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 'xen_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 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");
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 * Routines for managing virtual block devices (vbds).
174 static int vbd_translate(struct phys_req *req, struct blkif_st *blkif,
177 struct vbd *vbd = &blkif->vbd;
180 if ((operation != READ) && vbd->readonly)
183 if (unlikely((req->sector_number + req->nr_sects) > vbd_sz(vbd)))
186 req->dev = vbd->pdevice;
187 req->bdev = vbd->bdev;
194 static void vbd_resize(struct blkif_st *blkif)
196 struct vbd *vbd = &blkif->vbd;
197 struct xenbus_transaction xbt;
199 struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
200 unsigned long long new_size = vbd_sz(vbd);
202 printk(KERN_INFO "VBD Resize: Domid: %d, Device: (%d, %d)\n",
203 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
204 printk(KERN_INFO "VBD Resize: new size %llu\n", new_size);
205 vbd->size = new_size;
207 err = xenbus_transaction_start(&xbt);
209 printk(KERN_WARNING "Error starting transaction");
212 err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
213 (unsigned long long)vbd_sz(vbd));
215 printk(KERN_WARNING "Error writing new size");
219 * Write the current state; we will use this to synchronize
220 * the front-end. If the current state is "connected" the
221 * front-end will get the new size information online.
223 err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
225 printk(KERN_WARNING "Error writing the state");
229 err = xenbus_transaction_end(xbt, 0);
233 printk(KERN_WARNING "Error ending transaction");
235 xenbus_transaction_end(xbt, 1);
239 * Notification from the guest OS.
241 static void blkif_notify_work(struct blkif_st *blkif)
243 blkif->waiting_reqs = 1;
247 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
249 blkif_notify_work(dev_id);
254 * SCHEDULER FUNCTIONS
257 static void print_stats(struct blkif_st *blkif)
259 printk(KERN_DEBUG "%s: oo %3d | rd %4d | wr %4d | br %4d\n",
260 current->comm, blkif->st_oo_req,
261 blkif->st_rd_req, blkif->st_wr_req, blkif->st_br_req);
262 blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
263 blkif->st_rd_req = 0;
264 blkif->st_wr_req = 0;
265 blkif->st_oo_req = 0;
268 int xen_blkif_schedule(void *arg)
270 struct blkif_st *blkif = arg;
271 struct vbd *vbd = &blkif->vbd;
273 xen_blkif_get(blkif);
276 printk(KERN_DEBUG "%s: started\n", current->comm);
278 while (!kthread_should_stop()) {
279 struct blk_plug plug;
283 if (unlikely(vbd->size != vbd_sz(vbd)))
286 wait_event_interruptible(
288 blkif->waiting_reqs || kthread_should_stop());
289 wait_event_interruptible(
290 blkbk->pending_free_wq,
291 !list_empty(&blkbk->pending_free) ||
292 kthread_should_stop());
294 blkif->waiting_reqs = 0;
295 smp_mb(); /* clear flag *before* checking for work */
297 blk_start_plug(&plug);
299 if (do_block_io_op(blkif))
300 blkif->waiting_reqs = 1;
302 blk_finish_plug(&plug);
304 if (log_stats && time_after(jiffies, blkif->st_print))
311 printk(KERN_DEBUG "%s: exiting\n", current->comm);
313 blkif->xenblkd = NULL;
314 xen_blkif_put(blkif);
324 * Unmap the grant references, and also remove the M2P over-rides
325 * used in the 'pending_req'.
327 static void xen_blkbk_unmap(struct pending_req *req)
329 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
330 unsigned int i, invcount = 0;
331 grant_handle_t handle;
334 for (i = 0; i < req->nr_pages; i++) {
335 handle = pending_handle(req, i);
336 if (handle == BLKBACK_INVALID_HANDLE)
338 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
339 GNTMAP_host_map, handle);
340 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
344 ret = HYPERVISOR_grant_table_op(
345 GNTTABOP_unmap_grant_ref, unmap, invcount);
347 /* Note, we use invcount, so nr->pages, so we can't index
348 * using vaddr(req, i).
