xref: /linux/drivers/block/xen-blkback/blkback.c (revision bef9ae3d883ce908d8879fff0cd2c3971f5ee4b4)
1 /******************************************************************************
2  *
3  * Back-end of the driver for virtual block devices. This portion of the
4  * driver exports a 'unified' block-device interface that can be accessed
5  * by any operating system that implements a compatible front end. A
6  * reference front-end implementation can be found in:
7  *  drivers/block/xen-blkfront.c
8  *
9  * Copyright (c) 2003-2004, Keir Fraser & Steve Hand
10  * Copyright (c) 2005, Christopher Clark
11  *
12  * This program is free software; you can redistribute it and/or
13  * modify it under the terms of the GNU General Public License version 2
14  * as published by the Free Software Foundation; or, when distributed
15  * separately from the Linux kernel or incorporated into other
16  * software packages, subject to the following license:
17  *
18  * Permission is hereby granted, free of charge, to any person obtaining a copy
19  * of this source file (the "Software"), to deal in the Software without
20  * restriction, including without limitation the rights to use, copy, modify,
21  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
22  * and to permit persons to whom the Software is furnished to do so, subject to
23  * the following conditions:
24  *
25  * The above copyright notice and this permission notice shall be included in
26  * all copies or substantial portions of the Software.
27  *
28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
29  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
30  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
31  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
32  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
33  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
34  * IN THE SOFTWARE.
35  */
36 
37 #include <linux/spinlock.h>
38 #include <linux/kthread.h>
39 #include <linux/list.h>
40 #include <linux/delay.h>
41 #include <linux/freezer.h>
42 #include <linux/bitmap.h>
43 
44 #include <xen/events.h>
45 #include <xen/page.h>
46 #include <xen/xen.h>
47 #include <asm/xen/hypervisor.h>
48 #include <asm/xen/hypercall.h>
49 #include "common.h"
50 
51 /*
52  * These are rather arbitrary. They are fairly large because adjacent requests
53  * pulled from a communication ring are quite likely to end up being part of
54  * the same scatter/gather request at the disc.
55  *
56  * ** TRY INCREASING 'xen_blkif_reqs' IF WRITE SPEEDS SEEM TOO LOW **
57  *
58  * This will increase the chances of being able to write whole tracks.
59  * 64 should be enough to keep us competitive with Linux.
60  */
61 static int xen_blkif_reqs = 64;
62 module_param_named(reqs, xen_blkif_reqs, int, 0);
63 MODULE_PARM_DESC(reqs, "Number of blkback requests to allocate");
64 
65 /* Run-time switchable: /sys/module/blkback/parameters/ */
66 static unsigned int log_stats;
67 module_param(log_stats, int, 0644);
68 
69 /*
70  * Each outstanding request that we've passed to the lower device layers has a
71  * 'pending_req' allocated to it. Each buffer_head that completes decrements
72  * the pendcnt towards zero. When it hits zero, the specified domain has a
73  * response queued for it, with the saved 'id' passed back.
74  */
75 struct pending_req {
76 	struct xen_blkif	*blkif;
77 	u64			id;
78 	int			nr_pages;
79 	atomic_t		pendcnt;
80 	unsigned short		operation;
81 	int			status;
82 	struct list_head	free_list;
83 	DECLARE_BITMAP(unmap_seg, BLKIF_MAX_SEGMENTS_PER_REQUEST);
84 };
85 
86 #define BLKBACK_INVALID_HANDLE (~0)
87 
88 struct xen_blkbk {
89 	struct pending_req	*pending_reqs;
90 	/* List of all 'pending_req' available */
91 	struct list_head	pending_free;
92 	/* And its spinlock. */
93 	spinlock_t		pending_free_lock;
94 	wait_queue_head_t	pending_free_wq;
95 	/* The list of all pages that are available. */
96 	struct page		**pending_pages;
97 	/* And the grant handles that are available. */
98 	grant_handle_t		*pending_grant_handles;
99 };
100 
101 static struct xen_blkbk *blkbk;
102 
103 /*
104  * Maximum number of grant pages that can be mapped in blkback.
105  * BLKIF_MAX_SEGMENTS_PER_REQUEST * RING_SIZE is the maximum number of
106  * pages that blkback will persistently map.
107  * Currently, this is:
108  * RING_SIZE = 32 (for all known ring types)
109  * BLKIF_MAX_SEGMENTS_PER_REQUEST = 11
110  * sizeof(struct persistent_gnt) = 48
111  * So the maximum memory used to store the grants is:
112  * 32 * 11 * 48 = 16896 bytes
113  */
114 static inline unsigned int max_mapped_grant_pages(enum blkif_protocol protocol)
115 {
116 	switch (protocol) {
117 	case BLKIF_PROTOCOL_NATIVE:
118 		return __CONST_RING_SIZE(blkif, PAGE_SIZE) *
119 			   BLKIF_MAX_SEGMENTS_PER_REQUEST;
120 	case BLKIF_PROTOCOL_X86_32:
121 		return __CONST_RING_SIZE(blkif_x86_32, PAGE_SIZE) *
122 			   BLKIF_MAX_SEGMENTS_PER_REQUEST;
123 	case BLKIF_PROTOCOL_X86_64:
124 		return __CONST_RING_SIZE(blkif_x86_64, PAGE_SIZE) *
125 			   BLKIF_MAX_SEGMENTS_PER_REQUEST;
126 	default:
127 		BUG();
128 	}
129 	return 0;
130 }
131 
132 
133 /*
134  * Little helpful macro to figure out the index and virtual address of the
135  * pending_pages[..]. For each 'pending_req' we have have up to
136  * BLKIF_MAX_SEGMENTS_PER_REQUEST (11) pages. The seg would be from 0 through
137  * 10 and would index in the pending_pages[..].
