DAOS-1 test: test dfuse

src/cart/README.env

@@ -184,6 +184,12 @@ This file lists the environment variables used in CaRT.
    a file with the pattern <D_CLIENT_METRICS_DUMP_DIR>/<DAOS_JOBID>-<pid>.csv. As a
    convenience, setting this variable automatically sets D_CLIENT_METRICS_ENABLE=1.
 
+ . D_CLIENT_METRICS_MAX_ENTRY_CNT
+   Set maximum metrics entry count for client side metrics, which only works if
+   D_CLIENT_METRICS_ENABLE is set to 1. The value is set to 1024 by default, and it only
+   needs to be increased if there are more concurrent daos client threads, for example
+   thread-count for dfuse is more than 20.
+
    D_POST_INIT
    (server only) Controls the initial number of requests that are posted on context creation.
    When using a transport that supports multi-recv, also controls the maximum size

src/client/api/init.c

@@ -334,6 +334,8 @@ daos_fini(void)
 		D_GOTO(unlock, rc);
 	}
 
+	daos_hhash_fini();
+
 	/** clean up all registered per-module metrics */
 	daos_metrics_fini();
 #if BUILD_PIPELINE
@@ -354,7 +356,6 @@ daos_fini(void)
 	dc_job_fini();
 
 	pl_fini();
-	daos_hhash_fini();
 	daos_debug_fini();
 	module_initialized = 0;
 unlock:

src/client/api/metrics.c

@@ -20,11 +20,15 @@
 #include <gurt/telemetry_consumer.h>
 #include <gurt/telemetry_producer.h>
 
-#define INIT_JOB_NUM 1024
 bool daos_client_metric;
 bool daos_client_metric_retain;
+uint32_t max_entry_cnt = 1024;
 
-#define MAX_IDS_SIZE(num) (num * D_TM_METRIC_SIZE)
+static inline uint32_t
+client_metrics_shm_size(uint32_t num)
+{
+	return num * D_TM_METRIC_SIZE;
+}
 /* The client side metrics structure looks like
  * root/job_id/pid/....
  */
@@ -66,8 +70,9 @@ init_root(const char *name, pid_t pid, int flags)
 	key_t key;
 	int   rc;
 
+	d_getenv_uint(DAOS_CLIENT_METRICS_MAX_ENTRY_CNT, &max_entry_cnt);
 	key = d_tm_cli_pid_key(pid);
-	rc  = d_tm_init_with_name(key, MAX_IDS_SIZE(INIT_JOB_NUM), flags, name);
+	rc  = d_tm_init_with_name(key, client_metrics_shm_size(max_entry_cnt), flags, name);
 	if (rc != 0) {
 		DL_ERROR(rc, "failed to initialize root for %s.", name);
 		return rc;

src/client/dfuse/dfuse.h

@@ -137,7 +137,6 @@ struct dfuse_inode_entry;
  * when EOF is returned to the kernel.  If it's still present on release then it's freed then.
  */
 struct dfuse_pre_read {
-	pthread_mutex_t     dra_lock;
 	struct dfuse_event *dra_ev;
 	int                 dra_rc;
 };
@@ -149,8 +148,6 @@ struct dfuse_obj_hdl {
 	/** the DFS object handle.  Not created for directories. */
 	dfs_obj_t                *doh_obj;
 
-	struct dfuse_pre_read    *doh_readahead;
-
 	/** the inode entry for the file */
 	struct dfuse_inode_entry *doh_ie;
 
@@ -176,27 +173,26 @@ struct dfuse_obj_hdl {
 	 * to false), the expected position of the next read and a boolean for if EOF has been
 	 * detected.
 	 */
-	off_t                     doh_linear_read_pos;
-	bool                      doh_linear_read;
-	bool                      doh_linear_read_eof;
-
-	/** True if caching is enabled for this file. */
-	bool                      doh_caching;
-
+	off_t			doh_linear_read_pos;
+	uint64_t		doh_linear_read:1,
+				doh_linear_read_eof:1,
+				/** True if caching is enabled for this file. */
+				doh_caching:1,
 	/* True if the file handle is writable - used for cache invalidation */
-	bool                      doh_writeable;
-
+				doh_writeable:1,
 	/* Track possible kernel cache of readdir on this directory */
 	/* Set to true if there is any reason the kernel will not use this directory handle as the
 	 * basis for a readdir cache.  Includes if seekdir or rewind are used.
 	 */
-	bool                      doh_kreaddir_invalid;
+				doh_kreaddir_invalid:1,
 	/* Set to true if readdir calls are made on this handle */
-	bool                      doh_kreaddir_started;
+				doh_kreaddir_started:1,
 	/* Set to true if readdir calls reach EOF made on this handle */
-	bool                      doh_kreaddir_finished;
+				doh_kreaddir_finished:1,
+				doh_evict_on_close:1,
+				doh_readahead_inflight:1;
 
