More openib cleanup.. still note ready for public consumption ;-)
This commit was SVN r6565.
Этот коммит содержится в:
родитель
db4d993228
Коммит
946402b980
@ -17,8 +17,64 @@
|
||||
#ifndef MCA_BTL_BASE_ERROR_H
|
||||
#define MCA_BTL_BASE_ERROR_H
|
||||
|
||||
#define BTL_ERROR(fmt, args...) { \
|
||||
|
||||
#if defined(ACCEPT_C99) && __STDC_VERSION__ >= 199901L
|
||||
# define BTL_ERROR(fmt, ...) { \
|
||||
opal_output(0, "[%s:%d:%d] my_name: [%lu,%lu,%lu] " fmt, __FILE__, __LINE__, __func__, \
|
||||
ORTE_NAME_ARGS(orte_process_info.my_name), __VA_ARGS__); \
|
||||
}
|
||||
#else
|
||||
# if defined(__GNUC__) && !defined(__STDC__)
|
||||
#define BTL_ERROR(fmt, args...) { \
|
||||
opal_output(0, "[%s:%d:%d] my_name: [%lu,%lu,%lu]" fmt, __FILE__, __LINE__, __func__,\
|
||||
ORTE_NAME_ARGS(orte_process_info.my_name), ##args); \
|
||||
}
|
||||
#else
|
||||
static inline void BTL_ERROR(char *fmt, ... )
|
||||
{
|
||||
va_list list;
|
||||
va_start(list, fmt);
|
||||
opal_output(0, "[%s:%d:%d] my_name: [%lu,%lu,%lu]", fmt, __FILE__, __LINE__, __func__,
|
||||
ORTE_NAME_ARGS(orte_process_info.my_name), list);
|
||||
va_end(list);
|
||||
}
|
||||
#define BTL_ERROR printf
|
||||
#endif
|
||||
#endif
|
||||
#ifdef BTL_DEBUG_OUT
|
||||
#if defined(ACCEPT_C99) && __STDC_VERSION__ >= 199901L
|
||||
# define BTL_DEBUG_OUT(fmt, ...) { \
|
||||
opal_output(0, "[%s:%d:%d " fmt, __FILE__, __LINE__, __func__, __VA_ARGS__); \
|
||||
}
|
||||
#else
|
||||
# if defined(__GNUC__) && !defined(__STDC__)
|
||||
#define BTL_DEBUG_OUT(fmt, args...) { \
|
||||
opal_output(0, "[%s:%d:%d " fmt, __FILE__, __LINE__, __func__, ##args); \
|
||||
}
|
||||
}
|
||||
#else
|
||||
static inline void BTL_DEBUG_OUT(char *fmt, ... )
|
||||
{
|
||||
va_list list;
|
||||
va_start(list, fmt);
|
||||
opal_output(0, "[%s:%d:%d ", fmt, __FILE__, __LINE__, __func__, list);
|
||||
va_end(list);
|
||||
}
|
||||
#define BTL_DEBUG_OUT printf
|
||||
#endif
|
||||
#endif
|
||||
#else
|
||||
#if defined(ACCEPT_C99) && __STDC_VERSION__ >= 199901L
|
||||
# define BTL_DEBUG_OUT(fmt, ...)
|
||||
#else
|
||||
# if defined(__GNUC__) && !defined(__STDC__)
|
||||
#define BTL_DEBUG_OUT(fmt, args...)
|
||||
#else
|
||||
static inline void BTL_DEBUG_OUT(char *fmt, ... )
|
||||
{
|
||||
}
|
||||
#define BTL_DEBUG_OUT printf
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
#endif
|
||||
|
@ -63,6 +63,11 @@ mca_btl_openib_module_t mca_btl_openib_module = {
|
||||
|
||||
|
||||
|
||||
/*
|
||||
* add a proc to this btl module
|
||||
* creates an endpoint that is setup on the
|
||||
* first send to the endpoint
|
||||
*/
|
||||
int mca_btl_openib_add_procs(
|
||||
struct mca_btl_base_module_t* btl,
|
||||
size_t nprocs,
|
||||
@ -117,13 +122,17 @@ int mca_btl_openib_add_procs(
|
||||
return OMPI_SUCCESS;
|
||||
}
|
||||
|
||||
|
||||
/*
|
||||
* delete the proc as reachable from this btl module
|
||||
*/
|
||||
int mca_btl_openib_del_procs(struct mca_btl_base_module_t* btl,
|
||||
size_t nprocs,
|
||||
struct ompi_proc_t **procs,
|
||||
struct mca_btl_base_endpoint_t ** peers)
|
||||
{
|
||||
/* Stub */
|
||||
DEBUG_OUT("Stub\n");
|
||||
/* TODO */
|
||||
BTL_DEBUG_OUT("Stub\n");
|
||||
return OMPI_SUCCESS;
|
||||
}
|
||||
|
||||
@ -154,6 +163,9 @@ int mca_btl_openib_register(
|
||||
*
|
||||
* @param btl (IN) BTL module
|
||||
* @param size (IN) Request segment size.
|
||||
*
|
||||
* When allocating a segment we pull a pre-alllocated segment
|
||||
* from one of two free lists, an eager list and a max list
|
||||
*/
|
||||
mca_btl_base_descriptor_t* mca_btl_openib_alloc(
|
||||
struct mca_btl_base_module_t* btl,
|
||||
@ -185,6 +197,8 @@ mca_btl_base_descriptor_t* mca_btl_openib_alloc(
|
||||
/**
|
||||
* Return a segment
|
||||
*
|
||||
* Return the segment to the appropriate
|
||||
* preallocated segment list
|
||||
*/
|
||||
int mca_btl_openib_free(
|
||||
struct mca_btl_base_module_t* btl,
|
||||
@ -210,11 +224,27 @@ int mca_btl_openib_free(
|
||||
|
||||
|
||||
/**
|
||||
* Pack data and return a descriptor that can be
|
||||
* register user buffer or pack
|
||||
* data into pre-registered buffer and return a
|
||||
* descriptor that can be
|
||||
* used for send/put.
|
||||
*
|
||||
* @param btl (IN) BTL module
|
||||
* @param peer (IN) BTL peer addressing
|
||||
*
|
||||
* prepare source's behavior depends on the following:
|
||||
* Has a valid memory registration been passed to prepare_src?
|
||||
* if so we attempt to use the pre-registred user-buffer, if the memory registration
|
||||
* is to small (only a portion of the user buffer) then we must reregister the user buffer
|
||||
* Has the user requested the memory to be left pinned?
|
||||
* if so we insert the memory registration into a memory tree for later lookup, we
|
||||
* may also remove a previous registration if a MRU (most recently used) list of
|
||||
* registions is full, this prevents resources from being exhausted.
|
||||
* Is the requested size larger than the btl's max send size?
|
||||
* if so and we aren't asked to leave the registration pinned than we register the memory if
|
||||
* the users buffer is contiguous
|
||||
* Otherwise we choose from two free lists of pre-registered memory in which to pack the data into.
