First cut of sm coll component infrastrcutre (this is what took so

much time) and somewhat-lame implementation of barrier (need to
precompute some more stuff rather than calculate it every time).
Checkpointing so I can try this on another machine...

This commit was SVN r6985.

ompi/mca/coll/sm/coll_sm.h

@@ -26,12 +26,64 @@
 #include "opal/mca/mca.h"
 #include "ompi/mca/coll/coll.h"
 #include "ompi/mca/mpool/mpool.h"
+#include "ompi/mca/common/sm/common_sm_mmap.h"
 
 #if defined(c_plusplus) || defined(__cplusplus)
 extern "C" {
 #endif
 
-#define PUB(foo) mca_coll_sm##foo
+    /**
+     * Structure used within bootstrap mmap file for setting up a coll
+     * sm component on a communicator
+     */
+    struct mca_coll_sm_bootstrap_comm_setup_t {
+        /** Offset to be used in the data mpool for this comm's
+            collective sm operations -- use this value plus the base
+            of the mpool to obtain the pointer to this comm's
+            mca_coll_sm_mpool_area_t */
+        size_t smbcs_data_mpool_offset;
+
+        /** Number of segments in the data mpool area for this
+            communicator */
+        int smbcs_communicator_num_segments;
+
+        /** Number of processes in this communicator who have seen
+            this value already. */
+        int smbcs_count;
+    };
+    /**
+     * Convenience typedef
+     */
+    typedef struct mca_coll_sm_bootstrap_comm_setup_t 
+        mca_coll_sm_bootstrap_comm_setup_t;
+
+    /**
+     * Extension to the control structure in the bootstrap mmap file
+     */
+    struct mca_coll_sm_bootstrap_header_extension_t {
+        /** upper-level control structure */
+        mca_common_sm_file_header_t super;
+
+        /** Number of segments in the bootstrap mmap file */
+        int smbhe_num_segments;
+
+        /** Pointer to the start of the segments in the bootstrap area
+            (map->seg_data only points to just beyond the
+            mca_common_sm_file_header_t) */
+        mca_coll_sm_bootstrap_comm_setup_t *smbhe_segments;
+
+        /** Pointer to array containing smhe_num_segments CIDs for use
+            in bootstrap phase -- will always point immediately after
+            the end of this struct (i.e., still within this header,
+            but since it's variable size (set by MCA param), we can't
+            just have it here in the struct.  Bonk). */
+        uint32_t *smbhe_cids;
+    };
+    /**
+     * Convenience typedef
+     */
+    typedef struct mca_coll_sm_bootstrap_header_extension_t
+        mca_coll_sm_bootstrap_header_extension_t;
 
     /** 
      * Structure to hold the sm coll component.  First it holds the
@@ -39,33 +91,119 @@ extern "C" {
      * sm-coll-component-specific stuff (e.g., current MCA param
      * values). 
      */
-    typedef struct mca_coll_sm_component_t {
+    struct mca_coll_sm_component_t {
         /** Base coll component */
         mca_coll_base_component_1_0_0_t super;
 
-        /** Priority of this component */
+        /** MCA parameter: Priority of this component */
         int sm_priority;
-        /** Name of the mpool that this component will look for */
+
+        /** MCA parameter: Length of a cache line or page (in bytes) */
+        int sm_control_size;
+
+        /** MCA parameter: Name of shared memory control / bootstrap
+            mmap file */
+        char *sm_bootstrap_filename;
+
+        /** MCA parameter: Number of segments in the bootstrap file
+            (for use with setting up multiple comm's with sm
+            components simultaneously) */
+        int sm_bootstrap_num_segments;
+
+        /** MCA parameter: Name of the mpool that this component will
+            use */
         char *sm_mpool_name;
-        /** Mpool that will be used */
-        mca_mpool_base_module_t *sm_mpool;
-    } mca_coll_sm_component_t;
+
+        /** MCA parameter: Number of segments for each communicator in
+            the data mpool */
+        int sm_communicator_num_segments;
+
+        /** MCA parameter: Fragment size for data */
+        int sm_fragment_size;
+
+        /** MCA parameter: Degree of tree for tree-based collectives */
+        int sm_tree_degree;
+
+        /** Size of the bootstrap area -- calculated in
+            coll_sm_component.c */
+        size_t sm_bootstrap_size;
+
+        /** Size of the data mpool area -- calculated in
+            coll_sm_component.c */
+        size_t sm_data_mpool_size;
+
+        /** Data mpool that will be used */
+        mca_mpool_base_module_t *sm_data_mpool;
+
+        /** Whether we ended up creating the sm mpool or whether
+            someone else created it and we just found it */
+        bool sm_data_mpool_created;
+
+        /** Meta struct containing information about the bootstrap area */
+        mca_common_sm_mmap_t *sm_bootstrap_meta;
+    };
+    /**
+     * Convenience typedef
+     */
+    typedef struct mca_coll_sm_component_t mca_coll_sm_component_t;
 