138  */
139 static inline int vaddr_pagenr(struct pending_req *req, int seg)
140 {
141 	return (req - blkbk->pending_reqs) *
142 		BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
143 }
144 
145 #define pending_page(req, seg) pending_pages[vaddr_pagenr(req, seg)]
146 
147 static inline unsigned long vaddr(struct pending_req *req, int seg)
148 {
149 	unsigned long pfn = page_to_pfn(blkbk->pending_page(req, seg));
150 	return (unsigned long)pfn_to_kaddr(pfn);
151 }
152 
153 #define pending_handle(_req, _seg) \
154 	(blkbk->pending_grant_handles[vaddr_pagenr(_req, _seg)])
155 
156 
157 static int do_block_io_op(struct xen_blkif *blkif);
158 static int dispatch_rw_block_io(struct xen_blkif *blkif,
159 				struct blkif_request *req,
160 				struct pending_req *pending_req);
161 static void make_response(struct xen_blkif *blkif, u64 id,
162 			  unsigned short op, int st);
163 
164 #define foreach_grant(pos, rbtree, node) \
165 	for ((pos) = container_of(rb_first((rbtree)), typeof(*(pos)), node); \
166 	     &(pos)->node != NULL; \
167 	     (pos) = container_of(rb_next(&(pos)->node), typeof(*(pos)), node))
168 
169 
170 static void add_persistent_gnt(struct rb_root *root,
171 			       struct persistent_gnt *persistent_gnt)
172 {
173 	struct rb_node **new = &(root->rb_node), *parent = NULL;
174 	struct persistent_gnt *this;
175 
176 	/* Figure out where to put new node */
177 	while (*new) {
178 		this = container_of(*new, struct persistent_gnt, node);
179 
180 		parent = *new;
181 		if (persistent_gnt->gnt < this->gnt)
182 			new = &((*new)->rb_left);
183 		else if (persistent_gnt->gnt > this->gnt)
184 			new = &((*new)->rb_right);
185 		else {
186 			pr_alert(DRV_PFX " trying to add a gref that's already in the tree\n");
187 			BUG();
188 		}
189 	}
190 
191 	/* Add new node and rebalance tree. */
192 	rb_link_node(&(persistent_gnt->node), parent, new);
193 	rb_insert_color(&(persistent_gnt->node), root);
194 }
195 
196 static struct persistent_gnt *get_persistent_gnt(struct rb_root *root,
197 						 grant_ref_t gref)
198 {
199 	struct persistent_gnt *data;
200 	struct rb_node *node = root->rb_node;
201 
202 	while (node) {
203 		data = container_of(node, struct persistent_gnt, node);
204 
205 		if (gref < data->gnt)
206 			node = node->rb_left;
207 		else if (gref > data->gnt)
208 			node = node->rb_right;
209 		else
210 			return data;
211 	}
212 	return NULL;
213 }
214 
215 static void free_persistent_gnts(struct rb_root *root, unsigned int num)
216 {
217 	struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
218 	struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
219 	struct persistent_gnt *persistent_gnt;
220 	int ret = 0;
221 	int segs_to_unmap = 0;
222 
223 	foreach_grant(persistent_gnt, root, node) {
224 		BUG_ON(persistent_gnt->handle ==
225 			BLKBACK_INVALID_HANDLE);
226 		gnttab_set_unmap_op(&unmap[segs_to_unmap],
227 			(unsigned long) pfn_to_kaddr(page_to_pfn(
228 				persistent_gnt->page)),
229 			GNTMAP_host_map,
230 			persistent_gnt->handle);
231 
232 		pages[segs_to_unmap] = persistent_gnt->page;
233 		rb_erase(&persistent_gnt->node, root);
234 		kfree(persistent_gnt);
235 		num--;
236 
237 		if (++segs_to_unmap == BLKIF_MAX_SEGMENTS_PER_REQUEST ||
238 			!rb_next(&persistent_gnt->node)) {
239 			ret = gnttab_unmap_refs(unmap, NULL, pages,
240 				segs_to_unmap);
241 			BUG_ON(ret);
242 			segs_to_unmap = 0;
243 		}
244 	}
245 	BUG_ON(num != 0);
246 }
247 
248 /*
249  * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
250  */
251 static struct pending_req *alloc_req(void)
252 {
253 	struct pending_req *req = NULL;
254 	unsigned long flags;
255 
256 	spin_lock_irqsave(&blkbk->pending_free_lock, flags);
257 	if (!list_empty(&blkbk->pending_free)) {
258 		req = list_entry(blkbk->pending_free.next, struct pending_req,
259 				 free_list);
260 		list_del(&req->free_list);
261 	}
262 	spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
263 	return req;
264 }
265 
266 /*
267  * Return the 'pending_req' structure back to the freepool. We also
268  * wake up the thread if it was waiting for a free page.
