Whamcloud - gitweb
LU-3161 lod: Added proper error handling instead of LASSERT
[fs/lustre-release.git] / lustre / lod / lod_lov.c
index 37a5809..5d74c88 100644 (file)
@@ -23,7 +23,7 @@
  * Copyright  2009 Sun Microsystems, Inc. All rights reserved
  * Use is subject to license terms.
  *
- * Copyright (c) 2012, Intel Corporation.
+ * Copyright (c) 2012, 2013, Intel Corporation.
  */
 /*
  * lustre/lod/lod_lov.c
@@ -31,9 +31,6 @@
  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com> 
  */
 
-#ifndef EXPORT_SYMTAB
-# define EXPORT_SYMTAB
-#endif
 #define DEBUG_SUBSYSTEM S_MDS
 
 #include <obd_class.h>
@@ -243,15 +240,14 @@ int lod_add_device(const struct lu_env *env, struct lod_device *lod,
                                           OBD_CONNECT_MDS_MDS |
                                           OBD_CONNECT_FID |
                                           OBD_CONNECT_AT |
-                                          OBD_CONNECT_FULL20 |
-                                          OBD_CONNECT_PINGLESS;
+                                          OBD_CONNECT_FULL20;
                /* XXX set MDS-MDS flags, remove this when running this
                 * on client*/
                data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS;
                spin_lock(&imp->imp_lock);
                imp->imp_server_timeout = 1;
                spin_unlock(&imp->imp_lock);
-               imp->imp_client->cli_request_portal = MDS_MDS_PORTAL;
+               imp->imp_client->cli_request_portal = OUT_PORTAL;
                CDEBUG(D_OTHER, "%s: Set 'mds' portal and timeout\n",
                      obd->obd_name);
                ltd = &lod->lod_mdt_descs;
@@ -266,7 +262,7 @@ int lod_add_device(const struct lu_env *env, struct lod_device *lod,
        }
 
        LASSERT(obd->obd_lu_dev);
-       LASSERT(obd->obd_lu_dev->ld_site = lod->lod_dt_dev.dd_lu_dev.ld_site);
+       LASSERT(obd->obd_lu_dev->ld_site == lod->lod_dt_dev.dd_lu_dev.ld_site);
 
        ldev = obd->obd_lu_dev;
        d = lu2dt_dev(ldev);
@@ -505,9 +501,8 @@ int lod_generate_and_set_lovea(const struct lu_env *env,
        ENTRY;
 
        LASSERT(lo);
-       LASSERT(lo->ldo_stripenr > 0);
 
-       magic = lo->ldo_pool ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
+       magic = lo->ldo_pool != NULL ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
        lmm_size = lov_mds_md_size(lo->ldo_stripenr, magic);
        if (info->lti_ea_store_size < lmm_size) {
                rc = lod_ea_store_resize(info, lmm_size);
@@ -515,14 +510,19 @@ int lod_generate_and_set_lovea(const struct lu_env *env,
                        RETURN(rc);
        }
 
+       if (lo->ldo_pattern == 0) /* default striping */
+               lo->ldo_pattern = LOV_PATTERN_RAID0;
+
        lmm = info->lti_ea_store;
 
        lmm->lmm_magic = cpu_to_le32(magic);
-       lmm->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
-       fid_ostid_pack(fid, &lmm->lmm_oi);
-       ostid_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
+       lmm->lmm_pattern = cpu_to_le32(lo->ldo_pattern);
+       fid_to_lmm_oi(fid, &lmm->lmm_oi);
+       lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
        lmm->lmm_stripe_size = cpu_to_le32(lo->ldo_stripe_size);
        lmm->lmm_stripe_count = cpu_to_le16(lo->ldo_stripenr);
+       if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED)
+               lmm->lmm_stripe_count = cpu_to_le16(lo->ldo_released_stripenr);
        lmm->lmm_layout_gen = 0;
        if (magic == LOV_MAGIC_V1) {
                objs = &lmm->lmm_objects[0];
@@ -544,7 +544,7 @@ int lod_generate_and_set_lovea(const struct lu_env *env,
                LASSERT(lo->ldo_stripe[i]);
                fid = lu_object_fid(&lo->ldo_stripe[i]->do_lu);
 
-               rc = fid_ostid_pack(fid, &info->lti_ostid);
+               rc = fid_to_ostid(fid, &info->lti_ostid);
                LASSERT(rc == 0);
 
                ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
@@ -663,6 +663,30 @@ int lod_store_def_striping(const struct lu_env *env, struct dt_object *dt,
        RETURN(rc);
 }
 