350 for (i = 0; i < invcount; i++) {
351 ret = m2p_remove_override(
352 virt_to_page(unmap[i].host_addr), false);
354 printk(KERN_ALERT "Failed to remove M2P override for " \
355 "%lx\n", (unsigned long)unmap[i].host_addr);
360 static int xen_blkbk_map(struct blkif_request *req, struct pending_req *pending_req,
361 struct seg_buf seg[])
363 struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
365 int nseg = req->nr_segments;
367 /* Fill out preq.nr_sects with proper amount of sectors, and setup
368 * assign map[..] with the PFN of the page in our domain with the
369 * corresponding grant reference for each page.
371 for (i = 0; i < nseg; i++) {
374 flags = GNTMAP_host_map;
375 if (pending_req->operation != BLKIF_OP_READ)
376 flags |= GNTMAP_readonly;
377 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
378 req->u.rw.seg[i].gref, pending_req->blkif->domid);
381 ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
384 /* Now swizzel the MFN in our domain with the MFN from the other domain
385 * so that when we access vaddr(pending_req,i) it has the contents of
386 * the page from the other domain.
388 for (i = 0; i < nseg; i++) {
389 if (unlikely(map[i].status != 0)) {
390 DPRINTK("invalid buffer -- could not remap it\n");
391 map[i].handle = BLKBACK_INVALID_HANDLE;
395 pending_handle(pending_req, i) = map[i].handle;
400 ret = m2p_add_override(PFN_DOWN(map[i].dev_bus_addr),
401 blkbk->pending_page(pending_req, i), false);
403 printk(KERN_ALERT "Failed to install M2P override for"\
404 " %lx (ret: %d)\n", (unsigned long)
405 map[i].dev_bus_addr, ret);
406 /* We could switch over to GNTTABOP_copy */
410 seg[i].buf = map[i].dev_bus_addr |
411 (req->u.rw.seg[i].first_sect << 9);
417 * Completion callback on the bio's. Called as bh->b_end_io()
420 static void __end_block_io_op(struct pending_req *pending_req, int error)
422 /* An error fails the entire request. */
423 if ((pending_req->operation == BLKIF_OP_WRITE_BARRIER) &&
424 (error == -EOPNOTSUPP)) {
425 DPRINTK("blkback: write barrier op failed, not supported\n");
426 xen_blkbk_barrier(XBT_NIL, pending_req->blkif->be, 0);
427 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
429 DPRINTK("Buffer not up-to-date at end of operation, "
430 "error=%d\n", error);
431 pending_req->status = BLKIF_RSP_ERROR;
434 /* If all of the bio's have completed it is time to unmap
435 * the grant references associated with 'request' and provide
436 * the proper response on the ring.
438 if (atomic_dec_and_test(&pending_req->pendcnt)) {
439 xen_blkbk_unmap(pending_req);
440 make_response(pending_req->blkif, pending_req->id,
441 pending_req->operation, pending_req->status);
442 xen_blkif_put(pending_req->blkif);
443 free_req(pending_req);
450 static void end_block_io_op(struct bio *bio, int error)
452 __end_block_io_op(bio->bi_private, error);
459 * Function to copy the from the ring buffer the 'struct blkif_request'
460 * (which has the sectors we want, number of them, grant references, etc),
461 * and transmute it to the block API to hand it over to the proper block disk.
