2 * xen paravirt block device backend
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; under version 2 of the License.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License along
16 * with this program; if not, see <http://www.gnu.org/licenses/>.
29 #include <sys/ioctl.h>
30 #include <sys/types.h>
37 #include <xen/io/xenbus.h>
40 #include "block_int.h"
41 #include "qemu-char.h"
42 #include "xen_blkif.h"
43 #include "xen_backend.h"
45 /* ------------------------------------------------------------- */
47 static int syncwrite = 0;
48 static int batch_maps = 0;
50 static int max_requests = 32;
51 static int use_aio = 1;
53 /* ------------------------------------------------------------- */
55 #define BLOCK_SIZE 512
56 #define IOCB_COUNT (BLKIF_MAX_SEGMENTS_PER_REQUEST + 2)
69 uint32_t domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
70 uint32_t refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
72 void *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
79 struct XenBlkDev *blkdev;
80 QLIST_ENTRY(ioreq) list;
84 struct XenDevice xendev; /* must be first */
90 const char *fileproto;
97 blkif_back_rings_t rings;
102 QLIST_HEAD(inflight_head, ioreq) inflight;
103 QLIST_HEAD(finished_head, ioreq) finished;
104 QLIST_HEAD(freelist_head, ioreq) freelist;
106 int requests_inflight;
107 int requests_finished;
109 /* qemu block driver */
111 BlockDriverState *bs;
115 /* ------------------------------------------------------------- */
117 static struct ioreq *ioreq_start(struct XenBlkDev *blkdev)
119 struct ioreq *ioreq = NULL;
121 if (QLIST_EMPTY(&blkdev->freelist)) {
122 if (blkdev->requests_total >= max_requests)
124 /* allocate new struct */
125 ioreq = qemu_mallocz(sizeof(*ioreq));
126 ioreq->blkdev = blkdev;
127 blkdev->requests_total++;
128 qemu_iovec_init(&ioreq->v, BLKIF_MAX_SEGMENTS_PER_REQUEST);
130 /* get one from freelist */
131 ioreq = QLIST_FIRST(&blkdev->freelist);
132 QLIST_REMOVE(ioreq, list);
133 qemu_iovec_reset(&ioreq->v);
135 QLIST_INSERT_HEAD(&blkdev->inflight, ioreq, list);
136 blkdev->requests_inflight++;
142 static void ioreq_finish(struct ioreq *ioreq)
144 struct XenBlkDev *blkdev = ioreq->blkdev;
146 QLIST_REMOVE(ioreq, list);
147 QLIST_INSERT_HEAD(&blkdev->finished, ioreq, list);
148 blkdev->requests_inflight--;
149 blkdev->requests_finished++;
152 static void ioreq_release(struct ioreq *ioreq)
154 struct XenBlkDev *blkdev = ioreq->blkdev;
156 QLIST_REMOVE(ioreq, list);
157 memset(ioreq, 0, sizeof(*ioreq));
158 ioreq->blkdev = blkdev;
159 QLIST_INSERT_HEAD(&blkdev->freelist, ioreq, list);
160 blkdev->requests_finished--;
164 * translate request into iovec + start offset
165 * do sanity checks along the way
167 static int ioreq_parse(struct ioreq *ioreq)
169 struct XenBlkDev *blkdev = ioreq->blkdev;
174 xen_be_printf(&blkdev->xendev, 3,
175 "op %d, nr %d, handle %d, id %" PRId64 ", sector %" PRId64 "\n",
