1
1
openmpi/ompi/mca/btl/usnic/btl_usnic_send.h

283 строки
8.7 KiB
C
Исходник Обычный вид История

/*
* Copyright (c) 2013 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#ifndef BTL_USNIC_SEND_H
#define BTL_USNIC_SEND_H
#include <infiniband/verbs.h>
#include "btl_usnic.h"
#include "btl_usnic_frag.h"
#include "btl_usnic_ack.h"
#if MSGDEBUG1
#include "btl_usnic_util.h"
#endif
/*
* Check if conditions are right, and if so, put endpoint on
* list of endpoints that have sends to be done
*/
static inline void
ompi_btl_usnic_check_rts(
ompi_btl_usnic_endpoint_t *endpoint)
{
/*
* If endpoint not already ready,
* and has packets to send,
* and it has send credits,
* and its retransmission window is open,
* make it ready
*/
if (!endpoint->endpoint_ready_to_send &&
!opal_list_is_empty(&endpoint->endpoint_frag_send_queue) &&
endpoint->endpoint_send_credits > 0 &&
WINDOW_OPEN(endpoint)) {
opal_list_append(&endpoint->endpoint_module->endpoints_with_sends,
&endpoint->super);
endpoint->endpoint_ready_to_send = true;
#if MSGDEBUG1
opal_output(0, "make endpoint %p RTS\n", (void*)endpoint);
} else {
opal_output(0, "rts:%d empty:%d cred:%d open%d\n",
endpoint->endpoint_ready_to_send,
opal_list_is_empty(&endpoint->endpoint_frag_send_queue),
endpoint->endpoint_send_credits,
WINDOW_OPEN(endpoint));
#endif
}
}
#if MSGDEBUG2
static inline
int sge_total(struct ibv_send_wr *wr)
{
int i;
int len;
len=0;
for (i=0; i<wr->num_sge; ++i) {
len += wr->sg_list[i].length;
}
return len;
}
#endif
/*
* Common point for posting a segment to VERBS
*/
static inline void
ompi_btl_usnic_post_segment(
ompi_btl_usnic_module_t *module,
ompi_btl_usnic_endpoint_t *endpoint,
ompi_btl_usnic_send_segment_t *sseg)
{
struct ibv_send_wr *bad_wr;
ompi_btl_usnic_channel_t *channel;
struct ibv_send_wr *wr;
int ret;
#if MSGDEBUG1
opal_output(0, "post_send: type=%s, addr=%p, len=%d, payload=%d\n",
usnic_seg_type(sseg->ss_base.us_type),
(void*) sseg->ss_send_desc.sg_list->addr,
sge_total(&sseg->ss_send_desc),
sseg->ss_base.us_btl_header->payload_len);
/*ompi_btl_usnic_dump_hex((void *)(sseg->ss_send_desc.sg_list->addr + sizeof(ompi_btl_usnic_btl_header_t)), 16); */
#endif
/* set target address */
wr = &sseg->ss_send_desc;
wr->wr.ud.ah = endpoint->endpoint_remote_ah;
/* get channel and remote QPN */
channel = &module->mod_channels[sseg->ss_channel];
wr->wr.ud.remote_qpn =
endpoint->endpoint_remote_addr.qp_num[sseg->ss_channel];
/* Post segment to the NIC */
ret = ibv_post_send(channel->qp, &sseg->ss_send_desc, &bad_wr);
if (OPAL_UNLIKELY(0 != ret)) {
ompi_btl_usnic_util_abort("ibv_post_send() failed",
__FILE__, __LINE__, ret);
/* Never returns */
}
/* track # of time non-ACKs are posted */
if (sseg->ss_base.us_type != OMPI_BTL_USNIC_SEG_ACK) {
++sseg->ss_send_posted;
++sseg->ss_parent_frag->sf_seg_post_cnt;
}
/* consume a WQE */
--channel->sd_wqe;
/* Stats */
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
++module->stats.num_total_sends;
++channel->num_channel_sends;
}
/*
* Post a send to the verbs work queue
*/
static inline void
ompi_btl_usnic_endpoint_send_segment(
ompi_btl_usnic_module_t *module,
ompi_btl_usnic_send_segment_t *sseg)
{
ompi_btl_usnic_send_frag_t *frag;
ompi_btl_usnic_endpoint_t *endpoint;
uint16_t sfi;
frag = sseg->ss_parent_frag;
endpoint = frag->sf_endpoint;
/* Do we have room in the endpoint's sender window?
