static UCS_CLASS_INIT_FUNC(uct_ugni_smsg_iface_t, uct_md_h md, uct_worker_h worker,
const char *dev_name, size_t rx_headroom,
const uct_iface_config_t *tl_config)
{
uct_ugni_iface_config_t *config = ucs_derived_of(tl_config, uct_ugni_iface_config_t);
ucs_status_t status;
gni_return_t ugni_rc;
unsigned int bytes_per_mbox;
gni_smsg_attr_t smsg_attr;
pthread_mutex_lock(&uct_ugni_global_lock);
UCS_CLASS_CALL_SUPER_INIT(uct_ugni_iface_t, md, worker, dev_name, &uct_ugni_smsg_iface_ops,
&config->super UCS_STATS_ARG(NULL));
/* Setting initial configuration */
self->config.smsg_seg_size = 2048;
self->config.rx_headroom = rx_headroom;
self->config.smsg_max_retransmit = 16;
self->config.smsg_max_credit = 8;
self->smsg_id = 0;
smsg_attr.msg_type = GNI_SMSG_TYPE_MBOX_AUTO_RETRANSMIT;
smsg_attr.mbox_maxcredit = self->config.smsg_max_credit;
smsg_attr.msg_maxsize = self->config.smsg_seg_size;
ugni_rc = GNI_SmsgBufferSizeNeeded(&(smsg_attr), &bytes_per_mbox);
self->bytes_per_mbox = ucs_align_up_pow2(bytes_per_mbox, ucs_get_page_size());
if (ugni_rc != GNI_RC_SUCCESS) {
ucs_error("Smsg buffer size calculation failed");
status = UCS_ERR_INVALID_PARAM;
goto exit;
}
status = ucs_mpool_init(&self->free_desc,
0,
self->config.smsg_seg_size + sizeof(uct_ugni_smsg_desc_t),
0,
UCS_SYS_CACHE_LINE_SIZE, /* alignment */
128 , /* grow */
config->mpool.max_bufs, /* max buffers */
&uct_ugni_smsg_desc_mpool_ops,
"UGNI-SMSG-DESC");
if (UCS_OK != status) {
ucs_error("Desc Mpool creation failed");
goto exit;
}
status = ucs_mpool_init(&self->free_mbox,
0,
self->bytes_per_mbox + sizeof(uct_ugni_smsg_mbox_t),
sizeof(uct_ugni_smsg_mbox_t),
UCS_SYS_CACHE_LINE_SIZE, /* alignment */
128, /* grow */
config->mpool.max_bufs, /* max buffers */
&uct_ugni_smsg_mbox_mpool_ops,
"UGNI-SMSG-MBOX");
if (UCS_OK != status) {
ucs_error("Mbox Mpool creation failed");
goto clean_desc;
}
UCT_TL_IFACE_GET_TX_DESC(&self->super.super, &self->free_desc,
self->user_desc, self->user_desc = NULL);
status = ugni_smsg_activate_iface(self);
if (UCS_OK != status) {
ucs_error("Failed to activate the interface");
goto clean_mbox;
}
ugni_rc = GNI_SmsgSetMaxRetrans(self->super.nic_handle, self->config.smsg_max_retransmit);
if (ugni_rc != GNI_RC_SUCCESS) {
ucs_error("Smsg setting max retransmit count failed.");
status = UCS_ERR_INVALID_PARAM;
goto clean_iface;
}
/* TBD: eventually the uct_ugni_progress has to be moved to
* udt layer so each ugni layer will have own progress */
uct_worker_progress_register(worker, uct_ugni_smsg_progress, self);
pthread_mutex_unlock(&uct_ugni_global_lock);
return UCS_OK;
clean_iface:
ugni_smsg_deactivate_iface(self);
clean_desc:
ucs_mpool_put(self->user_desc);
ucs_mpool_cleanup(&self->free_desc, 1);
clean_mbox:
ucs_mpool_cleanup(&self->free_mbox, 1);
exit:
ucs_error("Failed to activate interface");
pthread_mutex_unlock(&uct_ugni_global_lock);
return status;
}
ucs_status_t uct_mem_alloc(size_t min_length, uct_alloc_method_t *methods,
unsigned num_methods, uct_pd_h *pds, unsigned num_pds,
const char *alloc_name, uct_allocated_memory_t *mem)