176 ioreq->req.operation, ioreq->req.nr_segments,
177 ioreq->req.handle, ioreq->req.id, ioreq->req.sector_number);
178 switch (ioreq->req.operation) {
180 ioreq->prot = PROT_WRITE; /* to memory */
182 case BLKIF_OP_WRITE_BARRIER:
184 ioreq->presync = ioreq->postsync = 1;
187 ioreq->prot = PROT_READ; /* from memory */
192 xen_be_printf(&blkdev->xendev, 0, "error: unknown operation (%d)\n",
193 ioreq->req.operation);
197 if (ioreq->req.operation != BLKIF_OP_READ && blkdev->mode[0] != 'w') {
198 xen_be_printf(&blkdev->xendev, 0, "error: write req for ro device\n");
202 ioreq->start = ioreq->req.sector_number * blkdev->file_blk;
203 for (i = 0; i < ioreq->req.nr_segments; i++) {
204 if (i == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
205 xen_be_printf(&blkdev->xendev, 0, "error: nr_segments too big\n");
208 if (ioreq->req.seg[i].first_sect > ioreq->req.seg[i].last_sect) {
209 xen_be_printf(&blkdev->xendev, 0, "error: first > last sector\n");
212 if (ioreq->req.seg[i].last_sect * BLOCK_SIZE >= XC_PAGE_SIZE) {
213 xen_be_printf(&blkdev->xendev, 0, "error: page crossing\n");
217 ioreq->domids[i] = blkdev->xendev.dom;
218 ioreq->refs[i] = ioreq->req.seg[i].gref;
220 mem = ioreq->req.seg[i].first_sect * blkdev->file_blk;
221 len = (ioreq->req.seg[i].last_sect - ioreq->req.seg[i].first_sect + 1) * blkdev->file_blk;
222 qemu_iovec_add(&ioreq->v, (void*)mem, len);
224 if (ioreq->start + ioreq->v.size > blkdev->file_size) {
225 xen_be_printf(&blkdev->xendev, 0, "error: access beyond end of file\n");
231 ioreq->status = BLKIF_RSP_ERROR;
235 static void ioreq_unmap(struct ioreq *ioreq)
237 int gnt = ioreq->blkdev->xendev.gnttabdev;
240 if (ioreq->v.niov == 0)
245 if (xc_gnttab_munmap(gnt, ioreq->pages, ioreq->v.niov) != 0)
246 xen_be_printf(&ioreq->blkdev->xendev, 0, "xc_gnttab_munmap failed: %s\n",
248 ioreq->blkdev->cnt_map -= ioreq->v.niov;
251 for (i = 0; i < ioreq->v.niov; i++) {
254 if (xc_gnttab_munmap(gnt, ioreq->page[i], 1) != 0)
255 xen_be_printf(&ioreq->blkdev->xendev, 0, "xc_gnttab_munmap failed: %s\n",
257 ioreq->blkdev->cnt_map--;
258 ioreq->page[i] = NULL;
263 static int ioreq_map(struct ioreq *ioreq)
265 int gnt = ioreq->blkdev->xendev.gnttabdev;
268 if (ioreq->v.niov == 0)
271 ioreq->pages = xc_gnttab_map_grant_refs
272 (gnt, ioreq->v.niov, ioreq->domids, ioreq->refs, ioreq->prot);
273 if (ioreq->pages == NULL) {
274 xen_be_printf(&ioreq->blkdev->xendev, 0,
275 "can't map %d grant refs (%s, %d maps)\n",
276 ioreq->v.niov, strerror(errno), ioreq->blkdev->cnt_map);
279 for (i = 0; i < ioreq->v.niov; i++)
280 ioreq->v.iov[i].iov_base = ioreq->pages + i * XC_PAGE_SIZE +
281 (uintptr_t)ioreq->v.iov[i].iov_base;
282 ioreq->blkdev->cnt_map += ioreq->v.niov;
284 for (i = 0; i < ioreq->v.niov; i++) {
285 ioreq->page[i] = xc_gnttab_map_grant_ref
286 (gnt, ioreq->domids[i], ioreq->refs[i], ioreq->prot);
