vhost: rework IOTLB messaging
[qemu/ar7.git] / hw / block / xen_disk.c
blob3a22805fbc180b0e35faa325973bec8d7cf7ed6f
1 /*
2 * xen paravirt block device backend
4 * (c) Gerd Hoffmann <kraxel@redhat.com>
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/>.
18 * Contributions after 2012-01-13 are licensed under the terms of the
19 * GNU GPL, version 2 or (at your option) any later version.
22 #include "qemu/osdep.h"
23 #include <sys/ioctl.h>
24 #include <sys/uio.h>
26 #include "hw/hw.h"
27 #include "hw/xen/xen_backend.h"
28 #include "xen_blkif.h"
29 #include "sysemu/blockdev.h"
30 #include "sysemu/block-backend.h"
31 #include "qapi/error.h"
32 #include "qapi/qmp/qdict.h"
33 #include "qapi/qmp/qstring.h"
35 /* ------------------------------------------------------------- */
37 static int batch_maps = 0;
39 static int max_requests = 32;
41 /* ------------------------------------------------------------- */
43 #define BLOCK_SIZE 512
44 #define IOCB_COUNT (BLKIF_MAX_SEGMENTS_PER_REQUEST + 2)
46 struct PersistentGrant {
47 void *page;
48 struct XenBlkDev *blkdev;
51 typedef struct PersistentGrant PersistentGrant;
53 struct PersistentRegion {
54 void *addr;
55 int num;
58 typedef struct PersistentRegion PersistentRegion;
60 struct ioreq {
61 blkif_request_t req;
62 int16_t status;
64 /* parsed request */
65 off_t start;
66 QEMUIOVector v;
67 int presync;
68 uint8_t mapped;
70 /* grant mapping */
71 uint32_t domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
72 uint32_t refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
73 int prot;
74 void *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
75 void *pages;
76 int num_unmap;
78 /* aio status */
79 int aio_inflight;
80 int aio_errors;
82 struct XenBlkDev *blkdev;
83 QLIST_ENTRY(ioreq) list;
84 BlockAcctCookie acct;
87 struct XenBlkDev {
88 struct XenDevice xendev; /* must be first */
89 char *params;
90 char *mode;
91 char *type;
92 char *dev;
93 char *devtype;
94 bool directiosafe;
95 const char *fileproto;
96 const char *filename;
97 int ring_ref;
98 void *sring;
99 int64_t file_blk;
100 int64_t file_size;
101 int protocol;
102 blkif_back_rings_t rings;
103 int more_work;
104 int cnt_map;
106 /* request lists */
107 QLIST_HEAD(inflight_head, ioreq) inflight;
108 QLIST_HEAD(finished_head, ioreq) finished;
109 QLIST_HEAD(freelist_head, ioreq) freelist;
110 int requests_total;
111 int requests_inflight;
112 int requests_finished;
114 /* Persistent grants extension */
115 gboolean feature_discard;
116 gboolean feature_persistent;
117 GTree *persistent_gnts;
118 GSList *persistent_regions;
119 unsigned int persistent_gnt_count;
120 unsigned int max_grants;
122 /* Grant copy */
123 gboolean feature_grant_copy;
125 /* qemu block driver */
126 DriveInfo *dinfo;
127 BlockBackend *blk;
128 QEMUBH *bh;
131 /* ------------------------------------------------------------- */
133 static void ioreq_reset(struct ioreq *ioreq)
135 memset(&ioreq->req, 0, sizeof(ioreq->req));
136 ioreq->status = 0;
137 ioreq->start = 0;
138 ioreq->presync = 0;
139 ioreq->mapped = 0;
141 memset(ioreq->domids, 0, sizeof(ioreq->domids));
142 memset(ioreq->refs, 0, sizeof(ioreq->refs));
143 ioreq->prot = 0;
144 memset(ioreq->page, 0, sizeof(ioreq->page));
145 ioreq->pages = NULL;
147 ioreq->aio_inflight = 0;
148 ioreq->aio_errors = 0;
150 ioreq->blkdev = NULL;
151 memset(&ioreq->list, 0, sizeof(ioreq->list));
152 memset(&ioreq->acct, 0, sizeof(ioreq->acct));
154 qemu_iovec_reset(&ioreq->v);
157 static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
159 uint ua = GPOINTER_TO_UINT(a);
160 uint ub = GPOINTER_TO_UINT(b);
161 return (ua > ub) - (ua < ub);
164 static void destroy_grant(gpointer pgnt)
166 PersistentGrant *grant = pgnt;
167 xengnttab_handle *gnt = grant->blkdev->xendev.gnttabdev;
169 if (xengnttab_unmap(gnt, grant->page, 1) != 0) {
170 xen_pv_printf(&grant->blkdev->xendev, 0,
171 "xengnttab_unmap failed: %s\n",
172 strerror(errno));
174 grant->blkdev->persistent_gnt_count--;
175 xen_pv_printf(&grant->blkdev->xendev, 3,
176 "unmapped grant %p\n", grant->page);
177 g_free(grant);
180 static void remove_persistent_region(gpointer data, gpointer dev)
182 PersistentRegion *region = data;
183 struct XenBlkDev *blkdev = dev;
184 xengnttab_handle *gnt = blkdev->xendev.gnttabdev;
186 if (xengnttab_unmap(gnt, region->addr, region->num) != 0) {
187 xen_pv_printf(&blkdev->xendev, 0,
188 "xengnttab_unmap region %p failed: %s\n",
189 region->addr, strerror(errno));
191 xen_pv_printf(&blkdev->xendev, 3,
192 "unmapped grant region %p with %d pages\n",
193 region->addr, region->num);
194 g_free(region);
197 static struct ioreq *ioreq_start(struct XenBlkDev *blkdev)
199 struct ioreq *ioreq = NULL;
