* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2011, 2014, Intel Corporation.
+ * Copyright (c) 2011, 2015, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
/* OI scrub should skip this inode. */
#define LDISKFS_STATE_LUSTRE_NOSCRUB 31
+#define LDISKFS_STATE_LUSTRE_DESTROY 30
/** Enable thandle usage statistics */
#define OSD_THANDLE_STATS (0)
#define ADMIN_USR "admin_quotafile_v2.usr"
#define ADMIN_GRP "admin_quotafile_v2.grp"
+/* Statfs space reservation for fragmentation and local objects */
+#define OSD_STATFS_RESERVED (1ULL << 23) /* 8MB */
+#define OSD_STATFS_RESERVED_SHIFT (7) /* reserve 0.78% of all space */
+
struct osd_directory {
struct iam_container od_container;
struct iam_descr od_descr;
struct osd_directory *oo_dir;
/** protects inode attributes. */
spinlock_t oo_guard;
+
+ __u32 oo_destroyed:1;
+
+ /* the i_flags in LMA */
+ __u32 oo_lma_flags;
/**
* Following two members are used to indicate the presence of dot and
* dotdot in the given directory. This is required for interop mode
#ifdef CONFIG_LOCKDEP
struct lockdep_map oo_dep_map;
#endif
+
+ struct list_head oo_xattr_list;
};
struct osd_obj_seq {
struct osd_mdobj_map {
struct dentry *omm_remote_parent;
};
-
-#define osd_ldiskfs_add_entry(handle, child, cinode, hlock) \
- ldiskfs_add_entry(handle, child, cinode, hlock)
+int osd_ldiskfs_add_entry(struct osd_thread_info *info,
+ handle_t *handle, struct dentry *child,
+ struct inode *inode, struct htree_lock *hlock);
#define OSD_OTABLE_IT_CACHE_SIZE 64
#define OSD_OTABLE_IT_CACHE_MASK (~(OSD_OTABLE_IT_CACHE_SIZE - 1))
ooi_waiting:1; /* it::next is waiting. */
};
+struct osd_obj_orphan {
+ struct list_head oor_list;
+ struct lu_env *oor_env; /* to identify "own" records */
+ __u32 oor_ino;
+};
+
/*
* osd device.
*/
od_igif_inoi:1,
od_check_ff:1,
od_is_ost:1,
- od_lma_self_repair:1;
+ od_index_in_idif:1;
- unsigned long od_capa_timeout;
- __u32 od_capa_alg;
__u32 od_dirent_journal;
- struct lustre_capa_key *od_capa_keys;
- struct hlist_head *od_capa_hash;
-
+ int od_index;
struct proc_dir_entry *od_proc_entry;
struct lprocfs_stats *od_stats;
* exceeds the osd_device::od_full_scrub_threshold_rate,
* then trigger OI scrub to scan the whole device. */
__u64 od_full_scrub_threshold_rate;
+
+ /* a list of orphaned agent inodes, protected with od_osfs_lock */
+ struct list_head od_orphan_list;
};
enum osd_full_scrub_ratio {
OSD_OT_WRITE = 7,
OSD_OT_INSERT = 8,
OSD_OT_DELETE = 9,
- OSD_OT_UPDATE = 10,
- OSD_OT_QUOTA = 11,
- OSD_OT_MAX = 12
+ OSD_OT_QUOTA = 10,
+ OSD_OT_MAX = 11
};
struct osd_thandle {
struct thandle ot_super;
handle_t *ot_handle;
struct ldiskfs_journal_cb_entry ot_jcb;
- struct list_head ot_dcb_list;
+ struct list_head ot_commit_dcb_list;
+ struct list_head ot_stop_dcb_list;
/* Link to the device, for debugging. */
struct lu_ref_link ot_dev_link;
unsigned short ot_credits;
unsigned short ot_id_cnt;
unsigned short ot_id_type;
+ unsigned int ot_remove_agents:1;
uid_t ot_id_array[OSD_MAX_UGID_CNT];
struct lquota_trans *ot_quota_trans;
#if OSD_THANDLE_STATS
* there would be one ext3 readdir for every mdd readdir page.