|
||||
*
|
||||
*/
|
||||
mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
struct mca_btl_base_module_t* btl,
|
||||
@ -229,7 +259,7 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
mca_btl_openib_frag_t* frag;
|
||||
mca_mpool_openib_registration_t * openib_reg;
|
||||
struct iovec iov;
|
||||
int32_t iov_count = 1;
|
||||
uint32_t iov_count = 1;
|
||||
size_t max_data = *size;
|
||||
int32_t free_after;
|
||||
int rc;
|
||||
@ -238,13 +268,13 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
openib_btl = (mca_btl_openib_module_t*) btl;
|
||||
openib_reg = (mca_mpool_openib_registration_t*) registration;
|
||||
|
||||
/** if the data fits in the eager limit and we aren't told to pinn then we
|
||||
simply pack, if the data fits in the eager limit and the data is non contiguous
|
||||
then we pack **/
|
||||
|
||||
|
||||
if(NULL != openib_reg && 0 == ompi_convertor_need_buffers(convertor)){
|
||||
bool is_leave_pinned = openib_reg->base_reg.is_leave_pinned;
|
||||
size_t reg_len;
|
||||
|
||||
/* the memory is already pinned and we have contiguous user data */
|
||||
|
||||
MCA_BTL_IB_FRAG_ALLOC_FRAG(btl, frag, rc);
|
||||
if(NULL == frag){
|
||||
return NULL;
|
||||
@ -254,15 +284,16 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
iov.iov_base = NULL;
|
||||
|
||||
ompi_convertor_pack(convertor, &iov, &iov_count, &max_data, &free_after);
|
||||
|
||||
/* first we will try to find this address in the memory tree (from MPI_Alloc_mem) */
|
||||
|
||||
|
||||
frag->segment.seg_len = max_data;
|
||||
frag->segment.seg_addr.pval = iov.iov_base;
|
||||
|
||||
size_t reg_len;
|
||||
|
||||
reg_len = (unsigned char*)openib_reg->base_reg.bound - (unsigned char*)iov.iov_base + 1;
|
||||
if(frag->segment.seg_len > reg_len) {
|
||||
|
||||
/* the pinned region is too small! we have to re-pinn it */
|
||||
|
||||
size_t new_len = openib_reg->base_reg.bound - openib_reg->base_reg.base + 1
|
||||
+ frag->segment.seg_len - reg_len;
|
||||
void * base_addr = openib_reg->base_reg.base;
|
||||
@ -299,12 +330,14 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
}
|
||||
|
||||
OBJ_RETAIN(openib_reg);
|
||||
|
||||
if(is_leave_pinned) {
|
||||
/* we should leave the memory pinned so put the memory on the MRU list */
|
||||
openib_reg->base_reg.is_leave_pinned = is_leave_pinned;
|
||||
opal_list_append(&openib_btl->reg_mru_list, (opal_list_item_t*) openib_reg);
|
||||
}
|
||||
}
|
||||
else if(is_leave_pinned) {
|
||||
} else if(is_leave_pinned) {
|
||||
/* the current memory region is large enough and we should leave the memory pinned */
|
||||
if(NULL == opal_list_remove_item(&openib_btl->reg_mru_list, (opal_list_item_t*) openib_reg)) {
|
||||
BTL_ERROR("error removing item from reg_mru_list");
|
||||
return NULL;
|
||||
@ -333,6 +366,10 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
ompi_convertor_need_buffers(convertor) == 0 &&
|
||||
reserve == 0)
|
||||
{
|
||||
/* The user buffer is contigous and we need to leave the buffer pinned or we are asked to send
|
||||
more than the max send size. Note that the memory was not already pinned because we have
|
||||
no registration information passed to us */
|
||||
|
||||
MCA_BTL_IB_FRAG_ALLOC_FRAG(btl, frag, rc);
|
||||
if(NULL == frag){
|
||||
return NULL;
|
||||
@ -350,14 +387,19 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
|
||||
|
||||
if(mca_btl_openib_component.leave_pinned) {
|
||||
/* so we need to leave pinned which means we must check and see if we
|
||||
have room on the MRU list */
|
||||
|
||||
if(mca_btl_openib_component.reg_mru_len <= openib_btl->reg_mru_list.opal_list_length ) {
|
||||
|
||||
/* we have the maximum number of entries on the MRU list, time to
|
||||
pull something off and make room. */
|
||||
|
||||
mca_mpool_openib_registration_t* old_reg =
|
||||
(mca_mpool_openib_registration_t*)
|
||||
opal_list_remove_last(&openib_btl->reg_mru_list);
|
||||
|
||||
if( NULL == old_reg) {
|
||||
opal_output(0,"%s:%d:%s error removing item from reg_mru_list", __FILE__, __LINE__, __func__);
|
||||
BTL_ERROR("error removing item from reg_mru_list");
|
||||
return NULL;
|
||||
}
|
||||
|
||||
@ -365,7 +407,7 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
rc = mca_mpool_base_remove((void*) old_reg->base_reg.base);
|
||||
|
||||
if(OMPI_SUCCESS != rc) {
|
||||
opal_output(0,"%s:%d:%s error removing memory region from memory pool tree", __FILE__, __LINE__, __func__);
|
||||
BTL_ERROR("error removing memory region from memory pool tree");
|
||||
return NULL;
|
||||
}
|
||||
|
||||
@ -390,7 +432,8 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
|
||||
opal_list_append(&openib_btl->reg_mru_list, (opal_list_item_t*) openib_reg);
|
||||
|
||||
} else {
|
||||
} else {
|
||||
/* we don't need to leave the memory pinned so just register it.. */
|
||||
openib_btl->ib_pool->mpool_register(openib_btl->ib_pool,
|
||||
iov.iov_base,
|
||||
max_data,
|
||||
@ -410,20 +453,23 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
frag->base.des_dst = NULL;
|
||||
frag->base.des_dst_cnt = 0;
|
||||
frag->openib_reg = openib_reg;
|
||||
DEBUG_OUT("frag->sg_entry.lkey = %lu .addr = %llu", frag->sg_entry.lkey, frag->sg_entry.addr);
|
||||
BTL_DEBUG_OUT("frag->sg_entry.lkey = %lu .addr = %llu", frag->sg_entry.lkey, frag->sg_entry.addr);
|
||||
|
||||
|
||||
return &frag->base;
|
||||
|
||||
} else if (max_data+reserve <= btl->btl_eager_limit) {
|
||||
|
||||
/* the data is small enough to fit in the eager frag and
|
||||
either we received no prepinned memory or leave pinned is
|
||||
not set
|
||||
*/
|
||||
MCA_BTL_IB_FRAG_ALLOC_EAGER(btl, frag, rc);
|
||||
if(NULL == frag) {
|
||||
return NULL;
|
||||
}
|
||||
|
||||
iov.iov_len = max_data;
|
||||
iov.iov_base = frag->segment.seg_addr.lval + reserve;
|
||||
iov.iov_base = (unsigned char*) frag->segment.seg_addr.lval + reserve;
|
||||
|
||||
rc = ompi_convertor_pack(convertor, &iov, &iov_count, &max_data, &free_after);
|
||||
*size = max_data;
|
||||
@ -484,6 +530,14 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_src(
|
||||
*
|
||||
* @param btl (IN) BTL module
|
||||
* @param peer (IN) BTL peer addressing
|
||||
* prepare dest's behavior depends on the following:
|
||||
* Has a valid memory registration been passed to prepare_src?
|
||||
* if so we attempt to use the pre-registred user-buffer, if the memory registration
|
||||
* is to small (only a portion of the user buffer) then we must reregister the user buffer
|
||||
* Has the user requested the memory to be left pinned?
|
||||
* if so we insert the memory registration into a memory tree for later lookup, we
|
||||
* may also remove a previous registration if a MRU (most recently used) list of
|
||||
* registions is full, this prevents resources from being exhausted.
