Skip to content

ikrit spml cleanup, mkey cache and assorted bug fixes #2354

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 14 commits into from
Nov 14, 2016
Merged

ikrit spml cleanup, mkey cache and assorted bug fixes #2354

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
2f91ce7
OSHMEM: mxm versions less than 2.0 are no longer supported
alex-mikheev Oct 25, 2016
0826e63
OSHMEM: spml_ikrit: makes quiet wait for get_nbi requests
alex-mikheev Oct 25, 2016
b5c7c7d
OSHMEM: memheap: disable oob if allgather mkey exchange is used
alex-mikheev Oct 26, 2016
df74d54
OSHMEM: spml ikrit: changes mxm_peers layout
alex-mikheev Oct 26, 2016
23c3dc8
OSHMEM: mxm: optimize mxm_peer layout.
alex-mikheev Oct 27, 2016
61bd59a
OSHMEM: fixes addr_acessible()
alex-mikheev Oct 27, 2016
defcc3d
OSHMEM: spml ikrit: get/put request cleanup
alex-mikheev Oct 27, 2016
7caa736
OSHMEM: fixes potential deadlock in shmem_lock()
alex-mikheev Nov 2, 2016
5c2f807
OSHMEM: fixes verbosity log level cal
alex-mikheev Nov 2, 2016
ff5095e
OSHMEM: adds support for mkey caching by spml
alex-mikheev Nov 2, 2016
f133d9b
oshmem: fixes comiplation errors in sshmem
alex-mikheev Nov 7, 2016
bf61961
oshmem: code review fixes
alex-mikheev Nov 8, 2016
48a7a0b
oshmem: lock: call opal_progress only when busy waiting
alex-mikheev Nov 10, 2016
864904e
oshmem: ucx: check status only if configured --with-oshmem-param-check
alex-mikheev Nov 10, 2016
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
75 changes: 68 additions & 7 deletions oshmem/mca/atomic/mxm/atomic_mxm.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@

/* This component does uses SPML:IKRIT */
#include "oshmem/mca/spml/ikrit/spml_ikrit.h"
#include "oshmem/runtime/runtime.h"


BEGIN_C_DECLS
Expand Down Expand Up @@ -60,16 +61,76 @@ struct mca_atomic_mxm_module_t {
typedef struct mca_atomic_mxm_module_t mca_atomic_mxm_module_t;
OBJ_CLASS_DECLARATION(mca_atomic_mxm_module_t);

END_C_DECLS

#if MXM_API >= MXM_VERSION(2,0)
static inline mxm_mem_key_t *to_mxm_mkey(sshmem_mkey_t *mkey) {
static inline uint8_t mca_atomic_mxm_order(size_t nlong)
{
if (OPAL_LIKELY(8 == nlong)) {
return 3;
}

if (OPAL_LIKELY(4 == nlong)) {
return 2;
}

if (2 == nlong) {
return 1;
}

if (0 == mkey->len) {
return &mxm_empty_mem_key;
if (1 == nlong) {
return 0;
}
return (mxm_mem_key_t *)mkey->u.data;

ATOMIC_ERROR("Type size must be 1/2/4 or 8 bytes.");
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}
#endif

static inline void mca_atomic_mxm_req_init(mxm_send_req_t *sreq, int pe, void *target, size_t nlong)
{
uint8_t nlong_order;
void *remote_addr;
mxm_mem_key_t *mkey;

nlong_order = mca_atomic_mxm_order(nlong);

mkey = mca_spml_ikrit_get_mkey(pe, target, MXM_PTL_RDMA, &remote_addr);

/* mxm request init */
sreq->base.state = MXM_REQ_NEW;
sreq->base.mq = mca_atomic_mxm_spml_self->mxm_mq;
sreq->base.conn = mca_atomic_mxm_spml_self->mxm_peers[pe].mxm_hw_rdma_conn;
sreq->base.completed_cb = NULL;
sreq->base.data_type = MXM_REQ_DATA_BUFFER;

sreq->base.data.buffer.memh = MXM_INVALID_MEM_HANDLE;
sreq->base.data.buffer.length = nlong;

sreq->op.atomic.remote_vaddr = (uintptr_t) remote_addr;
sreq->op.atomic.remote_mkey = mkey;
sreq->op.atomic.order = nlong_order;

