xref: /openbmc/linux/fs/nfsd/blocklayout.c (revision 0d3b051a)
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 #include "filecache.h"
19 
20 #define NFSDDBG_FACILITY	NFSDDBG_PNFS
21 
22 
23 static __be32
24 nfsd4_block_proc_layoutget(struct inode *inode, const struct svc_fh *fhp,
25 		struct nfsd4_layoutget *args)
26 {
27 	struct nfsd4_layout_seg *seg = &args->lg_seg;
28 	struct super_block *sb = inode->i_sb;
29 	u32 block_size = i_blocksize(inode);
30 	struct pnfs_block_extent *bex;
31 	struct iomap iomap;
32 	u32 device_generation = 0;
33 	int error;
34 
35 	if (seg->offset & (block_size - 1)) {
36 		dprintk("pnfsd: I/O misaligned\n");
37 		goto out_layoutunavailable;
38 	}
39 
40 	/*
41 	 * Some clients barf on non-zero block numbers for NONE or INVALID
42 	 * layouts, so make sure to zero the whole structure.
43 	 */
44 	error = -ENOMEM;
45 	bex = kzalloc(sizeof(*bex), GFP_KERNEL);
46 	if (!bex)
47 		goto out_error;
48 	args->lg_content = bex;
49 
50 	error = sb->s_export_op->map_blocks(inode, seg->offset, seg->length,
51 					    &iomap, seg->iomode != IOMODE_READ,
52 					    &device_generation);
53 	if (error) {
54 		if (error == -ENXIO)
55 			goto out_layoutunavailable;
56 		goto out_error;
57 	}
58 
59 	if (iomap.length < args->lg_minlength) {
60 		dprintk("pnfsd: extent smaller than minlength\n");
61 		goto out_layoutunavailable;
62 	}
63 
64 	switch (iomap.type) {
65 	case IOMAP_MAPPED:
66 		if (seg->iomode == IOMODE_READ)
67 			bex->es = PNFS_BLOCK_READ_DATA;
68 		else
69 			bex->es = PNFS_BLOCK_READWRITE_DATA;
70 		bex->soff = iomap.addr;
71 		break;
72 	case IOMAP_UNWRITTEN:
73 		if (seg->iomode & IOMODE_RW) {
74 			/*
75 			 * Crack monkey special case from section 2.3.1.
76 			 */
77 			if (args->lg_minlength == 0) {
78 				dprintk("pnfsd: no soup for you!\n");
79 				goto out_layoutunavailable;
80 			}
81 
82 			bex->es = PNFS_BLOCK_INVALID_DATA;
83 			bex->soff = iomap.addr;
84 			break;
85 		}
86 		fallthrough;
87 	case IOMAP_HOLE:
88 		if (seg->iomode == IOMODE_READ) {
89 			bex->es = PNFS_BLOCK_NONE_DATA;
90 			break;
91 		}
92 		fallthrough;
93 	case IOMAP_DELALLOC:
94 	default:
95 		WARN(1, "pnfsd: filesystem returned %d extent\n", iomap.type);
96 		goto out_layoutunavailable;
97 	}
98 
99 	error = nfsd4_set_deviceid(&bex->vol_id, fhp, device_generation);
100 	if (error)
101 		goto out_error;
102 	bex->foff = iomap.offset;
103 	bex->len = iomap.length;
104 
105 	seg->offset = iomap.offset;
106 	seg->length = iomap.length;
107 
108 	dprintk("GET: 0x%llx:0x%llx %d\n", bex->foff, bex->len, bex->es);
109 	return 0;
110 
111 out_error:
112 	seg->length = 0;
113 	return nfserrno(error);
114 out_layoutunavailable:
115 	seg->length = 0;
116 	return nfserr_layoutunavailable;
117 }
118 
119 static __be32
120 nfsd4_block_commit_blocks(struct inode *inode, struct nfsd4_layoutcommit *lcp,
121 		struct iomap *iomaps, int nr_iomaps)
122 {
123 	loff_t new_size = lcp->lc_last_wr + 1;
124 	struct iattr iattr = { .ia_valid = 0 };
125 	int error;
126 
127 	if (lcp->lc_mtime.tv_nsec == UTIME_NOW ||
128 	    timespec64_compare(&lcp->lc_mtime, &inode->i_mtime) < 0)
129 		lcp->lc_mtime = current_time(inode);
130 	iattr.ia_valid |= ATTR_ATIME | ATTR_CTIME | ATTR_MTIME;