|
||||
*/
|
||||
mca_btl_base_descriptor_t* mca_btl_openib_prepare_dst(
|
||||
struct mca_btl_base_module_t* btl,
|
||||
@ -525,13 +579,13 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_dst(
|
||||
|
||||
rc = mca_mpool_base_remove((void*) openib_reg->base_reg.base);
|
||||
if(OMPI_SUCCESS != rc) {
|
||||
opal_output(0,"%s:%d:%s error removing memory region from memory pool tree", __FILE__, __LINE__, __func__);
|
||||
BTL_ERROR("error removing memory region from memory pool tree");
|
||||
return NULL;
|
||||
}
|
||||
|
||||
if(is_leave_pinned) {
|
||||
if(NULL == opal_list_remove_item(&openib_btl->reg_mru_list, (opal_list_item_t*) openib_reg)) {
|
||||
opal_output(0,"%s:%d:%s error removing item from reg_mru_list", __FILE__, __LINE__, __func__);
|
||||
BTL_ERROR("error removing item from reg_mru_list");
|
||||
return NULL;
|
||||
}
|
||||
}
|
||||
@ -550,7 +604,7 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_dst(
|
||||
(mca_mpool_base_registration_t*) openib_reg);
|
||||
|
||||
if(OMPI_SUCCESS != rc) {
|
||||
opal_output(0,"%s:%d:%s error inserting memory region into memory pool tree", __FILE__, __LINE__, __func__);
|
||||
BTL_ERROR("error inserting memory region into memory pool tree");
|
||||
return NULL;
|
||||
}
|
||||
OBJ_RETAIN(openib_reg);
|
||||
@ -563,7 +617,7 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_dst(
|
||||
}
|
||||
else if(is_leave_pinned){
|
||||
if(NULL == opal_list_remove_item(&openib_btl->reg_mru_list, (opal_list_item_t*) openib_reg)) {
|
||||
opal_output(0,"%s:%d:%s error removing item from reg_mru_list", __FILE__, __LINE__, __func__);
|
||||
BTL_ERROR("error removing item from reg_mru_list");
|
||||
return NULL;
|
||||
}
|
||||
opal_list_append(&openib_btl->reg_mru_list, (opal_list_item_t*) openib_reg);
|
||||
@ -581,13 +635,13 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_dst(
|
||||
opal_list_remove_last(&openib_btl->reg_mru_list);
|
||||
|
||||
if( NULL == old_reg) {
|
||||
opal_output(0,"%s:%d:%s error removing item from reg_mru_list", __FILE__, __LINE__, __func__);
|
||||
BTL_ERROR("error removing item from reg_mru_list");
|
||||
return NULL;
|
||||
}
|
||||
|
||||
rc = mca_mpool_base_remove((void*) old_reg->base_reg.base);
|
||||
if(OMPI_SUCCESS !=rc ) {
|
||||
opal_output(0,"%s:%d:%s error removing memory region from memory pool tree", __FILE__, __LINE__, __func__);
|
||||
BTL_ERROR("error removing memory region from memory pool tree");
|
||||
return NULL;
|
||||
}
|
||||
|
||||
@ -607,7 +661,7 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_dst(
|
||||
(void*) (&openib_btl->super),
|
||||
(mca_mpool_base_registration_t*) openib_reg);
|
||||
if(OMPI_SUCCESS != rc){
|
||||
opal_output(0,"%s:%d:%s error inserting memory region into memory pool", __FILE__, __LINE__, __func__);
|
||||
BTL_ERROR("error inserting memory region into memory pool");
|
||||
return NULL;
|
||||
}
|
||||
|
||||
@ -638,7 +692,7 @@ mca_btl_base_descriptor_t* mca_btl_openib_prepare_dst(
|
||||
frag->base.des_src = NULL;
|
||||
frag->base.des_src_cnt = 0;
|
||||
frag->openib_reg = openib_reg;
|
||||
DEBUG_OUT("frag->sg_entry.lkey = %lu .addr = %llu frag->segment.seg_key.key32[0] = %lu" , frag->sg_entry.lkey, frag->sg_entry.addr, frag->segment.seg_key.key32[0]);
|
||||
BTL_DEBUG_OUT("frag->sg_entry.lkey = %lu .addr = %llu frag->segment.seg_key.key32[0] = %lu" , frag->sg_entry.lkey, frag->sg_entry.addr, frag->segment.seg_key.key32[0]);
|
||||
|
||||
return &frag->base;
|
||||
|
||||
@ -648,7 +702,8 @@ int mca_btl_openib_finalize(struct mca_btl_base_module_t* btl)
|
||||
{
|
||||
mca_btl_openib_module_t* openib_btl;
|
||||
openib_btl = (mca_btl_openib_module_t*) btl;
|
||||
|
||||
|
||||
#if 0
|
||||
if(openib_btl->send_free_eager.fl_num_allocated !=
|
||||
openib_btl->send_free_eager.super.opal_list_length){
|
||||
opal_output(0, "btl ib send_free_eager frags: %d allocated %d returned \n",
|
||||
@ -681,15 +736,13 @@ int mca_btl_openib_finalize(struct mca_btl_base_module_t* btl)
|
||||
openib_btl->recv_free_max.fl_num_allocated,
|
||||
openib_btl->recv_free_max.super.opal_list_length);
|
||||
}
|
||||
#endif
|
||||
|
||||
return OMPI_SUCCESS;
|
||||
}
|
||||
|
||||
/*
|
||||
* Initiate a send. If this is the first fragment, use the fragment
|
||||
* descriptor allocated with the send requests, otherwise obtain
|
||||
* one from the free list. Initialize the fragment and foward
|
||||
* on to the peer.
|
||||
* Initiate a send.
|
||||
*/
|
||||
|
||||
int mca_btl_openib_send(
|
||||
@ -721,82 +774,34 @@ int mca_btl_openib_put( mca_btl_base_module_t* btl,
|
||||
struct ibv_send_wr* bad_wr;
|
||||
mca_btl_openib_frag_t* frag = (mca_btl_openib_frag_t*) descriptor;
|
||||
frag->endpoint = endpoint;
|
||||
frag->sr_desc.opcode = IBV_WR_RDMA_WRITE;
|
||||
frag->sr_desc.send_flags = IBV_SEND_SIGNALED;
|
||||
frag->sr_desc.wr.rdma.remote_addr = (uintptr_t) frag->base.des_dst->seg_addr.pval;
|
||||
frag->sr_desc.wr.rdma.rkey = frag->base.des_dst->seg_key.key32[0];
|
||||
frag->wr_desc.sr_desc.opcode = IBV_WR_RDMA_WRITE;
|
||||
frag->wr_desc.sr_desc.send_flags = IBV_SEND_SIGNALED;
|
||||
frag->wr_desc.sr_desc.wr.rdma.remote_addr = (uintptr_t) frag->base.des_dst->seg_addr.pval;
|
||||
frag->wr_desc.sr_desc.wr.rdma.rkey = frag->base.des_dst->seg_key.key32[0];
|
||||
frag->sg_entry.addr = (uintptr_t) frag->base.des_src->seg_addr.pval;
|
||||
frag->sg_entry.length = frag->base.des_src->seg_len;
|
||||
|
||||
DEBUG_OUT("frag->sr_desc.wr.rdma.remote_addr = %llu .rkey = %lu frag->sg_entry.addr = %llu .length = %lu"
|
||||
, frag->sr_desc.wr.rdma.remote_addr
|
||||
, frag->sr_desc.wr.rdma.rkey
|
||||
, frag->sg_entry.addr
|
||||
, frag->sg_entry.length);
|
||||
BTL_DEBUG_OUT("frag->wr_desc.sr_desc.wr.rdma.remote_addr = %llu .rkey = %lu frag->sg_entry.addr = %llu .length = %lu"
|
||||
, frag->wr_desc.sr_desc.wr.rdma.remote_addr
|
||||
, frag->wr_desc.sr_desc.wr.rdma.rkey
|
||||
, frag->sg_entry.addr
|
||||
, frag->sg_entry.length);
|
||||
|
||||
if(ibv_post_send(endpoint->lcl_qp_low,
|
||||
&frag->sr_desc,
|
||||
&frag->wr_desc.sr_desc,
|
||||
&bad_wr)){
|
||||
opal_output(0, "%s: error posting send request errno says %s\n", __func__, strerror(errno));
|
||||
BTL_ERROR("error posting send request errno says %s", strerror(errno));
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
|
||||
mca_btl_openib_endpoint_post_rr(endpoint, 1);
|
||||
|
||||
|
||||
MCA_BTL_OPENIB_ENDPOINT_POST_RR_HIGH(endpoint, 1);
|
||||
MCA_BTL_OPENIB_ENDPOINT_POST_RR_LOW(endpoint, 1);
|
||||
|
||||
return OMPI_SUCCESS;
|
||||
|
||||
}
|
||||
|
||||
|
||||
|
||||
/*
|
||||
* Asynchronous event handler to detect unforseen
|
||||
* events. Usually, such events are catastrophic.
|
||||
* Should have a robust mechanism to handle these
|
||||
* events and abort the OMPI application if necessary.
