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 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/lov/lov_pack.c
38 * (Un)packing of OST/MDS requests
40 * Author: Andreas Dilger <adilger@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_LOV
45 #include <liblustre.h>
48 #include <lustre_net.h>
51 #include <obd_class.h>
52 #include <obd_support.h>
53 #include <lustre/lustre_user.h>
55 #include "lov_internal.h"
57 static void lov_dump_lmm_common(int level, void *lmmp)
59 struct lov_mds_md *lmm = lmmp;
61 CDEBUG(level, "objid "LPX64", magic 0x%08x, pattern %#x\n",
62 (__u64)le64_to_cpu(lmm->lmm_object_id),
63 le32_to_cpu(lmm->lmm_magic),
64 le32_to_cpu(lmm->lmm_pattern));
65 CDEBUG(level,"stripe_size %u, stripe_count %u\n",
66 le32_to_cpu(lmm->lmm_stripe_size),
67 le32_to_cpu(lmm->lmm_stripe_count));
70 static void lov_dump_lmm_objects(int level, struct lov_ost_data *lod,
75 if (stripe_count > LOV_V1_INSANE_STRIPE_COUNT) {
76 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
77 stripe_count, LOV_V1_INSANE_STRIPE_COUNT);
80 for (i = 0; i < stripe_count; ++i, ++lod) {
81 CDEBUG(level, "stripe %u idx %u subobj "LPX64"/"LPX64"\n", i,
82 le32_to_cpu(lod->l_ost_idx),
83 (__u64)le64_to_cpu(lod->l_object_gr),
84 (__u64)le64_to_cpu(lod->l_object_id));
88 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
90 lov_dump_lmm_common(level, lmm);
91 lov_dump_lmm_objects(level, lmm->lmm_objects,
92 le32_to_cpu(lmm->lmm_stripe_count));
95 void lov_dump_lmm_join(int level, struct lov_mds_md_join *lmmj)
97 lov_dump_lmm_common(level, &lmmj->lmmj_md);
98 CDEBUG(level, "extent_count %u\n",
99 le32_to_cpu(lmmj->lmmj_extent_count));
102 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
104 lov_dump_lmm_common(level, lmm);
105 CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
106 lov_dump_lmm_objects(level, lmm->lmm_objects,
107 le32_to_cpu(lmm->lmm_stripe_count));
110 void lov_dump_lmm(int level, void *lmm)
114 magic = ((struct lov_mds_md_v1 *)(lmm))->lmm_magic;
117 return lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)(lmm));
119 return lov_dump_lmm_join(level, (struct lov_mds_md_join *)(lmm));
121 return lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)(lmm));
123 CERROR("Cannot recognize lmm_magic %x", magic);
128 #define LMM_ASSERT(test) \
130 if (!(test)) lov_dump_lmm(D_ERROR, lmm); \
131 LASSERT(test); /* so we know what assertion failed */ \
134 /* Pack LOV object metadata for disk storage. It is packed in LE byte
135 * order and is opaque to the networking layer.
137 * XXX In the future, this will be enhanced to get the EA size from the
138 * underlying OSC device(s) to get their EA sizes so we can stack
139 * LOVs properly. For now lov_mds_md_size() just assumes one obd_id
142 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
143 struct lov_stripe_md *lsm)
145 struct obd_device *obd = class_exp2obd(exp);
146 struct lov_obd *lov = &obd->u.lov;
147 struct lov_mds_md_v1 *lmmv1;
148 struct lov_mds_md_v3 *lmmv3;
149 struct lov_oinfo *loi;
151 struct lov_ost_data_v1 *lmm_objects;
152 int lmm_size, lmm_magic;
157 lmm_magic = lsm->lsm_magic;
160 stripe_count = lov_get_stripecnt(lov,
161 lsm->lsm_stripe_count);
162 lsm->lsm_stripe_count = stripe_count;
164 stripe_count = lsm->lsm_stripe_count;
167 /* No needs to allocated more than LOV_MAX_STRIPE_COUNT.