131 	iattr.ia_atime = iattr.ia_ctime = iattr.ia_mtime = lcp->lc_mtime;
132 
133 	if (new_size > i_size_read(inode)) {
134 		iattr.ia_valid |= ATTR_SIZE;
135 		iattr.ia_size = new_size;
136 	}
137 
138 	error = inode->i_sb->s_export_op->commit_blocks(inode, iomaps,
139 			nr_iomaps, &iattr);
140 	kfree(iomaps);
141 	return nfserrno(error);
142 }
143 
144 #ifdef CONFIG_NFSD_BLOCKLAYOUT
145 static int
146 nfsd4_block_get_device_info_simple(struct super_block *sb,
147 		struct nfsd4_getdeviceinfo *gdp)
148 {
149 	struct pnfs_block_deviceaddr *dev;
150 	struct pnfs_block_volume *b;
151 
152 	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
153 		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
154 	if (!dev)
155 		return -ENOMEM;
156 	gdp->gd_device = dev;
157 
158 	dev->nr_volumes = 1;
159 	b = &dev->volumes[0];
160 
161 	b->type = PNFS_BLOCK_VOLUME_SIMPLE;
162 	b->simple.sig_len = PNFS_BLOCK_UUID_LEN;
163 	return sb->s_export_op->get_uuid(sb, b->simple.sig, &b->simple.sig_len,
164 			&b->simple.offset);
165 }
166 
167 static __be32
168 nfsd4_block_proc_getdeviceinfo(struct super_block *sb,
169 		struct svc_rqst *rqstp,
170 		struct nfs4_client *clp,
171 		struct nfsd4_getdeviceinfo *gdp)
172 {
173 	if (bdev_is_partition(sb->s_bdev))
174 		return nfserr_inval;
175 	return nfserrno(nfsd4_block_get_device_info_simple(sb, gdp));
176 }
177 
178 static __be32
179 nfsd4_block_proc_layoutcommit(struct inode *inode,
180 		struct nfsd4_layoutcommit *lcp)
181 {
182 	struct iomap *iomaps;
183 	int nr_iomaps;
184 
185 	nr_iomaps = nfsd4_block_decode_layoutupdate(lcp->lc_up_layout,
186 			lcp->lc_up_len, &iomaps, i_blocksize(inode));
187 	if (nr_iomaps < 0)
188 		return nfserrno(nr_iomaps);
189 
190 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
191 }
192 
193 const struct nfsd4_layout_ops bl_layout_ops = {
194 	/*
195 	 * Pretend that we send notification to the client.  This is a blatant
196 	 * lie to force recent Linux clients to cache our device IDs.
197 	 * We rarely ever change the device ID, so the harm of leaking deviceids
198 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
199 	 * in this regard, but I filed errata 4119 for this a while ago, and
200 	 * hopefully the Linux client will eventually start caching deviceids
201 	 * without this again.
202 	 */
203 	.notify_types		=
204 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
205 	.proc_getdeviceinfo	= nfsd4_block_proc_getdeviceinfo,
206 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
207 	.proc_layoutget		= nfsd4_block_proc_layoutget,
208 	.encode_layoutget	= nfsd4_block_encode_layoutget,
209 	.proc_layoutcommit	= nfsd4_block_proc_layoutcommit,
210 };
211 #endif /* CONFIG_NFSD_BLOCKLAYOUT */
212 
213 #ifdef CONFIG_NFSD_SCSILAYOUT
214 static int nfsd4_scsi_identify_device(struct block_device *bdev,
215 		struct pnfs_block_volume *b)
216 {
217 	struct request_queue *q = bdev->bd_disk->queue;
218 	struct request *rq;
219 	struct scsi_request *req;
220 	/*
221 	 * The allocation length (passed in bytes 3 and 4 of the INQUIRY
222 	 * command descriptor block) specifies the number of bytes that have
223 	 * been allocated for the data-in buffer.
