xref: /openbmc/linux/fs/nfsd/blocklayout.c (revision bb0eb050)
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 #include <scsi/scsi_request.h>
14 
15 #include "blocklayoutxdr.h"
16 #include "pnfs.h"
17 
18 #define NFSDDBG_FACILITY	NFSDDBG_PNFS
19 
20 
21 static __be32
22 nfsd4_block_proc_layoutget(struct inode *inode, const struct svc_fh *fhp,
23 		struct nfsd4_layoutget *args)
24 {
25 	struct nfsd4_layout_seg *seg = &args->lg_seg;
26 	struct super_block *sb = inode->i_sb;
27 	u32 block_size = i_blocksize(inode);
28 	struct pnfs_block_extent *bex;
29 	struct iomap iomap;
30 	u32 device_generation = 0;
31 	int error;
32 
33 	if (seg->offset & (block_size - 1)) {
34 		dprintk("pnfsd: I/O misaligned\n");
35 		goto out_layoutunavailable;
36 	}
37 
38 	/*
39 	 * Some clients barf on non-zero block numbers for NONE or INVALID
40 	 * layouts, so make sure to zero the whole structure.
41 	 */
42 	error = -ENOMEM;
43 	bex = kzalloc(sizeof(*bex), GFP_KERNEL);
44 	if (!bex)
45 		goto out_error;
46 	args->lg_content = bex;
47 
48 	error = sb->s_export_op->map_blocks(inode, seg->offset, seg->length,
49 					    &iomap, seg->iomode != IOMODE_READ,
50 					    &device_generation);
51 	if (error) {
52 		if (error == -ENXIO)
53 			goto out_layoutunavailable;
54 		goto out_error;
55 	}
56 
57 	if (iomap.length < args->lg_minlength) {
58 		dprintk("pnfsd: extent smaller than minlength\n");
59 		goto out_layoutunavailable;
60 	}
61 
62 	switch (iomap.type) {
63 	case IOMAP_MAPPED:
64 		if (seg->iomode == IOMODE_READ)
65 			bex->es = PNFS_BLOCK_READ_DATA;
66 		else
67 			bex->es = PNFS_BLOCK_READWRITE_DATA;
68 		bex->soff = (iomap.blkno << 9);
69 		break;
70 	case IOMAP_UNWRITTEN:
71 		if (seg->iomode & IOMODE_RW) {
72 			/*
73 			 * Crack monkey special case from section 2.3.1.
74 			 */
75 			if (args->lg_minlength == 0) {
76 				dprintk("pnfsd: no soup for you!\n");
77 				goto out_layoutunavailable;
78 			}
79 
80 			bex->es = PNFS_BLOCK_INVALID_DATA;
81 			bex->soff = (iomap.blkno << 9);
82 			break;
83 		}
84 		/*FALLTHRU*/
85 	case IOMAP_HOLE:
86 		if (seg->iomode == IOMODE_READ) {
87 			bex->es = PNFS_BLOCK_NONE_DATA;
88 			break;
89 		}
90 		/*FALLTHRU*/
91 	case IOMAP_DELALLOC:
92 	default:
93 		WARN(1, "pnfsd: filesystem returned %d extent\n", iomap.type);
94 		goto out_layoutunavailable;
95 	}
96 
97 	error = nfsd4_set_deviceid(&bex->vol_id, fhp, device_generation);
98 	if (error)
99 		goto out_error;
100 	bex->foff = iomap.offset;
101 	bex->len = iomap.length;
102 
103 	seg->offset = iomap.offset;
104 	seg->length = iomap.length;
105 
106 	dprintk("GET: 0x%llx:0x%llx %d\n", bex->foff, bex->len, bex->es);
107 	return 0;
108 
109 out_error:
110 	seg->length = 0;
111 	return nfserrno(error);
112 out_layoutunavailable:
113 	seg->length = 0;
114 	return nfserr_layoutunavailable;
115 }
116 
117 static __be32
118 nfsd4_block_commit_blocks(struct inode *inode, struct nfsd4_layoutcommit *lcp,
119 		struct iomap *iomaps, int nr_iomaps)
120 {
121 	loff_t new_size = lcp->lc_last_wr + 1;
122 	struct iattr iattr = { .ia_valid = 0 };
123 	int error;
124 
125 	if (lcp->lc_mtime.tv_nsec == UTIME_NOW ||
126 	    timespec_compare(&lcp->lc_mtime, &inode->i_mtime) < 0)
127 		lcp->lc_mtime = current_time(inode);
128 	iattr.ia_valid |= ATTR_ATIME | ATTR_CTIME | ATTR_MTIME;
129 	iattr.ia_atime = iattr.ia_ctime = iattr.ia_mtime = lcp->lc_mtime;
130 
131 	if (new_size > i_size_read(inode)) {
132 		iattr.ia_valid |= ATTR_SIZE;
133 		iattr.ia_size = new_size;
134 	}
135 
136 	error = inode->i_sb->s_export_op->commit_blocks(inode, iomaps,
137 			nr_iomaps, &iattr);
138 	kfree(iomaps);
139 	return nfserrno(error);
140 }
141 
142 #ifdef CONFIG_NFSD_BLOCKLAYOUT
143 static int
144 nfsd4_block_get_device_info_simple(struct super_block *sb,
145 		struct nfsd4_getdeviceinfo *gdp)