168 * Anyway, this is pretty inaccurate since ld_tgt_count now
169 * represents max index and we should rely on the actual number
171 stripe_count = min((__u32)LOV_MAX_STRIPE_COUNT,
172 lov->desc.ld_tgt_count);
175 lmm_magic = le32_to_cpu((*lmmp)->lmm_magic);
177 /* lsm == NULL and lmmp == NULL */
178 lmm_magic = LOV_MAGIC;
181 if ((lmm_magic != LOV_MAGIC_V1) &&
182 (lmm_magic != LOV_MAGIC_V3)) {
183 CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
184 lmm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
188 /* XXX LOV STACKING call into osc for sizes */
189 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
195 stripe_count = le32_to_cpu((*lmmp)->lmm_stripe_count);
196 lmm_size = lov_mds_md_size(stripe_count, le32_to_cpu((*lmmp)->lmm_magic));
197 OBD_FREE(*lmmp, lmm_size);
203 OBD_ALLOC(*lmmp, lmm_size);
208 CDEBUG(D_INFO, "lov_packmd: LOV_MAGIC 0x%08X, lmm_size = %d \n",
209 lmm_magic, lmm_size);
212 lmmv3 = (struct lov_mds_md_v3 *)*lmmp;
213 if (lmm_magic == LOV_MAGIC_V3)
214 lmmv3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3);
216 lmmv1->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
221 /* lmmv1 and lmmv3 point to the same struct and have the
224 lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
225 lmmv1->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
226 lmmv1->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
227 lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
228 lmmv1->lmm_stripe_count = cpu_to_le32(stripe_count);
229 if (lsm->lsm_magic == LOV_MAGIC_V3) {
230 strncpy(lmmv3->lmm_pool_name, lsm->lsm_pool_name, LOV_MAXPOOLNAME);
231 lmm_objects = lmmv3->lmm_objects;
233 lmm_objects = lmmv1->lmm_objects;
236 for (i = 0; i < stripe_count; i++) {
237 loi = lsm->lsm_oinfo[i];
238 /* XXX LOV STACKING call down to osc_packmd() to do packing */
239 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
240 lmmv1->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
241 lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
242 lmm_objects[i].l_object_gr = cpu_to_le64(loi->loi_gr);
243 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
244 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
250 /* Find the max stripecount we should use */
251 int lov_get_stripecnt(struct lov_obd *lov, __u32 stripe_count)
254 stripe_count = lov->desc.ld_default_stripe_count;
255 if (stripe_count > lov->desc.ld_active_tgt_count)
256 stripe_count = lov->desc.ld_active_tgt_count;
259 /* for now, we limit the stripe count directly, when bug 4424 is
260 * fixed this needs to be somewhat dynamic based on whether ext3
261 * can handle larger EA sizes. */
262 if (stripe_count > LOV_MAX_STRIPE_COUNT)
263 stripe_count = LOV_MAX_STRIPE_COUNT;
269 static int lov_verify_lmm(void *lmm, int lmm_bytes, int *stripe_count)
273 if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
277 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
278 le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
279 sz = lmm_bytes * 2 + 1;
280 OBD_ALLOC(buffer, sz);
281 if (buffer != NULL) {
284 for (i = 0; i < lmm_bytes; i++)
285 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
287 CERROR("%s\n", buffer);
288 OBD_FREE(buffer, sz);
292 rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
293 lmm_bytes, stripe_count);
297 int lov_alloc_memmd(struct lov_stripe_md **lsmp, int stripe_count,
298 int pattern, int magic)
303 CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
305 *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
307 CERROR("can't allocate lsmp, stripe_count %d\n", stripe_count);
311 spin_lock_init(&(*lsmp)->lsm_lock);
312 (*lsmp)->lsm_magic = magic;
313 (*lsmp)->lsm_stripe_count = stripe_count;
314 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
315 (*lsmp)->lsm_pattern = pattern;
316 (*lsmp)->lsm_pool_name[0] = '\0';
317 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
319 for (i = 0; i < stripe_count; i++)
320 loi_init((*lsmp)->lsm_oinfo[i]);
325 void lov_free_memmd(struct lov_stripe_md **lsmp)
327 struct lov_stripe_md *lsm = *lsmp;
329 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
330 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
336 /* Unpack LOV object metadata from disk storage. It is packed in LE byte
337 * order and is opaque to the networking layer.
