1 /* 2 * Copyright (c) 2014-2016 Christoph Hellwig. 3 */ 4 #include <linux/exportfs.h> 5 #include <linux/iomap.h> 6 #include <linux/genhd.h> 7 #include <linux/slab.h> 8 #include <linux/pr.h> 9 10 #include <linux/nfsd/debug.h> 11 #include <scsi/scsi_proto.h> 12 #include <scsi/scsi_common.h> 13 14 #include "blocklayoutxdr.h" 15 #include "pnfs.h" 16 17 #define NFSDDBG_FACILITY NFSDDBG_PNFS 18 19 20 static __be32 21 nfsd4_block_proc_layoutget(struct inode *inode, const struct svc_fh *fhp, 22 struct nfsd4_layoutget *args) 23 { 24 struct nfsd4_layout_seg *seg = &args->lg_seg; 25 struct super_block *sb = inode->i_sb; 26 u32 block_size = (1 << inode->i_blkbits); 27 struct pnfs_block_extent *bex; 28 struct iomap iomap; 29 u32 device_generation = 0; 30 int error; 31 32 if (seg->offset & (block_size - 1)) { 33 dprintk("pnfsd: I/O misaligned\n"); 34 goto out_layoutunavailable; 35 } 36 37 /* 38 * Some clients barf on non-zero block numbers for NONE or INVALID 39 * layouts, so make sure to zero the whole structure. 40 */ 41 error = -ENOMEM; 42 bex = kzalloc(sizeof(*bex), GFP_KERNEL); 43 if (!bex) 44 goto out_error; 45 args->lg_content = bex; 46 47 error = sb->s_export_op->map_blocks(inode, seg->offset, seg->length, 48 &iomap, seg->iomode != IOMODE_READ, 49 &device_generation); 50 if (error) { 51 if (error == -ENXIO) 52 goto out_layoutunavailable; 53 goto out_error; 54 } 55 56 if (iomap.length < args->lg_minlength) { 57 dprintk("pnfsd: extent smaller than minlength\n"); 58 goto out_layoutunavailable; 59 } 60 61 switch (iomap.type) { 62 case IOMAP_MAPPED: 63 if (seg->iomode == IOMODE_READ) 64 bex->es = PNFS_BLOCK_READ_DATA; 65 else 66 bex->es = PNFS_BLOCK_READWRITE_DATA; 67 bex->soff = (iomap.blkno << 9); 68 break; 69 case IOMAP_UNWRITTEN: 70 if (seg->iomode & IOMODE_RW) { 71 /* 72 * Crack monkey special case from section 2.3.1. 73 */ 74 if (args->lg_minlength == 0) { 75 dprintk("pnfsd: no soup for you!\n"); 76 goto out_layoutunavailable; 77 } 78 79 bex->es = PNFS_BLOCK_INVALID_DATA; 80 bex->soff = (iomap.blkno << 9); 81 break; 82 } 83 /*FALLTHRU*/ 84 case IOMAP_HOLE: 85 if (seg->iomode == IOMODE_READ) { 86 bex->es = PNFS_BLOCK_NONE_DATA; 87 break; 88 } 89 /*FALLTHRU*/ 90 case IOMAP_DELALLOC: 91 default: 92 WARN(1, "pnfsd: filesystem returned %d extent\n", iomap.type); 93 goto out_layoutunavailable; 94 } 95 96 error = nfsd4_set_deviceid(&bex->vol_id, fhp, device_generation); 97 if (error) 98 goto out_error; 99 bex->foff = iomap.offset; 100 bex->len = iomap.length; 101 102 seg->offset = iomap.offset; 103 seg->length = iomap.length; 104 105 dprintk("GET: 0x%llx:0x%llx %d\n", bex->foff, bex->len, bex->es); 106 return 0; 107 108 out_error: 109 seg->length = 0; 110 return nfserrno(error); 111 out_layoutunavailable: 112 seg->length = 0; 113 return nfserr_layoutunavailable; 114 } 115 116 static __be32 117 nfsd4_block_commit_blocks(struct inode *inode, struct nfsd4_layoutcommit *lcp, 118 struct iomap *iomaps, int nr_iomaps) 119 { 120 loff_t new_size = lcp->lc_last_wr + 1; 121 struct iattr iattr = { .ia_valid = 0 }; 122 int error; 123 124 if (lcp->lc_mtime.tv_nsec == UTIME_NOW || 125 timespec_compare(&lcp->lc_mtime, &inode->i_mtime) < 0) 126 lcp->lc_mtime = current_fs_time(inode->i_sb); 127 iattr.ia_valid |= ATTR_ATIME | ATTR_CTIME | ATTR_MTIME; 128 iattr.ia_atime = iattr.ia_ctime = iattr.ia_mtime = lcp->lc_mtime; 129 130 if (new_size > i_size_read(inode)) { 131 iattr.ia_valid |= ATTR_SIZE; 132 iattr.ia_size = new_size; 133 } 134 135 error = inode->i_sb->s_export_op->commit_blocks(inode, iomaps, 136 nr_iomaps, &iattr); 137 kfree(iomaps); 138 return nfserrno(error); 139 } 140 141 #ifdef CONFIG_NFSD_BLOCKLAYOUT 142 static int 143 nfsd4_block_get_device_info_simple(struct super_block *sb, 144 struct nfsd4_getdeviceinfo *gdp) 145 { 146 struct pnfs_block_deviceaddr *dev; 147 struct pnfs_block_volume *b; 148 149 dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) + 