1 /******************************************************************************
2  *
3  * Back-end of the driver for virtual block devices. This portion of the
4  * driver exports a 'unified' block-device interface that can be accessed
5  * by any operating system that implements a compatible front end. A
6  * reference front-end implementation can be found in:
7  *  drivers/block/xen-blkfront.c
8  *
9  * Copyright (c) 2003-2004, Keir Fraser & Steve Hand
10  * Copyright (c) 2005, Christopher Clark
11  *
12  * This program is free software; you can redistribute it and/or
13  * modify it under the terms of the GNU General Public License version 2
14  * as published by the Free Software Foundation; or, when distributed
15  * separately from the Linux kernel or incorporated into other
16  * software packages, subject to the following license:
17  *
18  * Permission is hereby granted, free of charge, to any person obtaining a copy
19  * of this source file (the "Software"), to deal in the Software without
20  * restriction, including without limitation the rights to use, copy, modify,
21  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
22  * and to permit persons to whom the Software is furnished to do so, subject to
23  * the following conditions:
24  *
25  * The above copyright notice and this permission notice shall be included in
26  * all copies or substantial portions of the Software.
27  *
28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
29  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
30  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
31  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
32  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
33  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
34  * IN THE SOFTWARE.
35  */
36 
37 #define pr_fmt(fmt) "xen-blkback: " fmt
38 
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>
45 
46 #include <xen/events.h>
47 #include <xen/page.h>
48 #include <xen/xen.h>
49 #include <asm/xen/hypervisor.h>
50 #include <asm/xen/hypercall.h>
51 #include <xen/balloon.h>
52 #include <xen/grant_table.h>
53 #include "common.h"
54 
55 /*
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.
59  *
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
62  * IO workloads.
63  */
64 
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");
69 
70 /*
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.
76  *
77  * When the list of persistent grants is full we clean it up using a LRU
78  * algorithm.
79  */
80 
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");
85 
86 /*
87  * Maximum number of rings/queues blkback supports, allow as many queues as there
88  * are CPUs if user has not specified a value.
89  */
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.");
95 
96 /*
97  * Maximum order of pages to be used for the shared ring between front and
98  * backend, 4KB page granularity is used.
99  */
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");
103 /*
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.
107  */
108 #define LRU_INTERVAL 100
109 
110 /*
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
113  * execution.
114  */
115 #define LRU_PERCENT_CLEAN 5
116 
117 /* Run-time switchable: /sys/module/blkback/parameters/ */
118 static unsigned int log_stats;
119 module_param(log_stats, int, 0644);
120 
121 #define BLKBACK_INVALID_HANDLE (~0)
122 
123 /* Number of free pages to remove on each call to gnttab_free_pages */
124 #define NUM_BATCH_FREE_PAGES 10
125 
126 static inline int get_free_page(struct xen_blkif_ring *ring, struct page **page)
127 {
128 	unsigned long flags;
129 
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);
135 	}
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);
141 
142 	return 0;
143 }
144 
145 static inline void put_free_pages(struct xen_blkif_ring *ring, struct page **page,
146                                   int num)
147 {
148 	unsigned long flags;
149 	int i;
150 
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);
156 }
157 
158 static inline void shrink_free_pagepool(struct xen_blkif_ring *ring, int num)
159 {
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;
163 	unsigned long flags;
164 
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,
169 		                                   struct page, lru);
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);
176 			num_pages = 0;
177 		}
178 	}
179 	spin_unlock_irqrestore(&ring->free_pages_lock, flags);
180 	if (num_pages != 0)
181 		gnttab_free_pages(num_pages, page);
182 }
183 
184 #define vaddr(page) ((unsigned long)pfn_to_kaddr(page_to_pfn(page)))
185 
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);
192 
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)
199 
200 
201 /*
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.
205  *
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.
210  */
211 static int add_persistent_gnt(struct xen_blkif_ring *ring,
212 			       struct persistent_gnt *persistent_gnt)
213 {
214 	struct rb_node **new = NULL, *parent = NULL;
215 	struct persistent_gnt *this;
216 	struct xen_blkif *blkif = ring->blkif;
217 
218 	if (ring->persistent_gnt_c >= xen_blkif_max_pgrants) {
219 		if (!blkif->vbd.overflow_max_grants)
220 			blkif->vbd.overflow_max_grants = 1;
221 		return -EBUSY;
222 	}
223 	/* Figure out where to put new node */
224 	new = &ring->persistent_gnts.rb_node;
225 	while (*new) {
226 		this = container_of(*new, struct persistent_gnt, node);
227 
228 		parent = *new;
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);
233 		else {
234 			pr_alert_ratelimited("trying to add a gref that's already in the tree\n");
235 			return -EINVAL;
236 		}
237 	}
238 
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);
246 	return 0;
247 }
248 
249 static struct persistent_gnt *get_persistent_gnt(struct xen_blkif_ring *ring,
250 						 grant_ref_t gref)
251 {
252 	struct persistent_gnt *data;
253 	struct rb_node *node = NULL;
254 
255 	node = ring->persistent_gnts.rb_node;
256 	while (node) {
257 		data = container_of(node, struct persistent_gnt, node);
258 
259 		if (gref < data->gnt)
260 			node = node->rb_left;
261 		else if (gref > data->gnt)
262 			node = node->rb_right;
263 		else {
264 			if(test_bit(PERSISTENT_GNT_ACTIVE, data->flags)) {
265 				pr_alert_ratelimited("requesting a grant already in use\n");
266 				return NULL;
267 			}
268 			set_bit(PERSISTENT_GNT_ACTIVE, data->flags);
269 			atomic_inc(&ring->persistent_gnt_in_use);
270 			return data;
271 		}
272 	}
273 	return NULL;
