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[..]. */
107 static inline int vaddr_pagenr(struct pending_req *req, int seg)
109 return (req - blkbk->pending_reqs) *
110 BLKIF_MAX_SEGMENTS_PER_REQUEST + seg;
113 #define pending_page(req, seg) pending_pages[vaddr_pagenr(req, seg)]
115 static inline unsigned long vaddr(struct pending_req *req, int seg)
117 unsigned long pfn = page_to_pfn(blkbk->pending_page(req, seg));
118 return (unsigned long)pfn_to_kaddr(pfn);
121 #define pending_handle(_req, _seg) \
122 (blkbk->pending_grant_handles[vaddr_pagenr(_req, _seg)])
125 static int do_block_io_op(struct blkif_st *blkif);
126 static int dispatch_rw_block_io(struct blkif_st *blkif,
127 struct blkif_request *req,
128 struct pending_req *pending_req);
129 static void make_response(struct blkif_st *blkif, u64 id,
130 unsigned short op, int st);
133 * Retrieve from the 'pending_reqs' a free pending_req structure to be used.
135 static struct pending_req *alloc_req(void)
137 struct pending_req *req = NULL;
140 spin_lock_irqsave(&blkbk->pending_free_lock, flags);
141 if (!list_empty(&blkbk->pending_free)) {
142 req = list_entry(blkbk->pending_free.next, struct pending_req,
144 list_del(&req->free_list);
146 spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
151 * Return the 'pending_req' structure back to the freepool. We also
152 * wake up the thread if it was waiting for a free page.
154 static void free_req(struct pending_req *req)
159 spin_lock_irqsave(&blkbk->pending_free_lock, flags);
160 was_empty = list_empty(&blkbk->pending_free);
161 list_add(&req->free_list, &blkbk->pending_free);
162 spin_unlock_irqrestore(&blkbk->pending_free_lock, flags);
164 wake_up(&blkbk->pending_free_wq);
168 * Routines for managing virtual block devices (vbds).
172 static int vbd_translate(struct phys_req *req, struct blkif_st *blkif,
175 struct vbd *vbd = &blkif->vbd;
178 if ((operation != READ) && vbd->readonly)
181 if (unlikely((req->sector_number + req->nr_sects) > vbd_sz(vbd)))
184 req->dev = vbd->pdevice;
185 req->bdev = vbd->bdev;
192 static void vbd_resize(struct blkif_st *blkif)
194 struct vbd *vbd = &blkif->vbd;
195 struct xenbus_transaction xbt;
197 struct xenbus_device *dev = xen_blkbk_xenbus(blkif->be);
198 unsigned long long new_size = vbd_sz(vbd);
200 printk(KERN_INFO "VBD Resize: Domid: %d, Device: (%d, %d)\n",
201 blkif->domid, MAJOR(vbd->pdevice), MINOR(vbd->pdevice));
202 printk(KERN_INFO "VBD Resize: new size %llu\n", new_size);
203 vbd->size = new_size;
205 err = xenbus_transaction_start(&xbt);
207 printk(KERN_WARNING "Error starting transaction");
210 err = xenbus_printf(xbt, dev->nodename, "sectors", "%llu",
211 (unsigned long long)vbd_sz(vbd));
213 printk(KERN_WARNING "Error writing new size");
217 * Write the current state; we will use this to synchronize
218 * the front-end. If the current state is "connected" the
219 * front-end will get the new size information online.
221 err = xenbus_printf(xbt, dev->nodename, "state", "%d", dev->state);
223 printk(KERN_WARNING "Error writing the state");
227 err = xenbus_transaction_end(xbt, 0);
231 printk(KERN_WARNING "Error ending transaction");
233 xenbus_transaction_end(xbt, 1);
237 * Notification from the guest OS.