201 if (QLIST_EMPTY(&blkdev->freelist)) {
202 if (blkdev->requests_total >= max_requests) {
203 goto out;
205 /* allocate new struct */
206 ioreq = g_malloc0(sizeof(*ioreq));
207 ioreq->blkdev = blkdev;
208 blkdev->requests_total++;
209 qemu_iovec_init(&ioreq->v, BLKIF_MAX_SEGMENTS_PER_REQUEST);
210 } else {
211 /* get one from freelist */
212 ioreq = QLIST_FIRST(&blkdev->freelist);
213 QLIST_REMOVE(ioreq, list);
215 QLIST_INSERT_HEAD(&blkdev->inflight, ioreq, list);
216 blkdev->requests_inflight++;
218 out:
219 return ioreq;
222 static void ioreq_finish(struct ioreq *ioreq)
224 struct XenBlkDev *blkdev = ioreq->blkdev;
226 QLIST_REMOVE(ioreq, list);
227 QLIST_INSERT_HEAD(&blkdev->finished, ioreq, list);
228 blkdev->requests_inflight--;
229 blkdev->requests_finished++;
232 static void ioreq_release(struct ioreq *ioreq, bool finish)
234 struct XenBlkDev *blkdev = ioreq->blkdev;
236 QLIST_REMOVE(ioreq, list);
237 ioreq_reset(ioreq);
238 ioreq->blkdev = blkdev;
239 QLIST_INSERT_HEAD(&blkdev->freelist, ioreq, list);
240 if (finish) {
241 blkdev->requests_finished--;
242 } else {
243 blkdev->requests_inflight--;
248 * translate request into iovec + start offset
249 * do sanity checks along the way
251 static int ioreq_parse(struct ioreq *ioreq)
253 struct XenBlkDev *blkdev = ioreq->blkdev;
254 uintptr_t mem;
255 size_t len;
256 int i;
258 xen_pv_printf(&blkdev->xendev, 3,
259 "op %d, nr %d, handle %d, id %" PRId64 ", sector %" PRId64 "\n",
260 ioreq->req.operation, ioreq->req.nr_segments,
261 ioreq->req.handle, ioreq->req.id, ioreq->req.sector_number);
262 switch (ioreq->req.operation) {
263 case BLKIF_OP_READ:
264 ioreq->prot = PROT_WRITE; /* to memory */
265 break;
266 case BLKIF_OP_FLUSH_DISKCACHE:
267 ioreq->presync = 1;
268 if (!ioreq->req.nr_segments) {
269 return 0;
271 /* fall through */
272 case BLKIF_OP_WRITE:
273 ioreq->prot = PROT_READ; /* from memory */
274 break;
275 case BLKIF_OP_DISCARD:
276 return 0;
277 default:
278 xen_pv_printf(&blkdev->xendev, 0, "error: unknown operation (%d)\n",
279 ioreq->req.operation);
280 goto err;
283 if (ioreq->req.operation != BLKIF_OP_READ && blkdev->mode[0] != 'w') {
284 xen_pv_printf(&blkdev->xendev, 0, "error: write req for ro device\n");
285 goto err;
288 ioreq->start = ioreq->req.sector_number * blkdev->file_blk;
289 for (i = 0; i < ioreq->req.nr_segments; i++) {
290 if (i == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
291 xen_pv_printf(&blkdev->xendev, 0, "error: nr_segments too big\n");
292 goto err;
294 if (ioreq->req.seg[i].first_sect > ioreq->req.seg[i].last_sect) {
295 xen_pv_printf(&blkdev->xendev, 0, "error: first > last sector\n");
296 goto err;
298 if (ioreq->req.seg[i].last_sect * BLOCK_SIZE >= XC_PAGE_SIZE) {
299 xen_pv_printf(&blkdev->xendev, 0, "error: page crossing\n");
300 goto err;
303 ioreq->domids[i] = blkdev->xendev.dom;
304 ioreq->refs[i] = ioreq->req.seg[i].gref;
306 mem = ioreq->req.seg[i].first_sect * blkdev->file_blk;
307 len = (ioreq->req.seg[i].last_sect - ioreq->req.seg[i].first_sect + 1) * blkdev->file_blk;
308 qemu_iovec_add(&ioreq->v, (void*)mem, len);
310 if (ioreq->start + ioreq->v.size > blkdev->file_size) {
311 xen_pv_printf(&blkdev->xendev, 0, "error: access beyond end of file\n");
312 goto err;
314 return 0;
316 err:
317 ioreq->status = BLKIF_RSP_ERROR;
318 return -1;
321 static void ioreq_unmap(struct ioreq *ioreq)
323 xengnttab_handle *gnt = ioreq->blkdev->xendev.gnttabdev;
324 int i;
326 if (ioreq->num_unmap == 0 || ioreq->mapped == 0) {
327 return;
329 if (batch_maps) {
330 if (!ioreq->pages) {
331 return;
333 if (xengnttab_unmap(gnt, ioreq->pages, ioreq->num_unmap) != 0) {
334 xen_pv_printf(&ioreq->blkdev->xendev, 0,
335 "xengnttab_unmap failed: %s\n",
336 strerror(errno));
338 ioreq->blkdev->cnt_map -= ioreq->num_unmap;
339 ioreq->pages = NULL;
340 } else {
341 for (i = 0; i < ioreq->num_unmap; i++) {
342 if (!ioreq->page[i]) {
343 continue;
345 if (xengnttab_unmap(gnt, ioreq->page[i], 1) != 0) {
346 xen_pv_printf(&ioreq->blkdev->xendev, 0,
347 "xengnttab_unmap failed: %s\n",
348 strerror(errno));
350 ioreq->blkdev->cnt_map--;
351 ioreq->page[i] = NULL;
354 ioreq->mapped = 0;
357 static int ioreq_map(struct ioreq *ioreq)
359 xengnttab_handle *gnt = ioreq->blkdev->xendev.gnttabdev;
360 uint32_t domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
361 uint32_t refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
362 void *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
363 int i, j, new_maps = 0;
364 PersistentGrant *grant;
365 PersistentRegion *region;
366 /* domids and refs variables will contain the information necessary
367 * to map the grants that are needed to fulfill this request.