-	bool                      doh_evict_on_close;
+	int			doh_flags;
 };
 
 /* Readdir support.
@@ -401,10 +397,18 @@ struct dfuse_event {
 	d_iov_t          de_iov;
 	d_sg_list_t      de_sgl;
 	d_list_t         de_list;
+
+	/* Position in a list of events, this will either be off ie->ie_open_reads or
+	 * de->de_read_slaves
+	 */
+	d_list_t         de_read_list;
+	/* List of slave events */
+	d_list_t         de_read_slaves;
 	struct dfuse_eq *de_eqt;
 	union {
 		struct dfuse_obj_hdl     *de_oh;
 		struct dfuse_inode_entry *de_ie;
+		struct read_chunk_data   *de_cd;
 	};
 	off_t  de_req_position; /**< The file position requested by fuse */
 	union {
@@ -474,6 +478,10 @@ struct dfuse_pool {
 	ACTION(OPEN)                                                                               \
 	ACTION(PRE_READ)                                                                           \
 	ACTION(READ)                                                                               \
+	ACTION(READ_EOF_M)                                                                         \
+	ACTION(READ_CON)                                                                           \
+	ACTION(READ_BUCKET)                                                                        \
+	ACTION(READ_BUCKET_M)                                                                      \
 	ACTION(WRITE)                                                                              \
 	ACTION(STATFS)
 
@@ -1011,6 +1019,13 @@ struct dfuse_inode_entry {
 
 	/* Entry on the evict list */
 	d_list_t                  ie_evict_entry;
+
+	/* ie read lock is used to protect ie_open_reads and ie_readahead */
+	d_list_t                  ie_open_reads;
+	struct dfuse_pre_read	  *ie_readahead;
+	pthread_mutex_t		  ie_read_lock;
+
+	struct read_chunk_core   *ie_chunk;
 };
 
 /* Flush write-back cache writes to a inode.  It does this by waiting for and then releasing an
@@ -1108,6 +1123,13 @@ dfuse_compute_inode(struct dfuse_cont *dfs,
 void
 dfuse_cache_evict_dir(struct dfuse_info *dfuse_info, struct dfuse_inode_entry *ie);
 
+/* Free any read chunk data for an inode.
+ *
+ * Returns true if feature was used.
+ */
+bool
+read_chunk_close(struct dfuse_inode_entry *ie);
+
 /* Metadata caching functions. */
 
 /* Mark the cache as up-to-date from now */
@@ -1170,8 +1192,13 @@ dfuse_cache_evict(struct dfuse_inode_entry *ie);
 bool
 dfuse_dcache_get_valid(struct dfuse_inode_entry *ie, double max_age);
 
+int
+dfuse_pre_read_init(struct dfuse_info *info, struct dfuse_inode_entry *ie,
+		    struct dfuse_event **evp);
+
 void
-dfuse_pre_read(struct dfuse_info *dfuse_info, struct dfuse_obj_hdl *oh);
+dfuse_pre_read(struct dfuse_info *dfuse_info, struct dfuse_obj_hdl *oh,
+	       struct dfuse_event *ev);
 
 int
 check_for_uns_ep(struct dfuse_info *dfuse_info, struct dfuse_inode_entry *ie, char *attr,
@@ -1183,6 +1210,7 @@ dfuse_ie_init(struct dfuse_info *dfuse_info, struct dfuse_inode_entry *ie);
 #define dfuse_ie_free(_di, _ie)                                                                    \
 	do {                                                                                       \
 		atomic_fetch_sub_relaxed(&(_di)->di_inode_count, 1);                               \
+		D_MUTEX_DESTROY(&_ie->ie_read_lock);					  	   \
 		D_FREE(_ie);                                                                       \
 	} while (0)
 

src/client/dfuse/dfuse_core.c

@@ -1044,7 +1044,6 @@ dfuse_mcache_get_valid(struct dfuse_inode_entry *ie, double max_age, double *tim
 		if (timeout)
 			*timeout = time_left;
 	}
-
 	return use;
 }
 
@@ -1237,7 +1236,7 @@ dfuse_open_handle_init(struct dfuse_info *dfuse_info, struct dfuse_obj_hdl *oh,
 {
 	oh->doh_dfs             = ie->ie_dfs->dfs_ns;
 	oh->doh_ie              = ie;
-	oh->doh_linear_read     = true;
+	oh->doh_linear_read     = 1;
 	oh->doh_linear_read_pos = 0;
 	atomic_init(&oh->doh_il_calls, 0);
 	atomic_init(&oh->doh_write_count, 0);
@@ -1254,7 +1253,9 @@ dfuse_ie_init(struct dfuse_info *dfuse_info, struct dfuse_inode_entry *ie)
 	atomic_init(&ie->ie_linear_read, true);
 	atomic_fetch_add_relaxed(&dfuse_info->di_inode_count, 1);
 	D_INIT_LIST_HEAD(&ie->ie_evict_entry);
+	D_INIT_LIST_HEAD(&ie->ie_open_reads);
 	D_RWLOCK_INIT(&ie->ie_wlock, 0);
+	D_MUTEX_INIT(&ie->ie_read_lock, 0);
 }
 
 void
@@ -1317,6 +1318,9 @@ dfuse_read_event_size(void *arg, size_t size)
 		ev->de_sgl.sg_nr       = 1;
 	}
 
