1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/lov/lov_pack.c
41 * (Un)packing of OST/MDS requests
43 * Author: Andreas Dilger <adilger@clusterfs.com>
46 #define DEBUG_SUBSYSTEM S_LOV
48 #include <liblustre.h>
51 #include <lustre_net.h>
54 #include <obd_class.h>
55 #include <obd_support.h>
56 #include <lustre/lustre_user.h>
58 #include "lov_internal.h"
60 static void lov_dump_lmm_common(int level, void *lmmp)
62 struct lov_mds_md *lmm = lmmp;
64 CDEBUG(level, "objid "LPX64", magic 0x%08x, pattern %#x\n",
65 (__u64)le64_to_cpu(lmm->lmm_object_id),
66 le32_to_cpu(lmm->lmm_magic),
67 le32_to_cpu(lmm->lmm_pattern));
68 CDEBUG(level,"stripe_size %u, stripe_count %u, layout_gen %u\n",
69 le32_to_cpu(lmm->lmm_stripe_size),
70 le16_to_cpu(lmm->lmm_stripe_count),
71 le16_to_cpu(lmm->lmm_layout_gen));
74 static void lov_dump_lmm_objects(int level, struct lov_ost_data *lod,
79 if (stripe_count > LOV_V1_INSANE_STRIPE_COUNT) {
80 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
81 stripe_count, LOV_V1_INSANE_STRIPE_COUNT);
84 for (i = 0; i < stripe_count; ++i, ++lod) {
85 CDEBUG(level, "stripe %u idx %u subobj "LPX64"/"LPX64"\n", i,
86 le32_to_cpu(lod->l_ost_idx),
87 (__u64)le64_to_cpu(lod->l_object_seq),
88 (__u64)le64_to_cpu(lod->l_object_id));
92 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
94 lov_dump_lmm_common(level, lmm);
95 lov_dump_lmm_objects(level, lmm->lmm_objects,
96 le16_to_cpu(lmm->lmm_stripe_count));
99 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
101 lov_dump_lmm_common(level, lmm);
102 CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
103 lov_dump_lmm_objects(level, lmm->lmm_objects,
104 le16_to_cpu(lmm->lmm_stripe_count));
107 void lov_dump_lmm(int level, void *lmm)
111 magic = ((struct lov_mds_md_v1 *)(lmm))->lmm_magic;
114 return lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)(lmm));
116 return lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)(lmm));
118 CERROR("Cannot recognize lmm_magic %x", magic);
123 #define LMM_ASSERT(test) \
125 if (!(test)) lov_dump_lmm(D_ERROR, lmm); \
126 LASSERT(test); /* so we know what assertion failed */ \
129 /* Pack LOV object metadata for disk storage. It is packed in LE byte
130 * order and is opaque to the networking layer.
132 * XXX In the future, this will be enhanced to get the EA size from the
133 * underlying OSC device(s) to get their EA sizes so we can stack
134 * LOVs properly. For now lov_mds_md_size() just assumes one obd_id
137 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
138 struct lov_stripe_md *lsm)
140 struct obd_device *obd = class_exp2obd(exp);
141 struct lov_obd *lov = &obd->u.lov;
142 struct lov_mds_md_v1 *lmmv1;
143 struct lov_mds_md_v3 *lmmv3;
145 struct lov_ost_data_v1 *lmm_objects;
146 int lmm_size, lmm_magic;
151 lmm_magic = lsm->lsm_magic;
154 lmm_magic = le32_to_cpu((*lmmp)->lmm_magic);
156 /* lsm == NULL and lmmp == NULL */
157 lmm_magic = LOV_MAGIC;
160 if ((lmm_magic != LOV_MAGIC_V1) &&
161 (lmm_magic != LOV_MAGIC_V3)) {
162 CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
163 lmm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
169 /* If we are just sizing the EA, limit the stripe count
170 * to the actual number of OSTs in this filesystem. */
172 stripe_count = lov_get_stripecnt(lov, lmm_magic,
173 lsm->lsm_stripe_count);
174 lsm->lsm_stripe_count = stripe_count;
176 stripe_count = lsm->lsm_stripe_count;
179 /* No need to allocate more than maximum supported stripes.