339 int lov_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
340 struct lov_mds_md *lmm, int lmm_bytes)
342 struct obd_device *obd = class_exp2obd(exp);
343 struct lov_obd *lov = &obd->u.lov;
344 int rc = 0, stripe_count, lsm_size;
348 /* If passed an MDS struct use values from there, otherwise defaults */
350 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
353 magic = le32_to_cpu(lmm->lmm_magic);
355 stripe_count = lov_get_stripecnt(lov, 0);
359 /* If we aren't passed an lsmp struct, we just want the size */
361 /* XXX LOV STACKING call into osc for sizes */
363 RETURN(lov_stripe_md_size(stripe_count));
366 /* If we are passed an allocated struct but nothing to unpack, free */
368 lov_free_memmd(lsmp);
372 lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
377 /* If we are passed a pointer but nothing to unpack, we only alloc */
381 LASSERT(lsm_op_find(magic) != NULL);
382 rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
384 lov_free_memmd(lsmp);
391 /* Configure object striping information on a new file.
393 * @lmmu is a pointer to a user struct with one or more of the fields set to
394 * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
395 * lmm_stripe_offset, and lmm_stripe_pattern. lmm_magic must be LOV_MAGIC.
396 * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
398 int lov_setstripe(struct obd_export *exp, struct lov_stripe_md **lsmp,
399 struct lov_user_md *lump)
401 struct obd_device *obd = class_exp2obd(exp);
402 struct lov_obd *lov = &obd->u.lov;
403 struct lov_user_md_v3 lumv3;
404 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
409 rc = copy_from_user(&lumv3, lump, sizeof(struct lov_user_md_v1));
413 switch (lumv1->lmm_magic) {
414 case LOV_USER_MAGIC_V3_SWABBED:
415 rc = copy_from_user(&lumv3, lump, sizeof(lumv3));
418 /* fall through to swab */
419 case LOV_USER_MAGIC_V1_SWABBED:
420 rc = lustre_swab_lov_user_md(lumv1);
422 case LOV_USER_MAGIC_V3:
423 rc = copy_from_user(&lumv3, lump, sizeof(lumv3));
425 case LOV_USER_MAGIC_V1:
428 CERROR("bad lsm magic %08x\n", lumv1->lmm_magic);
434 /* in the rest of the tests, as *lumv1 and lumv3 have the same
435 * fields, we use lumv1 to avoid code duplication */
437 if (lumv1->lmm_pattern == 0) {
438 lumv1->lmm_pattern = lov->desc.ld_pattern ?
439 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
442 if (lumv1->lmm_pattern != LOV_PATTERN_RAID0) {
443 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
448 /* 64kB is the largest common page size we see (ia64), and matches the
450 if (lumv1->lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
451 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
452 lumv1->lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
453 lumv1->lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
456 if ((lumv1->lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
457 (lumv1->lmm_stripe_offset !=
458 (typeof(lumv1->lmm_stripe_offset))(-1))) {
459 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
460 lumv1->lmm_stripe_offset, lov->desc.ld_tgt_count);
464 stripe_count = lov_get_stripecnt(lov, lumv1->lmm_stripe_count);
466 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3) {
467 struct pool_desc *pool;
469 pool = lov_find_pool(lov, lumv3.lmm_pool_name);
471 if (lumv1->lmm_stripe_offset !=
472 (typeof(lumv1->lmm_stripe_offset))(-1)) {
473 rc = lov_check_index_in_pool(
474 lumv1->lmm_stripe_offset, pool);
476 lov_pool_putref(pool);
481 if (stripe_count > pool_tgt_count(pool))
482 stripe_count = pool_tgt_count(pool);
484 lov_pool_putref(pool);
488 rc = lov_alloc_memmd(lsmp, stripe_count, lumv1->lmm_pattern,
494 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lumv1->lmm_stripe_offset;
495 (*lsmp)->lsm_stripe_size = lumv1->lmm_stripe_size;
497 if (lumv1->lmm_magic == LOV_USER_MAGIC_V3)
498 strncpy((*lsmp)->lsm_pool_name, lumv3.lmm_pool_name,
504 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
505 struct lov_user_md *lump)
509 struct obd_export *oexp;
510 struct lov_obd *lov = &exp->exp_obd->u.lov;
512 struct lov_user_ost_data_v1 *lmm_objects;
516 if (lump->lmm_magic == LOV_USER_MAGIC_V3)
517 lmm_objects = ((struct lov_user_md_v3 *)lump)->lmm_objects;
519 lmm_objects = lump->lmm_objects;
521 for (i = 0; i < lump->lmm_stripe_count; i++) {
522 __u32 len = sizeof(last_id);
523 oexp = lov->lov_tgts[lmm_objects[i].l_ost_idx]->ltd_exp;
524 rc = obd_get_info(oexp, sizeof(KEY_LAST_ID), KEY_LAST_ID,
525 &len, &last_id, NULL);
528 if (lmm_objects[i].l_object_id > last_id) {
529 CERROR("Setting EA for object > than last id on "
530 "ost idx %d "LPD64" > "LPD64" \n",
531 lmm_objects[i].l_ost_idx,
532 lmm_objects[i].l_object_id, last_id);
537 rc = lov_setstripe(exp, lsmp, lump);
541 for (i = 0; i < lump->lmm_stripe_count; i++) {
542 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
543 lmm_objects[i].l_ost_idx;
544 (*lsmp)->lsm_oinfo[i]->loi_id = lmm_objects[i].l_object_id;
545 (*lsmp)->lsm_oinfo[i]->loi_gr = lmm_objects[i].l_object_gr;
551 /* Retrieve object striping information.