{
uct_alloc_method_t *method;
uct_pd_attr_t pd_attr;
ucs_status_t status;
size_t alloc_length;
unsigned pd_index;
uct_mem_h memh;
uct_pd_h pd;
void *address;
int shmid;
if (min_length == 0) {
ucs_error("Allocation length cannot be 0");
return UCS_ERR_INVALID_PARAM;
}
if (num_methods == 0) {
ucs_error("No allocation methods provided");
return UCS_ERR_INVALID_PARAM;
}
for (method = methods; method < methods + num_methods; ++method) {
ucs_debug("trying allocation method %s", uct_alloc_method_names[*method]);
switch (*method) {
case UCT_ALLOC_METHOD_PD:
/* Allocate with one of the specified protection domains */
for (pd_index = 0; pd_index < num_pds; ++pd_index) {
pd = pds[pd_index];
status = uct_pd_query(pd, &pd_attr);
if (status != UCS_OK) {
ucs_error("Failed to query PD");
return status;
}
/* Check if PD supports allocation */
if (!(pd_attr.cap.flags & UCT_PD_FLAG_ALLOC)) {
continue;
}
/* Allocate memory using the PD.
* If the allocation fails, it's considered an error and we don't
* fall-back, because this PD already exposed support for memory
* allocation.
*/
alloc_length = min_length;
status = uct_pd_mem_alloc(pd, &alloc_length, &address,
alloc_name, &memh);
if (status != UCS_OK) {
ucs_error("failed to allocate %zu bytes using pd %s: %s",
alloc_length, pd->component->name,
ucs_status_string(status));
return status;
}
mem->pd = pd;
mem->memh = memh;
goto allocated;
}
break;
case UCT_ALLOC_METHOD_HEAP:
/* Allocate aligned memory using libc allocator */
alloc_length = min_length;
address = ucs_memalign(UCS_SYS_CACHE_LINE_SIZE, alloc_length
UCS_MEMTRACK_VAL);
if (address != NULL) {
goto allocated_without_pd;
}
ucs_debug("failed to allocate %zu bytes from the heap", alloc_length);
break;
case UCT_ALLOC_METHOD_MMAP:
/* Request memory from operating system using mmap() */
alloc_length = ucs_align_up_pow2(min_length, ucs_get_page_size());
address = ucs_mmap(NULL, alloc_length, PROT_READ|PROT_WRITE,
MAP_PRIVATE|MAP_ANON, -1, 0 UCS_MEMTRACK_VAL);
if (address != MAP_FAILED) {
goto allocated_without_pd;
}
ucs_debug("failed to mmap %zu bytes: %m", alloc_length);
break;
case UCT_ALLOC_METHOD_HUGE:
/* Allocate huge pages */
alloc_length = min_length;
status = ucs_sysv_alloc(&alloc_length, &address, SHM_HUGETLB, &shmid
UCS_MEMTRACK_VAL);
if (status == UCS_OK) {
goto allocated_without_pd;
}
ucs_debug("failed to allocate %zu bytes from hugetlb: %s",
min_length, ucs_status_string(status));
break;
default:
ucs_error("Invalid allocation method %d", *method);
return UCS_ERR_INVALID_PARAM;
}
}
ucs_debug("Could not allocate memory with any of the provided methods");
return UCS_ERR_NO_MEMORY;
allocated_without_pd:
mem->pd = NULL;
mem->memh = UCT_INVALID_MEM_HANDLE;
allocated:
ucs_debug("allocated %zu bytes at %p using %s", alloc_length, address,
(mem->pd == NULL) ? uct_alloc_method_names[*method]
: mem->pd->component->name);
mem->address = address;
mem->length = alloc_length;
mem->method = *method;
return UCS_OK;
}
/*
* Generic data-pointer posting function.
* Parameters which are not relevant to the opcode are ignored.
