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>
50 * These are rather arbitrary. They are fairly large because adjacent requests
51 * pulled from a communication ring are quite likely to end up being part of
52 * the same scatter/gather request at the disc.
54 * ** TRY INCREASING 'xen_blkif_reqs' IF WRITE SPEEDS SEEM TOO LOW **
56 * This will increase the chances of being able to write whole tracks.
57 * 64 should be enough to keep us competitive with Linux.
59 static int xen_blkif_reqs = 64;
60 module_param_named(reqs, xen_blkif_reqs, int, 0);
61 MODULE_PARM_DESC(reqs, "Number of blkback requests to allocate");
63 /* Run-time switchable: /sys/module/blkback/parameters/ */
64 static unsigned int log_stats;
65 module_param(log_stats, int, 0644);
68 * Each outstanding request that we've passed to the lower device layers has a
69 * 'pending_req' allocated to it. Each buffer_head that completes decrements
70 * the pendcnt towards zero. When it hits zero, the specified domain has a
71 * response queued for it, with the saved 'id' passed back.
74 struct xen_blkif *blkif;
78 unsigned short operation;
80 struct list_head free_list;
83 #define BLKBACK_INVALID_HANDLE (~0)
86 struct pending_req *pending_reqs;
87 /* List of all 'pending_req' available */
88 struct list_head pending_free;
89 /* And its spinlock. */
90 spinlock_t pending_free_lock;
91 wait_queue_head_t pending_free_wq;
92 /* The list of all pages that are available. */
93 struct page **pending_pages;
94 /* And the grant handles that are available. */
95 grant_handle_t *pending_grant_handles;
98 static struct xen_blkbk *blkbk;
101 * Little helpful macro to figure out the index and virtual address of the
102 * pending_pages[..]. For each 'pending_req' we have have up to
103 * BLKIF_MAX_SEGMENTS_PER_REQUEST (11) pages. The seg would be from 0 through
104 * 10 and would index in the pending_pages[..].
106 static inline int vaddr_pagenr(struct pending_req *req, int seg)
108 return (req - blkbk->pending_reqs) *
109 BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
112 #define pending_page(req, seg) pending_pages[vaddr_pagenr(req, seg)]
114 static inline unsigned long vaddr(struct pending_req *req, int seg)
116 unsigned long pfn = page_to_pfn(blkbk->pending_page(req, seg));
117 return (unsigned long)pfn_to_kaddr(pfn);
120 #define pending_handle(_req, _seg) \
121 (blkbk->pending_grant_handles[vaddr_pagenr(_req, _seg)])
124 static int do_block_io_op(struct xen_blkif *blkif);
125 static int dispatch_rw_block_io(struct xen_blkif *blkif,
126 struct blkif_request *req,
127 struct pending_req *pending_req);
128 static void make_response(struct xen_blkif *blkif, u64 id,
129 unsigned short op, int st);
132 * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
134 static struct pending_req *alloc_req(void)
136 struct pending_req *req = NULL;
139 spin_lock_irqsave(&blkbk->pending_free_lock, flags);
140 if (!list_empty(&blkbk->pending_free)) {
141 req = list_entry(blkbk->pending_free.next, struct pending_req,
143 list_del(&req->free_list);
145 spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
150 * Return the 'pending_req' structure back to the freepool. We also
151 * wake up the thread if it was waiting for a free page.
153 static void free_req(struct pending_req *req)
158 spin_lock_irqsave(&blkbk->pending_free_lock, flags);
159 was_empty = list_empty(&blkbk->pending_free);
160 list_add(&req->free_list, &blkbk->pending_free);
161 spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
163 wake_up(&blkbk->pending_free_wq);
167 * Routines for managing virtual block devices (vbds).
