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 static unsigned int debug_lvl;
66 module_param(log_stats, int, 0644);
67 module_param(debug_lvl, int, 0644);
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.
76 struct blkif_st *blkif;
80 unsigned short operation;
82 struct list_head free_list;
85 #define BLKBACK_INVALID_HANDLE (~0)
88 struct pending_req *pending_reqs;
89 /* List of all 'pending_req' available */
90 struct list_head pending_free;
91 /* And its spinlock. */
92 spinlock_t pending_free_lock;
93 wait_queue_head_t pending_free_wq;
94 /* The list of all pages that are available. */
95 struct page **pending_pages;
96 /* And the grant handles that are available. */
97 grant_handle_t *pending_grant_handles;
100 static struct xen_blkbk *blkbk;
103 * Little helpful macro to figure out the index and virtual address of the
104 * pending_pages[..]. For each 'pending_req' we have have up to
105 * BLKIF_MAX_SEGMENTS_PER_REQUEST (11) pages. The seg would be from 0 through
106 * 10 and would index in the pending_pages[..].
108 static inline int vaddr_pagenr(struct pending_req *req, int seg)
110 return (req - blkbk->pending_reqs) *
111 BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
114 #define pending_page(req, seg) pending_pages[vaddr_pagenr(req, seg)]
116 static inline unsigned long vaddr(struct pending_req *req, int seg)
118 unsigned long pfn = page_to_pfn(blkbk->pending_page(req, seg));
119 return (unsigned long)pfn_to_kaddr(pfn);
122 #define pending_handle(_req, _seg) \
123 (blkbk->pending_grant_handles[vaddr_pagenr(_req, _seg)])
126 static int do_block_io_op(struct blkif_st *blkif);
127 static int dispatch_rw_block_io(struct blkif_st *blkif,
128 struct blkif_request *req,
129 struct pending_req *pending_req);
130 static void make_response(struct blkif_st *blkif, u64 id,
131 unsigned short op, int st);
134 * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
136 static struct pending_req *alloc_req(void)
138 struct pending_req *req = NULL;
141 spin_lock_irqsave(&blkbk->pending_free_lock, flags);
142 if (!list_empty(&blkbk->pending_free)) {
143 req = list_entry(blkbk->pending_free.next, struct pending_req,
145 list_del(&req->free_list);
147 spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
152 * Return the 'pending_req' structure back to the freepool. We also
153 * wake up the thread if it was waiting for a free page.
155 static void free_req(struct pending_req *req)
160 spin_lock_irqsave(&blkbk->pending_free_lock, flags);
161 was_empty = list_empty(&blkbk->pending_free);
162 list_add(&req->free_list, &blkbk->pending_free);
163 spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
165 wake_up(&blkbk->pending_free_wq);
169 * Routines for managing virtual block devices (vbds).
171 static int vbd_translate(struct phys_req *req, struct blkif_st *blkif,
174 struct vbd *vbd = &blkif->vbd;
177 if ((operation != READ) && vbd->readonly)
180 if (unlikely((req->sector_number + req->nr_sects) > vbd_sz(vbd)))
183 req->dev = vbd->pdevice;
184 req->bdev = vbd->bdev;
191 static void vbd_resize(struct blkif_st *blkif)
193 struct vbd *vbd = &blkif->vbd;
194 struct xenbus_transaction xbt;
196 struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
197 unsigned long long new_size = vbd_sz(vbd);
199 printk(KERN_INFO "VBD Resize: Domid: %d, Device: (%d, %d)\n",
200 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
201 printk(KERN_INFO "VBD Resize: new size %llu\n", new_size);
202 vbd->size = new_size;
204 err = xenbus_transaction_start(&xbt);
206 printk(KERN_WARNING "Error starting transaction");
209 err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
210 (unsigned long long)vbd_sz(vbd));
212 printk(KERN_WARNING "Error writing new size");
216 * Write the current state; we will use this to synchronize
217 * the front-end. If the current state is "connected" the
218 * front-end will get the new size information online.
220 err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
222 printk(KERN_WARNING "Error writing the state");
226 err = xenbus_transaction_end(xbt, 0);
230 printk(KERN_WARNING "Error ending transaction");
232 xenbus_transaction_end(xbt, 1);
236 * Notification from the guest OS.
