1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright (c) Intel Corporation. All rights reserved. 3 * Copyright (c) 2019, 2021 Mellanox Technologies LTD. All rights reserved. 4 */ 5 6 #include "spdk/stdinc.h" 7 #include "spdk_cunit.h" 8 #include "common/lib/test_env.c" 9 #include "common/lib/test_rdma.c" 10 #include "nvmf/rdma.c" 11 #include "nvmf/transport.c" 12 13 #define RDMA_UT_UNITS_IN_MAX_IO 16 14 15 struct spdk_nvmf_transport_opts g_rdma_ut_transport_opts = { 16 .max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH, 17 .max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR, 18 .in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE, 19 .max_io_size = (SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE * RDMA_UT_UNITS_IN_MAX_IO), 20 .io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE, 21 .max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH, 22 .num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS, 23 }; 24 25 SPDK_LOG_REGISTER_COMPONENT(nvmf) 26 DEFINE_STUB(spdk_mem_map_set_translation, int, (struct spdk_mem_map *map, uint64_t vaddr, 27 uint64_t size, uint64_t translation), 0); 28 DEFINE_STUB(spdk_mem_map_clear_translation, int, (struct spdk_mem_map *map, uint64_t vaddr, 29 uint64_t size), 0); 30 DEFINE_STUB(spdk_mem_map_alloc, struct spdk_mem_map *, (uint64_t default_translation, 31 const struct spdk_mem_map_ops *ops, void *cb_ctx), NULL); 32 DEFINE_STUB(spdk_nvmf_qpair_disconnect, int, (struct spdk_nvmf_qpair *qpair, 33 nvmf_qpair_disconnect_cb cb_fn, void *ctx), 0); 34 DEFINE_STUB(spdk_nvmf_qpair_get_listen_trid, int, 35 (struct spdk_nvmf_qpair *qpair, struct spdk_nvme_transport_id *trid), 0); 36 DEFINE_STUB_V(spdk_mem_map_free, (struct spdk_mem_map **pmap)); 37 38 DEFINE_STUB_V(spdk_nvmf_ctrlr_data_init, (struct spdk_nvmf_transport_opts *opts, 39 struct spdk_nvmf_ctrlr_data *cdata)); 40 DEFINE_STUB_V(spdk_nvmf_request_exec, (struct spdk_nvmf_request *req)); 41 DEFINE_STUB(spdk_nvmf_request_complete, int, (struct spdk_nvmf_request *req), 0); 42 DEFINE_STUB(spdk_nvme_transport_id_compare, int, (const struct spdk_nvme_transport_id *trid1, 43 const struct spdk_nvme_transport_id *trid2), 0); 44 DEFINE_STUB_V(nvmf_ctrlr_abort_aer, (struct spdk_nvmf_ctrlr *ctrlr)); 45 DEFINE_STUB(spdk_nvmf_request_get_dif_ctx, bool, (struct spdk_nvmf_request *req, 46 struct spdk_dif_ctx *dif_ctx), false); 47 DEFINE_STUB_V(spdk_nvme_trid_populate_transport, (struct spdk_nvme_transport_id *trid, 48 enum spdk_nvme_transport_type trtype)); 49 DEFINE_STUB_V(spdk_nvmf_tgt_new_qpair, (struct spdk_nvmf_tgt *tgt, struct spdk_nvmf_qpair *qpair)); 50 DEFINE_STUB(nvmf_ctrlr_abort_request, int, (struct spdk_nvmf_request *req), 0); 51 DEFINE_STUB(spdk_nvme_transport_id_adrfam_str, const char *, (enum spdk_nvmf_adrfam adrfam), NULL); 52 DEFINE_STUB(ibv_dereg_mr, int, (struct ibv_mr *mr), 0); 53 DEFINE_STUB(ibv_resize_cq, int, (struct ibv_cq *cq, int cqe), 0); 54 DEFINE_STUB(spdk_mempool_lookup, struct spdk_mempool *, (const char *name), NULL); 55 56 /* ibv_reg_mr can be a macro, need to undefine it */ 57 #ifdef ibv_reg_mr 58 #undef ibv_reg_mr 59 #endif 60 61 DEFINE_RETURN_MOCK(ibv_reg_mr, struct ibv_mr *); 62 struct ibv_mr * 63 ibv_reg_mr(struct ibv_pd *pd, void *addr, size_t length, int access) 64 { 65 HANDLE_RETURN_MOCK(ibv_reg_mr); 66 if (length > 0) { 67 return &g_rdma_mr; 68 } else { 69 return NULL; 70 } 71 } 72 73 int 74 ibv_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr, 75 int attr_mask, struct ibv_qp_init_attr *init_attr) 76 { 77 if (qp == NULL) { 78 return -1; 79 } else { 80 attr->port_num = 80; 81 82 if (qp->state == IBV_QPS_ERR) { 83 attr->qp_state = 10; 84 } else { 85 attr->qp_state = IBV_QPS_INIT; 86 } 87 88 return 0; 89 } 90 } 91 92 const char * 93 spdk_nvme_transport_id_trtype_str(enum spdk_nvme_transport_type trtype) 94 { 95 switch (trtype) { 96 case SPDK_NVME_TRANSPORT_PCIE: 97 return "PCIe"; 98 case SPDK_NVME_TRANSPORT_RDMA: 99 return "RDMA"; 100 case SPDK_NVME_TRANSPORT_FC: 101 return "FC"; 102 default: 103 return NULL; 104 } 105 } 106 107 int 108 spdk_nvme_transport_id_populate_trstring(struct spdk_nvme_transport_id *trid, const char *trstring) 109 { 110 int len, i; 111 112 if (trstring == NULL) { 113 return -EINVAL; 114 } 115 116 len = strnlen(trstring, SPDK_NVMF_TRSTRING_MAX_LEN); 117 if (len == SPDK_NVMF_TRSTRING_MAX_LEN) { 118 return -EINVAL; 119 } 120 121 /* cast official trstring to uppercase version of input. */ 122 for (i = 0; i < len; i++) { 123 trid->trstring[i] = toupper(trstring[i]); 124 } 125 return 0; 126 } 127 128 static void 129 reset_nvmf_rdma_request(struct spdk_nvmf_rdma_request *rdma_req) 130 { 131 int i; 132 133 rdma_req->req.length = 0; 134 rdma_req->req.data_from_pool = false; 135 rdma_req->req.data = NULL; 136 rdma_req->data.wr.num_sge = 0; 137 rdma_req->data.wr.wr.rdma.remote_addr = 0; 138 rdma_req->data.wr.wr.rdma.rkey = 0; 139 rdma_req->offset = 0; 140 memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif)); 141 142 for (i = 0; i < SPDK_NVMF_MAX_SGL_ENTRIES; i++) { 143 rdma_req->req.iov[i].iov_base = 0; 144 rdma_req->req.iov[i].iov_len = 0; 145 rdma_req->req.buffers[i] = 0; 146 rdma_req->data.wr.sg_list[i].addr = 0; 147 rdma_req->data.wr.sg_list[i].length = 0; 148 rdma_req->data.wr.sg_list[i].lkey = 0; 149 } 150 rdma_req->req.iovcnt = 0; 151 if (rdma_req->req.stripped_data) { 152 free(rdma_req->req.stripped_data); 153 