xref: /openbmc/linux/fs/nfsd/blocklayout.c (revision f9a82c48)
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 	struct timespec ts;
125 	int error;
126 
127 	ts = timespec64_to_timespec(inode->i_mtime);
128 	if (lcp->lc_mtime.tv_nsec == UTIME_NOW ||
129 	    timespec_compare(&lcp->lc_mtime, &ts) < 0)
130 		lcp->lc_mtime = timespec64_to_timespec(current_time(inode));
131 	iattr.ia_valid |= ATTR_ATIME | ATTR_CTIME | ATTR_MTIME;
132 	iattr.ia_atime = iattr.ia_ctime = iattr.ia_mtime = timespec_to_timespec64(lcp->lc_mtime);
133 
134 	if (new_size > i_size_read(inode)) {
135 		iattr.ia_valid |= ATTR_SIZE;
136 		iattr.ia_size = new_size;
137 	}
138 
139 	error = inode->i_sb->s_export_op->commit_blocks(inode, iomaps,
140 			nr_iomaps, &iattr);
141 	kfree(iomaps);
142 	return nfserrno(error);
143 }
144 
145 #ifdef CONFIG_NFSD_BLOCKLAYOUT
146 static int
147 nfsd4_block_get_device_info_simple(struct super_block *sb,
148 		struct nfsd4_getdeviceinfo *gdp)
149 {
150 	struct pnfs_block_deviceaddr *dev;
151 	struct pnfs_block_volume *b;
152 
153 	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
154 		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
155 	if (!dev)
156 		return -ENOMEM;
157 	gdp->gd_device = dev;
158 
159 	dev->nr_volumes = 1;
160 	b = &dev->volumes[0];
161 
162 	b->type = PNFS_BLOCK_VOLUME_SIMPLE;
163 	b->simple.sig_len = PNFS_BLOCK_UUID_LEN;
164 	return sb->s_export_op->get_uuid(sb, b->simple.sig, &b->simple.sig_len,
165 			&b->simple.offset);
166 }
167 
168 static __be32
169 nfsd4_block_proc_getdeviceinfo(struct super_block *sb,
170 		struct svc_rqst *rqstp,
171 		struct nfs4_client *clp,
172 		struct nfsd4_getdeviceinfo *gdp)
173 {
174 	if (sb->s_bdev != sb->s_bdev->bd_contains)
175 		return nfserr_inval;
176 	return nfserrno(nfsd4_block_get_device_info_simple(sb, gdp));
177 }
178 
179 static __be32
180 nfsd4_block_proc_layoutcommit(struct inode *inode,
181 		struct nfsd4_layoutcommit *lcp)
182 {
183 	struct iomap *iomaps;
184 	int nr_iomaps;
185 
186 	nr_iomaps = nfsd4_block_decode_layoutupdate(lcp->lc_up_layout,
187 			lcp->lc_up_len, &iomaps, i_blocksize(inode));
188 	if (nr_iomaps < 0)
189 		return nfserrno(nr_iomaps);
190 
191 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
192 }
193 
194 const struct nfsd4_layout_ops bl_layout_ops = {
195 	/*
196 	 * Pretend that we send notification to the client.  This is a blatant
197 	 * lie to force recent Linux clients to cache our device IDs.
198 	 * We rarely ever change the device ID, so the harm of leaking deviceids
199 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
200 	 * in this regard, but I filed errata 4119 for this a while ago, and
201 	 * hopefully the Linux client will eventually start caching deviceids
202 	 * without this again.
203 	 */
204 	.notify_types		=
205 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
206 	.proc_getdeviceinfo	= nfsd4_block_proc_getdeviceinfo,
207 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
208 	.proc_layoutget		= nfsd4_block_proc_layoutget,
209 	.encode_layoutget	= nfsd4_block_encode_layoutget,
210 	.proc_layoutcommit	= nfsd4_block_proc_layoutcommit,
211 };
212 #endif /* CONFIG_NFSD_BLOCKLAYOUT */
213 
214 #ifdef CONFIG_NFSD_SCSILAYOUT
215 static int nfsd4_scsi_identify_device(struct block_device *bdev,
216 		struct pnfs_block_volume *b)
217 {
218 	struct request_queue *q = bdev->bd_disk->queue;
219 	struct request *rq;
220 	struct scsi_request *req;
221 	/*
222 	 * The allocation length (passed in bytes 3 and 4 of the INQUIRY
223 	 * command descriptor block) specifies the number of bytes that have
224 	 * been allocated for the data-in buffer.