     /**
-     * Structure for sm collective module, per communicator. The
-     * structure mainly stores memory pointers to the specific
-     * poritions in the shared memory area. Each shared memory area is
-     * reserved for special functions.  The shared memory is split
-     * between two types of areas. One is control section that stores
-     * shared flags used during synchronization, while other section
-     * is purely used to pass messages from one process to other. 
+     * Structure containing pointers to various arrays of data in the
+     * data mpool area (one of these indexes a single segment in the
+     * data mpool).  Nothing is hard-coded because all the array
+     * lengths and displacements of the pointers all depend on how
+     * many processes are in the communicator.
      */
-    typedef struct mca_coll_base_module_comm_t {
+    struct mca_coll_base_mpool_index_t {
+        /** Pointer to beginning of control fan-in data */
+        char *mcbmi_control_fan_in;
+        /** Pointer to beginning of control fan-out data */
+        char *mcbmi_control_fan_out;
+        /** Pointer to beginning of message data fan-in data */
+        char *mcbmi_data_fan_in;
+        /** Pointer to beginning of message data fan-out data */
+        char *mcbmi_data_fan_out;
+    };
+    typedef struct mca_coll_base_mpool_index_t mca_coll_base_mpool_index_t;
 
-        /* JMS fill in here */
-        int foo;
+    /**
+     * Structure for the sm coll module to hang off the communicator.
+     * Contains communicator-specific information, including pointers
+     * into the data mpool for this comm's sm collective operations
+     * area. 
+     */
+    struct mca_coll_base_comm_t {
+        /** If this process is the one that invoked mpool_alloc() for
+            the data segment, the return value will be in here.
+            Otherwise, it will be NULL (i.e., only the allocating
+            process will call free). */
+        void *mcb_data_mpool_malloc_addr;
+        /** Base of the data mpool */
+        char *mcb_mpool_base;
+        /** Offset into the data mpool where this comm's operations
+            area is */
+        size_t mcb_mpool_offset;
+        /** Pointer in the data mpool to the beginning of this comm's
+            operations area (i.e., mcb_mpool_base +
+            mcb_mpool_offset) */
+        char *mcb_mpool_area;
+        /** Number of segments in this comm's area in the data mpool */
+        int mcb_mpool_num_segments;
 
-    } mca_coll_base_module_comm_t;
+        /** Array of indexes into the mpool area (containing pointers
+            to each segments control and data areas).  This array will
+            be located immediately after the instance of this struct
+            in memory (i.e., so this struct and the array that this
+            member points to will be adjacent in memory). */
+        mca_coll_base_mpool_index_t **mcb_mpool_index;
+
+        /** Operation number (i.e., which segment number to use) */
+        int mcb_operation_count;
+    };
+    /**
+     * Convenience typedef
+     */
+    typedef struct mca_coll_base_comm_t mca_coll_base_comm_t;
         
     
     /**
@@ -88,10 +226,7 @@ extern "C" {
     int mca_coll_sm_comm_unquery(struct ompi_communicator_t *comm,
                                  struct mca_coll_base_comm_t *data);
 
-    const struct mca_coll_base_module_1_0_0_t *
-    mca_coll_sm_module_init(struct ompi_communicator_t *comm);
-
-    int mca_coll_sm_module_finalize(struct ompi_communicator_t *comm);
+    int mca_coll_sm_bootstrap_finalize(void);
 
     int mca_coll_sm_allgather_intra(void *sbuf, int scount, 
                                     struct ompi_datatype_t *sdtype, 

ompi/mca/coll/sm/coll_sm_barrier.c

@@ -17,8 +17,16 @@
 #include "ompi_config.h"
 
 #include "ompi/include/constants.h"
+#include "ompi/communicator/communicator.h"
+#include "ompi/mca/coll/coll.h"
+#include "opal/include/sys/atomic.h"
 #include "coll_sm.h"
 