274 }
275 
276 static void put_persistent_gnt(struct xen_blkif_ring *ring,
277                                struct persistent_gnt *persistent_gnt)
278 {
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);
284 }
285 
286 static void free_persistent_gnts(struct xen_blkif_ring *ring, struct rb_root *root,
287                                  unsigned int num)
288 {
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;
292 	struct rb_node *n;
293 	int segs_to_unmap = 0;
294 	struct gntab_unmap_queue_data unmap_data;
295 
296 	unmap_data.pages = pages;
297 	unmap_data.unmap_ops = unmap;
298 	unmap_data.kunmap_ops = NULL;
299 
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)),
306 			GNTMAP_host_map,
307 			persistent_gnt->handle);
308 
309 		pages[segs_to_unmap] = persistent_gnt->page;
310 
311 		if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST ||
312 			!rb_next(&persistent_gnt->node)) {
313 
314 			unmap_data.count = segs_to_unmap;
315 			BUG_ON(gnttab_unmap_refs_sync(&unmap_data));
316 
317 			put_free_pages(ring, pages, segs_to_unmap);
318 			segs_to_unmap = 0;
319 		}
320 
321 		rb_erase(&persistent_gnt->node, root);
322 		kfree(persistent_gnt);
323 		num--;
324 	}
325 	BUG_ON(num != 0);
326 }
327 
328 void xen_blkbk_unmap_purged_grants(struct work_struct *work)
329 {
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;
336 
337 	unmap_data.pages = pages;
338 	unmap_data.unmap_ops = unmap;
339 	unmap_data.kunmap_ops = NULL;
340 
341 	while(!list_empty(&ring->persistent_purge_list)) {
342 		persistent_gnt = list_first_entry(&ring->persistent_purge_list,
343 		                                  struct persistent_gnt,
344 		                                  remove_node);
345 		list_del(&persistent_gnt->remove_node);
346 
347 		gnttab_set_unmap_op(&unmap[segs_to_unmap],
348 			vaddr(persistent_gnt->page),
349 			GNTMAP_host_map,
350 			persistent_gnt->handle);
351 
352 		pages[segs_to_unmap] = persistent_gnt->page;
353 
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);
358 			segs_to_unmap = 0;
359 		}
360 		kfree(persistent_gnt);
361 	}
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);
366 	}
367 }
368 
369 static void purge_persistent_gnt(struct xen_blkif_ring *ring)
370 {
371 	struct persistent_gnt *persistent_gnt;
372 	struct rb_node *n;
373 	unsigned int num_clean, total;
374 	bool scan_used = false, clean_used = false;
375 	struct rb_root *root;
376 
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)) {
380 		goto out;
381 	}
382 
383 	if (work_busy(&ring->persistent_purge_work)) {
384 		pr_alert_ratelimited("Scheduled work from previous purge is still busy, cannot purge list\n");
385 		goto out;
386 	}
387 
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))))
393 		goto out;
394 
395 	/*
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
401          * number of grants.
402 	 */
403 
404 	total = num_clean;
405 
406 	pr_debug("Going to purge %u persistent grants\n", num_clean);
407 
408 	BUG_ON(!list_empty(&ring->persistent_purge_list));
409 	root = &ring->persistent_gnts;
410 purge_list:
411 	foreach_grant_safe(persistent_gnt, n, root, node) {
412 		BUG_ON(persistent_gnt->handle ==
413 			BLKBACK_INVALID_HANDLE);
414 
415 		if (clean_used) {
416 			clear_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags);
417 			continue;
418 		}
419 
420 		if (test_bit(PERSISTENT_GNT_ACTIVE, persistent_gnt->flags))
421 			continue;
422 		if (!scan_used &&
423 		    (test_bit(PERSISTENT_GNT_WAS_ACTIVE, persistent_gnt->flags)))
424 			continue;
425 
426 		rb_erase(&persistent_gnt->node, root);
427 		list_add(&persistent_gnt->remove_node,
428 			 &ring->persistent_purge_list);
429 		if (--num_clean == 0)
430 			goto finished;
431 	}
432 	/*
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
436 	 */
437 	if (!scan_used && !clean_used) {
438 		pr_debug("Still missing %u purged frames\n", num_clean);
439 		scan_used = true;
440 		goto purge_list;
441 	}
442 finished:
443 	if (!clean_used) {
444 		pr_debug("Finished scanning for grants to clean, removing used flag\n");
445 		clean_used = true;
446 		goto purge_list;
447 	}
448 
449 	ring->persistent_gnt_c -= (total - num_clean);
450 	ring->blkif->vbd.overflow_max_grants = 0;
451 
452 	/* We can defer this work */
453 	schedule_work(&ring->persistent_purge_work);
454 	pr_debug("Purged %u/%u\n", (total - num_clean), total);
455 
456 out:
457 	return;
458 }
459 
460 /*
461  * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
462  */
463 static struct pending_req *alloc_req(struct xen_blkif_ring *ring)
464 {
465 	struct pending_req *req = NULL;
466 	unsigned long flags;
467 
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,
471 				 free_list);
472 		list_del(&req->free_list);
473 	}
474 	spin_unlock_irqrestore(&ring->pending_free_lock, flags);
475 	return req;
476 }
477 
478 /*
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.
481  */
482 static void free_req(struct xen_blkif_ring *ring, struct pending_req *req)
483 {
484 	unsigned long flags;
485 	int was_empty;
486 
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);
491 	if (was_empty)
492 		wake_up(&ring->pending_free_wq);
493 }
494 
495 /*
496  * Routines for managing virtual block devices (vbds).
497  */
498 static int xen_vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
499 			     int operation)
500 {
501 	struct xen_vbd *vbd = &blkif->vbd;
502 	int rc = -EACCES;
503 
504 	if ((operation != READ) && vbd->readonly)
505 		goto out;
506 
507 	if (likely(req->nr_sects)) {
508 		blkif_sector_t end = req->sector_number + req->nr_sects;
509 
510 		if (unlikely(end < req->sector_number))
511 			goto out;
512 		if (unlikely(end > vbd_sz(vbd)))
513 			goto out;
514 	}
515 
516 	req->dev  = vbd->pdevice;
517 	req->bdev = vbd->bdev;
518 	rc = 0;
519 
520  out:
521 	return rc;
522 }
523 
524 static void xen_vbd_resize(struct xen_blkif *blkif)
525 {
526 	struct xen_vbd *vbd = &blkif->vbd;
527 	struct xenbus_transaction xbt;
528 	int err;
529 	struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
530 	unsigned long long new_size = vbd_sz(vbd);
531 
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;
536 again:
537 	err = xenbus_transaction_start(&xbt);
538 	if (err) {
539 		pr_warn("Error starting transaction\n");
540 		return;
541 	}
542 	err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
543 			    (unsigned long long)vbd_sz(vbd));
544 	if (err) {
545 		pr_warn("Error writing new size\n");
546 		goto abort;
547 	}
548 	/*
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.
