xref: /linux/fs/nfsd/blocklayout.c (revision b8d312aa075f33282565467662c4628dae0a2aff)
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/genhd.h>
8 #include <linux/slab.h>
9 #include <linux/pr.h>
10 
11 #include <linux/nfsd/debug.h>
12 #include <scsi/scsi_proto.h>
13 #include <scsi/scsi_common.h>
14 #include <scsi/scsi_request.h>
15 
16 #include "blocklayoutxdr.h"
17 #include "pnfs.h"
18 
19 #define NFSDDBG_FACILITY	NFSDDBG_PNFS
20 
21 
22 static __be32
23 nfsd4_block_proc_layoutget(struct inode *inode, const struct svc_fh *fhp,
24 		struct nfsd4_layoutget *args)
25 {
26 	struct nfsd4_layout_seg *seg = &args->lg_seg;
27 	struct super_block *sb = inode->i_sb;
28 	u32 block_size = i_blocksize(inode);
29 	struct pnfs_block_extent *bex;
30 	struct iomap iomap;
31 	u32 device_generation = 0;
32 	int error;
33 
34 	if (seg->offset & (block_size - 1)) {
35 		dprintk("pnfsd: I/O misaligned\n");
36 		goto out_layoutunavailable;
37 	}
38 
39 	/*
40 	 * Some clients barf on non-zero block numbers for NONE or INVALID
41 	 * layouts, so make sure to zero the whole structure.
42 	 */
43 	error = -ENOMEM;
44 	bex = kzalloc(sizeof(*bex), GFP_KERNEL);
45 	if (!bex)
46 		goto out_error;
47 	args->lg_content = bex;
48 
49 	error = sb->s_export_op->map_blocks(inode, seg->offset, seg->length,
50 					    &iomap, seg->iomode != IOMODE_READ,
51 					    &device_generation);
52 	if (error) {
53 		if (error == -ENXIO)
54 			goto out_layoutunavailable;
55 		goto out_error;
56 	}
57 
58 	if (iomap.length < args->lg_minlength) {
59 		dprintk("pnfsd: extent smaller than minlength\n");
60 		goto out_layoutunavailable;
61 	}
62 
63 	switch (iomap.type) {
64 	case IOMAP_MAPPED:
65 		if (seg->iomode == IOMODE_READ)
66 			bex->es = PNFS_BLOCK_READ_DATA;
67 		else
68 			bex->es = PNFS_BLOCK_READWRITE_DATA;
69 		bex->soff = iomap.addr;
70 		break;
71 	case IOMAP_UNWRITTEN:
72 		if (seg->iomode & IOMODE_RW) {
73 			/*
74 			 * Crack monkey special case from section 2.3.1.
75 			 */
76 			if (args->lg_minlength == 0) {
77 				dprintk("pnfsd: no soup for you!\n");
78 				goto out_layoutunavailable;
79 			}
80 
81 			bex->es = PNFS_BLOCK_INVALID_DATA;
82 			bex->soff = iomap.addr;
83 			break;
84 		}
85 		/*FALLTHRU*/
86 	case IOMAP_HOLE:
87 		if (seg->iomode == IOMODE_READ) {
88 			bex->es = PNFS_BLOCK_NONE_DATA;
89 			break;
90 		}
91 		/*FALLTHRU*/
92 	case IOMAP_DELALLOC:
93 	default:
94 		WARN(1, "pnfsd: filesystem returned %d extent\n", iomap.type);
95 		goto out_layoutunavailable;
96 	}
97 
98 	error = nfsd4_set_deviceid(&bex->vol_id, fhp, device_generation);
99 	if (error)
100 		goto out_error;
101 	bex->foff = iomap.offset;
102 	bex->len = iomap.length;
103 
104 	seg->offset = iomap.offset;
105 	seg->length = iomap.length;
106 
107 	dprintk("GET: 0x%llx:0x%llx %d\n", bex->foff, bex->len, bex->es);
108 	return 0;
109 
110 out_error:
111 	seg->length = 0;
112 	return nfserrno(error);