169 static int vbd_translate(struct phys_req *req, struct xen_blkif *blkif,
172 struct vbd *vbd = &blkif->vbd;
175 if ((operation != READ) && vbd->readonly)
178 if (unlikely((req->sector_number + req->nr_sects) > vbd_sz(vbd)))
181 req->dev = vbd->pdevice;
182 req->bdev = vbd->bdev;
189 static void vbd_resize(struct xen_blkif *blkif)
191 struct vbd *vbd = &blkif->vbd;
192 struct xenbus_transaction xbt;
194 struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
195 unsigned long long new_size = vbd_sz(vbd);
197 pr_info(DRV_PFX "VBD Resize: Domid: %d, Device: (%d, %d)\n",
198 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
199 pr_info(DRV_PFX "VBD Resize: new size %llu\n", new_size);
200 vbd->size = new_size;
202 err = xenbus_transaction_start(&xbt);
204 pr_warn(DRV_PFX "Error starting transaction");
207 err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
208 (unsigned long long)vbd_sz(vbd));
210 pr_warn(DRV_PFX "Error writing new size");
214 * Write the current state; we will use this to synchronize
215 * the front-end. If the current state is "connected" the
216 * front-end will get the new size information online.
218 err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
220 pr_warn(DRV_PFX "Error writing the state");
224 err = xenbus_transaction_end(xbt, 0);
228 pr_warn(DRV_PFX "Error ending transaction");
230 xenbus_transaction_end(xbt, 1);
234 * Notification from the guest OS.
236 static void blkif_notify_work(struct xen_blkif *blkif)
238 blkif->waiting_reqs = 1;
242 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
244 blkif_notify_work(dev_id);
249 * SCHEDULER FUNCTIONS
252 static void print_stats(struct xen_blkif *blkif)
254 pr_debug("xen-blkback (%s): oo %3d | rd %4d | wr %4d | f %4d\n",
255 current->comm, blkif->st_oo_req,
256 blkif->st_rd_req, blkif->st_wr_req, blkif->st_f_req);
257 blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
258 blkif->st_rd_req = 0;
259 blkif->st_wr_req = 0;
260 blkif->st_oo_req = 0;
263 int xen_blkif_schedule(void *arg)
265 struct xen_blkif *blkif = arg;
266 struct vbd *vbd = &blkif->vbd;
268 xen_blkif_get(blkif);
270 while (!kthread_should_stop()) {
273 if (unlikely(vbd->size != vbd_sz(vbd)))
276 wait_event_interruptible(
278 blkif->waiting_reqs || kthread_should_stop());
279 wait_event_interruptible(
280 blkbk->pending_free_wq,
281 !list_empty(&blkbk->pending_free) ||
282 kthread_should_stop());
284 blkif->waiting_reqs = 0;
285 smp_mb(); /* clear flag *before* checking for work */
287 if (do_block_io_op(blkif))
288 blkif->waiting_reqs = 1;
290 if (log_stats && time_after(jiffies, blkif->st_print))
297 blkif->xenblkd = NULL;
298 xen_blkif_put(blkif);
308 * Unmap the grant references, and also remove the M2P over-rides
309 * used in the 'pending_req'.
311 static void xen_blkbk_unmap(struct pending_req *req)
313 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
314 unsigned int i, invcount = 0;
315 grant_handle_t handle;
318 for (i = 0; i < req->nr_pages; i++) {
319 handle = pending_handle(req, i);
320 if (handle == BLKBACK_INVALID_HANDLE)
322 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
323 GNTMAP_host_map, handle);
324 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
328 ret = HYPERVISOR_grant_table_op(
329 GNTTABOP_unmap_grant_ref, unmap, invcount);
332 * Note, we use invcount, so nr->pages, so we can't index
333 * using vaddr(req, i).