552 	 */
553 	err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
554 	if (err) {
555 		pr_warn("Error writing the state\n");
556 		goto abort;
557 	}
558 
559 	err = xenbus_transaction_end(xbt, 0);
560 	if (err == -EAGAIN)
561 		goto again;
562 	if (err)
563 		pr_warn("Error ending transaction\n");
564 	return;
565 abort:
566 	xenbus_transaction_end(xbt, 1);
567 }
568 
569 /*
570  * Notification from the guest OS.
571  */
572 static void blkif_notify_work(struct xen_blkif_ring *ring)
573 {
574 	ring->waiting_reqs = 1;
575 	wake_up(&ring->wq);
576 }
577 
578 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
579 {
580 	blkif_notify_work(dev_id);
581 	return IRQ_HANDLED;
582 }
583 
584 /*
585  * SCHEDULER FUNCTIONS
586  */
587 
588 static void print_stats(struct xen_blkif_ring *ring)
589 {
590 	pr_info("(%s): oo %3llu  |  rd %4llu  |  wr %4llu  |  f %4llu"
591 		 "  |  ds %4llu | pg: %4u/%4d\n",
592 		 current->comm, ring->st_oo_req,
593 		 ring->st_rd_req, ring->st_wr_req,
594 		 ring->st_f_req, ring->st_ds_req,
595 		 ring->persistent_gnt_c,
596 		 xen_blkif_max_pgrants);
597 	ring->st_print = jiffies + msecs_to_jiffies(10 * 1000);
598 	ring->st_rd_req = 0;
599 	ring->st_wr_req = 0;
600 	ring->st_oo_req = 0;
601 	ring->st_ds_req = 0;
602 }
603 
604 int xen_blkif_schedule(void *arg)
605 {
606 	struct xen_blkif_ring *ring = arg;
607 	struct xen_blkif *blkif = ring->blkif;
608 	struct xen_vbd *vbd = &blkif->vbd;
609 	unsigned long timeout;
610 	int ret;
611 
612 	xen_blkif_get(blkif);
613 
614 	set_freezable();
615 	while (!kthread_should_stop()) {
616 		if (try_to_freeze())
617 			continue;
618 		if (unlikely(vbd->size != vbd_sz(vbd)))
619 			xen_vbd_resize(blkif);
620 
621 		timeout = msecs_to_jiffies(LRU_INTERVAL);
622 
623 		timeout = wait_event_interruptible_timeout(
624 			ring->wq,
625 			ring->waiting_reqs || kthread_should_stop(),
626 			timeout);
627 		if (timeout == 0)
628 			goto purge_gnt_list;
629 		timeout = wait_event_interruptible_timeout(
630 			ring->pending_free_wq,
631 			!list_empty(&ring->pending_free) ||
632 			kthread_should_stop(),
633 			timeout);
634 		if (timeout == 0)
635 			goto purge_gnt_list;
636 
637 		ring->waiting_reqs = 0;
638 		smp_mb(); /* clear flag *before* checking for work */
639 
640 		ret = do_block_io_op(ring);
641 		if (ret > 0)
642 			ring->waiting_reqs = 1;
643 		if (ret == -EACCES)
644 			wait_event_interruptible(ring->shutdown_wq,
645 						 kthread_should_stop());
646 
647 purge_gnt_list:
648 		if (blkif->vbd.feature_gnt_persistent &&
649 		    time_after(jiffies, ring->next_lru)) {
650 			purge_persistent_gnt(ring);
651 			ring->next_lru = jiffies + msecs_to_jiffies(LRU_INTERVAL);
652 		}
653 
654 		/* Shrink if we have more than xen_blkif_max_buffer_pages */
655 		shrink_free_pagepool(ring, xen_blkif_max_buffer_pages);
656 
657 		if (log_stats && time_after(jiffies, ring->st_print))
658 			print_stats(ring);
659 	}
660 
661 	/* Drain pending purge work */
662 	flush_work(&ring->persistent_purge_work);
663 
664 	if (log_stats)
665 		print_stats(ring);
666 
667 	ring->xenblkd = NULL;
668 	xen_blkif_put(blkif);
669 
670 	return 0;
671 }
672 
673 /*
674  * Remove persistent grants and empty the pool of free pages
675  */
676 void xen_blkbk_free_caches(struct xen_blkif_ring *ring)
677 {
678 	/* Free all persistent grant pages */
679 	if (!RB_EMPTY_ROOT(&ring->persistent_gnts))
680 		free_persistent_gnts(ring, &ring->persistent_gnts,
681 			ring->persistent_gnt_c);
682 
683 	BUG_ON(!RB_EMPTY_ROOT(&ring->persistent_gnts));
684 	ring->persistent_gnt_c = 0;
685 
686 	/* Since we are shutting down remove all pages from the buffer */
687 	shrink_free_pagepool(ring, 0 /* All */);
688 }
689 
690 static unsigned int xen_blkbk_unmap_prepare(
691 	struct xen_blkif_ring *ring,
692 	struct grant_page **pages,
693 	unsigned int num,
694 	struct gnttab_unmap_grant_ref *unmap_ops,
695 	struct page **unmap_pages)
696 {
697 	unsigned int i, invcount = 0;
698 
699 	for (i = 0; i < num; i++) {
700 		if (pages[i]->persistent_gnt != NULL) {
701 			put_persistent_gnt(ring, pages[i]->persistent_gnt);
702 			continue;
703 		}
704 		if (pages[i]->handle == BLKBACK_INVALID_HANDLE)
705 			continue;
706 		unmap_pages[invcount] = pages[i]->page;
707 		gnttab_set_unmap_op(&unmap_ops[invcount], vaddr(pages[i]->page),
708 				    GNTMAP_host_map, pages[i]->handle);
709 		pages[i]->handle = BLKBACK_INVALID_HANDLE;
710 		invcount++;
711        }
712 
713        return invcount;
714 }
715 
716 static void xen_blkbk_unmap_and_respond_callback(int result, struct gntab_unmap_queue_data *data)
717 {
718 	struct pending_req *pending_req = (struct pending_req *)(data->data);
719 	struct xen_blkif_ring *ring = pending_req->ring;
720 	struct xen_blkif *blkif = ring->blkif;
721 
722 	/* BUG_ON used to reproduce existing behaviour,
723 	   but is this the best way to deal with this? */
724 	BUG_ON(result);
725 
726 	put_free_pages(ring, data->pages, data->count);
727 	make_response(ring, pending_req->id,
728 		      pending_req->operation, pending_req->status);
729 	free_req(ring, pending_req);
730 	/*
731 	 * Make sure the request is freed before releasing blkif,
732 	 * or there could be a race between free_req and the
733 	 * cleanup done in xen_blkif_free during shutdown.
