Index: head/usr.sbin/bhyve/block_if.c =================================================================== --- head/usr.sbin/bhyve/block_if.c (revision 360228) +++ head/usr.sbin/bhyve/block_if.c (revision 360229) @@ -1,851 +1,857 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2013 Peter Grehan * All rights reserved. + * Copyright 2020 Joyent, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #ifndef WITHOUT_CAPSICUM #include #endif #include #include #include #include #include #include #ifndef WITHOUT_CAPSICUM #include #endif #include #include #include #include #include #include #include #include #include #include #include #include "bhyverun.h" #include "debug.h" #include "mevent.h" #include "block_if.h" #define BLOCKIF_SIG 0xb109b109 #define BLOCKIF_NUMTHR 8 #define BLOCKIF_MAXREQ (BLOCKIF_RING_MAX + BLOCKIF_NUMTHR) enum blockop { BOP_READ, BOP_WRITE, BOP_FLUSH, BOP_DELETE }; enum blockstat { BST_FREE, BST_BLOCK, BST_PEND, BST_BUSY, BST_DONE }; struct blockif_elem { TAILQ_ENTRY(blockif_elem) be_link; struct blockif_req *be_req; enum blockop be_op; enum blockstat be_status; pthread_t be_tid; off_t be_block; }; struct blockif_ctxt { int bc_magic; int bc_fd; int bc_ischr; int bc_isgeom; int bc_candelete; int bc_rdonly; off_t bc_size; int bc_sectsz; int bc_psectsz; int bc_psectoff; int bc_closing; pthread_t bc_btid[BLOCKIF_NUMTHR]; pthread_mutex_t bc_mtx; pthread_cond_t bc_cond; /* Request elements and free/pending/busy queues */ TAILQ_HEAD(, blockif_elem) bc_freeq; TAILQ_HEAD(, blockif_elem) bc_pendq; TAILQ_HEAD(, blockif_elem) bc_busyq; struct blockif_elem bc_reqs[BLOCKIF_MAXREQ]; }; static pthread_once_t blockif_once = PTHREAD_ONCE_INIT; struct blockif_sig_elem { pthread_mutex_t bse_mtx; pthread_cond_t bse_cond; int bse_pending; struct blockif_sig_elem *bse_next; }; static struct blockif_sig_elem *blockif_bse_head; static int blockif_enqueue(struct blockif_ctxt *bc, struct blockif_req *breq, enum blockop op) { struct blockif_elem *be, *tbe; off_t off; int i; be = TAILQ_FIRST(&bc->bc_freeq); assert(be != NULL); assert(be->be_status == BST_FREE); TAILQ_REMOVE(&bc->bc_freeq, be, be_link); be->be_req = breq; be->be_op = op; switch (op) { case BOP_READ: case BOP_WRITE: case BOP_DELETE: off = breq->br_offset; for (i = 0; i < breq->br_iovcnt; i++) off += breq->br_iov[i].iov_len; break; default: off = OFF_MAX; } be->be_block = off; TAILQ_FOREACH(tbe, &bc->bc_pendq, be_link) { if (tbe->be_block == breq->br_offset) break; } if (tbe == NULL) { TAILQ_FOREACH(tbe, &bc->bc_busyq, be_link) { if (tbe->be_block == breq->br_offset) break; } } if (tbe == NULL) be->be_status = BST_PEND; else be->be_status = BST_BLOCK; TAILQ_INSERT_TAIL(&bc->bc_pendq, be, be_link); return (be->be_status == BST_PEND); } static int blockif_dequeue(struct blockif_ctxt *bc, pthread_t t, struct blockif_elem **bep) { struct blockif_elem *be; TAILQ_FOREACH(be, &bc->bc_pendq, be_link) { if (be->be_status == BST_PEND) break; assert(be->be_status == BST_BLOCK); } if (be == NULL) return (0); TAILQ_REMOVE(&bc->bc_pendq, be, be_link); be->be_status = BST_BUSY; be->be_tid = t; TAILQ_INSERT_TAIL(&bc->bc_busyq, be, be_link); *bep = be; return (1); } static void blockif_complete(struct blockif_ctxt *bc, struct blockif_elem *be) { struct blockif_elem *tbe; if (be->be_status == BST_DONE || be->be_status == BST_BUSY) TAILQ_REMOVE(&bc->bc_busyq, be, be_link); else TAILQ_REMOVE(&bc->bc_pendq, be, be_link); TAILQ_FOREACH(tbe, &bc->bc_pendq, be_link) { if (tbe->be_req->br_offset == be->be_block) tbe->be_status = BST_PEND; } be->be_tid = 0; be->be_status = BST_FREE; be->be_req = NULL; TAILQ_INSERT_TAIL(&bc->bc_freeq, be, be_link); } static void blockif_proc(struct blockif_ctxt *bc, struct blockif_elem *be, uint8_t *buf) { struct blockif_req *br; off_t arg[2]; ssize_t