287 if (ioreq->page[i] == NULL) {
288 xen_be_printf(&ioreq->blkdev->xendev, 0,
289 "can't map grant ref %d (%s, %d maps)\n",
290 ioreq->refs[i], strerror(errno), ioreq->blkdev->cnt_map);
294 ioreq->v.iov[i].iov_base = ioreq->page[i] + (uintptr_t)ioreq->v.iov[i].iov_base;
295 ioreq->blkdev->cnt_map++;
301 static int ioreq_runio_qemu_sync(struct ioreq *ioreq)
303 struct XenBlkDev *blkdev = ioreq->blkdev;
307 if (ioreq_map(ioreq) == -1)
310 bdrv_flush(blkdev->bs);
312 switch (ioreq->req.operation) {
315 for (i = 0; i < ioreq->v.niov; i++) {
316 rc = bdrv_read(blkdev->bs, pos / BLOCK_SIZE,
317 ioreq->v.iov[i].iov_base,
318 ioreq->v.iov[i].iov_len / BLOCK_SIZE);
320 xen_be_printf(&blkdev->xendev, 0, "rd I/O error (%p, len %zd)\n",
321 ioreq->v.iov[i].iov_base,
322 ioreq->v.iov[i].iov_len);
325 len += ioreq->v.iov[i].iov_len;
326 pos += ioreq->v.iov[i].iov_len;
330 case BLKIF_OP_WRITE_BARRIER:
332 for (i = 0; i < ioreq->v.niov; i++) {
333 rc = bdrv_write(blkdev->bs, pos / BLOCK_SIZE,
334 ioreq->v.iov[i].iov_base,
335 ioreq->v.iov[i].iov_len / BLOCK_SIZE);
337 xen_be_printf(&blkdev->xendev, 0, "wr I/O error (%p, len %zd)\n",
338 ioreq->v.iov[i].iov_base,
339 ioreq->v.iov[i].iov_len);
342 len += ioreq->v.iov[i].iov_len;
343 pos += ioreq->v.iov[i].iov_len;
347 /* unknown operation (shouldn't happen -- parse catches this) */
352 bdrv_flush(blkdev->bs);
353 ioreq->status = BLKIF_RSP_OKAY;
360 ioreq->status = BLKIF_RSP_ERROR;
364 static void qemu_aio_complete(void *opaque, int ret)
366 struct ioreq *ioreq = opaque;
369 xen_be_printf(&ioreq->blkdev->xendev, 0, "%s I/O error\n",
370 ioreq->req.operation == BLKIF_OP_READ ? "read" : "write");
374 ioreq->aio_inflight--;
375 if (ioreq->aio_inflight > 0)
378 ioreq->status = ioreq->aio_errors ? BLKIF_RSP_ERROR : BLKIF_RSP_OKAY;
381 qemu_bh_schedule(ioreq->blkdev->bh);
384 static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
386 struct XenBlkDev *blkdev = ioreq->blkdev;
388 if (ioreq_map(ioreq) == -1)
391 ioreq->aio_inflight++;
393 bdrv_flush(blkdev->bs); /* FIXME: aio_flush() ??? */
395 switch (ioreq->req.operation) {
397 ioreq->aio_inflight++;
398 bdrv_aio_readv(blkdev->bs, ioreq->start / BLOCK_SIZE,
399 &ioreq->v, ioreq->v.size / BLOCK_SIZE,
400 qemu_aio_complete, ioreq);
403 case BLKIF_OP_WRITE_BARRIER:
404 ioreq->aio_inflight++;
405 bdrv_aio_writev(blkdev->bs, ioreq->start / BLOCK_SIZE,
406 &ioreq->v, ioreq->v.size / BLOCK_SIZE,
407 qemu_aio_complete, ioreq);
410 /* unknown operation (shouldn't happen -- parse catches this) */
415 bdrv_flush(blkdev->bs); /* FIXME: aio_flush() ??? */
416 qemu_aio_complete(ioreq, 0);
421 ioreq->status = BLKIF_RSP_ERROR;
425 static int blk_send_response_one(struct ioreq *ioreq)
427 struct XenBlkDev *blkdev = ioreq->blkdev;
429 int have_requests = 0;
430 blkif_response_t resp;
433 resp.id = ioreq->req.id;
434 resp.operation = ioreq->req.operation;