113 out_layoutunavailable:
114 	seg->length = 0;
115 	return nfserr_layoutunavailable;
116 }
117 
118 static __be32
119 nfsd4_block_commit_blocks(struct inode *inode, struct nfsd4_layoutcommit *lcp,
120 		struct iomap *iomaps, int nr_iomaps)
121 {
122 	loff_t new_size = lcp->lc_last_wr + 1;
123 	struct iattr iattr = { .ia_valid = 0 };
124 	int error;
125 
126 	if (lcp->lc_mtime.tv_nsec == UTIME_NOW ||
127 	    timespec64_compare(&lcp->lc_mtime, &inode->i_mtime) < 0)
128 		lcp->lc_mtime = current_time(inode);
129 	iattr.ia_valid |= ATTR_ATIME | ATTR_CTIME | ATTR_MTIME;
130 	iattr.ia_atime = iattr.ia_ctime = iattr.ia_mtime = lcp->lc_mtime;
131 
132 	if (new_size > i_size_read(inode)) {
133 		iattr.ia_valid |= ATTR_SIZE;
134 		iattr.ia_size = new_size;
135 	}
136 
137 	error = inode->i_sb->s_export_op->commit_blocks(inode, iomaps,
138 			nr_iomaps, &iattr);
139 	kfree(iomaps);
140 	return nfserrno(error);
141 }
142 
143 #ifdef CONFIG_NFSD_BLOCKLAYOUT
144 static int
145 nfsd4_block_get_device_info_simple(struct super_block *sb,
146 		struct nfsd4_getdeviceinfo *gdp)
147 {
148 	struct pnfs_block_deviceaddr *dev;
149 	struct pnfs_block_volume *b;
150 
151 	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
152 		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
153 	if (!dev)
154 		return -ENOMEM;
155 	gdp->gd_device = dev;
156 
157 	dev->nr_volumes = 1;
158 	b = &dev->volumes[0];
159 
160 	b->type = PNFS_BLOCK_VOLUME_SIMPLE;
161 	b->simple.sig_len = PNFS_BLOCK_UUID_LEN;
162 	return sb->s_export_op->get_uuid(sb, b->simple.sig, &b->simple.sig_len,
163 			&b->simple.offset);
164 }
165 
166 static __be32
167 nfsd4_block_proc_getdeviceinfo(struct super_block *sb,
168 		struct svc_rqst *rqstp,
169 		struct nfs4_client *clp,
170 		struct nfsd4_getdeviceinfo *gdp)
171 {
172 	if (sb->s_bdev != sb->s_bdev->bd_contains)
173 		return nfserr_inval;
174 	return nfserrno(nfsd4_block_get_device_info_simple(sb, gdp));
175 }
176 
177 static __be32
178 nfsd4_block_proc_layoutcommit(struct inode *inode,
179 		struct nfsd4_layoutcommit *lcp)
180 {
181 	struct iomap *iomaps;
182 	int nr_iomaps;
183 
184 	nr_iomaps = nfsd4_block_decode_layoutupdate(lcp->lc_up_layout,
185 			lcp->lc_up_len, &iomaps, i_blocksize(inode));
186 	if (nr_iomaps < 0)
187 		return nfserrno(nr_iomaps);
188 
189 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
190 }
191 
192 const struct nfsd4_layout_ops bl_layout_ops = {
193 	/*
194 	 * Pretend that we send notification to the client.  This is a blatant
195 	 * lie to force recent Linux clients to cache our device IDs.
196 	 * We rarely ever change the device ID, so the harm of leaking deviceids
197 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
198 	 * in this regard, but I filed errata 4119 for this a while ago, and
199 	 * hopefully the Linux client will eventually start caching deviceids
200 	 * without this again.