734 	 *
735 	 * NB: The fact that we might try to wake up pending_free_wq
736 	 * before drain_complete (in case there's a drain going on)
737 	 * it's not a problem with our current implementation
738 	 * because we can assure there's no thread waiting on
739 	 * pending_free_wq if there's a drain going on, but it has
740 	 * to be taken into account if the current model is changed.
741 	 */
742 	if (atomic_dec_and_test(&ring->inflight) && atomic_read(&blkif->drain)) {
743 		complete(&blkif->drain_complete);
744 	}
745 	xen_blkif_put(blkif);
746 }
747 
748 static void xen_blkbk_unmap_and_respond(struct pending_req *req)
749 {
750 	struct gntab_unmap_queue_data* work = &req->gnttab_unmap_data;
751 	struct xen_blkif_ring *ring = req->ring;
752 	struct grant_page **pages = req->segments;
753 	unsigned int invcount;
754 
755 	invcount = xen_blkbk_unmap_prepare(ring, pages, req->nr_segs,
756 					   req->unmap, req->unmap_pages);
757 
758 	work->data = req;
759 	work->done = xen_blkbk_unmap_and_respond_callback;
760 	work->unmap_ops = req->unmap;
761 	work->kunmap_ops = NULL;
762 	work->pages = req->unmap_pages;
763 	work->count = invcount;
764 
765 	gnttab_unmap_refs_async(&req->gnttab_unmap_data);
766 }
767 
768 
769 /*
770  * Unmap the grant references.
771  *
772  * This could accumulate ops up to the batch size to reduce the number
773  * of hypercalls, but since this is only used in error paths there's
774  * no real need.
775  */
776 static void xen_blkbk_unmap(struct xen_blkif_ring *ring,
777                             struct grant_page *pages[],
778                             int num)
779 {
780 	struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
781 	struct page *unmap_pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
782 	unsigned int invcount = 0;
783 	int ret;
784 
785 	while (num) {
786 		unsigned int batch = min(num, BLKIF_MAX_SEGMENTS_PER_REQUEST);
787 
788 		invcount = xen_blkbk_unmap_prepare(ring, pages, batch,
789 						   unmap, unmap_pages);
790 		if (invcount) {
791 			ret = gnttab_unmap_refs(unmap, NULL, unmap_pages, invcount);
792 			BUG_ON(ret);
793 			put_free_pages(ring, unmap_pages, invcount);
794 		}
795 		pages += batch;
796 		num -= batch;
797 	}
798 }
799 
800 static int xen_blkbk_map(struct xen_blkif_ring *ring,
801 			 struct grant_page *pages[],
802 			 int num, bool ro)
803 {
804 	struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
805 	struct page *pages_to_gnt[BLKIF_MAX_SEGMENTS_PER_REQUEST];
806 	struct persistent_gnt *persistent_gnt = NULL;
807 	phys_addr_t addr = 0;
808 	int i, seg_idx, new_map_idx;
809 	int segs_to_map = 0;
810 	int ret = 0;
811 	int last_map = 0, map_until = 0;
812 	int use_persistent_gnts;
813 	struct xen_blkif *blkif = ring->blkif;
814 
815 	use_persistent_gnts = (blkif->vbd.feature_gnt_persistent);
816 
817 	/*
818 	 * Fill out preq.nr_sects with proper amount of sectors, and setup
819 	 * assign map[..] with the PFN of the page in our domain with the
820 	 * corresponding grant reference for each page.
821 	 */
822 again:
823 	for (i = map_until; i < num; i++) {
824 		uint32_t flags;
825 
826 		if (use_persistent_gnts) {
827 			persistent_gnt = get_persistent_gnt(
828 				ring,
829 				pages[i]->gref);
830 		}
831 
832 		if (persistent_gnt) {
833 			/*
834 			 * We are using persistent grants and
835 			 * the grant is already mapped
836 			 */
837 			pages[i]->page = persistent_gnt->page;
838 			pages[i]->persistent_gnt = persistent_gnt;
839 		} else {
840 			if (get_free_page(ring, &pages[i]->page))
841 				goto out_of_memory;
842 			addr = vaddr(pages[i]->page);
843 			pages_to_gnt[segs_to_map] = pages[i]->page;
844 			pages[i]->persistent_gnt = NULL;
845 			flags = GNTMAP_host_map;
846 			if (!use_persistent_gnts && ro)
847 				flags |= GNTMAP_readonly;
848 			gnttab_set_map_op(&map[segs_to_map++], addr,
849 					  flags, pages[i]->gref,
850 					  blkif->domid);
851 		}
852 		map_until = i + 1;
853 		if (segs_to_map == BLKIF_MAX_SEGMENTS_PER_REQUEST)
854 			break;
855 	}
856 
857 	if (segs_to_map) {
858 		ret = gnttab_map_refs(map, NULL, pages_to_gnt, segs_to_map);
859 		BUG_ON(ret);
860 	}
861 
862 	/*
863 	 * Now swizzle the MFN in our domain with the MFN from the other domain
864 	 * so that when we access vaddr(pending_req,i) it has the contents of
865 	 * the page from the other domain.
