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 order of pages to be used for the shared ring between front and
88 * backend, 4KB page granularity is used.
90 unsigned int xen_blkif_max_ring_order = XENBUS_MAX_RING_GRANT_ORDER;
91 module_param_named(max_ring_page_order, xen_blkif_max_ring_order, int, S_IRUGO);
92 MODULE_PARM_DESC(max_ring_page_order, "Maximum order of pages to be used for the shared ring");
94 * The LRU mechanism to clean the lists of persistent grants needs to
95 * be executed periodically. The time interval between consecutive executions
96 * of the purge mechanism is set in ms.
98 #define LRU_INTERVAL 100
101 * When the persistent grants list is full we will remove unused grants
102 * from the list. The percent number of grants to be removed at each LRU
105 #define LRU_PERCENT_CLEAN 5
107 /* Run-time switchable: /sys/module/blkback/parameters/ */
108 static unsigned int log_stats;
109 module_param(log_stats, int, 0644);
111 #define BLKBACK_INVALID_HANDLE (~0)
113 /* Number of free pages to remove on each call to gnttab_free_pages */
114 #define NUM_BATCH_FREE_PAGES 10
116 static inline int get_free_page(struct xen_blkif *blkif, struct page **page)
120 spin_lock_irqsave(&blkif->free_pages_lock, flags);
121 if (list_empty(&blkif->free_pages)) {
122 BUG_ON(blkif->free_pages_num != 0);
123 spin_unlock_irqrestore(&blkif->free_pages_lock, flags);
124 return gnttab_alloc_pages(1, page);
126 BUG_ON(blkif->free_pages_num == 0);
127 page[0] = list_first_entry(&blkif->free_pages, struct page, lru);
128 list_del(&page[0]->lru);
129 blkif->free_pages_num--;
130 spin_unlock_irqrestore(&blkif->free_pages_lock, flags);
135 static inline void put_free_pages(struct xen_blkif *blkif, struct page **page,
141 spin_lock_irqsave(&blkif->free_pages_lock, flags);
142 for (i = 0; i < num; i++)
143 list_add(&page[i]->lru, &blkif->free_pages);
144 blkif->free_pages_num += num;
145 spin_unlock_irqrestore(&blkif->free_pages_lock, flags);
148 static inline void shrink_free_pagepool(struct xen_blkif *blkif, int num)
150 /* Remove requested pages in batches of NUM_BATCH_FREE_PAGES */
151 struct page *page[NUM_BATCH_FREE_PAGES];
152 unsigned int num_pages = 0;
155 spin_lock_irqsave(&blkif->free_pages_lock, flags);
156 while (blkif->free_pages_num > num) {
157 BUG_ON(list_empty(&blkif->free_pages));
158 page[num_pages] = list_first_entry(&blkif->free_pages,
160 list_del(&page[num_pages]->lru);
161 blkif->free_pages_num--;
162 if (++num_pages == NUM_BATCH_FREE_PAGES) {
163 spin_unlock_irqrestore(&blkif->free_pages_lock, flags);
164 gnttab_free_pages(num_pages, page);
165 spin_lock_irqsave(&blkif->free_pages_lock, flags);
169 spin_unlock_irqrestore(&blkif->free_pages_lock, flags);
171 gnttab_free_pages(num_pages, page);
174 #define vaddr(page) ((unsigned long)pfn_to_kaddr(page_to_pfn(page)))
176 static int do_block_io_op(struct xen_blkif *blkif, unsigned int *eoi_flags);
177 static int dispatch_rw_block_io(struct xen_blkif *blkif,
178 struct blkif_request *req,
179 struct pending_req *pending_req);
180 static void make_response(struct xen_blkif *blkif, u64 id,
181 unsigned short op, int st);
183 #define foreach_grant_safe(pos, n, rbtree, node) \
184 for ((pos) = container_of(rb_first((rbtree)), typeof(*(pos)), node), \
185 (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL; \
186 &(pos)->node != NULL; \
187 (pos) = container_of(n, typeof(*(pos)), node), \
188 (n) = (&(pos)->node != NULL) ? rb_next(&(pos)->node) : NULL)
192 * We don't need locking around the persistent grant helpers
193 * because blkback uses a single-thread for each backed, so we
194 * can be sure that this functions will never be called recursively.
196 * The only exception to that is put_persistent_grant, that can be called
197 * from interrupt context (by xen_blkbk_unmap), so we have to use atomic
198 * bit operations to modify the flags of a persistent grant and to count
199 * the number of used grants.