*/
-#define OSD_IT_EA_BUFSIZE (PAGE_CACHE_SIZE + PAGE_CACHE_SIZE/4)
+#define OSD_IT_EA_BUFSIZE (PAGE_SIZE + PAGE_SIZE/4)
/**
* This is iterator's in-memory data structure in interoperability
struct list_head oiq_list;
};
-#define MAX_BLOCKS_PER_PAGE (PAGE_CACHE_SIZE / 512)
+#define MAX_BLOCKS_PER_PAGE (PAGE_SIZE / 512)
struct osd_iobuf {
wait_queue_head_t dr_wait;
struct lu_buf dr_pg_buf;
struct page **dr_pages;
struct lu_buf dr_bl_buf;
- unsigned long *dr_blocks;
+ sector_t *dr_blocks;
unsigned long dr_start_time;
unsigned long dr_elapsed; /* how long io took */
struct osd_device *dr_dev;
unsigned int dr_init_at; /* the line iobuf was initialized */
};
+#define OSD_INS_CACHE_SIZE 8
+
struct osd_thread_info {
- const struct lu_env *oti_env;
- /**
- * used for index operations.
- */
- struct dentry oti_obj_dentry;
- struct dentry oti_child_dentry;
+ const struct lu_env *oti_env;
+ /**
+ * used for index operations.
+ */
+ struct dentry oti_obj_dentry;
+ struct dentry oti_child_dentry;
- /** dentry for Iterator context. */
- struct dentry oti_it_dentry;
- struct htree_lock *oti_hlock;
+ /** dentry for Iterator context. */
+ struct dentry oti_it_dentry;
- struct lu_fid oti_fid;
+ union {
+ /* fake struct file for osd_object_sync */
+ struct file oti_file;
+ /* osd_statfs() */
+ struct kstatfs oti_ksfs;
+ };
+
+ struct htree_lock *oti_hlock;
+
+ struct lu_fid oti_fid;
struct lu_fid oti_fid2;
struct lu_fid oti_fid3;
struct osd_inode_id oti_id;
* XXX temporary: for ->i_op calls.
*/
struct timespec oti_time;
- /*
- * XXX temporary: fake struct file for osd_object_sync
- */
- struct file oti_file;
- /*
- * XXX temporary: for capa operations.
- */
- struct lustre_capa_key oti_capa_key;
- struct lustre_capa oti_capa;
/** osd_device reference, initialized in osd_trans_start() and
used in osd_trans_stop() */
void *oti_it_ea_buf;
unsigned int oti_it_ea_buf_used:1;
- struct kstatfs oti_ksfs;
-
- /** IAM iterator for index operation. */
- struct iam_iterator oti_idx_it;
+ /* IAM iterator for index operation. */
+ struct iam_iterator oti_idx_it;
/** union to guarantee that ->oti_ipd[] has proper alignment. */
union {
+ char oti_name[48];
char oti_it_ipd[DX_IPD_MAX_SIZE];
long long oti_alignment_lieutenant;
};
struct osd_idmap_cache oti_cache;
+ /* dedicated OI cache for insert (which needs inum) */
+ struct osd_idmap_cache *oti_ins_cache;
+ int oti_ins_cache_size;
+ int oti_ins_cache_used;
+
int oti_r_locks;
int oti_w_locks;
int oti_txns;
struct lustre_mdt_attrs oti_mdt_attrs;
/* old LMA for compatibility */
char oti_mdt_attrs_old[LMA_OLD_SIZE];
+ struct filter_fid_old oti_ff;
+ struct filter_fid oti_ff_new;
};
/** 0-copy IO */
struct osd_iobuf oti_iobuf;
- struct inode oti_inode;
+ /* used to access objects in /O */
+ struct inode *oti_inode;
#define OSD_FID_REC_SZ 32
char oti_ldp[OSD_FID_REC_SZ];
char oti_ldp2[OSD_FID_REC_SZ];
/* used by quota code */
union {
-#ifdef HAVE_DQUOT_FS_DISK_QUOTA
+#if defined(HAVE_DQUOT_QC_DQBLK)
+ struct qc_dqblk oti_qdq;
+#elif defined(HAVE_DQUOT_FS_DISK_QUOTA)
struct fs_disk_quota oti_fdq;
#else
struct if_dqblk oti_dqblk;
/* Tracking for transaction credits, to allow debugging and optimizing