sreq->flags = 0;
}

static inline void mca_atomic_mxm_post(mxm_send_req_t *sreq)
{
mxm_error_t mxm_err;

mxm_err = mxm_req_send(sreq);
if (OPAL_UNLIKELY(MXM_OK != mxm_err)) {
ATOMIC_ERROR("mxm_req_send failed, mxm_error = %d",
mxm_err);
oshmem_shmem_abort(-1);
}

mxm_req_wait(&sreq->base);
if (OPAL_UNLIKELY(MXM_OK != sreq->base.error)) {
ATOMIC_ERROR("mxm_req_wait got non MXM_OK error: %d",
sreq->base.error);
oshmem_shmem_abort(-1);
}
}

END_C_DECLS

#endif /* MCA_ATOMIC_MXM_H */
80 changes: 2 additions & 78 deletions oshmem/mca/atomic/mxm/atomic_mxm_cswap.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,96 +31,20 @@ int mca_atomic_mxm_cswap(void *target,
size_t nlong,
int pe)
{
unsigned my_pe;
uint8_t nlong_order;
void *remote_addr;
int ptl_id;
mxm_send_req_t sreq;
mxm_error_t mxm_err;
sshmem_mkey_t *r_mkey;

my_pe = oshmem_my_proc_id();
ptl_id = -1;
mxm_err = MXM_OK;
mca_atomic_mxm_req_init(&sreq, pe, target, nlong);

switch (nlong) {
case 1:
nlong_order = 0;
break;
case 2:
nlong_order = 1;
break;
case 4:
nlong_order = 2;
break;
case 8:
nlong_order = 3;
break;
default:
ATOMIC_ERROR("[#%d] Type size must be 1/2/4 or 8 bytes.", my_pe);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}

ptl_id = OSHMEM_PROC_DATA(oshmem_proc_group_all(pe))->transport_ids[0];
if (MXM_PTL_SHM == ptl_id) {
ptl_id = MXM_PTL_RDMA;
}
r_mkey = mca_memheap_base_get_cached_mkey(pe, target, ptl_id, &remote_addr);
if (!r_mkey) {
ATOMIC_ERROR("[#%d] %p is not address of symmetric variable",
my_pe, target);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}

/* mxm request init */
sreq.base.state = MXM_REQ_NEW;
sreq.base.mq = mca_atomic_mxm_spml_self->mxm_mq;
sreq.base.conn = mca_atomic_mxm_spml_self->mxm_peers[pe]->mxm_hw_rdma_conn;
sreq.base.completed_cb = NULL;
sreq.base.data_type = MXM_REQ_DATA_BUFFER;

/* set data */
sreq.base.data.buffer.ptr = (void *) value;
sreq.base.data.buffer.length = nlong;
sreq.base.data.buffer.memh = MXM_INVALID_MEM_HANDLE;

sreq.op.atomic.remote_vaddr = (uintptr_t) remote_addr;
#if MXM_API < MXM_VERSION(2,0)
sreq.base.flags = 0;
sreq.op.atomic.remote_memh = MXM_INVALID_MEM_HANDLE;
#else
sreq.flags = 0;
sreq.op.atomic.remote_mkey = to_mxm_mkey(r_mkey);
#endif
sreq.op.atomic.order = nlong_order;

if (NULL == cond) {
sreq.opcode = MXM_REQ_OP_ATOMIC_SWAP;
} else {
#if MXM_API < MXM_VERSION(2,0)
memcpy(&sreq.op.atomic.value8, cond, nlong);
#else
memcpy(&sreq.op.atomic.value, cond, nlong);
#endif
sreq.opcode = MXM_REQ_OP_ATOMIC_CSWAP;
}

if (MXM_OK != (mxm_err = mxm_req_send(&sreq))) {
ATOMIC_ERROR("[#%d] mxm_req_send failed, mxm_error = %d",
my_pe, mxm_err);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
mca_atomic_mxm_post(&sreq);

mxm_req_wait(&sreq.base);
if (MXM_OK != sreq.base.error) {
ATOMIC_ERROR("[#%d] mxm_req_wait got non MXM_OK error: %d",
my_pe, sreq.base.error);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
memcpy(prev, value, nlong);