180 * Anyway, this is pretty inaccurate since ld_tgt_count now
181 * represents max index and we should rely on the actual number
183 stripe_count = lov_mds_md_stripecnt(lov->lov_ocd.ocd_max_easize,
185 if (stripe_count > lov->desc.ld_tgt_count)
186 stripe_count = lov->desc.ld_tgt_count;
189 /* XXX LOV STACKING call into osc for sizes */
190 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
196 stripe_count = le16_to_cpu((*lmmp)->lmm_stripe_count);
197 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
198 OBD_FREE_LARGE(*lmmp, lmm_size);
204 OBD_ALLOC_LARGE(*lmmp, lmm_size);
209 CDEBUG(D_INFO, "lov_packmd: LOV_MAGIC 0x%08X, lmm_size = %d \n",
210 lmm_magic, lmm_size);
213 lmmv3 = (struct lov_mds_md_v3 *)*lmmp;
214 if (lmm_magic == LOV_MAGIC_V3)
215 lmmv3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3);
217 lmmv1->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
222 /* lmmv1 and lmmv3 point to the same struct and have the
225 lmmv1->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
226 lmmv1->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
227 lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
228 lmmv1->lmm_stripe_count = cpu_to_le16(stripe_count);
229 lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
230 lmmv1->lmm_layout_gen = cpu_to_le16(lsm->lsm_layout_gen);
231 if (lsm->lsm_magic == LOV_MAGIC_V3) {
232 strncpy(lmmv3->lmm_pool_name, lsm->lsm_pool_name,
234 lmm_objects = lmmv3->lmm_objects;
236 lmm_objects = lmmv1->lmm_objects;
239 for (i = 0; i < stripe_count; i++) {
240 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
241 /* XXX LOV STACKING call down to osc_packmd() to do packing */
242 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
243 lmmv1->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
244 lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
245 lmm_objects[i].l_object_seq = cpu_to_le64(loi->loi_seq);
246 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
247 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
253 /* Find the max stripecount we should use */
254 __u16 lov_get_stripecnt(struct lov_obd *lov, __u32 magic, __u16 stripe_count)
256 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
259 stripe_count = lov->desc.ld_default_stripe_count;
260 if (stripe_count > lov->desc.ld_active_tgt_count)
261 stripe_count = lov->desc.ld_active_tgt_count;
265 /* stripe count is based on whether ldiskfs can handle
267 if (lov->lov_ocd.ocd_connect_flags & OBD_CONNECT_MAX_EASIZE &&
268 lov->lov_ocd.ocd_max_easize)
269 max_stripes = lov_mds_md_stripecnt(lov->lov_ocd.ocd_max_easize,
272 if (stripe_count > max_stripes)
273 stripe_count = max_stripes;
279 static int lov_verify_lmm(void *lmm, int lmm_bytes, __u16 *stripe_count)
283 if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
287 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
288 le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
289 sz = lmm_bytes * 2 + 1;
290 OBD_ALLOC_LARGE(buffer, sz);
291 if (buffer != NULL) {
294 for (i = 0; i < lmm_bytes; i++)
295 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
297 CERROR("%s\n", buffer);
298 OBD_FREE_LARGE(buffer, sz);
302 rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
303 lmm_bytes, stripe_count);
307 int lov_alloc_memmd(struct lov_stripe_md **lsmp, __u16 stripe_count,
308 int pattern, int magic)
313 CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
315 *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
317 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
321 cfs_spin_lock_init(&(*lsmp)->lsm_lock);
322 (*lsmp)->lsm_magic = magic;
323 (*lsmp)->lsm_stripe_count = stripe_count;
324 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
325 (*lsmp)->lsm_pattern = pattern;
326 (*lsmp)->lsm_pool_name[0] = '\0';
327 (*lsmp)->lsm_layout_gen = 0;
328 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
330 for (i = 0; i < stripe_count; i++)
331 loi_init((*lsmp)->lsm_oinfo[i]);
336 void lov_free_memmd(struct lov_stripe_md **lsmp)
338 struct lov_stripe_md *lsm = *lsmp;
340 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
341 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
347 /* Unpack LOV object metadata from disk storage. It is packed in LE byte
348 * order and is opaque to the networking layer.