+static int validate_lod_and_idx(struct lod_device *md, int idx)
+{
+       if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
+                    !cfs_bitmap_check(md->lod_ost_bitmap, idx))) {
+               CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
+                      md->lod_ost_descs.ltd_tgts_size);
+               return -EINVAL;
+       }
+
+       if (unlikely(OST_TGT(md, idx) == NULL)) {
+               CERROR("%s: bad lod_tgt_desc for idx: %d\n",
+                      lod2obd(md)->obd_name, idx);
+               return -EINVAL;
+       }
+
+       if (unlikely(OST_TGT(md, idx)->ltd_ost == NULL)) {
+               CERROR("%s: invalid lod device, for idx: %d\n",
+                      lod2obd(md)->obd_name , idx);
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
 /*
  * allocate array of objects pointers, find/create objects
  * stripenr and other fields should be initialized by this moment
@@ -679,7 +703,7 @@ int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
        int                      i, idx, rc = 0;
        ENTRY;
 
-       LASSERT(lo);
+       LASSERT(lo != NULL);
        LASSERT(lo->ldo_stripe == NULL);
        LASSERT(lo->ldo_stripenr > 0);
        LASSERT(lo->ldo_stripe_size > 0);
@@ -692,19 +716,19 @@ int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
        for (i = 0; i < lo->ldo_stripenr; i++) {
                ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
                idx = le64_to_cpu(objs[i].l_ost_idx);
-               fid_ostid_unpack(&info->lti_fid, &info->lti_ostid, idx);
+               rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
+               if (rc != 0)
+                       GOTO(out, rc);
                LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
                         PFID(&info->lti_fid));
-               /*
-                * XXX: assertion is left for testing, to make
-                * sure we never process requests till configuration
-                * is completed. to be changed to -EINVAL
-                */
-
                lod_getref(&md->lod_ost_descs);
-               LASSERT(cfs_bitmap_check(md->lod_ost_bitmap, idx));
-               LASSERT(OST_TGT(md,idx));
-               LASSERTF(OST_TGT(md,idx)->ltd_ost, "idx %d\n", idx);
+
+               rc = validate_lod_and_idx(md, idx);
+               if (unlikely(rc != 0)) {
+                       lod_putref(md, &md->lod_ost_descs);
+                       GOTO(out, rc);
+               }
+
                nd = &OST_TGT(md,idx)->ltd_ost->dd_lu_dev;
                lod_putref(md, &md->lod_ost_descs);
 
@@ -745,6 +769,7 @@ int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
        struct lov_mds_md_v1    *lmm;
        struct lov_ost_data_v1  *objs;
        __u32                    magic;
+       __u32                    pattern;
        int                      rc = 0;
        ENTRY;
 
@@ -754,15 +779,20 @@ int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
 
        lmm = (struct lov_mds_md_v1 *) buf->lb_buf;
        magic = le32_to_cpu(lmm->lmm_magic);
+       pattern = le32_to_cpu(lmm->lmm_pattern);
 
        if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)
                GOTO(out, rc = -EINVAL);
-       if (le32_to_cpu(lmm->lmm_pattern) != LOV_PATTERN_RAID0)
+       if (lov_pattern(pattern) != LOV_PATTERN_RAID0)
                GOTO(out, rc = -EINVAL);
 
+       lo->ldo_pattern = pattern;
        lo->ldo_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
-       lo->ldo_stripenr = le16_to_cpu(lmm->lmm_stripe_count);
        lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
+       lo->ldo_stripenr = le16_to_cpu(lmm->lmm_stripe_count);
+       /* released file stripenr fixup. */
+       if (pattern & LOV_PATTERN_F_RELEASED)
+               lo->ldo_stripenr = 0;
 
        LASSERT(buf->lb_len >= lov_mds_md_size(lo->ldo_stripenr, magic));
 
@@ -774,7 +804,8 @@ int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
                objs = &lmm->lmm_objects[0];
        }
 
-       rc = lod_initialize_objects(env, lo, objs);
+       if (lo->ldo_stripenr > 0)
+               rc = lod_initialize_objects(env, lo, objs);
 
 out:
        RETURN(rc);
@@ -797,13 +828,8 @@ int lod_load_striping(const struct lu_env *env, struct lod_object *lo)
         */
        dt_write_lock(env, next, 0);
        /* already initialized? */
-       if (lo->ldo_stripe) {
-               int i;
-               /* check validity */
-               for (i = 0; i < lo->ldo_stripenr; i++)
-                       LASSERTF(lo->ldo_stripe[i], "stripe %d is NULL\n", i);
+       if (lo->ldo_stripe != NULL)
                GOTO(out, rc = 0);
-       }
 
        if (!dt_object_exists(next))
                GOTO(out, rc = 0);
@@ -812,8 +838,6 @@ int lod_load_striping(const struct lu_env *env, struct lod_object *lo)
        if (!(lu_object_attr(lod2lu_obj(lo)) & S_IFREG))
                GOTO(out, rc = 0);
 