return OSHMEM_SUCCESS;
Expand Down
92 changes: 3 additions & 89 deletions oshmem/mca/atomic/mxm/atomic_mxm_fadd.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,106 +32,20 @@ int mca_atomic_mxm_fadd(void *target,
int pe,
struct oshmem_op_t *op)
{
unsigned my_pe;
uint8_t nlong_order;
void *remote_addr;
int ptl_id;
mxm_send_req_t sreq;
mxm_error_t mxm_err;
sshmem_mkey_t *r_mkey;
static char dummy_buf[8];

my_pe = oshmem_my_proc_id();
ptl_id = -1;
mxm_err = MXM_OK;
mca_atomic_mxm_req_init(&sreq, pe, target, nlong);

switch (nlong) {
case 1:
nlong_order = 0;
break;
case 2:
nlong_order = 1;
break;
case 4:
nlong_order = 2;
break;
case 8:
nlong_order = 3;
break;
default:
ATOMIC_ERROR("[#%d] Type size must be 1/2/4 or 8 bytes.", my_pe);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}

ptl_id = OSHMEM_PROC_DATA(oshmem_proc_group_all(pe))->transport_ids[0];
if (MXM_PTL_SHM == ptl_id) {
ptl_id = MXM_PTL_RDMA;
}
r_mkey = mca_memheap_base_get_cached_mkey(pe, target, ptl_id, &remote_addr);
if (!r_mkey) {
ATOMIC_ERROR("[#%d] %p is not address of symmetric variable",
my_pe, target);
oshmem_shmem_abort(-1);
return OSHMEM_ERR_BAD_PARAM;
}

/* mxm request init */
sreq.base.state = MXM_REQ_NEW;
sreq.base.mq = mca_atomic_mxm_spml_self->mxm_mq;
sreq.base.conn = mca_atomic_mxm_spml_self->mxm_peers[pe]->mxm_hw_rdma_conn;
sreq.base.completed_cb = NULL;
sreq.base.data_type = MXM_REQ_DATA_BUFFER;

sreq.op.atomic.remote_vaddr = (uintptr_t) remote_addr;
#if MXM_API < MXM_VERSION(2,0)
sreq.op.atomic.remote_memh = MXM_INVALID_MEM_HANDLE;
memcpy(&sreq.op.atomic.value8, value, nlong);
#else
sreq.op.atomic.remote_mkey = to_mxm_mkey(r_mkey);
memcpy(&sreq.op.atomic.value, value, nlong);
#endif
sreq.op.atomic.order = nlong_order;

/* Do we need atomic 'add' or atomic 'fetch and add'? */
sreq.opcode = MXM_REQ_OP_ATOMIC_FADD;
if (NULL == prev) {
sreq.base.data.buffer.ptr = dummy_buf;
sreq.base.data.buffer.length = nlong;
sreq.base.data.buffer.memh = MXM_INVALID_MEM_HANDLE;
#if MXM_API < MXM_VERSION(2,0)
sreq.base.flags = MXM_REQ_FLAG_SEND_SYNC;
sreq.opcode = MXM_REQ_OP_ATOMIC_ADD;
#else
sreq.flags = 0;
sreq.opcode = MXM_REQ_OP_ATOMIC_FADD;
#endif
} else {
sreq.base.data.buffer.ptr = prev;
sreq.base.data.buffer.length = nlong;
sreq.base.data.buffer.memh = MXM_INVALID_MEM_HANDLE;
#if MXM_API < MXM_VERSION(2,0)
sreq.base.flags = 0;
#else
sreq.flags = 0;
#endif

sreq.opcode = MXM_REQ_OP_ATOMIC_FADD;
}

if (MXM_OK != (mxm_err = mxm_req_send(&sreq))) {
ATOMIC_ERROR("[#%d] mxm_req_send failed, mxm_error = %d",
my_pe, mxm_err);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}

mxm_req_wait(&sreq.base);
if (MXM_OK != sreq.base.error) {
ATOMIC_ERROR("[#%d] mxm_req_wait got non MXM_OK error: %d",
my_pe, sreq.base.error);
oshmem_shmem_abort(-1);
return OSHMEM_ERROR;
}
mca_atomic_mxm_post(&sreq);

return OSHMEM_SUCCESS;
}
Loading