|
||||
*
|
||||
*/
|
||||
/* static void async_event_handler(VAPI_hca_hndl_t hca_hndl, */
|
||||
/* VAPI_event_record_t * event_p, */
|
||||
/* void *priv_data) */
|
||||
/* { */
|
||||
/* switch (event_p->type) { */
|
||||
/* case VAPI_QP_PATH_MIGRATED: */
|
||||
/* case VAPI_EEC_PATH_MIGRATED: */
|
||||
/* case VAPI_QP_COMM_ESTABLISHED: */
|
||||
/* case VAPI_EEC_COMM_ESTABLISHED: */
|
||||
/* case VAPI_SEND_QUEUE_DRAINED: */
|
||||
/* case VAPI_PORT_ACTIVE: */
|
||||
/* { */
|
||||
/* DEBUG_OUT("Got an asynchronous event: %s\n", */
|
||||
/* VAPI_event_record_sym(event_p->type)); */
|
||||
/* break; */
|
||||
/* } */
|
||||
/* case VAPI_CQ_ERROR: */
|
||||
/* case VAPI_LOCAL_WQ_INV_REQUEST_ERROR: */
|
||||
/* case VAPI_LOCAL_WQ_ACCESS_VIOL_ERROR: */
|
||||
/* case VAPI_LOCAL_WQ_CATASTROPHIC_ERROR: */
|
||||
/* case VAPI_PATH_MIG_REQ_ERROR: */
|
||||
/* case VAPI_LOCAL_EEC_CATASTROPHIC_ERROR: */
|
||||
/* case VAPI_LOCAL_CATASTROPHIC_ERROR: */
|
||||
/* case VAPI_PORT_ERROR: */
|
||||
/* { */
|
||||
/* opal_output(0, "Got an asynchronous event: %s (%s)", */
|
||||
/* VAPI_event_record_sym(event_p->type), */
|
||||
/* VAPI_event_syndrome_sym(event_p-> */
|
||||
/* syndrome)); */
|
||||
/* break; */
|
||||
/* } */
|
||||
/* default: */
|
||||
/* opal_output(0, "Warning!! Got an undefined " */
|
||||
/* "asynchronous event\n"); */
|
||||
/* } */
|
||||
|
||||
/* } */
|
||||
|
||||
|
||||
|
||||
|
||||
int mca_btl_openib_module_init(mca_btl_openib_module_t *openib_btl)
|
||||
{
|
||||
|
||||
@ -809,36 +814,30 @@ int mca_btl_openib_module_init(mca_btl_openib_module_t *openib_btl)
|
||||
|
||||
|
||||
if(NULL == openib_btl->ib_pd) {
|
||||
opal_output(0, "%s: error allocating pd for %s errno says %s\n",
|
||||
__func__,
|
||||
ibv_get_device_name(openib_btl->ib_dev),
|
||||
strerror(errno));
|
||||
BTL_ERROR("error allocating pd for %s errno says %s\n",
|
||||
ibv_get_device_name(openib_btl->ib_dev),
|
||||
strerror(errno));
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
|
||||
|
||||
/* Create the low and high priority queue pairs */
|
||||
openib_btl->ib_cq_low = ibv_create_cq(ctx, mca_btl_openib_component.ib_cq_size, NULL);
|
||||
|
||||
if(NULL == openib_btl->ib_cq_low) {
|
||||
opal_output(0, "%s: error creating low priority cq for %s errno says %s\n",
|
||||
__func__,
|
||||
ibv_get_device_name(openib_btl->ib_dev),
|
||||
strerror(errno));
|
||||
BTL_ERROR("error creating low priority cq for %s errno says %s\n",
|
||||
ibv_get_device_name(openib_btl->ib_dev),
|
||||
strerror(errno));
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
|
||||
openib_btl->ib_cq_high = ibv_create_cq(ctx, mca_btl_openib_component.ib_cq_size, NULL);
|
||||
|
||||
if(NULL == openib_btl->ib_cq_high) {
|
||||
opal_output(0, "%s: error creating high priority cq for %s errno says %s\n",
|
||||
__func__,
|
||||
ibv_get_device_name(openib_btl->ib_dev),
|
||||
strerror(errno));
|
||||
BTL_ERROR("error creating high priority cq for %s errno says %s\n",
|
||||
ibv_get_device_name(openib_btl->ib_dev),
|
||||
strerror(errno));
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
|
||||
/* TODO: EVAPI_set_qsync_event_handler? */
|
||||
|
||||
|
||||
|
||||
return OMPI_SUCCESS;
|
||||
}
|
||||
|
@ -32,7 +32,7 @@
|
||||
#include "mca/btl/btl.h"
|
||||
#include "opal/util/output.h"
|
||||
#include "mca/mpool/mpool.h"
|
||||
#include "btl_openib_error.h"
|
||||
#include "mca/btl/base/btl_base_error.h"
|
||||
|
||||
#include "mca/btl/btl.h"
|
||||
#include "mca/btl/base/base.h"
|
||||
|
@ -507,10 +507,9 @@ int mca_btl_openib_component_progress()
|
||||
for(i = 0; i < mca_btl_openib_component.ib_num_btls; i++) {
|
||||
|
||||
struct ibv_wc wc;
|
||||
memset(&wc, 0, sizeof(struct ibv_wc));
|
||||
|
||||
mca_btl_openib_module_t* openib_btl = &mca_btl_openib_component.openib_btls[i];
|
||||
|
||||
memset(&wc, 0, sizeof(struct ibv_wc));
|
||||
|
||||
/* we have two completion queues, one for "high" priority and one for "low".
|
||||
* we will check the high priority and process them until there are none left.
|
||||
* note that low priority messages are only processed one per progress call.
|
||||
@ -522,12 +521,12 @@ int mca_btl_openib_component_progress()
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
else if(wc.status != IBV_WC_SUCCESS) {
|
||||
BTL_ERROR("error polling CQ with status %d for wr_id %d\n",
|
||||
BTL_ERROR("error polling CQ with status %d for wr_id %llu\n",
|
||||
wc.status, wc.wr_id);
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
else if(1 == ne) {
|
||||
DEBUG_OUT("completion queue event says opcode is %d\n", wc.opcode);
|
||||
BTL_DEBUG_OUT("completion queue event says opcode is %d\n", wc.opcode);
|
||||
|
||||
/* Handle work completions */
|
||||
switch(wc.opcode) {
|
||||
@ -538,7 +537,7 @@ int mca_btl_openib_component_progress()
|
||||
case IBV_WC_RECV:
|
||||
/* Process a RECV */
|
||||
|
||||
DEBUG_OUT("Got an recv on the completion queue");
|
||||
BTL_DEBUG_OUT("Got an recv on the completion queue");
|
||||
frag = (mca_btl_openib_frag_t*) wc.wr_id;
|
||||
endpoint = (mca_btl_openib_endpoint_t*) frag->endpoint;
|
||||
frag->rc=OMPI_SUCCESS;
|
||||
@ -547,9 +546,6 @@ int mca_btl_openib_component_progress()
|
||||
((unsigned char*) frag->segment.seg_addr.pval - (unsigned char*) frag->hdr);
|
||||
|
||||
|
||||
OPAL_THREAD_ADD32(&endpoint->rr_posted_high, -1);
|
||||
|
||||
mca_btl_openib_endpoint_post_rr(((mca_btl_openib_frag_t*)wc.wr_id)->endpoint, 0);
|
||||
|
||||
/* advance the segment address past the header and subtract from the length..*/
|
||||
openib_btl->ib_reg[frag->hdr->tag].cbfunc(&openib_btl->super,
|
||||
@ -557,6 +553,8 @@ int mca_btl_openib_component_progress()
|
||||
&frag->base,
|
||||
openib_btl->ib_reg[frag->hdr->tag].cbdata);
|
||||
|
||||
OPAL_THREAD_ADD32(&endpoint->rr_posted_high, -1);
|
||||
MCA_BTL_OPENIB_ENDPOINT_POST_RR_HIGH(((mca_btl_openib_frag_t*)wc.wr_id)->endpoint, 0);
|
||||
OMPI_FREE_LIST_RETURN(&(openib_btl->recv_free_eager), (opal_list_item_t*) frag);
|
||||
count++;
|
||||
break;
|
||||
@ -588,8 +586,8 @@ int mca_btl_openib_component_progress()
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
else if(wc.status != IBV_WC_SUCCESS) {
|
||||