269  */
270 static void free_req(struct pending_req *req)
271 {
272 	unsigned long flags;
273 	int was_empty;
274 
275 	spin_lock_irqsave(&blkbk->pending_free_lock, flags);
276 	was_empty = list_empty(&blkbk->pending_free);
277 	list_add(&req->free_list, &blkbk->pending_free);
278 	spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
279 	if (was_empty)
280 		wake_up(&blkbk->pending_free_wq);
281 }
282 
283 /*
284  * Routines for managing virtual block devices (vbds).
285  */
286 static int xen_vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
287 			     int operation)
288 {
289 	struct xen_vbd *vbd = &blkif->vbd;
290 	int rc = -EACCES;
291 
292 	if ((operation != READ) && vbd->readonly)
293 		goto out;
294 
295 	if (likely(req->nr_sects)) {
296 		blkif_sector_t end = req->sector_number + req->nr_sects;
297 
298 		if (unlikely(end < req->sector_number))
299 			goto out;
300 		if (unlikely(end > vbd_sz(vbd)))
301 			goto out;
302 	}
303 
304 	req->dev  = vbd->pdevice;
305 	req->bdev = vbd->bdev;
306 	rc = 0;
307 
308  out:
309 	return rc;
310 }
311 
312 static void xen_vbd_resize(struct xen_blkif *blkif)
313 {
314 	struct xen_vbd *vbd = &blkif->vbd;
315 	struct xenbus_transaction xbt;
316 	int err;
317 	struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
318 	unsigned long long new_size = vbd_sz(vbd);
319 
320 	pr_info(DRV_PFX "VBD Resize: Domid: %d, Device: (%d, %d)\n",
321 		blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
322 	pr_info(DRV_PFX "VBD Resize: new size %llu\n", new_size);
323 	vbd->size = new_size;
324 again:
325 	err = xenbus_transaction_start(&xbt);
326 	if (err) {
327 		pr_warn(DRV_PFX "Error starting transaction");
328 		return;
329 	}
330 	err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
331 			    (unsigned long long)vbd_sz(vbd));
332 	if (err) {
333 		pr_warn(DRV_PFX "Error writing new size");
334 		goto abort;
335 	}
336 	/*
337 	 * Write the current state; we will use this to synchronize
338 	 * the front-end. If the current state is "connected" the
339 	 * front-end will get the new size information online.
340 	 */
341 	err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
342 	if (err) {
343 		pr_warn(DRV_PFX "Error writing the state");
344 		goto abort;
345 	}
346 
347 	err = xenbus_transaction_end(xbt, 0);
348 	if (err == -EAGAIN)
349 		goto again;
350 	if (err)
351 		pr_warn(DRV_PFX "Error ending transaction");
352 	return;
353 abort:
354 	xenbus_transaction_end(xbt, 1);
355 }
356 
357 /*
358  * Notification from the guest OS.
359  */
360 static void blkif_notify_work(struct xen_blkif *blkif)
361 {
362 	blkif->waiting_reqs = 1;
363 	wake_up(&blkif->wq);
364 }
365 
366 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
367 {
368 	blkif_notify_work(dev_id);
369 	return IRQ_HANDLED;
370 }
371 
372 /*
373  * SCHEDULER FUNCTIONS
374  */
375 
376 static void print_stats(struct xen_blkif *blkif)
377 {
378 	pr_info("xen-blkback (%s): oo %3d  |  rd %4d  |  wr %4d  |  f %4d"
379 		 "  |  ds %4d\n",
380 		 current->comm, blkif->st_oo_req,
381 		 blkif->st_rd_req, blkif->st_wr_req,
382 		 blkif->st_f_req, blkif->st_ds_req);
383 	blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
384 	blkif->st_rd_req = 0;
385 	blkif->st_wr_req = 0;
386 	blkif->st_oo_req = 0;
387 	blkif->st_ds_req = 0;
388 }
389 
390 int xen_blkif_schedule(void *arg)
391 {
392 	struct xen_blkif *blkif = arg;
393 	struct xen_vbd *vbd = &blkif->vbd;
394 
395 	xen_blkif_get(blkif);
396 
397 	while (!kthread_should_stop()) {
398 		if (try_to_freeze())
399 			continue;
400 		if (unlikely(vbd->size != vbd_sz(vbd)))
401 			xen_vbd_resize(blkif);
402 
403 		wait_event_interruptible(
404 			blkif->wq,
405 			blkif->waiting_reqs || kthread_should_stop());
406 		wait_event_interruptible(
407 			blkbk->pending_free_wq,
408 			!list_empty(&blkbk->pending_free) ||
409 			kthread_should_stop());
410 
411 		blkif->waiting_reqs = 0;
412 		smp_mb(); /* clear flag *before* checking for work */
413 
414 		if (do_block_io_op(blkif))
415 			blkif->waiting_reqs = 1;
416 
417 		if (log_stats && time_after(jiffies, blkif->st_print))
418 			print_stats(blkif);
419 	}
420 
421 	/* Free all persistent grant pages */
422 	if (!RB_EMPTY_ROOT(&blkif->persistent_gnts))
423 		free_persistent_gnts(&blkif->persistent_gnts,
424 			blkif->persistent_gnt_c);
425 
426 	BUG_ON(!RB_EMPTY_ROOT(&blkif->persistent_gnts));
427 	blkif->persistent_gnt_c = 0;
428 
429 	if (log_stats)
430 		print_stats(blkif);
431 
432 	blkif->xenblkd = NULL;
433 	xen_blkif_put(blkif);
434 
435 	return 0;
436 }
437 
438 struct seg_buf {
439 	unsigned long buf;
440 	unsigned int nsec;
441 };
442 /*
443  * Unmap the grant references, and also remove the M2P over-rides
444  * used in the 'pending_req'.