239 static void blkif_notify_work(struct blkif_st *blkif)
241 blkif->waiting_reqs = 1;
245 irqreturn_t xen_blkif_be_int(int irq, void *dev_id)
247 blkif_notify_work(dev_id);
252 * SCHEDULER FUNCTIONS
255 static void print_stats(struct blkif_st *blkif)
257 printk(KERN_DEBUG "%s: oo %3d | rd %4d | wr %4d | f %4d\n",
258 current->comm, blkif->st_oo_req,
259 blkif->st_rd_req, blkif->st_wr_req, blkif->st_f_req);
260 blkif->st_print = jiffies + msecs_to_jiffies(10 * 1000);
261 blkif->st_rd_req = 0;
262 blkif->st_wr_req = 0;
263 blkif->st_oo_req = 0;
266 int xen_blkif_schedule(void *arg)
268 struct blkif_st *blkif = arg;
269 struct vbd *vbd = &blkif->vbd;
271 xen_blkif_get(blkif);
274 printk(KERN_DEBUG "%s: started\n", current->comm);
276 while (!kthread_should_stop()) {
279 if (unlikely(vbd->size != vbd_sz(vbd)))
282 wait_event_interruptible(
284 blkif->waiting_reqs || kthread_should_stop());
285 wait_event_interruptible(
286 blkbk->pending_free_wq,
287 !list_empty(&blkbk->pending_free) ||
288 kthread_should_stop());
290 blkif->waiting_reqs = 0;
291 smp_mb(); /* clear flag *before* checking for work */
293 if (do_block_io_op(blkif))
294 blkif->waiting_reqs = 1;
296 if (log_stats && time_after(jiffies, blkif->st_print))
303 printk(KERN_DEBUG "%s: exiting\n", current->comm);
305 blkif->xenblkd = NULL;
306 xen_blkif_put(blkif);
316 * Unmap the grant references, and also remove the M2P over-rides
317 * used in the 'pending_req'.
319 static void xen_blkbk_unmap(struct pending_req *req)
321 struct gnttab_unmap_grant_ref unmap[BLKIF_MAX_SEGMENTS_PER_REQUEST];
322 unsigned int i, invcount = 0;
323 grant_handle_t handle;
326 for (i = 0; i < req->nr_pages; i++) {
327 handle = pending_handle(req, i);
328 if (handle == BLKBACK_INVALID_HANDLE)
330 gnttab_set_unmap_op(&unmap[invcount], vaddr(req, i),
331 GNTMAP_host_map, handle);
332 pending_handle(req, i) = BLKBACK_INVALID_HANDLE;
336 ret = HYPERVISOR_grant_table_op(
337 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 " \
347 "%lx\n", (unsigned long)unmap[i].host_addr);
352 static int xen_blkbk_map(struct blkif_request *req, struct pending_req *pending_req,
353 struct seg_buf seg[])
355 struct gnttab_map_grant_ref map[BLKIF_MAX_SEGMENTS_PER_REQUEST];
357 int nseg = req->nr_segments;
359 /* Fill out preq.nr_sects with proper amount of sectors, and setup
360 * assign map[..] with the PFN of the page in our domain with the
361 * corresponding grant reference for each page.
363 for (i = 0; i < nseg; i++) {
366 flags = GNTMAP_host_map;
367 if (pending_req->operation != BLKIF_OP_READ)
368 flags |= GNTMAP_readonly;
369 gnttab_set_map_op(&map[i], vaddr(pending_req, i), flags,
370 req->u.rw.seg[i].gref, pending_req->blkif->domid);
373 ret = HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, map, nseg);
376 /* Now swizzel the MFN in our domain with the MFN from the other domain
377 * so that when we access vaddr(pending_req,i) it has the contents of
378 * the page from the other domain.