201 static int add_persistent_gnt(struct xen_blkif *blkif,
202 struct persistent_gnt *persistent_gnt)
204 struct rb_node **new = NULL, *parent = NULL;
205 struct persistent_gnt *this;
207 if (blkif->persistent_gnt_c >= xen_blkif_max_pgrants) {
208 if (!blkif->vbd.overflow_max_grants)
209 blkif->vbd.overflow_max_grants = 1;
212 /* Figure out where to put new node */
213 new = &blkif->persistent_gnts.rb_node;
215 this = container_of(*new, struct persistent_gnt, node);
218 if (persistent_gnt->gnt < this->gnt)
219 new = &((*new)->rb_left);
220 else if (persistent_gnt->gnt > this->gnt)
221 new = &((*new)->rb_right);
223 pr_alert_ratelimited("trying to add a gref that's already in the tree\n");
228 bitmap_zero(persistent_gnt->flags, PERSISTENT_GNT_FLAGS_SIZE);
229 set_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags);
230 /* Add new node and rebalance tree. */
231 rb_link_node(&(persistent_gnt->node), parent, new);
232 rb_insert_color(&(persistent_gnt->node), &blkif->persistent_gnts);
233 blkif->persistent_gnt_c++;
234 atomic_inc(&blkif->persistent_gnt_in_use);
238 static struct persistent_gnt *get_persistent_gnt(struct xen_blkif *blkif,
241 struct persistent_gnt *data;
242 struct rb_node *node = NULL;
244 node = blkif->persistent_gnts.rb_node;
246 data = container_of(node, struct persistent_gnt, node);
248 if (gref < data->gnt)
249 node = node->rb_left;
250 else if (gref > data->gnt)
251 node = node->rb_right;
253 if(test_bit(PERSISTENT_GNT_ACTIVE, data->flags)) {
254 pr_alert_ratelimited("requesting a grant already in use\n");
257 set_bit(PERSISTENT_GNT_ACTIVE, data->flags);
258 atomic_inc(&blkif->persistent_gnt_in_use);
265 static void put_persistent_gnt(struct xen_blkif *blkif,
266 struct persistent_gnt *persistent_gnt)
268 if(!test_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags))
269 pr_alert_ratelimited("freeing a grant already unused\n");
270 set_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags);
271 clear_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags);
272 atomic_dec(&blkif->persistent_gnt_in_use);
275 static void free_persistent_gnts(struct xen_blkif *blkif, struct rb_root *root,
278 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
279 struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
280 struct persistent_gnt *persistent_gnt;
282 int segs_to_unmap = 0;
283 struct gntab_unmap_queue_data unmap_data;
285 unmap_data.pages = pages;
286 unmap_data.unmap_ops = unmap;
287 unmap_data.kunmap_ops = NULL;
289 foreach_grant_safe(persistent_gnt, n, root, node) {
290 BUG_ON(persistent_gnt->handle ==
291 BLKBACK_INVALID_HANDLE);
292 gnttab_set_unmap_op(&unmap[segs_to_unmap],
293 (unsigned long) pfn_to_kaddr(page_to_pfn(
294 persistent_gnt->page)),
296 persistent_gnt->handle);
298 pages[segs_to_unmap] = persistent_gnt->page;
300 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST ||
301 !rb_next(&persistent_gnt->node)) {
303 unmap_data.count = segs_to_unmap;
304 BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
306 put_free_pages(blkif, pages, segs_to_unmap);
310 rb_erase(&persistent_gnt->node, root);
311 kfree(persistent_gnt);
317 void xen_blkbk_unmap_purged_grants(struct work_struct *work)
319 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
320 struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
321 struct persistent_gnt *persistent_gnt;
322 int segs_to_unmap = 0;
323 struct xen_blkif *blkif = container_of(work, typeof(*blkif), persistent_purge_work);
324 struct gntab_unmap_queue_data unmap_data;
326 unmap_data.pages = pages;
327 unmap_data.unmap_ops = unmap;
328 unmap_data.kunmap_ops = NULL;
330 while(!list_empty(&blkif->persistent_purge_list)) {
331 persistent_gnt = list_first_entry(&blkif->persistent_purge_list,
332 struct persistent_gnt,
334 list_del(&persistent_gnt->remove_node);
336 gnttab_set_unmap_op(&unmap[segs_to_unmap],
337 vaddr(persistent_gnt->page),
339 persistent_gnt->handle);
341 pages[segs_to_unmap] = persistent_gnt->page;
343 if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
344 unmap_data.count = segs_to_unmap;
345 BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
346 put_free_pages(blkif, pages, segs_to_unmap);
349 kfree(persistent_gnt);
351 if (segs_to_unmap > 0) {
352 unmap_data.count = segs_to_unmap;
353 BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
354 put_free_pages(blkif, pages, segs_to_unmap);
358 static void purge_persistent_gnt(struct xen_blkif *blkif)
360 struct persistent_gnt *persistent_gnt;
362 unsigned int num_clean, total;
363 bool scan_used = false, clean_used = false;
364 struct rb_root *root;