445  */
446 static void xen_blkbk_unmap(struct pending_req *req)
447 {
448 	struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
449 	struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
450 	unsigned int i, invcount = 0;
451 	grant_handle_t handle;
452 	int ret;
453 
454 	for (i = 0; i < req->nr_pages; i++) {
455 		if (!test_bit(i, req->unmap_seg))
456 			continue;
457 		handle = pending_handle(req, i);
458 		if (handle == BLKBACK_INVALID_HANDLE)
459 			continue;
460 		gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
461 				    GNTMAP_host_map, handle);
462 		pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
463 		pages[invcount] = virt_to_page(vaddr(req, i));
464 		invcount++;
465 	}
466 
467 	ret = gnttab_unmap_refs(unmap, NULL, pages, invcount);
468 	BUG_ON(ret);
469 }
470 
471 static int xen_blkbk_map(struct blkif_request *req,
472 			 struct pending_req *pending_req,
473 			 struct seg_buf seg[],
474 			 struct page *pages[])
475 {
476 	struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
477 	struct persistent_gnt *persistent_gnts[BLKIF_MAX_SEGMENTS_PER_REQUEST];
478 	struct page *pages_to_gnt[BLKIF_MAX_SEGMENTS_PER_REQUEST];
479 	struct persistent_gnt *persistent_gnt = NULL;
480 	struct xen_blkif *blkif = pending_req->blkif;
481 	phys_addr_t addr = 0;
482 	int i, j;
483 	bool new_map;
484 	int nseg = req->u.rw.nr_segments;
485 	int segs_to_map = 0;
486 	int ret = 0;
487 	int use_persistent_gnts;
488 
489 	use_persistent_gnts = (blkif->vbd.feature_gnt_persistent);
490 
491 	BUG_ON(blkif->persistent_gnt_c >
492 		   max_mapped_grant_pages(pending_req->blkif->blk_protocol));
493 
494 	/*
495 	 * Fill out preq.nr_sects with proper amount of sectors, and setup
496 	 * assign map[..] with the PFN of the page in our domain with the
497 	 * corresponding grant reference for each page.
498 	 */
499 	for (i = 0; i < nseg; i++) {
500 		uint32_t flags;
501 
502 		if (use_persistent_gnts)
503 			persistent_gnt = get_persistent_gnt(
504 				&blkif->persistent_gnts,
505 				req->u.rw.seg[i].gref);
506 
507 		if (persistent_gnt) {
508 			/*
509 			 * We are using persistent grants and
510 			 * the grant is already mapped
511 			 */
512 			new_map = false;
513 		} else if (use_persistent_gnts &&
514 			   blkif->persistent_gnt_c <
515 			   max_mapped_grant_pages(blkif->blk_protocol)) {
516 			/*
517 			 * We are using persistent grants, the grant is
518 			 * not mapped but we have room for it
519 			 */
520 			new_map = true;
521 			persistent_gnt = kmalloc(
522 				sizeof(struct persistent_gnt),
523 				GFP_KERNEL);
524 			if (!persistent_gnt)
525 				return -ENOMEM;
526 			persistent_gnt->page = alloc_page(GFP_KERNEL);
527 			if (!persistent_gnt->page) {
528 				kfree(persistent_gnt);
529 				return -ENOMEM;
530 			}
531 			persistent_gnt->gnt = req->u.rw.seg[i].gref;
532 			persistent_gnt->handle = BLKBACK_INVALID_HANDLE;
533 
534 			pages_to_gnt[segs_to_map] =
535 				persistent_gnt->page;
536 			addr = (unsigned long) pfn_to_kaddr(
537 				page_to_pfn(persistent_gnt->page));
538 
539 			add_persistent_gnt(&blkif->persistent_gnts,
540 				persistent_gnt);
541 			blkif->persistent_gnt_c++;
542 			pr_debug(DRV_PFX " grant %u added to the tree of persistent grants, using %u/%u\n",
543 				 persistent_gnt->gnt, blkif->persistent_gnt_c,
544 				 max_mapped_grant_pages(blkif->blk_protocol));
545 		} else {
546 			/*
547 			 * We are either using persistent grants and
548 			 * hit the maximum limit of grants mapped,
549 			 * or we are not using persistent grants.