866 	 */
867 	for (seg_idx = last_map, new_map_idx = 0; seg_idx < map_until; seg_idx++) {
868 		if (!pages[seg_idx]->persistent_gnt) {
869 			/* This is a newly mapped grant */
870 			BUG_ON(new_map_idx >= segs_to_map);
871 			if (unlikely(map[new_map_idx].status != 0)) {
872 				pr_debug("invalid buffer -- could not remap it\n");
873 				put_free_pages(ring, &pages[seg_idx]->page, 1);
874 				pages[seg_idx]->handle = BLKBACK_INVALID_HANDLE;
875 				ret |= 1;
876 				goto next;
877 			}
878 			pages[seg_idx]->handle = map[new_map_idx].handle;
879 		} else {
880 			continue;
881 		}
882 		if (use_persistent_gnts &&
883 		    ring->persistent_gnt_c < xen_blkif_max_pgrants) {
884 			/*
885 			 * We are using persistent grants, the grant is
886 			 * not mapped but we might have room for it.
887 			 */
888 			persistent_gnt = kmalloc(sizeof(struct persistent_gnt),
889 				                 GFP_KERNEL);
890 			if (!persistent_gnt) {
891 				/*
892 				 * If we don't have enough memory to
893 				 * allocate the persistent_gnt struct
894 				 * map this grant non-persistenly
895 				 */
896 				goto next;
897 			}
898 			persistent_gnt->gnt = map[new_map_idx].ref;
899 			persistent_gnt->handle = map[new_map_idx].handle;
900 			persistent_gnt->page = pages[seg_idx]->page;
901 			if (add_persistent_gnt(ring,
902 			                       persistent_gnt)) {
903 				kfree(persistent_gnt);
904 				persistent_gnt = NULL;
905 				goto next;
906 			}
907 			pages[seg_idx]->persistent_gnt = persistent_gnt;
908 			pr_debug("grant %u added to the tree of persistent grants, using %u/%u\n",
909 				 persistent_gnt->gnt, ring->persistent_gnt_c,
910 				 xen_blkif_max_pgrants);
911 			goto next;
912 		}
913 		if (use_persistent_gnts && !blkif->vbd.overflow_max_grants) {
914 			blkif->vbd.overflow_max_grants = 1;
915 			pr_debug("domain %u, device %#x is using maximum number of persistent grants\n",
916 			         blkif->domid, blkif->vbd.handle);
917 		}
918 		/*
919 		 * We could not map this grant persistently, so use it as
920 		 * a non-persistent grant.
921 		 */
922 next:
923 		new_map_idx++;
924 	}
925 	segs_to_map = 0;
926 	last_map = map_until;
927 	if (map_until != num)
928 		goto again;
929 
930 	return ret;
931 
932 out_of_memory:
933 	pr_alert("%s: out of memory\n", __func__);
934 	put_free_pages(ring, pages_to_gnt, segs_to_map);
935 	return -ENOMEM;
936 }
937 
938 static int xen_blkbk_map_seg(struct pending_req *pending_req)
939 {
940 	int rc;
941 
942 	rc = xen_blkbk_map(pending_req->ring, pending_req->segments,
943 			   pending_req->nr_segs,
944 	                   (pending_req->operation != BLKIF_OP_READ));
945 
946 	return rc;
947 }
948 
949 static int xen_blkbk_parse_indirect(struct blkif_request *req,
950 				    struct pending_req *pending_req,
951 				    struct seg_buf seg[],
952 				    struct phys_req *preq)
953 {
954 	struct grant_page **pages = pending_req->indirect_pages;
955 	struct xen_blkif_ring *ring = pending_req->ring;
956 	int indirect_grefs, rc, n, nseg, i;
957 	struct blkif_request_segment *segments = NULL;
958 
959 	nseg = pending_req->nr_segs;
960 	indirect_grefs = INDIRECT_PAGES(nseg);
961 	BUG_ON(indirect_grefs > BLKIF_MAX_INDIRECT_PAGES_PER_REQUEST);
962 
963 	for (i = 0; i < indirect_grefs; i++)
964 		pages[i]->gref = req->u.indirect.indirect_grefs[i];
965 
966 	rc = xen_blkbk_map(ring, pages, indirect_grefs, true);
967 	if (rc)
968 		goto unmap;
969 
970 	for (n = 0, i = 0; n < nseg; n++) {
971 		uint8_t first_sect, last_sect;
972 
973 		if ((n % SEGS_PER_INDIRECT_FRAME) == 0) {
974 			/* Map indirect segments */
975 			if (segments)
976 				kunmap_atomic(segments);
977 			segments = kmap_atomic(pages[n/SEGS_PER_INDIRECT_FRAME]->page);
978 		}
979 		i = n % SEGS_PER_INDIRECT_FRAME;
980 
981 		pending_req->segments[n]->gref = segments[i].gref;
982 
983 		first_sect = READ_ONCE(segments[i].first_sect);
984 		last_sect = READ_ONCE(segments[i].last_sect);
985 		if (last_sect >= (XEN_PAGE_SIZE >> 9) || last_sect < first_sect) {
986 			rc = -EINVAL;
987 			goto unmap;
988 		}
989 
990 		seg[n].nsec = last_sect - first_sect + 1;
991 		seg[n].offset = first_sect << 9;
992 		preq->nr_sects += seg[n].nsec;
993 	}
994 
995 unmap:
996 	if (segments)
997 		kunmap_atomic(segments);
998 	xen_blkbk_unmap(ring, pages, indirect_grefs);
999 	return rc;
1000 }
1001 
1002 static int dispatch_discard_io(struct xen_blkif_ring *ring,
1003 				struct blkif_request *req)
1004 {
1005 	int err = 0;
1006 	int status = BLKIF_RSP_OKAY;
1007 	struct xen_blkif *blkif = ring->blkif;
1008 	struct block_device *bdev = blkif->vbd.bdev;
1009 	unsigned long secure;
1010 	struct phys_req preq;
1011 
1012 	xen_blkif_get(blkif);
1013 
1014 	preq.sector_number = req->u.discard.sector_number;
1015 	preq.nr_sects      = req->u.discard.nr_sectors;
1016 
1017 	err = xen_vbd_translate(&preq, blkif, WRITE);
1018 	if (err) {
1019 		pr_warn("access denied: DISCARD [%llu->%llu] on dev=%04x\n",
1020 			preq.sector_number,
1021 			preq.sector_number + preq.nr_sects, blkif->vbd.pdevice);
1022 		goto fail_response;
1023 	}
1024 	ring->st_ds_req++;
1025 
1026 	secure = (blkif->vbd.discard_secure &&
1027 		 (req->u.discard.flag & BLKIF_DISCARD_SECURE)) ?