335 for (i = 0; i < invcount; i++) {
336 ret = m2p_remove_override(
337 virt_to_page(unmap[i].host_addr), false);
339 pr_alert(DRV_PFX "Failed to remove M2P override for %lx\n",
340 (unsigned long)unmap[i].host_addr);
346 static int xen_blkbk_map(struct blkif_request *req,
347 struct pending_req *pending_req,
348 struct seg_buf seg[])
350 struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
352 int nseg = req->nr_segments;
356 * Fill out preq.nr_sects with proper amount of sectors, and setup
357 * assign map[..] with the PFN of the page in our domain with the
358 * corresponding grant reference for each page.
360 for (i = 0; i < nseg; i++) {
363 flags = GNTMAP_host_map;
364 if (pending_req->operation != BLKIF_OP_READ)
365 flags |= GNTMAP_readonly;
366 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
367 req->u.rw.seg[i].gref,
368 pending_req->blkif->domid);
371 ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
375 * Now swizzle the MFN in our domain with the MFN from the other domain
376 * so that when we access vaddr(pending_req,i) it has the contents of
377 * the page from the other domain.
379 for (i = 0; i < nseg; i++) {
380 if (unlikely(map[i].status != 0)) {
381 pr_debug(DRV_PFX "invalid buffer -- could not remap it\n");
382 map[i].handle = BLKBACK_INVALID_HANDLE;
386 pending_handle(pending_req, i) = map[i].handle;
391 ret = m2p_add_override(PFN_DOWN(map[i].dev_bus_addr),
392 blkbk->pending_page(pending_req, i), false);
394 pr_alert(DRV_PFX "Failed to install M2P override for %lx (ret: %d)\n",
395 (unsigned long)map[i].dev_bus_addr, ret);
396 /* We could switch over to GNTTABOP_copy */
400 seg[i].buf = map[i].dev_bus_addr |
401 (req->u.rw.seg[i].first_sect << 9);
407 * Completion callback on the bio's. Called as bh->b_end_io()
410 static void __end_block_io_op(struct pending_req *pending_req, int error)
412 /* An error fails the entire request. */
413 if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
414 (error == -EOPNOTSUPP)) {
415 pr_debug(DRV_PFX "flush diskcache op failed, not supported\n");
416 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->blkif->be, 0);
417 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
419 pr_debug(DRV_PFX "Buffer not up-to-date at end of operation,"
420 " error=%d\n", error);
421 pending_req->status = BLKIF_RSP_ERROR;
425 * If all of the bio's have completed it is time to unmap
426 * the grant references associated with 'request' and provide
427 * the proper response on the ring.
429 if (atomic_dec_and_test(&pending_req->pendcnt)) {
430 xen_blkbk_unmap(pending_req);
431 make_response(pending_req->blkif, pending_req->id,
432 pending_req->operation, pending_req->status);
433 xen_blkif_put(pending_req->blkif);
434 free_req(pending_req);
441 static void end_block_io_op(struct bio *bio, int error)
443 __end_block_io_op(bio->bi_private, error);
450 * Function to copy the from the ring buffer the 'struct blkif_request'
451 * (which has the sectors we want, number of them, grant references, etc),
452 * and transmute it to the block API to hand it over to the proper block disk.
454 static int do_block_io_op(struct xen_blkif *blkif)
456 union blkif_back_rings *blk_rings = &blkif->blk_rings;
457 struct blkif_request req;
458 struct pending_req *pending_req;
462 rc = blk_rings->common.req_cons;
463 rp = blk_rings->common.sring->req_prod;
464 rmb(); /* Ensure we see queued requests up to 'rp'. */
468 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
471 if (kthread_should_stop()) {
476 pending_req = alloc_req();
477 if (NULL == pending_req) {
483 switch (blkif->blk_protocol) {
484 case BLKIF_PROTOCOL_NATIVE:
485 memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
487 case BLKIF_PROTOCOL_X86_32:
488 blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
490 case BLKIF_PROTOCOL_X86_64:
491 blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
496 blk_rings->common.req_cons = ++rc; /* before make_response() */
498 /* Apply all sanity checks to /private copy/ of request. */
501 if (dispatch_rw_block_io(blkif, &req, pending_req))
504 /* Yield point for this unbounded loop. */
512 * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
513 * and call the 'submit_bio' to pass it to the underlying storage.