150 sizeof(struct pnfs_block_volume), GFP_KERNEL); 151 if (!dev) 152 return -ENOMEM; 153 gdp->gd_device = dev; 154 155 dev->nr_volumes = 1; 156 b = &dev->volumes[0]; 157 158 b->type = PNFS_BLOCK_VOLUME_SIMPLE; 159 b->simple.sig_len = PNFS_BLOCK_UUID_LEN; 160 return sb->s_export_op->get_uuid(sb, b->simple.sig, &b->simple.sig_len, 161 &b->simple.offset); 162 } 163 164 static __be32 165 nfsd4_block_proc_getdeviceinfo(struct super_block *sb, 166 struct svc_rqst *rqstp, 167 struct nfs4_client *clp, 168 struct nfsd4_getdeviceinfo *gdp) 169 { 170 if (sb->s_bdev != sb->s_bdev->bd_contains) 171 return nfserr_inval; 172 return nfserrno(nfsd4_block_get_device_info_simple(sb, gdp)); 173 } 174 175 static __be32 176 nfsd4_block_proc_layoutcommit(struct inode *inode, 177 struct nfsd4_layoutcommit *lcp) 178 { 179 struct iomap *iomaps; 180 int nr_iomaps; 181 182 nr_iomaps = nfsd4_block_decode_layoutupdate(lcp->lc_up_layout, 183 lcp->lc_up_len, &iomaps, 1 << inode->i_blkbits); 184 if (nr_iomaps < 0) 185 return nfserrno(nr_iomaps); 186 187 return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps); 188 } 189 190 const struct nfsd4_layout_ops bl_layout_ops = { 191 /* 192 * Pretend that we send notification to the client. This is a blatant 193 * lie to force recent Linux clients to cache our device IDs. 194 * We rarely ever change the device ID, so the harm of leaking deviceids 195 * for a while isn't too bad. Unfortunately RFC5661 is a complete mess 196 * in this regard, but I filed errata 4119 for this a while ago, and 197 * hopefully the Linux client will eventually start caching deviceids 198 * without this again. 199 */ 200 .notify_types = 201 NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE, 202 .proc_getdeviceinfo = nfsd4_block_proc_getdeviceinfo, 203 .encode_getdeviceinfo = nfsd4_block_encode_getdeviceinfo, 204 .proc_layoutget = nfsd4_block_proc_layoutget, 205 .encode_layoutget = nfsd4_block_encode_layoutget, 206 .proc_layoutcommit = nfsd4_block_proc_layoutcommit, 207 }; 208 #endif /* CONFIG_NFSD_BLOCKLAYOUT */ 209 210 #ifdef CONFIG_NFSD_SCSILAYOUT 211 static int nfsd4_scsi_identify_device(struct block_device *bdev, 212 struct pnfs_block_volume *b) 213 { 214 struct request_queue *q = bdev->bd_disk->queue; 215 struct request *rq; 216 size_t bufflen = 252, len, id_len; 217 u8 *buf, *d, type, assoc; 218 int error; 219 220 buf = kzalloc(bufflen, GFP_KERNEL); 221 if (!buf) 222 return -ENOMEM; 223 224 rq = blk_get_request(q, READ, GFP_KERNEL); 225 if (IS_ERR(rq)) { 226 error = -ENOMEM; 227 goto out_free_buf; 228 } 229 blk_rq_set_block_pc(rq); 230 231 error = blk_rq_map_kern(q, rq, buf, bufflen, GFP_KERNEL); 232 if (error) 233 goto out_put_request; 234 235 rq->cmd[0] = INQUIRY; 236 rq->cmd[1] = 1; 237 rq->cmd[2] = 0x83; 238 rq->cmd[3] = bufflen >> 8; 239 rq->cmd[4] = bufflen & 0xff; 240 rq->cmd_len = COMMAND_SIZE(INQUIRY); 241 242 error = blk_execute_rq(rq->q, NULL, rq, 1); 243 if (error) { 244 pr_err("pNFS: INQUIRY 0x83 failed with: %x\n", 245 rq->errors); 246 goto out_put_request; 247 } 248 249 len = (buf[2] << 8) + buf[3] + 4; 250 if (len > bufflen) { 251 pr_err("pNFS: INQUIRY 0x83 response invalid (len = %zd)\n", 252 len); 253 goto out_put_request; 254 } 255 256 d = buf + 4; 257 for (d = buf + 4; d < buf + len; d += id_len + 4) { 258 id_len = d[3]; 259 type = d[1] & 0xf; 260 assoc = (d[1] >> 4) & 0x3; 261 262 /* 263 * We only care about a EUI-64 and NAA designator types 264 * with LU association. 265 */ 266 if (assoc != 0x00) 267 continue; 268 if (type != 0x02 && type != 0x03) 269 continue; 270 if (id_len != 8 && id_len != 12 && id_len != 16) 271 continue; 272 273 b->scsi.code_set = PS_CODE_SET_BINARY; 274 b->scsi.designator_type = type == 0x02 ? 