366 if (blkif->persistent_gnt_c < xen_blkif_max_pgrants ||
367 (blkif->persistent_gnt_c == xen_blkif_max_pgrants &&
368 !blkif->vbd.overflow_max_grants)) {
372 if (work_busy(&blkif->persistent_purge_work)) {
373 pr_alert_ratelimited("Scheduled work from previous purge is still busy, cannot purge list\n");
377 num_clean = (xen_blkif_max_pgrants / 100) * LRU_PERCENT_CLEAN;
378 num_clean = blkif->persistent_gnt_c - xen_blkif_max_pgrants + num_clean;
379 num_clean = min(blkif->persistent_gnt_c, num_clean);
380 if ((num_clean == 0) ||
381 (num_clean > (blkif->persistent_gnt_c - atomic_read(&blkif->persistent_gnt_in_use))))
385 * At this point, we can assure that there will be no calls
386 * to get_persistent_grant (because we are executing this code from
387 * xen_blkif_schedule), there can only be calls to put_persistent_gnt,
388 * which means that the number of currently used grants will go down,
389 * but never up, so we will always be able to remove the requested
395 pr_debug("Going to purge %u persistent grants\n", num_clean);
397 BUG_ON(!list_empty(&blkif->persistent_purge_list));
398 root = &blkif->persistent_gnts;
400 foreach_grant_safe(persistent_gnt, n, root, node) {
401 BUG_ON(persistent_gnt->handle ==
402 BLKBACK_INVALID_HANDLE);
405 clear_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags);
409 if (test_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags))
412 (test_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags)))
415 rb_erase(&persistent_gnt->node, root);
416 list_add(&persistent_gnt->remove_node,
417 &blkif->persistent_purge_list);
418 if (--num_clean == 0)
422 * If we get here it means we also need to start cleaning
423 * grants that were used since last purge in order to cope
424 * with the requested num
426 if (!scan_used && !clean_used) {
427 pr_debug("Still missing %u purged frames\n", num_clean);
433 pr_debug("Finished scanning for grants to clean, removing used flag\n");
438 blkif->persistent_gnt_c -= (total - num_clean);
439 blkif->vbd.overflow_max_grants = 0;
441 /* We can defer this work */
442 schedule_work(&blkif->persistent_purge_work);
443 pr_debug("Purged %u/%u\n", (total - num_clean), total);
448 * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
450 static struct pending_req *alloc_req(struct xen_blkif *blkif)
452 struct pending_req *req = NULL;
455 spin_lock_irqsave(&blkif->pending_free_lock, flags);
456 if (!list_empty(&blkif->pending_free)) {
457 req = list_entry(blkif->pending_free.next, struct pending_req,
459 list_del(&req->free_list);
461 spin_unlock_irqrestore(&blkif->pending_free_lock, flags);
466 * Return the 'pending_req' structure back to the freepool. We also
467 * wake up the thread if it was waiting for a free page.
469 static void free_req(struct xen_blkif *blkif, struct pending_req *req)
474 spin_lock_irqsave(&blkif->pending_free_lock, flags);
475 was_empty = list_empty(&blkif->pending_free);
476 list_add(&req->free_list, &blkif->pending_free);
477 spin_unlock_irqrestore(&blkif->pending_free_lock, flags);
479 wake_up(&blkif->pending_free_wq);
483 * Routines for managing virtual block devices (vbds).
485 static int xen_vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
488 struct xen_vbd *vbd = &blkif->vbd;
491 if ((operation != READ) && vbd->readonly)
494 if (likely(req->nr_sects)) {
495 blkif_sector_t end = req->sector_number + req->nr_sects;
497 if (unlikely(end < req->sector_number))
499 if (unlikely(end > vbd_sz(vbd)))
503 req->dev = vbd->pdevice;
504 req->bdev = vbd->bdev;
511 static void xen_vbd_resize(struct xen_blkif *blkif)
513 struct xen_vbd *vbd = &blkif->vbd;
514 struct xenbus_transaction xbt;
516 struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
517 unsigned long long new_size = vbd_sz(vbd);
519 pr_info("VBD Resize: Domid: %d, Device: (%d, %d)\n",
520 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
521 pr_info("VBD Resize: new size %llu\n", new_size);
522 vbd->size = new_size;
524 err = xenbus_transaction_start(&xbt);
526 pr_warn("Error starting transaction\n");
529 err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
530 (unsigned long long)vbd_sz(vbd));
532 pr_warn("Error writing new size\n");
536 * Write the current state; we will use this to synchronize
537 * the front-end. If the current state is "connected" the
538 * front-end will get the new size information online.
540 err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
542 pr_warn("Error writing the state\n");
546 err = xenbus_transaction_end(xbt, 0);
550 pr_warn("Error ending transaction\n");
553 xenbus_transaction_end(xbt, 1);
557 * Notification from the guest OS.