rdma_req->req.stripped_data = NULL; 154 } 155 } 156 157 static void 158 test_spdk_nvmf_rdma_request_parse_sgl(void) 159 { 160 struct spdk_nvmf_rdma_transport rtransport; 161 struct spdk_nvmf_rdma_device device; 162 struct spdk_nvmf_rdma_request rdma_req = {}; 163 struct spdk_nvmf_rdma_recv recv; 164 struct spdk_nvmf_rdma_poll_group group; 165 struct spdk_nvmf_rdma_qpair rqpair; 166 struct spdk_nvmf_rdma_poller poller; 167 union nvmf_c2h_msg cpl; 168 union nvmf_h2c_msg cmd; 169 struct spdk_nvme_sgl_descriptor *sgl; 170 struct spdk_nvmf_transport_pg_cache_buf bufs[4]; 171 struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}}; 172 struct spdk_nvmf_rdma_request_data data; 173 int rc, i; 174 uint32_t sgl_length; 175 uintptr_t aligned_buffer_address; 176 177 data.wr.sg_list = data.sgl; 178 STAILQ_INIT(&group.group.buf_cache); 179 group.group.buf_cache_size = 0; 180 group.group.buf_cache_count = 0; 181 group.group.transport = &rtransport.transport; 182 poller.group = &group; 183 rqpair.poller = &poller; 184 rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 185 186 sgl = &cmd.nvme_cmd.dptr.sgl1; 187 rdma_req.recv = &recv; 188 rdma_req.req.cmd = &cmd; 189 rdma_req.req.rsp = &cpl; 190 rdma_req.data.wr.sg_list = rdma_req.data.sgl; 191 rdma_req.req.qpair = &rqpair.qpair; 192 rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST; 193 194 rtransport.transport.opts = g_rdma_ut_transport_opts; 195 rtransport.data_wr_pool = NULL; 196 rtransport.transport.data_buf_pool = NULL; 197 198 device.attr.device_cap_flags = 0; 199 sgl->keyed.key = 0xEEEE; 200 sgl->address = 0xFFFF; 201 rdma_req.recv->buf = (void *)0xDDDD; 202 203 /* Test 1: sgl type: keyed data block subtype: address */ 204 sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 205 sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 206 207 /* Part 1: simple I/O, one SGL smaller than the transport io unit size */ 208 MOCK_SET(spdk_mempool_get, (void *)0x2000); 209 reset_nvmf_rdma_request(&rdma_req); 210 sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2; 211 212 device.map = (void *)0x0; 213 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 214 CU_ASSERT(rc == 0); 215 CU_ASSERT(rdma_req.req.data_from_pool == true); 216 CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2); 217 CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000); 218 CU_ASSERT(rdma_req.data.wr.num_sge == 1); 219 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 220 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 221 CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000); 222 CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000); 223 CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2); 224 CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY); 225 226 /* Part 2: simple I/O, one SGL larger than the transport io unit size (equal to the max io size) */ 227 reset_nvmf_rdma_request(&rdma_req); 228 sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO; 229 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 230 231 CU_ASSERT(rc == 0); 232 CU_ASSERT(rdma_req.req.data_from_pool == true); 233 CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO); 234 CU_ASSERT(rdma_req.data.wr.num_sge == RDMA_UT_UNITS_IN_MAX_IO); 235 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 236 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 237 for (i = 0; i < RDMA_UT_UNITS_IN_MAX_IO; i++) { 238 CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000); 239 CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000); 240 CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size); 241 CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY); 242 } 243 244 /* Part 3: simple I/O one SGL larger than the transport max io size */ 245 reset_nvmf_rdma_request(&rdma_req); 246 sgl->keyed.length = rtransport.transport.opts.max_io_size * 2; 247 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 248 249 CU_ASSERT(rc == -1); 250 251 /* Part 4: Pretend there are no buffer pools */ 252 MOCK_SET(spdk_mempool_get, NULL); 253 reset_nvmf_rdma_request(&rdma_req); 254 sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO; 255 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 256 257 CU_ASSERT(rc == 0); 258 CU_ASSERT(rdma_req.req.data_from_pool == false); 259 CU_ASSERT(rdma_req.req.data == NULL); 260 CU_ASSERT(rdma_req.data.wr.num_sge == 0); 261 CU_ASSERT(rdma_req.req.buffers[0] == NULL); 262 CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0); 263 CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 0); 264 CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == 0); 265 266 rdma_req.recv->buf = (void *)0xDDDD; 267 /* Test 2: sgl type: keyed data block subtype: offset (in capsule data) */ 268 sgl->generic.