Merge pull request #13045 from tvegas1/oshmem_base_exchange.v5.0.x

oshmem/shmem: Allocate and exchange base segment address beforehand

Author: janjust

oshmem/mca/memheap/base/memheap_base_select.c

@@ -25,6 +25,8 @@
 #include "oshmem/mca/sshmem/base/base.h"
 #include "ompi/util/timings.h"
 
+#include <sys/mman.h>
+
 mca_memheap_base_config_t mca_memheap_base_config = {
     .device_nic_mem_seg_size = 0
 };
@@ -106,6 +108,136 @@ static size_t _memheap_size(void)
     return (size_t) memheap_align(oshmem_shmem_info_env.symmetric_heap_size);
 }
 
+static void *memheap_mmap_get(void *hint, size_t size)
+{
+    void *addr;
+
+    addr = mmap(hint, size,
+                PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+    if (addr == MAP_FAILED) {
+        return NULL;
+    }
+
+    return addr;
+}
+
+static int memheap_exchange_base_address(size_t size, void **address)
+{
+    int nprocs = oshmem_num_procs();
+    int need_sync = (*address == NULL);
+    void *base = NULL;
+    void *ptr = NULL;
+    int rc, i;
+    void **bases;
+
+    bases = calloc(nprocs, sizeof(*bases));
+    if (NULL == bases) {
+        return OSHMEM_ERROR;
+    }
+
+    if (oshmem_my_proc_id() == 0) {
+        ptr = memheap_mmap_get(NULL, size);
+        base = ptr;
+    }
+
+    rc = oshmem_shmem_bcast(&base, sizeof(base), 0);
+    if (OSHMEM_SUCCESS != rc) {
+        MEMHEAP_ERROR("Failed to exchange allocated vma for base segment "
+                      "(error %d)", rc);
+        goto out;
+    }
+
+    if (oshmem_my_proc_id() != 0) {
+        ptr = memheap_mmap_get(base, size);
+    }
+
+    MEMHEAP_VERBOSE(100, "#%d: exchange base address: base %p: %s",
+                    oshmem_my_proc_id(), base,
+                    (base == ptr)? "ok" : "unavailable");
+
+    *address = base;
+    if (need_sync) {
+        /* They all succeed or fail to allow fallback */
+        rc = oshmem_shmem_allgather(&ptr, bases, sizeof(ptr));
+        if (OSHMEM_SUCCESS != rc) {
+            MEMHEAP_ERROR("Failed to exchange selected vma for base segment "
+                          "(error %d)", rc);
+            goto out;
+        }
+
+        for (i = 0; i < nprocs; i++) {
+            if ((NULL == bases[i]) || (bases[i] != base)) {
+                *address = NULL;
+                break;
+            }
+        }
+    } else if (ptr != base) {
+        /* Any failure terminates the rank and others start teardown */
+        rc = OSHMEM_ERROR;
+    }
+
+out:
+    if (((OSHMEM_SUCCESS != rc) || (*address == NULL)) && (ptr != NULL)) {
+        (void)munmap(ptr, size);
+    }
+
+    free(bases);
+    return rc;
+}
+
+
+/*
+ * The returned mca_sshmem_base_start_address value is reserved by using
+ * mmap() for the expected size.
+ */
+static int memheap_base_segment_setup(size_t size)
+{
+    int rc;
+
+    if ((mca_sshmem_base_start_address == (void *)UINTPTR_MAX) ||
+        (mca_sshmem_base_start_address == NULL)) {
+        if (UINTPTR_MAX == 0xFFFFFFFF) {
+            /**
+             * if 32 bit we set sshmem_base_start_adress to 0
+             * to let OS allocate segment automatically
+             */
+            mca_sshmem_base_start_address = NULL;
+            return OSHMEM_SUCCESS;
+        }
+
+        rc = memheap_exchange_base_address(size, &mca_sshmem_base_start_address);
+        if (OSHMEM_SUCCESS != rc) {
+            MEMHEAP_ERROR("Failed to setup base segment address (error %d)", rc);
+            return rc;
+        }
+
+        if (NULL != mca_sshmem_base_start_address) {
+            goto done; /* Region is reserved */
+        }
+
+#if defined(__aarch64__)
+        mca_sshmem_base_start_address = (void*)0xAB0000000000;
+#else
+        mca_sshmem_base_start_address = (void*)0xFF000000;
+#endif
+    }
+
+    if (mca_sshmem_base_start_address != memheap_mmap_get(
+                                      mca_sshmem_base_start_address, size)) {
+        MEMHEAP_ERROR("Failed to create segment address %p/%zu",
+                      mca_sshmem_base_start_address, size);
+        return OSHMEM_ERROR;
+    }
+
+done:
+    if (oshmem_my_proc_id() == 0) {
+        MEMHEAP_VERBOSE(10, "Using symmetric segment address %p/%zu",
+                        mca_sshmem_base_start_address, size);
+    }
+
+    return OSHMEM_SUCCESS;
+}
+
 static memheap_context_t* _memheap_create(void)
 {
     int rc = OSHMEM_SUCCESS;
@@ -123,13 +255,18 @@ static memheap_context_t* _memheap_create(void)
 