BTL_ERROR("error polling CQ with status %d for wr_id %d",
|
||||
wc.status, wc.wr_id);
|
||||
BTL_ERROR("error polling CQ with status %d for wr_id %llu",
|
||||
wc.status, wc.wr_id);
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
else if(1 == ne) {
|
||||
@ -601,7 +599,7 @@ int mca_btl_openib_component_progress()
|
||||
|
||||
case IBV_WC_RECV:
|
||||
/* process a recv completion (this should only occur for a send not an rdma) */
|
||||
DEBUG_OUT( "%s:%d ib recv under redesign\n", __FILE__, __LINE__);
|
||||
BTL_DEBUG_OUT( "%s:%d ib recv under redesign\n", __FILE__, __LINE__);
|
||||
frag = (mca_btl_openib_frag_t*) wc.wr_id;
|
||||
endpoint = (mca_btl_openib_endpoint_t*) frag->endpoint;
|
||||
frag->rc=OMPI_SUCCESS;
|
||||
@ -611,17 +609,15 @@ int mca_btl_openib_component_progress()
|
||||
wc.byte_len-
|
||||
((unsigned char*) frag->segment.seg_addr.pval - (unsigned char*) frag->hdr);
|
||||
|
||||
OPAL_THREAD_ADD32(&endpoint->rr_posted_low, -1);
|
||||
|
||||
mca_btl_openib_endpoint_post_rr(((mca_btl_openib_frag_t*)wc.wr_id)->endpoint, 0);
|
||||
|
||||
openib_btl->ib_reg[frag->hdr->tag].cbfunc(&openib_btl->super,
|
||||
frag->hdr->tag,
|
||||
&frag->base,
|
||||
openib_btl->ib_reg[frag->hdr->tag].cbdata);
|
||||
|
||||
OPAL_THREAD_ADD32(&endpoint->rr_posted_low, -1);
|
||||
MCA_BTL_OPENIB_ENDPOINT_POST_RR_LOW(((mca_btl_openib_frag_t*)wc.wr_id)->endpoint, 0);
|
||||
OMPI_FREE_LIST_RETURN(&(openib_btl->recv_free_max), (opal_list_item_t*) frag);
|
||||
|
||||
count++;
|
||||
break;
|
||||
|
||||
|
@ -58,6 +58,10 @@ int mca_btl_openib_endpoint_qp_init_query(
|
||||
uint32_t port_num
|
||||
);
|
||||
|
||||
|
||||
/*
|
||||
* post a send to the work queue
|
||||
*/
|
||||
static inline int mca_btl_openib_endpoint_post_send(mca_btl_openib_module_t* openib_btl,
|
||||
mca_btl_openib_endpoint_t * endpoint,
|
||||
mca_btl_openib_frag_t * frag)
|
||||
@ -73,7 +77,7 @@ static inline int mca_btl_openib_endpoint_post_send(mca_btl_openib_module_t* ope
|
||||
ib_qp = endpoint->lcl_qp_low;
|
||||
}
|
||||
|
||||
frag->sr_desc.opcode = IBV_WR_SEND;
|
||||
frag->wr_desc.sr_desc.opcode = IBV_WR_SEND;
|
||||
|
||||
frag->sg_entry.length =
|
||||
frag->segment.seg_len +
|
||||
@ -81,18 +85,19 @@ static inline int mca_btl_openib_endpoint_post_send(mca_btl_openib_module_t* ope
|
||||
|
||||
|
||||
if(frag->sg_entry.length <= openib_btl->ib_inline_max) {
|
||||
/* frag->sr_desc.send_flags |= IBV_SEND_INLINE; */
|
||||
frag->wr_desc.sr_desc.send_flags |= IBV_SEND_INLINE;
|
||||
}
|
||||
|
||||
|
||||
if(ibv_post_send(ib_qp,
|
||||
&frag->sr_desc,
|
||||
&frag->wr_desc.sr_desc,
|
||||
&bad_wr)) {
|
||||
opal_output(0, "%s: error posting send request errno says %s\n", __func__, strerror(errno));
|
||||
BTL_ERROR("error posting send request errno says %s", strerror(errno));
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
mca_btl_openib_endpoint_post_rr(endpoint, 1);
|
||||
|
||||
MCA_BTL_OPENIB_ENDPOINT_POST_RR_HIGH(endpoint, 1);
|
||||
MCA_BTL_OPENIB_ENDPOINT_POST_RR_LOW(endpoint, 1);
|
||||
|
||||
return OMPI_SUCCESS;
|
||||
}
|
||||
|
||||
@ -200,7 +205,7 @@ static int mca_btl_openib_endpoint_send_connect_req(mca_btl_base_endpoint_t* end
|
||||
mca_btl_openib_endpoint_send_cb, NULL);
|
||||
|
||||
|
||||
DEBUG_OUT("Sending High Priority QP num = %d, Low Priority QP num = %d, LID = %d",
|
||||
BTL_DEBUG_OUT("Sending High Priority QP num = %d, Low Priority QP num = %d, LID = %d",
|
||||
endpoint->lcl_qp_high->qp_num,
|
||||
endpoint->lcl_qp_low->qp_num,
|
||||
endpoint->endpoint_btl->ib_port_attr->lid);
|
||||
@ -256,10 +261,7 @@ static int mca_btl_openib_endpoint_send_connect_ack(mca_btl_base_endpoint_t* end
|
||||
|
||||
/*
|
||||
* Set remote connection info
|
||||
*
|
||||
* XXX: Currently size is unutilized, this shall change
|
||||
* as soon as we add more info to be exchanged at connection
|
||||
* setup.
|
||||
* (from OOB connection)
|
||||
*
|
||||
*/
|
||||
static int mca_btl_openib_endpoint_set_remote_info(mca_btl_base_endpoint_t* endpoint, orte_buffer_t* buffer)
|
||||
@ -294,7 +296,7 @@ static int mca_btl_openib_endpoint_set_remote_info(mca_btl_base_endpoint_t* endp
|
||||
ORTE_ERROR_LOG(rc);
|
||||
return rc;
|
||||
}
|
||||
DEBUG_OUT("Received High Priority QP num = %d, Low Priority QP num %d, LID = %d",
|
||||
BTL_DEBUG_OUT("Received High Priority QP num = %d, Low Priority QP num %d, LID = %d",
|
||||
endpoint->rem_qp_num_high,
|
||||
endpoint->rem_qp_num_low,
|
||||
endpoint->rem_lid);
|
||||
@ -344,7 +346,7 @@ static int mca_btl_openib_endpoint_start_connect(mca_btl_base_endpoint_t* endpoi
|
||||
}
|
||||
endpoint->lcl_psn_low = lrand48() & 0xffffff;
|
||||
|
||||
DEBUG_OUT("Initialized High Priority QP num = %d, Low Priority QP num = %d, LID = %d",
|
||||
BTL_DEBUG_OUT("Initialized High Priority QP num = %d, Low Priority QP num = %d, LID = %d",
|
||||
endpoint->lcl_qp_high->qp_num,
|
||||
endpoint->lcl_qp_low->qp_num,
|
||||
openib_btl->ib_port_attr->lid);
|
||||
@ -375,8 +377,7 @@ static int mca_btl_openib_endpoint_reply_start_connect(mca_btl_openib_endpoint_t
|
||||
openib_btl->ib_cq_high,
|
||||
endpoint->lcl_qp_attr_high,
|
||||
&endpoint->lcl_qp_high))) {
|
||||
opal_output(0, "[%lu,%lu,%lu] %s:%d errcode %d\n",
|
||||
ORTE_NAME_ARGS(orte_process_info.my_name), __FILE__,__LINE__,rc);
|
||||
BTL_ERROR("error creating queue pair, error code %d", rc);
|
||||
return rc;
|
||||
}
|
||||
srand48(getpid() * time(NULL));
|
||||
@ -388,13 +389,12 @@ static int mca_btl_openib_endpoint_reply_start_connect(mca_btl_openib_endpoint_t
|
||||
openib_btl->ib_cq_low,
|
||||
endpoint->lcl_qp_attr_low,
|
||||
&endpoint->lcl_qp_low))) {
|
||||
opal_output(0, "[%lu,%lu,%lu] %s:%d errcode %d\n",
|
||||
ORTE_NAME_ARGS(orte_process_info.my_name), __FILE__,__LINE__,rc);
|
||||
BTL_ERROR("error createing queue pair, error code %d", rc);
|
||||
return rc;
|
||||
}
|
||||
endpoint->lcl_psn_low = lrand48() & 0xffffff;
|
||||
|
||||
DEBUG_OUT("Initialized High Priority QP num = %d, Low Priority QP num = %d, LID = %d",
|
||||
BTL_DEBUG_OUT("Initialized High Priority QP num = %d, Low Priority QP num = %d, LID = %d",
|
||||
endpoint->lcl_qp_high->qp_num,
|
||||
endpoint->lcl_qp_low->qp_num,
|
||||
openib_btl->ib_port_attr->lid);
|
||||
