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 #include <scsi/scsi_request.h> 14 15 #include "blocklayoutxdr.h" 16 #include "pnfs.h" 17 18 #define NFSDDBG_FACILITY NFSDDBG_PNFS 19 20 21 static __be32 22 nfsd4_block_proc_layoutget(struct inode *inode, const struct svc_fh *fhp, 23 struct nfsd4_layoutget *args) 24 { 25 struct nfsd4_layout_seg *seg = &args->lg_seg; 26 struct super_block *sb = inode->i_sb; 27 u32 block_size = i_blocksize(inode); 28 struct pnfs_block_extent *bex; 29 struct iomap iomap; 30 u32 device_generation = 0; 31 int error; 32 33 if (seg->offset & (block_size - 1)) { 34 dprintk("pnfsd: I/O misaligned\n"); 35 goto out_layoutunavailable; 36 } 37 38 /* 39 * Some clients barf on non-zero block numbers for NONE or INVALID 40 * layouts, so make sure to zero the whole structure. 41 */ 42 error = -ENOMEM; 43 bex = kzalloc(sizeof(*bex), GFP_KERNEL); 44 if (!bex) 45 goto out_error; 46 args->lg_content = bex; 47 48 error = sb->s_export_op->map_blocks(inode, seg->offset, seg->length, 49 &iomap, seg->iomode != IOMODE_READ, 50 &device_generation); 51 if (error) { 52 if (error == -ENXIO) 53 goto out_layoutunavailable; 54 goto out_error; 55 } 56 57 if (iomap.length < args->lg_minlength) { 58 dprintk("pnfsd: extent smaller than minlength\n"); 59 goto out_layoutunavailable; 60 } 61 62 switch (iomap.type) { 63 case IOMAP_MAPPED: 64 if (seg->iomode == IOMODE_READ) 65 bex->es = PNFS_BLOCK_READ_DATA; 66 else 67 bex->es = PNFS_BLOCK_READWRITE_DATA; 68 bex->soff = (iomap.blkno << 9); 69 break; 70 case IOMAP_UNWRITTEN: 71 if (seg->iomode & IOMODE_RW) { 72 /* 73 * Crack monkey special case from section 2.3.1. 74 */ 75 if (args->lg_minlength == 0) { 76 dprintk("pnfsd: no soup for you!\n"); 77 goto out_layoutunavailable; 78 } 79 80 bex->es = PNFS_BLOCK_INVALID_DATA; 81 bex->soff = (iomap.blkno << 9); 82 break; 83 } 84 /*FALLTHRU*/ 85 case IOMAP_HOLE: 86 if (seg->iomode == IOMODE_READ) { 87 bex->es = PNFS_BLOCK_NONE_DATA; 88 break; 89 } 90 /*FALLTHRU*/ 91 case IOMAP_DELALLOC: 92 default: 93 WARN(1, "pnfsd: filesystem returned %d extent\n", iomap.type); 94 goto out_layoutunavailable; 95 } 96 97 error = nfsd4_set_deviceid(&bex->vol_id, fhp, device_generation); 98 if (error) 99 goto out_error; 100 bex->foff = iomap.offset; 101 bex->len = iomap.length; 102 103 seg->offset = iomap.offset; 104 seg->length = iomap.length; 105 106 dprintk("GET: 0x%llx:0x%llx %d\n", bex->foff, bex->len, bex->es); 107 return 0; 108 109 out_error: 110 seg->length = 0; 111 return nfserrno(error); 112 out_layoutunavailable: 113 seg->length = 0; 114 return nfserr_layoutunavailable; 115 } 116 117 static __be32 118 nfsd4_block_commit_blocks(struct inode *inode, struct nfsd4_layoutcommit *lcp, 119 struct iomap *iomaps, int nr_iomaps) 120 { 121 loff_t new_size = lcp->lc_last_wr + 1; 122 struct iattr iattr = { .ia_valid = 0 }; 123 int error; 124 125 if (lcp->lc_mtime.tv_nsec == UTIME_NOW || 126 timespec_compare(&lcp->lc_mtime, &inode->i_mtime) < 0) 127 lcp->lc_mtime = current_time(inode); 128 iattr.ia_valid |= ATTR_ATIME | ATTR_CTIME | ATTR_MTIME; 129 iattr.ia_atime = iattr.ia_ctime = iattr.ia_mtime = lcp->lc_mtime; 130 131 if (new_size > i_size_read(inode)) { 132 iattr.ia_valid |= ATTR_SIZE; 133 iattr.ia_size = new_size; 134 } 135 136 error = inode->i_sb->s_export_op->commit_blocks(inode, iomaps, 137 nr_iomaps, &iattr); 138 kfree(iomaps); 139 return nfserrno(error); 140 } 141 142 #ifdef CONFIG_NFSD_BLOCKLAYOUT 143 static int 144 nfsd4_block_get_device_info_simple(struct super_block *sb, 145 struct nfsd4_getdeviceinfo *gdp) 146 { 147 struct pnfs_block_deviceaddr *dev; 148 struct pnfs_block_volume *b; 149 150 dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) + 151 sizeof(struct pnfs_block_volume), GFP_KERNEL); 152 if (!dev) 153 return -ENOMEM; 154 gdp->gd_device = dev; 155 156 dev->nr_volumes = 1; 157 b = &dev->volumes[0]; 158 159 b->type = PNFS_BLOCK_VOLUME_SIMPLE; 160 b->simple.sig_len = PNFS_BLOCK_UUID_LEN; 161 return sb->s_export_op->get_uuid(sb, b->simple.sig, &b->simple.sig_len, 162 &b->simple.offset); 163 } 164 165 static __be32 166 nfsd4_block_proc_getdeviceinfo(struct super_block *sb, 167 struct svc_rqst *rqstp, 168 struct nfs4_client *clp, 169 struct nfsd4_getdeviceinfo *gdp) 170 { 171 if (sb->s_bdev != sb->s_bdev->bd_contains) 172 return nfserr_inval; 173 return nfserrno(nfsd4_block_get_device_info_simple(sb, gdp)); 174 } 175 176 static __be32 177 nfsd4_block_proc_layoutcommit(struct inode *inode, 178 struct nfsd4_layoutcommit *lcp) 179 { 180 struct iomap *iomaps; 181 int nr_iomaps; 182 183 nr_iomaps = nfsd4_block_decode_layoutupdate(lcp->lc_up_layout, 184 lcp->lc_up_len, &iomaps, i_blocksize(inode)); 185 if (nr_iomaps < 0) 186 return nfserrno(nr_iomaps); 187 188 return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps); 189 } 190 191 const struct nfsd4_layout_ops bl_layout_ops = { 192 /* 193 * Pretend that we send notification to the client. This is a blatant 194 * lie to force recent Linux clients to cache our device IDs. 195 * We rarely ever change the device ID, so the harm of leaking deviceids 196 * for a while isn't too bad. Unfortunately RFC5661 is a complete mess 197 * in this regard, but I filed errata 4119 for this a while ago, and 198 * hopefully the Linux client will eventually start caching deviceids 199 * without this again. 200 */ 201 .notify_types = 202 NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE, 203 .proc_getdeviceinfo = nfsd4_block_proc_getdeviceinfo, 204 .encode_getdeviceinfo = nfsd4_block_encode_getdeviceinfo, 205 .proc_layoutget = nfsd4_block_proc_layoutget, 206 .encode_layoutget = nfsd4_block_encode_layoutget, 207 .proc_layoutcommit = nfsd4_block_proc_layoutcommit, 208 }; 209 #endif /* CONFIG_NFSD_BLOCKLAYOUT */ 210 211 #ifdef CONFIG_NFSD_SCSILAYOUT 212 static int nfsd4_scsi_identify_device(struct block_device *bdev, 213 struct pnfs_block_volume *b) 214 { 215 struct request_queue *q = bdev->bd_disk->queue; 216 struct request *rq; 217 struct scsi_request *req; 218 size_t bufflen = 252, len, id_len; 219 u8 *buf, *d, type, assoc; 220 int error; 221 222 if (WARN_ON_ONCE(!blk_queue_scsi_passthrough(q))) 223 return -EINVAL; 224 225 buf = kzalloc(bufflen, GFP_KERNEL); 226 if (!buf) 227 return -ENOMEM; 228 229 rq = blk_get_request(q, REQ_OP_SCSI_IN, GFP_KERNEL); 230 if (IS_ERR(rq)) { 231 error = -ENOMEM; 232 goto out_free_buf; 233 } 234 req = scsi_req(rq); 235 236 error = blk_rq_map_kern(q, rq, buf, bufflen, GFP_KERNEL); 237 if (error) 238 goto out_put_request; 239 240 req->cmd[0] = INQUIRY; 241 req->cmd[1] = 1; 242 req->cmd[2] = 0x83; 243 req->cmd[3] = bufflen >> 8; 244 req->cmd[4] = bufflen & 0xff; 245 req->cmd_len = COMMAND_SIZE(INQUIRY); 246 247 blk_execute_rq(rq->q, NULL, rq, 1); 248 if (req->result) { 249 pr_err("pNFS: INQUIRY 0x83 failed with: %x\n", 250 req->result); 251 error = -EIO; 252 goto out_put_request; 253 } 254 255 len = (buf[2] << 8) + buf[3] + 4; 256 if (len > bufflen) { 257 pr_err("pNFS: INQUIRY 0x83 response invalid (len = %zd)\n", 258 len); 259 goto out_put_request; 260 } 261 262 d = buf + 4; 263 for (d = buf + 4; d < buf + len; d += id_len + 4) { 264 id_len = d[3]; 265 type = d[1] & 0xf; 266 assoc = (d[1] >> 4) & 0x3; 267 268 /* 269 * We only care about a EUI-64 and NAA designator types 270 * with LU association. 271 */ 272 if (assoc != 0x00) 273 continue; 274 if (type != 0x02 && type != 0x03) 275 continue; 276 if (id_len != 8 && id_len != 12 && id_len != 16) 277 continue; 278 279 b->scsi.code_set = PS_CODE_SET_BINARY; 280 b->scsi.designator_type = type == 0x02 ? 