1028 		 BLKDEV_DISCARD_SECURE : 0;
1029 
1030 	err = blkdev_issue_discard(bdev, req->u.discard.sector_number,
1031 				   req->u.discard.nr_sectors,
1032 				   GFP_KERNEL, secure);
1033 fail_response:
1034 	if (err == -EOPNOTSUPP) {
1035 		pr_debug("discard op failed, not supported\n");
1036 		status = BLKIF_RSP_EOPNOTSUPP;
1037 	} else if (err)
1038 		status = BLKIF_RSP_ERROR;
1039 
1040 	make_response(ring, req->u.discard.id, req->operation, status);
1041 	xen_blkif_put(blkif);
1042 	return err;
1043 }
1044 
1045 static int dispatch_other_io(struct xen_blkif_ring *ring,
1046 			     struct blkif_request *req,
1047 			     struct pending_req *pending_req)
1048 {
1049 	free_req(ring, pending_req);
1050 	make_response(ring, req->u.other.id, req->operation,
1051 		      BLKIF_RSP_EOPNOTSUPP);
1052 	return -EIO;
1053 }
1054 
1055 static void xen_blk_drain_io(struct xen_blkif_ring *ring)
1056 {
1057 	struct xen_blkif *blkif = ring->blkif;
1058 
1059 	atomic_set(&blkif->drain, 1);
1060 	do {
1061 		if (atomic_read(&ring->inflight) == 0)
1062 			break;
1063 		wait_for_completion_interruptible_timeout(
1064 				&blkif->drain_complete, HZ);
1065 
1066 		if (!atomic_read(&blkif->drain))
1067 			break;
1068 	} while (!kthread_should_stop());
1069 	atomic_set(&blkif->drain, 0);
1070 }
1071 
1072 /*
1073  * Completion callback on the bio's. Called as bh->b_end_io()
1074  */
1075 
1076 static void __end_block_io_op(struct pending_req *pending_req, int error)
1077 {
1078 	/* An error fails the entire request. */
1079 	if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
1080 	    (error == -EOPNOTSUPP)) {
1081 		pr_debug("flush diskcache op failed, not supported\n");
1082 		xen_blkbk_flush_diskcache(XBT_NIL, pending_req->ring->blkif->be, 0);
1083 		pending_req->status = BLKIF_RSP_EOPNOTSUPP;
1084 	} else if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
1085 		    (error == -EOPNOTSUPP)) {
1086 		pr_debug("write barrier op failed, not supported\n");
1087 		xen_blkbk_barrier(XBT_NIL, pending_req->ring->blkif->be, 0);
1088 		pending_req->status = BLKIF_RSP_EOPNOTSUPP;
1089 	} else if (error) {
1090 		pr_debug("Buffer not up-to-date at end of operation,"
1091 			 " error=%d\n", error);
1092 		pending_req->status = BLKIF_RSP_ERROR;
1093 	}
1094 
1095 	/*
1096 	 * If all of the bio's have completed it is time to unmap
1097 	 * the grant references associated with 'request' and provide
1098 	 * the proper response on the ring.
1099 	 */
1100 	if (atomic_dec_and_test(&pending_req->pendcnt))
1101 		xen_blkbk_unmap_and_respond(pending_req);
1102 }
1103 
1104 /*
1105  * bio callback.
1106  */
1107 static void end_block_io_op(struct bio *bio)
1108 {
1109 	__end_block_io_op(bio->bi_private, bio->bi_error);
1110 	bio_put(bio);
1111 }
1112 
1113 
1114 
1115 /*
1116  * Function to copy the from the ring buffer the 'struct blkif_request'
1117  * (which has the sectors we want, number of them, grant references, etc),
1118  * and transmute  it to the block API to hand it over to the proper block disk.
1119  */
1120 static int
1121 __do_block_io_op(struct xen_blkif_ring *ring)
1122 {
1123 	union blkif_back_rings *blk_rings = &ring->blk_rings;
1124 	struct blkif_request req;
1125 	struct pending_req *pending_req;
1126 	RING_IDX rc, rp;
1127 	int more_to_do = 0;
1128 
1129 	rc = blk_rings->common.req_cons;
1130 	rp = blk_rings->common.sring->req_prod;
1131 	rmb(); /* Ensure we see queued requests up to 'rp'. */
1132 
1133 	if (RING_REQUEST_PROD_OVERFLOW(&blk_rings->common, rp)) {
1134 		rc = blk_rings->common.rsp_prod_pvt;
1135 		pr_warn("Frontend provided bogus ring requests (%d - %d = %d). Halting ring processing on dev=%04x\n",
1136 			rp, rc, rp - rc, ring->blkif->vbd.pdevice);
1137 		return -EACCES;
1138 	}
1139 	while (rc != rp) {
1140 
1141 		if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
1142 			break;
1143 
1144 		if (kthread_should_stop()) {
1145 			more_to_do = 1;
1146 			break;
1147 		}
1148 
1149 		pending_req = alloc_req(ring);
1150 		if (NULL == pending_req) {
1151 			ring->st_oo_req++;
1152 			more_to_do = 1;
1153 			break;
1154 		}
1155 
1156 		switch (ring->blkif->blk_protocol) {
1157 		case BLKIF_PROTOCOL_NATIVE:
1158 			memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
1159 			break;
1160 		case BLKIF_PROTOCOL_X86_32:
1161 			blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
1162 			break;
1163 		case BLKIF_PROTOCOL_X86_64:
1164 			blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
1165 			break;
1166 		default:
1167 			BUG();
1168 		}
1169 		blk_rings->common.req_cons = ++rc; /* before make_response() */
1170 
1171 		/* Apply all sanity checks to /private copy/ of request. */
1172 		barrier();
1173 
1174 		switch (req.operation) {
1175 		case BLKIF_OP_READ:
1176 		case BLKIF_OP_WRITE:
1177 		case BLKIF_OP_WRITE_BARRIER:
1178 		case BLKIF_OP_FLUSH_DISKCACHE:
1179 		case BLKIF_OP_INDIRECT:
1180 			if (dispatch_rw_block_io(ring, &req, pending_req))
1181 				goto done;
1182 			break;
1183 		case BLKIF_OP_DISCARD:
1184 			free_req(ring, pending_req);
1185 			if (dispatch_discard_io(ring, &req))
1186 				goto done;
1187 			break;
1188 		default:
1189 			if (dispatch_other_io(ring, &req, pending_req))
1190 				goto done;
1191 			break;
1192 		}
1193 
1194 		/* Yield point for this unbounded loop. */
1195 		cond_resched();
1196 	}
1197 done:
1198 	return more_to_do;
1199 }
1200 
1201 static int
1202 do_block_io_op(struct xen_blkif_ring *ring)
1203 {
1204 	union blkif_back_rings *blk_rings = &ring->blk_rings;
1205 	int more_to_do;
1206 
1207 	do {
1208 		more_to_do = __do_block_io_op(ring);
1209 		if (more_to_do)
1210 			break;
1211 
1212 		RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
1213 	} while (more_to_do);
1214 
1215 	return more_to_do;
1216 }
1217 /*
1218  * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
1219  * and call the 'submit_bio' to pass it to the underlying storage.
