1/*
  2 * Copyright (c) 2014-2016 Christoph Hellwig.
  3 */
  4#include <linux/exportfs.h>
  5#include <linux/genhd.h>
  6#include <linux/slab.h>
  7#include <linux/pr.h>
  8
  9#include <linux/nfsd/debug.h>
 10#include <scsi/scsi_proto.h>
 11#include <scsi/scsi_common.h>
 12
 13#include "blocklayoutxdr.h"
 14#include "pnfs.h"
 15
 16#define NFSDDBG_FACILITY	NFSDDBG_PNFS
 17
 18
 19static __be32
 20nfsd4_block_proc_layoutget(struct inode *inode, const struct svc_fh *fhp,
 21		struct nfsd4_layoutget *args)
 22{
 23	struct nfsd4_layout_seg *seg = &args->lg_seg;
 24	struct super_block *sb = inode->i_sb;
 25	u32 block_size = (1 << inode->i_blkbits);
 26	struct pnfs_block_extent *bex;
 27	struct iomap iomap;
 28	u32 device_generation = 0;
 29	int error;
 30
 31	if (seg->offset & (block_size - 1)) {
 32		dprintk("pnfsd: I/O misaligned\n");
 33		goto out_layoutunavailable;
 34	}
 35
 36	/*
 37	 * Some clients barf on non-zero block numbers for NONE or INVALID
 38	 * layouts, so make sure to zero the whole structure.
 39	 */
 40	error = -ENOMEM;
 41	bex = kzalloc(sizeof(*bex), GFP_KERNEL);
 42	if (!bex)
 43		goto out_error;
 44	args->lg_content = bex;
 45
 46	error = sb->s_export_op->map_blocks(inode, seg->offset, seg->length,
 47					    &iomap, seg->iomode != IOMODE_READ,
 48					    &device_generation);
 49	if (error) {
 50		if (error == -ENXIO)
 51			goto out_layoutunavailable;
 52		goto out_error;
 53	}
 54
 55	if (iomap.length < args->lg_minlength) {
 56		dprintk("pnfsd: extent smaller than minlength\n");
 57		goto out_layoutunavailable;
 58	}
 59
 60	switch (iomap.type) {
 61	case IOMAP_MAPPED:
 62		if (seg->iomode == IOMODE_READ)
 63			bex->es = PNFS_BLOCK_READ_DATA;
 64		else
 65			bex->es = PNFS_BLOCK_READWRITE_DATA;
 66		bex->soff = (iomap.blkno << 9);
 67		break;
 68	case IOMAP_UNWRITTEN:
 69		if (seg->iomode & IOMODE_RW) {
 70			/*
 71			 * Crack monkey special case from section 2.3.1.
 72			 */
 73			if (args->lg_minlength == 0) {
 74				dprintk("pnfsd: no soup for you!\n");
 75				goto out_layoutunavailable;
 76			}
 77
 78			bex->es = PNFS_BLOCK_INVALID_DATA;
 79			bex->soff = (iomap.blkno << 9);
 80			break;
 81		}
 82		/*FALLTHRU*/
 83	case IOMAP_HOLE:
 84		if (seg->iomode == IOMODE_READ) {
 85			bex->es = PNFS_BLOCK_NONE_DATA;
 86			break;
 87		}
 88		/*FALLTHRU*/
 89	case IOMAP_DELALLOC:
 90	default:
 91		WARN(1, "pnfsd: filesystem returned %d extent\n", iomap.type);
 92		goto out_layoutunavailable;
 93	}
 94
 95	error = nfsd4_set_deviceid(&bex->vol_id, fhp, device_generation);
 96	if (error)
 97		goto out_error;
 98	bex->foff = iomap.offset;
 99	bex->len = iomap.length;
100
101	seg->offset = iomap.offset;
102	seg->length = iomap.length;
103
104	dprintk("GET: 0x%llx:0x%llx %d\n", bex->foff, bex->len, bex->es);
105	return 0;
106
107out_error:
108	seg->length = 0;
109	return nfserrno(error);
110out_layoutunavailable:
111	seg->length = 0;
112	return nfserr_layoutunavailable;
113}
114
115static __be32
116nfsd4_block_commit_blocks(struct inode *inode, struct nfsd4_layoutcommit *lcp,
117		struct iomap *iomaps, int nr_iomaps)
118{
119	loff_t new_size = lcp->lc_last_wr + 1;
120	struct iattr iattr = { .ia_valid = 0 };
121	int error;
122
123	if (lcp->lc_mtime.tv_nsec == UTIME_NOW ||
124	    timespec_compare(&lcp->lc_mtime, &inode->i_mtime) < 0)
125		lcp->lc_mtime = current_fs_time(inode->i_sb);
126	iattr.ia_valid |= ATTR_ATIME | ATTR_CTIME | ATTR_MTIME;
127	iattr.ia_atime = iattr.ia_ctime = iattr.ia_mtime = lcp->lc_mtime;
128
129	if (new_size > i_size_read(inode)) {
130		iattr.ia_valid |= ATTR_SIZE;
131		iattr.ia_size = new_size;
132	}
133
134	error = inode->i_sb->s_export_op->commit_blocks(inode, iomaps,
135			nr_iomaps, &iattr);
136	kfree(iomaps);
137	return nfserrno(error);