238 static void blkif_notify_work(struct blkif_st *blkif)
240 blkif->waiting_reqs = 1;
244 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
246 blkif_notify_work(dev_id);
251 * SCHEDULER FUNCTIONS
254 static void print_stats(struct blkif_st *blkif)
256 printk(KERN_DEBUG "%s: oo %3d | rd %4d | wr %4d | f %4d\n",
257 current->comm, blkif->st_oo_req,
258 blkif->st_rd_req, blkif->st_wr_req, blkif->st_f_req);
259 blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
260 blkif->st_rd_req = 0;
261 blkif->st_wr_req = 0;
262 blkif->st_oo_req = 0;
265 int xen_blkif_schedule(void *arg)
267 struct blkif_st *blkif = arg;
268 struct vbd *vbd = &blkif->vbd;
270 xen_blkif_get(blkif);
273 printk(KERN_DEBUG "%s: started\n", current->comm);
275 while (!kthread_should_stop()) {
278 if (unlikely(vbd->size != vbd_sz(vbd)))
281 wait_event_interruptible(
283 blkif->waiting_reqs || kthread_should_stop());
284 wait_event_interruptible(
285 blkbk->pending_free_wq,
286 !list_empty(&blkbk->pending_free) ||
287 kthread_should_stop());
289 blkif->waiting_reqs = 0;
290 smp_mb(); /* clear flag *before* checking for work */
292 if (do_block_io_op(blkif))
293 blkif->waiting_reqs = 1;
295 if (log_stats && time_after(jiffies, blkif->st_print))
302 printk(KERN_DEBUG "%s: exiting\n", current->comm);
304 blkif->xenblkd = NULL;
305 xen_blkif_put(blkif);
315 * Unmap the grant references, and also remove the M2P over-rides
316 * used in the 'pending_req'.
318 static void xen_blkbk_unmap(struct pending_req *req)
320 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
321 unsigned int i, invcount = 0;
322 grant_handle_t handle;
325 for (i = 0; i < req->nr_pages; i++) {
326 handle = pending_handle(req, i);
327 if (handle == BLKBACK_INVALID_HANDLE)
329 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
330 GNTMAP_host_map, handle);
331 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
335 ret = HYPERVISOR_grant_table_op(
336 GNTTABOP_unmap_grant_ref, unmap, invcount);
339 * Note, we use invcount, so nr->pages, so we can't index
340 * using vaddr(req, i).
342 for (i = 0; i < invcount; i++) {
343 ret = m2p_remove_override(
344 virt_to_page(unmap[i].host_addr), false);
346 printk(KERN_ALERT "Failed to remove M2P override for %lx\n",
347 (unsigned long)unmap[i].host_addr);
353 static int xen_blkbk_map(struct blkif_request *req,
354 struct pending_req *pending_req,
355 struct seg_buf seg[])
357 struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
359 int nseg = req->nr_segments;
363 * Fill out preq.nr_sects with proper amount of sectors, and setup
364 * assign map[..] with the PFN of the page in our domain with the
365 * corresponding grant reference for each page.
367 for (i = 0; i < nseg; i++) {
370 flags = GNTMAP_host_map;
371 if (pending_req->operation != BLKIF_OP_READ)
372 flags |= GNTMAP_readonly;
373 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
374 req->u.rw.seg[i].gref,
375 pending_req->blkif->domid);
378 ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
382 * Now swizzle the MFN in our domain with the MFN from the other domain
383 * so that when we access vaddr(pending_req,i) it has the contents of
384 * the page from the other domain.
386 for (i = 0; i < nseg; i++) {
387 if (unlikely(map[i].status != 0)) {
388 DPRINTK("invalid buffer -- could not remap it\n");
389 map[i].handle = BLKBACK_INVALID_HANDLE;
393 pending_handle(pending_req, i) = map[i].handle;
398 ret = m2p_add_override(PFN_DOWN(map[i].dev_bus_addr),
399 blkbk->pending_page(pending_req, i), false);
401 printk(KERN_ALERT "Failed to install M2P override for"\
402 " %lx (ret: %d)\n", (unsigned long)
403 map[i].dev_bus_addr, ret);
404 /* We could switch over to GNTTABOP_copy */
408 seg[i].buf = map[i].dev_bus_addr |
409 (req->u.rw.seg[i].first_sect << 9);
415 * Completion callback on the bio's. Called as bh->b_end_io()
418 static void __end_block_io_op(struct pending_req *pending_req, int error)
420 /* An error fails the entire request. */
421 if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
422 (error == -EOPNOTSUPP)) {
423 DPRINTK("blkback: flush diskcache op failed, not supported\n");
424 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->blkif->be, 0);
425 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
427 DPRINTK("Buffer not up-to-date at end of operation, "
428 "error=%d\n", error);
429 pending_req->status = BLKIF_RSP_ERROR;
433 * If all of the bio's have completed it is time to unmap
434 * the grant references associated with 'request' and provide
435 * the proper response on the ring.