*
* +--------+-----+-------+--------+-------+
* SEND | CTRL | INL | am_id | am_hdr | DPSEG |
* +--------+-----+---+---+----+----+------+
* RDMA_WRITE | CTRL | RADDR | DPSEG |
* +--------+---------+--------+-------+
* ATOMIC | CTRL | RADDR | ATOMIC | DPSEG |
* +--------+---------+--------+-------+
*/
static UCS_F_ALWAYS_INLINE void
uct_rc_mlx5_ep_dptr_post(uct_rc_mlx5_ep_t *ep, unsigned opcode_flags,
const void *buffer, unsigned length, uint32_t *lkey_p,
/* SEND */ uint8_t am_id, const void *am_hdr, unsigned am_hdr_len,
/* RDMA/ATOMIC */ uint64_t remote_addr, uct_rkey_t rkey,
/* ATOMIC */ uint64_t compare_mask, uint64_t compare, uint64_t swap_add,
int signal)
{
struct mlx5_wqe_ctrl_seg *ctrl;
struct mlx5_wqe_raddr_seg *raddr;
struct mlx5_wqe_atomic_seg *atomic;
struct mlx5_wqe_data_seg *dptr;
struct mlx5_wqe_inl_data_seg *inl;
struct uct_ib_mlx5_atomic_masked_cswap32_seg *masked_cswap32;
struct uct_ib_mlx5_atomic_masked_fadd32_seg *masked_fadd32;
struct uct_ib_mlx5_atomic_masked_cswap64_seg *masked_cswap64;
uct_rc_mlx5_iface_t *iface;
uct_rc_hdr_t *rch;
unsigned wqe_size, inl_seg_size;
uint8_t opmod;
iface = ucs_derived_of(ep->super.super.super.iface, uct_rc_mlx5_iface_t);
if (!signal) {
signal = uct_rc_iface_tx_moderation(&iface->super, &ep->super,
MLX5_WQE_CTRL_CQ_UPDATE);
} else {
ucs_assert(signal == MLX5_WQE_CTRL_CQ_UPDATE);
}
opmod = 0;
ctrl = ep->tx.wq.curr;
switch (opcode_flags) {
case MLX5_OPCODE_SEND:
inl_seg_size = ucs_align_up_pow2(sizeof(*inl) + sizeof(*rch) + am_hdr_len,
UCT_IB_MLX5_WQE_SEG_SIZE);
ucs_assert(sizeof(*ctrl) + inl_seg_size + sizeof(*dptr) <=
UCT_RC_MLX5_MAX_BB * MLX5_SEND_WQE_BB);
ucs_assert(length + sizeof(*rch) + am_hdr_len <=
iface->super.super.config.seg_size);
/* Inline segment with AM ID and header */
inl = (void*)(ctrl + 1);
inl->byte_count = htonl((sizeof(*rch) + am_hdr_len) | MLX5_INLINE_SEG);
rch = (void*)(inl + 1);
rch->am_id = am_id;
uct_ib_mlx5_inline_copy(rch + 1, am_hdr, am_hdr_len, &ep->tx.wq);
/* Data segment with payload */
if (length == 0) {
wqe_size = sizeof(*ctrl) + inl_seg_size;
} else {
wqe_size = sizeof(*ctrl) + inl_seg_size + sizeof(*dptr);
dptr = (void*)(ctrl + 1) + inl_seg_size;
if (ucs_unlikely((void*)dptr >= ep->tx.wq.qend)) {
dptr = (void*)dptr - (ep->tx.wq.qend - ep->tx.wq.qstart);
}
ucs_assert((void*)dptr >= ep->tx.wq.qstart);
ucs_assert((void*)(dptr + 1) <= ep->tx.wq.qend);
uct_ib_mlx5_set_data_seg(dptr, buffer, length, *lkey_p);
}
break;
case MLX5_OPCODE_SEND|UCT_RC_MLX5_OPCODE_FLAG_RAW:
/* Data segment only */
ucs_assert(length < (2ul << 30));
ucs_assert(length <= iface->super.super.config.seg_size);
wqe_size = sizeof(*ctrl) + sizeof(*dptr);
uct_ib_mlx5_set_data_seg((void*)(ctrl + 1), buffer, length, *lkey_p);
break;
case MLX5_OPCODE_RDMA_READ:
case MLX5_OPCODE_RDMA_WRITE:
/* Set RDMA segment */
ucs_assert(length <= UCT_IB_MAX_MESSAGE_SIZE);
raddr = (void*)(ctrl + 1);
uct_rc_mlx5_ep_set_rdma_seg(raddr, remote_addr, rkey);
/* Data segment */
if (length == 0) {
wqe_size = sizeof(*ctrl) + sizeof(*raddr);
} else {
wqe_size = sizeof(*ctrl) + sizeof(*raddr) + sizeof(*dptr);
uct_ib_mlx5_set_data_seg((void*)(raddr + 1), buffer, length, *lkey_p);
}
break;
case MLX5_OPCODE_ATOMIC_FA:
case MLX5_OPCODE_ATOMIC_CS:
ucs_assert(length == sizeof(uint64_t));
raddr = (void*)(ctrl + 1);
uct_rc_mlx5_ep_set_rdma_seg(raddr, remote_addr, rkey);
atomic = (void*)(raddr + 1);
if (opcode_flags == MLX5_OPCODE_ATOMIC_CS) {
atomic->compare = compare;
}
atomic->swap_add = swap_add;
uct_ib_mlx5_set_data_seg((void*)(atomic + 1), buffer, length, *lkey_p);
wqe_size = sizeof(*ctrl) + sizeof(*raddr) + sizeof(*atomic) +
sizeof(*dptr);
break;
case MLX5_OPCODE_ATOMIC_MASKED_CS:
raddr = (void*)(ctrl + 1);
uct_rc_mlx5_ep_set_rdma_seg(raddr, remote_addr, rkey);
switch (length) {
case sizeof(uint32_t):
opmod = UCT_IB_MLX5_OPMOD_EXT_ATOMIC(2);
masked_cswap32 = (void*)(raddr + 1);
masked_cswap32->swap = swap_add;
masked_cswap32->compare = compare;
masked_cswap32->swap_mask = (uint32_t)-1;
masked_cswap32->compare_mask = compare_mask;
dptr = (void*)(masked_cswap32 + 1);
wqe_size = sizeof(*ctrl) + sizeof(*raddr) +
sizeof(*masked_cswap32) + sizeof(*dptr);
break;
case sizeof(uint64_t):
opmod = UCT_IB_MLX5_OPMOD_EXT_ATOMIC(3); /* Ext. atomic, size 2**3 */
masked_cswap64 = (void*)(raddr + 1);
masked_cswap64->swap = swap_add;
masked_cswap64->compare = compare;
masked_cswap64->swap_mask = (uint64_t)-1;
masked_cswap64->compare_mask = compare_mask;
dptr = (void*)(masked_cswap64 + 1);
wqe_size = sizeof(*ctrl) + sizeof(*raddr) +
sizeof(*masked_cswap64) + sizeof(*dptr);
/* Handle QP wrap-around. It cannot happen in the middle of
* masked-cswap segment, because it's still in the first BB.
*/
ucs_assert((void*)dptr <= ep->tx.wq.qend);
if (dptr == ep->tx.wq.qend) {
dptr = ep->tx.wq.qstart;
} else {
ucs_assert((void*)masked_cswap64 < ep->tx.wq.qend);
}
break;
default:
ucs_assert(0);
}
uct_ib_mlx5_set_data_seg(dptr, buffer, length, *lkey_p);
break;
case MLX5_OPCODE_ATOMIC_MASKED_FA:
ucs_assert(length == sizeof(uint32_t));
raddr = (void*)(ctrl + 1);
uct_rc_mlx5_ep_set_rdma_seg(raddr, remote_addr, rkey);
opmod = UCT_IB_MLX5_OPMOD_EXT_ATOMIC(2);
masked_fadd32 = (void*)(raddr + 1);
masked_fadd32->add = swap_add;
masked_fadd32->filed_boundary = 0;
uct_ib_mlx5_set_data_seg((void*)(masked_fadd32 + 1), buffer, length,
*lkey_p);
wqe_size = sizeof(*ctrl) + sizeof(*raddr) +
sizeof(*masked_fadd32) + sizeof(*dptr);
break;
default:
ucs_fatal("invalid send opcode");
}
uct_rc_mlx5_post_send(ep, ctrl, (opcode_flags & UCT_RC_MLX5_OPCODE_MASK),
opmod, signal, wqe_size);
}