550 			 */
551 			if (use_persistent_gnts &&
552 				!blkif->vbd.overflow_max_grants) {
553 				blkif->vbd.overflow_max_grants = 1;
554 				pr_alert(DRV_PFX " domain %u, device %#x is using maximum number of persistent grants\n",
555 					 blkif->domid, blkif->vbd.handle);
556 			}
557 			new_map = true;
558 			pages[i] = blkbk->pending_page(pending_req, i);
559 			addr = vaddr(pending_req, i);
560 			pages_to_gnt[segs_to_map] =
561 				blkbk->pending_page(pending_req, i);
562 		}
563 
564 		if (persistent_gnt) {
565 			pages[i] = persistent_gnt->page;
566 			persistent_gnts[i] = persistent_gnt;
567 		} else {
568 			persistent_gnts[i] = NULL;
569 		}
570 
571 		if (new_map) {
572 			flags = GNTMAP_host_map;
573 			if (!persistent_gnt &&
574 			    (pending_req->operation != BLKIF_OP_READ))
575 				flags |= GNTMAP_readonly;
576 			gnttab_set_map_op(&map[segs_to_map++], addr,
577 					  flags, req->u.rw.seg[i].gref,
578 					  blkif->domid);
579 		}
580 	}
581 
582 	if (segs_to_map) {
583 		ret = gnttab_map_refs(map, NULL, pages_to_gnt, segs_to_map);
584 		BUG_ON(ret);
585 	}
586 
587 	/*
588 	 * Now swizzle the MFN in our domain with the MFN from the other domain
589 	 * so that when we access vaddr(pending_req,i) it has the contents of
590 	 * the page from the other domain.
591 	 */
592 	bitmap_zero(pending_req->unmap_seg, BLKIF_MAX_SEGMENTS_PER_REQUEST);
593 	for (i = 0, j = 0; i < nseg; i++) {
594 		if (!persistent_gnts[i] ||
595 		    persistent_gnts[i]->handle == BLKBACK_INVALID_HANDLE) {
596 			/* This is a newly mapped grant */
597 			BUG_ON(j >= segs_to_map);
598 			if (unlikely(map[j].status != 0)) {
599 				pr_debug(DRV_PFX "invalid buffer -- could not remap it\n");
600 				map[j].handle = BLKBACK_INVALID_HANDLE;
601 				ret |= 1;
602 				if (persistent_gnts[i]) {
603 					rb_erase(&persistent_gnts[i]->node,
604 						 &blkif->persistent_gnts);
605 					blkif->persistent_gnt_c--;
606 					kfree(persistent_gnts[i]);
607 					persistent_gnts[i] = NULL;
608 				}
609 			}
610 		}
611 		if (persistent_gnts[i]) {
612 			if (persistent_gnts[i]->handle ==
613 			    BLKBACK_INVALID_HANDLE) {
614 				/*
615 				 * If this is a new persistent grant
616 				 * save the handler
617 				 */
618 				persistent_gnts[i]->handle = map[j].handle;
619 				persistent_gnts[i]->dev_bus_addr =
620 					map[j++].dev_bus_addr;
621 			}
622 			pending_handle(pending_req, i) =
623 				persistent_gnts[i]->handle;
624 
625 			if (ret)
626 				continue;
627 
628 			seg[i].buf = persistent_gnts[i]->dev_bus_addr |
629 				(req->u.rw.seg[i].first_sect << 9);
630 		} else {
631 			pending_handle(pending_req, i) = map[j].handle;
632 			bitmap_set(pending_req->unmap_seg, i, 1);
633 
634 			if (ret) {
635 				j++;
636 				continue;
637 			}
638 
639 			seg[i].buf = map[j++].dev_bus_addr |
640 				(req->u.rw.seg[i].first_sect << 9);
641 		}
642 	}
643 	return ret;
644 }
645 
646 static int dispatch_discard_io(struct xen_blkif *blkif,
647 				struct blkif_request *req)
648 {
649 	int err = 0;
650 	int status = BLKIF_RSP_OKAY;
651 	struct block_device *bdev = blkif->vbd.bdev;
652 	unsigned long secure;
653 
654 	blkif->st_ds_req++;
655 
656 	xen_blkif_get(blkif);
657 	secure = (blkif->vbd.discard_secure &&
658 		 (req->u.discard.flag & BLKIF_DISCARD_SECURE)) ?
659 		 BLKDEV_DISCARD_SECURE : 0;
660 
661 	err = blkdev_issue_discard(bdev, req->u.discard.sector_number,
662 				   req->u.discard.nr_sectors,
663 				   GFP_KERNEL, secure);
664 
665 	if (err == -EOPNOTSUPP) {
666 		pr_debug(DRV_PFX "discard op failed, not supported\n");
667 		status = BLKIF_RSP_EOPNOTSUPP;
668 	} else if (err)
669 		status = BLKIF_RSP_ERROR;
670 
671 	make_response(blkif, req->u.discard.id, req->operation, status);
672 	xen_blkif_put(blkif);
673 	return err;
674 }
675 
676 static void xen_blk_drain_io(struct xen_blkif *blkif)
677 {
678 	atomic_set(&blkif->drain, 1);
679 	do {
680 		/* The initial value is one, and one refcnt taken at the
681 		 * start of the xen_blkif_schedule thread. */
682 		if (atomic_read(&blkif->refcnt) <= 2)
683 			break;
684 		wait_for_completion_interruptible_timeout(
685 				&blkif->drain_complete, HZ);
686 
687 		if (!atomic_read(&blkif->drain))
688 			break;
689 	} while (!kthread_should_stop());
690 	atomic_set(&blkif->drain, 0);
691 }
692 
693 /*
694  * Completion callback on the bio's. Called as bh->b_end_io()
695  */
696 
697 static void __end_block_io_op(struct pending_req *pending_req, int error)
698 {
699 	/* An error fails the entire request. */
700 	if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
701 	    (error == -EOPNOTSUPP)) {
702 		pr_debug(DRV_PFX "flush diskcache op failed, not supported\n");
703 		xen_blkbk_flush_diskcache(XBT_NIL, pending_req->blkif->be, 0);
704 		pending_req->status = BLKIF_RSP_EOPNOTSUPP;
705 	} else if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
706 		    (error == -EOPNOTSUPP)) {
707 		pr_debug(DRV_PFX "write barrier op failed, not supported\n");
708 		xen_blkbk_barrier(XBT_NIL, pending_req->blkif->be, 0);
709 		pending_req->status = BLKIF_RSP_EOPNOTSUPP;
710 	} else if (error) {
711 		pr_debug(DRV_PFX "Buffer not up-to-date at end of operation,"
712 			 " error=%d\n", error);
713 		pending_req->status = BLKIF_RSP_ERROR;
714 	}
715 
716 	/*
717 	 * If all of the bio's have completed it is time to unmap
718 	 * the grant references associated with 'request' and provide
719 	 * the proper response on the ring.