437 if (atomic_dec_and_test(&pending_req->pendcnt)) {
438 xen_blkbk_unmap(pending_req);
439 make_response(pending_req->blkif, pending_req->id,
440 pending_req->operation, pending_req->status);
441 xen_blkif_put(pending_req->blkif);
442 free_req(pending_req);
449 static void end_block_io_op(struct bio *bio, int error)
451 __end_block_io_op(bio->bi_private, error);
458 * Function to copy the from the ring buffer the 'struct blkif_request'
459 * (which has the sectors we want, number of them, grant references, etc),
460 * and transmute it to the block API to hand it over to the proper block disk.
462 static int do_block_io_op(struct blkif_st *blkif)
464 union blkif_back_rings *blk_rings = &blkif->blk_rings;
465 struct blkif_request req;
466 struct pending_req *pending_req;
470 rc = blk_rings->common.req_cons;
471 rp = blk_rings->common.sring->req_prod;
472 rmb(); /* Ensure we see queued requests up to 'rp'. */
476 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
479 if (kthread_should_stop()) {
484 pending_req = alloc_req();
485 if (NULL == pending_req) {
491 switch (blkif->blk_protocol) {
492 case BLKIF_PROTOCOL_NATIVE:
493 memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
495 case BLKIF_PROTOCOL_X86_32:
496 blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
498 case BLKIF_PROTOCOL_X86_64:
499 blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
504 blk_rings->common.req_cons = ++rc; /* before make_response() */
506 /* Apply all sanity checks to /private copy/ of request. */
509 if (dispatch_rw_block_io(blkif, &req, pending_req))
512 /* Yield point for this unbounded loop. */
520 * Transmutation of the 'struct blkif_request' to a proper 'struct bio'
521 * and call the 'submit_bio' to pass it to the underlying storage.
523 static int dispatch_rw_block_io(struct blkif_st *blkif,
524 struct blkif_request *req,
525 struct pending_req *pending_req)
527 struct phys_req preq;
528 struct seg_buf seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
530 struct bio *bio = NULL;
531 struct bio *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
534 struct blk_plug plug;
536 switch (req->operation) {
543 operation = WRITE_ODIRECT;
545 case BLKIF_OP_FLUSH_DISKCACHE:
547 operation = WRITE_FLUSH;
549 * The frontend likes to set this to -1, which vbd_translate
552 req->u.rw.sector_number = 0;
554 case BLKIF_OP_WRITE_BARRIER:
556 operation = 0; /* make gcc happy */
561 /* Check that the number of segments is sane. */
562 nseg = req->nr_segments;
563 if (unlikely(nseg == 0 && operation != WRITE_FLUSH) ||
564 unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
565 DPRINTK("Bad number of segments in request (%d)\n", nseg);
566 /* Haven't submitted any bio's yet. */
570 preq.dev = req->handle;
571 preq.sector_number = req->u.rw.sector_number;
574 pending_req->blkif = blkif;
575 pending_req->id = req->id;
576 pending_req->operation = req->operation;
577 pending_req->status = BLKIF_RSP_OKAY;
578 pending_req->nr_pages = nseg;
580 for (i = 0; i < nseg; i++) {
581 seg[i].nsec = req->u.rw.seg[i].last_sect -
582 req->u.rw.seg[i].first_sect + 1;
583 if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
584 (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
586 preq.nr_sects += seg[i].nsec;
590 if (vbd_translate(&preq, blkif, operation) != 0) {
591 DPRINTK("access denied: %s of [%llu,%llu] on dev=%04x\n",
592 operation == READ ? "read" : "write",
594 preq.sector_number + preq.nr_sects, preq.dev);
599 * This check _MUST_ be done after vbd_translate as the preq.bdev
602 for (i = 0; i < nseg; i++) {
603 if (((int)preq.sector_number|(int)seg[i].nsec) &
604 ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
605 DPRINTK("Misaligned I/O request from domain %d",
612 * If we have failed at this point, we need to undo the M2P override,
613 * set gnttab_set_unmap_op on all of the grant references and perform
614 * the hypercall to unmap the grants - that is all done in
617 if (xen_blkbk_map(req, pending_req, seg))
620 /* This corresponding xen_blkif_put is done in __end_block_io_op */
621 xen_blkif_get(blkif);
623 for (i = 0; i < nseg; i++) {
624 while ((bio == NULL) ||
626 blkbk->pending_page(pending_req, i),
628 seg[i].buf & ~PAGE_MASK) == 0)) {
630 bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, nseg-i);