clen, len, off, boff, voff; int i, err; br = be->be_req; if (br->br_iovcnt <= 1) buf = NULL; err = 0; switch (be->be_op) { case BOP_READ: if (buf == NULL) { if ((len = preadv(bc->bc_fd, br->br_iov, br->br_iovcnt, br->br_offset)) < 0) err = errno; else br->br_resid -= len; break; } i = 0; off = voff = 0; while (br->br_resid > 0) { len = MIN(br->br_resid, MAXPHYS); if (pread(bc->bc_fd, buf, len, br->br_offset + off) < 0) { err = errno; break; } boff = 0; do { clen = MIN(len - boff, br->br_iov[i].iov_len - voff); memcpy(br->br_iov[i].iov_base + voff, buf + boff, clen); if (clen < br->br_iov[i].iov_len - voff) voff += clen; else { i++; voff = 0; } boff += clen; } while (boff < len); off += len; br->br_resid -= len; } break; case BOP_WRITE: if (bc->bc_rdonly) { err = EROFS; break; } if (buf == NULL) { if ((len = pwritev(bc->bc_fd, br->br_iov, br->br_iovcnt, br->br_offset)) < 0) err = errno; else br->br_resid -= len; break; } i = 0; off = voff = 0; while (br->br_resid > 0) { len = MIN(br->br_resid, MAXPHYS); boff = 0; do { clen = MIN(len - boff, br->br_iov[i].iov_len - voff); memcpy(buf + boff, br->br_iov[i].iov_base + voff, clen); if (clen < br->br_iov[i].iov_len - voff) voff += clen; else { i++; voff = 0; } boff += clen; } while (boff < len); if (pwrite(bc->bc_fd, buf, len, br->br_offset + off) < 0) { err = errno; break; } off += len; br->br_resid -= len; } break; case BOP_FLUSH: if (bc->bc_ischr) { if (ioctl(bc->bc_fd, DIOCGFLUSH)) err = errno; } else if (fsync(bc->bc_fd)) err = errno; break; case BOP_DELETE: if (!bc->bc_candelete) err = EOPNOTSUPP; else if (bc->bc_rdonly) err = EROFS; else if (bc->bc_ischr) { arg[0] = br->br_offset; arg[1] = br->br_resid; if (ioctl(bc->bc_fd, DIOCGDELETE, arg)) err = errno; else br->br_resid = 0; } else err = EOPNOTSUPP; break; default: err = EINVAL; break; } be->be_status = BST_DONE; (*br->br_callback)(br, err); } static void * blockif_thr(void *arg) { struct blockif_ctxt *bc; struct blockif_elem *be; pthread_t t; uint8_t *buf; bc = arg; if (bc->bc_isgeom) buf = malloc(MAXPHYS); else buf = NULL; t = pthread_self(); pthread_mutex_lock(&bc->bc_mtx); for (;;) { while (blockif_dequeue(bc, t, &be)) { pthread_mutex_unlock(&bc->bc_mtx); blockif_proc(bc, be, buf); pthread_mutex_lock(&bc->bc_mtx); blockif_complete(bc, be); } /* Check ctxt status here to see if exit requested */ if (bc->bc_closing) break; pthread_cond_wait(&bc->bc_cond, &bc->bc_mtx); } pthread_mutex_unlock(&bc->bc_mtx); if (buf) free(buf); pthread_exit(NULL); return (NULL); } static void blockif_sigcont_handler(int signal, enum ev_type type, void *arg) { struct blockif_sig_elem *bse; for (;;) { /* * Process the entire list even if not intended for * this thread. */ do { bse = blockif_bse_head; if (bse == NULL) return; } while (!atomic_cmpset_ptr((uintptr_t *)&blockif_bse_head, (uintptr_t)bse, (uintptr_t)bse->bse_next)); pthread_mutex_lock(&bse->bse_mtx); bse->bse_pending = 0; pthread_cond_signal(&bse->bse_cond); pthread_mutex_unlock(&bse->bse_mtx); } } static void blockif_init(void) { mevent_add(SIGCONT, EVF_SIGNAL, blockif_sigcont_handler, NULL); (void) signal(SIGCONT, SIG_IGN); } struct blockif_ctxt * blockif_open(const char *optstr, const char *ident) { char tname[MAXCOMLEN + 1]; char name[MAXPATHLEN]; char *nopt, *xopts, *cp; struct blockif_ctxt *bc; struct stat sbuf; struct diocgattr_arg arg; off_t size, psectsz, psectoff; int extra, fd, i, sectsz; int nocache, sync, ro, candelete, geom, ssopt, pssopt; + int nodelete; + #ifndef WITHOUT_CAPSICUM cap_rights_t rights; cap_ioctl_t cmds[] = { DIOCGFLUSH, DIOCGDELETE }; #endif pthread_once(&blockif_once, blockif_init); fd = -1; ssopt = 0; nocache = 0; sync = 0; ro = 0; + nodelete = 0; /* * The first element in the optstring is always a pathname. * Optional elements