+#if 0
+#define D(foo) printf foo
+#else
+#define D(foo)
+#endif
 
 /*
  *	barrier
@@ -29,5 +37,108 @@
  */
 int mca_coll_sm_barrier_intra(struct ompi_communicator_t *comm)
 {
-    return OMPI_ERR_NOT_IMPLEMENTED;
+    mca_coll_base_comm_t *data = comm->c_coll_selected_data;
+    uint32_t *my_control_in, *my_control_out;
+    uint32_t *parent_control_in;
+    int i, rank, start_rank, parent, num_children, segment;
+    char *control_in, *control_out;
+
+    segment = (++data->mcb_operation_count % data->mcb_mpool_num_segments);
+    control_in = data->mcb_mpool_index[segment]->mcbmi_control_fan_in;
+    control_out = data->mcb_mpool_index[segment]->mcbmi_control_fan_out;
+
+    /* THIS CAN BE PRECOMPUTED */
+    /* Figure out some identities */
+
+    rank = ompi_comm_rank(comm);
+    num_children = mca_coll_sm_component.sm_tree_degree;
+    parent = (rank - 1) / mca_coll_sm_component.sm_tree_degree;
+
+    /* Do we have children?  If so, how many? */
+
+    if ((rank * num_children) + 1 >= ompi_comm_size(comm)) {
+        /* Leaves */
+        num_children = 0;
+    } else {
+        int min_child = rank * num_children + 1;
+        int max_child = rank * num_children + num_children;
+        if (max_child >= ompi_comm_size(comm)) {
+            max_child = ompi_comm_size(comm) - 1;;
+        }
+        D(("rank %d: min child: %d, max child: %d\n", rank, min_child, max_child));
+        num_children = max_child - min_child + 1;
+    }
+    D(("rank %d: segment %d (opn count: %d), parent %d, num_children = %d\n", 
+       rank, segment, data->mcb_operation_count, parent, num_children));
+    fflush(stdout);
+
+    /* Pre-calculate some pointers */
+
+    my_control_in = (uint32_t *)
+        (control_in + (rank * mca_coll_sm_component.sm_control_size));
+    my_control_out = (uint32_t *)
+        (control_out + (rank * mca_coll_sm_component.sm_control_size));
+    *my_control_out = 0;
+
+    if (0 != rank) {
+        parent_control_in = (uint32_t *)
+            (control_in + (parent * mca_coll_sm_component.sm_control_size));
+    } else {
+        parent_control_in = NULL;
+    }
+
+    /* Fan in: wait for my children */
+
+    if (0 != num_children) {
+        D(("rank %d waiting for fan in from %d children...\n", rank, num_children));
+        while (*my_control_in != (uint32_t) num_children) {
+            opal_atomic_wmb();
+        }
+        *my_control_in = 0;
+        D(("rank %d got fan in\n", rank));
+    }
+
+    /* Fan in: send to my parent */
+
+    if (NULL != parent_control_in) {
+        D(("rank %d writing to parent\n", rank));
+        opal_atomic_add(parent_control_in, 1);
+        D(("rank %d wrote to parent: %d\n", rank, *parent_control_in));
+    }
+
+    /* Fan out: wait for my parent to write to me (don't poll on
+       parent's out buffer -- that would cause a lot of network
+       traffic / contention / faults / etc. -- this way, the children
+       poll on local memory and therefore only num_children messages
+       are sent across the network [vs. num_children *each* time all
+       the children poll] -- i.e., the memory is only being polled by
+       one process, and it is only changed *once* by an external
+       process) */
+
+    if (NULL != parent_control_in) {
+        D(("rank %d waiting for fan out from parent\n", rank));
+        while (0 == *my_control_out) {
+            opal_atomic_wmb();
+        }
+        D(("rank %d got fan out from parent\n", rank));
+    }
+
+    /* Fan out: send to my children */
+
+    start_rank = (rank * mca_coll_sm_component.sm_tree_degree) + 1;
+    for (i = 0; i < num_children; ++i) {
+        D(("rank %d writing fan out to child %d, rank %d (start %d, num_children %d)\n", rank, i, start_rank + i, start_rank, num_children));
+        *((uint32_t *) 
+          (control_out + ((start_rank + i) * mca_coll_sm_component.sm_control_size))) = 1;
+        
+    }
+    D(("rank %d done with barrier\n", rank));
+
+    /* All done!  End state of the control segment:
+
+       my_control_in: 0
+       my_control_out: 1
+    */
+
+    return OMPI_SUCCESS;
 }

ompi/mca/coll/sm/coll_sm_component.c

@@ -23,7 +23,9 @@
 #include "ompi_config.h"
 
 #include "ompi/include/constants.h"
-#include "mca/coll/coll.h"
+#include "ompi/communicator/communicator.h"
+#include "ompi/mca/coll/coll.h"
+#include "opal/util/show_help.h"
 #include "coll_sm.h"
 
 
@@ -91,12 +93,38 @@ mca_coll_sm_component_t mca_coll_sm_component = {
 
     /* sm-component specifc information */
 
-    /* priority */
+    /* (default) priority */
     75,
 
-    /* mpool name and instance */
+    /* (default) control unit size (bytes) */
+    64,
+
+    /* (default) bootstrap filename */
+    "coll-sm-bootstrap",
+
+    /* (default) number of segments in bootstrap file */
+    8,
+
+    /* (default) mpool name to use */
     "sm",
-    NULL
+
+    /* (default) number of segments for each communicator in the mpool
+       area */
+    2,
+
+    /* (default) fragment size */
+    8192,
+
+    /* (default) degree of tree for tree-based operations (must be <=
+       control unit size) */
+    4,
+
+    /* default values for non-MCA parameters */
+    0, /* bootstrap size -- filled in below */
+    0, /* mpool data size -- filled in below */
+    NULL, /* data mpool pointer */
+    false, /* whether this process created the data mpool */
+    NULL /* pointer to meta data about bootstrap area */
 };
 
 
@@ -105,19 +133,115 @@ mca_coll_sm_component_t mca_coll_sm_component = {
  */
 static int sm_open(void)
 {
-    int p, ival;
-    char *sval;
+    size_t size1, size2;
+    mca_base_component_t *c = &mca_coll_sm_component.super.collm_version;
+    mca_coll_sm_component_t *cs = &mca_coll_sm_component;
 
     /* If we want to be selected (i.e., all procs on one node), then
        we should have a high priority */
 
-    p = mca_base_param_register_int("coll", "sm", "priority", NULL, 75);
-    mca_base_param_lookup_int(p, &ival);
-    mca_coll_sm_component.sm_priority = ival;
+    mca_base_param_reg_int(c, "priority", 
+                           "Priority of the sm coll component",
+                           false, false, 
+                           cs->sm_priority,
+                           &cs->sm_priority);
 