720 	 */
721 	if (atomic_dec_and_test(&pending_req->pendcnt)) {
722 		xen_blkbk_unmap(pending_req);
723 		make_response(pending_req->blkif, pending_req->id,
724 			      pending_req->operation, pending_req->status);
725 		xen_blkif_put(pending_req->blkif);
726 		if (atomic_read(&pending_req->blkif->refcnt) <= 2) {
727 			if (atomic_read(&pending_req->blkif->drain))
728 				complete(&pending_req->blkif->drain_complete);
729 		}
730 		free_req(pending_req);
731 	}
732 }
733 
734 /*
735  * bio callback.
736  */
737 static void end_block_io_op(struct bio *bio, int error)
738 {
739 	__end_block_io_op(bio->bi_private, error);
740 	bio_put(bio);
741 }
742 
743 
744 
745 /*
746  * Function to copy the from the ring buffer the 'struct blkif_request'
747  * (which has the sectors we want, number of them, grant references, etc),
748  * and transmute  it to the block API to hand it over to the proper block disk.
749  */
750 static int
751 __do_block_io_op(struct xen_blkif *blkif)
752 {
753 	union blkif_back_rings *blk_rings = &blkif->blk_rings;
754 	struct blkif_request req;
755 	struct pending_req *pending_req;
756 	RING_IDX rc, rp;
757 	int more_to_do = 0;
758 
759 	rc = blk_rings->common.req_cons;
760 	rp = blk_rings->common.sring->req_prod;
761 	rmb(); /* Ensure we see queued requests up to 'rp'. */
762 
763 	while (rc != rp) {
764 
765 		if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
766 			break;
767 
768 		if (kthread_should_stop()) {
769 			more_to_do = 1;
770 			break;
771 		}
772 
773 		pending_req = alloc_req();
774 		if (NULL == pending_req) {
775 			blkif->st_oo_req++;
776 			more_to_do = 1;
777 			break;
778 		}
779 
780 		switch (blkif->blk_protocol) {
781 		case BLKIF_PROTOCOL_NATIVE:
782 			memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
783 			break;
784 		case BLKIF_PROTOCOL_X86_32:
785 			blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
786 			break;
787 		case BLKIF_PROTOCOL_X86_64:
788 			blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
789 			break;
790 		default:
791 			BUG();
792 		}
793 		blk_rings->common.req_cons = ++rc; /* before make_response() */
794 
795 		/* Apply all sanity checks to /private copy/ of request. */
796 		barrier();
797 		if (unlikely(req.operation == BLKIF_OP_DISCARD)) {
798 			free_req(pending_req);
799 			if (dispatch_discard_io(blkif, &req))
800 				break;
801 		} else if (dispatch_rw_block_io(blkif, &req, pending_req))
802 			break;
803 
804 		/* Yield point for this unbounded loop. */
805 		cond_resched();
806 	}
807 
808 	return more_to_do;
809 }
810 
811 static int
812 do_block_io_op(struct xen_blkif *blkif)
813 {
814 	union blkif_back_rings *blk_rings = &blkif->blk_rings;
815 	int more_to_do;
816 
817 	do {
818 		more_to_do = __do_block_io_op(blkif);
819 		if (more_to_do)
820 			break;
821 
822 		RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
823 	} while (more_to_do);
824 
825 	return more_to_do;
826 }
827 /*
828  * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
829  * and call the 'submit_bio' to pass it to the underlying storage.
