xref: /openbmc/linux/fs/nfsd/blocklayout.c (revision 77a87824)
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 
14 #include "blocklayoutxdr.h"
15 #include "pnfs.h"
16 
17 #define NFSDDBG_FACILITY	NFSDDBG_PNFS
18 
19 
20 static __be32
21 nfsd4_block_proc_layoutget(struct inode *inode, const struct svc_fh *fhp,
22 		struct nfsd4_layoutget *args)
23 {
24 	struct nfsd4_layout_seg *seg = &args->lg_seg;
25 	struct super_block *sb = inode->i_sb;
26 	u32 block_size = (1 << inode->i_blkbits);
27 	struct pnfs_block_extent *bex;
28 	struct iomap iomap;
29 	u32 device_generation = 0;
30 	int error;
31 
32 	if (seg->offset & (block_size - 1)) {
33 		dprintk("pnfsd: I/O misaligned\n");
34 		goto out_layoutunavailable;
35 	}
36 
37 	/*
38 	 * Some clients barf on non-zero block numbers for NONE or INVALID
39 	 * layouts, so make sure to zero the whole structure.
40 	 */
41 	error = -ENOMEM;
42 	bex = kzalloc(sizeof(*bex), GFP_KERNEL);
43 	if (!bex)
44 		goto out_error;
45 	args->lg_content = bex;
46 
47 	error = sb->s_export_op->map_blocks(inode, seg->offset, seg->length,
48 					    &iomap, seg->iomode != IOMODE_READ,
49 					    &device_generation);
50 	if (error) {
51 		if (error == -ENXIO)
52 			goto out_layoutunavailable;
53 		goto out_error;
54 	}
55 
56 	if (iomap.length < args->lg_minlength) {
57 		dprintk("pnfsd: extent smaller than minlength\n");
58 		goto out_layoutunavailable;
59 	}
60 
61 	switch (iomap.type) {
62 	case IOMAP_MAPPED:
63 		if (seg->iomode == IOMODE_READ)
64 			bex->es = PNFS_BLOCK_READ_DATA;
65 		else
66 			bex->es = PNFS_BLOCK_READWRITE_DATA;
67 		bex->soff = (iomap.blkno << 9);
68 		break;
69 	case IOMAP_UNWRITTEN:
70 		if (seg->iomode & IOMODE_RW) {
71 			/*
72 			 * Crack monkey special case from section 2.3.1.
73 			 */
74 			if (args->lg_minlength == 0) {
75 				dprintk("pnfsd: no soup for you!\n");
76 				goto out_layoutunavailable;
77 			}
78 
79 			bex->es = PNFS_BLOCK_INVALID_DATA;
80 			bex->soff = (iomap.blkno << 9);
81 			break;
82 		}
83 		/*FALLTHRU*/
84 	case IOMAP_HOLE:
85 		if (seg->iomode == IOMODE_READ) {
86 			bex->es = PNFS_BLOCK_NONE_DATA;
87 			break;
88 		}
89 		/*FALLTHRU*/
90 	case IOMAP_DELALLOC:
91 	default:
92 		WARN(1, "pnfsd: filesystem returned %d extent\n", iomap.type);
93 		goto out_layoutunavailable;
94 	}
95 
96 	error = nfsd4_set_deviceid(&bex->vol_id, fhp, device_generation);
97 	if (error)
98 		goto out_error;
99 	bex->foff = iomap.offset;
100 	bex->len = iomap.length;
101 
102 	seg->offset = iomap.offset;
103 	seg->length = iomap.length;
104 
105 	dprintk("GET: 0x%llx:0x%llx %d\n", bex->foff, bex->len, bex->es);
106 	return 0;
107 
108 out_error:
109 	seg->length = 0;
110 	return nfserrno(error);
111 out_layoutunavailable:
112 	seg->length = 0;
113 	return nfserr_layoutunavailable;
114 }
115 
116 static __be32
117 nfsd4_block_commit_blocks(struct inode *inode, struct nfsd4_layoutcommit *lcp,
118 		struct iomap *iomaps, int nr_iomaps)
119 {
120 	loff_t new_size = lcp->lc_last_wr + 1;
121 	struct iattr iattr = { .ia_valid = 0 };
122 	int error;
123 
124 	if (lcp->lc_mtime.tv_nsec == UTIME_NOW ||
125 	    timespec_compare(&lcp->lc_mtime, &inode->i_mtime) < 0)
126 		lcp->lc_mtime = current_fs_time(inode->i_sb);
127 	iattr.ia_valid |= ATTR_ATIME | ATTR_CTIME | ATTR_MTIME;
128 	iattr.ia_atime = iattr.ia_ctime = iattr.ia_mtime = lcp->lc_mtime;
129 
130 	if (new_size > i_size_read(inode)) {
131 		iattr.ia_valid |= ATTR_SIZE;
132 		iattr.ia_size = new_size;
133 	}
134 
135 	error = inode->i_sb->s_export_op->commit_blocks(inode, iomaps,
136 			nr_iomaps, &iattr);
137 	kfree(iomaps);
138 	return nfserrno(error);
139 }
140 
141 #ifdef CONFIG_NFSD_BLOCKLAYOUT
142 static int
143 nfsd4_block_get_device_info_simple(struct super_block *sb,