-    p = mca_base_param_register_string("coll", "sm", "mpool", NULL, "sm");
-    mca_base_param_lookup_string(p, &sval);
-    mca_coll_sm_component.sm_mpool_name = sval;
+    mca_base_param_reg_int(c, "control_size",
+                           "Length of the control data -- should usually be either a cache line on most SMPs, or a page on machine where pages that support direct memory affinity placement (in bytes)",
+                           false, false,
+                           cs->sm_control_size,
+                           &cs->sm_control_size);
+
+    mca_base_param_reg_string(c, "bootstrap_filename", 
+                              "Filename (in the Open MPI session directory) of the coll sm component bootstrap rendezvous mmap file",
+                              false, false,
+                              cs->sm_bootstrap_filename,
+                              &cs->sm_bootstrap_filename);
+
+    mca_base_param_reg_int(c, "bootstrap_num_segments",
+                           "Number of segments in the bootstrap file",
+                           false, false,
+                           cs->sm_bootstrap_num_segments,
+                           &cs->sm_bootstrap_num_segments);
+
+    mca_base_param_reg_int(c, "fragment_size",
+                           "Fragment size (in bytes) used for passing data through shared memory (will be rounded up to the nearest control_size size)",
+                           false, false,
+                           cs->sm_fragment_size,
+                           &cs->sm_fragment_size);
+    if (0 != (cs->sm_fragment_size % cs->sm_control_size)) {
+        cs->sm_fragment_size += cs->sm_control_size - 
+            (cs->sm_fragment_size % cs->sm_control_size);
+    }
+
+    mca_base_param_reg_string(c, "mpool", 
+                              "Name of the mpool component to use",
+                              false, false,
+                              cs->sm_mpool_name,
+                              &cs->sm_mpool_name);
+    
+    mca_base_param_reg_int(c, "communicator_num_segments",
+                           "Number of shared memory collective segments on each communicator",
+                           false, false,
+                           cs->sm_communicator_num_segments,
+                           &cs->sm_communicator_num_segments);
+
+    mca_base_param_reg_int(c, "tree_degree",
+                           "Degree of the tree for tree-based operations (must be <= control size and <= 255)",
+                           false, false,
+                           cs->sm_tree_degree,
+                           &cs->sm_tree_degree);
+    if (cs->sm_tree_degree > cs->sm_control_size) {
+        opal_show_help("help-coll-sm.txt", 
+                       "tree-degree-larger-than-control", true,
+                       cs->sm_tree_degree, cs->sm_control_size);
+        cs->sm_tree_degree = cs->sm_control_size;
+    }
+    if (cs->sm_tree_degree > 255) {
+        opal_show_help("help-coll-sm.txt", 
+                       "tree-degree-larger-than-255", true,
+                       cs->sm_tree_degree);
+        cs->sm_tree_degree = 255;
+    }
+
+    /* Size of the bootstrap shared mb
+emory area. */
+
+    size1 = 
+        sizeof(mca_coll_sm_bootstrap_header_extension_t) +
+        (mca_coll_sm_component.sm_bootstrap_num_segments *
+         sizeof(mca_coll_sm_bootstrap_comm_setup_t)) +
+        (sizeof(uint32_t) * mca_coll_sm_component.sm_bootstrap_num_segments);
+    mca_base_param_reg_int(c, "shared_mem_used_bootstrap",
+                           "Amount of shared memory used in the shared memory bootstrap area (in bytes)",
+                           false, true,
+                           size1, NULL);
+
+    /* Calculate how much space we need in the data mpool.  There are
+       several values to add (one of these for each segment):
+
+       - size of the control data:
+           - fan-in data (num_procs * control_size size)
+           - fan-out data (num_procs * control_size size)
+       - size of message data
+           - fan-in data (num_procs * (frag_size rounded up to
+             control_size size))
+           - fan-out data (num_procs * (frag_size rounded up
+             to control_size size))
+
+       So it's:
+
+           num_segs * ((num_procs * control_size * 2) + (num_procs * frag * 2))
+
+       Which reduces to:
+
+           num_segs * num_procs * 2 * (control_size + frag)
+
+       For this example, assume num_procs = 1.
+    */
+
+    size2 = cs->sm_communicator_num_segments * 2 *
+        (cs->sm_control_size + cs->sm_fragment_size);
 
     return OMPI_SUCCESS;
 }
@@ -133,5 +257,7 @@ static int sm_close(void)
         mca_coll_sm_component.sm_mpool_name = NULL;
     }
 
+    mca_coll_sm_bootstrap_finalize();
+
     return OMPI_SUCCESS;
 }

ompi/mca/coll/sm/coll_sm_module.c

@@ -17,6 +17,10 @@
 #include "ompi_config.h"
 
 #include <stdio.h>
+#ifdef HAVE_SCHED_H
+#include <sched.h>
+#endif
+#include <sys/mman.h>
 
 #include "mpi.h"
 #include "ompi/communicator/communicator.h"
@@ -27,6 +31,23 @@
 #include "coll_sm.h"
 
 
+/*
+ * Local functions
+ */
+static const struct mca_coll_base_module_1_0_0_t *
+    sm_module_init(struct ompi_communicator_t *comm);
+static int sm_module_finalize(struct ompi_communicator_t *comm);
+static bool have_local_peers(ompi_proc_t **procs, size_t size);
+static int bootstrap_init(void);
+static int bootstrap_comm(ompi_communicator_t *comm);
+
+
+/*
+ * Local variables
+ */
+static bool bootstrap_inited = false;
+
+
 /*
  * Linear set of collective algorithms
  */
@@ -34,8 +55,8 @@ static const mca_coll_base_module_1_0_0_t module = {
 