369 * After mapping the needed grants, the page array will contain the
370 * memory address of each granted page in the order specified in ioreq
371 * (disregarding if it's a persistent grant or not).
374 if (ioreq->v.niov == 0 || ioreq->mapped == 1) {
375 return 0;
377 if (ioreq->blkdev->feature_persistent) {
378 for (i = 0; i < ioreq->v.niov; i++) {
379 grant = g_tree_lookup(ioreq->blkdev->persistent_gnts,
380 GUINT_TO_POINTER(ioreq->refs[i]));
382 if (grant != NULL) {
383 page[i] = grant->page;
384 xen_pv_printf(&ioreq->blkdev->xendev, 3,
385 "using persistent-grant %" PRIu32 "\n",
386 ioreq->refs[i]);
387 } else {
388 /* Add the grant to the list of grants that
389 * should be mapped
391 domids[new_maps] = ioreq->domids[i];
392 refs[new_maps] = ioreq->refs[i];
393 page[i] = NULL;
394 new_maps++;
397 /* Set the protection to RW, since grants may be reused later
398 * with a different protection than the one needed for this request
400 ioreq->prot = PROT_WRITE | PROT_READ;
401 } else {
402 /* All grants in the request should be mapped */
403 memcpy(refs, ioreq->refs, sizeof(refs));
404 memcpy(domids, ioreq->domids, sizeof(domids));
405 memset(page, 0, sizeof(page));
406 new_maps = ioreq->v.niov;
409 if (batch_maps && new_maps) {
410 ioreq->pages = xengnttab_map_grant_refs
411 (gnt, new_maps, domids, refs, ioreq->prot);
412 if (ioreq->pages == NULL) {
413 xen_pv_printf(&ioreq->blkdev->xendev, 0,
414 "can't map %d grant refs (%s, %d maps)\n",
415 new_maps, strerror(errno), ioreq->blkdev->cnt_map);
416 return -1;
418 for (i = 0, j = 0; i < ioreq->v.niov; i++) {
419 if (page[i] == NULL) {
420 page[i] = ioreq->pages + (j++) * XC_PAGE_SIZE;
423 ioreq->blkdev->cnt_map += new_maps;
424 } else if (new_maps) {
425 for (i = 0; i < new_maps; i++) {
426 ioreq->page[i] = xengnttab_map_grant_ref
427 (gnt, domids[i], refs[i], ioreq->prot);
428 if (ioreq->page[i] == NULL) {
429 xen_pv_printf(&ioreq->blkdev->xendev, 0,
430 "can't map grant ref %d (%s, %d maps)\n",
431 refs[i], strerror(errno), ioreq->blkdev->cnt_map);
432 ioreq->mapped = 1;
433 ioreq_unmap(ioreq);
434 return -1;
436 ioreq->blkdev->cnt_map++;
438 for (i = 0, j = 0; i < ioreq->v.niov; i++) {
439 if (page[i] == NULL) {
440 page[i] = ioreq->page[j++];
444 if (ioreq->blkdev->feature_persistent && new_maps != 0 &&
445 (!batch_maps || (ioreq->blkdev->persistent_gnt_count + new_maps <=
446 ioreq->blkdev->max_grants))) {
448 * If we are using persistent grants and batch mappings only
449 * add the new maps to the list of persistent grants if the whole
450 * area can be persistently mapped.
452 if (batch_maps) {
453 region = g_malloc0(sizeof(*region));
454 region->addr = ioreq->pages;
455 region->num = new_maps;
456 ioreq->blkdev->persistent_regions = g_slist_append(
457 ioreq->blkdev->persistent_regions,
458 region);
460 while ((ioreq->blkdev->persistent_gnt_count < ioreq->blkdev->max_grants)
461 && new_maps) {
462 /* Go through the list of newly mapped grants and add as many
463 * as possible to the list of persistently mapped grants.
465 * Since we start at the end of ioreq->page(s), we only need
466 * to decrease new_maps to prevent this granted pages from
467 * being unmapped in ioreq_unmap.
