| 1 |
/* SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0-only */ |
| 2 |
/* SPDX-FileCopyrightText: Copyright Amazon.com, Inc. or its affiliates. All rights reserved. */ |
| 3 |
|
| 4 |
#include "efa_unit_tests.h" |
| 5 |
|
| 6 |
static void test_efa_mr_impl(struct efa_domain *efa_domain, struct fid_mr *mr, |
| 7 |
int mr_reg_count, int mr_reg_size, bool gdrcopy_flag) |
| 8 |
{ |
| 9 |
struct efa_mr *efa_mr; |
| 10 |
|
| 11 |
assert_int_equal(efa_domain->ibv_mr_reg_ct, mr_reg_count); |
| 12 |
assert_int_equal(efa_domain->ibv_mr_reg_sz, mr_reg_size); |
| 13 |
|
| 14 |
if (mr) { |
| 15 |
efa_mr = container_of(mr, struct efa_mr, mr_fid); |
| 16 |
if (cuda_is_gdrcopy_enabled()) { |
| 17 |
if (gdrcopy_flag) |
| 18 |
assert_true(efa_mr->peer.flags & |
| 19 |
OFI_HMEM_DATA_DEV_REG_HANDLE); |
| 20 |
else |
| 21 |
assert_false(efa_mr->peer.flags & |
| 22 |
OFI_HMEM_DATA_DEV_REG_HANDLE); |
| 23 |
} |
| 24 |
} |
| 25 |
} |
| 26 |
|
| 27 |
void test_efa_rdm_mr_reg_host_memory(struct efa_resource **state) |
| 28 |
{ |
| 29 |
struct efa_resource *resource = *state; |
| 30 |
struct efa_domain *efa_domain; |
| 31 |
size_t mr_size = 64; |
| 32 |
void *buf; |
| 33 |
struct fid_mr *mr = NULL; |
| 34 |
|
| 35 |
efa_unit_test_resource_construct(resource, FI_EP_RDM, EFA_FABRIC_NAME); |
| 36 |
|
| 37 |
efa_domain = container_of(resource->domain, struct efa_domain, |
| 38 |
util_domain.domain_fid); |
| 39 |
|
| 40 |
buf = malloc(mr_size); |
| 41 |
assert_non_null(buf); |
| 42 |
|
| 43 |
test_efa_mr_impl(efa_domain, mr, 0, 0, false); |
| 44 |
|
| 45 |
assert_int_equal(fi_mr_reg(resource->domain, buf, mr_size, |
| 46 |
FI_SEND | FI_RECV, 0, 0, 0, &mr, NULL), |
| 47 |
0); |
| 48 |
|
| 49 |
/* No GDRCopy registration for host memory */ |
| 50 |
test_efa_mr_impl(efa_domain, mr, 1, mr_size, false); |
| 51 |
|
| 52 |
assert_int_equal(fi_close(&mr->fid), 0); |
| 53 |
test_efa_mr_impl(efa_domain, NULL, 0, 0, false); |
| 54 |
free(buf); |
| 55 |
} |
| 56 |
|
| 57 |
#if HAVE_CUDA |
| 58 |
void test_efa_rdm_mr_reg_cuda_memory(struct efa_resource **state) |
| 59 |
{ |
| 60 |
struct efa_resource *resource = *state; |
| 61 |
struct efa_domain *efa_domain; |
| 62 |
size_t mr_size = 64; |
| 63 |
void *buf; |
| 64 |
struct fid_mr *mr = NULL; |
| 65 |
struct fi_mr_attr mr_reg_attr = { 0 }; |
| 66 |
struct iovec iovec; |
| 67 |
int err; |
| 68 |
|
| 69 |
if (hmem_ops[FI_HMEM_CUDA].initialized && |
| 70 |
g_efa_hmem_info[FI_HMEM_CUDA].p2p_supported_by_device) { |
| 71 |
resource->hints = efa_unit_test_alloc_hints_hmem( |
| 72 |
FI_EP_RDM, EFA_FABRIC_NAME); |
| 73 |
efa_unit_test_resource_construct_with_hints(resource, FI_EP_RDM, |
| 74 |
FI_VERSION(2, 0), |
| 75 |
resource->hints, |
| 76 |
true, true); |
| 77 |
|
| 78 |
efa_domain = container_of(resource->domain, struct efa_domain, |
| 79 |
util_domain.domain_fid); |
| 80 |
test_efa_mr_impl(efa_domain, mr, 0, 0, false); |
| 81 |
|
| 82 |
err = ofi_cudaMalloc(&buf, mr_size); |