1220  */
1221 static int dispatch_rw_block_io(struct xen_blkif_ring *ring,
1222 				struct blkif_request *req,
1223 				struct pending_req *pending_req)
1224 {
1225 	struct phys_req preq;
1226 	struct seg_buf *seg = pending_req->seg;
1227 	unsigned int nseg;
1228 	struct bio *bio = NULL;
1229 	struct bio **biolist = pending_req->biolist;
1230 	int i, nbio = 0;
1231 	int operation;
1232 	struct blk_plug plug;
1233 	bool drain = false;
1234 	struct grant_page **pages = pending_req->segments;
1235 	unsigned short req_operation;
1236 
1237 	req_operation = req->operation == BLKIF_OP_INDIRECT ?
1238 			req->u.indirect.indirect_op : req->operation;
1239 
1240 	if ((req->operation == BLKIF_OP_INDIRECT) &&
1241 	    (req_operation != BLKIF_OP_READ) &&
1242 	    (req_operation != BLKIF_OP_WRITE)) {
1243 		pr_debug("Invalid indirect operation (%u)\n", req_operation);
1244 		goto fail_response;
1245 	}
1246 
1247 	switch (req_operation) {
1248 	case BLKIF_OP_READ:
1249 		ring->st_rd_req++;
1250 		operation = READ;
1251 		break;
1252 	case BLKIF_OP_WRITE:
1253 		ring->st_wr_req++;
1254 		operation = WRITE_ODIRECT;
1255 		break;
1256 	case BLKIF_OP_WRITE_BARRIER:
1257 		drain = true;
1258 	case BLKIF_OP_FLUSH_DISKCACHE:
1259 		ring->st_f_req++;
1260 		operation = WRITE_FLUSH;
1261 		break;
1262 	default:
1263 		operation = 0; /* make gcc happy */
1264 		goto fail_response;
1265 		break;
1266 	}
1267 
1268 	/* Check that the number of segments is sane. */
1269 	nseg = req->operation == BLKIF_OP_INDIRECT ?
1270 	       req->u.indirect.nr_segments : req->u.rw.nr_segments;
1271 
1272 	if (unlikely(nseg == 0 && operation != WRITE_FLUSH) ||
1273 	    unlikely((req->operation != BLKIF_OP_INDIRECT) &&
1274 		     (nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) ||
1275 	    unlikely((req->operation == BLKIF_OP_INDIRECT) &&
1276 		     (nseg > MAX_INDIRECT_SEGMENTS))) {
1277 		pr_debug("Bad number of segments in request (%d)\n", nseg);
1278 		/* Haven't submitted any bio's yet. */
1279 		goto fail_response;
1280 	}
1281 
1282 	preq.nr_sects      = 0;
1283 
1284 	pending_req->ring      = ring;
1285 	pending_req->id        = req->u.rw.id;
1286 	pending_req->operation = req_operation;
1287 	pending_req->status    = BLKIF_RSP_OKAY;
1288 	pending_req->nr_segs   = nseg;
1289 
1290 	if (req->operation != BLKIF_OP_INDIRECT) {
1291 		preq.dev               = req->u.rw.handle;
1292 		preq.sector_number     = req->u.rw.sector_number;
1293 		for (i = 0; i < nseg; i++) {
1294 			pages[i]->gref = req->u.rw.seg[i].gref;
1295 			seg[i].nsec = req->u.rw.seg[i].last_sect -
1296 				req->u.rw.seg[i].first_sect + 1;
1297 			seg[i].offset = (req->u.rw.seg[i].first_sect << 9);
1298 			if ((req->u.rw.seg[i].last_sect >= (XEN_PAGE_SIZE >> 9)) ||
1299 			    (req->u.rw.seg[i].last_sect <
1300 			     req->u.rw.seg[i].first_sect))
1301 				goto fail_response;
1302 			preq.nr_sects += seg[i].nsec;
1303 		}
1304 	} else {
1305 		preq.dev               = req->u.indirect.handle;
1306 		preq.sector_number     = req->u.indirect.sector_number;
1307 		if (xen_blkbk_parse_indirect(req, pending_req, seg, &preq))
1308 			goto fail_response;
1309 	}
1310 
1311 	if (xen_vbd_translate(&preq, ring->blkif, operation) != 0) {
1312 		pr_debug("access denied: %s of [%llu,%llu] on dev=%04x\n",
1313 			 operation == READ ? "read" : "write",
1314 			 preq.sector_number,
1315 			 preq.sector_number + preq.nr_sects,
1316 			 ring->blkif->vbd.pdevice);
1317 		goto fail_response;
1318 	}
1319 
1320 	/*
1321 	 * This check _MUST_ be done after xen_vbd_translate as the preq.bdev
1322 	 * is set there.
1323 	 */
1324 	for (i = 0; i < nseg; i++) {
1325 		if (((int)preq.sector_number|(int)seg[i].nsec) &
1326 		    ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
1327 			pr_debug("Misaligned I/O request from domain %d\n",
1328 				 ring->blkif->domid);
1329 			goto fail_response;
1330 		}
1331 	}
1332 
1333 	/* Wait on all outstanding I/O's and once that has been completed
1334 	 * issue the WRITE_FLUSH.