     OPAL_TIMING_ENV_NEXT(timing, "_memheap_size()");
 
-    /* Inititialize symmetric area */
-    if (OSHMEM_SUCCESS == rc) {
-        rc = mca_memheap_base_alloc_init(&mca_memheap_base_map,
-                                         user_size + MEMHEAP_BASE_PRIVATE_SIZE, 0,
-                                         "regular_mem");
+    /* Locate and reserve symmetric area */
+    rc = memheap_base_segment_setup(user_size + MEMHEAP_BASE_PRIVATE_SIZE);
+    if (OSHMEM_SUCCESS != rc) {
+        MEMHEAP_ERROR("Failed to negotiate base segment addres");
+        return NULL;
     }
 
+    /* Initialize symmetric area */
+    rc = mca_memheap_base_alloc_init(&mca_memheap_base_map,
+                                     user_size + MEMHEAP_BASE_PRIVATE_SIZE, 0,
+                                     "regular_mem");
+
     OPAL_TIMING_ENV_NEXT(timing, "mca_memheap_base_alloc_init()");
 
     /* Initialize atomic symmetric area */

oshmem/mca/memheap/base/memheap_base_static.c

@@ -25,13 +25,17 @@
 #include <pthread.h>
 
 static int _check_perms(const char *perm);
+static int _check_non_static_segment(const map_segment_t *mem_segs,
+                                     int n_segment,
+                                     const void *start, const void *end);
 static int _check_address(void *start, void **end);
 static int _check_pathname(uint64_t inode, const char *pathname);
 
 int mca_memheap_base_static_init(mca_memheap_map_t *map)
 {
     /* read and parse segments from /proc/self/maps */
     int ret = OSHMEM_SUCCESS;
+    int n_segments = map->n_segments;
     uint64_t total_mem = 0;
     void* start;
     void* end;
@@ -54,14 +58,6 @@ int mca_memheap_base_static_init(mca_memheap_map_t *map)
         return OSHMEM_ERROR;
     }
 
-#ifdef __linux__
-    extern unsigned _end;
-    if (mca_sshmem_base_start_address < (uintptr_t)&_end) {
-        MEMHEAP_VERBOSE(1, "sshmem base start address is inside data region"
-                        " (%p < %p)", mca_sshmem_base_start_address, &_end);
-    }
-#endif
-
     while (NULL != fgets(line, sizeof(line), fp)) {
         if (3 > sscanf(line,
                "%llx-%llx %s %llx %s %llx %s",
@@ -77,6 +73,12 @@ int mca_memheap_base_static_init(mca_memheap_map_t *map)
             goto out;
         }
 