@ -407,15 +407,13 @@ static int mca_btl_openib_endpoint_reply_start_connect(mca_btl_openib_endpoint_t
|
||||
|
||||
rc = mca_btl_openib_endpoint_connect(endpoint);
|
||||
if(rc != OMPI_SUCCESS) {
|
||||
opal_output(0, "[%lu,%lu,%lu] %s:%d errcode %d\n",
|
||||
ORTE_NAME_ARGS(orte_process_info.my_name), __FILE__,__LINE__,rc);
|
||||
BTL_ERROR("error in endpoint connect error code is %d", rc);
|
||||
return rc;
|
||||
}
|
||||
|
||||
/* Send connection info over to remote endpoint */
|
||||
if(OMPI_SUCCESS != (rc = mca_btl_openib_endpoint_send_connect_req(endpoint))) {
|
||||
opal_output(0, "[%lu,%lu,%lu] %s:%d errcode %d\n",
|
||||
ORTE_NAME_ARGS(orte_process_info.my_name), __FILE__,__LINE__,rc);
|
||||
BTL_ERROR("error in endpoint send connect request error code is %d", rc);
|
||||
return rc;
|
||||
}
|
||||
return OMPI_SUCCESS;
|
||||
@ -492,8 +490,7 @@ static void mca_btl_openib_endpoint_recv(
|
||||
|
||||
mca_btl_openib_endpoint_set_remote_info(ib_endpoint, buffer);
|
||||
if(OMPI_SUCCESS != (rc = mca_btl_openib_endpoint_connect(ib_endpoint))) {
|
||||
opal_output(0, "[%lu,%lu,%lu] %s:%d errcode %d\n",
|
||||
ORTE_NAME_ARGS(orte_process_info.my_name), __FILE__,__LINE__,rc);
|
||||
BTL_ERROR("endpoint connect error: %d", rc);
|
||||
break;
|
||||
}
|
||||
|
||||
@ -505,7 +502,6 @@ static void mca_btl_openib_endpoint_recv(
|
||||
break;
|
||||
|
||||
case MCA_BTL_IB_CONNECT_ACK:
|
||||
DEBUG_OUT("Got a connect ack from %d\n", endpoint->vpid);
|
||||
|
||||
mca_btl_openib_endpoint_connected(ib_endpoint);
|
||||
break;
|
||||
@ -514,7 +510,7 @@ static void mca_btl_openib_endpoint_recv(
|
||||
|
||||
break;
|
||||
default :
|
||||
opal_output(0, "Connected -> Connecting not possible.\n");
|
||||
BTL_ERROR("Invalid endpoint state %d", endpoint_state);
|
||||
}
|
||||
|
||||
break;
|
||||
@ -526,6 +522,9 @@ static void mca_btl_openib_endpoint_recv(
|
||||
mca_btl_openib_post_recv();
|
||||
}
|
||||
|
||||
/*
|
||||
* Post the OOB recv (for receiving the peers information)
|
||||
*/
|
||||
void mca_btl_openib_post_recv()
|
||||
{
|
||||
|
||||
@ -556,7 +555,7 @@ int mca_btl_openib_endpoint_send(
|
||||
switch(endpoint->endpoint_state) {
|
||||
case MCA_BTL_IB_CONNECTING:
|
||||
|
||||
DEBUG_OUT("Queing because state is connecting");
|
||||
BTL_DEBUG_OUT("Queing because state is connecting");
|
||||
|
||||
opal_list_append(&endpoint->pending_send_frags,
|
||||
(opal_list_item_t *)frag);
|
||||
@ -566,7 +565,7 @@ int mca_btl_openib_endpoint_send(
|
||||
|
||||
case MCA_BTL_IB_CONNECT_ACK:
|
||||
|
||||
DEBUG_OUT("Queuing because waiting for ack");
|
||||
BTL_DEBUG_OUT("Queuing because waiting for ack");
|
||||
|
||||
opal_list_append(&endpoint->pending_send_frags,
|
||||
(opal_list_item_t *)frag);
|
||||
@ -576,13 +575,10 @@ int mca_btl_openib_endpoint_send(
|
||||
|
||||
case MCA_BTL_IB_CLOSED:
|
||||
|
||||
DEBUG_OUT("Connection to endpoint closed ... connecting ...");
|
||||
|
||||
BTL_DEBUG_OUT("Connection to endpoint closed ... connecting ...");
|
||||
opal_list_append(&endpoint->pending_send_frags,
|
||||
(opal_list_item_t *)frag);
|
||||
|
||||
rc = mca_btl_openib_endpoint_start_connect(endpoint);
|
||||
|
||||
break;
|
||||
|
||||
case MCA_BTL_IB_FAILED:
|
||||
@ -593,15 +589,11 @@ int mca_btl_openib_endpoint_send(
|
||||
case MCA_BTL_IB_CONNECTED:
|
||||
{
|
||||
openib_btl = endpoint->endpoint_btl;
|
||||
|
||||
|
||||
DEBUG_OUT("Send to : %d, len : %d, frag : %p",
|
||||
endpoint->endpoint_proc->proc_guid.vpid,
|
||||
frag->sg_entry.length,
|
||||
frag);
|
||||
|
||||
BTL_DEBUG_OUT("Send to : %d, len : %lu, frag : %llu",
|
||||
endpoint->endpoint_proc->proc_guid.vpid,
|
||||
frag->sg_entry.length,
|
||||
(unsigned long long) frag);
|
||||
rc = mca_btl_openib_endpoint_post_send(openib_btl, endpoint, frag);
|
||||
|
||||
break;
|
||||
}
|
||||
|
||||
@ -613,7 +605,10 @@ int mca_btl_openib_endpoint_send(
|
||||
|
||||
return rc;
|
||||
}
|
||||
|
||||
/*
|
||||
* Progress send frags, any frags that have been queued up before the connection
|
||||
* was established need to be progressed.
|
||||
*/
|
||||
void mca_btl_openib_progress_send_frags(mca_btl_openib_endpoint_t* endpoint)
|
||||
{
|
||||
opal_list_item_t *frag_item;
|
||||
@ -635,7 +630,7 @@ void mca_btl_openib_progress_send_frags(mca_btl_openib_endpoint_t* endpoint)
|
||||
/* We need to post this one */
|
||||
|
||||
if(OMPI_SUCCESS != mca_btl_openib_endpoint_post_send(openib_btl, endpoint, frag))
|
||||
opal_output(0, "error in mca_btl_openib_endpoint_send");
|
||||
BTL_ERROR("Error posting send");
|
||||
}
|
||||
}
|
||||
|
||||
@ -685,12 +680,17 @@ int mca_btl_openib_endpoint_connect(
|
||||
return rc;
|
||||
}
|
||||
|
||||
mca_btl_openib_endpoint_post_rr(endpoint, 0);
|
||||
MCA_BTL_OPENIB_ENDPOINT_POST_RR_HIGH(endpoint, 0);
|
||||
MCA_BTL_OPENIB_ENDPOINT_POST_RR_LOW(endpoint, 0);
|
||||
|
||||
return OMPI_SUCCESS;
|
||||
}
|
||||
|
||||
|
||||
/*
|
||||
* Create the queue pair note that this is just the initial
|
||||
* queue pair creation and we need to get the remote queue pair
|
||||
* info from the peer before the qp is usable,
|
||||
*/
|
||||
|
||||
int mca_btl_openib_endpoint_create_qp(
|
||||
mca_btl_openib_module_t* openib_btl,
|
||||
@ -718,7 +718,7 @@ int mca_btl_openib_endpoint_create_qp(
|
||||
my_qp = ibv_create_qp(pd, &qp_init_attr);
|
||||
|
||||
if(NULL == my_qp) {
|
||||
opal_output(0, "%s: error creating qp errno says %s\n", __func__, strerror(errno));
|
||||
BTL_ERROR("error creating qp errno says %s", strerror(errno));
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
(*qp) = my_qp;
|
||||
@ -737,7 +737,7 @@ int mca_btl_openib_endpoint_create_qp(
|
||||
IBV_QP_PKEY_INDEX |
|
||||
IBV_QP_PORT |
|
||||
IBV_QP_ACCESS_FLAGS )) {
|
||||
opal_output(0, "%s: error modifying qp to INIT errno says %s\n", __func__, strerror(errno));
|
||||
BTL_ERROR("error modifying qp to INIT errno says %s", strerror(errno));
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
}
|
||||
@ -745,6 +745,11 @@ int mca_btl_openib_endpoint_create_qp(
|
||||
return OMPI_SUCCESS;
|
||||
}
|
||||
|
||||
/*
|
||||
* The queue pair has been created and we have received the remote
|
||||
* queue pair information from the peer so we init this queue pair
|
||||
* and are ready to roll.
