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 #include <linux/spinlock.h>
38 #include <linux/kthread.h>
39 #include <linux/list.h>
40 #include <linux/delay.h>
41 #include <linux/freezer.h>
43 #include <xen/events.h>
45 #include <asm/xen/hypervisor.h>
46 #include <asm/xen/hypercall.h>
49 #define WRITE_BARRIER (REQ_WRITE | REQ_FLUSH | REQ_FUA)
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.
56 * ** TRY INCREASING 'blkif_reqs' IF WRITE SPEEDS SEEM TOO LOW **
58 * This will increase the chances of being able to write whole tracks.
59 * 64 should be enough to keep us competitive with Linux.
61 static int blkif_reqs = 64;
62 module_param_named(reqs, blkif_reqs, int, 0);
63 MODULE_PARM_DESC(reqs, "Number of blkback requests to allocate");
65 /* Run-time switchable: /sys/module/blkback/parameters/ */
66 static unsigned int log_stats;
67 static unsigned int debug_lvl;
68 module_param(log_stats, int, 0644);
69 module_param(debug_lvl, int, 0644);
72 * Each outstanding request that we've passed to the lower device layers has a
73 * 'pending_req' allocated to it. Each buffer_head that completes decrements
74 * the pendcnt towards zero. When it hits zero, the specified domain has a
75 * response queued for it, with the saved 'id' passed back.
78 struct blkif_st *blkif;
82 unsigned short operation;
84 struct list_head free_list;
87 #define BLKBACK_INVALID_HANDLE (~0)
90 struct pending_req *pending_reqs;
91 /* List of all 'pending_req' available */
92 struct list_head pending_free;
93 /* And its spinlock. */
94 spinlock_t pending_free_lock;
95 wait_queue_head_t pending_free_wq;
96 /* The list of all pages that are available. */
97 struct page **pending_pages;
98 /* And the grant handles that are available. */
99 grant_handle_t *pending_grant_handles;
102 static struct xen_blkbk *blkbk;
105 * Little helpful macro to figure out the index and virtual address of the
106 * pending_pages[..]. For each 'pending_req' we have have up to
107 * BLKIF_MAX_SEGMENTS_PER_REQUEST (11) pages. The seg would be from 0 through
108 * 10 and would index in the pending_pages[..]. */
109 static inline int vaddr_pagenr(struct pending_req *req, int seg)
111 return (req - blkbk->pending_reqs) *
112 BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
115 #define pending_page(req, seg) pending_pages[vaddr_pagenr(req, seg)]
117 static inline unsigned long vaddr(struct pending_req *req, int seg)
119 unsigned long pfn = page_to_pfn(blkbk->pending_page(req, seg));
120 return (unsigned long)pfn_to_kaddr(pfn);
123 #define pending_handle(_req, _seg) \
124 (blkbk->pending_grant_handles[vaddr_pagenr(_req, _seg)])
127 static int do_block_io_op(struct blkif_st *blkif);
128 static void dispatch_rw_block_io(struct blkif_st *blkif,
129 struct blkif_request *req,
130 struct pending_req *pending_req);
131 static void make_response(struct blkif_st *blkif, u64 id,
132 unsigned short op, int st);
135 * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
137 static struct pending_req *alloc_req(void)
139 struct pending_req *req = NULL;
142 spin_lock_irqsave(&blkbk->pending_free_lock, flags);
143 if (!list_empty(&blkbk->pending_free)) {
144 req = list_entry(blkbk->pending_free.next, struct pending_req,
146 list_del(&req->free_list);
148 spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
153 * Return the 'pending_req' structure back to the freepool. We also
154 * wake up the thread if it was waiting for a free page.
156 static void free_req(struct pending_req *req)
161 spin_lock_irqsave(&blkbk->pending_free_lock, flags);
162 was_empty = list_empty(&blkbk->pending_free);
163 list_add(&req->free_list, &blkbk->pending_free);
164 spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
166 wake_up(&blkbk->pending_free_wq);
170 * Notification from the guest OS.
