rumpblk.c revision 1.1.2.2
1 1.1.2.2 mjf /* $NetBSD: rumpblk.c,v 1.1.2.2 2009/01/17 13:29:38 mjf Exp $ */ 2 1.1.2.2 mjf 3 1.1.2.2 mjf /* 4 1.1.2.2 mjf * Copyright (c) 2009 Antti Kantee. All Rights Reserved. 5 1.1.2.2 mjf * 6 1.1.2.2 mjf * Development of this software was supported by the 7 1.1.2.2 mjf * Finnish Cultural Foundation. 8 1.1.2.2 mjf * 9 1.1.2.2 mjf * Redistribution and use in source and binary forms, with or without 10 1.1.2.2 mjf * modification, are permitted provided that the following conditions 11 1.1.2.2 mjf * are met: 12 1.1.2.2 mjf * 1. Redistributions of source code must retain the above copyright 13 1.1.2.2 mjf * notice, this list of conditions and the following disclaimer. 14 1.1.2.2 mjf * 2. Redistributions in binary form must reproduce the above copyright 15 1.1.2.2 mjf * notice, this list of conditions and the following disclaimer in the 16 1.1.2.2 mjf * documentation and/or other materials provided with the distribution. 17 1.1.2.2 mjf * 18 1.1.2.2 mjf * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS 19 1.1.2.2 mjf * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 20 1.1.2.2 mjf * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 21 1.1.2.2 mjf * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 22 1.1.2.2 mjf * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 23 1.1.2.2 mjf * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 24 1.1.2.2 mjf * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25 1.1.2.2 mjf * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26 1.1.2.2 mjf * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27 1.1.2.2 mjf * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 28 1.1.2.2 mjf * SUCH DAMAGE. 29 1.1.2.2 mjf */ 30 1.1.2.2 mjf 31 1.1.2.2 mjf /* 32 1.1.2.2 mjf * Block device emulation. Presents a block device interface and 33 1.1.2.2 mjf * uses rumpuser system calls to satisfy I/O requests. 34 1.1.2.2 mjf */ 35 1.1.2.2 mjf 36 1.1.2.2 mjf #include <sys/cdefs.h> 37 1.1.2.2 mjf __KERNEL_RCSID(0, "$NetBSD: rumpblk.c,v 1.1.2.2 2009/01/17 13:29:38 mjf Exp $"); 38 1.1.2.2 mjf 39 1.1.2.2 mjf #include <sys/param.h> 40 1.1.2.2 mjf #include <sys/buf.h> 41 1.1.2.2 mjf #include <sys/conf.h> 42 1.1.2.2 mjf #include <sys/disklabel.h> 43 1.1.2.2 mjf #include <sys/fcntl.h> 44 1.1.2.2 mjf #include <sys/kmem.h> 45 1.1.2.2 mjf #include <sys/malloc.h> 46 1.1.2.2 mjf #include <sys/stat.h> 47 1.1.2.2 mjf 48 1.1.2.2 mjf #include <rump/rumpuser.h> 49 1.1.2.2 mjf 50 1.1.2.2 mjf #include "rump_private.h" 51 1.1.2.2 mjf #include "rump_vfs_private.h" 52 1.1.2.2 mjf 53 1.1.2.2 mjf #define RUMPBLK_SIZE 16 54 1.1.2.2 mjf static struct rblkdev { 55 1.1.2.2 mjf char *rblk_path; 56 1.1.2.2 mjf int rblk_fd; 57 1.1.2.2 mjf 58 1.1.2.2 mjf struct partition *rblk_curpi; 59 1.1.2.2 mjf struct partition rblk_pi; 60 1.1.2.2 mjf struct disklabel rblk_dl; 61 1.1.2.2 