469 grant = g_malloc0(sizeof(*grant));
470 new_maps--;
471 if (batch_maps) {
472 grant->page = ioreq->pages + (new_maps) * XC_PAGE_SIZE;
473 } else {
474 grant->page = ioreq->page[new_maps];
476 grant->blkdev = ioreq->blkdev;
477 xen_pv_printf(&ioreq->blkdev->xendev, 3,
478 "adding grant %" PRIu32 " page: %p\n",
479 refs[new_maps], grant->page);
480 g_tree_insert(ioreq->blkdev->persistent_gnts,
481 GUINT_TO_POINTER(refs[new_maps]),
482 grant);
483 ioreq->blkdev->persistent_gnt_count++;
485 assert(!batch_maps || new_maps == 0);
487 for (i = 0; i < ioreq->v.niov; i++) {
488 ioreq->v.iov[i].iov_base += (uintptr_t)page[i];
490 ioreq->mapped = 1;
491 ioreq->num_unmap = new_maps;
492 return 0;
495 #if CONFIG_XEN_CTRL_INTERFACE_VERSION >= 40800
497 static void ioreq_free_copy_buffers(struct ioreq *ioreq)
499 int i;
501 for (i = 0; i < ioreq->v.niov; i++) {
502 ioreq->page[i] = NULL;
505 qemu_vfree(ioreq->pages);
508 static int ioreq_init_copy_buffers(struct ioreq *ioreq)
510 int i;
512 if (ioreq->v.niov == 0) {
513 return 0;
516 ioreq->pages = qemu_memalign(XC_PAGE_SIZE, ioreq->v.niov * XC_PAGE_SIZE);
518 for (i = 0; i < ioreq->v.niov; i++) {
519 ioreq->page[i] = ioreq->pages + i * XC_PAGE_SIZE;
520 ioreq->v.iov[i].iov_base = ioreq->page[i];
523 return 0;
526 static int ioreq_grant_copy(struct ioreq *ioreq)
528 xengnttab_handle *gnt = ioreq->blkdev->xendev.gnttabdev;
529 xengnttab_grant_copy_segment_t segs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
530 int i, count, rc;
531 int64_t file_blk = ioreq->blkdev->file_blk;
533 if (ioreq->v.niov == 0) {
534 return 0;
537 count = ioreq->v.niov;
539 for (i = 0; i < count; i++) {
540 if (ioreq->req.operation == BLKIF_OP_READ) {
541 segs[i].flags = GNTCOPY_dest_gref;
542 segs[i].dest.foreign.ref = ioreq->refs[i];
543 segs[i].dest.foreign.domid = ioreq->domids[i];
544 segs[i].dest.foreign.offset = ioreq->req.seg[i].first_sect * file_blk;
545 segs[i].source.virt = ioreq->v.iov[i].iov_base;
546 } else {
547 segs[i].flags = GNTCOPY_source_gref;
548 segs[i].source.foreign.ref = ioreq->refs[i];
549 segs[i].source.foreign.domid = ioreq->domids[i];
550 segs[i].source.foreign.offset = ioreq->req.seg[i].first_sect * file_blk;
551 segs[i].dest.virt = ioreq->v.iov[i].iov_base;
553 segs[i].len = (ioreq->req.seg[i].last_sect
554 - ioreq->req.seg[i].first_sect + 1) * file_blk;
557 rc = xengnttab_grant_copy(gnt, count, segs);
559 if (rc) {
560 xen_pv_printf(&ioreq->blkdev->xendev, 0,
561 "failed to copy data %d\n", rc);
562 ioreq->aio_errors++;
563 return -1;
566 for (i = 0; i < count; i++) {
567 if (segs[i].status != GNTST_okay) {
568 xen_pv_printf(&ioreq->blkdev->xendev, 3,
569 "failed to copy data %d for gref %d, domid %d\n",
570 segs[i].status, ioreq->refs[i], ioreq->domids[i]);
571 ioreq->aio_errors++;
572 rc = -1;
576 return rc;
578 #else
579 static void ioreq_free_copy_buffers(struct ioreq *ioreq)
581 abort();
584 static int ioreq_init_copy_buffers(struct ioreq *ioreq)
586 abort();
589 static int ioreq_grant_copy(struct ioreq *ioreq)
591 abort();
593 #endif
595 static int ioreq_runio_qemu_aio(struct ioreq *ioreq);
597 static void qemu_aio_complete(void *opaque, int ret)
599 struct ioreq *ioreq = opaque;
601 if (ret != 0) {
602 xen_pv_printf(&ioreq->blkdev->xendev, 0, "%s I/O error\n",
603 ioreq->req.operation == BLKIF_OP_READ ? "read" : "write");
604 ioreq->aio_errors++;
607 ioreq->aio_inflight--;
608 if (ioreq->presync) {
609 ioreq->presync = 0;
610 ioreq_runio_qemu_aio(ioreq);
611 return;
613 if (ioreq->aio_inflight > 0) {
614 return;
617 if (ioreq->blkdev->feature_grant_copy) {
618 switch (ioreq->req.operation) {
619 case BLKIF_OP_READ:
620 /* in case of failure ioreq->aio_errors is increased */
621 if (ret == 0) {
622 ioreq_grant_copy(ioreq);
624 ioreq_free_copy_buffers(ioreq);
625 break;
626 case BLKIF_OP_WRITE:
627 case BLKIF_OP_FLUSH_DISKCACHE:
628 if (!ioreq->req.nr_segments) {
629 break;
631 ioreq_free_copy_buffers(ioreq);
632 break;
633 default:
634 break;
638 ioreq->status = ioreq->aio_errors ? BLKIF_RSP_ERROR : BLKIF_RSP_OKAY;
639 if (!ioreq->blkdev->feature_grant_copy) {
640 ioreq_unmap(ioreq);
642 ioreq_finish(ioreq);
643 switch (ioreq->req.operation) {
644 case BLKIF_OP_WRITE:
645 case BLKIF_OP_FLUSH_DISKCACHE:
646 if (!ioreq->req.nr_segments) {
647 break;
649 case BLKIF_OP_READ:
650 if (ioreq->status == BLKIF_RSP_OKAY) {
651 block_acct_done(blk_get_stats(ioreq->blkdev->blk), &ioreq->acct);
652 } else {
653 block_acct_failed(blk_get_stats(ioreq->blkdev->blk), &ioreq->acct);
655 break;
656 case BLKIF_OP_DISCARD:
657 default:
658 break;
660 qemu_bh_schedule(ioreq->blkdev->bh);
663 static bool blk_split_discard(struct ioreq *ioreq, blkif_sector_t sector_number,
664 uint64_t nr_sectors)
666 struct XenBlkDev *blkdev = ioreq->blkdev;
667 int64_t byte_offset;
668 int byte_chunk;
669 uint64_t byte_remaining, limit;
670 uint64_t sec_start = sector_number;
671 uint64_t sec_count = nr_sectors;
673 /* Wrap around, or overflowing byte limit? */
674 if (sec_start + sec_count < sec_count ||
675 sec_start + sec_count > INT64_MAX >> BDRV_SECTOR_BITS) {
676 return false;
679 limit = BDRV_REQUEST_MAX_SECTORS << BDRV_SECTOR_BITS;
680 byte_offset = sec_start << BDRV_SECTOR_BITS;
681 byte_remaining = sec_count << BDRV_SECTOR_BITS;
683 do {
684 byte_chunk = byte_remaining > limit ? limit : byte_remaining;
685 ioreq->aio_inflight++;
686 blk_aio_pdiscard(blkdev->blk, byte_offset, byte_chunk,
687 qemu_aio_complete, ioreq);
688 byte_remaining -= byte_chunk;
689 byte_offset += byte_chunk;
690 } while (byte_remaining > 0);
692 return true;
695 static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
697 struct XenBlkDev *blkdev = ioreq->blkdev;
699 if (ioreq->blkdev->feature_grant_copy) {
700 ioreq_init_copy_buffers(ioreq);
701 if (ioreq->req.nr_segments && (ioreq->req.operation == BLKIF_OP_WRITE ||
702 ioreq->req.operation == BLKIF_OP_FLUSH_DISKCACHE) &&
703 ioreq_grant_copy(ioreq)) {
704 ioreq_free_copy_buffers(ioreq);
705 goto err;
707 } else {
708 if (ioreq->req.nr_segments && ioreq_map(ioreq)) {
709 goto err;
713 ioreq->aio_inflight++;
714 if (ioreq->presync) {
715 blk_aio_flush(ioreq->blkdev->blk, qemu_aio_complete, ioreq);
716 return 0;
719 switch (ioreq->req.operation) {
720 case BLKIF_OP_READ:
721 block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
722 ioreq->v.size, BLOCK_ACCT_READ);
723 ioreq->aio_inflight++;
724 blk_aio_preadv(blkdev->blk, ioreq->start, &ioreq->v, 0,
725 qemu_aio_complete, ioreq);
726 break;
727 case BLKIF_OP_WRITE:
728 case BLKIF_OP_FLUSH_DISKCACHE:
729 if (!ioreq->req.nr_segments) {
730 break;
733 block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
734 ioreq->v.size,
735 ioreq->req.operation == BLKIF_OP_WRITE ?
736 BLOCK_ACCT_WRITE : BLOCK_ACCT_FLUSH);
737 ioreq->aio_inflight++;
738 blk_aio_pwritev(blkdev->blk, ioreq->start, &ioreq->v, 0,
739 qemu_aio_complete, ioreq);
740 break;
741 case BLKIF_OP_DISCARD:
743 struct blkif_request_discard *req = (void *)&ioreq->req;
744 if (!blk_split_discard(ioreq, req->sector_number, req->nr_sectors)) {
745 goto err;
747 break;
749 default:
750 /* unknown operation (shouldn't happen -- parse catches this) */
751 if (!ioreq->blkdev->feature_grant_copy) {
752 ioreq_unmap(ioreq);
754 goto err;
757 qemu_aio_complete(ioreq, 0);
759 return 0;
761 err:
762 ioreq_finish(ioreq);
763 ioreq->status = BLKIF_RSP_ERROR;
764 return -1;
767 static int blk_send_response_one(struct ioreq *ioreq)
769 struct XenBlkDev *blkdev = ioreq->blkdev;
770 int send_notify = 0;
771 int have_requests = 0;
772 blkif_response_t resp;
773 void *dst;
775 resp.id = ioreq->req.id;
776 resp.operation = ioreq->req.operation;
777 resp.status = ioreq->status;
779 /* Place on the response ring for the relevant domain. */
780 switch (blkdev->protocol) {
781 case BLKIF_PROTOCOL_NATIVE:
782 dst = RING_GET_RESPONSE(&blkdev->rings.native, blkdev->rings.native.rsp_prod_pvt);
783 break;
784 case BLKIF_PROTOCOL_X86_32:
785 dst = RING_GET_RESPONSE(&blkdev->rings.x86_32_part,
786 blkdev->rings.x86_32_part.rsp_prod_pvt);
787 break;
788 case BLKIF_PROTOCOL_X86_64:
789 dst = RING_GET_RESPONSE(&blkdev->rings.x86_64_part,
790 blkdev->rings.x86_64_part.rsp_prod_pvt);
791 break;
792 default:
793 dst = NULL;
794 return 0;
796 memcpy(dst, &resp, sizeof(resp));
797 blkdev->rings.common.rsp_prod_pvt++;
799 RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blkdev->rings.common, send_notify);
800 if (blkdev->rings.common.rsp_prod_pvt == blkdev->rings.common.req_cons) {
802 * Tail check for pending requests. Allows frontend to avoid
803 * notifications if requests are already in flight (lower
804 * overheads and promotes batching).