559 static void blkif_notify_work(struct xen_blkif *blkif)
561 blkif->waiting_reqs = 1;
565 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
567 blkif_notify_work(dev_id);
572 * SCHEDULER FUNCTIONS
575 static void print_stats(struct xen_blkif *blkif)
577 pr_info("(%s): oo %3llu | rd %4llu | wr %4llu | f %4llu"
578 " | ds %4llu | pg: %4u/%4d\n",
579 current->comm, blkif->st_oo_req,
580 blkif->st_rd_req, blkif->st_wr_req,
581 blkif->st_f_req, blkif->st_ds_req,
582 blkif->persistent_gnt_c,
583 xen_blkif_max_pgrants);
584 blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
585 blkif->st_rd_req = 0;
586 blkif->st_wr_req = 0;
587 blkif->st_oo_req = 0;
588 blkif->st_ds_req = 0;
591 int xen_blkif_schedule(void *arg)
593 struct xen_blkif *blkif = arg;
594 struct xen_vbd *vbd = &blkif->vbd;
595 unsigned long timeout;
598 unsigned int eoi_flags = XEN_EOI_FLAG_SPURIOUS;
600 while (!kthread_should_stop()) {
603 if (unlikely(vbd->size != vbd_sz(vbd)))
604 xen_vbd_resize(blkif);
606 timeout = msecs_to_jiffies(LRU_INTERVAL);
608 timeout = wait_event_interruptible_timeout(
610 blkif->waiting_reqs || kthread_should_stop(),
614 timeout = wait_event_interruptible_timeout(
615 blkif->pending_free_wq,
616 !list_empty(&blkif->pending_free) ||
617 kthread_should_stop(),
622 do_eoi = blkif->waiting_reqs;
624 blkif->waiting_reqs = 0;
625 smp_mb(); /* clear flag *before* checking for work */
627 ret = do_block_io_op(blkif, &eoi_flags);
629 blkif->waiting_reqs = 1;
631 wait_event_interruptible(blkif->shutdown_wq,
632 kthread_should_stop());
634 if (do_eoi && !blkif->waiting_reqs) {
635 xen_irq_lateeoi(blkif->irq, eoi_flags);
636 eoi_flags |= XEN_EOI_FLAG_SPURIOUS;
640 if (blkif->vbd.feature_gnt_persistent &&
641 time_after(jiffies, blkif->next_lru)) {
642 purge_persistent_gnt(blkif);
643 blkif->next_lru = jiffies + msecs_to_jiffies(LRU_INTERVAL);
646 /* Shrink if we have more than xen_blkif_max_buffer_pages */
647 shrink_free_pagepool(blkif, xen_blkif_max_buffer_pages);
649 if (log_stats && time_after(jiffies, blkif->st_print))
653 /* Drain pending purge work */
654 flush_work(&blkif->persistent_purge_work);
659 blkif->xenblkd = NULL;
665 * Remove persistent grants and empty the pool of free pages
667 void xen_blkbk_free_caches(struct xen_blkif *blkif)
669 /* Free all persistent grant pages */
670 if (!RB_EMPTY_ROOT(&blkif->persistent_gnts))
671 free_persistent_gnts(blkif, &blkif->persistent_gnts,
672 blkif->persistent_gnt_c);
674 BUG_ON(!RB_EMPTY_ROOT(&blkif->persistent_gnts));
675 blkif->persistent_gnt_c = 0;
677 /* Since we are shutting down remove all pages from the buffer */
678 shrink_free_pagepool(blkif, 0 /* All */);
681 static unsigned int xen_blkbk_unmap_prepare(
682 struct xen_blkif *blkif,
683 struct grant_page **pages,
685 struct gnttab_unmap_grant_ref *unmap_ops,
686 struct page **unmap_pages)
688 unsigned int i, invcount = 0;
690 for (i = 0; i < num; i++) {
691 if (pages[i]->persistent_gnt != NULL) {
692 put_persistent_gnt(blkif, pages[i]->persistent_gnt);
695 if (pages[i]->handle == BLKBACK_INVALID_HANDLE)
697 unmap_pages[invcount] = pages[i]->page;
698 gnttab_set_unmap_op(&unmap_ops[invcount], vaddr(pages[i]->page),
699 GNTMAP_host_map, pages[i]->handle);
700 pages[i]->handle = BLKBACK_INVALID_HANDLE;
707 static void xen_blkbk_unmap_and_respond_callback(int result, struct gntab_unmap_queue_data *data)
709 struct pending_req* pending_req = (struct pending_req*) (data->data);
710 struct xen_blkif *blkif = pending_req->blkif;
712 /* BUG_ON used to reproduce existing behaviour,
713 but is this the best way to deal with this? */
716 put_free_pages(blkif, data->pages, data->count);
717 make_response(blkif, pending_req->id,
718 pending_req->operation, pending_req->status);
719 free_req(blkif, pending_req);
721 * Make sure the request is freed before releasing blkif,
722 * or there could be a race between free_req and the
723 * cleanup done in xen_blkif_free during shutdown.
725 * NB: The fact that we might try to wake up pending_free_wq
726 * before drain_complete (in case there's a drain going on)
727 * it's not a problem with our current implementation
728 * because we can assure there's no thread waiting on
729 * pending_free_wq if there's a drain going on, but it has
730 * to be taken into account if the current model is changed.
732 if (atomic_dec_and_test(&blkif->inflight) && atomic_read(&blkif->drain)) {
733 complete(&blkif->drain_complete);
735 xen_blkif_put(blkif);
738 static void xen_blkbk_unmap_and_respond(struct pending_req *req)
740 struct gntab_unmap_queue_data* work = &req->gnttab_unmap_data;
741 struct xen_blkif *blkif = req->blkif;
742 struct grant_page **pages = req->segments;
743 unsigned int invcount;
745 invcount = xen_blkbk_unmap_prepare(blkif, pages, req->nr_segs,
746 req->unmap, req->unmap_pages);
749 work->done = xen_blkbk_unmap_and_respond_callback;
750 work->unmap_ops = req->unmap;
751 work->kunmap_ops = NULL;
752 work->pages = req->unmap_pages;
753 work->count = invcount;
755 gnttab_unmap_refs_async(&req->gnttab_unmap_data);
760 * Unmap the grant references.