* cases where a large number of credits are being allocated for
* single transaction. */
+ unsigned int oti_credits_before;
unsigned short oti_declare_ops[OSD_OT_MAX];
- unsigned short oti_declare_ops_rb[OSD_OT_MAX];
unsigned short oti_declare_ops_cred[OSD_OT_MAX];
- bool oti_rollback;
-
- char oti_name[48];
- union {
- struct filter_fid_old oti_ff;
- struct filter_fid oti_ff_new;
- };
+ unsigned short oti_declare_ops_used[OSD_OT_MAX];
};
extern int ldiskfs_pdo;
int osd_procfs_init(struct osd_device *osd, const char *name);
int osd_procfs_fini(struct osd_device *osd);
void osd_brw_stats_update(struct osd_device *osd, struct osd_iobuf *iobuf);
+#if LUSTRE_VERSION_CODE < OBD_OCD_VERSION(3, 0, 52, 0)
+int osd_register_proc_index_in_idif(struct osd_device *osd);
+#endif
#endif
int osd_statfs(const struct lu_env *env, struct dt_device *dev,
struct obd_statfs *sfs);
-int osd_object_auth(const struct lu_env *env, struct dt_object *dt,
- struct lustre_capa *capa, __u64 opc);
struct inode *osd_iget(struct osd_thread_info *info, struct osd_device *dev,
struct osd_inode_id *id);
int osd_ea_fid_set(struct osd_thread_info *info, struct inode *inode,
const struct dt_rec *rec,
union lquota_rec *quota_rec);
void osd_quota_unpack(struct osd_object *obj, const struct dt_rec *rec);
-int osd_quota_migration(const struct lu_env *env, struct dt_object *dt,
- const struct dt_index_features *feat);
-
-static inline bool is_quota_glb_feat(const struct dt_index_features *feat)
-{
- return (feat == &dt_quota_iusr_features ||
- feat == &dt_quota_busr_features ||
- feat == &dt_quota_igrp_features ||
- feat == &dt_quota_bgrp_features) ? true : false;
-}
+int osd_quota_migration(const struct lu_env *env, struct dt_object *dt);
#ifndef HAVE_I_UID_READ
static inline uid_t i_uid_read(const struct inode *inode)
}
#endif
+#ifdef LDISKFS_HT_MISC
+# define osd_journal_start_sb(sb, type, nblock) \
+ ldiskfs_journal_start_sb(sb, type, nblock)
+# define osd_ldiskfs_append(handle, inode, nblock) \
+ ldiskfs_append(handle, inode, nblock)
+# define osd_ldiskfs_find_entry(dir, name, de, inlined, lock) \
+ (__ldiskfs_find_entry(dir, name, de, inlined, lock) ?: \
+ ERR_PTR(-ENOENT))
+# define osd_journal_start(inode, type, nblocks) \
+ ldiskfs_journal_start(inode, type, nblocks)
+# define osd_transaction_size(dev) \
+ (osd_journal(dev)->j_max_transaction_buffers / 2)
+#else
+# define LDISKFS_HT_MISC 0
+# define osd_journal_start_sb(sb, type, nblock) \
+ ldiskfs_journal_start_sb(sb, nblock)
+
+static inline struct buffer_head *osd_ldiskfs_append(handle_t *handle,
+ struct inode *inode,
+ ldiskfs_lblk_t *nblock)
+{
+ struct buffer_head *bh;
+ int err = 0;
+
+ bh = ldiskfs_append(handle, inode, nblock, &err);
+ if (bh == NULL)
+ bh = ERR_PTR(err);
+
+ return bh;
+}
+
+# define osd_ldiskfs_find_entry(dir, name, de, inlined, lock) \
+ (__ldiskfs_find_entry(dir, name, de, lock) ?: \
+ ERR_PTR(-ENOENT))
+# define osd_journal_start(inode, type, nblocks) \
+ ldiskfs_journal_start(inode, nblocks)
+# define osd_transaction_size(dev) \
+ (osd_journal(dev)->j_max_transaction_buffers)
+#endif
+
/*
* Invariants, assertions.
