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
 22static __be32
 23nfsd4_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
110out_error:
111	seg->length = 0;
112	return nfserrno(error);
113out_layoutunavailable:
114	seg->length = 0;
115	return nfserr_layoutunavailable;
116}
117
118static __be32
119nfsd4_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	    timespec_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
144static int
145nfsd4_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
166static __be32
167nfsd4_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
177static __be32
178nfsd4_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
192const 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
213static 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	size_t bufflen = 252, len, id_len;
220	u8 *buf, *d, type, assoc;
221	int error;
222
223	if (WARN_ON_ONCE(!blk_queue_scsi_passthrough(q)))
224		return -EINVAL;
225
226	buf = kzalloc(bufflen, GFP_KERNEL);
227	if (!buf)
228		return -ENOMEM;
229
230	rq = blk_get_request(q, REQ_OP_SCSI_IN, GFP_KERNEL);
231	if (IS_ERR(rq)) {
232		error = -ENOMEM;
233		goto out_free_buf;
234	}
235	req = scsi_req(rq);
236
237	error = blk_rq_map_kern(q, rq, buf, bufflen, GFP_KERNEL);
238	if (error)
239		goto out_put_request;
240
241	req->cmd[0] = INQUIRY;
242	req->cmd[1] = 1;
243	req->cmd[2] = 0x83;
244	req->cmd[3] = bufflen >> 8;
245	req->cmd[4] = bufflen & 0xff;
246	req->cmd_len = COMMAND_SIZE(INQUIRY);
247
248	blk_execute_rq(rq->q, NULL, rq, 1);
249	if (req->result) {
250		pr_err("pNFS: INQUIRY 0x83 failed with: %x\n",
251			req->result);
252		error = -EIO;
253		goto out_put_request;
254	}
255
256	len = (buf[2] << 8) + buf[3] + 4;
257	if (len > bufflen) {
258		pr_err("pNFS: INQUIRY 0x83 response invalid (len = %zd)\n",
259			len);
260		goto out_put_request;
261	}
262
263	d = buf + 4;
264	for (d = buf + 4; d < buf + len; d += id_len + 4) {
265		id_len = d[3];
266		type = d[1] & 0xf;
267		assoc = (d[1] >> 4) & 0x3;
268
269		/*
270		 * We only care about a EUI-64 and NAA designator types
271		 * with LU association.
272		 */
273		if (assoc != 0x00)
274			continue;
275		if (type != 0x02 && type != 0x03)
276			continue;
277		if (id_len != 8 && id_len != 12 && id_len != 16)
278			continue;
279
280		b->scsi.code_set = PS_CODE_SET_BINARY;
281		b->scsi.designator_type = type == 0x02 ?
282			PS_DESIGNATOR_EUI64 : PS_DESIGNATOR_NAA;
283		b->scsi.designator_len = id_len;
284		memcpy(b->scsi.designator, d + 4, id_len);
285
286		/*
287		 * If we found a 8 or 12 byte descriptor continue on to
288		 * see if a 16 byte one is available.  If we find a
289		 * 16 byte descriptor we're done.
290		 */
291		if (id_len == 16)
292			break;
293	}
294
295out_put_request:
296	blk_put_request(rq);
297out_free_buf:
298	kfree(buf);
299	return error;
300}
301
302#define NFSD_MDS_PR_KEY		0x0100000000000000ULL
303
304/*
305 * We use the client ID as a unique key for the reservations.
306 * This allows us to easily fence a client when recalls fail.
307 */
308static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp)
309{
310	return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id;
311}
312
313static int
314nfsd4_block_get_device_info_scsi(struct super_block *sb,
315		struct nfs4_client *clp,
316		struct nfsd4_getdeviceinfo *gdp)
317{
318	struct pnfs_block_deviceaddr *dev;
319	struct pnfs_block_volume *b;
320	const struct pr_ops *ops;
321	int error;
322
323	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
324		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
325	if (!dev)
326		return -ENOMEM;
327	gdp->gd_device = dev;
328
329	dev->nr_volumes = 1;
330	b = &dev->volumes[0];
331
332	b->type = PNFS_BLOCK_VOLUME_SCSI;
333	b->scsi.pr_key = nfsd4_scsi_pr_key(clp);
334
335	error = nfsd4_scsi_identify_device(sb->s_bdev, b);
336	if (error)
337		return error;
338
339	ops = sb->s_bdev->bd_disk->fops->pr_ops;
340	if (!ops) {
341		pr_err("pNFS: device %s does not support PRs.\n",
342			sb->s_id);
343		return -EINVAL;
344	}
345
346	error = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true);
347	if (error) {
348		pr_err("pNFS: failed to register key for device %s.\n",
349			sb->s_id);
350		return -EINVAL;
351	}
352
353	error = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY,
354			PR_EXCLUSIVE_ACCESS_REG_ONLY, 0);
355	if (error) {
356		pr_err("pNFS: failed to reserve device %s.\n",
357			sb->s_id);
358		return -EINVAL;
359	}
360
361	return 0;
362}
363
364static __be32
365nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb,
366		struct svc_rqst *rqstp,
367		struct nfs4_client *clp,
368		struct nfsd4_getdeviceinfo *gdp)
369{
370	if (sb->s_bdev != sb->s_bdev->bd_contains)
371		return nfserr_inval;
372	return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp));
373}
374static __be32
375nfsd4_scsi_proc_layoutcommit(struct inode *inode,
376		struct nfsd4_layoutcommit *lcp)
377{
378	struct iomap *iomaps;
379	int nr_iomaps;
380
381	nr_iomaps = nfsd4_scsi_decode_layoutupdate(lcp->lc_up_layout,
382			lcp->lc_up_len, &iomaps, i_blocksize(inode));
383	if (nr_iomaps < 0)
384		return nfserrno(nr_iomaps);
385
386	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
387}
388
389static void
390nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls)
391{
392	struct nfs4_client *clp = ls->ls_stid.sc_client;
393	struct block_device *bdev = ls->ls_file->f_path.mnt->mnt_sb->s_bdev;
394
395	bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY,
396			nfsd4_scsi_pr_key(clp), 0, true);
397}
398
399const struct nfsd4_layout_ops scsi_layout_ops = {
400	/*
401	 * Pretend that we send notification to the client.  This is a blatant
402	 * lie to force recent Linux clients to cache our device IDs.
403	 * We rarely ever change the device ID, so the harm of leaking deviceids
404	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
405	 * in this regard, but I filed errata 4119 for this a while ago, and
406	 * hopefully the Linux client will eventually start caching deviceids
407	 * without this again.
408	 */
409	.notify_types		=
410			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
411	.proc_getdeviceinfo	= nfsd4_scsi_proc_getdeviceinfo,
412	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
413	.proc_layoutget		= nfsd4_block_proc_layoutget,
414	.encode_layoutget	= nfsd4_block_encode_layoutget,
415	.proc_layoutcommit	= nfsd4_scsi_proc_layoutcommit,
416	.fence_client		= nfsd4_scsi_fence_client,
417};
418#endif /* CONFIG_NFSD_SCSILAYOUT */