xref: /linux/fs/nfsd/blocklayout.c (revision 75a9b40f3b14d1cc3771c463d32b71cf4e558246)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (c) 2014-2016 Christoph Hellwig.
4  */
5 #include <linux/exportfs.h>
6 #include <linux/iomap.h>
7 #include <linux/slab.h>
8 #include <linux/pr.h>
9 
10 #include <linux/nfsd/debug.h>
11 
12 #include "blocklayoutxdr.h"
13 #include "pnfs.h"
14 #include "filecache.h"
15 #include "vfs.h"
16 
17 #define NFSDDBG_FACILITY	NFSDDBG_PNFS
18 
19 
20 /*
21  * Get an extent from the file system that starts at offset or below
22  * and may be shorter than the requested length.
23  */
24 static __be32
25 nfsd4_block_map_extent(struct inode *inode, const struct svc_fh *fhp,
26 		u64 offset, u64 length, u32 iomode, u64 minlength,
27 		struct pnfs_block_extent *bex)
28 {
29 	struct super_block *sb = inode->i_sb;
30 	struct iomap iomap;
31 	u32 device_generation = 0;
32 	int error;
33 
34 	error = sb->s_export_op->map_blocks(inode, offset, length, &iomap,
35 			iomode != IOMODE_READ, &device_generation);
36 	if (error) {
37 		if (error == -ENXIO)
38 			return nfserr_layoutunavailable;
39 		return nfserrno(error);
40 	}
41 
42 	switch (iomap.type) {
43 	case IOMAP_MAPPED:
44 		if (iomode == IOMODE_READ)
45 			bex->es = PNFS_BLOCK_READ_DATA;
46 		else
47 			bex->es = PNFS_BLOCK_READWRITE_DATA;
48 		bex->soff = iomap.addr;
49 		break;
50 	case IOMAP_UNWRITTEN:
51 		if (iomode & IOMODE_RW) {
52 			/*
53 			 * Crack monkey special case from section 2.3.1.
54 			 */
55 			if (minlength == 0) {
56 				dprintk("pnfsd: no soup for you!\n");
57 				return nfserr_layoutunavailable;
58 			}
59 
60 			bex->es = PNFS_BLOCK_INVALID_DATA;
61 			bex->soff = iomap.addr;
62 			break;
63 		}
64 		fallthrough;
65 	case IOMAP_HOLE:
66 		if (iomode == IOMODE_READ) {
67 			bex->es = PNFS_BLOCK_NONE_DATA;
68 			break;
69 		}
70 		fallthrough;
71 	case IOMAP_DELALLOC:
72 	default:
73 		WARN(1, "pnfsd: filesystem returned %d extent\n", iomap.type);
74 		return nfserr_layoutunavailable;
75 	}
76 
77 	error = nfsd4_set_deviceid(&bex->vol_id, fhp, device_generation);
78 	if (error)
79 		return nfserrno(error);
80 
81 	bex->foff = iomap.offset;
82 	bex->len = iomap.length;
83 	return nfs_ok;
84 }
85 
86 static __be32
87 nfsd4_block_proc_layoutget(struct svc_rqst *rqstp, struct inode *inode,
88 		const struct svc_fh *fhp, struct nfsd4_layoutget *args)
89 {
90 	struct nfsd4_layout_seg *seg = &args->lg_seg;
91 	struct pnfs_block_layout *bl;
92 	struct pnfs_block_extent *first_bex, *last_bex;
93 	u64 offset = seg->offset, length = seg->length;
94 	u32 i, nr_extents_max, block_size = i_blocksize(inode);
95 	__be32 nfserr;
96 
97 	if (locks_in_grace(SVC_NET(rqstp)))
98 		return nfserr_grace;
99 
100 	nfserr = nfserr_layoutunavailable;
101 	if (seg->offset & (block_size - 1)) {
102 		dprintk("pnfsd: I/O misaligned\n");
103 		goto out_error;
104 	}
105 
106 	/*
107 	 * RFC 8881, section 3.3.17:
108 	 *   The layout4 data type defines a layout for a file.