830  */
831 static int dispatch_rw_block_io(struct xen_blkif *blkif,
832 				struct blkif_request *req,
833 				struct pending_req *pending_req)
834 {
835 	struct phys_req preq;
836 	struct seg_buf seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
837 	unsigned int nseg;
838 	struct bio *bio = NULL;
839 	struct bio *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
840 	int i, nbio = 0;
841 	int operation;
842 	struct blk_plug plug;
843 	bool drain = false;
844 	struct page *pages[BLKIF_MAX_SEGMENTS_PER_REQUEST];
845 
846 	switch (req->operation) {
847 	case BLKIF_OP_READ:
848 		blkif->st_rd_req++;
849 		operation = READ;
850 		break;
851 	case BLKIF_OP_WRITE:
852 		blkif->st_wr_req++;
853 		operation = WRITE_ODIRECT;
854 		break;
855 	case BLKIF_OP_WRITE_BARRIER:
856 		drain = true;
857 	case BLKIF_OP_FLUSH_DISKCACHE:
858 		blkif->st_f_req++;
859 		operation = WRITE_FLUSH;
860 		break;
861 	default:
862 		operation = 0; /* make gcc happy */
863 		goto fail_response;
864 		break;
865 	}
866 
867 	/* Check that the number of segments is sane. */
868 	nseg = req->u.rw.nr_segments;
869 
870 	if (unlikely(nseg == 0 && operation != WRITE_FLUSH) ||
871 	    unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
872 		pr_debug(DRV_PFX "Bad number of segments in request (%d)\n",
873 			 nseg);
874 		/* Haven't submitted any bio's yet. */
875 		goto fail_response;
876 	}
877 
878 	preq.dev           = req->u.rw.handle;
879 	preq.sector_number = req->u.rw.sector_number;
880 	preq.nr_sects      = 0;
881 
882 	pending_req->blkif     = blkif;
883 	pending_req->id        = req->u.rw.id;
884 	pending_req->operation = req->operation;
885 	pending_req->status    = BLKIF_RSP_OKAY;
886 	pending_req->nr_pages  = nseg;
887 
888 	for (i = 0; i < nseg; i++) {
889 		seg[i].nsec = req->u.rw.seg[i].last_sect -
890 			req->u.rw.seg[i].first_sect + 1;
891 		if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
892 		    (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
893 			goto fail_response;
894 		preq.nr_sects += seg[i].nsec;
895 
896 	}
897 
898 	if (xen_vbd_translate(&preq, blkif, operation) != 0) {
899 		pr_debug(DRV_PFX "access denied: %s of [%llu,%llu] on dev=%04x\n",
900 			 operation == READ ? "read" : "write",
901 			 preq.sector_number,
902 			 preq.sector_number + preq.nr_sects, preq.dev);
903 		goto fail_response;
904 	}
905 
906 	/*
907 	 * This check _MUST_ be done after xen_vbd_translate as the preq.bdev
908 	 * is set there.
909 	 */
910 	for (i = 0; i < nseg; i++) {
911 		if (((int)preq.sector_number|(int)seg[i].nsec) &
912 		    ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
913 			pr_debug(DRV_PFX "Misaligned I/O request from domain %d",
914 				 blkif->domid);
915 			goto fail_response;
916 		}
917 	}
918 
919 	/* Wait on all outstanding I/O's and once that has been completed
920 	 * issue the WRITE_FLUSH.
921 	 */
922 	if (drain)
923 		xen_blk_drain_io(pending_req->blkif);
924 
925 	/*
926 	 * If we have failed at this point, we need to undo the M2P override,
927 	 * set gnttab_set_unmap_op on all of the grant references and perform
928 	 * the hypercall to unmap the grants - that is all done in
929 	 * xen_blkbk_unmap.
930 	 */
931 	if (xen_blkbk_map(req, pending_req, seg, pages))
932 		goto fail_flush;
933 
934 	/*
935 	 * This corresponding xen_blkif_put is done in __end_block_io_op, or
936 	 * below (in "!bio") if we are handling a BLKIF_OP_DISCARD.
937 	 */
938 	xen_blkif_get(blkif);
939 
940 	for (i = 0; i < nseg; i++) {
941 		while ((bio == NULL) ||
942 		       (bio_add_page(bio,
943 				     pages[i],
944 				     seg[i].nsec << 9,
945 				     seg[i].buf & ~PAGE_MASK) == 0)) {
946 
947 			bio = bio_alloc(GFP_KERNEL, nseg-i);
948 			if (unlikely(bio == NULL))
949 				goto fail_put_bio;
950 
951 			biolist[nbio++] = bio;
952 			bio->bi_bdev    = preq.bdev;
953 			bio->bi_private = pending_req;
954 			bio->bi_end_io  = end_block_io_op;
955 			bio->bi_sector  = preq.sector_number;
956 		}
957 
958 		preq.sector_number += seg[i].nsec;
959 	}
960 
961 	/* This will be hit if the operation was a flush or discard. */
962 	if (!bio) {
963 		BUG_ON(operation != WRITE_FLUSH);
964 
965 		bio = bio_alloc(GFP_KERNEL, 0);
966 		if (unlikely(bio == NULL))
967 			goto fail_put_bio;
968 
969 		biolist[nbio++] = bio;
970 		bio->bi_bdev    = preq.bdev;
971 		bio->bi_private = pending_req;
972 		bio->bi_end_io  = end_block_io_op;
973 	}
974 
975 	/*
976 	 * We set it one so that the last submit_bio does not have to call
977 	 * atomic_inc.