Sender window:
|-------- WINDOW_SIZE ----------|
+---------------------------------+
| next_seq_to_send |
| somewhere in this range |
^+---------------------------------+
|
+-- ack_seq_rcvd: one less than the window left edge
Assuming that next_seq_to_send is > ack_seq_rcvd (verified
by assert), then the good condition to send is:
next_seq_to_send <= ack_seq_rcvd + WINDOW_SIZE
And therefore the bad condition is
next_seq_to_send > ack_seq_rcvd + WINDOW_SIZE
*/
assert(endpoint->endpoint_next_seq_to_send >
endpoint->endpoint_ack_seq_rcvd);
assert(WINDOW_OPEN(endpoint));
/* Assign sequence number and increment */
sseg->ss_base.us_btl_header->seq = endpoint->endpoint_next_seq_to_send++;
/* Fill in remote address to indicate PUT or not */
sseg->ss_base.us_btl_header->put_addr =
frag->sf_base.uf_dst_seg[0].seg_addr.pval;
/* piggy-back an ACK if needed */
ompi_btl_usnic_piggyback_ack(endpoint, sseg);
#if MSGDEBUG1
{
uint8_t mac[6];
char mac_str1[128];
char mac_str2[128];
ompi_btl_usnic_sprintf_mac(mac_str1, module->local_addr.mac);
ompi_btl_usnic_gid_to_mac(&endpoint->endpoint_remote_addr.gid, mac);
ompi_btl_usnic_sprintf_mac(mac_str2, mac);
opal_output(0, "--> Sending %s: seq: %" UDSEQ ", sender: 0x%016lx from device %s MAC %s, qp %u, seg %p, room %d, wc len %u, remote MAC %s, qp %u",
(sseg->ss_parent_frag->sf_base.uf_type == OMPI_BTL_USNIC_FRAG_LARGE_SEND)?
"CHUNK" : "FRAG",
sseg->ss_base.us_btl_header->seq,
sseg->ss_base.us_btl_header->sender,
endpoint->endpoint_module->device->name,
mac_str1, module->local_addr.qp_num[sseg->ss_channel],
(void*)sseg, sseg->ss_hotel_room,
sseg->ss_base.us_sg_entry[0].length,
mac_str2, endpoint->endpoint_remote_addr.qp_num[sseg->ss_channel]);
}
#endif
/* do the actual send */
ompi_btl_usnic_post_segment(module, endpoint, sseg);
/* Track this header by stashing in an array on the endpoint that
is the same length as the sender's window (i.e., WINDOW_SIZE).
To find a unique slot in this array, use (seq % WINDOW_SIZE).