225 	 * 252 is the highest one-byte value that is a multiple of 4.
226 	 * 65532 is the highest two-byte value that is a multiple of 4.
227 	 */
228 	size_t bufflen = 252, maxlen = 65532, len, id_len;
229 	u8 *buf, *d, type, assoc;
230 	int retries = 1, error;
231 
232 	if (WARN_ON_ONCE(!blk_queue_scsi_passthrough(q)))
233 		return -EINVAL;
234 
235 again:
236 	buf = kzalloc(bufflen, GFP_KERNEL);
237 	if (!buf)
238 		return -ENOMEM;
239 
240 	rq = blk_get_request(q, REQ_OP_SCSI_IN, 0);
241 	if (IS_ERR(rq)) {
242 		error = -ENOMEM;
243 		goto out_free_buf;
244 	}
245 	req = scsi_req(rq);
246 
247 	error = blk_rq_map_kern(q, rq, buf, bufflen, GFP_KERNEL);
248 	if (error)
249 		goto out_put_request;
250 
251 	req->cmd[0] = INQUIRY;
252 	req->cmd[1] = 1;
253 	req->cmd[2] = 0x83;
254 	req->cmd[3] = bufflen >> 8;
255 	req->cmd[4] = bufflen & 0xff;
256 	req->cmd_len = COMMAND_SIZE(INQUIRY);
257 
258 	blk_execute_rq(rq->q, NULL, rq, 1);
259 	if (req->result) {
260 		pr_err("pNFS: INQUIRY 0x83 failed with: %x\n",
261 			req->result);
262 		error = -EIO;
263 		goto out_put_request;
264 	}
265 
266 	len = (buf[2] << 8) + buf[3] + 4;
267 	if (len > bufflen) {
268 		if (len <= maxlen && retries--) {
269 			blk_put_request(rq);
270 			kfree(buf);
271 			bufflen = len;
272 			goto again;
273 		}
274 		pr_err("pNFS: INQUIRY 0x83 response invalid (len = %zd)\n",
275 			len);
276 		goto out_put_request;
277 	}
278 
279 	d = buf + 4;
280 	for (d = buf + 4; d < buf + len; d += id_len + 4) {
281 		id_len = d[3];
282 		type = d[1] & 0xf;
283 		assoc = (d[1] >> 4) & 0x3;
284 
285 		/*
286 		 * We only care about a EUI-64 and NAA designator types
287 		 * with LU association.
288 		 */
289 		if (assoc != 0x00)
290 			continue;
291 		if (type != 0x02 && type != 0x03)
292 			continue;
293 		if (id_len != 8 && id_len != 12 && id_len != 16)
294 			continue;
295 
296 		b->scsi.code_set = PS_CODE_SET_BINARY;
297 		b->scsi.designator_type = type == 0x02 ?
298 			PS_DESIGNATOR_EUI64 : PS_DESIGNATOR_NAA;
299 		b->scsi.designator_len = id_len;
300 		memcpy(b->scsi.designator, d + 4, id_len);
301 
302 		/*
303 		 * If we found a 8 or 12 byte descriptor continue on to
304 		 * see if a 16 byte one is available.  If we find a
305 		 * 16 byte descriptor we're done.
306 		 */
307 		if (id_len == 16)
308 			break;
309 	}
310 
311 out_put_request:
312 	blk_put_request(rq);
313 out_free_buf:
314 	kfree(buf);
315 	return error;
316 }
317 
318 #define NFSD_MDS_PR_KEY		0x0100000000000000ULL
319 
320 /*
321  * We use the client ID as a unique key for the reservations.