1335 	 */
1336 	if (drain)
1337 		xen_blk_drain_io(pending_req->ring);
1338 
1339 	/*
1340 	 * If we have failed at this point, we need to undo the M2P override,
1341 	 * set gnttab_set_unmap_op on all of the grant references and perform
1342 	 * the hypercall to unmap the grants - that is all done in
1343 	 * xen_blkbk_unmap.
1344 	 */
1345 	if (xen_blkbk_map_seg(pending_req))
1346 		goto fail_flush;
1347 
1348 	/*
1349 	 * This corresponding xen_blkif_put is done in __end_block_io_op, or
1350 	 * below (in "!bio") if we are handling a BLKIF_OP_DISCARD.
1351 	 */
1352 	xen_blkif_get(ring->blkif);
1353 	atomic_inc(&ring->inflight);
1354 
1355 	for (i = 0; i < nseg; i++) {
1356 		while ((bio == NULL) ||
1357 		       (bio_add_page(bio,
1358 				     pages[i]->page,
1359 				     seg[i].nsec << 9,
1360 				     seg[i].offset) == 0)) {
1361 
1362 			int nr_iovecs = min_t(int, (nseg-i), BIO_MAX_PAGES);
1363 			bio = bio_alloc(GFP_KERNEL, nr_iovecs);
1364 			if (unlikely(bio == NULL))
1365 				goto fail_put_bio;
1366 
1367 			biolist[nbio++] = bio;
1368 			bio->bi_bdev    = preq.bdev;
1369 			bio->bi_private = pending_req;
1370 			bio->bi_end_io  = end_block_io_op;
1371 			bio->bi_iter.bi_sector  = preq.sector_number;
1372 		}
1373 
1374 		preq.sector_number += seg[i].nsec;
1375 	}
1376 
1377 	/* This will be hit if the operation was a flush or discard. */
1378 	if (!bio) {
1379 		BUG_ON(operation != WRITE_FLUSH);
1380 
1381 		bio = bio_alloc(GFP_KERNEL, 0);
1382 		if (unlikely(bio == NULL))
1383 			goto fail_put_bio;
1384 
1385 		biolist[nbio++] = bio;
1386 		bio->bi_bdev    = preq.bdev;
1387 		bio->bi_private = pending_req;
1388 		bio->bi_end_io  = end_block_io_op;
1389 	}
1390 
1391 	atomic_set(&pending_req->pendcnt, nbio);
1392 	blk_start_plug(&plug);
1393 
1394 	for (i = 0; i < nbio; i++)
1395 		submit_bio(operation, biolist[i]);
1396 
1397 	/* Let the I/Os go.. */
1398 	blk_finish_plug(&plug);
1399 
1400 	if (operation == READ)
1401 		ring->st_rd_sect += preq.nr_sects;
1402 	else if (operation & WRITE)
1403 		ring->st_wr_sect += preq.nr_sects;
1404 
1405 	return 0;
1406 
1407  fail_flush:
1408 	xen_blkbk_unmap(ring, pending_req->segments,
1409 	                pending_req->nr_segs);
1410  fail_response:
1411 	/* Haven't submitted any bio's yet. */
1412 	make_response(ring, req->u.rw.id, req_operation, BLKIF_RSP_ERROR);
1413 	free_req(ring, pending_req);
1414 	msleep(1); /* back off a bit */
1415 	return -EIO;
1416 
1417  fail_put_bio:
1418 	for (i = 0; i < nbio; i++)
1419 		bio_put(biolist[i]);
1420 	atomic_set(&pending_req->pendcnt, 1);
1421 	__end_block_io_op(pending_req, -EINVAL);
1422 	msleep(1); /* back off a bit */
1423 	return -EIO;
1424 }
1425 
1426 
1427 
1428 /*
1429  * Put a response on the ring on how the operation fared.
1430  */
1431 static void make_response(struct xen_blkif_ring *ring, u64 id,
1432 			  unsigned short op, int st)
1433 {
1434 	struct blkif_response  resp;
1435 	unsigned long     flags;
1436 	union blkif_back_rings *blk_rings;
1437 	int notify;
1438 
1439 	resp.id        = id;
1440 	resp.operation = op;
1441 	resp.status    = st;
1442 
1443 	spin_lock_irqsave(&ring->blk_ring_lock, flags);
1444 	blk_rings = &ring->blk_rings;
1445 	/* Place on the response ring for the relevant domain. */
1446 	switch (ring->blkif->blk_protocol) {
1447 	case BLKIF_PROTOCOL_NATIVE:
1448 		memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
1449 		       &resp, sizeof(resp));
1450 		break;
1451 	case BLKIF_PROTOCOL_X86_32:
1452 		memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
1453 		       &resp, sizeof(resp));
1454 		break;
1455 	case BLKIF_PROTOCOL_X86_64:
1456 		memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
1457 		       &resp, sizeof(resp));
1458 		break;
1459 	default:
1460 		BUG();
1461 	}
1462 	blk_rings->common.rsp_prod_pvt++;
1463 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
1464 	spin_unlock_irqrestore(&ring->blk_ring_lock, flags);
1465 	if (notify)
1466 		notify_remote_via_irq(ring->irq);
1467 }
1468 
1469 static int __init xen_blkif_init(void)
1470 {
1471 	int rc = 0;
1472 
1473 	if (!xen_domain())
1474 		return -ENODEV;
1475 
1476 	if (xen_blkif_max_ring_order > XENBUS_MAX_RING_GRANT_ORDER) {
1477 		pr_info("Invalid max_ring_order (%d), will use default max: %d.\n",
1478 			xen_blkif_max_ring_order, XENBUS_MAX_RING_GRANT_ORDER);
1479 		xen_blkif_max_ring_order = XENBUS_MAX_RING_GRANT_ORDER;
1480 	}
1481 
1482 	if (xenblk_max_queues == 0)
1483 		xenblk_max_queues = num_online_cpus();
1484 
1485 	rc = xen_blkif_interface_init();
1486 	if (rc)
1487 		goto failed_init;
1488 
1489 	rc = xen_blkif_xenbus_init();
1490 	if (rc)
1491 		goto failed_init;
1492 
1493  failed_init:
1494 	return rc;
1495 }
1496 
1497 module_init(xen_blkif_init);
1498 
1499 MODULE_LICENSE("Dual BSD/GPL");
1500 MODULE_ALIAS("xen-backend:vbd");
1501