146 {
147 	struct pnfs_block_deviceaddr *dev;
148 	struct pnfs_block_volume *b;
149 
150 	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
151 		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
152 	if (!dev)
153 		return -ENOMEM;
154 	gdp->gd_device = dev;
155 
156 	dev->nr_volumes = 1;
157 	b = &dev->volumes[0];
158 
159 	b->type = PNFS_BLOCK_VOLUME_SIMPLE;
160 	b->simple.sig_len = PNFS_BLOCK_UUID_LEN;
161 	return sb->s_export_op->get_uuid(sb, b->simple.sig, &b->simple.sig_len,
162 			&b->simple.offset);
163 }
164 
165 static __be32
166 nfsd4_block_proc_getdeviceinfo(struct super_block *sb,
167 		struct svc_rqst *rqstp,
168 		struct nfs4_client *clp,
169 		struct nfsd4_getdeviceinfo *gdp)
170 {
171 	if (sb->s_bdev != sb->s_bdev->bd_contains)
172 		return nfserr_inval;
173 	return nfserrno(nfsd4_block_get_device_info_simple(sb, gdp));
174 }
175 
176 static __be32
177 nfsd4_block_proc_layoutcommit(struct inode *inode,
178 		struct nfsd4_layoutcommit *lcp)
179 {
180 	struct iomap *iomaps;
181 	int nr_iomaps;
182 
183 	nr_iomaps = nfsd4_block_decode_layoutupdate(lcp->lc_up_layout,
184 			lcp->lc_up_len, &iomaps, i_blocksize(inode));
185 	if (nr_iomaps < 0)
186 		return nfserrno(nr_iomaps);
187 
188 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
189 }
190 
191 const struct nfsd4_layout_ops bl_layout_ops = {
192 	/*
193 	 * Pretend that we send notification to the client.  This is a blatant
194 	 * lie to force recent Linux clients to cache our device IDs.
195 	 * We rarely ever change the device ID, so the harm of leaking deviceids
196 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
197 	 * in this regard, but I filed errata 4119 for this a while ago, and
198 	 * hopefully the Linux client will eventually start caching deviceids
199 	 * without this again.
200 	 */
201 	.notify_types		=
202 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
203 	.proc_getdeviceinfo	= nfsd4_block_proc_getdeviceinfo,
204 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
205 	.proc_layoutget		= nfsd4_block_proc_layoutget,
206 	.encode_layoutget	= nfsd4_block_encode_layoutget,
207 	.proc_layoutcommit	= nfsd4_block_proc_layoutcommit,
208 };
209 #endif /* CONFIG_NFSD_BLOCKLAYOUT */
210 
211 #ifdef CONFIG_NFSD_SCSILAYOUT
212 static int nfsd4_scsi_identify_device(struct block_device *bdev,
213 		struct pnfs_block_volume *b)
214 {
215 	struct request_queue *q = bdev->bd_disk->queue;
216 	struct request *rq;
217 	struct scsi_request *req;
218 	size_t bufflen = 252, len, id_len;
219 	u8 *buf, *d, type, assoc;
220 	int error;
221 
222 	buf = kzalloc(bufflen, GFP_KERNEL);
223 	if (!buf)
224 		return -ENOMEM;
225 
226 	rq = blk_get_request(q, REQ_OP_SCSI_IN, GFP_KERNEL);
227 	if (IS_ERR(rq)) {
228 		error = -ENOMEM;
229 		goto out_free_buf;
230 	}
231 	req = scsi_req(rq);
232 	scsi_req_init(rq);
233 
234 	error = blk_rq_map_kern(q, rq, buf, bufflen, GFP_KERNEL);
235 	if (error)
236 		goto out_put_request;
237 
238 	req->cmd[0] = INQUIRY;
239 	req->cmd[1] = 1;
240 	req->cmd[2] = 0x83;
241 	req->cmd[3] = bufflen >> 8;
242 	req->cmd[4] = bufflen & 0xff;
243 	req->cmd_len = COMMAND_SIZE(INQUIRY);
244 
245 	blk_execute_rq(rq->q, NULL, rq, 1);
246 	if (req->result) {
247 		pr_err("pNFS: INQUIRY 0x83 failed with: %x\n",
248 			req->result);
249 		error = -EIO;
250 		goto out_put_request;
251 	}
252 
253 	len = (buf[2] << 8) + buf[3] + 4;
254 	if (len > bufflen) {
255 		pr_err("pNFS: INQUIRY 0x83 response invalid (len = %zd)\n",
256 			len);
257 		goto out_put_request;
258 	}
259 
260 	d = buf + 4;
261 	for (d = buf + 4; d < buf + len; d += id_len + 4) {
262 		id_len = d[3];
263 		type = d[1] & 0xf;
264 		assoc = (d[1] >> 4) & 0x3;
265 
266 		/*
267 		 * We only care about a EUI-64 and NAA designator types
268 		 * with LU association.