435 resp.status = ioreq->status;
437 /* Place on the response ring for the relevant domain. */
438 switch (blkdev->protocol) {
439 case BLKIF_PROTOCOL_NATIVE:
440 dst = RING_GET_RESPONSE(&blkdev->rings.native, blkdev->rings.native.rsp_prod_pvt);
442 case BLKIF_PROTOCOL_X86_32:
443 dst = RING_GET_RESPONSE(&blkdev->rings.x86_32_part,
444 blkdev->rings.x86_32_part.rsp_prod_pvt);
446 case BLKIF_PROTOCOL_X86_64:
447 dst = RING_GET_RESPONSE(&blkdev->rings.x86_64_part,
448 blkdev->rings.x86_64_part.rsp_prod_pvt);
453 memcpy(dst, &resp, sizeof(resp));
454 blkdev->rings.common.rsp_prod_pvt++;
456 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blkdev->rings.common, send_notify);
457 if (blkdev->rings.common.rsp_prod_pvt == blkdev->rings.common.req_cons) {
459 * Tail check for pending requests. Allows frontend to avoid
460 * notifications if requests are already in flight (lower
461 * overheads and promotes batching).
463 RING_FINAL_CHECK_FOR_REQUESTS(&blkdev->rings.common, have_requests);
464 } else if (RING_HAS_UNCONSUMED_REQUESTS(&blkdev->rings.common)) {
473 /* walk finished list, send outstanding responses, free requests */
474 static void blk_send_response_all(struct XenBlkDev *blkdev)
479 while (!QLIST_EMPTY(&blkdev->finished)) {
480 ioreq = QLIST_FIRST(&blkdev->finished);
481 send_notify += blk_send_response_one(ioreq);
482 ioreq_release(ioreq);
485 xen_be_send_notify(&blkdev->xendev);
488 static int blk_get_request(struct XenBlkDev *blkdev, struct ioreq *ioreq, RING_IDX rc)
490 switch (blkdev->protocol) {
491 case BLKIF_PROTOCOL_NATIVE:
492 memcpy(&ioreq->req, RING_GET_REQUEST(&blkdev->rings.native, rc),
495 case BLKIF_PROTOCOL_X86_32:
496 blkif_get_x86_32_req(&ioreq->req,
497 RING_GET_REQUEST(&blkdev->rings.x86_32_part, rc));
499 case BLKIF_PROTOCOL_X86_64:
500 blkif_get_x86_64_req(&ioreq->req,
501 RING_GET_REQUEST(&blkdev->rings.x86_64_part, rc));
507 static void blk_handle_requests(struct XenBlkDev *blkdev)
512 blkdev->more_work = 0;
514 rc = blkdev->rings.common.req_cons;
515 rp = blkdev->rings.common.sring->req_prod;
516 xen_rmb(); /* Ensure we see queued requests up to 'rp'. */
519 blk_send_response_all(blkdev);
521 /* pull request from ring */
522 if (RING_REQUEST_CONS_OVERFLOW(&blkdev->rings.common, rc))
524 ioreq = ioreq_start(blkdev);
529 blk_get_request(blkdev, ioreq, rc);
530 blkdev->rings.common.req_cons = ++rc;
533 if (ioreq_parse(ioreq) != 0) {
534 if (blk_send_response_one(ioreq))
535 xen_be_send_notify(&blkdev->xendev);
536 ioreq_release(ioreq);
541 /* run i/o in aio mode */
542 ioreq_runio_qemu_aio(ioreq);
544 /* run i/o in sync mode */
545 ioreq_runio_qemu_sync(ioreq);
549 blk_send_response_all(blkdev);
551 if (blkdev->more_work && blkdev->requests_inflight < max_requests)
552 qemu_bh_schedule(blkdev->bh);
555 /* ------------------------------------------------------------- */
557 static void blk_bh(void *opaque)