follow */ nopt = xopts = strdup(optstr); while (xopts != NULL) { cp = strsep(&xopts, ","); if (cp == nopt) /* file or device pathname */ continue; else if (!strcmp(cp, "nocache")) nocache = 1; + else if (!strcmp(cp, "nodelete")) + nodelete = 1; else if (!strcmp(cp, "sync") || !strcmp(cp, "direct")) sync = 1; else if (!strcmp(cp, "ro")) ro = 1; else if (sscanf(cp, "sectorsize=%d/%d", &ssopt, &pssopt) == 2) ; else if (sscanf(cp, "sectorsize=%d", &ssopt) == 1) pssopt = ssopt; else { EPRINTLN("Invalid device option \"%s\"", cp); goto err; } } extra = 0; if (nocache) extra |= O_DIRECT; if (sync) extra |= O_SYNC; fd = open(nopt, (ro ? O_RDONLY : O_RDWR) | extra); if (fd < 0 && !ro) { /* Attempt a r/w fail with a r/o open */ fd = open(nopt, O_RDONLY | extra); ro = 1; } if (fd < 0) { warn("Could not open backing file: %s", nopt); goto err; } if (fstat(fd, &sbuf) < 0) { warn("Could not stat backing file %s", nopt); goto err; } #ifndef WITHOUT_CAPSICUM cap_rights_init(&rights, CAP_FSYNC, CAP_IOCTL, CAP_READ, CAP_SEEK, CAP_WRITE); if (ro) cap_rights_clear(&rights, CAP_FSYNC, CAP_WRITE); if (caph_rights_limit(fd, &rights) == -1) errx(EX_OSERR, "Unable to apply rights for sandbox"); #endif /* * Deal with raw devices */ size = sbuf.st_size; sectsz = DEV_BSIZE; psectsz = psectoff = 0; candelete = geom = 0; if (S_ISCHR(sbuf.st_mode)) { if (ioctl(fd, DIOCGMEDIASIZE, &size) < 0 || ioctl(fd, DIOCGSECTORSIZE, §sz)) { perror("Could not fetch dev blk/sector size"); goto err; } assert(size != 0); assert(sectsz != 0); if (ioctl(fd, DIOCGSTRIPESIZE, &psectsz) == 0 && psectsz > 0) ioctl(fd, DIOCGSTRIPEOFFSET, &psectoff); strlcpy(arg.name, "GEOM::candelete", sizeof(arg.name)); arg.len = sizeof(arg.value.i); - if (ioctl(fd, DIOCGATTR, &arg) == 0) + if (nodelete == 0 && ioctl(fd, DIOCGATTR, &arg) == 0) candelete = arg.value.i; if (ioctl(fd, DIOCGPROVIDERNAME, name) == 0) geom = 1; } else psectsz = sbuf.st_blksize; #ifndef WITHOUT_CAPSICUM if (caph_ioctls_limit(fd, cmds, nitems(cmds)) == -1) errx(EX_OSERR, "Unable to apply rights for sandbox"); #endif if (ssopt != 0) { if (!powerof2(ssopt) || !powerof2(pssopt) || ssopt < 512 || ssopt > pssopt) { EPRINTLN("Invalid sector size %d/%d", ssopt, pssopt); goto err; } /* * Some backend drivers (e.g. cd0, ada0) require that the I/O * size be a multiple of the device's sector size. * * Validate that the emulated sector size complies with this * requirement. */ if (S_ISCHR(sbuf.st_mode)) { if (ssopt < sectsz || (ssopt % sectsz) != 0) { EPRINTLN("Sector size %d incompatible " "with underlying device sector size %d", ssopt, sectsz); goto err; } } sectsz = ssopt; psectsz = pssopt; psectoff = 0; } bc = calloc(1, sizeof(struct blockif_ctxt)); if (bc == NULL) { perror("calloc"); goto err; } bc->bc_magic = BLOCKIF_SIG; bc->bc_fd = fd; bc->bc_ischr = S_ISCHR(sbuf.st_mode); bc->bc_isgeom = geom; bc->bc_candelete = candelete; bc->bc_rdonly = ro; bc->bc_size = size; bc->bc_sectsz = sectsz; bc->bc_psectsz = psectsz; bc->bc_psectoff = psectoff; pthread_mutex_init(&bc->bc_mtx, NULL); pthread_cond_init(&bc->bc_cond, NULL); TAILQ_INIT(&bc->bc_freeq); TAILQ_INIT(&bc->bc_pendq); TAILQ_INIT(&bc->bc_busyq); for (i = 0; i < BLOCKIF_MAXREQ; i++) { bc->bc_reqs[i].be_status = BST_FREE; TAILQ_INSERT_HEAD(&bc->bc_freeq, &bc->bc_reqs[i], be_link); } for (i = 0; i < BLOCKIF_NUMTHR; i++) { pthread_create(&bc->bc_btid[i], NULL, blockif_thr, bc); snprintf(tname, sizeof(tname), "blk-%s-%d", ident, i); pthread_set_name_np(bc->bc_btid[i], tname); } return (bc); err: if (fd >= 0) close(fd); free(nopt); return (NULL); } static int blockif_request(struct blockif_ctxt *bc, struct blockif_req *breq, enum blockop op) { int err; err = 0; pthread_mutex_lock(&bc->bc_mtx); if (!TAILQ_EMPTY(&bc->bc_freeq)) { /* * Enqueue and inform the block