172 static void blkif_notify_work(struct blkif_st *blkif)
174 blkif->waiting_reqs = 1;
178 irqreturn_t blkif_be_int(int irq, void *dev_id)
180 blkif_notify_work(dev_id);
185 * SCHEDULER FUNCTIONS
188 static void print_stats(struct blkif_st *blkif)
190 printk(KERN_DEBUG "%s: oo %3d | rd %4d | wr %4d | br %4d\n",
191 current->comm, blkif->st_oo_req,
192 blkif->st_rd_req, blkif->st_wr_req, blkif->st_br_req);
193 blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
194 blkif->st_rd_req = 0;
195 blkif->st_wr_req = 0;
196 blkif->st_oo_req = 0;
199 int blkif_schedule(void *arg)
201 struct blkif_st *blkif = arg;
202 struct vbd *vbd = &blkif->vbd;
207 printk(KERN_DEBUG "%s: started\n", current->comm);
209 while (!kthread_should_stop()) {
212 if (unlikely(vbd->size != vbd_size(vbd)))
215 wait_event_interruptible(
217 blkif->waiting_reqs || kthread_should_stop());
218 wait_event_interruptible(
219 blkbk->pending_free_wq,
220 !list_empty(&blkbk->pending_free) ||
221 kthread_should_stop());
223 blkif->waiting_reqs = 0;
224 smp_mb(); /* clear flag *before* checking for work */
226 if (do_block_io_op(blkif))
227 blkif->waiting_reqs = 1;
229 if (log_stats && time_after(jiffies, blkif->st_print))
236 printk(KERN_DEBUG "%s: exiting\n", current->comm);
238 blkif->xenblkd = NULL;
249 * Unmap the grant references, and also remove the M2P over-rides
250 * used in the 'pending_req'.
252 static void xen_blkbk_unmap(struct pending_req *req)
254 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
255 unsigned int i, invcount = 0;
256 grant_handle_t handle;
259 for (i = 0; i < req->nr_pages; i++) {
260 handle = pending_handle(req, i);
261 if (handle == BLKBACK_INVALID_HANDLE)
263 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
264 GNTMAP_host_map, handle);
265 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
269 ret = HYPERVISOR_grant_table_op(
270 GNTTABOP_unmap_grant_ref, unmap, invcount);
272 /* Note, we use invcount, so nr->pages, so we can't index
273 * using vaddr(req, i).
275 for (i = 0; i < invcount; i++) {
276 ret = m2p_remove_override(
277 virt_to_page(unmap[i].host_addr), false);
279 printk(KERN_ALERT "Failed to remove M2P override for " \
280 "%lx\n", (unsigned long)unmap[i].host_addr);
285 static int xen_blkbk_map(struct blkif_request *req, struct pending_req *pending_req,
286 struct seg_buf seg[])
288 struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
290 int nseg = req->nr_segments;
292 /* Fill out preq.nr_sects with proper amount of sectors, and setup
293 * assign map[..] with the PFN of the page in our domain with the
294 * corresponding grant reference for each page.
296 for (i = 0; i < nseg; i++) {
299 flags = GNTMAP_host_map;
300 if (pending_req->operation != BLKIF_OP_READ)
301 flags |= GNTMAP_readonly;
302 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
303 req->u.rw.seg[i].gref, pending_req->blkif->domid);
306 ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
309 /* Now swizzel the MFN in our domain with the MFN from the other domain
310 * so that when we access vaddr(pending_req,i) it has the contents of
311 * the page from the other domain.
313 for (i = 0; i < nseg; i++) {
314 if (unlikely(map[i].status != 0)) {
315 DPRINTK("invalid buffer -- could not remap it\n");
316 map[i].handle = BLKBACK_INVALID_HANDLE;
320 pending_handle(pending_req, i) = map[i].handle;
325 ret = m2p_add_override(PFN_DOWN(map[i].dev_bus_addr),
326 blkbk->pending_page(pending_req, i), false);
328 printk(KERN_ALERT "Failed to install M2P override for"\
329 " %lx (ret: %d)\n", (unsigned long)
330 map[i].dev_bus_addr, ret);
331 /* We could switch over to GNTTABOP_copy */
335 seg[i].buf = map[i].dev_bus_addr |
336 (req->u.rw.seg[i].first_sect << 9);
342 * Completion callback on the bio's. Called as bh->b_end_io()
345 static void __end_block_io_op(struct pending_req *pending_req, int error)
347 /* An error fails the entire request. */
348 if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
349 (error == -EOPNOTSUPP)) {
350 DPRINTK("blkback: write barrier op failed, not supported\n");
351 blkback_barrier(XBT_NIL, pending_req->blkif->be, 0);
352 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
354 DPRINTK("Buffer not up-to-date at end of operation, "
355 "error=%d\n", error);
356 pending_req->status = BLKIF_RSP_ERROR;
359 /* If all of the bio's have completed it is time to unmap
360 * the grant references associated with 'request' and provide
361 * the proper response on the ring.
