blocklayout.c

  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/slab.h>
  8#include <linux/pr.h>
  9
 10#include <linux/nfsd/debug.h>
 11
 12#include "blocklayoutxdr.h"
 13#include "pnfs.h"
 14#include "filecache.h"
 15#include "vfs.h"
 16#include "trace.h"
 17
 18#define NFSDDBG_FACILITY	NFSDDBG_PNFS
 19
 20
 21/*
 22 * Get an extent from the file system that starts at offset or below
 23 * and may be shorter than the requested length.
 24 */
 25static __be32
 26nfsd4_block_map_extent(struct inode *inode, const struct svc_fh *fhp,
 27		u64 offset, u64 length, u32 iomode, u64 minlength,
 28		struct pnfs_block_extent *bex)
 29{
 30	struct super_block *sb = inode->i_sb;
 31	struct iomap iomap;
 32	u32 device_generation = 0;
 33	int error;
 34
 35	error = sb->s_export_op->map_blocks(inode, offset, length, &iomap,
 36			iomode != IOMODE_READ, &device_generation);
 37	if (error) {
 38		if (error == -ENXIO)
 39			return nfserr_layoutunavailable;
 40		return nfserrno(error);
 41	}
 42
 43	switch (iomap.type) {
 44	case IOMAP_MAPPED:
 45		if (iomode == IOMODE_READ)
 46			bex->es = PNFS_BLOCK_READ_DATA;
 47		else
 48			bex->es = PNFS_BLOCK_READWRITE_DATA;
 49		bex->soff = iomap.addr;
 50		break;
 51	case IOMAP_UNWRITTEN:
 52		if (iomode & IOMODE_RW) {
 53			/*
 54			 * Crack monkey special case from section 2.3.1.
 55			 */
 56			if (minlength == 0) {
 57				dprintk("pnfsd: no soup for you!\n");
 58				return nfserr_layoutunavailable;
 59			}
 60
 61			bex->es = PNFS_BLOCK_INVALID_DATA;
 62			bex->soff = iomap.addr;
 63			break;
 64		}
 65		fallthrough;
 66	case IOMAP_HOLE:
 67		if (iomode == IOMODE_READ) {
 68			bex->es = PNFS_BLOCK_NONE_DATA;
 69			break;
 70		}
 71		fallthrough;
 72	case IOMAP_DELALLOC:
 73	default:
 74		WARN(1, "pnfsd: filesystem returned %d extent\n", iomap.type);
 75		return nfserr_layoutunavailable;
 76	}
 77
 78	error = nfsd4_set_deviceid(&bex->vol_id, fhp, device_generation);
 79	if (error)
 80		return nfserrno(error);
 81
 82	bex->foff = iomap.offset;
 83	bex->len = iomap.length;
 84	return nfs_ok;
 85}
 86
 87static __be32
 88nfsd4_block_proc_layoutget(struct svc_rqst *rqstp, struct inode *inode,
 89		const struct svc_fh *fhp, struct nfsd4_layoutget *args)
 90{
 91	struct nfsd4_layout_seg *seg = &args->lg_seg;
 92	struct pnfs_block_layout *bl;
 93	struct pnfs_block_extent *first_bex, *last_bex;
 94	u64 offset = seg->offset, length = seg->length;
 95	u32 i, nr_extents_max, block_size = i_blocksize(inode);
 96	__be32 nfserr;
 97
 98	if (locks_in_grace(SVC_NET(rqstp)))
 99		return nfserr_grace;
100
101	nfserr = nfserr_layoutunavailable;
102	if (seg->offset & (block_size - 1)) {
103		dprintk("pnfsd: I/O misaligned\n");
104		goto out_error;
105	}
106
107	/*
108	 * RFC 8881, section 3.3.17:
109	 *   The layout4 data type defines a layout for a file.
110	 *
111	 * RFC 8881, section 18.43.3:
112	 *   The loga_maxcount field specifies the maximum layout size
113	 *   (in bytes) that the client can handle. If the size of the
114	 *   layout structure exceeds the size specified by maxcount,
115	 *   the metadata server will return the NFS4ERR_TOOSMALL error.