+        if (OSHMEM_ERROR == _check_non_static_segment(
+                                                   map->mem_segs, n_segments,
+                                                   start, end)) {
+            continue;
+        }
+
         if (OSHMEM_ERROR == _check_address(start, &end))
             continue;
 
@@ -138,6 +140,26 @@ static int _check_perms(const char *perms)
     return OSHMEM_ERROR;
 }
 
+static int _check_non_static_segment(const map_segment_t *mem_segs,
+                                     int n_segment,
+                                     const void *start, const void *end)
+{
+    int i;
+
+    for (i = 0; i < n_segment; i++) {
+        if ((start <= mem_segs[i].super.va_base) &&
+            (mem_segs[i].super.va_base < end)) {
+            MEMHEAP_VERBOSE(100,
+                            "non static segment: %p-%p already exists as %p-%p",
+                            start, end, mem_segs[i].super.va_base,
+                            mem_segs[i].super.va_end);
+            return OSHMEM_ERROR;
+        }
+    }
+
+    return OSHMEM_SUCCESS;
+}
+
 static int _check_address(void *start, void **end)
 {
     /* FIXME Linux specific code */
@@ -148,11 +170,9 @@ static int _check_address(void *start, void **end)
     /**
      * SGI shmem only supports globals&static in main program.
      * It does not support them in shared objects or in dlopen()
-     * (Clarified on PGAS 2011 tutorial)
+     * (Clarified on PGAS 2011 tutorial).
      *
-     * So ignored any maps that start higher then process _end
-     * FIXME: make sure we do not register symmetric heap twice
-     * if we decide to allow shared objects
+     * So ignored any maps that start higher then process _end.
      */
     if ((uintptr_t)start > data_end) {
         MEMHEAP_VERBOSE(100,

oshmem/mca/sshmem/base/sshmem_base_open.c

@@ -31,17 +31,7 @@
  * globals
  */
 
-/**
- * if 32 bit we set sshmem_base_start_address to 0
- * to let OS allocate segment automatically
- */
-#if UINTPTR_MAX == 0xFFFFFFFF
-void *mca_sshmem_base_start_address = (void*)0;
-#elif defined(__aarch64__)
-void* mca_sshmem_base_start_address = (void*)0xAB0000000000;
-#else
-void* mca_sshmem_base_start_address = (void*)0xFF000000;
-#endif
+void *mca_sshmem_base_start_address = UINTPTR_MAX;
 
 char * mca_sshmem_base_backing_file_dir = NULL;
 

oshmem/mca/sshmem/sysv/sshmem_sysv_module.c

@@ -170,6 +170,7 @@ segment_create(map_segment_t *ds_buf,
     }
 
     /* Attach to the segment */
+    (void)munmap(mca_sshmem_base_start_address, size);
     addr = shmat(shmid, (void *) mca_sshmem_base_start_address, 0);
     if (addr == (void *) -1L) {
         opal_show_help("help-oshmem-sshmem.txt",

oshmem/runtime/oshmem_shmem_exchange.c

@@ -16,6 +16,15 @@
 #include "oshmem/runtime/runtime.h"
 #include "oshmem/runtime/params.h"
 
+int oshmem_shmem_bcast(void *buf, int elem_size, int root)
+{
+    int rc;
+
+    rc = PMPI_Bcast(buf, elem_size, MPI_BYTE, root, oshmem_comm_world);
+
+    return rc;
+}
+
 int oshmem_shmem_allgather(void *send_buf, void *rcv_buf, int elem_size)
 {
     int rc;

oshmem/runtime/runtime.h

@@ -128,6 +128,11 @@ int oshmem_shmem_finalize(void);
  */
 OSHMEM_DECLSPEC int oshmem_shmem_abort(int errcode);
 
+/**
+ * Broadcast between all PEs
+ */
+OSHMEM_DECLSPEC int oshmem_shmem_bcast(void *buf, int elem_size, int root);
+
 /**
  * Allgather between all PEs
  */