363 if (atomic_dec_and_test(&pending_req->pendcnt)) {
364 xen_blkbk_unmap(pending_req);
365 make_response(pending_req->blkif, pending_req->id,
366 pending_req->operation, pending_req->status);
367 blkif_put(pending_req->blkif);
368 free_req(pending_req);
375 static void end_block_io_op(struct bio *bio, int error)
377 __end_block_io_op(bio->bi_private, error);
384 * Function to copy the from the ring buffer the 'struct blkif_request'
385 * (which has the sectors we want, number of them, grant references, etc),
386 * and transmute it to the block API to hand it over to the proper block disk.
388 static int do_block_io_op(struct blkif_st *blkif)
390 union blkif_back_rings *blk_rings = &blkif->blk_rings;
391 struct blkif_request req;
392 struct pending_req *pending_req;
396 rc = blk_rings->common.req_cons;
397 rp = blk_rings->common.sring->req_prod;
398 rmb(); /* Ensure we see queued requests up to 'rp'. */
402 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
405 if (kthread_should_stop()) {
410 pending_req = alloc_req();
411 if (NULL == pending_req) {
417 switch (blkif->blk_protocol) {
418 case BLKIF_PROTOCOL_NATIVE:
419 memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
421 case BLKIF_PROTOCOL_X86_32:
422 blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
424 case BLKIF_PROTOCOL_X86_64:
425 blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
430 blk_rings->common.req_cons = ++rc; /* before make_response() */
432 /* Apply all sanity checks to /private copy/ of request. */
435 switch (req.operation) {
438 dispatch_rw_block_io(blkif, &req, pending_req);
440 case BLKIF_OP_WRITE_BARRIER:
445 dispatch_rw_block_io(blkif, &req, pending_req);
448 /* A good sign something is wrong: sleep for a while to
449 * avoid excessive CPU consumption by a bad guest. */
451 DPRINTK("error: unknown block io operation [%d]\n",
453 make_response(blkif, req.id, req.operation,
455 free_req(pending_req);
459 /* Yield point for this unbounded loop. */
467 * Transumation of the 'struct blkif_request' to a proper 'struct bio'
468 * and call the 'submit_bio' to pass it to the underlaying storage.
470 static void dispatch_rw_block_io(struct blkif_st *blkif,
471 struct blkif_request *req,
472 struct pending_req *pending_req)
474 struct phys_req preq;
475 struct seg_buf seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
477 struct bio *bio = NULL;
478 struct bio *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
481 struct blk_plug plug;
483 switch (req->operation) {
490 case BLKIF_OP_WRITE_BARRIER:
491 operation = WRITE_BARRIER;
494 operation = 0; /* make gcc happy */
498 /* Check that the number of segments is sane. */
499 nseg = req->nr_segments;
500 if (unlikely(nseg == 0 && operation != WRITE_BARRIER) ||
501 unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
502 DPRINTK("Bad number of segments in request (%d)\n", nseg);
503 /* Haven't submitted any bio's yet. */
507 preq.dev = req->handle;
508 preq.sector_number = req->u.rw.sector_number;
511 pending_req->blkif = blkif;
512 pending_req->id = req->id;
513 pending_req->operation = req->operation;
514 pending_req->status = BLKIF_RSP_OKAY;
515 pending_req->nr_pages = nseg;
517 for (i = 0; i < nseg; i++) {
518 seg[i].nsec = req->u.rw.seg[i].last_sect -
519 req->u.rw.seg[i].first_sect + 1;
520 if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
521 (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
523 preq.nr_sects += seg[i].nsec;
527 if (vbd_translate(&preq, blkif, operation) != 0) {
528 DPRINTK("access denied: %s of [%llu,%llu] on dev=%04x\n",
529 operation == READ ? "read" : "write",
531 preq.sector_number + preq.nr_sects, preq.dev);
534 /* This check _MUST_ be done after vbd_translate as the preq.bdev
536 for (i = 0; i < nseg; i++) {
537 if (((int)preq.sector_number|(int)seg[i].nsec) &
538 ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
539 DPRINTK("Misaligned I/O request from domain %d",
544 /* If we have failed at this point, we need to undo the M2P override,
545 * set gnttab_set_unmap_op on all of the grant references and perform
546 * the hypercall to unmap the grants - that is all done in
549 if (xen_blkbk_map(req, pending_req, seg))
552 /* This corresponding blkif_put is done in __end_block_io_op */
555 for (i = 0; i < nseg; i++) {
556 while ((bio == NULL) ||
558 blkbk->pending_page(pending_req, i),