i/o thread * that there is work available */ if (blockif_enqueue(bc, breq, op)) pthread_cond_signal(&bc->bc_cond); } else { /* * Callers are not allowed to enqueue more than * the specified blockif queue limit. Return an * error to indicate that the queue length has been * exceeded. */ err = E2BIG; } pthread_mutex_unlock(&bc->bc_mtx); return (err); } int blockif_read(struct blockif_ctxt *bc, struct blockif_req *breq) { assert(bc->bc_magic == BLOCKIF_SIG); return (blockif_request(bc, breq, BOP_READ)); } int blockif_write(struct blockif_ctxt *bc, struct blockif_req *breq) { assert(bc->bc_magic == BLOCKIF_SIG); return (blockif_request(bc, breq, BOP_WRITE)); } int blockif_flush(struct blockif_ctxt *bc, struct blockif_req *breq) { assert(bc->bc_magic == BLOCKIF_SIG); return (blockif_request(bc, breq, BOP_FLUSH)); } int blockif_delete(struct blockif_ctxt *bc, struct blockif_req *breq) { assert(bc->bc_magic == BLOCKIF_SIG); return (blockif_request(bc, breq, BOP_DELETE)); } int blockif_cancel(struct blockif_ctxt *bc, struct blockif_req *breq) { struct blockif_elem *be; assert(bc->bc_magic == BLOCKIF_SIG); pthread_mutex_lock(&bc->bc_mtx); /* * Check pending requests. */ TAILQ_FOREACH(be, &bc->bc_pendq, be_link) { if (be->be_req == breq) break; } if (be != NULL) { /* * Found it. */ blockif_complete(bc, be); pthread_mutex_unlock(&bc->bc_mtx); return (0); } /* * Check in-flight requests. */ TAILQ_FOREACH(be, &bc->bc_busyq, be_link) { if (be->be_req == breq) break; } if (be == NULL) { /* * Didn't find it. */ pthread_mutex_unlock(&bc->bc_mtx); return (EINVAL); } /* * Interrupt the processing thread to force it return * prematurely via it's normal callback path. */ while (be->be_status == BST_BUSY) { struct blockif_sig_elem bse, *old_head; pthread_mutex_init(&bse.bse_mtx, NULL); pthread_cond_init(&bse.bse_cond, NULL); bse.bse_pending = 1; do { old_head = blockif_bse_head; bse.bse_next = old_head; } while (!atomic_cmpset_ptr((uintptr_t *)&blockif_bse_head, (uintptr_t)old_head, (uintptr_t)&bse)); pthread_kill(be->be_tid, SIGCONT); pthread_mutex_lock(&bse.bse_mtx); while (bse.bse_pending) pthread_cond_wait(&bse.bse_cond, &bse.bse_mtx); pthread_mutex_unlock(&bse.bse_mtx); } pthread_mutex_unlock(&bc->bc_mtx); /* * The processing thread has been interrupted. Since it's not * clear if the callback has been invoked yet, return EBUSY. */ return (EBUSY); } int blockif_close(struct blockif_ctxt *bc) { void *jval; int i; assert(bc->bc_magic == BLOCKIF_SIG); /* * Stop the block i/o thread */ pthread_mutex_lock(&bc->bc_mtx); bc->bc_closing = 1; pthread_mutex_unlock(&bc->bc_mtx); pthread_cond_broadcast(&bc->bc_cond); for (i = 0; i < BLOCKIF_NUMTHR; i++) pthread_join(bc->bc_btid[i], &jval); /* XXX Cancel queued i/o's ??? */ /* * Release resources */ bc->bc_magic = 0; close(bc->bc_fd); free(bc); return (0); } /* * Return virtual C/H/S values for a given block. Use the algorithm * outlined in the VHD specification to calculate values. */ void blockif_chs(struct blockif_ctxt *bc, uint16_t *c, uint8_t *h, uint8_t *s) { off_t sectors; /* total sectors of the block dev */ off_t hcyl; /* cylinders times heads */ uint16_t secpt; /* sectors per track */ uint8_t heads; assert(bc->bc_magic == BLOCKIF_SIG); sectors = bc->bc_size / bc->bc_sectsz; /* Clamp the size to the largest possible with CHS */ if (sectors > 65535UL*16*255) sectors = 65535UL*16*255; if (sectors >= 65536UL*16*63) { secpt = 255; heads = 16; hcyl = sectors / secpt; } else { secpt = 17; hcyl = sectors / secpt; heads = (hcyl + 1023) / 1024; if (heads < 4) heads = 4; if (hcyl >= (heads * 1024) || heads > 16) { secpt = 31; heads = 16; hcyl = sectors / secpt; } if (hcyl >= (heads * 1024)) { secpt = 63; heads = 16; hcyl = sectors / secpt; } } *c = hcyl / heads; *h = heads; *s = secpt; } /* * Accessors */ off_t blockif_size(struct blockif_ctxt *bc) { assert(bc->bc_magic == BLOCKIF_SIG); return (bc->bc_size); } int blockif_sectsz(struct blockif_ctxt *bc) { assert(bc->bc_magic == BLOCKIF_SIG); return (bc->bc_sectsz); } void blockif_psectsz(struct blockif_ctxt *bc, int *size, int *off) { assert(bc->bc_magic == BLOCKIF_SIG); *size = bc->bc_psectsz; *off = bc->bc_psectoff; } int blockif_queuesz(struct blockif_ctxt *bc) { assert(bc->bc_magic == BLOCKIF_SIG); return (BLOCKIF_MAXREQ - 1); } int blockif_is_ro(struct blockif_ctxt *bc) { assert(bc->bc_magic == BLOCKIF_SIG); return (bc->bc_rdonly); } int blockif_candelete(struct blockif_ctxt *bc) { assert(bc->bc_magic == BLOCKIF_SIG); return (bc->bc_candelete); } Index: head/usr.sbin/bhyve/pci_virtio_block.c =================================================================== --- head/usr.sbin/bhyve/pci_virtio_block.c (revision 360228) +++ head/usr.sbin/bhyve/pci_virtio_block.c (revision 360229) @@ -1,425 +1,528 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2011 NetApp, Inc. * All rights reserved. - * Copyright (c) 2019 Joyent, Inc. + * Copyright 2020 Joyent, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "bhyverun.h" #include "debug.h" #include "pci_emul.h" #include "virtio.h" #include "block_if.h" -#define VTBLK_RINGSZ 128 +#define VTBLK_BSIZE 512 +#define VTBLK_RINGSZ 128 _Static_assert(VTBLK_RINGSZ <= BLOCKIF_RING_MAX, "Each ring entry must be able to queue a request"); -#define VTBLK_S_OK 0 -#define VTBLK_S_IOERR 1 +#define VTBLK_S_OK 0 +#define VTBLK_S_IOERR 1 #define VTBLK_S_UNSUPP 2 #define VTBLK_BLK_ID_BYTES 20 + 1 /* Capability bits */ -#define VTBLK_F_SEG_MAX (1 << 2) /* Maximum request segments */ -#define VTBLK_F_BLK_SIZE (1 << 6) /* cfg block size valid */ -#define VTBLK_F_FLUSH (1 << 9) /* Cache flush support */ -#define VTBLK_F_TOPOLOGY (1 << 10) /* Optimal I/O alignment */ +#define VTBLK_F_BARRIER (1 << 0) /* Does host support barriers? */ +#define VTBLK_F_SIZE_MAX (1 << 1) /* Indicates maximum segment size */ +#define VTBLK_F_SEG_MAX (1 << 2) /* Indicates maximum # of segments */ +#define VTBLK_F_GEOMETRY (1 << 4) /* Legacy geometry available */ +#define VTBLK_F_RO (1 << 5) /* Disk is read-only */ +#define VTBLK_F_BLK_SIZE (1 << 6) /* Block size of disk is available*/ +#define VTBLK_F_SCSI (1 << 7) /* Supports scsi command passthru */ +#define VTBLK_F_FLUSH (1 << 9) /* Writeback mode enabled after reset */ +#define VTBLK_F_WCE (1 << 9) /* Legacy alias for FLUSH */ +#define VTBLK_F_TOPOLOGY (1 << 10) /* Topology information is available */ +#define VTBLK_F_CONFIG_WCE (1 << 11) /* Writeback mode available in config */ +#define VTBLK_F_MQ (1 << 12) /* Multi-Queue */ +#define VTBLK_F_DISCARD (1 << 13) /* Trim blocks */ +#define VTBLK_F_WRITE_ZEROES (1 << 14) /* Write zeros */ /* * Host capabilities */ -#define VTBLK_S_HOSTCAPS \ +#define VTBLK_S_HOSTCAPS \ ( VTBLK_F_SEG_MAX | \ VTBLK_F_BLK_SIZE | \ VTBLK_F_FLUSH | \ VTBLK_F_TOPOLOGY | \ VIRTIO_RING_F_INDIRECT_DESC ) /* indirect descriptors */ /* + * The current blockif_delete() interface only allows a single delete + * request at a time. + */ +#define VTBLK_MAX_DISCARD_SEG 1 + +/* + * An arbitrary limit to prevent excessive latency due to large + * delete requests. + */ +#define VTBLK_MAX_DISCARD_SECT ((16 << 20) / VTBLK_BSIZE) /* 16 MiB */ + +/* * Config space "registers" */ struct vtblk_config { uint64_t vbc_capacity; uint32_t vbc_size_max; uint32_t vbc_seg_max; struct { uint16_t cylinders; uint8_t heads; uint8_t sectors; } vbc_geometry; uint32_t vbc_blk_size; struct { uint8_t physical_block_exp; uint8_t alignment_offset; uint16_t min_io_size; uint32_t opt_io_size; } vbc_topology; uint8_t vbc_writeback; + uint8_t unused0[1]; + uint16_t num_queues; + uint32_t