144 		struct nfsd4_getdeviceinfo *gdp)
145 {
146 	struct pnfs_block_deviceaddr *dev;
147 	struct pnfs_block_volume *b;
148 
149 	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
150 		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
151 	if (!dev)
152 		return -ENOMEM;
153 	gdp->gd_device = dev;
154 
155 	dev->nr_volumes = 1;
156 	b = &dev->volumes[0];
157 
158 	b->type = PNFS_BLOCK_VOLUME_SIMPLE;
159 	b->simple.sig_len = PNFS_BLOCK_UUID_LEN;
160 	return sb->s_export_op->get_uuid(sb, b->simple.sig, &b->simple.sig_len,
161 			&b->simple.offset);
162 }
163 
164 static __be32
165 nfsd4_block_proc_getdeviceinfo(struct super_block *sb,
166 		struct nfs4_client *clp,
167 		struct nfsd4_getdeviceinfo *gdp)
168 {
169 	if (sb->s_bdev != sb->s_bdev->bd_contains)
170 		return nfserr_inval;
171 	return nfserrno(nfsd4_block_get_device_info_simple(sb, gdp));
172 }
173 
174 static __be32
175 nfsd4_block_proc_layoutcommit(struct inode *inode,
176 		struct nfsd4_layoutcommit *lcp)
177 {
178 	struct iomap *iomaps;
179 	int nr_iomaps;
180 
181 	nr_iomaps = nfsd4_block_decode_layoutupdate(lcp->lc_up_layout,
182 			lcp->lc_up_len, &iomaps, 1 << inode->i_blkbits);
183 	if (nr_iomaps < 0)
184 		return nfserrno(nr_iomaps);
185 
186 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
187 }
188 
189 const struct nfsd4_layout_ops bl_layout_ops = {
190 	/*
191 	 * Pretend that we send notification to the client.  This is a blatant
192 	 * lie to force recent Linux clients to cache our device IDs.
193 	 * We rarely ever change the device ID, so the harm of leaking deviceids
194 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
195 	 * in this regard, but I filed errata 4119 for this a while ago, and
196 	 * hopefully the Linux client will eventually start caching deviceids
197 	 * without this again.
198 	 */
199 	.notify_types		=
200 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
201 	.proc_getdeviceinfo	= nfsd4_block_proc_getdeviceinfo,
202 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
203 	.proc_layoutget		= nfsd4_block_proc_layoutget,
204 	.encode_layoutget	= nfsd4_block_encode_layoutget,
205 	.proc_layoutcommit	= nfsd4_block_proc_layoutcommit,
206 };
207 #endif /* CONFIG_NFSD_BLOCKLAYOUT */
208 
209 #ifdef CONFIG_NFSD_SCSILAYOUT
210 static int nfsd4_scsi_identify_device(struct block_device *bdev,
211 		struct pnfs_block_volume *b)
212 {
213 	struct request_queue *q = bdev->bd_disk->queue;
214 	struct request *rq;
215 	size_t bufflen = 252, len, id_len;
216 	u8 *buf, *d, type, assoc;
217 	int error;
218 
219 	buf = kzalloc(bufflen, GFP_KERNEL);
220 	if (!buf)
221 		return -ENOMEM;
222 
223 	rq = blk_get_request(q, READ, GFP_KERNEL);
224 	if (IS_ERR(rq)) {
225 		error = -ENOMEM;
226 		goto out_free_buf;
227 	}
228 	blk_rq_set_block_pc(rq);
229 
230 	error = blk_rq_map_kern(q, rq, buf, bufflen, GFP_KERNEL);
231 	if (error)
232 		goto out_put_request;
233 
234 	rq->cmd[0] = INQUIRY;
235 	rq->cmd[1] = 1;
236 	rq->cmd[2] = 0x83;
237 	rq->cmd[3] = bufflen >> 8;
238 	rq->cmd[4] = bufflen & 0xff;
239 	rq->cmd_len = COMMAND_SIZE(INQUIRY);
240 
241 	error = blk_execute_rq(rq->q, NULL, rq, 1);
242 	if (error) {
243 		pr_err("pNFS: INQUIRY 0x83 failed with: %x\n",
244 			rq->errors);
245 		goto out_put_request;
246 	}
247 
248 	len = (buf[2] << 8) + buf[3] + 4;
249 	if (len > bufflen) {
250 		pr_err("pNFS: INQUIRY 0x83 response invalid (len = %zd)\n",
251 			len);
252 		goto out_put_request;
253 	}
254 
255 	d = buf + 4;
256 	for (d = buf + 4; d < buf + len; d += id_len + 4) {
257 		id_len = d[3];
258 		type = d[1] & 0xf;
259 		assoc = (d[1] >> 4) & 0x3;
260 
261 		/*
262 		 * We only care about a EUI-64 and NAA designator types
263 		 * with LU association.