*/
return osd->od_dt_dev.dd_lu_dev.ld_obd->obd_name;
}
+static inline bool osd_is_ea_inode(struct inode *inode)
+{
+ return !!(LDISKFS_I(inode)->i_flags & LDISKFS_EA_INODE_FL);
+}
+
extern const struct dt_body_operations osd_body_ops;
extern struct lu_context_key osd_key;
bag->ic_descr->id_ops->id_ipd_free(ipd);
}
+int osd_calc_bkmap_credits(struct super_block *sb, struct inode *inode,
+ const loff_t size, const loff_t pos,
+ const int blocks);
+
int osd_ldiskfs_read(struct inode *inode, void *buf, int size, loff_t *offs);
int osd_ldiskfs_write_record(struct inode *inode, void *buf, int bufsize,
int write_NUL, loff_t *offs, handle_t *handle);
extern int osd_trans_declare_op2rb[];
extern int ldiskfs_track_declares_assert;
+void osd_trans_dump_creds(const struct lu_env *env, struct thandle *th);
static inline void osd_trans_declare_op(const struct lu_env *env,
struct osd_thandle *oh,
LASSERT(op < OSD_OT_MAX);
} else {
CWARN("%s: Invalid operation index %d\n",
- osd_name(oti->oti_dev), op);
+ osd_name(osd_dt_dev(oh->ot_super.th_dev)), op);
libcfs_debug_dumpstack(NULL);
}
} else {
struct osd_thread_info *oti = osd_oti_get(env);
struct osd_thandle *oh = container_of(th, struct osd_thandle,
ot_super);
- unsigned int rb;
+ unsigned int rb, left;
LASSERT(oh->ot_handle != NULL);
if (unlikely(op >= OSD_OT_MAX)) {
LASSERT(op < OSD_OT_MAX);
else {
CWARN("%s: Invalid operation index %d\n",
- osd_name(oti->oti_dev), op);
+ osd_name(osd_dt_dev(oh->ot_super.th_dev)), op);
libcfs_debug_dumpstack(NULL);
return;
}
}
- if (likely(!oti->oti_rollback && oti->oti_declare_ops[op] > 0)) {
- oti->oti_declare_ops[op]--;
- oti->oti_declare_ops_rb[op]++;
- } else {
- /* all future updates are considered rollback */
- oti->oti_rollback = true;
- rb = osd_trans_declare_op2rb[op];
- if (unlikely(rb >= OSD_OT_MAX)) {
- if (unlikely(ldiskfs_track_declares_assert))
- LASSERTF(rb < OSD_OT_MAX, "rb = %u\n", rb);
- else {
- CWARN("%s: Invalid rollback index %d\n",
- osd_name(oti->oti_dev), rb);
- libcfs_debug_dumpstack(NULL);
- return;
- }
- }
- if (unlikely(oti->oti_declare_ops_rb[rb] == 0)) {
+ /* find rollback (or reverse) operation for the given one
+ * such an operation doesn't require additional credits
+ * as the same set of blocks are modified */
+ rb = osd_trans_declare_op2rb[op];
+
+ /* check whether credits for this operation were reserved at all */
+ if (unlikely(oti->oti_declare_ops_cred[op] == 0 &&
+ oti->oti_declare_ops_cred[rb] == 0)) {
+ /* the API is not perfect yet: CREATE does REF_ADD internally
+ * while DESTROY does not. To rollback CREATE the callers
+ * needs to call REF_DEL+DESTROY which is hard to detect using
+ * a simple table of rollback operations */
+ if (op == OSD_OT_REF_DEL &&
+ oti->oti_declare_ops_cred[OSD_OT_CREATE] > 0)
+ goto proceed;
+ if (op == OSD_OT_REF_ADD &&
+ oti->oti_declare_ops_cred[OSD_OT_DESTROY] > 0)
+ goto proceed;
+ osd_trans_dump_creds(env, th);
+ CERROR("%s: op = %d, rb = %d\n",
+ osd_name(osd_dt_dev(oh->ot_super.th_dev)), op, rb);
+ if (unlikely(ldiskfs_track_declares_assert))
+ LBUG();
+ }
+
+proceed:
+ /* remember how many credits we have unused before the operation */
+ oti->oti_credits_before = oh->ot_handle->h_buffer_credits;
+ left = oti->oti_declare_ops_cred[op] - oti->oti_declare_ops_used[op];
+ if (unlikely(oti->oti_credits_before < left)) {
+ osd_trans_dump_creds(env, th);
+ CERROR("%s: op = %d, rb = %d\n",
+ osd_name(osd_dt_dev(oh->ot_super.th_dev)), op, rb);
+ /* on a very small fs (testing?) it's possible that
+ * the transaction can't fit 1/4 of journal, so we
+ * just request less credits (see osd_trans_start()).