201 	 */
202 	.notify_types		=
203 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
204 	.proc_getdeviceinfo	= nfsd4_block_proc_getdeviceinfo,
205 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
206 	.proc_layoutget		= nfsd4_block_proc_layoutget,
207 	.encode_layoutget	= nfsd4_block_encode_layoutget,
208 	.proc_layoutcommit	= nfsd4_block_proc_layoutcommit,
209 };
210 #endif /* CONFIG_NFSD_BLOCKLAYOUT */
211 
212 #ifdef CONFIG_NFSD_SCSILAYOUT
213 static int nfsd4_scsi_identify_device(struct block_device *bdev,
214 		struct pnfs_block_volume *b)
215 {
216 	struct request_queue *q = bdev->bd_disk->queue;
217 	struct request *rq;
218 	struct scsi_request *req;
219 	/*
220 	 * The allocation length (passed in bytes 3 and 4 of the INQUIRY
221 	 * command descriptor block) specifies the number of bytes that have
222 	 * been allocated for the data-in buffer.
223 	 * 252 is the highest one-byte value that is a multiple of 4.
224 	 * 65532 is the highest two-byte value that is a multiple of 4.
225 	 */
226 	size_t bufflen = 252, maxlen = 65532, len, id_len;
227 	u8 *buf, *d, type, assoc;
228 	int retries = 1, error;
229 
230 	if (WARN_ON_ONCE(!blk_queue_scsi_passthrough(q)))
231 		return -EINVAL;
232 
233 again:
234 	buf = kzalloc(bufflen, GFP_KERNEL);
235 	if (!buf)
236 		return -ENOMEM;
237 
238 	rq = blk_get_request(q, REQ_OP_SCSI_IN, 0);
239 	if (IS_ERR(rq)) {
240 		error = -ENOMEM;
241 		goto out_free_buf;
242 	}
243 	req = scsi_req(rq);
244 
245 	error = blk_rq_map_kern(q, rq, buf, bufflen, GFP_KERNEL);
246 	if (error)
247 		goto out_put_request;
248 
249 	req->cmd[0] = INQUIRY;
250 	req->cmd[1] = 1;
251 	req->cmd[2] = 0x83;
252 	req->cmd[3] = bufflen >> 8;
253 	req->cmd[4] = bufflen & 0xff;
254 	req->cmd_len = COMMAND_SIZE(INQUIRY);
255 
256 	blk_execute_rq(rq->q, NULL, rq, 1);
257 	if (req->result) {
258 		pr_err("pNFS: INQUIRY 0x83 failed with: %x\n",
259 			req->result);
260 		error = -EIO;
261 		goto out_put_request;
262 	}
263 
264 	len = (buf[2] << 8) + buf[3] + 4;
265 	if (len > bufflen) {
266 		if (len <= maxlen && retries--) {
267 			blk_put_request(rq);
268 			kfree(buf);
269 			bufflen = len;
270 			goto again;
271 		}
272 		pr_err("pNFS: INQUIRY 0x83 response invalid (len = %zd)\n",
273 			len);
274 		goto out_put_request;
275 	}
276 
277 	d = buf + 4;
278 	for (d = buf + 4; d < buf + len; d += id_len + 4) {
279 		id_len = d[3];
280 		type = d[1] & 0xf;
281 		assoc = (d[1] >> 4) & 0x3;
282 
283 		/*
284 		 * We only care about a EUI-64 and NAA designator types
285 		 * with LU association.
286 		 */
287 		if (assoc != 0x00)
288 			continue;
289 		if (type != 0x02 && type != 0x03)
290 			continue;
291 		if (id_len != 8 && id_len != 12 && id_len != 16)
292 			continue;
293 
294 		b->scsi.code_set = PS_CODE_SET_BINARY;
295 		b->scsi.designator_type = type == 0x02 ?
296 			PS_DESIGNATOR_EUI64 : PS_DESIGNATOR_NAA;
297 		b->scsi.designator_len = id_len;
298 		memcpy(b->scsi.designator, d + 4, id_len);
299 
300 		/*
301 		 * If we found a 8 or 12 byte descriptor continue on to
302 		 * see if a 16 byte one is available.  If we find a
303 		 * 16 byte descriptor we're done.
304 		 */
305 		if (id_len == 16)
306 			break;
307 	}
308 
309 out_put_request:
310 	blk_put_request(rq);
311 out_free_buf:
312 	kfree(buf);
313 	return error;
314 }
315 
316 #define NFSD_MDS_PR_KEY		0x0100000000000000ULL
317 
318 /*
319  * We use the client ID as a unique key for the reservations.