269 		 */
270 		if (assoc != 0x00)
271 			continue;
272 		if (type != 0x02 && type != 0x03)
273 			continue;
274 		if (id_len != 8 && id_len != 12 && id_len != 16)
275 			continue;
276 
277 		b->scsi.code_set = PS_CODE_SET_BINARY;
278 		b->scsi.designator_type = type == 0x02 ?
279 			PS_DESIGNATOR_EUI64 : PS_DESIGNATOR_NAA;
280 		b->scsi.designator_len = id_len;
281 		memcpy(b->scsi.designator, d + 4, id_len);
282 
283 		/*
284 		 * If we found a 8 or 12 byte descriptor continue on to
285 		 * see if a 16 byte one is available.  If we find a
286 		 * 16 byte descriptor we're done.
287 		 */
288 		if (id_len == 16)
289 			break;
290 	}
291 
292 out_put_request:
293 	blk_put_request(rq);
294 out_free_buf:
295 	kfree(buf);
296 	return error;
297 }
298 
299 #define NFSD_MDS_PR_KEY		0x0100000000000000ULL
300 
301 /*
302  * We use the client ID as a unique key for the reservations.
303  * This allows us to easily fence a client when recalls fail.
304  */
305 static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp)
306 {
307 	return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id;
308 }
309 
310 static int
311 nfsd4_block_get_device_info_scsi(struct super_block *sb,
312 		struct nfs4_client *clp,
313 		struct nfsd4_getdeviceinfo *gdp)
314 {
315 	struct pnfs_block_deviceaddr *dev;
316 	struct pnfs_block_volume *b;
317 	const struct pr_ops *ops;
318 	int error;
319 
320 	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
321 		      sizeof(struct pnfs_block_volume), 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 	error = nfsd4_scsi_identify_device(sb->s_bdev, b);
333 	if (error)
334 		return error;
335 
336 	ops = sb->s_bdev->bd_disk->fops->pr_ops;
337 	if (!ops) {
338 		pr_err("pNFS: device %s does not support PRs.\n",
339 			sb->s_id);
340 		return -EINVAL;
341 	}
342 
343 	error = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true);
344 	if (error) {
345 		pr_err("pNFS: failed to register key for device %s.\n",
346 			sb->s_id);
347 		return -EINVAL;
348 	}
349 
350 	error = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY,
351 			PR_EXCLUSIVE_ACCESS_REG_ONLY, 0);
352 	if (error) {
353 		pr_err("pNFS: failed to reserve device %s.\n",
354 			sb->s_id);
355 		return -EINVAL;
356 	}
357 
358 	return 0;
359 }
360 
361 static __be32
362 nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb,
363 		struct svc_rqst *rqstp,
364 		struct nfs4_client *clp,
365 		struct nfsd4_getdeviceinfo *gdp)
366 {
367 	if (sb->s_bdev != sb->s_bdev->bd_contains)
368 		return nfserr_inval;
369 	return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp));
370 }
371 static __be32
372 nfsd4_scsi_proc_layoutcommit(struct inode *inode,
373 		struct nfsd4_layoutcommit *lcp)
374 {
375 	struct iomap *iomaps;
376 	int nr_iomaps;
377 
378 	nr_iomaps = nfsd4_scsi_decode_layoutupdate(lcp->lc_up_layout,
379 			lcp->lc_up_len, &iomaps, i_blocksize(inode));
380 	if (nr_iomaps < 0)
381 		return nfserrno(nr_iomaps);
382 
383 	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
384 }
385 
386 static void
387 nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls)
388 {
389 	struct nfs4_client *clp = ls->ls_stid.sc_client;
390 	struct block_device *bdev = ls->ls_file->f_path.mnt->mnt_sb->s_bdev;
391 
392 	bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY,
393 			nfsd4_scsi_pr_key(clp), 0, true);
394 }
395 
396 const struct nfsd4_layout_ops scsi_layout_ops = {
397 	/*
398 	 * Pretend that we send notification to the client.  This is a blatant
399 	 * lie to force recent Linux clients to cache our device IDs.
400 	 * We rarely ever change the device ID, so the harm of leaking deviceids
401 	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
402 	 * in this regard, but I filed errata 4119 for this a while ago, and
403 	 * hopefully the Linux client will eventually start caching deviceids
404 	 * without this again.
405 	 */
406 	.notify_types		=
407 			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
408 	.proc_getdeviceinfo	= nfsd4_scsi_proc_getdeviceinfo,
409 	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
410 	.proc_layoutget		= nfsd4_block_proc_layoutget,
411 	.encode_layoutget	= nfsd4_block_encode_layoutget,
412 	.proc_layoutcommit	= nfsd4_scsi_proc_layoutcommit,
413 	.fence_client		= nfsd4_scsi_fence_client,
414 };
415 #endif /* CONFIG_NFSD_SCSILAYOUT */
416