224 	 * 252 is the highest one-byte value that is a multiple of 4.
225 	 * 65532 is the highest two-byte value that is a multiple of 4.
226 	 */
227 	size_t bufflen = 252, maxlen = 65532, len, id_len;
228 	u8 *buf, *d, type, assoc;
229 	int retries = 1, error;
230 
231 	if (WARN_ON_ONCE(!blk_queue_scsi_passthrough(q)))
232 		return -EINVAL;
233 
234 again:
235 	buf = kzalloc(bufflen, GFP_KERNEL);
236 	if (!buf)
237 		return -ENOMEM;
238 
239 	rq = blk_get_request(q, REQ_OP_SCSI_IN, 0);
240 	if (IS_ERR(rq)) {
241 		error = -ENOMEM;
242 		goto out_free_buf;
243 	}
244 	req = scsi_req(rq);
245 
246 	error = blk_rq_map_kern(q, rq, buf, bufflen, GFP_KERNEL);
247 	if (error)
248 		goto out_put_request;
249 
250 	req->cmd[0] = INQUIRY;
251 	req->cmd[1] = 1;
252 	req->cmd[2] = 0x83;
253 	req->cmd[3] = bufflen >> 8;
254 	req->cmd[4] = bufflen & 0xff;
255 	req->cmd_len = COMMAND_SIZE(INQUIRY);
256 
257 	blk_execute_rq(rq->q, NULL, rq, 1);
258 	if (req->result) {
259 		pr_err("pNFS: INQUIRY 0x83 failed with: %x\n",
260 			req->result);
261 		error = -EIO;
262 		goto out_put_request;
263 	}
264 
265 	len = (buf[2] << 8) + buf[3] + 4;
266 	if (len > bufflen) {
267 		if (len <= maxlen && retries--) {
268 			blk_put_request(rq);
269 			kfree(buf);
270 			bufflen = len;
271 			goto again;
272 		}
273 		pr_err("pNFS: INQUIRY 0x83 response invalid (len = %zd)\n",
274 			len);
275 		goto out_put_request;
276 	}
277 
278 	d = buf + 4;
279 	for (d = buf + 4; d < buf + len; d += id_len + 4) {
280 		id_len = d[3];
281 		type = d[1] & 0xf;
282 		assoc = (d[1] >> 4) & 0x3;
283 
284 		/*
285 		 * We only care about a EUI-64 and NAA designator types
286 		 * with LU association.
287 		 */
288 		if (assoc != 0x00)
289 			continue;
290 		if (type != 0x02 && type != 0x03)
291 			continue;
292 		if (id_len != 8 && id_len != 12 && id_len != 16)
293 			continue;
294 
295 		b->scsi.code_set = PS_CODE_SET_BINARY;
296 		b->scsi.designator_type = type == 0x02 ?
297 			PS_DESIGNATOR_EUI64 : PS_DESIGNATOR_NAA;
298 		b->scsi.designator_len = id_len;
299 		memcpy(b->scsi.designator, d + 4, id_len);
300 
301 		/*
302 		 * If we found a 8 or 12 byte descriptor continue on to
303 		 * see if a 16 byte one is available.  If we find a
304 		 * 16 byte descriptor we're done.
305 		 */
306 		if (id_len == 16)
307 			break;
308 	}
309 
310 out_put_request:
311 	blk_put_request(rq);
312 out_free_buf:
313 	kfree(buf);
314 	return error;
315 }
316 
317 #define NFSD_MDS_PR_KEY		0x0100000000000000ULL
318 
319 /*
320  * We use the client ID as a unique key for the reservations.
321  * This allows us to easily fence a client when recalls fail.