-       LASSERT(lo->ldo_stripenr == 0);
-
        rc = lod_get_lov_ea(env, lo);
        if (rc <= 0)
                GOTO(out, rc);
@@ -834,94 +858,130 @@ int lod_verify_striping(struct lod_device *d, const struct lu_buf *buf,
                        int specific)
 {
        struct lov_user_md_v1   *lum;
-       struct lov_user_md_v3   *v3 = NULL;
+       struct lov_user_md_v3   *lum3;
        struct pool_desc        *pool = NULL;
-       int                      rc;
+       __u32                    magic;
+       __u32                    stripe_size;
+       __u16                    stripe_count;
+       __u16                    stripe_offset;
+       size_t                   lum_size;
+       int                      rc = 0;
        ENTRY;
 
        lum = buf->lb_buf;
 
-       if (lum->lmm_magic != LOV_USER_MAGIC_V1 &&
-           lum->lmm_magic != LOV_USER_MAGIC_V3 &&
-           lum->lmm_magic != LOV_MAGIC_V1_DEF &&
-           lum->lmm_magic != LOV_MAGIC_V3_DEF) {
-               CDEBUG(D_IOCTL, "bad userland LOV MAGIC: %#x\n",
-                      lum->lmm_magic);
-               RETURN(-EINVAL);
+       LASSERT(sizeof(*lum) < sizeof(*lum3));
+
+       if (buf->lb_len < sizeof(*lum)) {
+               CDEBUG(D_IOCTL, "buf len %zd too small for lov_user_md\n",
+                      buf->lb_len);
+               GOTO(out, rc = -EINVAL);
+       }
+
+       magic = le32_to_cpu(lum->lmm_magic);
+       if (magic != LOV_USER_MAGIC_V1 &&
+           magic != LOV_USER_MAGIC_V3 &&
+           magic != LOV_MAGIC_V1_DEF &&
+           magic != LOV_MAGIC_V3_DEF) {
+               CDEBUG(D_IOCTL, "bad userland LOV MAGIC: %#x\n", magic);
+               GOTO(out, rc = -EINVAL);
        }
 
-       if ((specific && lum->lmm_pattern != LOV_PATTERN_RAID0) ||
-           (specific == 0 && lum->lmm_pattern != 0)) {
+       if ((specific && le32_to_cpu(lum->lmm_pattern) != LOV_PATTERN_RAID0) ||
+           (!specific && lum->lmm_pattern != 0)) {
                CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
-                      lum->lmm_pattern);
-               RETURN(-EINVAL);
+                      le32_to_cpu(lum->lmm_pattern));
+               GOTO(out, rc = -EINVAL);
        }
 
        /* 64kB is the largest common page size we see (ia64), and matches the
         * check in lfs */
-       if (lum->lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
-               CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
-                      lum->lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
-               RETURN(-EINVAL);
+       stripe_size = le32_to_cpu(lum->lmm_stripe_size);
+       if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
+               CDEBUG(D_IOCTL, "stripe size %u not a multiple of %u\n",
+                      stripe_size, LOV_MIN_STRIPE_SIZE);
+               GOTO(out, rc = -EINVAL);
        }
 
        /* an offset of -1 is treated as a "special" valid offset */
-       if (lum->lmm_stripe_offset != (typeof(lum->lmm_stripe_offset))(-1)) {
+       stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
+       if (stripe_offset != (typeof(stripe_offset))-1) {
                /* if offset is not within valid range [0, osts_size) */
-               if (lum->lmm_stripe_offset >= d->lod_osts_size) {
+               if (stripe_offset >= d->lod_osts_size) {
                        CDEBUG(D_IOCTL, "stripe offset %u >= bitmap size %u\n",
-                              lum->lmm_stripe_offset, d->lod_osts_size);
-                       RETURN(-EINVAL);
+                              stripe_offset, d->lod_osts_size);
+                       GOTO(out, rc = -EINVAL);
                }
 
                /* if lmm_stripe_offset is *not* in bitmap */
-               if (!cfs_bitmap_check(d->lod_ost_bitmap,
-                                     lum->lmm_stripe_offset)) {
+               if (!cfs_bitmap_check(d->lod_ost_bitmap, stripe_offset)) {
                        CDEBUG(D_IOCTL, "stripe offset %u not in bitmap\n",
-                              lum->lmm_stripe_offset);
-                       RETURN(-EINVAL);
+                              stripe_offset);
+                       GOTO(out, rc = -EINVAL);
                }
        }
 