281 PS_DESIGNATOR_EUI64 : PS_DESIGNATOR_NAA; 282 b->scsi.designator_len = id_len; 283 memcpy(b->scsi.designator, d + 4, id_len); 284 285 /* 286 * If we found a 8 or 12 byte descriptor continue on to 287 * see if a 16 byte one is available. If we find a 288 * 16 byte descriptor we're done. 289 */ 290 if (id_len == 16) 291 break; 292 } 293 294 out_put_request: 295 blk_put_request(rq); 296 out_free_buf: 297 kfree(buf); 298 return error; 299 } 300 301 #define NFSD_MDS_PR_KEY 0x0100000000000000ULL 302 303 /* 304 * We use the client ID as a unique key for the reservations. 305 * This allows us to easily fence a client when recalls fail. 306 */ 307 static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp) 308 { 309 return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id; 310 } 311 312 static int 313 nfsd4_block_get_device_info_scsi(struct super_block *sb, 314 struct nfs4_client *clp, 315 struct nfsd4_getdeviceinfo *gdp) 316 { 317 struct pnfs_block_deviceaddr *dev; 318 struct pnfs_block_volume *b; 319 const struct pr_ops *ops; 320 int error; 321 322 dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) + 323 sizeof(struct pnfs_block_volume), GFP_KERNEL); 324 if (!dev) 325 return -ENOMEM; 326 gdp->gd_device = dev; 327 328 dev->nr_volumes = 1; 329 b = &dev->volumes[0]; 330 331 b->type = PNFS_BLOCK_VOLUME_SCSI; 332 b->scsi.pr_key = nfsd4_scsi_pr_key(clp); 333 334 error = nfsd4_scsi_identify_device(sb->s_bdev, b); 335 if (error) 336 return error; 337 338 ops = sb->s_bdev->bd_disk->fops->pr_ops; 339 if (!ops) { 340 pr_err("pNFS: device %s does not support PRs.\n", 341 sb->s_id); 342 return -EINVAL; 343 } 344 345 error = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true); 346 if (error) { 347 pr_err("pNFS: failed to register key for device %s.\n", 348 sb->s_id); 349 return -EINVAL; 350 } 351 352 error = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY, 353 PR_EXCLUSIVE_ACCESS_REG_ONLY, 0); 354 if (error) { 355 pr_err("pNFS: failed to reserve device %s.\n", 356 sb->s_id); 357 return -EINVAL; 358 } 359 360 return 0; 361 } 362 363 static __be32 364 nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb, 365 struct svc_rqst *rqstp, 366 struct nfs4_client *clp, 367 struct nfsd4_getdeviceinfo *gdp) 368 { 369 if (sb->s_bdev != sb->s_bdev->bd_contains) 370 return nfserr_inval; 371 return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp)); 372 } 373 static __be32 374 nfsd4_scsi_proc_layoutcommit(struct inode *inode, 375 struct nfsd4_layoutcommit *lcp) 376 { 377 struct iomap *iomaps; 378 int nr_iomaps; 379 380 nr_iomaps = nfsd4_scsi_decode_layoutupdate(lcp->lc_up_layout, 381 lcp->lc_up_len, &iomaps, i_blocksize(inode)); 382 if (nr_iomaps < 0) 383 return nfserrno(nr_iomaps); 384 385 return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps); 386 } 387 388 static void 389 nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls) 390 { 391 struct nfs4_client *clp = ls->ls_stid.sc_client; 392 struct block_device *bdev = ls->ls_file->f_path.mnt->mnt_sb->s_bdev; 393 394 bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY, 395 nfsd4_scsi_pr_key(clp), 0, true); 396 } 397 398 const struct nfsd4_layout_ops scsi_layout_ops = { 399 /* 400 * Pretend that we send notification to the client. This is a blatant 401 * lie to force recent Linux clients to cache our device IDs. 402 * We rarely ever change the device ID, so the harm of leaking deviceids 403 * for a while isn't too bad. Unfortunately RFC5661 is a complete mess 404 * in this regard, but I filed errata 4119 for this a while ago, and 405 * hopefully the Linux client will eventually start caching deviceids 406 * without this again. 407 */ 408 .notify_types = 409 NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE, 410 .proc_getdeviceinfo = nfsd4_scsi_proc_getdeviceinfo, 411 .encode_getdeviceinfo = nfsd4_block_encode_getdeviceinfo, 412 .proc_layoutget = nfsd4_block_proc_layoutget, 413 .encode_layoutget = nfsd4_block_encode_layoutget, 414 .proc_layoutcommit = nfsd4_scsi_proc_layoutcommit, 415 .fence_client = nfsd4_scsi_fence_client, 416 }; 417 #endif /* CONFIG_NFSD_SCSILAYOUT */ 418