515 static int dispatch_rw_block_io(struct xen_blkif *blkif,
516 struct blkif_request *req,
517 struct pending_req *pending_req)
519 struct phys_req preq;
520 struct seg_buf seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
522 struct bio *bio = NULL;
523 struct bio *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
526 struct blk_plug plug;
528 switch (req->operation) {
535 operation = WRITE_ODIRECT;
537 case BLKIF_OP_FLUSH_DISKCACHE:
539 operation = WRITE_FLUSH;
541 * The frontend likes to set this to -1, which vbd_translate
544 req->u.rw.sector_number = 0;
546 case BLKIF_OP_WRITE_BARRIER:
548 operation = 0; /* make gcc happy */
553 /* Check that the number of segments is sane. */
554 nseg = req->nr_segments;
555 if (unlikely(nseg == 0 && operation != WRITE_FLUSH) ||
556 unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
557 pr_debug(DRV_PFX "Bad number of segments in request (%d)\n",
559 /* Haven't submitted any bio's yet. */
563 preq.dev = req->handle;
564 preq.sector_number = req->u.rw.sector_number;
567 pending_req->blkif = blkif;
568 pending_req->id = req->id;
569 pending_req->operation = req->operation;
570 pending_req->status = BLKIF_RSP_OKAY;
571 pending_req->nr_pages = nseg;
573 for (i = 0; i < nseg; i++) {
574 seg[i].nsec = req->u.rw.seg[i].last_sect -
575 req->u.rw.seg[i].first_sect + 1;
576 if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
577 (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
579 preq.nr_sects += seg[i].nsec;
583 if (vbd_translate(&preq, blkif, operation) != 0) {
584 pr_debug(DRV_PFX "access denied: %s of [%llu,%llu] on dev=%04x\n",
585 operation == READ ? "read" : "write",
587 preq.sector_number + preq.nr_sects, preq.dev);
592 * This check _MUST_ be done after vbd_translate as the preq.bdev
595 for (i = 0; i < nseg; i++) {
596 if (((int)preq.sector_number|(int)seg[i].nsec) &
597 ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
598 pr_debug(DRV_PFX "Misaligned I/O request from domain %d",
605 * If we have failed at this point, we need to undo the M2P override,
606 * set gnttab_set_unmap_op on all of the grant references and perform
607 * the hypercall to unmap the grants - that is all done in
610 if (xen_blkbk_map(req, pending_req, seg))
613 /* This corresponding xen_blkif_put is done in __end_block_io_op */
614 xen_blkif_get(blkif);
616 for (i = 0; i < nseg; i++) {
617 while ((bio == NULL) ||
619 blkbk->pending_page(pending_req, i),
621 seg[i].buf & ~PAGE_MASK) == 0)) {
623 bio = bio_alloc(GFP_KERNEL, nseg-i);
624 if (unlikely(bio == NULL))
627 biolist[nbio++] = bio;
628 bio->bi_bdev = preq.bdev;
629 bio->bi_private = pending_req;
630 bio->bi_end_io = end_block_io_op;
631 bio->bi_sector = preq.sector_number;
634 preq.sector_number += seg[i].nsec;
637 /* This will be hit if the operation was a flush. */
639 BUG_ON(operation != WRITE_FLUSH);
641 bio = bio_alloc(GFP_KERNEL, 0);
642 if (unlikely(bio == NULL))
645 biolist[nbio++] = bio;
646 bio->bi_bdev = preq.bdev;
647 bio->bi_private = pending_req;
648 bio->bi_end_io = end_block_io_op;
652 * We set it one so that the last submit_bio does not have to call
655 atomic_set(&pending_req->pendcnt, nbio);
657 /* Get a reference count for the disk queue and start sending I/O */
658 blk_start_plug(&plug);
660 for (i = 0; i < nbio; i++)
661 submit_bio(operation, biolist[i]);
663 /* Let the I/Os go.. */
664 blk_finish_plug(&plug);
666 if (operation == READ)
667 blkif->st_rd_sect += preq.nr_sects;
668 else if (operation == WRITE || operation == WRITE_FLUSH)
669 blkif->st_wr_sect += preq.nr_sects;
674 xen_blkbk_unmap(pending_req);
676 /* Haven't submitted any bio's yet. */
677 make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
678 free_req(pending_req);
679 msleep(1); /* back off a bit */
683 for (i = 0; i < nbio; i++)
685 __end_block_io_op(pending_req, -EINVAL);
686 msleep(1); /* back off a bit */
693 * Put a response on the ring on how the operation fared.