320  * This allows us to easily fence a client when recalls fail.
321  */
322 static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp)
323 {
324 	return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id;
325 }
326 
327 static int
328 nfsd4_block_get_device_info_scsi(struct super_block *sb,
329 		struct nfs4_client *clp,
330 		struct nfsd4_getdeviceinfo *gdp)
331 {
332 	struct pnfs_block_deviceaddr *dev;
333 	struct pnfs_block_volume *b;
334 	const struct pr_ops *ops;
335 	int error;
336 
337 	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
338 		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
339 	if (!dev)
340 		return -ENOMEM;
341 	gdp->gd_device = dev;
342 
343 	dev->nr_volumes = 1;
344 	b = &dev->volumes[0];
345 
346 	b->type = PNFS_BLOCK_VOLUME_SCSI;
347 	b->scsi.pr_key = nfsd4_scsi_pr_key(clp);
348 
349 	error = nfsd4_scsi_identify_device(sb->s_bdev, b);
350 	if (error)
351 		return error;
352 
353 	ops = sb->s_bdev->bd_disk->fops->pr_ops;
354 	if (!ops) {
355 		pr_err("pNFS: device %s does not support PRs.\n",
356 			sb->s_id);
357 		return -EINVAL;
358 	}
359 
360 	error = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true);
361 	if (error) {
362 		pr_err("pNFS: failed to register key for device %s.\n",
363 			sb->s_id);
364 		return -EINVAL;
365 	}
366 
367 	error = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY,
368 			PR_EXCLUSIVE_ACCESS_REG_ONLY, 0);
369 	if (error) {
370 		pr_err("pNFS: failed to reserve device %s.\n",
371 			sb->s_id);
372 		return -EINVAL;
373 	}
374 
375 	return 0;
376 }
377 
378 static __be32
379 nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb,
380 		struct svc_rqst *rqstp,
381 		struct nfs4_client *clp,
382 		struct nfsd4_getdeviceinfo *gdp)
383 {
384 	if (sb->s_bdev != sb->s_bdev->bd_contains)
385 		return nfserr_inval;
386 	return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp));
387 }
388 static __be32
389 nfsd4_scsi_proc_layoutcommit(struct inode *inode,
390 		struct nfsd4_layoutcommit *lcp)
391 {
392 	struct iomap *iomaps;
393 	int nr_iomaps;
394 
395 	nr_iomaps = nfsd4_scsi_decode_layoutupdate(lcp->lc_up_layout,
396 			lcp->lc_up_len, &iomaps, i_blocksize(inode));
397 	if (nr_iomaps < 0)
398 		return nfserrno(nr_iomaps);
399 
400 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
401 }
402 
403 static void
404 nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls)
405 {
406 	struct nfs4_client *clp = ls->ls_stid.sc_client;
407 	struct block_device *bdev = ls->ls_file->f_path.mnt->mnt_sb->s_bdev;
408 
409 	bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY,
410 			nfsd4_scsi_pr_key(clp), 0, true);
411 }
412 
413 const struct nfsd4_layout_ops scsi_layout_ops = {
414 	/*
415 	 * Pretend that we send notification to the client.  This is a blatant
416 	 * lie to force recent Linux clients to cache our device IDs.
417 	 * We rarely ever change the device ID, so the harm of leaking deviceids
418 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
419 	 * in this regard, but I filed errata 4119 for this a while ago, and
420 	 * hopefully the Linux client will eventually start caching deviceids
421 	 * without this again.
422 	 */
423 	.notify_types		=
424 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
425 	.proc_getdeviceinfo	= nfsd4_scsi_proc_getdeviceinfo,
426 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
427 	.proc_layoutget		= nfsd4_block_proc_layoutget,
428 	.encode_layoutget	= nfsd4_block_encode_layoutget,
429 	.proc_layoutcommit	= nfsd4_scsi_proc_layoutcommit,
430 	.fence_client		= nfsd4_scsi_fence_client,
431 };
432 #endif /* CONFIG_NFSD_SCSILAYOUT */
433