116	 */
117	nfserr = nfserr_toosmall;
118	if (args->lg_maxcount < PNFS_BLOCK_LAYOUT4_SIZE +
119				PNFS_BLOCK_EXTENT_SIZE)
120		goto out_error;
121
122	/*
123	 * Limit the maximum layout size to avoid allocating
124	 * a large buffer on the server for each layout request.
125	 */
126	nr_extents_max = (min(args->lg_maxcount, PAGE_SIZE) -
127			  PNFS_BLOCK_LAYOUT4_SIZE) / PNFS_BLOCK_EXTENT_SIZE;
128
129	/*
130	 * Some clients barf on non-zero block numbers for NONE or INVALID
131	 * layouts, so make sure to zero the whole structure.
132	 */
133	nfserr = nfserrno(-ENOMEM);
134	bl = kzalloc(struct_size(bl, extents, nr_extents_max), GFP_KERNEL);
135	if (!bl)
136		goto out_error;
137	bl->nr_extents = nr_extents_max;
138	args->lg_content = bl;
139
140	for (i = 0; i < bl->nr_extents; i++) {
141		struct pnfs_block_extent *bex = bl->extents + i;
142		u64 bex_length;
143
144		nfserr = nfsd4_block_map_extent(inode, fhp, offset, length,
145				seg->iomode, args->lg_minlength, bex);
146		if (nfserr != nfs_ok)
147			goto out_error;
148
149		bex_length = bex->len - (offset - bex->foff);
150		if (bex_length >= length) {
151			bl->nr_extents = i + 1;
152			break;
153		}
154
155		offset = bex->foff + bex->len;
156		length -= bex_length;
157	}
158
159	first_bex = bl->extents;
160	last_bex = bl->extents + bl->nr_extents - 1;
161
162	nfserr = nfserr_layoutunavailable;
163	length = last_bex->foff + last_bex->len - seg->offset;
164	if (length < args->lg_minlength) {
165		dprintk("pnfsd: extent smaller than minlength\n");
166		goto out_error;
167	}
168
169	seg->offset = first_bex->foff;
170	seg->length = last_bex->foff - first_bex->foff + last_bex->len;
171	return nfs_ok;
172
173out_error:
174	seg->length = 0;
175	return nfserr;
176}
177
178static __be32
179nfsd4_block_commit_blocks(struct inode *inode, struct nfsd4_layoutcommit *lcp,
180		struct iomap *iomaps, int nr_iomaps)
181{
182	struct timespec64 mtime = inode_get_mtime(inode);
183	struct iattr iattr = { .ia_valid = 0 };
184	int error;
185
186	if (lcp->lc_mtime.tv_nsec == UTIME_NOW ||
187	    timespec64_compare(&lcp->lc_mtime, &mtime) < 0)
188		lcp->lc_mtime = current_time(inode);
189	iattr.ia_valid |= ATTR_ATIME | ATTR_CTIME | ATTR_MTIME;
190	iattr.ia_atime = iattr.ia_ctime = iattr.ia_mtime = lcp->lc_mtime;
191
192	if (lcp->lc_size_chg) {
193		iattr.ia_valid |= ATTR_SIZE;
194		iattr.ia_size = lcp->lc_newsize;
195	}
196
197	error = inode->i_sb->s_export_op->commit_blocks(inode, iomaps,
198			nr_iomaps, &iattr);
199	kfree(iomaps);
200	return nfserrno(error);
201}
202
203#ifdef CONFIG_NFSD_BLOCKLAYOUT
204static int
205nfsd4_block_get_device_info_simple(struct super_block *sb,
206		struct nfsd4_getdeviceinfo *gdp)
207{
208	struct pnfs_block_deviceaddr *dev;
209	struct pnfs_block_volume *b;
210
211	dev = kzalloc(struct_size(dev, volumes, 1), GFP_KERNEL);
212	if (!dev)
213		return -ENOMEM;
214	gdp->gd_device = dev;
215
216	dev->nr_volumes = 1;
217	b = &dev->volumes[0];
218
219	b->type = PNFS_BLOCK_VOLUME_SIMPLE;
220	b->simple.sig_len = PNFS_BLOCK_UUID_LEN;
221	return sb->s_export_op->get_uuid(sb, b->simple.sig, &b->simple.sig_len,
222			&b->simple.offset);
223}
224
225static __be32