-       if (lum->lmm_magic == LOV_USER_MAGIC_V3)
-               v3 = buf->lb_buf;
+       stripe_count = le16_to_cpu(lum->lmm_stripe_count);
+       if (magic == LOV_USER_MAGIC_V1 || magic == LOV_MAGIC_V1_DEF)
+               lum_size = offsetof(struct lov_user_md_v1,
+                                   lmm_objects[stripe_count]);
+       else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_MAGIC_V3_DEF)
+               lum_size = offsetof(struct lov_user_md_v3,
+                                   lmm_objects[stripe_count]);
+       else
+               LBUG();
+
+       if (specific && buf->lb_len != lum_size) {
+               CDEBUG(D_IOCTL, "invalid buf len %zd for lov_user_md with "
+                      "magic %#x and stripe_count %u\n",
+                      buf->lb_len, magic, stripe_count);
+               GOTO(out, rc = -EINVAL);
+       }
 
-       if (v3)
-               /* In the function below, .hs_keycmp resolves to
-                * pool_hashkey_keycmp() */
-               /* coverity[overrun-buffer-val] */
-               pool = lod_find_pool(d, v3->lmm_pool_name);
+       if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_MAGIC_V3_DEF))
+               goto out;
 
-       if (pool != NULL) {
-               __u16 offs = v3->lmm_stripe_offset;
+       lum3 = buf->lb_buf;
+       if (buf->lb_len < sizeof(*lum3)) {
+               CDEBUG(D_IOCTL, "buf len %zd too small for lov_user_md_v3\n",
+                      buf->lb_len);
+               GOTO(out, rc = -EINVAL);
+       }
 
-               if (offs != (typeof(v3->lmm_stripe_offset))(-1)) {
-                       rc = lod_check_index_in_pool(offs, pool);
-                       if (rc < 0) {
-                               lod_pool_putref(pool);
-                               RETURN(-EINVAL);
-                       }
-               }
+       /* In the function below, .hs_keycmp resolves to
+        * pool_hashkey_keycmp() */
+       /* coverity[overrun-buffer-val] */
+       pool = lod_find_pool(d, lum3->lmm_pool_name);
+       if (pool == NULL)
+               goto out;
+
+       if (stripe_offset != (typeof(stripe_offset))-1) {
+               rc = lod_check_index_in_pool(stripe_offset, pool);
+               if (rc < 0)
+                       GOTO(out, rc = -EINVAL);
+       }
 
-               if (specific && lum->lmm_stripe_count > pool_tgt_count(pool)) {
-                       CDEBUG(D_IOCTL,
-                              "stripe count %u > # OSTs %u in the pool\n",
-                              lum->lmm_stripe_count, pool_tgt_count(pool));
-                       lod_pool_putref(pool);
-                       RETURN(-EINVAL);
-               }
+       if (specific && stripe_count > pool_tgt_count(pool)) {
+               CDEBUG(D_IOCTL,
+                      "stripe count %u > # OSTs %u in the pool\n",
+                      stripe_count, pool_tgt_count(pool));
+               GOTO(out, rc = -EINVAL);
+       }
 
+out:
+       if (pool != NULL)
                lod_pool_putref(pool);
-       }
 
-       RETURN(0);
+       RETURN(rc);
 }
 
 void lod_fix_desc_stripe_size(__u64 *val)
 {
-       if (*val < LOV_DEFAULT_STRIPE_SIZE) {
-               LCONSOLE_WARN("Increasing default stripe size to min %u\n",
-                             LOV_DEFAULT_STRIPE_SIZE);
+       if (*val < LOV_MIN_STRIPE_SIZE) {
+               if (*val != 0)
+                       LCONSOLE_INFO("Increasing default stripe size to "
+                                     "minimum value %u\n",
+                                     LOV_DEFAULT_STRIPE_SIZE);
                *val = LOV_DEFAULT_STRIPE_SIZE;
        } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
                *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
@@ -1018,7 +1078,7 @@ int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
        OBD_ALLOC_PTR(lod->lod_qos.lq_statfs_data);
        if (NULL == lod->lod_qos.lq_statfs_data)
                RETURN(-ENOMEM);
-       cfs_waitq_init(&lod->lod_qos.lq_statfs_waitq);
+       init_waitqueue_head(&lod->lod_qos.lq_statfs_waitq);
 
        /* Set up OST pool environment */
        lod->lod_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,