Merge pull request #1991 from hoopoepg/topic/v1.2.x

UCX: backports from 1.3/master

config/m4/ucm.m4

@@ -42,3 +42,8 @@ esac
 
 AM_CONDITIONAL([HAVE_UCM_PTMALLOC286],[test "x$HAVE_UCM_PTMALLOC286" = "xyes"])
 
+AC_CHECK_FUNCS([malloc_get_state malloc_set_state],
+               [],
+               [],
+               [#include <stdlib.h>])
+

src/ucp/api/ucp.h

@@ -14,8 +14,10 @@
 #include <ucs/type/thread_mode.h>
 #include <ucs/type/cpu_set.h>
 #include <ucs/config/types.h>
+#include <ucs/sys/compiler_def.h>
 #include <stdio.h>
 
+BEGIN_C_DECLS
 
 /**
  * @defgroup UCP_API Unified Communication Protocol (UCP) API
@@ -2446,5 +2448,6 @@ ucs_status_t ucp_worker_flush(ucp_worker_h worker);
  * UCP hello world client / server example utility.
  */
 
+END_C_DECLS
 
 #endif

src/ucp/api/ucp_compat.h

@@ -10,7 +10,9 @@
 
 
 #include <ucp/api/ucp_def.h>
+#include <ucs/sys/compiler_def.h>
 
+BEGIN_C_DECLS
 
 /**
  * @ingroup UCP_COMM
@@ -32,5 +34,6 @@ void ucp_request_release(void *request);
  */
 void ucp_ep_destroy(ucp_ep_h ep);
 
+END_C_DECLS
 
 #endif

src/ucp/wireup/wireup.c

@@ -452,7 +452,7 @@ ucs_status_t ucp_wireup_init_lanes(ucp_ep_h ep, unsigned address_count,
     key.reachable_md_map |= ucp_ep_config(ep)->key.reachable_md_map;
 
     new_cfg_index = ucp_worker_get_ep_config(worker, &key);
-    if ((ep->cfg_index == new_cfg_index)) {
+    if (ep->cfg_index == new_cfg_index) {
         return UCS_OK; /* No change */
     }
 

src/ucs/algorithm/crc.h

@@ -10,6 +10,10 @@
 #include <stddef.h>
 #include <stdint.h>
 
+#include <ucs/sys/compiler_def.h>
+
+BEGIN_C_DECLS
+
 /**
  * Calculate CRC16 of an arbitrary buffer.
  *
@@ -26,5 +30,6 @@ uint16_t ucs_crc16(const void *buffer, size_t size);
  */
 uint16_t ucs_crc16_string(const char *s);
 
+END_C_DECLS
 
 #endif

src/ucs/algorithm/qsort_r.h

@@ -37,6 +37,9 @@
 
 #include <stddef.h>
 
+#include <ucs/sys/compiler_def.h>
+
+BEGIN_C_DECLS
 
 /**
  * Compare callback for @ref qsort_r.
@@ -59,5 +62,6 @@ typedef int ucs_qsort_r_compare_cb_t(const void *elem1, const void *elem2,
 void ucs_qsort_r(void *base, size_t nmemb, size_t size,
                  ucs_qsort_r_compare_cb_t *compare, void *arg);
 
+END_C_DECLS
 
 #endif

src/ucs/arch/aarch64/cpu.h

@@ -11,6 +11,7 @@
 #include "config.h"
 #include <time.h>
 #include <sys/times.h>
+#include <ucs/sys/compiler_def.h>
 #include <ucs/arch/generic/cpu.h>
 #if __ARM_NEON
 #include <arm_neon.h>
@@ -19,6 +20,8 @@
 
 #define UCS_ARCH_CACHE_LINE_SIZE 64
 
+BEGIN_C_DECLS
+
 /**
  * Assume the worst - weak memory ordering.
  */
@@ -63,8 +66,6 @@ static inline int ucs_arch_get_cpu_flag()
     return UCS_CPU_FLAG_UNKNOWN;
 }
 
-#endif
-
 static inline void ucs_arch_wait_mem(void *address)
 {
     unsigned long tmp;
@@ -73,3 +74,8 @@ static inline void ucs_arch_wait_mem(void *address)
                           : "=&r"(tmp)
                           : "r"(address));
 }
+
+END_C_DECLS
+
+#endif
+

src/ucs/arch/ppc64/cpu.h

@@ -13,9 +13,11 @@
 #ifdef HAVE_SYS_PLATFORM_PPC_H
 #  include <sys/platform/ppc.h>
 #endif
+#include <ucs/sys/compiler_def.h>
 #include <ucs/arch/generic/cpu.h>
 #include <stdint.h>
 
+BEGIN_C_DECLS
 
 #define UCS_ARCH_CACHE_LINE_SIZE 128
 
@@ -53,4 +55,6 @@ double ucs_arch_get_clocks_per_sec();
 