*/
sfi = WINDOW_SIZE_MOD(sseg->ss_base.us_btl_header->seq);
endpoint->endpoint_sent_segs[sfi] = sseg;
sseg->ss_ack_pending = true;
/* bookkeeping */
--endpoint->endpoint_send_credits;
/* Stats */
if (sseg->ss_parent_frag->sf_base.uf_type
== OMPI_BTL_USNIC_FRAG_LARGE_SEND) {
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
++module->stats.num_chunk_sends;
} else {
Move all usNIC stats to _stats.c|h and export them as MPI_T pvars. This commit moves all the module stats into their own struct so that the stats only need to appear as a single line in the module_t definition, and then moves all the logic for reporting the stats into btl_usnic_stats.c|h. Further, the stats are now exported as MPI_T_BIND_NO_OBJECT entities (i.e., not bound to any particular MPI handle), and are marked as READONLY and CONTINUOUS. They currently all default to verbose level 5 ("Application tuner / detailed", according to https://svn.open-mpi.org/trac/ompi/wiki/MCAParamLevels). Most of the statistics are counters, but a small number are high watermark values. Due to how counters are reported via MPI_T, none of the counters are exported through MPI_T if the MCA param btl_usnic_stats_relative=1 (i.e., the module resets the stats back to zero at a given frequency). When MPI_T_pvar_handle_alloc() is invoked on any of these pvars, it will return a count that is equal to the number of active usnic BTL modules. The values returned for any given pvar (e.g., num_total_sends) are an array containing one value for each active usnic BTL module. The ordering of values in the array is both consistent across all usnic pvars and stable throughout a single job: array slot 0 corresponds to module X, array slot 1 corresponds to module Y, etc. Mapping which array slot corresponds to which underlying Linux usnic_X device works as follows: * The btl_usnic_devices MPI_T state pvar is associated with a btl_usnic_device MPI_T enum, and be obtained via MPI_T_pvar_get_info(). * If all usNIC pvars are of length N, the values [0,N) in the btl_usnic_device enum are associated with strings of the corresponding underlying Linux device. For exampe, to look up which Linux device is reported in all usNIC pvars' array slot 1, look up the int value 1 in the btl_usnic_devices enum. Its corresponding string value is underlying Linux device name (e.g., "usnic_1"). cmr=v1.7.4:subject="usnic BTL MPI_T pvars" This commit was SVN r29545.
2013-10-29 02:23:08 +04:00
++module->stats.num_frag_sends;
}
/* If we have room in the sender's window, we also have room in
endpoint hotel */
opal_hotel_checkin_with_res(&endpoint->endpoint_hotel, sseg,
&sseg->ss_hotel_room);
}
/*
* This enqueues a fragment send into the system. A send of a fragment
* may result in the sending of multiple segments
*/
static inline int
ompi_btl_usnic_endpoint_enqueue_frag(
ompi_btl_usnic_endpoint_t *endpoint,
ompi_btl_usnic_send_frag_t *frag)
{
#if MSGDEBUG1
opal_output(0, "enq_frag: frag=%p, endpoint=%p, %s, len=%lu\n",
(void*)frag, (void*)endpoint,
usnic_frag_type(frag->sf_base.uf_type),
(long unsigned)frag->sf_base.uf_base.des_src->seg_len);
if (frag->sf_base.uf_type == OMPI_BTL_USNIC_FRAG_LARGE_SEND) {
ompi_btl_usnic_large_send_frag_t *lfrag;
lfrag = (ompi_btl_usnic_large_send_frag_t *)frag;
opal_output(0, " large size=%zd\n", lfrag->lsf_base.sf_size);
}
#endif
/* add to tail of in-progress list */
opal_list_append(&endpoint->endpoint_frag_send_queue,
&frag->sf_base.uf_base.super.super);
/* possibly make this endpoint ready to send again */
ompi_btl_usnic_check_rts(endpoint);
return OMPI_SUCCESS;
}
static inline void
ompi_btl_usnic_release_send_segment(
ompi_btl_usnic_module_t *module,
ompi_btl_usnic_send_frag_t *frag,
ompi_btl_usnic_send_segment_t *sseg)
{
/* We only return CHUNK segments because they are the only send-style
* segments that are dynamically allocated.
*/
if (sseg->ss_base.us_type == OMPI_BTL_USNIC_SEG_CHUNK) {
ompi_btl_usnic_chunk_segment_return(module, sseg);
}
}
void ompi_btl_usnic_frag_complete(ompi_btl_usnic_send_frag_t *frag);
void ompi_btl_usnic_frag_send_complete(ompi_btl_usnic_module_t *module,
ompi_btl_usnic_send_segment_t *sseg);
void ompi_btl_usnic_chunk_send_complete(ompi_btl_usnic_module_t *module,
ompi_btl_usnic_send_segment_t *sseg);
int ompi_btl_usnic_send_slower( ompi_btl_usnic_module_t *module,
ompi_btl_usnic_endpoint_t *endpoint,
ompi_btl_usnic_send_frag_t *frag,
mca_btl_base_tag_t tag);
#endif /* BTL_USNIC_SEND_H */