762 * This could accumulate ops up to the batch size to reduce the number
763 * of hypercalls, but since this is only used in error paths there's
766 static void xen_blkbk_unmap(struct xen_blkif *blkif,
767 struct grant_page *pages[],
770 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
771 struct page *unmap_pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
772 unsigned int invcount = 0;
776 unsigned int batch = min(num, BLKIF_MAX_SEGMENTS_PER_REQUEST);
778 invcount = xen_blkbk_unmap_prepare(blkif, pages, batch,
781 ret = gnttab_unmap_refs(unmap, NULL, unmap_pages, invcount);
783 put_free_pages(blkif, unmap_pages, invcount);
790 static int xen_blkbk_map(struct xen_blkif *blkif,
791 struct grant_page *pages[],
794 struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
795 struct page *pages_to_gnt[BLKIF_MAX_SEGMENTS_PER_REQUEST];
796 struct persistent_gnt *persistent_gnt = NULL;
797 phys_addr_t addr = 0;
798 int i, seg_idx, new_map_idx;
801 int last_map = 0, map_until = 0;
802 int use_persistent_gnts;
804 use_persistent_gnts = (blkif->vbd.feature_gnt_persistent);
807 * Fill out preq.nr_sects with proper amount of sectors, and setup
808 * assign map[..] with the PFN of the page in our domain with the
809 * corresponding grant reference for each page.
812 for (i = map_until; i < num; i++) {
815 if (use_persistent_gnts)
816 persistent_gnt = get_persistent_gnt(
820 if (persistent_gnt) {
822 * We are using persistent grants and
823 * the grant is already mapped
825 pages[i]->page = persistent_gnt->page;
826 pages[i]->persistent_gnt = persistent_gnt;
828 if (get_free_page(blkif, &pages[i]->page)) {
829 put_free_pages(blkif, pages_to_gnt, segs_to_map);
833 addr = vaddr(pages[i]->page);
834 pages_to_gnt[segs_to_map] = pages[i]->page;
835 pages[i]->persistent_gnt = NULL;
836 flags = GNTMAP_host_map;
837 if (!use_persistent_gnts && ro)
838 flags |= GNTMAP_readonly;
839 gnttab_set_map_op(&map[segs_to_map++], addr,
840 flags, pages[i]->gref,
844 if (segs_to_map == BLKIF_MAX_SEGMENTS_PER_REQUEST)
849 ret = gnttab_map_refs(map, NULL, pages_to_gnt, segs_to_map);
852 * Now swizzle the MFN in our domain with the MFN from the other domain
853 * so that when we access vaddr(pending_req,i) it has the contents of
854 * the page from the other domain.
856 for (seg_idx = last_map, new_map_idx = 0; seg_idx < map_until; seg_idx++) {
857 if (!pages[seg_idx]->persistent_gnt) {
858 /* This is a newly mapped grant */
859 BUG_ON(new_map_idx >= segs_to_map);
860 if (unlikely(map[new_map_idx].status != 0)) {
861 pr_debug("invalid buffer -- could not remap it\n");
862 put_free_pages(blkif, &pages[seg_idx]->page, 1);
863 pages[seg_idx]->handle = BLKBACK_INVALID_HANDLE;
867 pages[seg_idx]->handle = map[new_map_idx].handle;
871 if (use_persistent_gnts &&
872 blkif->persistent_gnt_c < xen_blkif_max_pgrants) {
874 * We are using persistent grants, the grant is
875 * not mapped but we might have room for it.
877 persistent_gnt = kmalloc(sizeof(struct persistent_gnt),
879 if (!persistent_gnt) {
881 * If we don't have enough memory to
882 * allocate the persistent_gnt struct
883 * map this grant non-persistenly
887 persistent_gnt->gnt = map[new_map_idx].ref;
888 persistent_gnt->handle = map[new_map_idx].handle;
889 persistent_gnt->page = pages[seg_idx]->page;
890 if (add_persistent_gnt(blkif,
892 kfree(persistent_gnt);
893 persistent_gnt = NULL;
896 pages[seg_idx]->persistent_gnt = persistent_gnt;
897 pr_debug("grant %u added to the tree of persistent grants, using %u/%u\n",
898 persistent_gnt->gnt, blkif->persistent_gnt_c,
899 xen_blkif_max_pgrants);
902 if (use_persistent_gnts && !blkif->vbd.overflow_max_grants) {
903 blkif->vbd.overflow_max_grants = 1;
904 pr_debug("domain %u, device %#x is using maximum number of persistent grants\n",
905 blkif->domid, blkif->vbd.handle);
908 * We could not map this grant persistently, so use it as
909 * a non-persistent grant.