109 	 *
110 	 * RFC 8881, section 18.43.3:
111 	 *   The loga_maxcount field specifies the maximum layout size
112 	 *   (in bytes) that the client can handle. If the size of the
113 	 *   layout structure exceeds the size specified by maxcount,
114 	 *   the metadata server will return the NFS4ERR_TOOSMALL error.
115 	 */
116 	nfserr = nfserr_toosmall;
117 	if (args->lg_maxcount < PNFS_BLOCK_LAYOUT4_SIZE +
118 				PNFS_BLOCK_EXTENT_SIZE)
119 		goto out_error;
120 
121 	/*
122 	 * Limit the maximum layout size to avoid allocating
123 	 * a large buffer on the server for each layout request.
124 	 */
125 	nr_extents_max = (min(args->lg_maxcount, PAGE_SIZE) -
126 			  PNFS_BLOCK_LAYOUT4_SIZE) / PNFS_BLOCK_EXTENT_SIZE;
127 
128 	/*
129 	 * Some clients barf on non-zero block numbers for NONE or INVALID
130 	 * layouts, so make sure to zero the whole structure.
131 	 */
132 	nfserr = nfserrno(-ENOMEM);
133 	bl = kzalloc(struct_size(bl, extents, nr_extents_max), GFP_KERNEL);
134 	if (!bl)
135 		goto out_error;
136 	bl->nr_extents = nr_extents_max;
137 	args->lg_content = bl;
138 
139 	for (i = 0; i < bl->nr_extents; i++) {
140 		struct pnfs_block_extent *bex = bl->extents + i;
141 		u64 bex_length;
142 
143 		nfserr = nfsd4_block_map_extent(inode, fhp, offset, length,
144 				seg->iomode, args->lg_minlength, bex);
145 		if (nfserr != nfs_ok)
146 			goto out_error;
147 
148 		bex_length = bex->len - (offset - bex->foff);
149 		if (bex_length >= length) {
150 			bl->nr_extents = i + 1;
151 			break;
152 		}
153 
154 		offset = bex->foff + bex->len;
155 		length -= bex_length;
156 	}
157 
158 	first_bex = bl->extents;
159 	last_bex = bl->extents + bl->nr_extents - 1;
160 
161 	nfserr = nfserr_layoutunavailable;
162 	length = last_bex->foff + last_bex->len - seg->offset;
163 	if (length < args->lg_minlength) {
164 		dprintk("pnfsd: extent smaller than minlength\n");
165 		goto out_error;
166 	}
167 
168 	seg->offset = first_bex->foff;
169 	seg->length = last_bex->foff - first_bex->foff + last_bex->len;
170 	return nfs_ok;
171 
172 out_error:
173 	seg->length = 0;
174 	return nfserr;
175 }
176 
177 static __be32
178 nfsd4_block_commit_blocks(struct inode *inode, struct nfsd4_layoutcommit *lcp,
179 		struct iomap *iomaps, int nr_iomaps)
180 {
181 	struct timespec64 mtime = inode_get_mtime(inode);
182 	struct iattr iattr = { .ia_valid = 0 };
183 	int error;
184 
185 	if (lcp->lc_mtime.tv_nsec == UTIME_NOW ||
186 	    timespec64_compare(&lcp->lc_mtime, &mtime) < 0)
187 		lcp->lc_mtime = current_time(inode);
188 	iattr.ia_valid |= ATTR_ATIME | ATTR_CTIME | ATTR_MTIME;
189 	iattr.ia_atime = iattr.ia_ctime = iattr.ia_mtime = lcp->lc_mtime;
190 
191 	if (lcp->lc_size_chg) {
192 		iattr.ia_valid |= ATTR_SIZE;
193 		iattr.ia_size = lcp->lc_newsize;
194 	}
195 