380 for (i = 0; i < nseg; i++) {
381 if (unlikely(map[i].status != 0)) {
382 DPRINTK("invalid buffer -- could not remap it\n");
383 map[i].handle = BLKBACK_INVALID_HANDLE;
387 pending_handle(pending_req, i) = map[i].handle;
392 ret = m2p_add_override(PFN_DOWN(map[i].dev_bus_addr),
393 blkbk->pending_page(pending_req, i), false);
395 printk(KERN_ALERT "Failed to install M2P override for"\
396 " %lx (ret: %d)\n", (unsigned long)
397 map[i].dev_bus_addr, ret);
398 /* We could switch over to GNTTABOP_copy */
402 seg[i].buf = map[i].dev_bus_addr |
403 (req->u.rw.seg[i].first_sect << 9);
409 * Completion callback on the bio's. Called as bh->b_end_io()
412 static void __end_block_io_op(struct pending_req *pending_req, int error)
414 /* An error fails the entire request. */
415 if ((pending_req->operation == BLKIF_OP_FLUSH_DISKCACHE) &&
416 (error == -EOPNOTSUPP)) {
417 DPRINTK("blkback: flush diskcache op failed, not supported\n");
418 xen_blkbk_flush_diskcache(XBT_NIL, pending_req->blkif->be, 0);
419 pending_req->status = BLKIF_RSP_EOPNOTSUPP;
421 DPRINTK("Buffer not up-to-date at end of operation, "
422 "error=%d\n", error);
423 pending_req->status = BLKIF_RSP_ERROR;
426 /* If all of the bio's have completed it is time to unmap
427 * the grant references associated with 'request' and provide
428 * the proper response on the ring.
430 if (atomic_dec_and_test(&pending_req->pendcnt)) {
431 xen_blkbk_unmap(pending_req);
432 make_response(pending_req->blkif, pending_req->id,
433 pending_req->operation, pending_req->status);
434 xen_blkif_put(pending_req->blkif);
435 free_req(pending_req);
442 static void end_block_io_op(struct bio *bio, int error)
444 __end_block_io_op(bio->bi_private, error);
451 * Function to copy the from the ring buffer the 'struct blkif_request'
452 * (which has the sectors we want, number of them, grant references, etc),
453 * and transmute it to the block API to hand it over to the proper block disk.
455 static int do_block_io_op(struct blkif_st *blkif)
457 union blkif_back_rings *blk_rings = &blkif->blk_rings;
458 struct blkif_request req;
459 struct pending_req *pending_req;
463 rc = blk_rings->common.req_cons;
464 rp = blk_rings->common.sring->req_prod;
465 rmb(); /* Ensure we see queued requests up to 'rp'. */
469 if (RING_REQUEST_CONS_OVERFLOW(&blk_rings->common, rc))
472 if (kthread_should_stop()) {
477 pending_req = alloc_req();
478 if (NULL == pending_req) {
484 switch (blkif->blk_protocol) {
485 case BLKIF_PROTOCOL_NATIVE:
486 memcpy(&req, RING_GET_REQUEST(&blk_rings->native, rc), sizeof(req));
488 case BLKIF_PROTOCOL_X86_32:
489 blkif_get_x86_32_req(&req, RING_GET_REQUEST(&blk_rings->x86_32, rc));
491 case BLKIF_PROTOCOL_X86_64:
492 blkif_get_x86_64_req(&req, RING_GET_REQUEST(&blk_rings->x86_64, rc));
497 blk_rings->common.req_cons = ++rc; /* before make_response() */
499 /* Apply all sanity checks to /private copy/ of request. */
502 if (dispatch_rw_block_io(blkif, &req, pending_req))
505 /* Yield point for this unbounded loop. */
513 * Transumation of the 'struct blkif_request' to a proper 'struct bio'
514 * and call the 'submit_bio' to pass it to the underlaying storage.