mjf } minors[RUMPBLK_SIZE]; 62 1.1.2.2 mjf 63 1.1.2.2 mjf dev_type_open(rumpblk_open); 64 1.1.2.2 mjf dev_type_close(rumpblk_close); 65 1.1.2.2 mjf dev_type_read(rumpblk_read); 66 1.1.2.2 mjf dev_type_write(rumpblk_write); 67 1.1.2.2 mjf dev_type_ioctl(rumpblk_ioctl); 68 1.1.2.2 mjf dev_type_strategy(rumpblk_strategy); 69 1.1.2.2 mjf dev_type_dump(rumpblk_dump); 70 1.1.2.2 mjf dev_type_size(rumpblk_size); 71 1.1.2.2 mjf 72 1.1.2.2 mjf static const struct bdevsw rumpblk_bdevsw = { 73 1.1.2.2 mjf rumpblk_open, rumpblk_close, rumpblk_strategy, rumpblk_ioctl, 74 1.1.2.2 mjf nodump, nosize, D_DISK 75 1.1.2.2 mjf }; 76 1.1.2.2 mjf 77 1.1.2.2 mjf static const struct cdevsw rumpblk_cdevsw = { 78 1.1.2.2 mjf rumpblk_open, rumpblk_close, rumpblk_read, rumpblk_write, 79 1.1.2.2 mjf rumpblk_ioctl, nostop, notty, nopoll, nommap, nokqfilter, D_DISK 80 1.1.2.2 mjf }; 81 1.1.2.2 mjf 82 1.1.2.2 mjf /* XXX: not mpsafe */ 83 1.1.2.2 mjf 84 1.1.2.2 mjf int 85 1.1.2.2 mjf rumpblk_init() 86 1.1.2.2 mjf { 87 1.1.2.2 mjf int rumpblk = RUMPBLK; 88 1.1.2.2 mjf 89 1.1.2.2 mjf return devsw_attach("rumpblk", &rumpblk_bdevsw, &rumpblk, 90 1.1.2.2 mjf &rumpblk_cdevsw, &rumpblk); 91 1.1.2.2 mjf } 92 1.1.2.2 mjf 93 1.1.2.2 mjf int 94 1.1.2.2 mjf rumpblk_register(const char *path) 95 1.1.2.2 mjf { 96 1.1.2.2 mjf size_t len; 97 1.1.2.2 mjf int i; 98 1.1.2.2 mjf 99 1.1.2.2 mjf for (i = 0; i < RUMPBLK_SIZE; i++) 100 1.1.2.2 mjf if (minors[i].rblk_path && strcmp(minors[i].rblk_path, path)==0) 101 1.1.2.2 mjf return i; 102 1.1.2.2 mjf 103 1.1.2.2 mjf for (i = 0; i < RUMPBLK_SIZE; i++) 104 1.1.2.2 mjf if (minors[i].rblk_path == NULL) 105 1.1.2.2 mjf break; 106 1.1.2.2 mjf if (i == RUMPBLK_SIZE) 107 1.1.2.2 mjf return -1; 108 1.1.2.2 mjf 109 1.1.2.2 mjf len = strlen(path); 110 1.1.2.2 mjf minors[i].rblk_path = malloc(len+1, M_TEMP, M_WAITOK); 111 1.1.2.2 mjf strcpy(minors[i].rblk_path, path); 112 1.1.2.2 mjf minors[i].rblk_fd = -1; 113 1.1.2.2 mjf return i; 114 1.1.2.2 mjf } 115 1.1.2.2 mjf 116 1.1.2.2 mjf int 117 1.1.2.2 mjf rumpblk_open(dev_t dev, int flag, int fmt, struct lwp *l) 118 1.1.2.2 mjf { 119 1.1.2.2 mjf struct rblkdev *rblk = &minors[minor(dev)]; 120 1.1.2.2 mjf struct stat sb; 121 1.1.2.2 mjf int error, fd; 122 1.1.2.2 mjf 123 1.1.2.2 mjf KASSERT(rblk->rblk_fd == -1); 124 1.1.2.2 mjf fd = rumpuser_open(rblk->rblk_path, OFLAGS(flag), &error); 125 1.1.2.2 mjf if (error) 126 1.1.2.2 mjf return error; 127 1.1.2.2 mjf 128 1.1.2.2 mjf /* 129 1.1.2.2 mjf * Setup partition info. First try the usual. */ 130 1.1.2.2 mjf if (rumpuser_ioctl(fd, DIOCGDINFO, &rblk->rblk_dl, &error) != -1) { 131 1.1.2.2 mjf /* 132 1.1.2.2 mjf * If that works, use it. We still need to guess 133 1.1.2.2 mjf * which partition we are on. 134 1.1.2.2 mjf */ 135 1.1.2.2 mjf rblk->rblk_curpi = &rblk->rblk_dl.d_partitions[0]; 136 1.1.2.2 mjf } else { 137 1.1.2.2 