max_discard_sectors; + uint32_t max_discard_seg; + uint32_t discard_sector_alignment; + uint32_t max_write_zeroes_sectors; + uint32_t max_write_zeroes_seg; + uint8_t write_zeroes_may_unmap; + uint8_t unused1[3]; } __packed; /* * Fixed-size block header */ struct virtio_blk_hdr { #define VBH_OP_READ 0 #define VBH_OP_WRITE 1 +#define VBH_OP_SCSI_CMD 2 +#define VBH_OP_SCSI_CMD_OUT 3 #define VBH_OP_FLUSH 4 #define VBH_OP_FLUSH_OUT 5 #define VBH_OP_IDENT 8 +#define VBH_OP_DISCARD 11 +#define VBH_OP_WRITE_ZEROES 13 + #define VBH_FLAG_BARRIER 0x80000000 /* OR'ed into vbh_type */ uint32_t vbh_type; uint32_t vbh_ioprio; uint64_t vbh_sector; } __packed; /* * Debug printf */ static int pci_vtblk_debug; -#define DPRINTF(params) if (pci_vtblk_debug) PRINTLN params -#define WPRINTF(params) PRINTLN params +#define DPRINTF(params) if (pci_vtblk_debug) PRINTLN params +#define WPRINTF(params) PRINTLN params struct pci_vtblk_ioreq { struct blockif_req io_req; struct pci_vtblk_softc *io_sc; uint8_t *io_status; uint16_t io_idx; }; +struct virtio_blk_discard_write_zeroes { + uint64_t sector; + uint32_t num_sectors; + struct { + uint32_t unmap:1; + uint32_t reserved:31; + } flags; +}; + /* * Per-device softc */ struct pci_vtblk_softc { struct virtio_softc vbsc_vs; pthread_mutex_t vsc_mtx; struct vqueue_info vbsc_vq; struct vtblk_config vbsc_cfg; + struct virtio_consts vbsc_consts; struct blockif_ctxt *bc; char vbsc_ident[VTBLK_BLK_ID_BYTES]; struct pci_vtblk_ioreq vbsc_ios[VTBLK_RINGSZ]; }; static void pci_vtblk_reset(void *); static void pci_vtblk_notify(void *, struct vqueue_info *); static int pci_vtblk_cfgread(void *, int, int, uint32_t *); static int pci_vtblk_cfgwrite(void *, int, int, uint32_t); static struct virtio_consts vtblk_vi_consts = { "vtblk", /* our name */ 1, /* we support 1 virtqueue */ sizeof(struct vtblk_config), /* config reg size */ pci_vtblk_reset, /* reset */ pci_vtblk_notify, /* device-wide qnotify */ pci_vtblk_cfgread, /* read PCI config */ pci_vtblk_cfgwrite, /* write PCI config */ NULL, /* apply negotiated features */ VTBLK_S_HOSTCAPS, /* our capabilities */ }; static void pci_vtblk_reset(void *vsc) { struct pci_vtblk_softc *sc = vsc; DPRINTF(("vtblk: device reset requested !")); vi_reset_dev(&sc->vbsc_vs); } static void -pci_vtblk_done(struct blockif_req *br, int err) +pci_vtblk_done_locked(struct pci_vtblk_ioreq *io, int err) { - struct pci_vtblk_ioreq *io = br->br_param; struct pci_vtblk_softc *sc = io->io_sc; /* convert errno into a virtio block error return */ if (err == EOPNOTSUPP || err == ENOSYS) *io->io_status = VTBLK_S_UNSUPP; else if (err != 0) *io->io_status = VTBLK_S_IOERR; else *io->io_status = VTBLK_S_OK; /* * Return the descriptor back to the host. * We wrote 1 byte (our status) to host. */ - pthread_mutex_lock(&sc->vsc_mtx); vq_relchain(&sc->vbsc_vq, io->io_idx, 1); vq_endchains(&sc->vbsc_vq, 0); +} + +static void +pci_vtblk_done(struct blockif_req *br, int err) +{ + struct pci_vtblk_ioreq *io = br->br_param; + struct pci_vtblk_softc *sc = io->io_sc; + + pthread_mutex_lock(&sc->vsc_mtx); + pci_vtblk_done_locked(io, err); pthread_mutex_unlock(&sc->vsc_mtx); } static void pci_vtblk_proc(struct pci_vtblk_softc *sc, struct vqueue_info *vq) { struct virtio_blk_hdr *vbh; struct pci_vtblk_ioreq *io; int i, n; int err; ssize_t iolen; int writeop, type; struct iovec iov[BLOCKIF_IOV_MAX + 2]; uint16_t idx, flags[BLOCKIF_IOV_MAX + 2]; + struct virtio_blk_discard_write_zeroes *discard; n = vq_getchain(vq, &idx, iov, BLOCKIF_IOV_MAX + 2, flags); /* * The first descriptor will be the read-only fixed header, * and the last is for status (hence +2 above and below). * The remaining iov's are the actual data I/O vectors. * * XXX - note - this fails on crash dump, which does a * VIRTIO_BLK_T_FLUSH with a zero transfer