322  */
323 static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp)
324 {
325 	return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id;
326 }
327 
328 static int
329 nfsd4_block_get_device_info_scsi(struct super_block *sb,
330 		struct nfs4_client *clp,
331 		struct nfsd4_getdeviceinfo *gdp)
332 {
333 	struct pnfs_block_deviceaddr *dev;
334 	struct pnfs_block_volume *b;
335 	const struct pr_ops *ops;
336 	int error;
337 
338 	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
339 		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
340 	if (!dev)
341 		return -ENOMEM;
342 	gdp->gd_device = dev;
343 
344 	dev->nr_volumes = 1;
345 	b = &dev->volumes[0];
346 
347 	b->type = PNFS_BLOCK_VOLUME_SCSI;
348 	b->scsi.pr_key = nfsd4_scsi_pr_key(clp);
349 
350 	error = nfsd4_scsi_identify_device(sb->s_bdev, b);
351 	if (error)
352 		return error;
353 
354 	ops = sb->s_bdev->bd_disk->fops->pr_ops;
355 	if (!ops) {
356 		pr_err("pNFS: device %s does not support PRs.\n",
357 			sb->s_id);
358 		return -EINVAL;
359 	}
360 
361 	error = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true);
362 	if (error) {
363 		pr_err("pNFS: failed to register key for device %s.\n",
364 			sb->s_id);
365 		return -EINVAL;
366 	}
367 
368 	error = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY,
369 			PR_EXCLUSIVE_ACCESS_REG_ONLY, 0);
370 	if (error) {
371 		pr_err("pNFS: failed to reserve device %s.\n",
372 			sb->s_id);
373 		return -EINVAL;
374 	}
375 
376 	return 0;
377 }
378 
379 static __be32
380 nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb,
381 		struct svc_rqst *rqstp,
382 		struct nfs4_client *clp,
383 		struct nfsd4_getdeviceinfo *gdp)
384 {
385 	if (bdev_is_partition(sb->s_bdev))
386 		return nfserr_inval;
387 	return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp));
388 }
389 static __be32
390 nfsd4_scsi_proc_layoutcommit(struct inode *inode,
391 		struct nfsd4_layoutcommit *lcp)
392 {
393 	struct iomap *iomaps;
394 	int nr_iomaps;
395 
396 	nr_iomaps = nfsd4_scsi_decode_layoutupdate(lcp->lc_up_layout,
397 			lcp->lc_up_len, &iomaps, i_blocksize(inode));
398 	if (nr_iomaps < 0)
399 		return nfserrno(nr_iomaps);
400 
401 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
402 }
403 
404 static void
405 nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls)
406 {
407 	struct nfs4_client *clp = ls->ls_stid.sc_client;
408 	struct block_device *bdev = ls->ls_file->nf_file->f_path.mnt->mnt_sb->s_bdev;
409 
410 	bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY,
411 			nfsd4_scsi_pr_key(clp), 0, true);
412 }
413 
414 const struct nfsd4_layout_ops scsi_layout_ops = {
415 	/*
416 	 * Pretend that we send notification to the client.  This is a blatant
417 	 * lie to force recent Linux clients to cache our device IDs.
418 	 * We rarely ever change the device ID, so the harm of leaking deviceids
419 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
420 	 * in this regard, but I filed errata 4119 for this a while ago, and
421 	 * hopefully the Linux client will eventually start caching deviceids
422 	 * without this again.
423 	 */
424 	.notify_types		=
425 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
426 	.proc_getdeviceinfo	= nfsd4_scsi_proc_getdeviceinfo,
427 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
428 	.proc_layoutget		= nfsd4_block_proc_layoutget,
429 	.encode_layoutget	= nfsd4_block_encode_layoutget,
430 	.proc_layoutcommit	= nfsd4_scsi_proc_layoutcommit,
431 	.fence_client		= nfsd4_scsi_fence_client,
432 };
433 #endif /* CONFIG_NFSD_SCSILAYOUT */
434