mjf /* 138 1.1.2.2 mjf * If that didn't work, assume were a regular file 139 1.1.2.2 mjf * and just try to fake the info the best we can. 140 1.1.2.2 mjf */ 141 1.1.2.2 mjf memset(&rblk->rblk_dl, 0, sizeof(rblk->rblk_dl)); 142 1.1.2.2 mjf 143 1.1.2.2 mjf if (rumpuser_stat(rblk->rblk_path, &sb, &error) == -1) { 144 1.1.2.2 mjf int dummy; 145 1.1.2.2 mjf 146 1.1.2.2 mjf rumpuser_close(fd, &dummy); 147 1.1.2.2 mjf return error; 148 1.1.2.2 mjf } 149 1.1.2.2 mjf rblk->rblk_pi.p_size = sb.st_size >> DEV_BSHIFT; 150 1.1.2.2 mjf rblk->rblk_dl.d_secsize = DEV_BSIZE; 151 1.1.2.2 mjf rblk->rblk_curpi = &rblk->rblk_pi; 152 1.1.2.2 mjf } 153 1.1.2.2 mjf rblk->rblk_fd = fd; 154 1.1.2.2 mjf 155 1.1.2.2 mjf return 0; 156 1.1.2.2 mjf } 157 1.1.2.2 mjf 158 1.1.2.2 mjf int 159 1.1.2.2 mjf rumpblk_close(dev_t dev, int flag, int fmt, struct lwp *l) 160 1.1.2.2 mjf { 161 1.1.2.2 mjf struct rblkdev *rblk = &minors[minor(dev)]; 162 1.1.2.2 mjf int dummy; 163 1.1.2.2 mjf 164 1.1.2.2 mjf rumpuser_close(rblk->rblk_fd, &dummy); 165 1.1.2.2 mjf rblk->rblk_fd = -1; 166 1.1.2.2 mjf 167 1.1.2.2 mjf return 0; 168 1.1.2.2 mjf } 169 1.1.2.2 mjf 170 1.1.2.2 mjf int 171 1.1.2.2 mjf rumpblk_ioctl(dev_t dev, u_long xfer, void *addr, int flag, struct lwp *l) 172 1.1.2.2 mjf { 173 1.1.2.2 mjf struct rblkdev *rblk = &minors[minor(dev)]; 174 1.1.2.2 mjf int rv, error; 175 1.1.2.2 mjf 176 1.1.2.2 mjf if (xfer == DIOCGPART) { 177 1.1.2.2 mjf struct partinfo *pi = (struct partinfo *)addr; 178 1.1.2.2 mjf 179 1.1.2.2 mjf pi->part = rblk->rblk_curpi; 180 1.1.2.2 mjf pi->disklab = &rblk->rblk_dl; 181 1.1.2.2 mjf 182 1.1.2.2 mjf return 0; 183 1.1.2.2 mjf } 184 1.1.2.2 mjf 185 1.1.2.2 mjf rv = rumpuser_ioctl(rblk->rblk_fd, xfer, addr, &error); 186 1.1.2.2 mjf if (rv == -1) 187 1.1.2.2 mjf return error; 188 1.1.2.2 mjf 189 1.1.2.2 mjf return 0; 190 1.1.2.2 mjf } 191 1.1.2.2 mjf 192 1.1.2.2 mjf int 193 1.1.2.2 mjf rumpblk_read(dev_t dev, struct uio *uio, int flags) 194 1.1.2.2 mjf { 195 1.1.2.2 mjf 196 1.1.2.2 mjf panic("%s: unimplemented", __func__); 197 1.1.2.2 mjf } 198 1.1.2.2 mjf 199 1.1.2.2 mjf int 200 1.1.2.2 mjf rumpblk_write(dev_t dev, struct uio *uio, int flags) 201 1.1.2.2 mjf { 202 1.1.2.2 mjf 203 1.1.2.2 mjf panic("%s: unimplemented", __func__); 204 1.1.2.2 mjf } 205 1.1.2.2 mjf 206 1.1.2.2 mjf void 207 1.1.2.2 mjf rumpblk_strategy(struct buf *bp) 208 1.1.2.2 mjf { 209 1.1.2.2 mjf struct rblkdev *rblk = &minors[minor(bp->b_dev)]; 210 1.1.2.2 mjf off_t off; 211 1.1.2.2 mjf int async; 212 1.1.2.2 mjf 213 1.1.2.2 mjf off = bp->b_blkno << DEV_BSHIFT; 214 1.1.2.2 mjf DPRINTF(("rumpblk_strategy: 0x%x bytes %s off 0x%" PRIx64 215 1.1.2.2 mjf " (0x%" PRIx64 " - 0x%" PRIx64")\n", 216 1.1.2.2 mjf bp->b_bcount, BUF_ISREAD(bp) "READ" : "WRITE", 217 1.1.2.2 mjf off, off, (off + bp->b_bcount))); 218 1.1.2.2 mjf 219 1.1.2.2 mjf /* 220 1.1.2.2 mjf * Do