+	/* D_INIT_LIST_HEAD(&ev->de_read_list); */
+	D_INIT_LIST_HEAD(&ev->de_read_slaves);
+
 	rc = daos_event_init(&ev->de_ev, ev->de_eqt->de_eq, NULL);
 	if (rc != -DER_SUCCESS) {
 		return false;

src/client/dfuse/dfuse_fuseops.c

@@ -1,5 +1,5 @@
 /**
- * (C) Copyright 2016-2023 Intel Corporation.
+ * (C) Copyright 2016-2024 Intel Corporation.
  *
  * SPDX-License-Identifier: BSD-2-Clause-Patent
  */
@@ -88,7 +88,7 @@ dfuse_fuse_init(void *arg, struct fuse_conn_info *conn)
 	DFUSE_TRA_INFO(dfuse_info, "kernel readdir cache support compiled in");
 
 	conn->want |= FUSE_CAP_READDIRPLUS;
-	conn->want |= FUSE_CAP_READDIRPLUS_AUTO;
+	conn->want &= ~FUSE_CAP_READDIRPLUS_AUTO;
 
 #ifdef FUSE_CAP_CACHE_SYMLINKS
 	conn->want |= FUSE_CAP_CACHE_SYMLINKS;
@@ -197,7 +197,7 @@ df_ll_setattr(fuse_req_t req, fuse_ino_t ino, struct stat *attr, int to_set,
 
 	if (handle) {
 		inode                   = handle->doh_ie;
-		handle->doh_linear_read = false;
+		handle->doh_linear_read = 0;
 		DFUSE_IE_STAT_ADD(inode, DS_FSETATTR);
 
 	} else {

src/client/dfuse/ops/create.c

@@ -1,5 +1,5 @@
 /**
- * (C) Copyright 2016-2023 Intel Corporation.
+ * (C) Copyright 2016-2024 Intel Corporation.
  *
  * SPDX-License-Identifier: BSD-2-Clause-Patent
  */
@@ -170,7 +170,7 @@ dfuse_cb_create(fuse_req_t req, struct dfuse_inode_entry *parent, const char *na
 	if (flags & O_APPEND)
 		flags &= ~O_APPEND;
 
-	oh->doh_linear_read = false;
+	oh->doh_linear_read = 0;
 
 	if (!dfuse_info->di_multi_user) {
 		rc = _dfuse_mode_update(req, parent, &mode);
@@ -192,7 +192,7 @@ dfuse_cb_create(fuse_req_t req, struct dfuse_inode_entry *parent, const char *na
 	if (rc)
 		D_GOTO(release, rc);
 
-	oh->doh_writeable = true;
+	oh->doh_writeable = 1;
 
 	if (dfs->dfc_data_timeout != 0 || ie->ie_dfs->dfc_data_otoc) {
 		if (fi->flags & O_DIRECT)
@@ -205,7 +205,7 @@ dfuse_cb_create(fuse_req_t req, struct dfuse_inode_entry *parent, const char *na
 		fi_out.direct_io = 0;
 
 	if (!fi_out.direct_io)
-		oh->doh_caching = true;
+		oh->doh_caching = 1;
 
 	fi_out.fh = (uint64_t)oh;
 

src/client/dfuse/ops/ioctl.c

@@ -353,7 +353,7 @@ handle_cont_query_ioctl(struct dfuse_obj_hdl *oh, fuse_req_t req, const void *in
 static void
 handle_cont_evict_ioctl(fuse_req_t req, struct dfuse_obj_hdl *oh)
 {
-	oh->doh_evict_on_close = true;
+	oh->doh_evict_on_close = 1;
 
 	handle_cont_qe_ioctl_helper(oh, req, NULL);
 }

src/client/dfuse/ops/lookup.c

@@ -266,6 +266,7 @@ dfuse_cb_lookup(fuse_req_t req, struct dfuse_inode_entry *parent,
 	DFUSE_TRA_UP(ie, parent, "inode");
 
 	dfuse_ie_init(dfuse_info, ie);
+	dfuse_dcache_set_time(ie);
 
 	ie->ie_parent = parent->ie_stat.st_ino;
 	ie->ie_dfs = parent->ie_dfs;