915 last_map = map_until;
916 if (!ret && map_until != num)
920 for (i = last_map; i < num; i++) {
921 /* Don't zap current batch's valid persistent grants. */
923 pages[i]->persistent_gnt = NULL;
924 pages[i]->handle = BLKBACK_INVALID_HANDLE;
930 static int xen_blkbk_map_seg(struct pending_req *pending_req)
934 rc = xen_blkbk_map(pending_req->blkif, pending_req->segments,
935 pending_req->nr_segs,
936 (pending_req->operation != BLKIF_OP_READ));
941 static int xen_blkbk_parse_indirect(struct blkif_request *req,
942 struct pending_req *pending_req,
943 struct seg_buf seg[],
944 struct phys_req *preq)
946 struct grant_page **pages = pending_req->indirect_pages;
947 struct xen_blkif *blkif = pending_req->blkif;
948 int indirect_grefs, rc, n, nseg, i;
949 struct blkif_request_segment *segments = NULL;
951 nseg = pending_req->nr_segs;
952 indirect_grefs = INDIRECT_PAGES(nseg);
953 BUG_ON(indirect_grefs > BLKIF_MAX_INDIRECT_PAGES_PER_REQUEST);
955 for (i = 0; i < indirect_grefs; i++)
956 pages[i]->gref = req->u.indirect.indirect_grefs[i];
958 rc = xen_blkbk_map(blkif, pages, indirect_grefs, true);
962 for (n = 0, i = 0; n < nseg; n++) {
963 uint8_t first_sect, last_sect;
965 if ((n % SEGS_PER_INDIRECT_FRAME) == 0) {
966 /* Map indirect segments */
968 kunmap_atomic(segments);
969 segments = kmap_atomic(pages[n/SEGS_PER_INDIRECT_FRAME]->page);
971 i = n % SEGS_PER_INDIRECT_FRAME;
973 pending_req->segments[n]->gref = segments[i].gref;
975 first_sect = READ_ONCE(segments[i].first_sect);
976 last_sect = READ_ONCE(segments[i].last_sect);
977 if (last_sect >= (XEN_PAGE_SIZE >> 9) || last_sect < first_sect) {
982 seg[n].nsec = last_sect - first_sect + 1;
983 seg[n].offset = first_sect << 9;
984 preq->nr_sects += seg[n].nsec;
989 kunmap_atomic(segments);
990 xen_blkbk_unmap(blkif, pages, indirect_grefs);
994 static int dispatch_discard_io(struct xen_blkif *blkif,
995 struct blkif_request *req)
998 int status = BLKIF_RSP_OKAY;
999 struct block_device *bdev = blkif->vbd.bdev;
1000 unsigned long secure;
1001 struct phys_req preq;
1003 xen_blkif_get(blkif);
1005 preq.sector_number = req->u.discard.sector_number;
1006 preq.nr_sects = req->u.discard.nr_sectors;
1008 err = xen_vbd_translate(&preq, blkif, WRITE);
1010 pr_warn("access denied: DISCARD [%llu->%llu] on dev=%04x\n",
1012 preq.sector_number + preq.nr_sects, blkif->vbd.pdevice);
1017 secure = (blkif->vbd.discard_secure &&
1018 (req->u.discard.flag & BLKIF_DISCARD_SECURE)) ?
1019 BLKDEV_DISCARD_SECURE : 0;
1021 err = blkdev_issue_discard(bdev, req->u.discard.sector_number,
1022 req->u.discard.nr_sectors,
1023 GFP_KERNEL, secure);
1025 if (err == -EOPNOTSUPP) {
1026 pr_debug("discard op failed, not supported\n");
1027 status = BLKIF_RSP_EOPNOTSUPP;
1029 status = BLKIF_RSP_ERROR;
1031 make_response(blkif, req->u.discard.id, req->operation, status);
1032 xen_blkif_put(blkif);
1036 static int dispatch_other_io(struct xen_blkif *blkif,
1037 struct blkif_request *req,
1038 struct pending_req *pending_req)
1040 free_req(blkif, pending_req);
1041 make_response(blkif, req->u.other.id, req->operation,
1042 BLKIF_RSP_EOPNOTSUPP);
1046 static void xen_blk_drain_io(struct xen_blkif *blkif)
1048 atomic_set(&blkif->drain, 1);
1050 if (atomic_read(&blkif->inflight) == 0)
1052 wait_for_completion_interruptible_timeout(
1053 &blkif->drain_complete, HZ);
1055 if (!atomic_read(&blkif->drain))
1057 } while (!kthread_should_stop());
1058 atomic_set(&blkif->drain, 0);
1062 * Completion callback on the bio's. Called as bh->b_end_io()
1065 static void __end_block_io_op(struct pending_req *pending_req, int error)
1067 /* An error fails the entire request. */
1068 if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
1069 (error == -EOPNOTSUPP)) {
1070 pr_debug("flush diskcache op failed, not supported\n");
1071 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->blkif->be, 0);
1072 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
1073 } else if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
1074 (error == -EOPNOTSUPP)) {
1075 pr_debug("write barrier op failed, not supported\n");
1076 xen_blkbk_barrier(XBT_NIL, pending_req->blkif->be, 0);
1077 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
1079 pr_debug("Buffer not up-to-date at end of operation,"
1080 " error=%d\n", error);
1081 pending_req->status = BLKIF_RSP_ERROR;
1085 * If all of the bio's have completed it is time to unmap
1086 * the grant references associated with 'request' and provide
1087 * the proper response on the ring.
1089 if (atomic_dec_and_test(&pending_req->pendcnt))
1090 xen_blkbk_unmap_and_respond(pending_req);
1096 static void end_block_io_op(struct bio *bio)
1098 __end_block_io_op(bio->bi_private, bio->bi_error);
1105 * Function to copy the from the ring buffer the 'struct blkif_request'
1106 * (which has the sectors we want, number of them, grant references, etc),
1107 * and transmute it to the block API to hand it over to the proper block disk.