463 static int do_block_io_op(struct blkif_st *blkif)
465 union blkif_back_rings *blk_rings = &blkif->blk_rings;
466 struct blkif_request req;
467 struct pending_req *pending_req;
471 rc = blk_rings->common.req_cons;
472 rp = blk_rings->common.sring->req_prod;
473 rmb(); /* Ensure we see queued requests up to 'rp'. */
477 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
480 if (kthread_should_stop()) {
485 pending_req = alloc_req();
486 if (NULL == pending_req) {
492 switch (blkif->blk_protocol) {
493 case BLKIF_PROTOCOL_NATIVE:
494 memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
496 case BLKIF_PROTOCOL_X86_32:
497 blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
499 case BLKIF_PROTOCOL_X86_64:
500 blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
505 blk_rings->common.req_cons = ++rc; /* before make_response() */
507 /* Apply all sanity checks to /private copy/ of request. */
510 switch (req.operation) {
513 dispatch_rw_block_io(blkif, &req, pending_req);
515 case BLKIF_OP_WRITE_BARRIER:
520 dispatch_rw_block_io(blkif, &req, pending_req);
523 /* A good sign something is wrong: sleep for a while to
524 * avoid excessive CPU consumption by a bad guest. */
526 DPRINTK("error: unknown block io operation [%d]\n",
528 make_response(blkif, req.id, req.operation,
530 free_req(pending_req);
534 /* Yield point for this unbounded loop. */
542 * Transumation of the 'struct blkif_request' to a proper 'struct bio'
543 * and call the 'submit_bio' to pass it to the underlaying storage.
545 static void dispatch_rw_block_io(struct blkif_st *blkif,
546 struct blkif_request *req,
547 struct pending_req *pending_req)
549 struct phys_req preq;
550 struct seg_buf seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
552 struct bio *bio = NULL;
553 struct bio *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
557 switch (req->operation) {
562 operation = WRITE_ODIRECT;
564 case BLKIF_OP_WRITE_BARRIER:
565 operation = WRITE_BARRIER;
568 operation = 0; /* make gcc happy */
572 /* Check that the number of segments is sane. */
573 nseg = req->nr_segments;
574 if (unlikely(nseg == 0 && operation != WRITE_BARRIER) ||
575 unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
576 DPRINTK("Bad number of segments in request (%d)\n", nseg);
577 /* Haven't submitted any bio's yet. */
581 preq.dev = req->handle;
582 preq.sector_number = req->u.rw.sector_number;
585 pending_req->blkif = blkif;
586 pending_req->id = req->id;
587 pending_req->operation = req->operation;
588 pending_req->status = BLKIF_RSP_OKAY;
589 pending_req->nr_pages = nseg;
591 for (i = 0; i < nseg; i++) {
592 seg[i].nsec = req->u.rw.seg[i].last_sect -
593 req->u.rw.seg[i].first_sect + 1;
594 if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
595 (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
597 preq.nr_sects += seg[i].nsec;
601 if (vbd_translate(&preq, blkif, operation) != 0) {
602 DPRINTK("access denied: %s of [%llu,%llu] on dev=%04x\n",
603 operation == READ ? "read" : "write",
605 preq.sector_number + preq.nr_sects, preq.dev);
608 /* This check _MUST_ be done after vbd_translate as the preq.bdev
610 for (i = 0; i < nseg; i++) {
611 if (((int)preq.sector_number|(int)seg[i].nsec) &
612 ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
613 DPRINTK("Misaligned I/O request from domain %d",
618 /* If we have failed at this point, we need to undo the M2P override,
619 * set gnttab_set_unmap_op on all of the grant references and perform
620 * the hypercall to unmap the grants - that is all done in
623 if (xen_blkbk_map(req, pending_req, seg))
626 /* This corresponding blkif_put is done in __end_block_io_op */
627 xen_blkif_get(blkif);
629 for (i = 0; i < nseg; i++) {
630 while ((bio == NULL) ||
632 blkbk->pending_page(pending_req, i),
634 seg[i].buf & ~PAGE_MASK) == 0)) {
636 bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, nseg-i);
637 if (unlikely(bio == NULL))
640 bio->bi_bdev = preq.bdev;
641 bio->bi_private = pending_req;
642 bio->bi_end_io = end_block_io_op;
643 bio->bi_sector = preq.sector_number;
646 preq.sector_number += seg[i].nsec;
649 /* This will be hit if the operation was a barrier. */
651 BUG_ON(operation != WRITE_BARRIER);
652 bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, 0);
653 if (unlikely(bio == NULL))
656 bio->bi_bdev = preq.bdev;
657 bio->bi_private = pending_req;
658 bio->bi_end_io = end_block_io_op;
663 /* We set it one so that the last submit_bio does not have to call
666 atomic_set(&pending_req->pendcnt, nbio);
668 for (i = 0; i < nbio; i++)
669 submit_bio(operation, biolist[i]);
671 if (operation == READ)
672 blkif->st_rd_sect += preq.nr_sects;
673 else if (operation == WRITE || operation == WRITE_BARRIER)
674 blkif->st_wr_sect += preq.nr_sects;
679 xen_blkbk_unmap(pending_req);
681 /* Haven't submitted any bio's yet. */
682 make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
683 free_req(pending_req);
684 msleep(1); /* back off a bit */
688 for (i = 0; i < (nbio-1); i++)
690 __end_block_io_op(pending_req, -EINVAL);
691 msleep(1); /* back off a bit */
698 * Put a response on the ring on how the operation fared.