I/O. We have different paths for async and sync I/O. 221 1.1.2.2 mjf * Async I/O is done by passing a request to rumpuser where 222 1.1.2.2 mjf * it is executed. The rumpuser routine then calls 223 1.1.2.2 mjf * biodone() to signal any waiters in the kernel. I/O's are 224 1.1.2.2 mjf * executed in series. Technically executing them in parallel 225 1.1.2.2 mjf * would produce better results, but then we'd need either 226 1.1.2.2 mjf * more threads or posix aio. Maybe worth investigating 227 1.1.2.2 mjf * this later. 228 1.1.2.2 mjf * 229 1.1.2.2 mjf * Synchronous I/O is done directly in the context mainly to 230 1.1.2.2 mjf * avoid unnecessary scheduling with the I/O thread. 231 1.1.2.2 mjf */ 232 1.1.2.2 mjf async = bp->b_flags & B_ASYNC; 233 1.1.2.2 mjf if (async && rump_threads) { 234 1.1.2.2 mjf struct rumpuser_aio *rua; 235 1.1.2.2 mjf 236 1.1.2.2 mjf rua = kmem_alloc(sizeof(struct rumpuser_aio), KM_SLEEP); 237 1.1.2.2 mjf rua->rua_fd = rblk->rblk_fd; 238 1.1.2.2 mjf rua->rua_data = bp->b_data; 239 1.1.2.2 mjf rua->rua_dlen = bp->b_bcount; 240 1.1.2.2 mjf rua->rua_off = off; 241 1.1.2.2 mjf rua->rua_bp = bp; 242 1.1.2.2 mjf rua->rua_op = BUF_ISREAD(bp); 243 1.1.2.2 mjf 244 1.1.2.2 mjf rumpuser_mutex_enter(&rumpuser_aio_mtx); 245 1.1.2.2 mjf 246 1.1.2.2 mjf /* 247 1.1.2.2 mjf * Check if our buffer is full. Doing it this way 248 1.1.2.2 mjf * throttles the I/O a bit if we have a massive 249 1.1.2.2 mjf * async I/O burst. 250 1.1.2.2 mjf * 251 1.1.2.2 mjf * XXX: this actually leads to deadlocks with spl() 252 1.1.2.2 mjf * (caller maybe be at splbio() legally for async I/O), 253 1.1.2.2 mjf * so for now set N_AIOS high and FIXXXME some day. 254 1.1.2.2 mjf */ 255 1.1.2.2 mjf if ((rumpuser_aio_head+1) % N_AIOS == rumpuser_aio_tail) { 256 1.1.2.2 mjf kmem_free(rua, sizeof(*rua)); 257 1.1.2.2 mjf rumpuser_mutex_exit(&rumpuser_aio_mtx); 258 1.1.2.2 mjf goto syncfallback; 259 1.1.2.2 mjf } 260 1.1.2.2 mjf 261 1.1.2.2 mjf /* insert into queue & signal */ 262 1.1.2.2 mjf rumpuser_aios[rumpuser_aio_head] = rua; 263 1.1.2.2 mjf rumpuser_aio_head = (rumpuser_aio_head+1) % (N_AIOS-1); 264 1.1.2.2 mjf rumpuser_cv_signal(&rumpuser_aio_cv); 265 1.1.2.2 mjf rumpuser_mutex_exit(&rumpuser_aio_mtx); 266 1.1.2.2 mjf } else { 267 1.1.2.2 mjf syncfallback: 268 1.1.2.2 mjf if (BUF_ISREAD(bp)) { 269 1.1.2.2 mjf rumpuser_read_bio(rblk->rblk_fd, bp->b_data, 270 1.1.2.2 mjf bp->b_bcount, off, rump_biodone, bp); 271 1.1.2.2 mjf } else { 272 1.1.2.2 mjf rumpuser_write_bio(rblk->rblk_fd, bp->b_data, 273 1.1.2.2 mjf bp->b_bcount, off, rump_biodone, bp); 274 1.1.2.2 mjf } 275 1.1.2.2 mjf if (!async) { 276 1.1.2.2 mjf int error; 277 1.1.2.2 mjf 278 1.1.2.2 mjf if (BUF_ISWRITE(bp)) 279 1.1.2.2 mjf rumpuser_fsync(rblk->rblk_fd, &error); 280 1.1.2.2 mjf } 281 1.1.2.2 mjf } 282 1.1.2.2 mjf } 283
Indexes created Fri Oct 03 02:09:57 GMT 2025