196 	error = inode->i_sb->s_export_op->commit_blocks(inode, iomaps,
197 			nr_iomaps, &iattr);
198 	kfree(iomaps);
199 	return nfserrno(error);
200 }
201 
202 #ifdef CONFIG_NFSD_BLOCKLAYOUT
203 static int
204 nfsd4_block_get_device_info_simple(struct super_block *sb,
205 		struct nfsd4_getdeviceinfo *gdp)
206 {
207 	struct pnfs_block_deviceaddr *dev;
208 	struct pnfs_block_volume *b;
209 
210 	dev = kzalloc(struct_size(dev, volumes, 1), GFP_KERNEL);
211 	if (!dev)
212 		return -ENOMEM;
213 	gdp->gd_device = dev;
214 
215 	dev->nr_volumes = 1;
216 	b = &dev->volumes[0];
217 
218 	b->type = PNFS_BLOCK_VOLUME_SIMPLE;
219 	b->simple.sig_len = PNFS_BLOCK_UUID_LEN;
220 	return sb->s_export_op->get_uuid(sb, b->simple.sig, &b->simple.sig_len,
221 			&b->simple.offset);
222 }
223 
224 static __be32
225 nfsd4_block_proc_getdeviceinfo(struct super_block *sb,
226 		struct svc_rqst *rqstp,
227 		struct nfs4_client *clp,
228 		struct nfsd4_getdeviceinfo *gdp)
229 {
230 	if (bdev_is_partition(sb->s_bdev))
231 		return nfserr_inval;
232 	return nfserrno(nfsd4_block_get_device_info_simple(sb, gdp));
233 }
234 
235 static __be32
236 nfsd4_block_proc_layoutcommit(struct inode *inode, struct svc_rqst *rqstp,
237 		struct nfsd4_layoutcommit *lcp)
238 {
239 	struct iomap *iomaps;
240 	int nr_iomaps;
241 	__be32 nfserr;
242 
243 	rqstp->rq_arg = lcp->lc_up_layout;
244 	svcxdr_init_decode(rqstp);
245 
246 	nfserr = nfsd4_block_decode_layoutupdate(&rqstp->rq_arg_stream,
247 			&iomaps, &nr_iomaps, i_blocksize(inode));
248 	if (nfserr != nfs_ok)
249 		return nfserr;
250 
251 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
252 }
253 
254 const struct nfsd4_layout_ops bl_layout_ops = {
255 	/*
256 	 * Pretend that we send notification to the client.  This is a blatant
257 	 * lie to force recent Linux clients to cache our device IDs.
258 	 * We rarely ever change the device ID, so the harm of leaking deviceids
259 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
260 	 * in this regard, but I filed errata 4119 for this a while ago, and
261 	 * hopefully the Linux client will eventually start caching deviceids
262 	 * without this again.
263 	 */
264 	.notify_types		=
265 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
266 	.proc_getdeviceinfo	= nfsd4_block_proc_getdeviceinfo,
267 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
268 	.proc_layoutget		= nfsd4_block_proc_layoutget,
269 	.encode_layoutget	= nfsd4_block_encode_layoutget,
270 	.proc_layoutcommit	= nfsd4_block_proc_layoutcommit,
271 };
272 #endif /* CONFIG_NFSD_BLOCKLAYOUT */
273 
274 #ifdef CONFIG_NFSD_SCSILAYOUT
275 #define NFSD_MDS_PR_KEY		0x0100000000000000ULL
276 
277 /*
278  * We use the client ID as a unique key for the reservations.
279  * This allows us to easily fence a client when recalls fail.