|
||||
*/
|
||||
int mca_btl_openib_endpoint_qp_init_query(
|
||||
mca_btl_openib_module_t* openib_btl,
|
||||
struct ibv_qp* qp,
|
||||
@ -778,7 +783,7 @@ int mca_btl_openib_endpoint_qp_init_query(
|
||||
IBV_QP_RQ_PSN |
|
||||
IBV_QP_MAX_DEST_RD_ATOMIC |
|
||||
IBV_QP_MIN_RNR_TIMER)) {
|
||||
opal_output(0, "%s: error modifing QP to RTR errno says %s\n", __func__, strerror(errno));
|
||||
BTL_ERROR("error modifing QP to RTR errno says %s", strerror(errno));
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
attr->qp_state = IBV_QPS_RTS;
|
||||
@ -794,7 +799,7 @@ int mca_btl_openib_endpoint_qp_init_query(
|
||||
IBV_QP_RNR_RETRY |
|
||||
IBV_QP_SQ_PSN |
|
||||
IBV_QP_MAX_QP_RD_ATOMIC)) {
|
||||
opal_output(0, "%s: error modifying QP to RTS errno says %s\n", __func__, strerror(errno));
|
||||
BTL_ERROR("error modifying QP to RTS errno says %s", strerror(errno));
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
return OMPI_SUCCESS;
|
||||
|
@ -26,7 +26,7 @@
|
||||
#include "btl_openib.h"
|
||||
#include <errno.h>
|
||||
#include <string.h>
|
||||
|
||||
#include "mca/btl/base/btl_base_error.h"
|
||||
extern int errno;
|
||||
|
||||
#if defined(c_plusplus) || defined(__cplusplus)
|
||||
@ -133,87 +133,66 @@ void mca_btl_openib_post_recv(void);
|
||||
|
||||
void mca_btl_openib_progress_send_frags(mca_btl_openib_endpoint_t*);
|
||||
|
||||
static inline int mca_btl_openib_endpoint_post_rr_sub(int cnt,
|
||||
mca_btl_openib_endpoint_t* endpoint,
|
||||
ompi_free_list_t* frag_list,
|
||||
uint32_t* rr_posted,
|
||||
struct ibv_qp* qp
|
||||
)
|
||||
{
|
||||
|
||||
int rc, i;
|
||||
opal_list_item_t* item;
|
||||
mca_btl_openib_frag_t* frag;
|
||||
mca_btl_openib_module_t *openib_btl = endpoint->endpoint_btl;
|
||||
struct ibv_recv_wr* bad_wr;
|
||||
struct ibv_recv_wr* rr_desc_post = openib_btl->rr_desc_post;
|
||||
|
||||
/* prepare frags and post receive requests, given, this is ugly,
|
||||
* if openib doesn't plan on supporting a post_list method than
|
||||
* this should be changed to simply loop through and post receives
|
||||
* without bothering with the rr_desc_post array as it is not needed
|
||||
*/
|
||||
for(i = 0; i < cnt; i++) {
|
||||
OMPI_FREE_LIST_WAIT(frag_list, item, rc);
|
||||
frag = (mca_btl_openib_frag_t*) item;
|
||||
frag->endpoint = endpoint;
|
||||
frag->sg_entry.length = frag->size +
|
||||
((unsigned char*) frag->segment.seg_addr.pval-
|
||||
(unsigned char*) frag->hdr);
|
||||
|
||||
rr_desc_post[i] = frag->rr_desc;
|
||||
}
|
||||
|
||||
for(i=0; i< cnt; i++){
|
||||
if(ibv_post_recv(qp,
|
||||
&rr_desc_post[i],
|
||||
&bad_wr)) {
|
||||
opal_output(0, "%s: error posting receive errno says %s\n", __func__, strerror(errno));
|
||||
return OMPI_ERROR;
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
OPAL_THREAD_ADD32(rr_posted, cnt);
|
||||
return OMPI_SUCCESS;
|
||||
#define MCA_BTL_OPENIB_ENDPOINT_POST_RR_HIGH(post_rr_high_endpoint, \
|
||||
post_rr_high_additional) \
|
||||
{ \
|
||||
mca_btl_openib_module_t * post_rr_high_openib_btl = post_rr_high_endpoint->endpoint_btl; \
|
||||
OPAL_THREAD_LOCK(&post_rr_high_openib_btl->ib_lock); \
|
||||
if(post_rr_high_endpoint->rr_posted_high <= mca_btl_openib_component.ib_rr_buf_min+post_rr_high_additional && \
|
||||
post_rr_high_endpoint->rr_posted_high < mca_btl_openib_component.ib_rr_buf_max){ \
|
||||
MCA_BTL_OPENIB_ENDPOINT_POST_RR_SUB(mca_btl_openib_component.ib_rr_buf_max - \
|
||||
post_rr_high_endpoint->rr_posted_high, \
|
||||
post_rr_high_endpoint, \
|
||||
&post_rr_high_openib_btl->recv_free_eager, \
|
||||
&post_rr_high_endpoint->rr_posted_high, \
|
||||
post_rr_high_endpoint->lcl_qp_high); \
|
||||
} \
|
||||
OPAL_THREAD_UNLOCK(&post_rr_high_openib_btl->ib_lock); \
|
||||
}
|
||||
|
||||
static inline int mca_btl_openib_endpoint_post_rr( mca_btl_openib_endpoint_t * endpoint, int additional){
|
||||
mca_btl_openib_module_t * openib_btl = endpoint->endpoint_btl;
|
||||
int rc;
|
||||
OPAL_THREAD_LOCK(&openib_btl->ib_lock);
|
||||
#define MCA_BTL_OPENIB_ENDPOINT_POST_RR_LOW(post_rr_low_endpoint, \
|
||||
post_rr_low_additional) { \
|
||||
mca_btl_openib_module_t * post_rr_low_openib_btl = post_rr_low_endpoint->endpoint_btl; \
|
||||
OPAL_THREAD_LOCK(&post_rr_low_openib_btl->ib_lock); \
|
||||
if(post_rr_low_endpoint->rr_posted_low <= mca_btl_openib_component.ib_rr_buf_min+post_rr_low_additional && \
|
||||
post_rr_low_endpoint->rr_posted_low < mca_btl_openib_component.ib_rr_buf_max){ \
|
||||
MCA_BTL_OPENIB_ENDPOINT_POST_RR_SUB(mca_btl_openib_component.ib_rr_buf_max - \
|
||||
post_rr_low_endpoint->rr_posted_low, \
|
||||
post_rr_low_endpoint, \
|
||||
&post_rr_low_openib_btl->recv_free_max, \
|
||||
&post_rr_low_endpoint->rr_posted_low, \
|
||||
post_rr_low_endpoint->lcl_qp_low \
|
||||
); } \
|
||||
OPAL_THREAD_UNLOCK(&post_rr_low_openib_btl->ib_lock); \
|
||||
}
|
||||
|
||||
if(endpoint->rr_posted_high <= mca_btl_openib_component.ib_rr_buf_min+additional && endpoint->rr_posted_high < mca_btl_openib_component.ib_rr_buf_max){
|
||||
|
||||
rc = mca_btl_openib_endpoint_post_rr_sub(mca_btl_openib_component.ib_rr_buf_max - endpoint->rr_posted_high,
|
||||
endpoint,
|
||||
&openib_btl->recv_free_eager,
|
||||
&endpoint->rr_posted_high,
|
||||
endpoint->lcl_qp_high
|
||||
);
|
||||
if(rc != OMPI_SUCCESS){
|
||||
OPAL_THREAD_UNLOCK(&openib_btl->ib_lock);
|
||||
return rc;
|
||||
}
|
||||
}
|
||||
if(endpoint->rr_posted_low <= mca_btl_openib_component.ib_rr_buf_min+additional && endpoint->rr_posted_low < mca_btl_openib_component.ib_rr_buf_max){
|
||||
|
||||
rc = mca_btl_openib_endpoint_post_rr_sub(mca_btl_openib_component.ib_rr_buf_max - endpoint->rr_posted_low,
|
||||
endpoint,
|
||||
&openib_btl->recv_free_max,
|
||||
&endpoint->rr_posted_low,