559 struct XenBlkDev *blkdev = opaque;
560 blk_handle_requests(blkdev);
563 static void blk_alloc(struct XenDevice *xendev)
565 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
567 QLIST_INIT(&blkdev->inflight);
568 QLIST_INIT(&blkdev->finished);
569 QLIST_INIT(&blkdev->freelist);
570 blkdev->bh = qemu_bh_new(blk_bh, blkdev);
571 if (xen_mode != XEN_EMULATE)
575 static int blk_init(struct XenDevice *xendev)
577 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
578 int index, qflags, have_barriers, info = 0;
581 /* read xenstore entries */
582 if (blkdev->params == NULL) {
583 blkdev->params = xenstore_read_be_str(&blkdev->xendev, "params");
584 h = strchr(blkdev->params, ':');
586 blkdev->fileproto = blkdev->params;
587 blkdev->filename = h+1;
590 blkdev->fileproto = "<unset>";
591 blkdev->filename = blkdev->params;
594 if (blkdev->mode == NULL)
595 blkdev->mode = xenstore_read_be_str(&blkdev->xendev, "mode");
596 if (blkdev->type == NULL)
597 blkdev->type = xenstore_read_be_str(&blkdev->xendev, "type");
598 if (blkdev->dev == NULL)
599 blkdev->dev = xenstore_read_be_str(&blkdev->xendev, "dev");
600 if (blkdev->devtype == NULL)
601 blkdev->devtype = xenstore_read_be_str(&blkdev->xendev, "device-type");
603 /* do we have all we need? */
604 if (blkdev->params == NULL ||
605 blkdev->mode == NULL ||
606 blkdev->type == NULL ||
611 if (strcmp(blkdev->mode, "w") == 0) {
612 qflags = BDRV_O_RDWR;
615 info |= VDISK_READONLY;
619 if (blkdev->devtype && !strcmp(blkdev->devtype, "cdrom"))
622 /* init qemu block driver */
623 index = (blkdev->xendev.dev - 202 * 256) / 16;
624 blkdev->dinfo = drive_get(IF_XEN, 0, index);
625 if (!blkdev->dinfo) {
626 /* setup via xenbus -> create new block driver instance */
627 xen_be_printf(&blkdev->xendev, 2, "create new bdrv (xenbus setup)\n");
628 blkdev->bs = bdrv_new(blkdev->dev);
630 if (bdrv_open(blkdev->bs, blkdev->filename, qflags,
631 bdrv_find_whitelisted_format(blkdev->fileproto))
633 bdrv_delete(blkdev->bs);
640 /* setup via qemu cmdline -> already setup for us */
641 xen_be_printf(&blkdev->xendev, 2, "get configured bdrv (cmdline setup)\n");
642 blkdev->bs = blkdev->dinfo->bdrv;
644 blkdev->file_blk = BLOCK_SIZE;
645 blkdev->file_size = bdrv_getlength(blkdev->bs);
646 if (blkdev->file_size < 0) {
647 xen_be_printf(&blkdev->xendev, 1, "bdrv_getlength: %d (%s) | drv %s\n",
648 (int)blkdev->file_size, strerror(-blkdev->file_size),
649 blkdev->bs->drv ? blkdev->bs->drv->format_name : "-");
650 blkdev->file_size = 0;
652 have_barriers = blkdev->bs->drv && blkdev->bs->drv->bdrv_flush ? 1 : 0;
654 xen_be_printf(xendev, 1, "type \"%s\", fileproto \"%s\", filename \"%s\","
655 " size %" PRId64 " (%" PRId64 " MB)\n",
656 blkdev->type, blkdev->fileproto, blkdev->filename,
657 blkdev->file_size, blkdev->file_size >> 20);
660 xenstore_write_be_int(&blkdev->xendev, "feature-barrier", have_barriers);
