_req);
1188 if (dispatch_discard_io(ring, &req))
1192 if (dispatch_other_io(ring, &req, pending_req))
1197 /* Yield point for this unbounded loop. */
1205 do_block_io_op(struct xen_blkif_ring *ring)
1207 union blkif_back_rings *blk_rings = &ring->blk_rings;
1211 more_to_do = __do_block_io_op(ring);
1215 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
1216 } while (more_to_do);
1221 * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
1222 * and call the 'submit_bio' to pass it to the underlying storage.
1224 static int dispatch_rw_block_io(struct xen_blkif_ring *ring,
1225 struct blkif_request *req,
1226 struct pending_req *pending_req)
1228 struct phys_req preq;
1229 struct seg_buf *seg = pending_req->seg;
1231 struct bio *bio = NULL;
1232 struct bio **biolist = pending_req->biolist;
1235 int operation_flags = 0;
1236 struct blk_plug plug;
1238 struct grant_page **pages = pending_req->segments;
1239 unsigned short req_operation;
1241 req_operation = req->operation == BLKIF_OP_INDIRECT ?
1242 req->u.indirect.indirect_op : req->operation;
1244 if ((req->operation == BLKIF_OP_INDIRECT) &&
1245 (req_operation != BLKIF_OP_READ) &&
1246 (req_operation != BLKIF_OP_WRITE)) {
1247 pr_debug("Invalid indirect operation (%u)\n", req_operation);
1251 switch (req_operation) {
1254 operation = REQ_OP_READ;
1256 case BLKIF_OP_WRITE:
1258 operation = REQ_OP_WRITE;
1259 operation_flags = REQ_SYNC | REQ_IDLE;
1261 case BLKIF_OP_WRITE_BARRIER:
1264 case BLKIF_OP_FLUSH_DISKCACHE:
1266 operation = REQ_OP_WRITE;
1267 operation_flags = REQ_PREFLUSH;
1270 operation = 0; /* make gcc happy */
1275 /* Check that the number of segments is sane. */
1276 nseg = req->operation == BLKIF_OP_INDIRECT ?
1277 req->u.indirect.nr_segments : req->u.rw.nr_segments;
1279 if (unlikely(nseg == 0 && operation_flags != REQ_PREFLUSH) ||
1280 unlikely((req->operation != BLKIF_OP_INDIRECT) &&
1281 (nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) ||
1282 unlikely((req->operation == BLKIF_OP_INDIRECT) &&
1283 (nseg > MAX_INDIRECT_SEGMENTS))) {
1284 pr_debug("Bad number of segments in request (%d)\n", nseg);
1285 /* Haven't submitted any bio's yet. */
1291 pending_req->ring = ring;
1292 pending_req->id = req->u.rw.id;
1293 pending_req->operation = req_operation;
1294 pending_req->status = BLKIF_RSP_OKAY;
1295 pending_req->nr_segs = nseg;
1297 if (req->operation != BLKIF_OP_INDIRECT) {
1298 preq.dev = req->u.rw.handle;
1299 preq.sector_number = req->u.rw.sector_number;
1300 for (i = 0; i < nseg; i++) {
1301 pages[i]->gref = req->u.rw.seg[i].gref;
1302 seg[i].nsec = req->u.rw.seg[i].last_sect -
1303 req->u.rw.seg[i].first_sect + 1;
1304 seg[i].offset = (req->u.rw.seg[i].first_sect << 9);
1305 if ((req->u.rw.seg[i].last_sect >= (XEN_PAGE_SIZE >> 9)) ||
1306 (req->u.rw.seg[i].last_sect <
1307 req->u.rw.seg[i].first_sect))
1309 preq.nr_sects += seg[i].nsec;
1312 preq.dev = req->u.indirect.handle;
1313 preq.sector_number = req->u.indirect.sector_number;
1314 if (xen_blkbk_parse_indirect(req, pending_req, seg, &preq))
1318 if (xen_vbd_translate(&preq, ring->blkif, operation) != 0) {
1319 pr_debug("access denied: %s of [%llu,%llu] on dev=%04x\n",
1320 operation == REQ_OP_READ ? "read" : "write",
1322 preq.sector_number + preq.nr_sects,
1323 ring->blkif->vbd.pdevice);
1328 * This check _MUST_ be done after xen_vbd_translate as the preq.bdev
1331 for (i = 0; i < nseg; i++) {
1332 if (((int)preq.sector_number|(int)seg[i].nsec) &
1333 ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
1334 pr_debug("Misaligned I/O request from domain %d\n",
1335 ring->blkif->domid);
1340 /* Wait on all outstanding I/O's and once that has been completed
1344 xen_blk_drain_io(pending_req->ring);
1347 * If we have failed at this point, we need to undo the M2P override,
1348 * set gnttab_set_unmap_op on all of the grant references and perform
1349 * the hypercall to unmap the grants - that is all done in
1352 if (xen_blkbk_map_seg(pending_req))
1356 * This corresponding xen_blkif_put is done in __end_block_io_op, or
1357 * below (in "!bio") if we are handling a BLKIF_OP_DISCARD.