|
||||
endpoint->lcl_qp_low
|
||||
);
|
||||
if(rc != OMPI_SUCCESS) {
|
||||
OPAL_THREAD_UNLOCK(&openib_btl->ib_lock);
|
||||
return rc;
|
||||
}
|
||||
|
||||
}
|
||||
OPAL_THREAD_UNLOCK(&openib_btl->ib_lock);
|
||||
return OMPI_SUCCESS;
|
||||
|
||||
|
||||
#define MCA_BTL_OPENIB_ENDPOINT_POST_RR_SUB(post_rr_sub_cnt, \
|
||||
post_rr_sub_endpoint, \
|
||||
post_rr_sub_frag_list, \
|
||||
post_rr_sub_rr_posted, \
|
||||
post_rr_sub_qp ) \
|
||||
{\
|
||||
uint32_t post_rr_sub_i; \
|
||||
int post_rr_sub_rc; \
|
||||
opal_list_item_t* post_rr_sub_item; \
|
||||
mca_btl_openib_frag_t* post_rr_sub_frag; \
|
||||
struct ibv_recv_wr* post_rr_sub_bad_wr; \
|
||||
for(post_rr_sub_i = 0; post_rr_sub_i < post_rr_sub_cnt; post_rr_sub_i++) { \
|
||||
OMPI_FREE_LIST_WAIT(post_rr_sub_frag_list, post_rr_sub_item, post_rr_sub_rc); \
|
||||
post_rr_sub_frag = (mca_btl_openib_frag_t*) post_rr_sub_item; \
|
||||
post_rr_sub_frag->endpoint = post_rr_sub_endpoint; \
|
||||
post_rr_sub_frag->sg_entry.length = post_rr_sub_frag->size + \
|
||||
((unsigned char*) post_rr_sub_frag->segment.seg_addr.pval- \
|
||||
(unsigned char*) post_rr_sub_frag->hdr); \
|
||||
if(ibv_post_recv(post_rr_sub_qp, \
|
||||
&post_rr_sub_frag->wr_desc.rr_desc, \
|
||||
&post_rr_sub_bad_wr)) { \
|
||||
BTL_ERROR("error posting receive errno says %s\n", strerror(errno)); \
|
||||
return OMPI_ERROR; \
|
||||
}\
|
||||
}\
|
||||
OPAL_THREAD_ADD32(post_rr_sub_rr_posted, post_rr_sub_cnt); \
|
||||
}
|
||||
|
||||
#define DUMP_ENDPOINT(endpoint_ptr) { \
|
||||
|
@ -1,37 +0,0 @@
|
||||
/*
|
||||
* 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
|
||||
*
|
||||
* $HEADER$
|
||||
*/
|
||||
|
||||
#ifndef MCA_BTL_IB_ERROR_H
|
||||
#define MCA_BTL_IB_ERROR_H
|
||||
|
||||
#include <infiniband/verbs.h>
|
||||
|
||||
/*
|
||||
*
|
||||
*
|
||||
*/
|
||||
|
||||
/* Debug Print */
|
||||
#if 0
|
||||
#define DEBUG_OUT(fmt, args...) { \
|
||||
opal_output(0, "[%s:%d:%s] " fmt, __FILE__, __LINE__, __func__, \
|
||||
##args); \
|
||||
}
|
||||
#else
|
||||
#define DEBUG_OUT(fmt, args...)
|
||||
#endif
|
||||
|
||||
#endif
|
@ -1,3 +1,19 @@
|
||||
/*
|
||||
* 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
|
||||
*
|
||||
* $HEADER$
|
||||
*/
|
||||
|
||||
#include "btl_openib_frag.h"
|
||||
#include "mca/mpool/openib/mpool_openib.h"
|
||||
|
||||
@ -45,12 +61,12 @@ static void mca_btl_openib_send_frag_common_constructor(mca_btl_openib_frag_t* f
|
||||
frag->base.des_dst = NULL;
|
||||
frag->base.des_dst_cnt = 0;
|
||||
|
||||
frag->sr_desc.wr_id = (uint64_t) frag;
|
||||
frag->sr_desc.sg_list = &frag->sg_entry;
|
||||
frag->sr_desc.num_sge = 1;
|
||||
frag->sr_desc.opcode = IBV_WR_SEND;
|
||||
frag->sr_desc.send_flags = IBV_SEND_SIGNALED;
|
||||
frag->sr_desc.next = NULL;
|
||||
frag->wr_desc.sr_desc.wr_id = (uint64_t) frag;
|
||||
frag->wr_desc.sr_desc.sg_list = &frag->sg_entry;
|
||||
frag->wr_desc.sr_desc.num_sge = 1;
|
||||
frag->wr_desc.sr_desc.opcode = IBV_WR_SEND;
|
||||
frag->wr_desc.sr_desc.send_flags = IBV_SEND_SIGNALED;
|
||||
frag->wr_desc.sr_desc.next = NULL;
|
||||
}
|
||||
|
||||
static void mca_btl_openib_recv_frag_common_constructor(mca_btl_openib_frag_t* frag)
|
||||
@ -62,10 +78,10 @@ static void mca_btl_openib_recv_frag_common_constructor(mca_btl_openib_frag_t* f
|
||||
frag->base.des_src = NULL;
|
||||
frag->base.des_src_cnt = 0;
|
||||
|
||||
frag->rr_desc.wr_id = (uint64_t) frag;
|
||||
frag->rr_desc.sg_list = &frag->sg_entry;
|
||||
frag->rr_desc.num_sge = 1;
|
||||
frag->rr_desc.next = NULL;
|
||||
frag->wr_desc.rr_desc.wr_id = (uint64_t) frag;
|
||||
frag->wr_desc.rr_desc.sg_list = &frag->sg_entry;
|
||||
frag->wr_desc.rr_desc.num_sge = 1;
|
||||
frag->wr_desc.rr_desc.next = NULL;
|
||||
}
|
||||
|
||||
static void mca_btl_openib_send_frag_eager_constructor(mca_btl_openib_frag_t* frag)
|
||||
|
@ -51,11 +51,10 @@ struct mca_btl_openib_frag_t {
|
||||
mca_btl_openib_frag_type_t type;
|
||||
size_t size;
|
||||
int rc;
|
||||
|
||||
union{
|
||||
struct ibv_recv_wr rr_desc;
|
||||
struct ibv_send_wr sr_desc;
|
||||
};
|
||||
} wr_desc;
|
||||
struct ibv_sge sg_entry;
|
||||
struct ibv_mr *mr;
|
||||
mca_btl_openib_header_t *hdr;
|
||||
|
@ -75,16 +75,12 @@ static mca_btl_openib_proc_t* mca_btl_openib_proc_lookup_ompi(ompi_proc_t* ompi_
|
||||
ib_proc != (mca_btl_openib_proc_t*)
|
||||
opal_list_get_end(&mca_btl_openib_component.ib_procs);
|
||||
ib_proc = (mca_btl_openib_proc_t*)opal_list_get_next(ib_proc)) {
|
||||
|
||||
if(ib_proc->proc_ompi == ompi_proc) {
|
||||
OPAL_THREAD_UNLOCK(&mca_btl_openib_component.ib_lock);
|
||||
return ib_proc;
|
||||
}
|
||||
|
||||
}
|
||||
|
||||
OPAL_THREAD_UNLOCK(&mca_btl_openib_component.ib_lock);
|
||||
|
||||
return NULL;
|
||||
}
|
||||
|
||||
@ -105,19 +101,15 @@ mca_btl_openib_proc_t* mca_btl_openib_proc_create(ompi_proc_t* ompi_proc)
|
||||
module_proc = mca_btl_openib_proc_lookup_ompi(ompi_proc);
|
||||
|
||||
if(module_proc != NULL) {
|
||||
|
||||
/* Gotcha! */
|
||||
return module_proc;
|
||||
}
|
||||
|
||||
/* Oops! First time, gotta create a new IB proc
|
||||
* out of the ompi_proc ... */
|
||||
|
||||
module_proc = OBJ_NEW(mca_btl_openib_proc_t);
|
||||
|
||||
/* Initialize number of peer */
|
||||
module_proc->proc_endpoint_count = 0;
|
||||
|
||||
module_proc->proc_ompi = ompi_proc;
|
||||
|
||||
/* build a unique identifier (of arbitrary
|
||||
@ -156,6 +148,5 @@ int mca_btl_openib_proc_insert(mca_btl_openib_proc_t* module_proc,
|
||||
/* insert into endpoint array */
|
||||
module_endpoint->endpoint_proc = module_proc;
|
||||
module_proc->proc_endpoints[module_proc->proc_endpoint_count++] = module_endpoint;
|
||||
|
||||
return OMPI_SUCCESS;
|
||||
}
|
||||
|
Загрузка…
x
Ссылка в новой задаче
Block a user