978 	 */
979 	atomic_set(&pending_req->pendcnt, nbio);
980 
981 	/* Get a reference count for the disk queue and start sending I/O */
982 	blk_start_plug(&plug);
983 
984 	for (i = 0; i < nbio; i++)
985 		submit_bio(operation, biolist[i]);
986 
987 	/* Let the I/Os go.. */
988 	blk_finish_plug(&plug);
989 
990 	if (operation == READ)
991 		blkif->st_rd_sect += preq.nr_sects;
992 	else if (operation & WRITE)
993 		blkif->st_wr_sect += preq.nr_sects;
994 
995 	return 0;
996 
997  fail_flush:
998 	xen_blkbk_unmap(pending_req);
999  fail_response:
1000 	/* Haven't submitted any bio's yet. */
1001 	make_response(blkif, req->u.rw.id, req->operation, BLKIF_RSP_ERROR);
1002 	free_req(pending_req);
1003 	msleep(1); /* back off a bit */
1004 	return -EIO;
1005 
1006  fail_put_bio:
1007 	for (i = 0; i < nbio; i++)
1008 		bio_put(biolist[i]);
1009 	__end_block_io_op(pending_req, -EINVAL);
1010 	msleep(1); /* back off a bit */
1011 	return -EIO;
1012 }
1013 
1014 
1015 
1016 /*
1017  * Put a response on the ring on how the operation fared.
1018  */
1019 static void make_response(struct xen_blkif *blkif, u64 id,
1020 			  unsigned short op, int st)
1021 {
1022 	struct blkif_response  resp;
1023 	unsigned long     flags;
1024 	union blkif_back_rings *blk_rings = &blkif->blk_rings;
1025 	int notify;
1026 
1027 	resp.id        = id;
1028 	resp.operation = op;
1029 	resp.status    = st;
1030 
1031 	spin_lock_irqsave(&blkif->blk_ring_lock, flags);
1032 	/* Place on the response ring for the relevant domain. */
1033 	switch (blkif->blk_protocol) {
1034 	case BLKIF_PROTOCOL_NATIVE:
1035 		memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
1036 		       &resp, sizeof(resp));
1037 		break;
1038 	case BLKIF_PROTOCOL_X86_32:
1039 		memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
1040 		       &resp, sizeof(resp));
1041 		break;
1042 	case BLKIF_PROTOCOL_X86_64:
1043 		memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
1044 		       &resp, sizeof(resp));
1045 		break;
1046 	default:
1047 		BUG();
1048 	}
1049 	blk_rings->common.rsp_prod_pvt++;
1050 	RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
1051 	spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
1052 	if (notify)
1053 		notify_remote_via_irq(blkif->irq);
1054 }
1055 
1056 static int __init xen_blkif_init(void)
1057 {
1058 	int i, mmap_pages;
1059 	int rc = 0;
1060 
1061 	if (!xen_domain())
1062 		return -ENODEV;
1063 
1064 	blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
1065 	if (!blkbk) {
1066 		pr_alert(DRV_PFX "%s: out of memory!\n", __func__);
1067 		return -ENOMEM;
1068 	}
1069 
1070 	mmap_pages = xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
1071 
1072 	blkbk->pending_reqs          = kzalloc(sizeof(blkbk->pending_reqs[0]) *
1073 					xen_blkif_reqs, GFP_KERNEL);
1074 	blkbk->pending_grant_handles = kmalloc(sizeof(blkbk->pending_grant_handles[0]) *
1075 					mmap_pages, GFP_KERNEL);
1076 	blkbk->pending_pages         = kzalloc(sizeof(blkbk->pending_pages[0]) *
1077 					mmap_pages, GFP_KERNEL);
1078 
1079 	if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
1080 	    !blkbk->pending_pages) {
1081 		rc = -ENOMEM;
1082 		goto out_of_memory;
1083 	}
1084 
1085 	for (i = 0; i < mmap_pages; i++) {
1086 		blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
1087 		blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
1088 		if (blkbk->pending_pages[i] == NULL) {
1089 			rc = -ENOMEM;
1090 			goto out_of_memory;
1091 		}
1092 	}
1093 	rc = xen_blkif_interface_init();
1094 	if (rc)
1095 		goto failed_init;
1096 
1097 	INIT_LIST_HEAD(&blkbk->pending_free);
1098 	spin_lock_init(&blkbk->pending_free_lock);
1099 	init_waitqueue_head(&blkbk->pending_free_wq);
1100 
1101 	for (i = 0; i < xen_blkif_reqs; i++)
1102 		list_add_tail(&blkbk->pending_reqs[i].free_list,
1103 			      &blkbk->pending_free);
1104 
1105 	rc = xen_blkif_xenbus_init();
1106 	if (rc)
1107 		goto failed_init;
1108 
1109 	return 0;
1110 
1111  out_of_memory:
1112 	pr_alert(DRV_PFX "%s: out of memory\n", __func__);
1113  failed_init:
1114 	kfree(blkbk->pending_reqs);
1115 	kfree(blkbk->pending_grant_handles);
1116 	if (blkbk->pending_pages) {
1117 		for (i = 0; i < mmap_pages; i++) {
1118 			if (blkbk->pending_pages[i])
1119 				__free_page(blkbk->pending_pages[i]);
1120 		}
1121 		kfree(blkbk->pending_pages);
1122 	}
1123 	kfree(blkbk);
1124 	blkbk = NULL;
1125 	return rc;
1126 }
1127 
1128 module_init(xen_blkif_init);
1129 
1130 MODULE_LICENSE("Dual BSD/GPL");
1131 MODULE_ALIAS("xen-backend:vbd");
1132