516 static int dispatch_rw_block_io(struct blkif_st *blkif,
517 struct blkif_request *req,
518 struct pending_req *pending_req)
520 struct phys_req preq;
521 struct seg_buf seg[BLKIF_MAX_SEGMENTS_PER_REQUEST];
523 struct bio *bio = NULL;
524 struct bio *biolist[BLKIF_MAX_SEGMENTS_PER_REQUEST];
527 struct blk_plug plug;
529 switch (req->operation) {
536 operation = WRITE_ODIRECT;
538 case BLKIF_OP_FLUSH_DISKCACHE:
540 operation = WRITE_FLUSH;
541 /* The frontend likes to set this to -1, which vbd_translate
543 req->u.rw.sector_number = 0;
545 case BLKIF_OP_WRITE_BARRIER:
547 operation = 0; /* make gcc happy */
552 /* Check that the number of segments is sane. */
553 nseg = req->nr_segments;
554 if (unlikely(nseg == 0 && operation != WRITE_FLUSH) ||
555 unlikely(nseg > BLKIF_MAX_SEGMENTS_PER_REQUEST)) {
556 DPRINTK("Bad number of segments in request (%d)\n", nseg);
557 /* Haven't submitted any bio's yet. */
561 preq.dev = req->handle;
562 preq.sector_number = req->u.rw.sector_number;
565 pending_req->blkif = blkif;
566 pending_req->id = req->id;
567 pending_req->operation = req->operation;
568 pending_req->status = BLKIF_RSP_OKAY;
569 pending_req->nr_pages = nseg;
571 for (i = 0; i < nseg; i++) {
572 seg[i].nsec = req->u.rw.seg[i].last_sect -
573 req->u.rw.seg[i].first_sect + 1;
574 if ((req->u.rw.seg[i].last_sect >= (PAGE_SIZE >> 9)) ||
575 (req->u.rw.seg[i].last_sect < req->u.rw.seg[i].first_sect))
577 preq.nr_sects += seg[i].nsec;
581 if (vbd_translate(&preq, blkif, operation) != 0) {
582 DPRINTK("access denied: %s of [%llu,%llu] on dev=%04x\n",
583 operation == READ ? "read" : "write",
585 preq.sector_number + preq.nr_sects, preq.dev);
588 /* This check _MUST_ be done after vbd_translate as the preq.bdev
590 for (i = 0; i < nseg; i++) {
591 if (((int)preq.sector_number|(int)seg[i].nsec) &
592 ((bdev_logical_block_size(preq.bdev) >> 9) - 1)) {
593 DPRINTK("Misaligned I/O request from domain %d",
598 /* If we have failed at this point, we need to undo the M2P override,
599 * set gnttab_set_unmap_op on all of the grant references and perform
600 * the hypercall to unmap the grants - that is all done in
603 if (xen_blkbk_map(req, pending_req, seg))
606 /* This corresponding blkif_put is done in __end_block_io_op */
607 xen_blkif_get(blkif);
609 for (i = 0; i < nseg; i++) {
610 while ((bio == NULL) ||
612 blkbk->pending_page(pending_req, i),
614 seg[i].buf & ~PAGE_MASK) == 0)) {
616 bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, nseg-i);
617 if (unlikely(bio == NULL))
620 bio->bi_bdev = preq.bdev;
621 bio->bi_private = pending_req;
622 bio->bi_end_io = end_block_io_op;
623 bio->bi_sector = preq.sector_number;
626 preq.sector_number += seg[i].nsec;
629 /* This will be hit if the operation was a barrier. */
631 BUG_ON(operation != WRITE_FLUSH);
632 bio = biolist[nbio++] = bio_alloc(GFP_KERNEL, 0);
633 if (unlikely(bio == NULL))
636 bio->bi_bdev = preq.bdev;
637 bio->bi_private = pending_req;
638 bio->bi_end_io = end_block_io_op;
642 /* We set it one so that the last submit_bio does not have to call
645 atomic_set(&pending_req->pendcnt, nbio);
647 /* Get a reference count for the disk queue and start sending I/O */
648 blk_start_plug(&plug);
650 for (i = 0; i < nbio; i++)
651 submit_bio(operation, biolist[i]);
653 blk_finish_plug(&plug);
654 /* Let the I/Os go.. */
656 if (operation == READ)
657 blkif->st_rd_sect += preq.nr_sects;
658 else if (operation == WRITE || operation == WRITE_FLUSH)
659 blkif->st_wr_sect += preq.nr_sects;
664 xen_blkbk_unmap(pending_req);
666 /* Haven't submitted any bio's yet. */
667 make_response(blkif, req->id, req->operation, BLKIF_RSP_ERROR);
668 free_req(pending_req);
669 msleep(1); /* back off a bit */
673 for (i = 0; i < (nbio-1); i++)
675 __end_block_io_op(pending_req, -EINVAL);
676 msleep(1); /* back off a bit */
683 * Put a response on the ring on how the operation fared.