length */ assert(n >= 2 && n <= BLOCKIF_IOV_MAX + 2); io = &sc->vbsc_ios[idx]; assert((flags[0] & VRING_DESC_F_WRITE) == 0); assert(iov[0].iov_len == sizeof(struct virtio_blk_hdr)); - vbh = iov[0].iov_base; + vbh = (struct virtio_blk_hdr *)iov[0].iov_base; memcpy(&io->io_req.br_iov, &iov[1], sizeof(struct iovec) * (n - 2)); io->io_req.br_iovcnt = n - 2; - io->io_req.br_offset = vbh->vbh_sector * DEV_BSIZE; - io->io_status = iov[--n].iov_base; + io->io_req.br_offset = vbh->vbh_sector * VTBLK_BSIZE; + io->io_status = (uint8_t *)iov[--n].iov_base; assert(iov[n].iov_len == 1); assert(flags[n] & VRING_DESC_F_WRITE); /* * XXX * The guest should not be setting the BARRIER flag because * we don't advertise the capability. */ type = vbh->vbh_type & ~VBH_FLAG_BARRIER; - writeop = (type == VBH_OP_WRITE); + writeop = (type == VBH_OP_WRITE || type == VBH_OP_DISCARD); iolen = 0; for (i = 1; i < n; i++) { /* * - write op implies read-only descriptor, * - read/ident op implies write-only descriptor, * therefore test the inverse of the descriptor bit * to the op. */ assert(((flags[i] & VRING_DESC_F_WRITE) == 0) == writeop); iolen += iov[i].iov_len; } io->io_req.br_resid = iolen; DPRINTF(("virtio-block: %s op, %zd bytes, %d segs, offset %ld", - writeop ? "write" : "read/ident", iolen, i - 1, + writeop ? "write/discard" : "read/ident", iolen, i - 1, io->io_req.br_offset)); switch (type) { case VBH_OP_READ: err = blockif_read(sc->bc, &io->io_req); break; case VBH_OP_WRITE: err = blockif_write(sc->bc, &io->io_req); break; + case VBH_OP_DISCARD: + /* + * We currently only support a single request, if the guest + * has submitted a request that doesn't conform to the + * requirements, we return a error. + */ + if (iov[1].iov_len != sizeof (*discard)) { + pci_vtblk_done_locked(io, EINVAL); + return; + } + + /* The segments to discard are provided rather than data */ + discard = (struct virtio_blk_discard_write_zeroes *) + iov[1].iov_base; + + /* + * virtio v1.1 5.2.6.2: + * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP + * for discard and write zeroes commands if any unknown flag is + * set. Furthermore, the device MUST set the status byte to + * VIRTIO_BLK_S_UNSUPP for discard commands if the unmap flag + * is set. + * + * Currently there are no known flags for a DISCARD request. + */ + if (discard->flags.unmap != 0 || discard->flags.reserved != 0) { + pci_vtblk_done_locked(io, ENOTSUP); + return; + } + + /* Make sure the request doesn't exceed our size limit */ + if (discard->num_sectors > VTBLK_MAX_DISCARD_SECT) { + pci_vtblk_done_locked(io, EINVAL); + return; + } + + io->io_req.br_offset = discard->sector * VTBLK_BSIZE; + io->io_req.br_resid = discard->num_sectors * VTBLK_BSIZE; + err = blockif_delete(sc->bc, &io->io_req); + break; case VBH_OP_FLUSH: case VBH_OP_FLUSH_OUT: err = blockif_flush(sc->bc, &io->io_req); break; case VBH_OP_IDENT: /* Assume a single buffer */ /* S/n equal to buffer is not zero-terminated. */ memset(iov[1].iov_base, 0, iov[1].iov_len); strncpy(iov[1].iov_base, sc->vbsc_ident, MIN(iov[1].iov_len, sizeof(sc->vbsc_ident))); - pci_vtblk_done(&io->io_req, 0); + pci_vtblk_done_locked(io, 0); return; default: - pci_vtblk_done(&io->io_req, EOPNOTSUPP); + pci_vtblk_done_locked(io, EOPNOTSUPP); return; } assert(err == 0); } static void pci_vtblk_notify(void *vsc, struct vqueue_info *vq) { struct pci_vtblk_softc *sc = vsc; while (vq_has_descs(vq)) pci_vtblk_proc(sc, vq); } static int pci_vtblk_init(struct vmctx *ctx, struct pci_devinst *pi, char *opts) { char bident[sizeof("XX:X:X")]; struct blockif_ctxt *bctxt; MD5_CTX mdctx; u_char digest[16]; struct pci_vtblk_softc *sc; off_t size; int i, sectsz, sts, sto; if (opts == NULL) { WPRINTF(("virtio-block: backing device required")); return (1); } /* * The