     /* Initialization / finalization functions */
 
-    mca_coll_sm_module_init,
-    mca_coll_sm_module_finalize,
+    sm_module_init,
+    sm_module_finalize,
 
     /* Collective function pointers */
 
@@ -61,12 +82,59 @@ static const mca_coll_base_module_1_0_0_t module = {
 /*
  * Initial query function that is invoked during MPI_INIT, allowing
  * this component to disqualify itself if it doesn't support the
- * required level of thread support.
+ * required level of thread support.  This function is invoked exactly
+ * once.
  */
 int mca_coll_sm_init_query(bool enable_progress_threads,
                            bool enable_mpi_threads)
 {
-    /* Nothing to do */
+    int ret;
+#if 0
+    /* JMS: Arrgh.  Unfortunately, we don't have this information by
+       the time this is invoked -- the GPR compound command doesn't
+       fire until after coll_base_open() in ompi_mpi_init(). */
+
+    ompi_proc_t **procs;
+    size_t size;
+
+    /* Check to see if anyone is local on this machine.  If not, don't
+       bother doing anything else. */
+
+    procs = ompi_proc_all(&size);
+    if (NULL == procs || 0 == size) {
+        return OMPI_ERROR;
+    }
+    if (!have_local_peers(procs, size)) {
+        return OMPI_ERROR;
+    }
+    free(procs);
+#endif
+
+    /* Ok, we have local peers.  So setup the bootstrap file */
+
+    if (OMPI_SUCCESS != (ret = bootstrap_init())) {
+        return ret;
+    }
+
+    /* Can we get an mpool allocation?  See if there was one created
+       already.  If not, try to make one. */
+
+    mca_coll_sm_component.sm_data_mpool = 
+        mca_mpool_base_module_lookup(mca_coll_sm_component.sm_mpool_name);
+    if (NULL == mca_coll_sm_component.sm_data_mpool) {
+        mca_coll_sm_component.sm_data_mpool = 
+            mca_mpool_base_module_create(mca_coll_sm_component.sm_mpool_name,
+                                         NULL, NULL);
+        if (NULL == mca_coll_sm_component.sm_data_mpool) {
+            mca_coll_sm_bootstrap_finalize();
+            return OMPI_ERR_OUT_OF_RESOURCE;
+        }
+        mca_coll_sm_component.sm_data_mpool_created = true;
+    } else {
+        mca_coll_sm_component.sm_data_mpool_created = false;
+    }
+
+    /* Alles gut */
 
     return OMPI_SUCCESS;
 }
@@ -81,12 +149,13 @@ const mca_coll_base_module_1_0_0_t *
 mca_coll_sm_comm_query(struct ompi_communicator_t *comm, int *priority,
                        struct mca_coll_base_comm_t **data)
 {
-    int i;
-
     /* If we're intercomm, or if there's only one process in the
-       communicator, we don't want to run */
+       communicator, or if not all the processes in the communicator
+       are not on this node, then we don't want to run */
 
-    if (OMPI_COMM_IS_INTER(comm) || 1 == ompi_comm_size(comm)) {
+    if (OMPI_COMM_IS_INTER(comm) || 1 == ompi_comm_size(comm) ||
+        !have_local_peers(comm->c_local_group->grp_proc_pointers,
+                          ompi_comm_size(comm))) {
 	return NULL;
     }
 
@@ -94,30 +163,6 @@ mca_coll_sm_comm_query(struct ompi_communicator_t *comm, int *priority,
 
     *priority = mca_coll_sm_component.sm_priority;
     
-    /* We only want to run if all the processes in the communicator
-       are on the same node */
-
-    for (i = 0; i < ompi_comm_size(comm); ++i) {
-        if (0 == (comm->c_local_group->grp_proc_pointers[i]->proc_flags &
-                  OMPI_PROC_FLAG_LOCAL)) {
-            return NULL;
-        }
-    }
-
-    /* Can we get an mpool allocation?  See if there was one created
-       already.  If not, try to make one. */
-
-    mca_coll_sm_component.sm_mpool = 
-        mca_mpool_base_module_lookup(mca_coll_sm_component.sm_mpool_name);
-    if (NULL == mca_coll_sm_component.sm_mpool) {
-        mca_coll_sm_component.sm_mpool = 
-            mca_mpool_base_module_create(mca_coll_sm_component.sm_mpool_name,
-                                         NULL, NULL);
-        if (NULL == mca_coll_sm_component.sm_mpool) {
-            return NULL;
-        }
-    }
-
     /* All is good -- return a module */
 
     return &module;
@@ -130,9 +175,6 @@ mca_coll_sm_comm_query(struct ompi_communicator_t *comm, int *priority,
 int mca_coll_sm_comm_unquery(struct ompi_communicator_t *comm,
                              struct mca_coll_base_comm_t *data)
 {
-    /* JMS */
-    /* Remove mpool query, if we got one */
-
     return OMPI_SUCCESS;
 }
 