1110 __do_block_io_op(struct xen_blkif *blkif, unsigned int *eoi_flags)
1112 union blkif_back_rings *blk_rings = &blkif->blk_rings;
1113 struct blkif_request req;
1114 struct pending_req *pending_req;
1118 rc = blk_rings->common.req_cons;
1119 rp = blk_rings->common.sring->req_prod;
1120 rmb(); /* Ensure we see queued requests up to 'rp'. */
1122 if (RING_REQUEST_PROD_OVERFLOW(&blk_rings->common, rp)) {
1123 rc = blk_rings->common.rsp_prod_pvt;
1124 pr_warn("Frontend provided bogus ring requests (%d - %d = %d). Halting ring processing on dev=%04x\n",
1125 rp, rc, rp - rc, blkif->vbd.pdevice);
1130 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
1133 /* We've seen a request, so clear spurious eoi flag. */
1134 *eoi_flags &= ~XEN_EOI_FLAG_SPURIOUS;
1136 if (kthread_should_stop()) {
1141 pending_req = alloc_req(blkif);
1142 if (NULL == pending_req) {
1148 switch (blkif->blk_protocol) {
1149 case BLKIF_PROTOCOL_NATIVE:
1150 memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
1152 case BLKIF_PROTOCOL_X86_32:
1153 blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
1155 case BLKIF_PROTOCOL_X86_64:
1156 blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
1161 blk_rings->common.req_cons = ++rc; /* before make_response() */
1163 /* Apply all sanity checks to /private copy/ of request. */
1166 switch (req.operation) {
1168 case BLKIF_OP_WRITE:
1169 case BLKIF_OP_WRITE_BARRIER:
1170 case BLKIF_OP_FLUSH_DISKCACHE:
1171 case BLKIF_OP_INDIRECT:
1172 if (dispatch_rw_block_io(blkif, &req, pending_req))
1175 case BLKIF_OP_DISCARD:
1176 free_req(blkif, pending_req);
1177 if (dispatch_discard_io(blkif, &req))
1181 if (dispatch_other_io(blkif, &req, pending_req))
1186 /* Yield point for this unbounded loop. */
1194 do_block_io_op(struct xen_blkif *blkif, unsigned int *eoi_flags)
1196 union blkif_back_rings *blk_rings = &blkif->blk_rings;
1200 more_to_do = __do_block_io_op(blkif, eoi_flags);
1204 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
1205 } while (more_to_do);
1210 * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
1211 * and call the 'submit_bio' to pass it to the underlying storage.
1213 static int dispatch_rw_block_io(struct xen_blkif *blkif,
1214 struct blkif_request *req,
1215 struct pending_req *pending_req)
1217 struct phys_req preq;
1218 struct seg_buf *seg = pending_req->seg;
1220 struct bio *bio = NULL;
1221 struct bio **biolist = pending_req->biolist;
1224 struct blk_plug plug;
1226 struct grant_page **pages = pending_req->segments;
1227 unsigned short req_operation;
1229 req_operation = req->operation == BLKIF_OP_INDIRECT ?
1230 req->u.indirect.indirect_op : req->operation;
1232 if ((req->operation == BLKIF_OP_INDIRECT) &&
1233 (req_operation != BLKIF_OP_READ) &&
1234 (req_operation != BLKIF_OP_WRITE)) {
1235 pr_debug("Invalid indirect operation (%u)\n", req_operation);
1239 switch (req_operation) {
1244 case BLKIF_OP_WRITE:
1246 operation = WRITE_ODIRECT;
1248 case BLKIF_OP_WRITE_BARRIER:
1250 case BLKIF_OP_FLUSH_DISKCACHE:
1252 operation = WRITE_FLUSH;
1255 operation = 0; /* make gcc happy */
1260 /* Check that the number of segments is sane. */
1261 nseg = req->operation == BLKIF_OP_INDIRECT ?
1262 req->u.indirect.nr_segments : req->u.rw.nr_segments;
1264 if (unlikely(nseg == 0 && operation != WRITE_FLUSH) ||
1265 unlikely((req->operation != BLKIF_OP_INDIRECT) &&
1266 (nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) ||
1267 unlikely((req->operation == BLKIF_OP_INDIRECT) &&
1268 (nseg > MAX_INDIRECT_SEGMENTS))) {
1269 pr_debug("Bad number of segments in request (%d)\n", nseg);
1270 /* Haven't submitted any bio's yet. */
1276 pending_req->blkif = blkif;
1277 pending_req->id = req->u.rw.id;
1278 pending_req->operation = req_operation;
1279 pending_req->status = BLKIF_RSP_OKAY;
1280 pending_req->nr_segs = nseg;
1282 if (req->operation != BLKIF_OP_INDIRECT) {
1283 preq.dev = req->u.rw.handle;
1284 preq.sector_number = req->u.rw.sector_number;
1285 for (i = 0; i < nseg; i++) {
1286 pages[i]->gref = req->u.rw.seg[i].gref;
1287 seg[i].nsec = req->u.rw.seg[i].last_sect -
1288 req->u.rw.seg[i].first_sect + 1;
1289 seg[i].offset = (req->u.rw.seg[i].first_sect << 9);
1290 if ((req->u.rw.seg[i].last_sect >= (XEN_PAGE_SIZE >> 9)) ||
1291 (req->u.rw.seg[i].last_sect <
1292 req->u.rw.seg[i].first_sect))
1294 preq.nr_sects += seg[i].nsec;
1297 preq.dev = req->u.indirect.handle;
1298 preq.sector_number = req->u.indirect.sector_number;
1299 if (xen_blkbk_parse_indirect(req, pending_req, seg, &preq))
1303 if (xen_vbd_translate(&preq, blkif, operation) != 0) {
1304 pr_debug("access denied: %s of [%llu,%llu] on dev=%04x\n",
1305 operation == READ ? "read" : "write",
1307 preq.sector_number + preq.nr_sects,
1308 blkif->vbd.pdevice);
1313 * This check _MUST_ be done after xen_vbd_translate as the preq.bdev
1316 for (i = 0; i < nseg; i++) {
1317 if (((int)preq.sector_number|(int)seg[i].nsec) &
1318 ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
1319 pr_debug("Misaligned I/O request from domain %d\n",
1325 /* Wait on all outstanding I/O's and once that has been completed
1326 * issue the WRITE_FLUSH.