350 int lov_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
351 struct lov_mds_md *lmm, int lmm_bytes)
353 struct obd_device *obd = class_exp2obd(exp);
354 struct lov_obd *lov = &obd->u.lov;
355 int rc = 0, lsm_size;
360 /* If passed an MDS struct use values from there, otherwise defaults */
362 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
365 magic = le32_to_cpu(lmm->lmm_magic);
368 stripe_count = lov_get_stripecnt(lov, magic, 0);
371 /* If we aren't passed an lsmp struct, we just want the size */
373 /* XXX LOV STACKING call into osc for sizes */
375 RETURN(lov_stripe_md_size(stripe_count));
377 /* If we are passed an allocated struct but nothing to unpack, free */
379 lov_free_memmd(lsmp);
383 lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
388 /* If we are passed a pointer but nothing to unpack, we only alloc */
392 LASSERT(lsm_op_find(magic) != NULL);
393 rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
395 lov_free_memmd(lsmp);
402 static int __lov_setstripe(struct obd_export *exp, int max_lmm_size,
403 struct lov_stripe_md **lsmp,
404 struct lov_user_md *lump)
406 struct obd_device *obd = class_exp2obd(exp);
407 struct lov_obd *lov = &obd->u.lov;
408 char buffer[sizeof(struct lov_user_md_v3)];
409 struct lov_user_md_v3 *lumv3 = (struct lov_user_md_v3 *)&buffer[0];
410 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&buffer[0];
416 rc = lov_lum_swab_if_needed(lumv3, &lmm_magic, lump);
420 /* in the rest of the tests, as *lumv1 and lumv3 have the same
421 * fields, we use lumv1 to avoid code duplication */
423 if (lumv1->lmm_pattern == 0) {
424 lumv1->lmm_pattern = lov->desc.ld_pattern ?
425 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
428 if (lumv1->lmm_pattern != LOV_PATTERN_RAID0) {
429 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
434 /* 64kB is the largest common page size we see (ia64), and matches the
436 if (lumv1->lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
437 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
438 lumv1->lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
439 lumv1->lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
442 if ((lumv1->lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
443 (lumv1->lmm_stripe_offset !=
444 (typeof(lumv1->lmm_stripe_offset))(-1))) {
445 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
446 lumv1->lmm_stripe_offset, lov->desc.ld_tgt_count);
449 stripe_count = lov_get_stripecnt(lov, lmm_magic,
450 lumv1->lmm_stripe_count);
453 int max_stripes = (max_lmm_size -
454 lov_mds_md_size(0, lmm_magic)) /
455 sizeof(struct lov_ost_data_v1);
456 if (unlikely(max_stripes < stripe_count)) {
457 CDEBUG(D_IOCTL, "stripe count reset from %d to %d\n",
458 stripe_count, max_stripes);
459 stripe_count = max_stripes;
463 if (lmm_magic == LOV_USER_MAGIC_V3) {
464 struct pool_desc *pool;
466 pool = lov_find_pool(lov, lumv3->lmm_pool_name);
468 if (lumv3->lmm_stripe_offset !=
469 (typeof(lumv3->lmm_stripe_offset))(-1)) {
470 rc = lov_check_index_in_pool(
471 lumv3->lmm_stripe_offset, pool);
473 lov_pool_putref(pool);
478 if (stripe_count > pool_tgt_count(pool))
479 stripe_count = pool_tgt_count(pool);
481 lov_pool_putref(pool);
485 rc = lov_alloc_memmd(lsmp, stripe_count, lumv1->lmm_pattern, lmm_magic);
488 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lumv1->lmm_stripe_offset;
489 (*lsmp)->lsm_stripe_size = lumv1->lmm_stripe_size;
490 if (lmm_magic == LOV_USER_MAGIC_V3)
491 strncpy((*lsmp)->lsm_pool_name, lumv3->lmm_pool_name,
499 /* Configure object striping information on a new file.
501 * @lmmu is a pointer to a user struct with one or more of the fields set to
502 * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
503 * lmm_stripe_offset, and lmm_stripe_pattern. lmm_magic must be LOV_MAGIC.
504 * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
506 int lov_setstripe(struct obd_export *exp, int max_lmm_size,
507 struct lov_stripe_md **lsmp, struct lov_user_md *lump)
515 rc = __lov_setstripe(exp, max_lmm_size, lsmp, lump);
520 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
521 struct lov_user_md *lump)
525 struct obd_export *oexp;
526 struct lov_obd *lov = &exp->exp_obd->u.lov;
528 struct lov_user_ost_data_v1 *lmm_objects;
532 if (lump->lmm_magic == LOV_USER_MAGIC_V3)
533 lmm_objects = ((struct lov_user_md_v3 *)lump)->lmm_objects;
535 lmm_objects = lump->lmm_objects;
537 for (i = 0; i < lump->lmm_stripe_count; i++) {
538 __u32 len = sizeof(last_id);
539 oexp = lov->lov_tgts[lmm_objects[i].l_ost_idx]->ltd_exp;
540 rc = obd_get_info(oexp, sizeof(KEY_LAST_ID), KEY_LAST_ID,
541 &len, &last_id, NULL);
544 if (lmm_objects[i].l_object_id > last_id) {
545 CERROR("Setting EA for object > than last id on "
546 "ost idx %d "LPD64" > "LPD64" \n",
547 lmm_objects[i].l_ost_idx,
548 lmm_objects[i].l_object_id, last_id);
553 rc = lov_setstripe(exp, 0, lsmp, lump);
557 for (i = 0; i < lump->lmm_stripe_count; i++) {
558 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
559 lmm_objects[i].l_ost_idx;
560 (*lsmp)->lsm_oinfo[i]->loi_id = lmm_objects[i].l_object_id;
561 (*lsmp)->lsm_oinfo[i]->loi_seq = lmm_objects[i].l_object_seq;
567 /* Retrieve object striping information.