+ * ignore the same case here */
+ rb = osd_transaction_size(osd_dt_dev(th->th_dev));
+ if (unlikely(oh->ot_credits < rb)) {
if (unlikely(ldiskfs_track_declares_assert))
- LASSERTF(oti->oti_declare_ops_rb[rb] > 0,
- "rb = %u\n", rb);
- else {
- CWARN("%s: Overflow in tracking declares for "
- "index, rb = %d\n",
- osd_name(oti->oti_dev), rb);
- libcfs_debug_dumpstack(NULL);
- return;
- }
+ LBUG();
}
- oti->oti_declare_ops_rb[rb]--;
}
}
-static inline void osd_trans_declare_rb(const struct lu_env *env,
- struct thandle *th, unsigned int op)
+static inline void osd_trans_exec_check(const struct lu_env *env,
+ struct thandle *th,
+ unsigned int op)
{
struct osd_thread_info *oti = osd_oti_get(env);
struct osd_thandle *oh = container_of(th, struct osd_thandle,
ot_super);
+ int used, over, quota;
- LASSERT(oh->ot_handle != NULL);
- if (unlikely(op >= OSD_OT_MAX)) {
- if (unlikely(ldiskfs_track_declares_assert))
- LASSERT(op < OSD_OT_MAX);
- else {
- CWARN("%s: Invalid operation index %d\n",
- osd_name(oti->oti_dev), op);
- libcfs_debug_dumpstack(NULL);
- }
+ /* how many credits have been used by the operation */
+ used = oti->oti_credits_before - oh->ot_handle->h_buffer_credits;
+
+ if (unlikely(used < 0)) {
+ /* if some block was allocated and released in the same
+ * transaction, then it won't be a part of the transaction
+ * and delta can be negative */
+ return;
+ }
+ if (used == 0) {
+ /* rollback operations (e.g. when we destroy just created
+ * object) should not consume any credits. there is no point
+ * to confuse the checks below */
+ return;
+ }
+
+ oti->oti_declare_ops_used[op] += used;
+ if (oti->oti_declare_ops_used[op] <= oti->oti_declare_ops_cred[op])
+ return;
+
+ /* we account quota for a whole transaction and any operation can
+ * consume corresponding credits */
+ over = oti->oti_declare_ops_used[op] -
+ oti->oti_declare_ops_cred[op];
+ quota = oti->oti_declare_ops_cred[OSD_OT_QUOTA] -
+ oti->oti_declare_ops_used[OSD_OT_QUOTA];
+ if (over <= quota) {
+ /* probably that credits were consumed by
+ * quota indirectly (in the depths of ldiskfs) */
+ oti->oti_declare_ops_used[OSD_OT_QUOTA] += over;
+ oti->oti_declare_ops_used[op] -= over;
} else {
- oti->oti_declare_ops_rb[op]++;
+ CWARN("op %d: used %u, used now %u, reserved %u\n",
+ op, oti->oti_declare_ops_used[op], used,
+ oti->oti_declare_ops_cred[op]);
+ osd_trans_dump_creds(env, th);
+ if (unlikely(ldiskfs_track_declares_assert))
+ LBUG();
}
}
static inline int is_32bit_api(void)
{
#ifdef CONFIG_COMPAT
- return is_compat_task();
+ return in_compat_syscall();
#else
return (BITS_PER_LONG == 32);
#endif
return dev->od_mdt_map->omm_remote_parent->d_inode->i_ino;
}
-#ifdef JOURNAL_START_HAS_3ARGS
-# define osd_journal_start_sb(sb, type, nblock) \
- ldiskfs_journal_start_sb(sb, type, nblock)
-# define osd_ldiskfs_append(handle, inode, nblock, err) \
- ldiskfs_append(handle, inode, nblock)
-# define osd_ldiskfs_find_entry(dir, name, de, inlined, lock) \
- ldiskfs_find_entry(dir, name, de, inlined, lock)
-# define osd_journal_start(inode, type, nblocks) \
- ldiskfs_journal_start(inode, type, nblocks)
-# define osd_transaction_size(dev) \
- (osd_journal(dev)->j_max_transaction_buffers / 2)
+/**
+ * ext4_bread/ldiskfs_bread has either 5 or 4 parameters. The error
+ * return code has been removed and integrated into the pointer in the
+ * kernel 3.18.
+ */
+static inline struct buffer_head *__ldiskfs_bread(handle_t *handle,
+ struct inode *inode,
+ ldiskfs_lblk_t block,
+ int create)
+{
+#ifdef HAVE_EXT4_BREAD_4ARGS
+ return ldiskfs_bread(handle, inode, block, create);
#else
-# define LDISKFS_HT_MISC 0
-# define osd_journal_start_sb(sb, type, nblock) \
- ldiskfs_journal_start_sb(sb, nblock)
-# define osd_ldiskfs_append(handle, inode, nblock, err) \
- ldiskfs_append(handle, inode, nblock, err)
-# define osd_ldiskfs_find_entry(dir, name, de, inlined, lock) \
- ldiskfs_find_entry(dir, name, de, lock)
-# define osd_journal_start(inode, type, nblocks) \
- ldiskfs_journal_start(inode, nblocks)
-# define osd_transaction_size(dev) \
- (osd_journal(dev)->j_max_transaction_buffers)
+ struct buffer_head *bh;
+ int error = 0;
+
+ bh = ldiskfs_bread(handle, inode, block, create, &error);
+ if (bh == NULL && error != 0)
+ bh = ERR_PTR(error);
+
+ return bh;
#endif
+}
void ldiskfs_inc_count(handle_t *handle, struct inode *inode);
void ldiskfs_dec_count(handle_t *handle, struct inode *inode);