 #define ucs_arch_wait_mem ucs_arch_generic_wait_mem
 
+END_C_DECLS
+
 #endif

src/ucs/arch/x86_64/cpu.h

@@ -10,6 +10,7 @@
 
 #include <ucs/sys/compiler.h>
 #include <ucs/arch/generic/cpu.h>
+#include <ucs/sys/compiler_def.h>
 #include <stdint.h>
 
 #ifdef __SSE4_1__
@@ -19,6 +20,7 @@
 #  include <immintrin.h>
 #endif
 
+BEGIN_C_DECLS
 
 #define UCS_ARCH_CACHE_LINE_SIZE 64
 
@@ -48,6 +50,7 @@ ucs_cpu_flag_t ucs_arch_get_cpu_flag() UCS_F_NOOPTIMIZE;
 
 #define ucs_arch_wait_mem ucs_arch_generic_wait_mem
 
+END_C_DECLS
 
 #endif
 

src/ucs/async/async.h

@@ -11,10 +11,12 @@
 #include "signal.h"
 #include "async_fwd.h"
 
+#include <ucs/sys/compiler_def.h>
 #include <ucs/datastruct/mpmc.h>
 #include <ucs/time/time.h>
 #include <ucs/debug/log.h>
 
+BEGIN_C_DECLS
 
 /**
  * Async event context. Manages timer and fd notifications.
@@ -119,4 +121,6 @@ static inline int ucs_async_check_miss(ucs_async_context_t *async)
     } while (0)
 
 
+END_C_DECLS
+
 #endif

src/ucs/async/async_fwd.h

@@ -10,6 +10,7 @@
 #include <ucs/config/types.h>
 #include <ucs/time/time_def.h>
 
+BEGIN_C_DECLS
 
 typedef struct ucs_async_context ucs_async_context_t;
 
@@ -121,5 +122,6 @@ void ucs_async_poll(ucs_async_context_t *async);
 
 void __ucs_async_poll_missed(ucs_async_context_t *async);
 
+END_C_DECLS
 
 #endif

src/ucs/config/global_opts.c

@@ -29,9 +29,7 @@ ucs_global_opts_t ucs_global_opts = {
     .stats_dest            = "",
     .tuning_path           = "",
     .memtrack_dest         = "",
-    .stats_dest            = "",
     .stats_trigger         = "exit",
-    .memtrack_dest         = "",
     .profile_mode          = 0,
     .profile_file          = "",
     .stats_filter          = { NULL, 0 },

src/ucs/config/global_opts.h

@@ -11,9 +11,11 @@
 
 #include <ucs/stats/stats_fwd.h>
 #include <ucs/type/status.h>
+#include <ucs/sys/compiler_def.h>
 #include <stddef.h>
 #include <stdio.h>
 
+BEGIN_C_DECLS
 
 
 /**
@@ -108,4 +110,6 @@ ucs_status_t ucs_global_opts_clone(void *dst);
 void ucs_global_opts_release();
 void ucs_global_opts_print(FILE *stream, ucs_config_print_flags_t print_flags);
 
+END_C_DECLS
+
 #endif

src/ucs/config/parser.h

@@ -11,12 +11,14 @@
 
 #include <ucs/type/status.h>
 #include <ucs/sys/math.h>
+#include <ucs/sys/compiler_def.h>
 #include <stdio.h>
 
 
 #define UCS_CONFIG_PREFIX      "UCX_"
 #define UCS_CONFIG_ARRAY_MAX   128
 
+BEGIN_C_DECLS
 
 /*
  * Configuration varaibles syntax:
@@ -326,4 +328,6 @@ size_t ucs_config_memunits_get(size_t config_size, size_t auto_size,
 int ucs_config_names_search(ucs_config_names_array_t config_names,
                             const char *str);
 
+END_C_DECLS
+
 #endif

src/ucs/datastruct/callbackq.h

@@ -11,6 +11,9 @@
 #include <ucs/async/async_fwd.h>
 #include <ucs/datastruct/list_types.h>
 #include <stdint.h>
+#include <ucs/sys/compiler_def.h>
+
+BEGIN_C_DECLS
 
 /*
  *
@@ -236,6 +239,7 @@ void ucs_callbackq_add_slow_path(ucs_callbackq_t *cbq,
 void ucs_callbackq_remove_slow_path(ucs_callbackq_t *cbq,
                                     ucs_callbackq_slow_elem_t* elem);
 
+END_C_DECLS
 
 /**
  * Remove all slow path elements with a given callback function from the list.

src/ucs/datastruct/mpool.h

@@ -11,6 +11,9 @@
 #include <ucs/debug/log.h>
 #include <ucs/debug/memtrack.h>
 #include <ucs/type/status.h>
+#include <ucs/sys/compiler_def.h>
+
+BEGIN_C_DECLS
 
 
 typedef struct ucs_mpool_chunk   ucs_mpool_chunk_t;
@@ -235,5 +238,6 @@ void ucs_mpool_chunk_munmap(ucs_mpool_t *mp, void *chunk);
 ucs_status_t ucs_mpool_hugetlb_malloc(ucs_mpool_t *mp, size_t *size_p, void **chunk_p);
 void ucs_mpool_hugetlb_free(ucs_mpool_t *mp, void *chunk);
 