226nfsd4_block_proc_getdeviceinfo(struct super_block *sb,
227		struct svc_rqst *rqstp,
228		struct nfs4_client *clp,
229		struct nfsd4_getdeviceinfo *gdp)
230{
231	if (bdev_is_partition(sb->s_bdev))
232		return nfserr_inval;
233	return nfserrno(nfsd4_block_get_device_info_simple(sb, gdp));
234}
235
236static __be32
237nfsd4_block_proc_layoutcommit(struct inode *inode, struct svc_rqst *rqstp,
238		struct nfsd4_layoutcommit *lcp)
239{
240	struct iomap *iomaps;
241	int nr_iomaps;
242	__be32 nfserr;
243
244	rqstp->rq_arg = lcp->lc_up_layout;
245	svcxdr_init_decode(rqstp);
246
247	nfserr = nfsd4_block_decode_layoutupdate(&rqstp->rq_arg_stream,
248			&iomaps, &nr_iomaps, i_blocksize(inode));
249	if (nfserr != nfs_ok)
250		return nfserr;
251
252	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
253}
254
255const struct nfsd4_layout_ops bl_layout_ops = {
256	/*
257	 * Pretend that we send notification to the client.  This is a blatant
258	 * lie to force recent Linux clients to cache our device IDs.
259	 * We rarely ever change the device ID, so the harm of leaking deviceids
260	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
261	 * in this regard, but I filed errata 4119 for this a while ago, and
262	 * hopefully the Linux client will eventually start caching deviceids
263	 * without this again.
264	 */
265	.notify_types		=
266			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
267	.proc_getdeviceinfo	= nfsd4_block_proc_getdeviceinfo,
268	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
269	.proc_layoutget		= nfsd4_block_proc_layoutget,
270	.encode_layoutget	= nfsd4_block_encode_layoutget,
271	.proc_layoutcommit	= nfsd4_block_proc_layoutcommit,
272};
273#endif /* CONFIG_NFSD_BLOCKLAYOUT */
274
275#ifdef CONFIG_NFSD_SCSILAYOUT
276#define NFSD_MDS_PR_KEY		0x0100000000000000ULL
277
278/*
279 * We use the client ID as a unique key for the reservations.
280 * This allows us to easily fence a client when recalls fail.
281 */
282static u64 nfsd4_scsi_pr_key(struct nfs4_client *clp)
283{
284	return ((u64)clp->cl_clientid.cl_boot << 32) | clp->cl_clientid.cl_id;
285}
286
287static const u8 designator_types[] = {
288	PS_DESIGNATOR_EUI64,
289	PS_DESIGNATOR_NAA,
290};
291
292static int
293nfsd4_block_get_unique_id(struct gendisk *disk, struct pnfs_block_volume *b)
294{
295	int ret, i;
296
297	for (i = 0; i < ARRAY_SIZE(designator_types); i++) {
298		u8 type = designator_types[i];
299
300		ret = disk->fops->get_unique_id(disk, b->scsi.designator, type);
301		if (ret > 0) {
302			b->scsi.code_set = PS_CODE_SET_BINARY;
303			b->scsi.designator_type = type;
304			b->scsi.designator_len = ret;
305			return 0;
306		}
307	}
308
309	return -EINVAL;
310}
311
312static int
313nfsd4_block_get_device_info_scsi(struct super_block *sb,
314		struct nfs4_client *clp,
315		struct nfsd4_getdeviceinfo *gdp)
316{
317	struct pnfs_block_deviceaddr *dev;
318	struct pnfs_block_volume *b;
319	const struct pr_ops *ops;
320	int ret;
321
322	dev = kzalloc(struct_size(dev, volumes, 1), GFP_KERNEL);
323	if (!dev)
324		return -ENOMEM;
325	gdp->gd_device = dev;
326
327	dev->nr_volumes = 1;
328	b = &dev->volumes[0];
329