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK; 269 sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 270 271 /* Part 1: Normal I/O smaller than in capsule data size no offset */ 272 reset_nvmf_rdma_request(&rdma_req); 273 sgl->address = 0; 274 sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size; 275 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 276 277 CU_ASSERT(rc == 0); 278 CU_ASSERT(rdma_req.req.data == (void *)0xDDDD); 279 CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.in_capsule_data_size); 280 CU_ASSERT(rdma_req.req.data_from_pool == false); 281 282 /* Part 2: I/O offset + length too large */ 283 reset_nvmf_rdma_request(&rdma_req); 284 sgl->address = rtransport.transport.opts.in_capsule_data_size; 285 sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size; 286 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 287 288 CU_ASSERT(rc == -1); 289 290 /* Part 3: I/O too large */ 291 reset_nvmf_rdma_request(&rdma_req); 292 sgl->address = 0; 293 sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size * 2; 294 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 295 296 CU_ASSERT(rc == -1); 297 298 /* Test 3: Multi SGL */ 299 sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT; 300 sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 301 sgl->address = 0; 302 rdma_req.recv->buf = (void *)&sgl_desc; 303 MOCK_SET(spdk_mempool_get, &data); 304 305 /* part 1: 2 segments each with 1 wr. */ 306 reset_nvmf_rdma_request(&rdma_req); 307 sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor); 308 for (i = 0; i < 2; i++) { 309 sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 310 sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 311 sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size; 312 sgl_desc[i].address = 0x4000 + i * rtransport.transport.opts.io_unit_size; 313 sgl_desc[i].keyed.key = 0x44; 314 } 315 316 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 317 318 CU_ASSERT(rc == 0); 319 CU_ASSERT(rdma_req.req.data_from_pool == true); 320 CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 2); 321 CU_ASSERT(rdma_req.data.wr.num_sge == 1); 322 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44); 323 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000); 324 CU_ASSERT(rdma_req.data.wr.next == &data.wr); 325 CU_ASSERT(data.wr.wr.rdma.rkey == 0x44); 326 CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size); 327 CU_ASSERT(data.wr.num_sge == 1); 328 CU_ASSERT(data.wr.next == &rdma_req.rsp.wr); 329 330 /* part 2: 2 segments, each with 1 wr containing 8 sge_elements */ 331 reset_nvmf_rdma_request(&rdma_req); 332 sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor); 333 for (i = 0; i < 2; i++) { 334 sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 335 sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 336 sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size * 8; 337 sgl_desc[i].address = 0x4000 + i * 8 * rtransport.transport.opts.io_unit_size; 338 sgl_desc[i].keyed.key = 0x44; 339 } 340 341 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 342 343 CU_ASSERT(rc == 0); 344 CU_ASSERT(rdma_req.req.data_from_pool == true); 345 CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16); 346 CU_ASSERT(rdma_req.req.iovcnt == 16); 347 CU_ASSERT(rdma_req.data.wr.num_sge == 8); 348 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44); 349 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000); 350 CU_ASSERT(rdma_req.data.wr.next == &data.wr); 351 CU_ASSERT(data.wr.wr.rdma.rkey == 0x44); 352 CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 8); 353 CU_ASSERT(data.wr.num_sge == 8); 354 CU_ASSERT(data.wr.next == &rdma_req.rsp.wr); 355 356 /* part 3: 2 segments, one very large, one very small */ 357 reset_nvmf_rdma_request(&rdma_req); 358 for (i = 0; i < 2; i++) { 359 sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 360 sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 361 sgl_desc[i].keyed.key = 0x44; 362 } 363 364 sgl_desc[0].keyed.length = rtransport.transport.opts.io_unit_size * 15 + 365 rtransport.transport.opts.io_unit_size / 2; 366 sgl_desc[0].address = 0x4000; 367 sgl_desc[1].keyed.length = rtransport.transport.opts.io_unit_size / 2; 368 sgl_desc[1].address = 0x4000 + rtransport.transport.opts.io_unit_size * 15 + 369 rtransport.transport.opts.io_unit_size / 2; 370 371 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 372 373 CU_ASSERT(rc == 0); 374 CU_ASSERT(rdma_req.req.data_from_pool == true); 375 CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16); 376 CU_ASSERT(rdma_req.req.iovcnt == 16); 377 CU_ASSERT(rdma_req.data.wr.num_sge == 16); 378 for (i = 0; i < 15; i++) { 379 CU_ASSERT(rdma_req.data.sgl[i].length == rtransport.transport.opts.io_unit_size); 380 } 381 CU_ASSERT(rdma_req.data.sgl[15].length == rtransport.transport.opts.io_unit_size / 2); 382 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44); 383 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000); 384 CU_ASSERT(rdma_req.data.wr.next == &data.wr); 385 CU_ASSERT(data.wr.wr.rdma.rkey == 0x44); 386 CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 15 + 387 rtransport.transport.opts.io_unit_size / 2); 388 CU_ASSERT(data.sgl[0].length == rtransport.transport.opts.io_unit_size / 2); 389 CU_ASSERT(data.wr.num_sge == 1); 390 CU_ASSERT(data.wr.next == &rdma_req.rsp.wr); 391 392 /* part 4: 2 SGL descriptors, each length is transport buffer / 2 393 * 1 transport buffers should be allocated */ 394 reset_nvmf_rdma_request(&rdma_req); 395 aligned_buffer_address = ((uintptr_t)(&data) + NVMF_DATA_BUFFER_MASK) & ~NVMF_DATA_BUFFER_MASK; 396 sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor); 397 sgl_length = rtransport.transport.opts.io_unit_size / 2; 398 for (i = 0; i < 2; i++) { 399 sgl_desc[i].keyed.length = sgl_length; 400 sgl_desc[i].address = 0x4000 + i * sgl_length; 401 } 402 403 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 404 405 CU_ASSERT(rc == 0); 406 CU_ASSERT(rdma_req.req.data_from_pool == true); 407 CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size); 408 CU_ASSERT(rdma_req.req.iovcnt == 1); 409 410 CU_ASSERT(rdma_req.data.sgl[0].length == sgl_length); 411 /* We mocked mempool_get to return address of data variable. Mempool is used 412 * to get both additional WRs and data buffers, so data points to &data */ 413 CU_ASSERT(rdma_req.data.sgl[0].addr == aligned_buffer_address); 414 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44); 415 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000); 416 CU_ASSERT(rdma_req.data.wr.num_sge == 1); 417 CU_ASSERT(rdma_req.data.wr.next == &data.wr); 418 419 CU_ASSERT(data.wr.wr.rdma.rkey == 0x44); 420 CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + sgl_length); 421 CU_ASSERT(data.sgl[0].length == sgl_length); 422 CU_ASSERT(data.sgl[0].addr == aligned_buffer_address + sgl_length); 423 CU_ASSERT(data.wr.num_sge == 1); 424 425 /* Test 4: use PG buffer cache */ 426 sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 427 sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 428 sgl->address = 0xFFFF; 429 rdma_req.recv->buf = (void *)0xDDDD; 430 sgl->keyed.key = 0xEEEE; 431 432 for (i = 0; i < 4; i++) { 433 STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link); 434 } 435 436 /* part 1: use the four buffers from the pg cache */ 437 group.group.buf_cache_size = 4; 438 group.group.buf_cache_count = 4; 439 MOCK_SET(spdk_mempool_get, (void *)0x2000); 440 reset_nvmf_rdma_request(&rdma_req); 441 sgl->keyed.length = rtransport.transport.opts.io_unit_size * 4; 442 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 443 444 SPDK_CU_ASSERT_FATAL(rc == 0); 445 CU_ASSERT(rdma_req.req.data_from_pool == true); 446 CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4); 447 CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) & 448 ~NVMF_DATA_BUFFER_MASK)); 449 CU_ASSERT(rdma_req.data.wr.num_sge == 4); 450 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 451 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 452 CU_ASSERT(group.group.buf_cache_count == 0); 453 CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache)); 454 for (i = 0; i < 4; i++) { 455 CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]); 456 CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) & 457 ~NVMF_DATA_BUFFER_MASK)); 458 CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size); 459 } 460 461 /* part 2: now that we have used the buffers from the cache, try again. We should get mempool buffers. */ 462 reset_nvmf_rdma_request(&rdma_req); 463 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 464 465 SPDK_CU_ASSERT_FATAL(rc == 0); 466 CU_ASSERT(rdma_req.req.data_from_pool == true); 467 CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4); 468 CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000); 469 CU_ASSERT(rdma_req.data.wr.num_sge == 4); 470 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 471 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 472 CU_ASSERT(group.group.buf_cache_count == 0); 473 CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache)); 474 for (i = 0; i < 4; i++) { 475 CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000); 476 CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000); 477 CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size); 478 CU_ASSERT(group.group.buf_cache_count == 0); 479 } 480 481 /* part 3: half and half */ 482 group.group.buf_cache_count = 2; 483 484 for (i = 0; i < 2; i++) { 485 STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link); 486 } 487 reset_nvmf_rdma_request(&rdma_req); 488 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 489 490 SPDK_CU_ASSERT_FATAL(rc == 0); 491 CU_ASSERT(rdma_req.req.data_from_pool == true); 492 CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4); 493 CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) & 494 ~NVMF_DATA_BUFFER_MASK)); 495 CU_ASSERT(rdma_req.data.wr.num_sge == 4); 496 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 497 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 498 CU_ASSERT(group.group.buf_cache_count == 0); 499 for (i = 0; i < 2; i++) { 500 CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]); 501 CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) & 502 ~NVMF_DATA_BUFFER_MASK)); 503 CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size); 504 } 505 for (i = 2; i < 4; i++) { 506 CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000); 507 CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000); 508 CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size); 509 } 510 511 reset_nvmf_rdma_request(&rdma_req); 512 } 513 514 static struct spdk_nvmf_rdma_recv * 515 create_recv(struct spdk_nvmf_rdma_qpair *rqpair, enum spdk_nvme_nvm_opcode opc) 516 { 517 struct spdk_nvmf_rdma_recv *rdma_recv; 518 union nvmf_h2c_msg *cmd; 519 struct spdk_nvme_sgl_descriptor *sgl; 520 521 rdma_recv = calloc(1, sizeof(*rdma_recv)); 522 rdma_recv->qpair = rqpair; 523 cmd = calloc(1, sizeof(*cmd)); 524 rdma_recv->sgl[0].addr = (uintptr_t)cmd; 525 cmd->nvme_cmd.opc = opc; 526 sgl = &cmd->nvme_cmd.dptr.sgl1; 527 sgl->keyed.key = 0xEEEE; 528 sgl->address = 0xFFFF; 529 sgl->keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 530 sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 531 sgl->keyed.length = 1; 532 533 return rdma_recv; 534 } 535 536 static void 537 free_recv(struct spdk_nvmf_rdma_recv *rdma_recv) 538 { 539 free((void *)rdma_recv->sgl[0].addr); 540 free(rdma_recv); 541 } 542 543 static struct spdk_nvmf_rdma_request * 544 create_req(struct spdk_nvmf_rdma_qpair *rqpair, 545 struct spdk_nvmf_rdma_recv *rdma_recv) 546 { 547 struct spdk_nvmf_rdma_request *rdma_req; 548 union nvmf_c2h_msg *cpl; 549 550 rdma_req = calloc(1, sizeof(*rdma_req)); 551 rdma_req->recv = rdma_recv; 552 rdma_req->req.qpair = &rqpair->qpair; 553 rdma_req->state = RDMA_REQUEST_STATE_NEW; 554 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr; 555 rdma_req->data.wr.sg_list = rdma_req->data.sgl; 556 cpl = calloc(1, sizeof(*cpl)); 557 rdma_req->rsp.sgl[0].addr = (uintptr_t)cpl; 558 rdma_req->req.rsp = cpl; 559 560 return rdma_req; 561 } 562 563 static void 564 free_req(struct spdk_nvmf_rdma_request *rdma_req) 565 { 566 free((void *)rdma_req->rsp.sgl[0].addr); 567 free(rdma_req); 568 } 569 570 static void 571 qpair_reset(struct spdk_nvmf_rdma_qpair *rqpair, 572 struct spdk_nvmf_rdma_poller *poller, 573 struct spdk_nvmf_rdma_device *device, 574 struct spdk_nvmf_rdma_resources *resources, 575 struct spdk_nvmf_transport *transport) 576 { 577 memset(rqpair, 0, sizeof(*rqpair)); 578 STAILQ_INIT(&rqpair->pending_rdma_write_queue); 579 STAILQ_INIT(&rqpair->pending_rdma_read_queue); 580 rqpair->poller = poller; 581 rqpair->device = device; 582 rqpair->resources = resources; 583 rqpair->qpair.qid = 1; 584 rqpair->ibv_state = IBV_QPS_RTS; 585 rqpair->qpair.state = SPDK_NVMF_QPAIR_ACTIVE; 586 rqpair->max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 587 rqpair->max_send_depth = 16; 588 rqpair->max_read_depth = 16; 589 rqpair->qpair.transport = transport; 590 } 591 592 static void 593 poller_reset(struct spdk_nvmf_rdma_poller *poller, 594 struct spdk_nvmf_rdma_poll_group *group) 595 { 596 memset(poller, 0, sizeof(*poller)); 597 STAILQ_INIT(&poller->qpairs_pending_recv); 598 STAILQ_INIT(&poller->qpairs_pending_send); 599 poller->group = group; 600 } 601 602 static void 603 test_spdk_nvmf_rdma_request_process(void) 604 { 605 struct spdk_nvmf_rdma_transport rtransport = {}; 606 struct spdk_nvmf_rdma_poll_group group = {}; 607 struct spdk_nvmf_rdma_poller poller = {}; 608 struct spdk_nvmf_rdma_device device = {}; 609 struct spdk_nvmf_rdma_resources resources = {}; 610 struct spdk_nvmf_rdma_qpair rqpair = {}; 611 struct spdk_nvmf_rdma_recv *rdma_recv; 612 struct spdk_nvmf_rdma_request *rdma_req; 613 bool progress; 614 615 STAILQ_INIT(&group.group.buf_cache); 616 STAILQ_INIT(&group.group.pending_buf_queue); 617 group.group.buf_cache_size = 0; 618 group.group.buf_cache_count = 0; 619 poller_reset(&poller, &group); 620 qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport); 621 622 rtransport.transport.opts = g_rdma_ut_transport_opts; 623 rtransport.transport.data_buf_pool = spdk_mempool_create("test_data_pool", 16, 128, 0, 0); 624 rtransport.data_wr_pool = spdk_mempool_create("test_wr_pool", 128, 625 sizeof(struct spdk_nvmf_rdma_request_data), 626 0, 0); 627 MOCK_CLEAR(spdk_mempool_get); 628 629 device.attr.device_cap_flags = 0; 630 device.map = (void *)0x0; 631 632 /* Test 1: single SGL READ request */ 633 rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_READ); 634 rdma_req = create_req(&rqpair, rdma_recv); 635 rqpair.current_recv_depth = 1; 636 /* NEW -> EXECUTING */ 637 progress = nvmf_rdma_request_process(&rtransport, rdma_req); 638 CU_ASSERT(progress == true); 639 CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING); 640 CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST); 641 /* EXECUTED -> TRANSFERRING_C2H */ 642 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 643 progress = nvmf_rdma_request_process(&rtransport, rdma_req); 644 CU_ASSERT(progress == true); 645 CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST); 646 CU_ASSERT(rdma_req->recv == NULL); 647 /* COMPLETED -> FREE */ 648 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 649 progress = nvmf_rdma_request_process(&rtransport, rdma_req); 650 CU_ASSERT(progress == true); 651 CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE); 652 653 free_recv(rdma_recv); 654 free_req(rdma_req); 655 poller_reset(&poller, &group); 656 qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport); 657 658 /* Test 2: single SGL WRITE request */ 659 rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE); 660 rdma_req = create_req(&rqpair, rdma_recv); 661 rqpair.current_recv_depth = 1; 662 /* NEW -> TRANSFERRING_H2C */ 663 progress = nvmf_rdma_request_process(&rtransport, rdma_req); 664 CU_ASSERT(progress == true); 665 CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 666 CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER); 667 STAILQ_INIT(&poller.qpairs_pending_send); 668 /* READY_TO_EXECUTE -> EXECUTING */ 669 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 670 progress = nvmf_rdma_request_process(&rtransport, rdma_req); 671 CU_ASSERT(progress == true); 672 CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING); 673 /* EXECUTED -> COMPLETING */ 674 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED; 675 progress = nvmf_rdma_request_process(&rtransport, rdma_req); 676 CU_ASSERT(progress == true); 677 CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING); 678 CU_ASSERT(rdma_req->recv == NULL); 679 /* COMPLETED -> FREE */ 680 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED; 681 progress = nvmf_rdma_request_process(&rtransport, rdma_req); 682 CU_ASSERT(progress == true); 683 CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE); 684 685 free_recv(rdma_recv); 686 free_req(rdma_req); 687 poller_reset(&poller, &group); 688 qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport); 689 690 /* Test 3: WRITE+WRITE ibv_send batching */ 691 { 692 struct spdk_nvmf_rdma_recv *recv1, *recv2; 693 struct spdk_nvmf_rdma_request *req1, *req2; 694 recv1 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE); 695 req1 = create_req(&rqpair, recv1); 696 recv2 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE); 697 req2 = create_req(&rqpair, recv2); 698 699 /* WRITE 1: NEW -> TRANSFERRING_H2C */ 700 rqpair.current_recv_depth = 1; 701 nvmf_rdma_request_process(&rtransport, req1); 702 CU_ASSERT(req1->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 703 704 /* WRITE 2: NEW -> TRANSFERRING_H2C */ 705 rqpair.current_recv_depth = 2; 706 nvmf_rdma_request_process(&rtransport, req2); 707 CU_ASSERT(req2->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER); 708 709 STAILQ_INIT(&poller.qpairs_pending_send); 710 711 /* WRITE 1 completes before WRITE 2 has finished RDMA reading */ 712 /* WRITE 1: READY_TO_EXECUTE -> EXECUTING */ 713 req1->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 714 nvmf_rdma_request_process(&rtransport, req1); 715 CU_ASSERT(req1->state == RDMA_REQUEST_STATE_EXECUTING); 716 /* WRITE 1: EXECUTED -> COMPLETING */ 717 req1->state = RDMA_REQUEST_STATE_EXECUTED; 718 nvmf_rdma_request_process(&rtransport, req1); 719 CU_ASSERT(req1->state == RDMA_REQUEST_STATE_COMPLETING); 720 STAILQ_INIT(&poller.qpairs_pending_send); 721 /* WRITE 1: COMPLETED -> FREE */ 722 req1->state = RDMA_REQUEST_STATE_COMPLETED; 723 nvmf_rdma_request_process(&rtransport, req1); 724 CU_ASSERT(req1->state == RDMA_REQUEST_STATE_FREE); 725 726 /* Now WRITE 2 has finished reading and completes */ 727 /* WRITE 2: COMPLETED -> FREE */ 728 /* WRITE 2: READY_TO_EXECUTE -> EXECUTING */ 729 req2->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE; 730 nvmf_rdma_request_process(&rtransport, req2); 731 CU_ASSERT(req2->state == RDMA_REQUEST_STATE_EXECUTING); 732 /* WRITE 1: EXECUTED -> COMPLETING */ 733 req2->state = RDMA_REQUEST_STATE_EXECUTED; 734 nvmf_rdma_request_process(&rtransport, req2); 735 CU_ASSERT(req2->state == RDMA_REQUEST_STATE_COMPLETING); 736 STAILQ_INIT(&poller.qpairs_pending_send); 737 /* WRITE 1: COMPLETED -> FREE */ 738 req2->state = RDMA_REQUEST_STATE_COMPLETED; 739 nvmf_rdma_request_process(&rtransport, req2); 740 CU_ASSERT(req2->state == RDMA_REQUEST_STATE_FREE); 741 742 free_recv(recv1); 743 free_req(req1); 744 free_recv(recv2); 745 free_req(req2); 746 poller_reset(&poller, &group); 747 qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport); 748 } 749 750 /* Test 4, invalid command, check xfer type */ 751 { 752 struct spdk_nvmf_rdma_recv *rdma_recv_inv; 753 struct spdk_nvmf_rdma_request *rdma_req_inv; 754 /* construct an opcode that specifies BIDIRECTIONAL transfer */ 755 uint8_t opc = 0x10 | SPDK_NVME_DATA_BIDIRECTIONAL; 756 757 rdma_recv_inv = create_recv(&rqpair, opc); 758 rdma_req_inv = create_req(&rqpair, rdma_recv_inv); 759 760 /* NEW -> RDMA_REQUEST_STATE_COMPLETING */ 761 rqpair.current_recv_depth = 1; 762 progress = nvmf_rdma_request_process(&rtransport, rdma_req_inv); 763 CU_ASSERT(progress == true); 764 CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_COMPLETING); 765 CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sct == SPDK_NVME_SCT_GENERIC); 766 CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sc == SPDK_NVME_SC_INVALID_OPCODE); 767 768 /* RDMA_REQUEST_STATE_COMPLETED -> FREE */ 769 rdma_req_inv->state = RDMA_REQUEST_STATE_COMPLETED; 770 nvmf_rdma_request_process(&rtransport, rdma_req_inv); 771 CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_FREE); 772 773 free_recv(rdma_recv_inv); 774 free_req(rdma_req_inv); 775 poller_reset(&poller, &group); 776 qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport); 777 } 778 779 spdk_mempool_free(rtransport.transport.data_buf_pool); 780 spdk_mempool_free(rtransport.data_wr_pool); 781 } 782 783 #define TEST_GROUPS_COUNT 5 784 static void 785 test_nvmf_rdma_get_optimal_poll_group(void) 786 { 787 struct spdk_nvmf_rdma_transport rtransport = {}; 788 struct spdk_nvmf_transport *transport = &rtransport.transport; 789 struct spdk_nvmf_rdma_qpair rqpair = {}; 790 struct spdk_nvmf_transport_poll_group *groups[TEST_GROUPS_COUNT]; 791 struct spdk_nvmf_rdma_poll_group *rgroups[TEST_GROUPS_COUNT]; 792 struct spdk_nvmf_transport_poll_group *result; 793 struct spdk_nvmf_poll_group group = {}; 794 uint32_t i; 795 796 rqpair.qpair.transport = transport; 797 TAILQ_INIT(&rtransport.poll_groups); 798 799 for (i = 0; i < TEST_GROUPS_COUNT; i++) { 800 groups[i] = nvmf_rdma_poll_group_create(transport, NULL); 801 CU_ASSERT(groups[i] != NULL); 802 groups[i]->group = &group; 803 rgroups[i] = SPDK_CONTAINEROF(groups[i], struct spdk_nvmf_rdma_poll_group, group); 804 groups[i]->transport = transport; 805 } 806 CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[0]); 807 CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[0]); 808 809 /* Emulate connection of %TEST_GROUPS_COUNT% initiators - each creates 1 admin and 1 io qp */ 810 for (i = 0; i < TEST_GROUPS_COUNT; i++) { 811 rqpair.qpair.qid = 0; 812 