695 static void make_response(struct xen_blkif *blkif, u64 id,
696 unsigned short op, int st)
698 struct blkif_response resp;
700 union blkif_back_rings *blk_rings = &blkif->blk_rings;
708 spin_lock_irqsave(&blkif->blk_ring_lock, flags);
709 /* Place on the response ring for the relevant domain. */
710 switch (blkif->blk_protocol) {
711 case BLKIF_PROTOCOL_NATIVE:
712 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
713 &resp, sizeof(resp));
715 case BLKIF_PROTOCOL_X86_32:
716 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
717 &resp, sizeof(resp));
719 case BLKIF_PROTOCOL_X86_64:
720 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
721 &resp, sizeof(resp));
726 blk_rings->common.rsp_prod_pvt++;
727 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
728 if (blk_rings->common.rsp_prod_pvt == blk_rings->common.req_cons) {
730 * Tail check for pending requests. Allows frontend to avoid
731 * notifications if requests are already in flight (lower
732 * overheads and promotes batching).
734 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
736 } else if (RING_HAS_UNCONSUMED_REQUESTS(&blk_rings->common)) {
740 spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
743 blkif_notify_work(blkif);
745 notify_remote_via_irq(blkif->irq);
748 static int __init xen_blkif_init(void)
753 if (!xen_pv_domain())
756 blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
758 pr_alert(DRV_PFX "%s: out of memory!\n", __func__);
762 mmap_pages = xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
764 blkbk->pending_reqs = kmalloc(sizeof(blkbk->pending_reqs[0]) *
765 xen_blkif_reqs, GFP_KERNEL);
766 blkbk->pending_grant_handles = kzalloc(sizeof(blkbk->pending_grant_handles[0]) *
767 mmap_pages, GFP_KERNEL);
768 blkbk->pending_pages = kzalloc(sizeof(blkbk->pending_pages[0]) *
769 mmap_pages, GFP_KERNEL);
771 if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
772 !blkbk->pending_pages) {
777 for (i = 0; i < mmap_pages; i++) {
778 blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
779 blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
780 if (blkbk->pending_pages[i] == NULL) {
785 rc = xen_blkif_interface_init();
789 memset(blkbk->pending_reqs, 0, sizeof(blkbk->pending_reqs));
791 INIT_LIST_HEAD(&blkbk->pending_free);
792 spin_lock_init(&blkbk->pending_free_lock);
793 init_waitqueue_head(&blkbk->pending_free_wq);
795 for (i = 0; i < xen_blkif_reqs; i++)
796 list_add_tail(&blkbk->pending_reqs[i].free_list,
797 &blkbk->pending_free);
799 rc = xen_blkif_xenbus_init();
806 pr_alert(DRV_PFX "%s: out of memory\n", __func__);
808 kfree(blkbk->pending_reqs);
809 kfree(blkbk->pending_grant_handles);
810 for (i = 0; i < mmap_pages; i++) {
811 if (blkbk->pending_pages[i])
812 __free_page(blkbk->pending_pages[i]);
814 kfree(blkbk->pending_pages);
820 module_init(xen_blkif_init);
822 MODULE_LICENSE("Dual BSD/GPL");