806 RING_FINAL_CHECK_FOR_REQUESTS(&blkdev->rings.common, have_requests);
807 } else if (RING_HAS_UNCONSUMED_REQUESTS(&blkdev->rings.common)) {
808 have_requests = 1;
811 if (have_requests) {
812 blkdev->more_work++;
814 return send_notify;
817 /* walk finished list, send outstanding responses, free requests */
818 static void blk_send_response_all(struct XenBlkDev *blkdev)
820 struct ioreq *ioreq;
821 int send_notify = 0;
823 while (!QLIST_EMPTY(&blkdev->finished)) {
824 ioreq = QLIST_FIRST(&blkdev->finished);
825 send_notify += blk_send_response_one(ioreq);
826 ioreq_release(ioreq, true);
828 if (send_notify) {
829 xen_pv_send_notify(&blkdev->xendev);
833 static int blk_get_request(struct XenBlkDev *blkdev, struct ioreq *ioreq, RING_IDX rc)
835 switch (blkdev->protocol) {
836 case BLKIF_PROTOCOL_NATIVE:
837 memcpy(&ioreq->req, RING_GET_REQUEST(&blkdev->rings.native, rc),
838 sizeof(ioreq->req));
839 break;
840 case BLKIF_PROTOCOL_X86_32:
841 blkif_get_x86_32_req(&ioreq->req,
842 RING_GET_REQUEST(&blkdev->rings.x86_32_part, rc));
843 break;
844 case BLKIF_PROTOCOL_X86_64:
845 blkif_get_x86_64_req(&ioreq->req,
846 RING_GET_REQUEST(&blkdev->rings.x86_64_part, rc));
847 break;
849 /* Prevent the compiler from accessing the on-ring fields instead. */
850 barrier();
851 return 0;
854 static void blk_handle_requests(struct XenBlkDev *blkdev)
856 RING_IDX rc, rp;
857 struct ioreq *ioreq;
859 blkdev->more_work = 0;
861 rc = blkdev->rings.common.req_cons;
862 rp = blkdev->rings.common.sring->req_prod;
863 xen_rmb(); /* Ensure we see queued requests up to 'rp'. */
865 blk_send_response_all(blkdev);
866 while (rc != rp) {
867 /* pull request from ring */
868 if (RING_REQUEST_CONS_OVERFLOW(&blkdev->rings.common, rc)) {
869 break;
871 ioreq = ioreq_start(blkdev);
872 if (ioreq == NULL) {
873 blkdev->more_work++;
874 break;
876 blk_get_request(blkdev, ioreq, rc);
877 blkdev->rings.common.req_cons = ++rc;
879 /* parse them */
880 if (ioreq_parse(ioreq) != 0) {
882 switch (ioreq->req.operation) {
883 case BLKIF_OP_READ:
884 block_acct_invalid(blk_get_stats(blkdev->blk),
885 BLOCK_ACCT_READ);
886 break;
887 case BLKIF_OP_WRITE:
888 block_acct_invalid(blk_get_stats(blkdev->blk),
889 BLOCK_ACCT_WRITE);
890 break;
891 case BLKIF_OP_FLUSH_DISKCACHE:
892 block_acct_invalid(blk_get_stats(blkdev->blk),
893 BLOCK_ACCT_FLUSH);
894 default:
895 break;
898 if (blk_send_response_one(ioreq)) {
899 xen_pv_send_notify(&blkdev->xendev);
901 ioreq_release(ioreq, false);
902 continue;
905 ioreq_runio_qemu_aio(ioreq);
908 if (blkdev->more_work && blkdev->requests_inflight < max_requests) {
909 qemu_bh_schedule(blkdev->bh);
913 /* ------------------------------------------------------------- */
915 static void blk_bh(void *opaque)
917 struct XenBlkDev *blkdev = opaque;
918 blk_handle_requests(blkdev);
922 * We need to account for the grant allocations requiring contiguous
923 * chunks; the worst case number would be
924 * max_req * max_seg + (max_req - 1) * (max_seg - 1) + 1,
925 * but in order to keep things simple just use
926 * 2 * max_req * max_seg.
928 #define MAX_GRANTS(max_req, max_seg) (2 * (max_req) * (max_seg))
930 static void blk_alloc(struct XenDevice *xendev)
932 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
934 QLIST_INIT(&blkdev->inflight);
935 QLIST_INIT(&blkdev->finished);
936 QLIST_INIT(&blkdev->freelist);
937 blkdev->bh = qemu_bh_new(blk_bh, blkdev);
938 if (xen_mode != XEN_EMULATE) {
939 batch_maps = 1;
941 if (xengnttab_set_max_grants(xendev->gnttabdev,
942 MAX_GRANTS(max_requests, BLKIF_MAX_SEGMENTS_PER_REQUEST)) < 0) {
943 xen_pv_printf(xendev, 0, "xengnttab_set_max_grants failed: %s\n",
944 strerror(errno));
948 static void blk_parse_discard(struct XenBlkDev *blkdev)
950 int enable;
952 blkdev->feature_discard = true;
954 if (xenstore_read_be_int(&blkdev->xendev, "discard-enable", &enable) == 0) {
955 blkdev->feature_discard = !!enable;
958 if (blkdev->feature_discard) {
959 xenstore_write_be_int(&blkdev->xendev, "feature-discard", 1);
963 static int blk_init(struct XenDevice *xendev)
965 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
966 int info = 0;
967 char *directiosafe = NULL;
969 /* read xenstore entries */
970 if (blkdev->params == NULL) {
971 char *h = NULL;
972 blkdev->params = xenstore_read_be_str(&blkdev->xendev, "params");
973 if (blkdev->params != NULL) {
974 h = strchr(blkdev->params, ':');
976 if (h != NULL) {
977 blkdev->fileproto = blkdev->params;
978 blkdev->filename = h+1;
979 *h = 0;
980 } else {
981 blkdev->fileproto = "<unset>";
982 blkdev->filename = blkdev->params;
985 if (!strcmp("aio", blkdev->fileproto)) {
986 blkdev->fileproto = "raw";