supplied backing file has to exist */ snprintf(bident, sizeof(bident), "%d:%d", pi->pi_slot, pi->pi_func); bctxt = blockif_open(opts, bident); if (bctxt == NULL) { perror("Could not open backing file"); return (1); } size = blockif_size(bctxt); sectsz = blockif_sectsz(bctxt); blockif_psectsz(bctxt, &sts, &sto); sc = calloc(1, sizeof(struct pci_vtblk_softc)); sc->bc = bctxt; for (i = 0; i < VTBLK_RINGSZ; i++) { struct pci_vtblk_ioreq *io = &sc->vbsc_ios[i]; io->io_req.br_callback = pci_vtblk_done; io->io_req.br_param = io; io->io_sc = sc; io->io_idx = i; } + bcopy(&vtblk_vi_consts, &sc->vbsc_consts, sizeof (vtblk_vi_consts)); + if (blockif_candelete(sc->bc)) + sc->vbsc_consts.vc_hv_caps |= VTBLK_F_DISCARD; + pthread_mutex_init(&sc->vsc_mtx, NULL); /* init virtio softc and virtqueues */ - vi_softc_linkup(&sc->vbsc_vs, &vtblk_vi_consts, sc, pi, &sc->vbsc_vq); + vi_softc_linkup(&sc->vbsc_vs, &sc->vbsc_consts, sc, pi, &sc->vbsc_vq); sc->vbsc_vs.vs_mtx = &sc->vsc_mtx; sc->vbsc_vq.vq_qsize = VTBLK_RINGSZ; /* sc->vbsc_vq.vq_notify = we have no per-queue notify */ /* * Create an identifier for the backing file. Use parts of the * md5 sum of the filename */ MD5Init(&mdctx); MD5Update(&mdctx, opts, strlen(opts)); MD5Final(digest, &mdctx); snprintf(sc->vbsc_ident, VTBLK_BLK_ID_BYTES, "BHYVE-%02X%02X-%02X%02X-%02X%02X", digest[0], digest[1], digest[2], digest[3], digest[4], digest[5]); /* setup virtio block config space */ - sc->vbsc_cfg.vbc_capacity = size / DEV_BSIZE; /* 512-byte units */ + sc->vbsc_cfg.vbc_capacity = size / VTBLK_BSIZE; /* 512-byte units */ sc->vbsc_cfg.vbc_size_max = 0; /* not negotiated */ /* * If Linux is presented with a seg_max greater than the virtio queue * size, it can stumble into situations where it violates its own * invariants and panics. For safety, we keep seg_max clamped, paying * heed to the two extra descriptors needed for the header and status * of a request. */ sc->vbsc_cfg.vbc_seg_max = MIN(VTBLK_RINGSZ - 2, BLOCKIF_IOV_MAX); sc->vbsc_cfg.vbc_geometry.cylinders = 0; /* no geometry */ sc->vbsc_cfg.vbc_geometry.heads = 0; sc->vbsc_cfg.vbc_geometry.sectors = 0; sc->vbsc_cfg.vbc_blk_size = sectsz; sc->vbsc_cfg.vbc_topology.physical_block_exp = (sts > sectsz) ? (ffsll(sts / sectsz) - 1) : 0; sc->vbsc_cfg.vbc_topology.alignment_offset = (sto != 0) ? ((sts - sto) / sectsz) : 0; sc->vbsc_cfg.vbc_topology.min_io_size = 0; sc->vbsc_cfg.vbc_topology.opt_io_size = 0; sc->vbsc_cfg.vbc_writeback = 0; + sc->vbsc_cfg.max_discard_sectors = VTBLK_MAX_DISCARD_SECT; + sc->vbsc_cfg.max_discard_seg = VTBLK_MAX_DISCARD_SEG; + sc->vbsc_cfg.discard_sector_alignment = sectsz / VTBLK_BSIZE; /* * Should we move some of this into virtio.c? Could * have the device, class, and subdev_0 as fields in * the virtio constants structure. */ pci_set_cfgdata16(pi, PCIR_DEVICE, VIRTIO_DEV_BLOCK); pci_set_cfgdata16(pi, PCIR_VENDOR, VIRTIO_VENDOR); pci_set_cfgdata8(pi, PCIR_CLASS, PCIC_STORAGE); pci_set_cfgdata16(pi, PCIR_SUBDEV_0, VIRTIO_TYPE_BLOCK); pci_set_cfgdata16(pi, PCIR_SUBVEND_0, VIRTIO_VENDOR); if (vi_intr_init(&sc->vbsc_vs, 1, fbsdrun_virtio_msix())) { blockif_close(sc->bc); free(sc); return (1); } vi_set_io_bar(&sc->vbsc_vs, 0); return (0); } static int pci_vtblk_cfgwrite(void *vsc, int offset, int size, uint32_t value) { DPRINTF(("vtblk: write to readonly reg %d", offset)); return (1); } static int pci_vtblk_cfgread(void *vsc, int offset, int size, uint32_t *retval) { struct pci_vtblk_softc *sc = vsc; void *ptr; /* our caller has already verified offset and size */ ptr = (uint8_t *)&sc->vbsc_cfg + offset; memcpy(retval, ptr, size); return (0); } struct pci_devemu pci_de_vblk = { .pe_emu = "virtio-blk", .pe_init = pci_vtblk_init, .pe_barwrite = vi_pci_write, .pe_barread = vi_pci_read }; PCI_EMUL_SET(pci_de_vblk);