138}
139
140#ifdef CONFIG_NFSD_BLOCKLAYOUT
141static int
142nfsd4_block_get_device_info_simple(struct super_block *sb,
143		struct nfsd4_getdeviceinfo *gdp)
144{
145	struct pnfs_block_deviceaddr *dev;
146	struct pnfs_block_volume *b;
147
148	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
149		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
150	if (!dev)
151		return -ENOMEM;
152	gdp->gd_device = dev;
153
154	dev->nr_volumes = 1;
155	b = &dev->volumes[0];
156
157	b->type = PNFS_BLOCK_VOLUME_SIMPLE;
158	b->simple.sig_len = PNFS_BLOCK_UUID_LEN;
159	return sb->s_export_op->get_uuid(sb, b->simple.sig, &b->simple.sig_len,
160			&b->simple.offset);
161}
162
163static __be32
164nfsd4_block_proc_getdeviceinfo(struct super_block *sb,
165		struct nfs4_client *clp,
166		struct nfsd4_getdeviceinfo *gdp)
167{
168	if (sb->s_bdev != sb->s_bdev->bd_contains)
169		return nfserr_inval;
170	return nfserrno(nfsd4_block_get_device_info_simple(sb, gdp));
171}
172
173static __be32
174nfsd4_block_proc_layoutcommit(struct inode *inode,
175		struct nfsd4_layoutcommit *lcp)
176{
177	struct iomap *iomaps;
178	int nr_iomaps;
179
180	nr_iomaps = nfsd4_block_decode_layoutupdate(lcp->lc_up_layout,
181			lcp->lc_up_len, &iomaps, 1 << inode->i_blkbits);
182	if (nr_iomaps < 0)
183		return nfserrno(nr_iomaps);
184
185	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
186}
187
188const struct nfsd4_layout_ops bl_layout_ops = {
189	/*
190	 * Pretend that we send notification to the client.  This is a blatant
191	 * lie to force recent Linux clients to cache our device IDs.
192	 * We rarely ever change the device ID, so the harm of leaking deviceids
193	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
194	 * in this regard, but I filed errata 4119 for this a while ago, and
195	 * hopefully the Linux client will eventually start caching deviceids
196	 * without this again.
197	 */
198	.notify_types		=
199			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
200	.proc_getdeviceinfo	= nfsd4_block_proc_getdeviceinfo,
201	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
202	.proc_layoutget		= nfsd4_block_proc_layoutget,
203	.encode_layoutget	= nfsd4_block_encode_layoutget,
204	.proc_layoutcommit	= nfsd4_block_proc_layoutcommit,
205};
206#endif /* CONFIG_NFSD_BLOCKLAYOUT */
207
208#ifdef CONFIG_NFSD_SCSILAYOUT
209static int nfsd4_scsi_identify_device(struct block_device *bdev,
210		struct pnfs_block_volume *b)
211{
212	struct request_queue *q = bdev->bd_disk->queue;
213	struct request *rq;
214	size_t bufflen = 252, len, id_len;
215	u8 *buf, *d, type, assoc;
216	int error;
217
218	buf = kzalloc(bufflen, GFP_KERNEL);
219	if (!buf)
220		return -ENOMEM;
221
222	rq = blk_get_request(q, READ, GFP_KERNEL);
223	if (IS_ERR(rq)) {
224		error = -ENOMEM;
225		goto out_free_buf;
226	}
227	blk_rq_set_block_pc(rq);
228
229	error = blk_rq_map_kern(q, rq, buf, bufflen, GFP_KERNEL);
230	if (error)
231		goto out_put_request;
232
233	rq->cmd[0] = INQUIRY;
234	rq->cmd[1] = 1;
235	rq->cmd[2] = 0x83;
236	rq->cmd[3] = bufflen >> 8;
237	rq->cmd[4] = bufflen & 0xff;
238	rq->cmd_len = COMMAND_SIZE(INQUIRY);
239
240	error = blk_execute_rq(rq->q, NULL, rq, 1);
241	if (error) {
242		pr_err("pNFS: INQUIRY 0x83 failed with: %x\n",
243			rq->errors);
244		goto out_put_request;
245	}
246
247	len = (buf[2] << 8) + buf[3] + 4;
248	if (len > bufflen) {
249		pr_err("pNFS: INQUIRY 0x83 response invalid (len = %zd)\n",
250			len);
251		goto out_put_request;
252	}
253
254	d = buf + 4;
255	for (d = buf + 4; d < buf + len; d += id_len + 4) {
256		id_len = d[3];
257		type = d[1] & 0xf;
258		assoc = (d[1] >> 4) & 0x3;
259
260		/*
261		 * We only care about a EUI-64 and NAA designator types
262		 * with LU association.