1359 xen_blkif_get(ring->blkif);
1360 atomic_inc(&ring->inflight);
1362 for (i = 0; i < nseg; i++) {
1363 while ((bio == NULL) ||
1367 seg[i].offset) == 0)) {
1369 int nr_iovecs = min_t(int, (nseg-i), BIO_MAX_PAGES);
1370 bio = bio_alloc(GFP_KERNEL, nr_iovecs);
1371 if (unlikely(bio == NULL))
1374 biolist[nbio++] = bio;
1375 bio_set_dev(bio, preq.bdev);
1376 bio->bi_private = pending_req;
1377 bio->bi_end_io = end_block_io_op;
1378 bio->bi_iter.bi_sector = preq.sector_number;
1379 bio_set_op_attrs(bio, operation, operation_flags);
1382 preq.sector_number += seg[i].nsec;
1385 /* This will be hit if the operation was a flush or discard. */
1387 BUG_ON(operation_flags != REQ_PREFLUSH);
1389 bio = bio_alloc(GFP_KERNEL, 0);
1390 if (unlikely(bio == NULL))
1393 biolist[nbio++] = bio;
1394 bio_set_dev(bio, preq.bdev);
1395 bio->bi_private = pending_req;
1396 bio->bi_end_io = end_block_io_op;
1397 bio_set_op_attrs(bio, operation, operation_flags);
1400 atomic_set(&pending_req->pendcnt, nbio);
1401 blk_start_plug(&plug);
1403 for (i = 0; i < nbio; i++)
1404 submit_bio(biolist[i]);
1406 /* Let the I/Os go.. */
1407 blk_finish_plug(&plug);
1409 if (operation == REQ_OP_READ)
1410 ring->st_rd_sect += preq.nr_sects;
1411 else if (operation == REQ_OP_WRITE)
1412 ring->st_wr_sect += preq.nr_sects;
1417 xen_blkbk_unmap(ring, pending_req->segments,
1418 pending_req->nr_segs);
1420 /* Haven't submitted any bio's yet. */
1421 make_response(ring, req->u.rw.id, req_operation, BLKIF_RSP_ERROR);
1422 free_req(ring, pending_req);
1423 msleep(1); /* back off a bit */
1427 for (i = 0; i < nbio; i++)
1428 bio_put(biolist[i]);
1429 atomic_set(&pending_req->pendcnt, 1);
1430 __end_block_io_op(pending_req, BLK_STS_RESOURCE);
1431 msleep(1); /* back off a bit */
1438 * Put a response on the ring on how the operation fared.
1440 static void make_response(struct xen_blkif_ring *ring, u64 id,
1441 unsigned short op, int st)
1443 struct blkif_response *resp;
1444 unsigned long flags;
1445 union blkif_back_rings *blk_rings;
1448 spin_lock_irqsave(&ring->blk_ring_lock, flags);
1449 blk_rings = &ring->blk_rings;
1450 /* Place on the response ring for the relevant domain. */
1451 switch (ring->blkif->blk_protocol) {
1452 case BLKIF_PROTOCOL_NATIVE:
1453 resp = RING_GET_RESPONSE(&blk_rings->native,
1454 blk_rings->native.rsp_prod_pvt);
1456 case BLKIF_PROTOCOL_X86_32:
1457 resp = RING_GET_RESPONSE(&blk_rings->x86_32,
1458 blk_rings->x86_32.rsp_prod_pvt);
1460 case BLKIF_PROTOCOL_X86_64:
1461 resp = RING_GET_RESPONSE(&blk_rings->x86_64,
1462 blk_rings->x86_64.rsp_prod_pvt);
1469 resp->operation = op;
1472 blk_rings->common.rsp_prod_pvt++;
1473 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
1474 spin_unlock_irqrestore(&ring->blk_ring_lock, flags);
1476 notify_remote_via_irq(ring->irq);
1479 static int __init xen_blkif_init(void)
1486 if (xen_blkif_max_ring_order > XENBUS_MAX_RING_GRANT_ORDER) {
1487 pr_info("Invalid max_ring_order (%d), will use default max: %d.\n",
1488 xen_blkif_max_ring_order, XENBUS_MAX_RING_GRANT_ORDER);
1489 xen_blkif_max_ring_order = XENBUS_MAX_RING_GRANT_ORDER;
1492 if (xenblk_max_queues == 0)
1493 xenblk_max_queues = num_online_cpus();
1495 rc = xen_blkif_interface_init();
1499 rc = xen_blkif_xenbus_init();
1507 module_init(xen_blkif_init);
1509 static void __exit xen_blkif_fini(void)
1511 xen_blkif_xenbus_fini();
1512 xen_blkif_interface_fini();
1515 module_exit(xen_blkif_fini);
1517 MODULE_LICENSE("Dual BSD/GPL");
1518 MODULE_ALIAS("xen-backend:vbd");