@@ -140,10 +182,77 @@ int mca_coll_sm_comm_unquery(struct ompi_communicator_t *comm,
 /*
  * Init module on the communicator
  */
-const struct mca_coll_base_module_1_0_0_t *
-mca_coll_sm_module_init(struct ompi_communicator_t *comm)
+static const struct mca_coll_base_module_1_0_0_t *
+sm_module_init(struct ompi_communicator_t *comm)
 {
-    /* JMS */
+    int i;
+    mca_coll_base_comm_t *data;
+    size_t control_size, data_size, total_size;
+    mca_coll_sm_component_t *c = &mca_coll_sm_component;
+
+    /* Allocate data to hang off the communicator.  The memory we
+       alloc will be laid out as follows:
+
+       1. mca_coll_base_comm_t
+       2. array of num_segments pointers to
+          mca_coll_base_mpool_index_t instances
+       3. array of num_segments mca_coll_base_mpool_index_t instances
+          (pointed to by the array in 2)
+    */
+
+    data = malloc(sizeof(mca_coll_base_comm_t) + 
+                  (mca_coll_sm_component.sm_bootstrap_num_segments * 
+                   (sizeof(mca_coll_base_mpool_index_t *) +
+                    sizeof(mca_coll_base_mpool_index_t))));
+    if (NULL == data) {
+        return NULL;
+    }
+    data->mcb_data_mpool_malloc_addr = NULL;
+
+    /* Setup #2: set the array to point immediately beyond the
+       mca_coll_base_comm_t */
+    data->mcb_mpool_index = (mca_coll_base_mpool_index_t**) (data + 1);
+    /* Setup the first pointer in #2 to point to the first instance in
+       #3 */
+    data->mcb_mpool_index[0] = (mca_coll_base_mpool_index_t*)
+        (((char*) data->mcb_mpool_index) +
+         (mca_coll_sm_component.sm_bootstrap_num_segments * 
+          sizeof(mca_coll_base_mpool_index_t *)));
+    /* Now setup the rest of the pointers in #2 to point to their
+       corresponding instances in #3 */
+    for (i = 1; i < mca_coll_sm_component.sm_bootstrap_num_segments; ++i) {
+        data->mcb_mpool_index[i] = data->mcb_mpool_index[i - 1] + 1;
+    }
+
+    comm->c_coll_selected_data = data;
+
+    /* Bootstrap this communicator; set us up with the global mpool,
+       etc. */
+
+    if (OMPI_SUCCESS != bootstrap_comm(comm)) {
+        return NULL;
+    }
+
+    /* Once the communicator is bootstrapped, setup the
+       mca_coll_base_mpool_index_t pointers to point to the
+       appropriate places in the mpool. */
+
+    control_size = ompi_comm_size(comm) * c->sm_control_size;
+    data_size = ompi_comm_size(comm) * c->sm_fragment_size;
+    total_size = (control_size + data_size) * 2;
+    for (i = 0; i < data->mcb_mpool_num_segments; ++i) {
+        data->mcb_mpool_index[i]->mcbmi_control_fan_in = 
+            data->mcb_mpool_base + data->mcb_mpool_offset +
+            (i * total_size);
+        data->mcb_mpool_index[i]->mcbmi_control_fan_out =
+            data->mcb_mpool_index[i]->mcbmi_control_fan_in + control_size;
+        data->mcb_mpool_index[i]->mcbmi_data_fan_in =
+            data->mcb_mpool_index[i]->mcbmi_control_fan_out + control_size;
+        data->mcb_mpool_index[i]->mcbmi_data_fan_out =
+            data->mcb_mpool_index[i]->mcbmi_data_fan_in + data_size;
+    }
+
+    /* All done */
 