275 PS_DESIGNATOR_EUI64 : PS_DESIGNATOR_NAA; 276 b->scsi.designator_len = id_len; 277 memcpy(b->scsi.designator, d + 4, id_len); 278 279 /* 280 * If we found a 8 or 12 byte descriptor continue on to 281 * see if a 16 byte one is available. If we find a 282 * 16 byte descriptor we're done. 283 */ 284 if (id_len == 16) 285 break; 286 } 287 288 out_put_request: 289 blk_put_request(rq); 290 out_free_buf: 291 kfree(buf); 292 return error; 293 } 294 295 #define NFSD_MDS_PR_KEY 0x0100000000000000ULL 296 297 /* 298 * We use the client ID as a unique key for the reservations. 299 * This allows us to easily fence a client when recalls fail. 300 */ 301 static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp) 302 { 303 return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id; 304 } 305 306 static int 307 nfsd4_block_get_device_info_scsi(struct super_block *sb, 308 struct nfs4_client *clp, 309 struct nfsd4_getdeviceinfo *gdp) 310 { 311 struct pnfs_block_deviceaddr *dev; 312 struct pnfs_block_volume *b; 313 const struct pr_ops *ops; 314 int error; 315 316 dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) + 317 sizeof(struct pnfs_block_volume), GFP_KERNEL); 318 if (!dev) 319 return -ENOMEM; 320 gdp->gd_device = dev; 321 322 dev->nr_volumes = 1; 323 b = &dev->volumes[0]; 324 325 b->type = PNFS_BLOCK_VOLUME_SCSI; 326 b->scsi.pr_key = nfsd4_scsi_pr_key(clp); 327 328 error = nfsd4_scsi_identify_device(sb->s_bdev, b); 329 if (error) 330 return error; 331 332 ops = sb->s_bdev->bd_disk->fops->pr_ops; 333 if (!ops) { 334 pr_err("pNFS: device %s does not support PRs.\n", 335 sb->s_id); 336 return -EINVAL; 337 } 338 339 error = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true); 340 if (error) { 341 pr_err("pNFS: failed to register key for device %s.\n", 342 sb->s_id); 343 return -EINVAL; 344 } 345 346 error = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY, 347 PR_EXCLUSIVE_ACCESS_REG_ONLY, 0); 348 if (error) { 349 pr_err("pNFS: failed to reserve device %s.\n", 350 sb->s_id); 351 return -EINVAL; 352 } 353 354 return 0; 355 } 356 357 static __be32 358 nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb, 359 struct svc_rqst *rqstp, 360 struct nfs4_client *clp, 361 struct nfsd4_getdeviceinfo *gdp) 362 { 363 if (sb->s_bdev != sb->s_bdev->bd_contains) 364 return nfserr_inval; 365 return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp)); 366 } 367 static __be32 368 nfsd4_scsi_proc_layoutcommit(struct inode *inode, 369 struct nfsd4_layoutcommit *lcp) 370 { 371 struct iomap *iomaps; 372 int nr_iomaps; 373 374 nr_iomaps = nfsd4_scsi_decode_layoutupdate(lcp->lc_up_layout, 375 lcp->lc_up_len, &iomaps, 1 << inode->i_blkbits); 376 if (nr_iomaps < 0) 377 return nfserrno(nr_iomaps); 378 379 return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps); 380 } 381 382 static void 383 nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls) 384 { 385 struct nfs4_client *clp = ls->ls_stid.sc_client; 386 struct block_device *bdev = ls->ls_file->f_path.mnt->mnt_sb->s_bdev; 387 388 bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY, 389 nfsd4_scsi_pr_key(clp), 0, true); 390 } 391 392 const struct nfsd4_layout_ops scsi_layout_ops = { 393 /* 394 * Pretend that we send notification to the client. This is a blatant 395 * lie to force recent Linux clients to cache our device IDs. 396 * We rarely ever change the device ID, so the harm of leaking deviceids 397 * for a while isn't too bad. Unfortunately RFC5661 is a complete mess 398 * in this regard, but I filed errata 4119 for this a while ago, and 399 * hopefully the Linux client will eventually start caching deviceids 400 * without this again. 401 */ 402 .notify_types = 403 NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE, 404 .proc_getdeviceinfo = nfsd4_scsi_proc_getdeviceinfo, 405 .encode_getdeviceinfo = nfsd4_block_encode_getdeviceinfo, 406 .proc_layoutget = nfsd4_block_proc_layoutget, 407 .encode_layoutget = nfsd4_block_encode_layoutget, 408 .proc_layoutcommit = nfsd4_scsi_proc_layoutcommit, 409 .fence_client = nfsd4_scsi_fence_client, 410 }; 411 #endif /* CONFIG_NFSD_SCSILAYOUT */ 412