263		 */
264		if (assoc != 0x00)
265			continue;
266		if (type != 0x02 && type != 0x03)
267			continue;
268		if (id_len != 8 && id_len != 12 && id_len != 16)
269			continue;
270
271		b->scsi.code_set = PS_CODE_SET_BINARY;
272		b->scsi.designator_type = type == 0x02 ?
273			PS_DESIGNATOR_EUI64 : PS_DESIGNATOR_NAA;
274		b->scsi.designator_len = id_len;
275		memcpy(b->scsi.designator, d + 4, id_len);
276
277		/*
278		 * If we found a 8 or 12 byte descriptor continue on to
279		 * see if a 16 byte one is available.  If we find a
280		 * 16 byte descriptor we're done.
281		 */
282		if (id_len == 16)
283			break;
284	}
285
286out_put_request:
287	blk_put_request(rq);
288out_free_buf:
289	kfree(buf);
290	return error;
291}
292
293#define NFSD_MDS_PR_KEY		0x0100000000000000
294
295/*
296 * We use the client ID as a unique key for the reservations.
297 * This allows us to easily fence a client when recalls fail.
298 */
299static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp)
300{
301	return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id;
302}
303
304static int
305nfsd4_block_get_device_info_scsi(struct super_block *sb,
306		struct nfs4_client *clp,
307		struct nfsd4_getdeviceinfo *gdp)
308{
309	struct pnfs_block_deviceaddr *dev;
310	struct pnfs_block_volume *b;
311	const struct pr_ops *ops;
312	int error;
313
314	dev = kzalloc(sizeof(struct pnfs_block_deviceaddr) +
315		      sizeof(struct pnfs_block_volume), GFP_KERNEL);
316	if (!dev)
317		return -ENOMEM;
318	gdp->gd_device = dev;
319
320	dev->nr_volumes = 1;
321	b = &dev->volumes[0];
322
323	b->type = PNFS_BLOCK_VOLUME_SCSI;
324	b->scsi.pr_key = nfsd4_scsi_pr_key(clp);
325
326	error = nfsd4_scsi_identify_device(sb->s_bdev, b);
327	if (error)
328		return error;
329
330	ops = sb->s_bdev->bd_disk->fops->pr_ops;
331	if (!ops) {
332		pr_err("pNFS: device %s does not support PRs.\n",
333			sb->s_id);
334		return -EINVAL;
335	}
336
337	error = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true);
338	if (error) {
339		pr_err("pNFS: failed to register key for device %s.\n",
340			sb->s_id);
341		return -EINVAL;
342	}
343
344	error = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY,
345			PR_EXCLUSIVE_ACCESS_REG_ONLY, 0);
346	if (error) {
347		pr_err("pNFS: failed to reserve device %s.\n",
348			sb->s_id);
349		return -EINVAL;
350	}
351
352	return 0;
353}
354
355static __be32
356nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb,
357		struct nfs4_client *clp,
358		struct nfsd4_getdeviceinfo *gdp)
359{
360	if (sb->s_bdev != sb->s_bdev->bd_contains)
361		return nfserr_inval;
362	return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp));
363}
364static __be32
365nfsd4_scsi_proc_layoutcommit(struct inode *inode,
366		struct nfsd4_layoutcommit *lcp)
367{
368	struct iomap *iomaps;
369	int nr_iomaps;
370
371	nr_iomaps = nfsd4_scsi_decode_layoutupdate(lcp->lc_up_layout,
372			lcp->lc_up_len, &iomaps, 1 << inode->i_blkbits);
373	if (nr_iomaps < 0)
374		return nfserrno(nr_iomaps);
375
376	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
377}
378
379static void
380nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls)
381{
382	struct nfs4_client *clp = ls->ls_stid.sc_client;
383	struct block_device *bdev = ls->ls_file->f_path.mnt->mnt_sb->s_bdev;
384
385	bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY,
386			nfsd4_scsi_pr_key(clp), 0, true);
387}
388
389const struct nfsd4_layout_ops scsi_layout_ops = {
390	/*
391	 * Pretend that we send notification to the client.  This is a blatant
392	 * lie to force recent Linux clients to cache our device IDs.
393	 * We rarely ever change the device ID, so the harm of leaking deviceids
394	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
395	 * in this regard, but I filed errata 4119 for this a while ago, and
396	 * hopefully the Linux client will eventually start caching deviceids
397	 * without this again.
398	 */
399	.notify_types		=
400			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
401	.proc_getdeviceinfo	= nfsd4_scsi_proc_getdeviceinfo,
402	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
403	.proc_layoutget		= nfsd4_block_proc_layoutget,
404	.encode_layoutget	= nfsd4_block_encode_layoutget,
405	.proc_layoutcommit	= nfsd4_scsi_proc_layoutcommit,
406	.fence_client		= nfsd4_scsi_fence_client,
407};
408#endif /* CONFIG_NFSD_SCSILAYOUT */