result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair); 813 CU_ASSERT(result == groups[i]); 814 CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]); 815 CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i]); 816 817 rqpair.qpair.qid = 1; 818 result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair); 819 CU_ASSERT(result == groups[i]); 820 CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]); 821 CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]); 822 } 823 /* wrap around, admin/io pg point to the first pg 824 Destroy all poll groups except of the last one */ 825 for (i = 0; i < TEST_GROUPS_COUNT - 1; i++) { 826 nvmf_rdma_poll_group_destroy(groups[i]); 827 CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[i + 1]); 828 CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i + 1]); 829 } 830 831 CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]); 832 CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]); 833 834 /* Check that pointers to the next admin/io poll groups are not changed */ 835 rqpair.qpair.qid = 0; 836 result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair); 837 CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]); 838 CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]); 839 CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]); 840 841 rqpair.qpair.qid = 1; 842 result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair); 843 CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]); 844 CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]); 845 CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]); 846 847 /* Remove the last poll group, check that pointers are NULL */ 848 nvmf_rdma_poll_group_destroy(groups[TEST_GROUPS_COUNT - 1]); 849 CU_ASSERT(rtransport.conn_sched.next_admin_pg == NULL); 850 CU_ASSERT(rtransport.conn_sched.next_io_pg == NULL); 851 852 /* Request optimal poll group, result must be NULL */ 853 rqpair.qpair.qid = 0; 854 result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair); 855 CU_ASSERT(result == NULL); 856 857 rqpair.qpair.qid = 1; 858 result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair); 859 CU_ASSERT(result == NULL); 860 } 861 #undef TEST_GROUPS_COUNT 862 863 static void 864 test_spdk_nvmf_rdma_request_parse_sgl_with_md(void) 865 { 866 struct spdk_nvmf_rdma_transport rtransport; 867 struct spdk_nvmf_rdma_device device; 868 struct spdk_nvmf_rdma_request rdma_req = {}; 869 struct spdk_nvmf_rdma_recv recv; 870 struct spdk_nvmf_rdma_poll_group group; 871 struct spdk_nvmf_rdma_qpair rqpair; 872 struct spdk_nvmf_rdma_poller poller; 873 union nvmf_c2h_msg cpl; 874 union nvmf_h2c_msg cmd; 875 struct spdk_nvme_sgl_descriptor *sgl; 876 struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}}; 877 char data_buffer[8192]; 878 struct spdk_nvmf_rdma_request_data *data = (struct spdk_nvmf_rdma_request_data *)data_buffer; 879 char data2_buffer[8192]; 880 struct spdk_nvmf_rdma_request_data *data2 = (struct spdk_nvmf_rdma_request_data *)data2_buffer; 881 const uint32_t data_bs = 512; 882 const uint32_t md_size = 8; 883 int rc, i; 884 void *aligned_buffer; 885 886 data->wr.sg_list = data->sgl; 887 STAILQ_INIT(&group.group.buf_cache); 888 group.group.buf_cache_size = 0; 889 group.group.buf_cache_count = 0; 890 group.group.transport = &rtransport.transport; 891 poller.group = &group; 892 rqpair.poller = &poller; 893 rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES; 894 895 sgl = &cmd.nvme_cmd.dptr.sgl1; 896 rdma_req.recv = &recv; 897 rdma_req.req.cmd = &cmd; 898 rdma_req.req.rsp = &cpl; 899 rdma_req.data.wr.sg_list = rdma_req.data.sgl; 900 rdma_req.req.qpair = &rqpair.qpair; 901 rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST; 902 903 rtransport.transport.opts = g_rdma_ut_transport_opts; 904 rtransport.data_wr_pool = NULL; 905 rtransport.transport.data_buf_pool = NULL; 906 907 device.attr.device_cap_flags = 0; 908 device.map = NULL; 909 sgl->keyed.key = 0xEEEE; 910 sgl->address = 0xFFFF; 911 rdma_req.recv->buf = (void *)0xDDDD; 912 913 /* Test 1: sgl type: keyed data block subtype: address */ 914 sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 915 sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 916 917 /* Part 1: simple I/O, one SGL smaller than the transport io unit size, block size 512 */ 918 MOCK_SET(spdk_mempool_get, (void *)0x2000); 919 reset_nvmf_rdma_request(&rdma_req); 920 spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false, 921 SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 922 0, 0, 0, 0, 0); 923 rdma_req.req.dif_enabled = true; 924 rtransport.transport.opts.io_unit_size = data_bs * 8; 925 rdma_req.req.qpair->transport = &rtransport.transport; 926 sgl->keyed.length = data_bs * 4; 927 928 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 929 930 CU_ASSERT(rc == 0); 931 CU_ASSERT(rdma_req.req.data_from_pool == true); 932 CU_ASSERT(rdma_req.req.length == data_bs * 4); 933 CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length); 934 CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4); 935 CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000); 936 CU_ASSERT(rdma_req.data.wr.num_sge == 1); 937 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 938 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 939 CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000); 940 941 CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000); 