553 * @lump is a pointer to an in-core struct with lmm_ost_count indicating
554 * the maximum number of OST indices which will fit in the user buffer.
555 * lmm_magic must be LOV_USER_MAGIC.
557 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
558 struct lov_user_md *lump)
560 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
561 struct lov_user_md_v3 lum;
562 struct lov_mds_md *lmmk = NULL;
569 /* we only need the header part from user space to get lmm_magic and
570 * lmm_stripe_count, (the header part is common to v1 and v3) */
571 lum_size = sizeof(struct lov_user_md_v1);
572 rc = copy_from_user(&lum, lump, lum_size);
575 /* if v3 we just have to update the lum_size */
576 if (lum.lmm_magic == LOV_USER_MAGIC_V3)
577 lum_size = sizeof(struct lov_user_md_v3);
579 if ((lum.lmm_magic != LOV_USER_MAGIC_V1) &&
580 (lum.lmm_magic != LOV_USER_MAGIC_V3))
583 if (lum.lmm_stripe_count &&
584 (lum.lmm_stripe_count < lsm->lsm_stripe_count)) {
585 /* Return right size of stripe to user */
586 lum.lmm_stripe_count = lsm->lsm_stripe_count;
587 rc = copy_to_user(lump, &lum, lum_size);
590 rc = lov_packmd(exp, &lmmk, lsm);
596 /* FIXME: Bug 1185 - copy fields properly when structs change */
597 /* struct lov_user_md_v3 and struct lov_mds_md_v3 must be the same */
598 LASSERT(sizeof(lum) == sizeof(struct lov_mds_md_v3));
599 LASSERT(sizeof(lum.lmm_objects[0]) == sizeof(lmmk->lmm_objects[0]));
601 if ((cpu_to_le32(LOV_MAGIC) != LOV_MAGIC) &&
602 ((lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V1)) ||
603 (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)))) {
604 lustre_swab_lov_user_md_objects((struct lov_user_md*)lmmk);
605 lustre_swab_lov_mds_md(lmmk);
607 if (lum.lmm_magic == LOV_USER_MAGIC) {
608 /* User request for v1, we need skip lmm_pool_name */
609 if (lmmk->lmm_magic == LOV_MAGIC_V3) {
610 memmove(((struct lov_mds_md_v1*)lmmk)->lmm_objects,
611 ((struct lov_mds_md_v3*)lmmk)->lmm_objects,
612 lmmk->lmm_stripe_count *
613 sizeof(struct lov_ost_data_v1));
614 lmm_size -= LOV_MAXPOOLNAME;
617 /* if v3 we just have to update the lum_size */
618 lum_size = sizeof(struct lov_user_md_v3);
620 /* User wasn't expecting this many OST entries */
621 if (lum.lmm_stripe_count == 0)
623 else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count)
626 * Have a difference between lov_mds_md & lov_user_md.
627 * So we have to re-order the data before copy to user.
629 lum.lmm_stripe_count = lmmk->lmm_stripe_count;
630 ((struct lov_user_md*)lmmk)->lmm_stripe_offset = 0;
631 ((struct lov_user_md*)lmmk)->lmm_stripe_count = lum.lmm_stripe_count;
632 if (copy_to_user(lump, lmmk, lmm_size))
635 /* Restore the LE endian to ensure obd_free_diskmd works well */
636 if (cpu_to_le32(LOV_MAGIC) != LOV_MAGIC) {
637 lmmk->lmm_stripe_count = lum.lmm_stripe_count;
638 lustre_swab_lov_mds_md(lmmk);
641 obd_free_diskmd(exp, &lmmk);