685 static void make_response(struct blkif_st *blkif, u64 id,
686 unsigned short op, int st)
688 struct blkif_response resp;
690 union blkif_back_rings *blk_rings = &blkif->blk_rings;
698 spin_lock_irqsave(&blkif->blk_ring_lock, flags);
699 /* Place on the response ring for the relevant domain. */
700 switch (blkif->blk_protocol) {
701 case BLKIF_PROTOCOL_NATIVE:
702 memcpy(RING_GET_RESPONSE(&blk_rings->native, blk_rings->native.rsp_prod_pvt),
703 &resp, sizeof(resp));
705 case BLKIF_PROTOCOL_X86_32:
706 memcpy(RING_GET_RESPONSE(&blk_rings->x86_32, blk_rings->x86_32.rsp_prod_pvt),
707 &resp, sizeof(resp));
709 case BLKIF_PROTOCOL_X86_64:
710 memcpy(RING_GET_RESPONSE(&blk_rings->x86_64, blk_rings->x86_64.rsp_prod_pvt),
711 &resp, sizeof(resp));
716 blk_rings->common.rsp_prod_pvt++;
717 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blk_rings->common, notify);
718 if (blk_rings->common.rsp_prod_pvt == blk_rings->common.req_cons) {
720 * Tail check for pending requests. Allows frontend to avoid
721 * notifications if requests are already in flight (lower
722 * overheads and promotes batching).
724 RING_FINAL_CHECK_FOR_REQUESTS(&blk_rings->common, more_to_do);
726 } else if (RING_HAS_UNCONSUMED_REQUESTS(&blk_rings->common)) {
730 spin_unlock_irqrestore(&blkif->blk_ring_lock, flags);
733 blkif_notify_work(blkif);
735 notify_remote_via_irq(blkif->irq);
738 static int __init xen_blkif_init(void)
743 if (!xen_pv_domain())
746 blkbk = kzalloc(sizeof(struct xen_blkbk), GFP_KERNEL);
748 printk(KERN_ALERT "%s: out of memory!\n", __func__);
752 mmap_pages = xen_blkif_reqs * BLKIF_MAX_SEGMENTS_PER_REQUEST;
754 blkbk->pending_reqs = kmalloc(sizeof(blkbk->pending_reqs[0]) *
755 xen_blkif_reqs, GFP_KERNEL);
756 blkbk->pending_grant_handles = kzalloc(sizeof(blkbk->pending_grant_handles[0]) *
757 mmap_pages, GFP_KERNEL);
758 blkbk->pending_pages = kzalloc(sizeof(blkbk->pending_pages[0]) *
759 mmap_pages, GFP_KERNEL);
761 if (!blkbk->pending_reqs || !blkbk->pending_grant_handles ||
762 !blkbk->pending_pages) {
767 for (i = 0; i < mmap_pages; i++) {
768 blkbk->pending_grant_handles[i] = BLKBACK_INVALID_HANDLE;
769 blkbk->pending_pages[i] = alloc_page(GFP_KERNEL);
770 if (blkbk->pending_pages[i] == NULL) {
775 rc = xen_blkif_interface_init();
779 memset(blkbk->pending_reqs, 0, sizeof(blkbk->pending_reqs));
781 INIT_LIST_HEAD(&blkbk->pending_free);
782 spin_lock_init(&blkbk->pending_free_lock);
783 init_waitqueue_head(&blkbk->pending_free_wq);
785 for (i = 0; i < xen_blkif_reqs; i++)
786 list_add_tail(&blkbk->pending_reqs[i].free_list,
787 &blkbk->pending_free);
789 rc = xen_blkif_xenbus_init();
796 printk(KERN_ERR "%s: out of memory\n", __func__);
798 kfree(blkbk->pending_reqs);
799 kfree(blkbk->pending_grant_handles);
800 for (i = 0; i < mmap_pages; i++) {
801 if (blkbk->pending_pages[i])
802 __free_page(blkbk->pending_pages[i]);
804 kfree(blkbk->pending_pages);
810 module_init(xen_blkif_init);
812 MODULE_LICENSE("Dual BSD/GPL");