631 if (unlikely(bio == NULL))
634 bio->bi_bdev = preq.bdev;
635 bio->bi_private = pending_req;
636 bio->bi_end_io = end_block_io_op;
637 bio->bi_sector = preq.sector_number;
640 preq.sector_number += seg[i].nsec;
643 /* This will be hit if the operation was a flush. */
645 BUG_ON(operation != WRITE_FLUSH);
646 bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, 0);
647 if (unlikely(bio == NULL))
650 bio->bi_bdev = preq.bdev;
651 bio->bi_private = pending_req;
652 bio->bi_end_io = end_block_io_op;
656 * We set it one so that the last submit_bio does not have to call
659 atomic_set(&pending_req->pendcnt, nbio);
661 /* Get a reference count for the disk queue and start sending I/O */
662 blk_start_plug(&plug);
664 for (i = 0; i < nbio; i++)
665 submit_bio(operation, biolist[i]);
667 /* Let the I/Os go.. */
668 blk_finish_plug(&plug);
670 if (operation == READ)
671 blkif->st_rd_sect += preq.nr_sects;
672 else if (operation == WRITE || operation == WRITE_FLUSH)
673 blkif->st_wr_sect += preq.nr_sects;
678 xen_blkbk_unmap(pending_req);
680 /* Haven't submitted any bio's yet. */
681 make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
682 free_req(pending_req);
683 msleep(1); /* back off a bit */
687 for (i = 0; i < (nbio-1); i++)
689 __end_block_io_op(pending_req, -EINVAL);
690 msleep(1); /* back off a bit */
697 * Put a response on the ring on how the operation fared.
699 static void make_response(struct blkif_st *blkif, u64 id,
700 unsigned short op, int st)
702 struct blkif_response resp;
704 union blkif_back_rings *blk_rings = &blkif->blk_rings;
712 spin_lock_irqsave(&blkif->blk_ring_lock, flags);
713 /* Place on the response ring for the relevant domain. */
714 switch (blkif->blk_protocol) {
715 case BLKIF_PROTOCOL_NATIVE:
716 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
717 &resp, sizeof(resp));
719 case BLKIF_PROTOCOL_X86_32:
720 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
721 &resp, sizeof(resp));
723 case BLKIF_PROTOCOL_X86_64:
724 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
725 &resp, sizeof(resp));
730 blk_rings->common.rsp_prod_pvt++;
731 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
732 if (blk_rings->common.rsp_prod_pvt == blk_rings->common.req_cons) {
734 * Tail check for pending requests. Allows frontend to avoid
735 * notifications if requests are already in flight (lower
736 * overheads and promotes batching).
738 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
740 } else if (RING_HAS_UNCONSUMED_REQUESTS(&blk_rings->common)) {
744 spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
747 blkif_notify_work(blkif);
749 notify_remote_via_irq(blkif->irq);
752 static int __init xen_blkif_init(void)
757 if (!xen_pv_domain())
760 blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
762 printk(KERN_ALERT "%s: out of memory!\n", __func__);
766 mmap_pages = xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
768 blkbk->pending_reqs = kmalloc(sizeof(blkbk->pending_reqs[0]) *
769 xen_blkif_reqs, GFP_KERNEL);
770 blkbk->pending_grant_handles = kzalloc(sizeof(blkbk->pending_grant_handles[0]) *
771 mmap_pages, GFP_KERNEL);
772 blkbk->pending_pages = kzalloc(sizeof(blkbk->pending_pages[0]) *
773 mmap_pages, GFP_KERNEL);
775 if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
776 !blkbk->pending_pages) {
781 for (i = 0; i < mmap_pages; i++) {
782 blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
783 blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
784 if (blkbk->pending_pages[i] == NULL) {
789 rc = xen_blkif_interface_init();
793 memset(blkbk->pending_reqs, 0, sizeof(blkbk->pending_reqs));
795 INIT_LIST_HEAD(&blkbk->pending_free);
796 spin_lock_init(&blkbk->pending_free_lock);
797 init_waitqueue_head(&blkbk->pending_free_wq);
799 for (i = 0; i < xen_blkif_reqs; i++)
800 list_add_tail(&blkbk->pending_reqs[i].free_list,
801 &blkbk->pending_free);
803 rc = xen_blkif_xenbus_init();
810 printk(KERN_ERR "%s: out of memory\n", __func__);
812 kfree(blkbk->pending_reqs);
813 kfree(blkbk->pending_grant_handles);
814 for (i = 0; i < mmap_pages; i++) {
815 if (blkbk->pending_pages[i])
816 __free_page(blkbk->pending_pages[i]);
818 kfree(blkbk->pending_pages);
824 module_init(xen_blkif_init);
826 MODULE_LICENSE("Dual BSD/GPL");