     return &module;
 }
@@ -152,9 +261,310 @@ mca_coll_sm_module_init(struct ompi_communicator_t *comm)
 /*
  * Finalize module on the communicator
  */
-int mca_coll_sm_module_finalize(struct ompi_communicator_t *comm)
+static int sm_module_finalize(struct ompi_communicator_t *comm)
 {
-    /* JMS */
+    mca_coll_base_comm_t *data;
+
+    /* Free the space in the data mpool and the data hanging off the
+       communicator */
+
+    data = comm->c_coll_selected_data;
+    if (NULL != data) {
+        /* If this was the process that allocated the space in the
+           data mpool, then this is the process that frees it */
+
+        if (NULL != data->mcb_data_mpool_malloc_addr) {
+            mca_coll_sm_component.sm_data_mpool->mpool_free(mca_coll_sm_component.sm_data_mpool,
+                                                       data->mcb_data_mpool_malloc_addr, NULL);
+        }
+
+        /* Now free the data hanging off the communicator */
+
+        free(data);
+    }
+
+    return OMPI_SUCCESS;
+}
+
+
+static bool have_local_peers(ompi_proc_t **procs, size_t size)
+{
+    size_t i;
+
+    for (i = 0; i < size; ++i) {
+        if (0 == (procs[i]->proc_flags & OMPI_PROC_FLAG_LOCAL)) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+
+static int bootstrap_init(void)
+{
+    int i;
+    size_t size;
+    char *fullpath;
+    mca_common_sm_mmap_t *meta;
+    mca_coll_sm_bootstrap_header_extension_t *bshe;
+
+    /* Create/open the sm coll bootstrap mmap.  Make it have enough
+       space for the top-level control structure and
+       sm_bootstrap_num_segments per-communicator setup struct's
+       (i.e., enough for sm_bootstrap_num_segments communicators to
+       simultaneously set themselves up)  */
+
+    if (NULL == mca_coll_sm_component.sm_bootstrap_filename) {
+        return OMPI_ERROR;
+    }
+    orte_proc_info();
+    asprintf(&fullpath, "%s/%s", orte_process_info.job_session_dir,
+             mca_coll_sm_component.sm_bootstrap_filename);
+    if (NULL == fullpath) {
+        return OMPI_ERR_OUT_OF_RESOURCE;
+    }
+    size = 
+        sizeof(mca_coll_sm_bootstrap_header_extension_t) +
+        (mca_coll_sm_component.sm_bootstrap_num_segments *
+         sizeof(mca_coll_sm_bootstrap_comm_setup_t)) +
+        (sizeof(uint32_t) * mca_coll_sm_component.sm_bootstrap_num_segments);
+
+    mca_coll_sm_component.sm_bootstrap_meta = meta =
+        mca_common_sm_mmap_init(size, fullpath,
+                                sizeof(mca_coll_sm_bootstrap_header_extension_t),
+                                8);
+    if (NULL == meta) {
+        return OMPI_ERR_OUT_OF_RESOURCE;
+    }
+    free(fullpath);
+
+    /* set the pointer to the bootstrap control structure */
+    bshe = (mca_coll_sm_bootstrap_header_extension_t *) meta->map_seg;
+
+    /* Lock the bootstrap control structure.  If it's not already
+       initialized, then we're the first one in and we setup the data
+       structures */
+
+    opal_atomic_lock(&bshe->super.seg_lock);
+    opal_atomic_wmb();
+    if (!bshe->super.seg_inited) {
+        bshe->smbhe_num_segments = 
+            mca_coll_sm_component.sm_bootstrap_num_segments;
+        bshe->smbhe_segments = (mca_coll_sm_bootstrap_comm_setup_t *)
+            (((char *) bshe) + 
+             sizeof(mca_coll_sm_bootstrap_header_extension_t) +
+             (sizeof(uint32_t) * 
+              mca_coll_sm_component.sm_bootstrap_num_segments));
+        bshe->smbhe_cids = (uint32_t *)
+            (((char *) bshe) + sizeof(*bshe));
+        for (i = 0; i < bshe->smbhe_num_segments; ++i) {
+            bshe->smbhe_cids[i] = INT_MAX;
+        }
+
+        bshe->super.seg_inited = true;
+    }
+    opal_atomic_unlock(&bshe->super.seg_lock);
+
+    /* All done */
+
+    bootstrap_inited = true;
+    return OMPI_SUCCESS;
+}
+
+
+static int bootstrap_comm(ompi_communicator_t *comm)
+{
+    int i, empty_index, err;
+    bool found;
+    mca_coll_sm_component_t *c = &mca_coll_sm_component;
+    mca_coll_sm_bootstrap_header_extension_t *bshe;
+    mca_coll_sm_bootstrap_comm_setup_t *bscs;
+    mca_coll_base_comm_t *data = comm->c_coll_selected_data;
+
+    /* Is our CID in the CIDs array?  If not, loop until we can find
+       an open slot in the array to use in the bootstrap to setup our
+       communicator. */
+
+    bshe = (mca_coll_sm_bootstrap_header_extension_t *) 
+        c->sm_bootstrap_meta->map_seg;
+    bscs = bshe->smbhe_segments;
+    opal_atomic_lock(&bshe->super.seg_lock);
+    while (1) {
+        opal_atomic_wmb();
+        found = false;
+        empty_index = -1;
+        for (i = 0; i < bshe->smbhe_num_segments; ++i) {
+            if (comm->c_contextid == bshe->smbhe_cids[i]) {
+                found = true;
+                break;
+            } else if (INT_MAX == bshe->smbhe_cids[i] && -1 == empty_index) {
+                empty_index = i;
+            }
+        }
+
+        /* Did we find our CID? */
+
+        if (found) {
+            break;
+        }
+
+        /* Nope.  Did we find an empty slot?  If so, initialize that
+           slot and its corresponding segment for our CID.  Get an
+           mpool allocation big enough to handle all the shared memory
+           collective stuff. */
+
+        else if (-1 != empty_index) {
+            char *tmp;
+            size_t size;
+
+            i = empty_index;
+            bshe->smbhe_cids[i] = comm->c_contextid;
+
+            bscs[i].smbcs_communicator_num_segments = 
+                c->sm_communicator_num_segments;
+            bscs[i].smbcs_count = ompi_comm_size(comm);
+
+            /* Calculate how much space we need in the data mpool.
+               There are several values to add (one of these for each
+               segment):
+
+               - size of the control data:
+                   - fan-in data (num_procs * control size)
+                   - fan-out data (num_procs * control size)
+               - size of message data
+                   - fan-in data (num_procs * (frag_size rounded up to
+                     control size))