988 if (!strcmp("vhd", blkdev->fileproto)) {
989 blkdev->fileproto = "vpc";
991 if (blkdev->mode == NULL) {
992 blkdev->mode = xenstore_read_be_str(&blkdev->xendev, "mode");
994 if (blkdev->type == NULL) {
995 blkdev->type = xenstore_read_be_str(&blkdev->xendev, "type");
997 if (blkdev->dev == NULL) {
998 blkdev->dev = xenstore_read_be_str(&blkdev->xendev, "dev");
1000 if (blkdev->devtype == NULL) {
1001 blkdev->devtype = xenstore_read_be_str(&blkdev->xendev, "device-type");
1003 directiosafe = xenstore_read_be_str(&blkdev->xendev, "direct-io-safe");
1004 blkdev->directiosafe = (directiosafe && atoi(directiosafe));
1006 /* do we have all we need? */
1007 if (blkdev->params == NULL ||
1008 blkdev->mode == NULL ||
1009 blkdev->type == NULL ||
1010 blkdev->dev == NULL) {
1011 goto out_error;
1014 /* read-only ? */
1015 if (strcmp(blkdev->mode, "w")) {
1016 info |= VDISK_READONLY;
1019 /* cdrom ? */
1020 if (blkdev->devtype && !strcmp(blkdev->devtype, "cdrom")) {
1021 info |= VDISK_CDROM;
1024 blkdev->file_blk = BLOCK_SIZE;
1026 /* fill info
1027 * blk_connect supplies sector-size and sectors
1029 xenstore_write_be_int(&blkdev->xendev, "feature-flush-cache", 1);
1030 xenstore_write_be_int(&blkdev->xendev, "feature-persistent", 1);
1031 xenstore_write_be_int(&blkdev->xendev, "info", info);
1033 blk_parse_discard(blkdev);
1035 g_free(directiosafe);
1036 return 0;
1038 out_error:
1039 g_free(blkdev->params);
1040 blkdev->params = NULL;
1041 g_free(blkdev->mode);
1042 blkdev->mode = NULL;
1043 g_free(blkdev->type);
1044 blkdev->type = NULL;
1045 g_free(blkdev->dev);
1046 blkdev->dev = NULL;
1047 g_free(blkdev->devtype);
1048 blkdev->devtype = NULL;
1049 g_free(directiosafe);
1050 blkdev->directiosafe = false;
1051 return -1;
1054 static int blk_connect(struct XenDevice *xendev)
1056 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1057 int pers, index, qflags;
1058 bool readonly = true;
1059 bool writethrough = true;
1061 /* read-only ? */
1062 if (blkdev->directiosafe) {
1063 qflags = BDRV_O_NOCACHE | BDRV_O_NATIVE_AIO;
1064 } else {
1065 qflags = 0;
1066 writethrough = false;
1068 if (strcmp(blkdev->mode, "w") == 0) {
1069 qflags |= BDRV_O_RDWR;
1070 readonly = false;
1072 if (blkdev->feature_discard) {
1073 qflags |= BDRV_O_UNMAP;
1076 /* init qemu block driver */
1077 index = (blkdev->xendev.dev - 202 * 256) / 16;
1078 blkdev->dinfo = drive_get(IF_XEN, 0, index);
1079 if (!blkdev->dinfo) {
1080 Error *local_err = NULL;
1081 QDict *options = NULL;
1083 if (strcmp(blkdev->fileproto, "<unset>")) {
1084 options = qdict_new();
1085 qdict_put_str(options, "driver", blkdev->fileproto);
1088 /* setup via xenbus -> create new block driver instance */
1089 xen_pv_printf(&blkdev->xendev, 2, "create new bdrv (xenbus setup)\n");
1090 blkdev->blk = blk_new_open(blkdev->filename, NULL, options,
1091 qflags, &local_err);
1092 if (!blkdev->blk) {
1093 xen_pv_printf(&blkdev->xendev, 0, "error: %s\n",
1094 error_get_pretty(local_err));
1095 error_free(local_err);
1096 return -1;
1098 blk_set_enable_write_cache(blkdev->blk, !writethrough);
1099 } else {
1100 /* setup via qemu cmdline -> already setup for us */
1101 xen_pv_printf(&blkdev->xendev, 2,
1102 "get configured bdrv (cmdline setup)\n");
1103 blkdev->blk = blk_by_legacy_dinfo(blkdev->dinfo);
1104 if (blk_is_read_only(blkdev->blk) && !readonly) {
1105 xen_pv_printf(&blkdev->xendev, 0, "Unexpected read-only drive");
1106 blkdev->blk = NULL;
1107 return -1;
1109 /* blkdev->blk is not create by us, we get a reference
1110 * so we can blk_unref() unconditionally */
1111 blk_ref(blkdev->blk);
1113 blk_attach_dev_legacy(blkdev->blk, blkdev);
1114 blkdev->file_size = blk_getlength(blkdev->blk);
1115 if (blkdev->file_size < 0) {
1116 BlockDriverState *bs = blk_bs(blkdev->blk);
1117 const char *drv_name = bs ? bdrv_get_format_name(bs) : NULL;
1118 xen_pv_printf(&blkdev->xendev, 1, "blk_getlength: %d (%s) | drv %s\n",
1119 (int)blkdev->file_size, strerror(-blkdev->file_size),
1120 drv_name ?: "-");
1121 blkdev->file_size = 0;
1124 xen_pv_printf(xendev, 1, "type \"%s\", fileproto \"%s\", filename \"%s\","
1125 " size %" PRId64 " (%" PRId64 " MB)\n",
1126 blkdev->type, blkdev->fileproto, blkdev->filename,
1127 blkdev->file_size, blkdev->file_size >> 20);
1129 /* Fill in number of sector size and number of sectors */
1130 xenstore_write_be_int(&blkdev->xendev, "sector-size", blkdev->file_blk);
1131 xenstore_write_be_int64(&blkdev->xendev, "sectors",
1132 blkdev->file_size / blkdev->file_blk);
1134 if (xenstore_read_fe_int(&blkdev->xendev, "ring-ref", &blkdev->ring_ref) == -1) {
1135 return -1;