700 static void make_response(struct blkif_st *blkif, u64 id,
701 unsigned short op, int st)
703 struct blkif_response resp;
705 union blkif_back_rings *blk_rings = &blkif->blk_rings;
713 spin_lock_irqsave(&blkif->blk_ring_lock, flags);
714 /* Place on the response ring for the relevant domain. */
715 switch (blkif->blk_protocol) {
716 case BLKIF_PROTOCOL_NATIVE:
717 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
718 &resp, sizeof(resp));
720 case BLKIF_PROTOCOL_X86_32:
721 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
722 &resp, sizeof(resp));
724 case BLKIF_PROTOCOL_X86_64:
725 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
726 &resp, sizeof(resp));
731 blk_rings->common.rsp_prod_pvt++;
732 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
733 if (blk_rings->common.rsp_prod_pvt == blk_rings->common.req_cons) {
735 * Tail check for pending requests. Allows frontend to avoid
736 * notifications if requests are already in flight (lower
737 * overheads and promotes batching).
739 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
741 } else if (RING_HAS_UNCONSUMED_REQUESTS(&blk_rings->common)) {
745 spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
748 blkif_notify_work(blkif);
750 notify_remote_via_irq(blkif->irq);
753 static int __init xen_blkif_init(void)
758 if (!xen_pv_domain())
761 blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
763 printk(KERN_ALERT "%s: out of memory!\n", __func__);
767 mmap_pages = xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
769 blkbk->pending_reqs = kmalloc(sizeof(blkbk->pending_reqs[0]) *
770 xen_blkif_reqs, GFP_KERNEL);
771 blkbk->pending_grant_handles = kzalloc(sizeof(blkbk->pending_grant_handles[0]) *
772 mmap_pages, GFP_KERNEL);
773 blkbk->pending_pages = kzalloc(sizeof(blkbk->pending_pages[0]) *
774 mmap_pages, GFP_KERNEL);
776 if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
777 !blkbk->pending_pages) {
782 for (i = 0; i < mmap_pages; i++) {
783 blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
784 blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
785 if (blkbk->pending_pages[i] == NULL) {
790 rc = xen_blkif_interface_init();
794 memset(blkbk->pending_reqs, 0, sizeof(blkbk->pending_reqs));
796 INIT_LIST_HEAD(&blkbk->pending_free);
797 spin_lock_init(&blkbk->pending_free_lock);
798 init_waitqueue_head(&blkbk->pending_free_wq);
800 for (i = 0; i < xen_blkif_reqs; i++)
801 list_add_tail(&blkbk->pending_reqs[i].free_list,
802 &blkbk->pending_free);
804 rc = xen_blkif_xenbus_init();
811 printk(KERN_ERR "%s: out of memory\n", __func__);
813 kfree(blkbk->pending_reqs);
814 kfree(blkbk->pending_grant_handles);
815 for (i = 0; i < mmap_pages; i++) {
816 if (blkbk->pending_pages[i])
817 __free_page(blkbk->pending_pages[i]);
819 kfree(blkbk->pending_pages);
825 module_init(xen_blkif_init);
827 MODULE_LICENSE("Dual BSD/GPL");