280  */
281 static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp)
282 {
283 	return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id;
284 }
285 
286 static const u8 designator_types[] = {
287 	PS_DESIGNATOR_EUI64,
288 	PS_DESIGNATOR_NAA,
289 };
290 
291 static int
292 nfsd4_block_get_unique_id(struct gendisk *disk, struct pnfs_block_volume *b)
293 {
294 	int ret, i;
295 
296 	for (i = 0; i < ARRAY_SIZE(designator_types); i++) {
297 		u8 type = designator_types[i];
298 
299 		ret = disk->fops->get_unique_id(disk, b->scsi.designator, type);
300 		if (ret > 0) {
301 			b->scsi.code_set = PS_CODE_SET_BINARY;
302 			b->scsi.designator_type = type;
303 			b->scsi.designator_len = ret;
304 			return 0;
305 		}
306 	}
307 
308 	return -EINVAL;
309 }
310 
311 static int
312 nfsd4_block_get_device_info_scsi(struct super_block *sb,
313 		struct nfs4_client *clp,
314 		struct nfsd4_getdeviceinfo *gdp)
315 {
316 	struct pnfs_block_deviceaddr *dev;
317 	struct pnfs_block_volume *b;
318 	const struct pr_ops *ops;
319 	int ret;
320 
321 	dev = kzalloc(struct_size(dev, volumes, 1), GFP_KERNEL);
322 	if (!dev)
323 		return -ENOMEM;
324 	gdp->gd_device = dev;
325 
326 	dev->nr_volumes = 1;
327 	b = &dev->volumes[0];
328 
329 	b->type = PNFS_BLOCK_VOLUME_SCSI;
330 	b->scsi.pr_key = nfsd4_scsi_pr_key(clp);
331 
332 	ret = nfsd4_block_get_unique_id(sb->s_bdev->bd_disk, b);
333 	if (ret < 0)
334 		goto out_free_dev;
335 
336 	ret = -EINVAL;
337 	ops = sb->s_bdev->bd_disk->fops->pr_ops;
338 	if (!ops) {
339 		pr_err("pNFS: device %s does not support PRs.\n",
340 			sb->s_id);
341 		goto out_free_dev;
342 	}
343 
344 	ret = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true);
345 	if (ret) {
346 		pr_err("pNFS: failed to register key for device %s.\n",
347 			sb->s_id);
348 		goto out_free_dev;
349 	}
350 
351 	ret = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY,
352 			PR_EXCLUSIVE_ACCESS_REG_ONLY, 0);
353 	if (ret) {
354 		pr_err("pNFS: failed to reserve device %s.\n",
355 			sb->s_id);
356 		goto out_free_dev;
357 	}
358 
359 	return 0;
360 
361 out_free_dev:
362 	kfree(dev);
363 	gdp->gd_device = NULL;
364 	return ret;
365 }
366 
367 static __be32
368 nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb,
369 		struct svc_rqst *rqstp,
370 		struct nfs4_client *clp,
371 		struct nfsd4_getdeviceinfo *gdp)
372 {
373 	if (bdev_is_partition(sb->s_bdev))
374 		return nfserr_inval;
375 	return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp));
376 }
377 static __be32
378 nfsd4_scsi_proc_layoutcommit(struct inode *inode, struct svc_rqst *rqstp,
379 		struct nfsd4_layoutcommit *lcp)
380 {
381 	struct iomap *iomaps;
382 	int nr_iomaps;
383 	__be32 nfserr;
384 
385 	rqstp->rq_arg = lcp->lc_up_layout;
386 	svcxdr_init_decode(rqstp);
387 
388 	nfserr = nfsd4_scsi_decode_layoutupdate(&rqstp->rq_arg_stream,
389 			&iomaps, &nr_iomaps, i_blocksize(inode));
390 	if (nfserr != nfs_ok)
391 		return nfserr;
392 
393 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
394 }
395 
396 static void
397 nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls, struct nfsd_file *file)
398 {
399 	struct nfs4_client *clp = ls->ls_stid.sc_client;
400 	struct block_device *bdev = file->nf_file->f_path.mnt->mnt_sb->s_bdev;
401 
402 	bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY,
403 			nfsd4_scsi_pr_key(clp), 0, true);
404 }
405 
406 const struct nfsd4_layout_ops scsi_layout_ops = {
407 	/*
408 	 * Pretend that we send notification to the client.  This is a blatant
409 	 * lie to force recent Linux clients to cache our device IDs.
410 	 * We rarely ever change the device ID, so the harm of leaking deviceids
411 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
412 	 * in this regard, but I filed errata 4119 for this a while ago, and
413 	 * hopefully the Linux client will eventually start caching deviceids
414 	 * without this again.
415 	 */
416 	.notify_types		=
417 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
418 	.proc_getdeviceinfo	= nfsd4_scsi_proc_getdeviceinfo,
419 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
420 	.proc_layoutget		= nfsd4_block_proc_layoutget,
421 	.encode_layoutget	= nfsd4_block_encode_layoutget,
422 	.proc_layoutcommit	= nfsd4_scsi_proc_layoutcommit,
423 	.fence_client		= nfsd4_scsi_fence_client,
424 };
425 #endif /* CONFIG_NFSD_SCSILAYOUT */
426