322  * This allows us to easily fence a client when recalls fail.
323  */
324 static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp)
325 {
326 	return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id;
327 }
328 
329 static int
330 nfsd4_block_get_device_info_scsi(struct super_block *sb,
331 		struct nfs4_client *clp,
332 		struct nfsd4_getdeviceinfo *gdp)
333 {
334 	struct pnfs_block_deviceaddr *dev;
335 	struct pnfs_block_volume *b;
336 	const struct pr_ops *ops;
337 	int error;
338 
339 	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
340 		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
341 	if (!dev)
342 		return -ENOMEM;
343 	gdp->gd_device = dev;
344 
345 	dev->nr_volumes = 1;
346 	b = &dev->volumes[0];
347 
348 	b->type = PNFS_BLOCK_VOLUME_SCSI;
349 	b->scsi.pr_key = nfsd4_scsi_pr_key(clp);
350 
351 	error = nfsd4_scsi_identify_device(sb->s_bdev, b);
352 	if (error)
353 		return error;
354 
355 	ops = sb->s_bdev->bd_disk->fops->pr_ops;
356 	if (!ops) {
357 		pr_err("pNFS: device %s does not support PRs.\n",
358 			sb->s_id);
359 		return -EINVAL;
360 	}
361 
362 	error = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true);
363 	if (error) {
364 		pr_err("pNFS: failed to register key for device %s.\n",
365 			sb->s_id);
366 		return -EINVAL;
367 	}
368 
369 	error = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY,
370 			PR_EXCLUSIVE_ACCESS_REG_ONLY, 0);
371 	if (error) {
372 		pr_err("pNFS: failed to reserve device %s.\n",
373 			sb->s_id);
374 		return -EINVAL;
375 	}
376 
377 	return 0;
378 }
379 
380 static __be32
381 nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb,
382 		struct svc_rqst *rqstp,
383 		struct nfs4_client *clp,
384 		struct nfsd4_getdeviceinfo *gdp)
385 {
386 	if (sb->s_bdev != sb->s_bdev->bd_contains)
387 		return nfserr_inval;
388 	return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp));
389 }
390 static __be32
391 nfsd4_scsi_proc_layoutcommit(struct inode *inode,
392 		struct nfsd4_layoutcommit *lcp)
393 {
394 	struct iomap *iomaps;
395 	int nr_iomaps;
396 
397 	nr_iomaps = nfsd4_scsi_decode_layoutupdate(lcp->lc_up_layout,
398 			lcp->lc_up_len, &iomaps, i_blocksize(inode));
399 	if (nr_iomaps < 0)
400 		return nfserrno(nr_iomaps);
401 
402 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
403 }
404 
405 static void
406 nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls)
407 {
408 	struct nfs4_client *clp = ls->ls_stid.sc_client;
409 	struct block_device *bdev = ls->ls_file->f_path.mnt->mnt_sb->s_bdev;
410 
411 	bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY,
412 			nfsd4_scsi_pr_key(clp), 0, true);
413 }
414 
415 const struct nfsd4_layout_ops scsi_layout_ops = {
416 	/*
417 	 * Pretend that we send notification to the client.  This is a blatant
418 	 * lie to force recent Linux clients to cache our device IDs.
419 	 * We rarely ever change the device ID, so the harm of leaking deviceids
420 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
421 	 * in this regard, but I filed errata 4119 for this a while ago, and
422 	 * hopefully the Linux client will eventually start caching deviceids
423 	 * without this again.
424 	 */
425 	.notify_types		=
426 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
427 	.proc_getdeviceinfo	= nfsd4_scsi_proc_getdeviceinfo,
428 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
429 	.proc_layoutget		= nfsd4_block_proc_layoutget,
430 	.encode_layoutget	= nfsd4_block_encode_layoutget,
431 	.proc_layoutcommit	= nfsd4_scsi_proc_layoutcommit,
432 	.fence_client		= nfsd4_scsi_fence_client,
433 };
434 #endif /* CONFIG_NFSD_SCSILAYOUT */
435