| 83 |
assert_int_equal(err, 0); |
| 84 |
assert_non_null(buf); |
| 85 |
|
| 86 |
mr_reg_attr.access = FI_SEND | FI_RECV; |
| 87 |
mr_reg_attr.iface = FI_HMEM_CUDA; |
| 88 |
iovec.iov_base = buf; |
| 89 |
iovec.iov_len = mr_size; |
| 90 |
mr_reg_attr.mr_iov = &iovec; |
| 91 |
mr_reg_attr.iov_count = 1; |
| 92 |
|
| 93 |
err = fi_mr_regattr(resource->domain, &mr_reg_attr, 0, &mr); |
| 94 |
assert_int_equal(err, 0); |
| 95 |
|
| 96 |
/* FI_MR_DMABUF flag was not set, so GDRCopy should be registered if available */ |
| 97 |
test_efa_mr_impl(efa_domain, mr, 1, mr_size, true); |
| 98 |
|
| 99 |
assert_int_equal(fi_close(&mr->fid), 0); |
| 100 |
test_efa_mr_impl(efa_domain, NULL, 0, 0, false); |
| 101 |
|
| 102 |
err = ofi_cudaFree(buf); |
| 103 |
assert_int_equal(err, 0); |
| 104 |
} |
| 105 |
} |
| 106 |
#else |
| 107 |
void test_efa_rdm_mr_reg_cuda_memory(struct efa_resource **state) |
| 108 |
{ |
| 109 |
skip(); |
| 110 |
} |
| 111 |
#endif |
| 112 |
|
| 113 |
#if HAVE_CUDA |
| 114 |
void test_efa_direct_mr_reg_no_gdrcopy(struct efa_resource **state) |
| 115 |
{ |
| 116 |
struct efa_resource *resource = *state; |
| 117 |
struct efa_domain *efa_domain; |
| 118 |
size_t mr_size = 64; |
| 119 |
void *buf; |
| 120 |
struct fid_mr *mr = NULL; |
| 121 |
struct fi_mr_attr mr_reg_attr = { 0 }; |
| 122 |
struct iovec iovec; |
| 123 |
int err; |
| 124 |
|
| 125 |
if (g_efa_hmem_info[FI_HMEM_CUDA].initialized && |
| 126 |
g_efa_hmem_info[FI_HMEM_CUDA].p2p_supported_by_device) { |
| 127 |
resource->hints = efa_unit_test_alloc_hints_hmem( |
| 128 |
FI_EP_RDM, EFA_DIRECT_FABRIC_NAME); |
| 129 |
efa_unit_test_resource_construct_with_hints(resource, FI_EP_RDM, |
| 130 |
FI_VERSION(2, 0), |
| 131 |
resource->hints, |
| 132 |
true, true); |
| 133 |
|
| 134 |
efa_domain = container_of(resource->domain, struct efa_domain, |
| 135 |
util_domain.domain_fid); |
| 136 |
test_efa_mr_impl(efa_domain, mr, 0, 0, false); |
| 137 |
|
| 138 |
err = ofi_cudaMalloc(&buf, mr_size); |
| 139 |
assert_int_equal(err, 0); |
| 140 |
assert_non_null(buf); |
| 141 |
|
| 142 |
mr_reg_attr.access = FI_SEND | FI_RECV; |
| 143 |
mr_reg_attr.iface = FI_HMEM_CUDA; |
| 144 |
iovec.iov_base = buf; |
| 145 |
iovec.iov_len = mr_size; |
| 146 |
mr_reg_attr.mr_iov = &iovec; |
| 147 |
mr_reg_attr.iov_count = 1; |
| 148 |
|
| 149 |
err = fi_mr_regattr(resource->domain, &mr_reg_attr, 0, &mr); |
| 150 |
assert_int_equal(err, 0); |
| 151 |
|
| 152 |
/* no GDRCopy in the efa-direct path */ |
| 153 |
test_efa_mr_impl(efa_domain, mr, 1, mr_size, false); |
| 154 |
|
| 155 |
assert_int_equal(fi_close(&mr->fid), 0); |
| 156 |
test_efa_mr_impl(efa_domain, NULL, 0, 0, false); |
| 157 |
|
| 158 |
err = ofi_cudaFree(buf); |
| 159 |
assert_int_equal(err, 0); |
| 160 |
} |
| 161 |
} |
| 162 |
#else |
| 163 |
void test_efa_direct_mr_reg_no_gdrcopy(struct efa_resource **state) |
| 164 |
{ |
| 165 |
skip(); |
| 166 |
} |
| 167 |
#endif |