569 * @lump is a pointer to an in-core struct with lmm_ost_count indicating
570 * the maximum number of OST indices which will fit in the user buffer.
571 * lmm_magic must be LOV_USER_MAGIC.
573 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
574 struct lov_user_md *lump)
577 * XXX huge struct allocated on stack.
579 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
580 struct lov_user_md_v3 lum;
581 struct lov_mds_md *lmmk = NULL;
591 * "Switch to kernel segment" to allow copying from kernel space by
592 * copy_{to,from}_user().
597 /* we only need the header part from user space to get lmm_magic and
598 * lmm_stripe_count, (the header part is common to v1 and v3) */
599 lum_size = sizeof(struct lov_user_md_v1);
600 if (cfs_copy_from_user(&lum, lump, lum_size))
601 GOTO(out_set, rc = -EFAULT);
602 else if ((lum.lmm_magic != LOV_USER_MAGIC) &&
603 (lum.lmm_magic != LOV_USER_MAGIC_V3))
604 GOTO(out_set, rc = -EINVAL);
606 if (lum.lmm_stripe_count &&
607 (lum.lmm_stripe_count < lsm->lsm_stripe_count)) {
608 /* Return right size of stripe to user */
609 lum.lmm_stripe_count = lsm->lsm_stripe_count;
610 rc = cfs_copy_to_user(lump, &lum, lum_size);
611 GOTO(out_set, rc = -EOVERFLOW);
613 rc = lov_packmd(exp, &lmmk, lsm);
619 /* FIXME: Bug 1185 - copy fields properly when structs change */
620 /* struct lov_user_md_v3 and struct lov_mds_md_v3 must be the same */
621 CLASSERT(sizeof(lum) == sizeof(struct lov_mds_md_v3));
622 CLASSERT(sizeof lum.lmm_objects[0] == sizeof lmmk->lmm_objects[0]);
624 if ((cpu_to_le32(LOV_MAGIC) != LOV_MAGIC) &&
625 ((lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V1)) ||
626 (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)))) {
627 lustre_swab_lov_mds_md(lmmk);
628 lustre_swab_lov_user_md_objects(
629 (struct lov_user_ost_data*)lmmk->lmm_objects,
630 lmmk->lmm_stripe_count);
632 if (lum.lmm_magic == LOV_USER_MAGIC) {
633 /* User request for v1, we need skip lmm_pool_name */
634 if (lmmk->lmm_magic == LOV_MAGIC_V3) {
635 memmove((char*)(&lmmk->lmm_stripe_count) +
636 sizeof(lmmk->lmm_stripe_count),
637 ((struct lov_mds_md_v3*)lmmk)->lmm_objects,
638 lmmk->lmm_stripe_count *
639 sizeof(struct lov_ost_data_v1));
640 lmm_size -= LOV_MAXPOOLNAME;
643 /* if v3 we just have to update the lum_size */
644 lum_size = sizeof(struct lov_user_md_v3);
647 /* User wasn't expecting this many OST entries */
648 if (lum.lmm_stripe_count == 0)
650 else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count)
651 GOTO(out_set, rc = -EOVERFLOW);
653 * Have a difference between lov_mds_md & lov_user_md.
654 * So we have to re-order the data before copy to user.
656 lum.lmm_stripe_count = lmmk->lmm_stripe_count;
657 lum.u.lum_layout_gen = lmmk->lmm_layout_gen;
658 ((struct lov_user_md*)lmmk)->u.lum_layout_gen = lum.u.lum_layout_gen;
659 ((struct lov_user_md*)lmmk)->lmm_stripe_count = lum.lmm_stripe_count;
660 if (cfs_copy_to_user(lump, lmmk, lmm_size))
663 obd_free_diskmd(exp, &lmmk);