1137 if (xenstore_read_fe_int(&blkdev->xendev, "event-channel",
1138 &blkdev->xendev.remote_port) == -1) {
1139 return -1;
1141 if (xenstore_read_fe_int(&blkdev->xendev, "feature-persistent", &pers)) {
1142 blkdev->feature_persistent = FALSE;
1143 } else {
1144 blkdev->feature_persistent = !!pers;
1147 if (!blkdev->xendev.protocol) {
1148 blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
1149 } else if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_NATIVE) == 0) {
1150 blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
1151 } else if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_32) == 0) {
1152 blkdev->protocol = BLKIF_PROTOCOL_X86_32;
1153 } else if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_64) == 0) {
1154 blkdev->protocol = BLKIF_PROTOCOL_X86_64;
1155 } else {
1156 blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
1159 blkdev->sring = xengnttab_map_grant_ref(blkdev->xendev.gnttabdev,
1160 blkdev->xendev.dom,
1161 blkdev->ring_ref,
1162 PROT_READ | PROT_WRITE);
1163 if (!blkdev->sring) {
1164 return -1;
1166 blkdev->cnt_map++;
1168 switch (blkdev->protocol) {
1169 case BLKIF_PROTOCOL_NATIVE:
1171 blkif_sring_t *sring_native = blkdev->sring;
1172 BACK_RING_INIT(&blkdev->rings.native, sring_native, XC_PAGE_SIZE);
1173 break;
1175 case BLKIF_PROTOCOL_X86_32:
1177 blkif_x86_32_sring_t *sring_x86_32 = blkdev->sring;
1179 BACK_RING_INIT(&blkdev->rings.x86_32_part, sring_x86_32, XC_PAGE_SIZE);
1180 break;
1182 case BLKIF_PROTOCOL_X86_64:
1184 blkif_x86_64_sring_t *sring_x86_64 = blkdev->sring;
1186 BACK_RING_INIT(&blkdev->rings.x86_64_part, sring_x86_64, XC_PAGE_SIZE);
1187 break;
1191 if (blkdev->feature_persistent) {
1192 /* Init persistent grants */
1193 blkdev->max_grants = max_requests * BLKIF_MAX_SEGMENTS_PER_REQUEST;
1194 blkdev->persistent_gnts = g_tree_new_full((GCompareDataFunc)int_cmp,
1195 NULL, NULL,
1196 batch_maps ?
1197 (GDestroyNotify)g_free :
1198 (GDestroyNotify)destroy_grant);
1199 blkdev->persistent_regions = NULL;
1200 blkdev->persistent_gnt_count = 0;
1203 xen_be_bind_evtchn(&blkdev->xendev);
1205 blkdev->feature_grant_copy =
1206 (xengnttab_grant_copy(blkdev->xendev.gnttabdev, 0, NULL) == 0);
1208 xen_pv_printf(&blkdev->xendev, 3, "grant copy operation %s\n",
1209 blkdev->feature_grant_copy ? "enabled" : "disabled");
1211 xen_pv_printf(&blkdev->xendev, 1, "ok: proto %s, ring-ref %d, "
1212 "remote port %d, local port %d\n",
1213 blkdev->xendev.protocol, blkdev->ring_ref,
1214 blkdev->xendev.remote_port, blkdev->xendev.local_port);
1215 return 0;
1218 static void blk_disconnect(struct XenDevice *xendev)
1220 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1222 if (blkdev->blk) {
1223 blk_detach_dev(blkdev->blk, blkdev);
1224 blk_unref(blkdev->blk);
1225 blkdev->blk = NULL;
1227 xen_pv_unbind_evtchn(&blkdev->xendev);
1229 if (blkdev->sring) {
1230 xengnttab_unmap(blkdev->xendev.gnttabdev, blkdev->sring, 1);
1231 blkdev->cnt_map--;
1232 blkdev->sring = NULL;
1236 * Unmap persistent grants before switching to the closed state
1237 * so the frontend can free them.
1239 * In the !batch_maps case g_tree_destroy will take care of unmapping
1240 * the grant, but in the batch_maps case we need to iterate over every
1241 * region in persistent_regions and unmap it.
1243 if (blkdev->feature_persistent) {
1244 g_tree_destroy(blkdev->persistent_gnts);
1245 assert(batch_maps || blkdev->persistent_gnt_count == 0);
1246 if (batch_maps) {
1247 blkdev->persistent_gnt_count = 0;
1248 g_slist_foreach(blkdev->persistent_regions,
1249 (GFunc)remove_persistent_region, blkdev);
1250 g_slist_free(blkdev->persistent_regions);
1252 blkdev->feature_persistent = false;
1256 static int blk_free(struct XenDevice *xendev)
1258 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1259 struct ioreq *ioreq;
1261 if (blkdev->blk || blkdev->sring) {
1262 blk_disconnect(xendev);
1265 while (!QLIST_EMPTY(&blkdev->freelist)) {
1266 ioreq = QLIST_FIRST(&blkdev->freelist);
1267 QLIST_REMOVE(ioreq, list);
1268 qemu_iovec_destroy(&ioreq->v);
1269 g_free(ioreq);
1272 g_free(blkdev->params);
1273 g_free(blkdev->mode);
1274 g_free(blkdev->type);
1275 g_free(blkdev->dev);
1276 g_free(blkdev->devtype);
1277 qemu_bh_delete(blkdev->bh);
1278 return 0;
1281 static void blk_event(struct XenDevice *xendev)
1283 struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1285 qemu_bh_schedule(blkdev->bh);
1288 struct XenDevOps xen_blkdev_ops = {
1289 .size = sizeof(struct XenBlkDev),
1290 .flags = DEVOPS_FLAG_NEED_GNTDEV,
1291 .alloc = blk_alloc,
1292 .init = blk_init,
1293 .initialise = blk_connect,
1294 .disconnect = blk_disconnect,
1295 .event = blk_event,
1296 .free = blk_free,