264 		 */
265 		if (assoc != 0x00)
266 			continue;
267 		if (type != 0x02 && type != 0x03)
268 			continue;
269 		if (id_len != 8 && id_len != 12 && id_len != 16)
270 			continue;
271 
272 		b->scsi.code_set = PS_CODE_SET_BINARY;
273 		b->scsi.designator_type = type == 0x02 ?
274 			PS_DESIGNATOR_EUI64 : PS_DESIGNATOR_NAA;
275 		b->scsi.designator_len = id_len;
276 		memcpy(b->scsi.designator, d + 4, id_len);
277 
278 		/*
279 		 * If we found a 8 or 12 byte descriptor continue on to
280 		 * see if a 16 byte one is available.  If we find a
281 		 * 16 byte descriptor we're done.
282 		 */
283 		if (id_len == 16)
284 			break;
285 	}
286 
287 out_put_request:
288 	blk_put_request(rq);
289 out_free_buf:
290 	kfree(buf);
291 	return error;
292 }
293 
294 #define NFSD_MDS_PR_KEY		0x0100000000000000ULL
295 
296 /*
297  * We use the client ID as a unique key for the reservations.
298  * This allows us to easily fence a client when recalls fail.
299  */
300 static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp)
301 {
302 	return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id;
303 }
304 
305 static int
306 nfsd4_block_get_device_info_scsi(struct super_block *sb,
307 		struct nfs4_client *clp,
308 		struct nfsd4_getdeviceinfo *gdp)
309 {
310 	struct pnfs_block_deviceaddr *dev;
311 	struct pnfs_block_volume *b;
312 	const struct pr_ops *ops;
313 	int error;
314 
315 	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
316 		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
317 	if (!dev)
318 		return -ENOMEM;
319 	gdp->gd_device = dev;
320 
321 	dev->nr_volumes = 1;
322 	b = &dev->volumes[0];
323 
324 	b->type = PNFS_BLOCK_VOLUME_SCSI;
325 	b->scsi.pr_key = nfsd4_scsi_pr_key(clp);
326 
327 	error = nfsd4_scsi_identify_device(sb->s_bdev, b);
328 	if (error)
329 		return error;
330 
331 	ops = sb->s_bdev->bd_disk->fops->pr_ops;
332 	if (!ops) {
333 		pr_err("pNFS: device %s does not support PRs.\n",
334 			sb->s_id);
335 		return -EINVAL;
336 	}
337 
338 	error = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true);
339 	if (error) {
340 		pr_err("pNFS: failed to register key for device %s.\n",
341 			sb->s_id);
342 		return -EINVAL;
343 	}
344 
345 	error = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY,
346 			PR_EXCLUSIVE_ACCESS_REG_ONLY, 0);
347 	if (error) {
348 		pr_err("pNFS: failed to reserve device %s.\n",
349 			sb->s_id);
350 		return -EINVAL;
351 	}
352 
353 	return 0;
354 }
355 
356 static __be32
357 nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb,
358 		struct nfs4_client *clp,
359 		struct nfsd4_getdeviceinfo *gdp)
360 {
361 	if (sb->s_bdev != sb->s_bdev->bd_contains)
362 		return nfserr_inval;
363 	return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp));
364 }
365 static __be32
366 nfsd4_scsi_proc_layoutcommit(struct inode *inode,
367 		struct nfsd4_layoutcommit *lcp)
368 {
369 	struct iomap *iomaps;
370 	int nr_iomaps;
371 
372 	nr_iomaps = nfsd4_scsi_decode_layoutupdate(lcp->lc_up_layout,
373 			lcp->lc_up_len, &iomaps, 1 << inode->i_blkbits);
374 	if (nr_iomaps < 0)
375 		return nfserrno(nr_iomaps);
376 
377 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
378 }
379 
380 static void
381 nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls)
382 {
383 	struct nfs4_client *clp = ls->ls_stid.sc_client;
384 	struct block_device *bdev = ls->ls_file->f_path.mnt->mnt_sb->s_bdev;
385 
386 	bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY,
387 			nfsd4_scsi_pr_key(clp), 0, true);
388 }
389 
390 const struct nfsd4_layout_ops scsi_layout_ops = {
391 	/*
392 	 * Pretend that we send notification to the client.  This is a blatant
393 	 * lie to force recent Linux clients to cache our device IDs.
394 	 * We rarely ever change the device ID, so the harm of leaking deviceids
395 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
396 	 * in this regard, but I filed errata 4119 for this a while ago, and
397 	 * hopefully the Linux client will eventually start caching deviceids
398 	 * without this again.
399 	 */
400 	.notify_types		=
401 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
402 	.proc_getdeviceinfo	= nfsd4_scsi_proc_getdeviceinfo,
403 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
404 	.proc_layoutget		= nfsd4_block_proc_layoutget,
405 	.encode_layoutget	= nfsd4_block_encode_layoutget,
406 	.proc_layoutcommit	= nfsd4_scsi_proc_layoutcommit,
407 	.fence_client		= nfsd4_scsi_fence_client,
408 };
409 #endif /* CONFIG_NFSD_SCSILAYOUT */
410