1329 xen_blk_drain_io(pending_req->blkif);
1332 * If we have failed at this point, we need to undo the M2P override,
1333 * set gnttab_set_unmap_op on all of the grant references and perform
1334 * the hypercall to unmap the grants - that is all done in
1337 if (xen_blkbk_map_seg(pending_req))
1341 * This corresponding xen_blkif_put is done in __end_block_io_op, or
1342 * below (in "!bio") if we are handling a BLKIF_OP_DISCARD.
1344 xen_blkif_get(blkif);
1345 atomic_inc(&blkif->inflight);
1347 for (i = 0; i < nseg; i++) {
1348 while ((bio == NULL) ||
1352 seg[i].offset) == 0)) {
1354 int nr_iovecs = min_t(int, (nseg-i), BIO_MAX_PAGES);
1355 bio = bio_alloc(GFP_KERNEL, nr_iovecs);
1356 if (unlikely(bio == NULL))
1359 biolist[nbio++] = bio;
1360 bio->bi_bdev = preq.bdev;
1361 bio->bi_private = pending_req;
1362 bio->bi_end_io = end_block_io_op;
1363 bio->bi_iter.bi_sector = preq.sector_number;
1366 preq.sector_number += seg[i].nsec;
1369 /* This will be hit if the operation was a flush or discard. */
1371 BUG_ON(operation != WRITE_FLUSH);
1373 bio = bio_alloc(GFP_KERNEL, 0);
1374 if (unlikely(bio == NULL))
1377 biolist[nbio++] = bio;
1378 bio->bi_bdev = preq.bdev;
1379 bio->bi_private = pending_req;
1380 bio->bi_end_io = end_block_io_op;
1383 atomic_set(&pending_req->pendcnt, nbio);
1384 blk_start_plug(&plug);
1386 for (i = 0; i < nbio; i++)
1387 submit_bio(operation, biolist[i]);
1389 /* Let the I/Os go.. */
1390 blk_finish_plug(&plug);
1392 if (operation == READ)
1393 blkif->st_rd_sect += preq.nr_sects;
1394 else if (operation & WRITE)
1395 blkif->st_wr_sect += preq.nr_sects;
1400 xen_blkbk_unmap(blkif, pending_req->segments,
1401 pending_req->nr_segs);
1403 /* Haven't submitted any bio's yet. */
1404 make_response(blkif, req->u.rw.id, req_operation, BLKIF_RSP_ERROR);
1405 free_req(blkif, pending_req);
1406 msleep(1); /* back off a bit */
1410 for (i = 0; i < nbio; i++)
1411 bio_put(biolist[i]);
1412 atomic_set(&pending_req->pendcnt, 1);
1413 __end_block_io_op(pending_req, -EINVAL);
1414 msleep(1); /* back off a bit */
1421 * Put a response on the ring on how the operation fared.
1423 static void make_response(struct xen_blkif *blkif, u64 id,
1424 unsigned short op, int st)
1426 struct blkif_response *resp;
1427 unsigned long flags;
1428 union blkif_back_rings *blk_rings = &blkif->blk_rings;
1431 spin_lock_irqsave(&blkif->blk_ring_lock, flags);
1432 /* Place on the response ring for the relevant domain. */
1433 switch (blkif->blk_protocol) {
1434 case BLKIF_PROTOCOL_NATIVE:
1435 resp = RING_GET_RESPONSE(&blk_rings->native,
1436 blk_rings->native.rsp_prod_pvt);
1438 case BLKIF_PROTOCOL_X86_32:
1439 resp = RING_GET_RESPONSE(&blk_rings->x86_32,
1440 blk_rings->x86_32.rsp_prod_pvt);
1442 case BLKIF_PROTOCOL_X86_64:
1443 resp = RING_GET_RESPONSE(&blk_rings->x86_64,
1444 blk_rings->x86_64.rsp_prod_pvt);
1451 resp->operation = op;
1454 blk_rings->common.rsp_prod_pvt++;
1455 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
1456 spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
1458 notify_remote_via_irq(blkif->irq);
1461 static int __init xen_blkif_init(void)
1468 if (xen_blkif_max_ring_order > XENBUS_MAX_RING_GRANT_ORDER) {
1469 pr_info("Invalid max_ring_order (%d), will use default max: %d.\n",
1470 xen_blkif_max_ring_order, XENBUS_MAX_RING_GRANT_ORDER);
1471 xen_blkif_max_ring_order = XENBUS_MAX_RING_GRANT_ORDER;
1474 rc = xen_blkif_interface_init();
1478 rc = xen_blkif_xenbus_init();
1486 module_init(xen_blkif_init);
1488 MODULE_LICENSE("Dual BSD/GPL");
1489 MODULE_ALIAS("xen-backend:vbd");