330	b->type = PNFS_BLOCK_VOLUME_SCSI;
331	b->scsi.pr_key = nfsd4_scsi_pr_key(clp);
332
333	ret = nfsd4_block_get_unique_id(sb->s_bdev->bd_disk, b);
334	if (ret < 0)
335		goto out_free_dev;
336
337	ret = -EINVAL;
338	ops = sb->s_bdev->bd_disk->fops->pr_ops;
339	if (!ops) {
340		pr_err("pNFS: device %s does not support PRs.\n",
341			sb->s_id);
342		goto out_free_dev;
343	}
344
345	ret = ops->pr_register(sb->s_bdev, 0, NFSD_MDS_PR_KEY, true);
346	if (ret) {
347		pr_err("pNFS: failed to register key for device %s.\n",
348			sb->s_id);
349		goto out_free_dev;
350	}
351
352	ret = ops->pr_reserve(sb->s_bdev, NFSD_MDS_PR_KEY,
353			PR_EXCLUSIVE_ACCESS_REG_ONLY, 0);
354	if (ret) {
355		pr_err("pNFS: failed to reserve device %s.\n",
356			sb->s_id);
357		goto out_free_dev;
358	}
359
360	return 0;
361
362out_free_dev:
363	kfree(dev);
364	gdp->gd_device = NULL;
365	return ret;
366}
367
368static __be32
369nfsd4_scsi_proc_getdeviceinfo(struct super_block *sb,
370		struct svc_rqst *rqstp,
371		struct nfs4_client *clp,
372		struct nfsd4_getdeviceinfo *gdp)
373{
374	if (bdev_is_partition(sb->s_bdev))
375		return nfserr_inval;
376	return nfserrno(nfsd4_block_get_device_info_scsi(sb, clp, gdp));
377}
378static __be32
379nfsd4_scsi_proc_layoutcommit(struct inode *inode, struct svc_rqst *rqstp,
380		struct nfsd4_layoutcommit *lcp)
381{
382	struct iomap *iomaps;
383	int nr_iomaps;
384	__be32 nfserr;
385
386	rqstp->rq_arg = lcp->lc_up_layout;
387	svcxdr_init_decode(rqstp);
388
389	nfserr = nfsd4_scsi_decode_layoutupdate(&rqstp->rq_arg_stream,
390			&iomaps, &nr_iomaps, i_blocksize(inode));
391	if (nfserr != nfs_ok)
392		return nfserr;
393
394	return nfsd4_block_commit_blocks(inode, lcp, iomaps, nr_iomaps);
395}
396
397static void
398nfsd4_scsi_fence_client(struct nfs4_layout_stateid *ls, struct nfsd_file *file)
399{
400	struct nfs4_client *clp = ls->ls_stid.sc_client;
401	struct block_device *bdev = file->nf_file->f_path.mnt->mnt_sb->s_bdev;
402	int status;
403
404	status = bdev->bd_disk->fops->pr_ops->pr_preempt(bdev, NFSD_MDS_PR_KEY,
405			nfsd4_scsi_pr_key(clp),
406			PR_EXCLUSIVE_ACCESS_REG_ONLY, true);
407	trace_nfsd_pnfs_fence(clp, bdev->bd_disk->disk_name, status);
408}
409
410const struct nfsd4_layout_ops scsi_layout_ops = {
411	/*
412	 * Pretend that we send notification to the client.  This is a blatant
413	 * lie to force recent Linux clients to cache our device IDs.
414	 * We rarely ever change the device ID, so the harm of leaking deviceids
415	 * for a while isn't too bad.  Unfortunately RFC5661 is a complete mess
416	 * in this regard, but I filed errata 4119 for this a while ago, and
417	 * hopefully the Linux client will eventually start caching deviceids
418	 * without this again.
419	 */
420	.notify_types		=
421			NOTIFY_DEVICEID4_DELETE | NOTIFY_DEVICEID4_CHANGE,
422	.proc_getdeviceinfo	= nfsd4_scsi_proc_getdeviceinfo,
423	.encode_getdeviceinfo	= nfsd4_block_encode_getdeviceinfo,
424	.proc_layoutget		= nfsd4_block_proc_layoutget,
425	.encode_layoutget	= nfsd4_block_encode_layoutget,
426	.proc_layoutcommit	= nfsd4_scsi_proc_layoutcommit,
427	.fence_client		= nfsd4_scsi_fence_client,
428};
429#endif /* CONFIG_NFSD_SCSILAYOUT */