661 xenstore_write_be_int(&blkdev->xendev, "info", info);
662 xenstore_write_be_int(&blkdev->xendev, "sector-size", blkdev->file_blk);
663 xenstore_write_be_int(&blkdev->xendev, "sectors",
664 blkdev->file_size / blkdev->file_blk);
668 static int blk_connect(struct XenDevice *xendev)
670 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
672 if (xenstore_read_fe_int(&blkdev->xendev, "ring-ref", &blkdev->ring_ref) == -1)
674 if (xenstore_read_fe_int(&blkdev->xendev, "event-channel",
675 &blkdev->xendev.remote_port) == -1)
678 blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
679 if (blkdev->xendev.protocol) {
680 if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_32) == 0)
681 blkdev->protocol = BLKIF_PROTOCOL_X86_32;
682 if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_64) == 0)
683 blkdev->protocol = BLKIF_PROTOCOL_X86_64;
686 blkdev->sring = xc_gnttab_map_grant_ref(blkdev->xendev.gnttabdev,
689 PROT_READ | PROT_WRITE);
694 switch (blkdev->protocol) {
695 case BLKIF_PROTOCOL_NATIVE:
697 blkif_sring_t *sring_native = blkdev->sring;
698 BACK_RING_INIT(&blkdev->rings.native, sring_native, XC_PAGE_SIZE);
701 case BLKIF_PROTOCOL_X86_32:
703 blkif_x86_32_sring_t *sring_x86_32 = blkdev->sring;
705 BACK_RING_INIT(&blkdev->rings.x86_32_part, sring_x86_32, XC_PAGE_SIZE);
708 case BLKIF_PROTOCOL_X86_64:
710 blkif_x86_64_sring_t *sring_x86_64 = blkdev->sring;
712 BACK_RING_INIT(&blkdev->rings.x86_64_part, sring_x86_64, XC_PAGE_SIZE);
717 xen_be_bind_evtchn(&blkdev->xendev);
719 xen_be_printf(&blkdev->xendev, 1, "ok: proto %s, ring-ref %d, "
720 "remote port %d, local port %d\n",
721 blkdev->xendev.protocol, blkdev->ring_ref,
722 blkdev->xendev.remote_port, blkdev->xendev.local_port);
726 static void blk_disconnect(struct XenDevice *xendev)
728 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
731 if (!blkdev->dinfo) {
732 /* close/delete only if we created it ourself */
733 bdrv_close(blkdev->bs);
734 bdrv_delete(blkdev->bs);
738 xen_be_unbind_evtchn(&blkdev->xendev);
741 xc_gnttab_munmap(blkdev->xendev.gnttabdev, blkdev->sring, 1);
743 blkdev->sring = NULL;
747 static int blk_free(struct XenDevice *xendev)
749 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
752 while (!QLIST_EMPTY(&blkdev->freelist)) {
753 ioreq = QLIST_FIRST(&blkdev->freelist);
754 QLIST_REMOVE(ioreq, list);
755 qemu_iovec_destroy(&ioreq->v);
759 qemu_free(blkdev->params);
760 qemu_free(blkdev->mode);
761 qemu_free(blkdev->type);
762 qemu_free(blkdev->dev);
763 qemu_free(blkdev->devtype);
764 qemu_bh_delete(blkdev->bh);
768 static void blk_event(struct XenDevice *xendev)
770 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
772 qemu_bh_schedule(blkdev->bh);
775 struct XenDevOps xen_blkdev_ops = {
776 .size = sizeof(struct XenBlkDev),
777 .flags = DEVOPS_FLAG_NEED_GNTDEV,
780 .connect = blk_connect,
781 .disconnect = blk_disconnect,