560 seg[i].buf & ~PAGE_MASK) == 0)) {
562 bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, nseg-i);
563 if (unlikely(bio == NULL))
566 bio->bi_bdev = preq.bdev;
567 bio->bi_private = pending_req;
568 bio->bi_end_io = end_block_io_op;
569 bio->bi_sector = preq.sector_number;
572 preq.sector_number += seg[i].nsec;
575 /* This will be hit if the operation was a barrier. */
577 BUG_ON(operation != WRITE_BARRIER);
578 bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, 0);
579 if (unlikely(bio == NULL))
582 bio->bi_bdev = preq.bdev;
583 bio->bi_private = pending_req;
584 bio->bi_end_io = end_block_io_op;
589 /* We set it one so that the last submit_bio does not have to call
592 atomic_set(&pending_req->pendcnt, nbio);
594 /* Get a reference count for the disk queue and start sending I/O */
595 blk_start_plug(&plug);
597 for (i = 0; i < nbio; i++)
598 submit_bio(operation, biolist[i]);
600 blk_finish_plug(&plug);
601 /* Let the I/Os go.. */
603 if (operation == READ)
604 blkif->st_rd_sect += preq.nr_sects;
605 else if (operation == WRITE || operation == WRITE_BARRIER)
606 blkif->st_wr_sect += preq.nr_sects;
611 xen_blkbk_unmap(pending_req);
613 /* Haven't submitted any bio's yet. */
614 make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
615 free_req(pending_req);
616 msleep(1); /* back off a bit */
620 for (i = 0; i < (nbio-1); i++)
622 __end_block_io_op(pending_req, -EINVAL);
623 msleep(1); /* back off a bit */
630 * Put a response on the ring on how the operation fared.
632 static void make_response(struct blkif_st *blkif, u64 id,
633 unsigned short op, int st)
635 struct blkif_response resp;
637 union blkif_back_rings *blk_rings = &blkif->blk_rings;
645 spin_lock_irqsave(&blkif->blk_ring_lock, flags);
646 /* Place on the response ring for the relevant domain. */
647 switch (blkif->blk_protocol) {
648 case BLKIF_PROTOCOL_NATIVE:
649 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
650 &resp, sizeof(resp));
652 case BLKIF_PROTOCOL_X86_32:
653 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
654 &resp, sizeof(resp));
656 case BLKIF_PROTOCOL_X86_64:
657 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
658 &resp, sizeof(resp));
663 blk_rings->common.rsp_prod_pvt++;
664 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
665 if (blk_rings->common.rsp_prod_pvt == blk_rings->common.req_cons) {
667 * Tail check for pending requests. Allows frontend to avoid
668 * notifications if requests are already in flight (lower
669 * overheads and promotes batching).
671 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
673 } else if (RING_HAS_UNCONSUMED_REQUESTS(&blk_rings->common)) {
677 spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
680 blkif_notify_work(blkif);
682 notify_remote_via_irq(blkif->irq);
685 static int __init blkif_init(void)
690 if (!xen_pv_domain())
693 blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
695 printk(KERN_ALERT "%s: out of memory!\n", __func__);
699 mmap_pages = blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
701 blkbk->pending_reqs = kmalloc(sizeof(blkbk->pending_reqs[0]) *
702 blkif_reqs, GFP_KERNEL);
703 blkbk->pending_grant_handles = kzalloc(sizeof(blkbk->pending_grant_handles[0]) *
704 mmap_pages, GFP_KERNEL);
705 blkbk->pending_pages = kzalloc(sizeof(blkbk->pending_pages[0]) *
706 mmap_pages, GFP_KERNEL);
708 if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
709 !blkbk->pending_pages) {
714 for (i = 0; i < mmap_pages; i++) {
715 blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
716 blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
717 if (blkbk->pending_pages[i] == NULL) {
722 rc = blkif_interface_init();
726 memset(blkbk->pending_reqs, 0, sizeof(blkbk->pending_reqs));
728 INIT_LIST_HEAD(&blkbk->pending_free);
729 spin_lock_init(&blkbk->pending_free_lock);
730 init_waitqueue_head(&blkbk->pending_free_wq);
732 for (i = 0; i < blkif_reqs; i++)
733 list_add_tail(&blkbk->pending_reqs[i].free_list,
734 &blkbk->pending_free);
736 rc = blkif_xenbus_init();
743 printk(KERN_ERR "%s: out of memory\n", __func__);
745 kfree(blkbk->pending_reqs);
746 kfree(blkbk->pending_grant_handles);
747 for (i = 0; i < mmap_pages; i++) {
748 if (blkbk->pending_pages[i])
749 __free_page(blkbk->pending_pages[i]);
751 kfree(blkbk->pending_pages);
757 module_init(blkif_init);
759 MODULE_LICENSE("Dual BSD/GPL");