+END_C_DECLS
 
 #endif /* MPOOL_H_ */

src/ucs/debug/log.h

@@ -16,6 +16,7 @@
 #include <ucs/config/global_opts.h>
 #include <stdint.h>
 
+BEGIN_C_DECLS
 
 #define ucs_log_enabled(_level) \
     ucs_unlikely(((_level) <= UCS_MAX_LOG_LEVEL) && ((_level) <= (ucs_global_opts.log_level)))
@@ -144,5 +145,7 @@ const char *ucs_log_dump_hex(const void* data, size_t length, char *buf, size_t
 void ucs_log_push_handler(ucs_log_func_t handler);
 void ucs_log_pop_handler();
 
+END_C_DECLS
+
 #endif
 

src/ucs/debug/memtrack.h

@@ -11,11 +11,14 @@
 #  include "config.h"
 #endif
 
+#include <ucs/sys/compiler_def.h>
+
 #include <sys/types.h>
 #include <malloc.h>
 #include <stdlib.h>
 #include <stdio.h>
 
+BEGIN_C_DECLS
 
 enum {
     UCS_MEMTRACK_STAT_ALLOCATION_COUNT,
@@ -158,5 +161,6 @@ char *ucs_strdup(const char *src, const char *name);
 
 #endif /* ENABLE_MEMTRACK */
 
+END_C_DECLS
 
 #endif

src/ucs/debug/profile.c

@@ -32,9 +32,9 @@ static void ucs_profile_file_write_data(int fd, void *data, size_t size)
 {
     ssize_t written = write(fd, data, size);
     if (written < 0) {
-        ucs_warn("failed to write %Zu bytes to profiling file: %m", size);
+        ucs_warn("failed to write %zu bytes to profiling file: %m", size);
     } else if (size != written) {
-        ucs_warn("wrote only %Zd of %Zu bytes to profiling file: %m",
+        ucs_warn("wrote only %zd of %zu bytes to profiling file: %m",
                  written, size);
     }
 }

src/ucs/stats/client_server.c

@@ -333,36 +333,36 @@ ucs_stats_server_entity_update(ucs_stats_server_h server, stats_entity_t *entity
     frag_hole_t *hole, *new_hole;
     void *frag_start, *frag_end, *hole_end;
 
-    ucs_debug("From %s:%d - timestamp %"PRIu64", %Zu..%Zu / %Zu",
+    ucs_debug("From %s:%d - timestamp %"PRIu64", %zu..%zu / %zu",
               inet_ntoa(entity->in_addr.sin_addr), ntohs(entity->in_addr.sin_port),
               timestamp, frag_offset, frag_offset + frag_size, total_size);
 
     if (timestamp < entity->timestamp) {
         ucs_debug("Dropping - old timestamp");
         return 0;
     } else if (timestamp > entity->timestamp) {
-        ucs_debug("New timestamp, resetting buffer with size %Zu", total_size);
+        ucs_debug("New timestamp, resetting buffer with size %zu", total_size);
         entity->timestamp = timestamp;
         ucs_stats_server_entity_reset_buffer(entity, total_size);
     } else {
         /* Make sure all packets in this timestamp have the same 'total_size' */
         if (entity->buffer_size != total_size) {
-            ucs_error("Total size in the packet is %Zu, but expected is %Zu",
+            ucs_error("Total size in the packet is %zu, but expected is %zu",
                       total_size, entity->buffer_size);
         }
     }
 
     hole = find_frag_hole(entity, frag_size, frag_offset);
     if (hole == NULL) {
-        ucs_error("cannot fill fragment (offset %Zu size %Zu)", frag_offset, frag_size);
+        ucs_error("cannot fill fragment (offset %zu size %zu)", frag_offset, frag_size);
         return UCS_ERR_MESSAGE_TRUNCATED;
     }
 
     frag_start = entity->inprogress_buffer + frag_offset;
     frag_end   = entity->inprogress_buffer + frag_offset + frag_size;
     hole_end   = (void*)hole + hole->size;
 
-    ucs_debug("inserting into a hole of %Zu..%Zu",
+    ucs_debug("inserting into a hole of %zu..%zu",
               (void*)hole - entity->inprogress_buffer,
               hole_end    - entity->inprogress_buffer);
 
@@ -414,7 +414,7 @@ ucs_stats_server_update_context(ucs_stats_server_h server, struct sockaddr_in *s
 
     /* Validate fragment size */
     if (pkt_len != pkt->frag_size + sizeof(ucs_stats_packet_hdr_t)) {
-        ucs_error("Invalid receive size: expected %Zu, got %Zu",
+        ucs_error("Invalid receive size: expected %zu, got %zu",
                   pkt->frag_size + sizeof(ucs_stats_packet_hdr_t), pkt_len);
         return UCS_ERR_MESSAGE_TRUNCATED;
     }