+                   - fan-out data (num_procs * (frag_size rounded up
+                     to control size))
+
+               So it's:
+
+               num_segs * ((num_procs * cachline * 2) + (num_procs *
+                                                         frag * 2))
+
+               Which reduces to:
+
+               num_segs * num_procs * 2 * (control + frag)
+            */
+
+            size = c->sm_communicator_num_segments * ompi_comm_size(comm) * 2 *
+                (c->sm_control_size + c->sm_fragment_size);
+
+            data->mcb_data_mpool_malloc_addr = tmp =
+                c->sm_data_mpool->mpool_alloc(c->sm_data_mpool, size, 
+                                         sizeof(double), NULL);
+            if (NULL == tmp) {
+                /* Cleanup before returning; allow other processes in
+                   this communicator to learn of the failure.  Note
+                   that by definition, bscs[i].smbcs_count won't be
+                   zero after the decrement (because there must be >=2
+                   processes in this communicator, or the self coll
+                   component would have been chosen), so we don't need
+                   to do that cleanup. */
+                bscs[i].smbcs_data_mpool_offset = 0;
+                bscs[i].smbcs_communicator_num_segments = 0;
+                --bscs[i].smbcs_count;
+                opal_atomic_unlock(&bshe->super.seg_lock);
+                return OMPI_ERR_OUT_OF_RESOURCE;
+            }
+
+            /* Zero it out.  We really only need to zero out the
+               control structures, but zeroing out the entire this is
+               easy and "good enough". */
+
+            memset(tmp, 0, size);
+
+            /* Calculate the offset and put it in the bootstrap
+               area */
+
+            bscs[i].smbcs_data_mpool_offset = (size_t) 
+                (tmp - ((char *) c->sm_data_mpool->mpool_base(c->sm_data_mpool)));
+
+            /* Now setup memory affinity such that various control
+               data pages are local to their target processors, if
+               supported. */
+
+            /* JMS continue here */
+
+            break;
+        }
+
+        /* Bad luck all around -- we didn't find our CID in the array
+           and there were no empty slots.  So give up the lock and let
+           some other processes / threads in there to try to free up
+           some slots, and then try again once we have reacquired the
+           lock. */
+
+        else {
+            opal_atomic_unlock(&bshe->super.seg_lock);
+#ifdef HAVE_SCHED_YIELD
+            sched_yield();
+#endif
+            opal_atomic_lock(&bshe->super.seg_lock);
+        }
+    }
+
+    /* Check to see if there was an error while allocating the shared
+       memory */
+    if (0 == bscs[i].smbcs_communicator_num_segments) {
+        err = OMPI_ERR_OUT_OF_RESOURCE;
+    }
+
+    /* Look at the comm_setup_t section (in the data segment of the
+       bootstrap) and fill in the values on our communicator */
+
+    else {
+        err = OMPI_SUCCESS;
+        data->mcb_mpool_base = c->sm_data_mpool->mpool_base(c->sm_data_mpool);
+        data->mcb_mpool_offset = bscs[i].smbcs_data_mpool_offset;
+        data->mcb_mpool_area = data->mcb_mpool_base + data->mcb_mpool_offset;
+        data->mcb_mpool_num_segments = bscs[i].smbcs_communicator_num_segments;
+        data->mcb_operation_count = 0;
+    }
+
+    /* If the count is now zero, then we're finished with this section
+       in the bootstrap segment, and we should release it for others
+       to use */
+
+    --bscs[i].smbcs_count;
+    if (0 == bscs[i].smbcs_count) {
+        bscs[i].smbcs_data_mpool_offset = 0;
+        bshe->smbhe_cids[i] = INT_MAX;
+    }
+
+    /* All done */
+    
+    opal_atomic_unlock(&bshe->super.seg_lock);
+    return err;
+}
+
+
+/*
+ * This function is not static and has a prefix-rule-enabled name
+ * because it gets called from the component.  This is only called
+ * once -- no need for reference counting or thread protection.
+ */
+int mca_coll_sm_bootstrap_finalize(void)
+{
+    mca_common_sm_mmap_t *meta;
+
+    if (bootstrap_inited) {
+        meta = mca_coll_sm_component.sm_bootstrap_meta;
+
+        /* Free the area in the mpool that we were using */
+        if (mca_coll_sm_component.sm_data_mpool_created) {
+            /* JMS: there does not yet seem to be any opposite to
+               mca_mpool_base_module_create()... */
+        }
+
+        /* Free the entire bootstrap area (no need to zero out
+           anything in here -- all data structures are referencing
+           within the bootstrap area, so the one top-level unmap does
+           it all) */
+
+        munmap(meta->map_seg, meta->map_size);
+    }
 
     return OMPI_SUCCESS;
 }

ompi/mca/coll/sm/help-coll-sm.txt

@@ -0,0 +1,31 @@
+# -*- text -*-
+#
+# Copyright (c) 2004-2005 The Trustees of Indiana University.
+#                         All rights reserved.
+# Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
+#                         All rights reserved.
+# Copyright (c) 2004-2005 High Performance Computing Center Stuttgart, 
+#                         University of Stuttgart.  All rights reserved.
+# Copyright (c) 2004-2005 The Regents of the University of California.
+#                         All rights reserved.
+# $COPYRIGHT$
+# 
+# Additional copyrights may follow
+# 
+# $HEADER$
+#
+# This is the US/English general help file for Open RTE's orterun.
+#
+[tree-degree-larger-than-control]
+The specified shared memory collective tree degree
+(coll_sm_tree_degree = %d) is too large.  It must be less than or
+equal to the control size (coll_sm_control_size = %d).
+
+Automatically adjusting the tree degree to be equal to the control
+size and continuing...
+[tree-degree-larger-than-255]
+The specified shared memory collective tree degree
+(coll_sm_tree_degree = %d) is too large.  It must be less than or
+equal to 255.
+
+Automatically adjusting the tree degree to be 255 and continuing...