942 CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rdma_req.req.length); 943 CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY); 944 945 /* Part 2: simple I/O, one SGL equal to io unit size, io_unit_size is not aligned with md_size, 946 block size 512 */ 947 MOCK_SET(spdk_mempool_get, (void *)0x2000); 948 reset_nvmf_rdma_request(&rdma_req); 949 spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false, 950 SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 951 0, 0, 0, 0, 0); 952 rdma_req.req.dif_enabled = true; 953 rtransport.transport.opts.io_unit_size = data_bs * 4; 954 sgl->keyed.length = data_bs * 4; 955 956 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 957 958 CU_ASSERT(rc == 0); 959 CU_ASSERT(rdma_req.req.data_from_pool == true); 960 CU_ASSERT(rdma_req.req.length == data_bs * 4); 961 CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length); 962 CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4); 963 CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000); 964 CU_ASSERT(rdma_req.data.wr.num_sge == 5); 965 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 966 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 967 CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000); 968 969 for (i = 0; i < 3; ++i) { 970 CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size)); 971 CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs); 972 CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY); 973 } 974 CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size)); 975 CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488); 976 CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY); 977 978 /* 2nd buffer consumed */ 979 CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000); 980 CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24); 981 CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY); 982 983 /* Part 3: simple I/O, one SGL equal io unit size, io_unit_size is equal to block size 512 bytes */ 984 MOCK_SET(spdk_mempool_get, (void *)0x2000); 985 reset_nvmf_rdma_request(&rdma_req); 986 spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false, 987 SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 988 0, 0, 0, 0, 0); 989 rdma_req.req.dif_enabled = true; 990 rtransport.transport.opts.io_unit_size = data_bs; 991 sgl->keyed.length = data_bs; 992 993 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 994 995 CU_ASSERT(rc == 0); 996 CU_ASSERT(rdma_req.req.data_from_pool == true); 997 CU_ASSERT(rdma_req.req.length == data_bs); 998 CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length); 999 CU_ASSERT(rdma_req.req.dif.elba_length == data_bs + md_size); 1000 CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000); 1001 CU_ASSERT(rdma_req.data.wr.num_sge == 1); 1002 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 1003 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 1004 CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000); 1005 1006 CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000); 1007 CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs); 1008 CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY); 1009 1010 CU_ASSERT(rdma_req.req.iovcnt == 2); 1011 CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)((unsigned long)0x2000)); 1012 CU_ASSERT(rdma_req.req.iov[0].iov_len == data_bs); 1013 /* 2nd buffer consumed for metadata */ 1014 CU_ASSERT(rdma_req.req.iov[1].iov_base == (void *)((unsigned long)0x2000)); 1015 CU_ASSERT(rdma_req.req.iov[1].iov_len == md_size); 1016 1017 /* Part 4: simple I/O, one SGL equal io unit size, io_unit_size is aligned with md_size, 1018 block size 512 */ 1019 MOCK_SET(spdk_mempool_get, (void *)0x2000); 1020 reset_nvmf_rdma_request(&rdma_req); 1021 spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false, 1022 SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 1023 0, 0, 0, 0, 0); 1024 rdma_req.req.dif_enabled = true; 1025 rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4; 1026 sgl->keyed.length = data_bs * 4; 1027 1028 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 1029 1030 CU_ASSERT(rc == 0); 1031 CU_ASSERT(rdma_req.req.data_from_pool == true); 1032 CU_ASSERT(rdma_req.req.length == data_bs * 4); 1033 CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length); 1034 CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4); 1035 CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000); 1036 CU_ASSERT(rdma_req.data.wr.num_sge == 1); 1037 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 1038 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 1039 CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000); 1040 1041 CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000); 1042 CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rdma_req.req.length); 1043 CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY); 1044 1045 /* Part 5: simple I/O, one SGL equal to 2x io unit size, io_unit_size is aligned with md_size, 1046 block size 512 */ 1047 MOCK_SET(spdk_mempool_get, (void *)0x2000); 1048 reset_nvmf_rdma_request(&rdma_req); 1049 spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false, 1050 SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 1051 0, 0, 0, 0, 0); 1052 rdma_req.req.dif_enabled = true; 1053 rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 2; 1054 sgl->keyed.length = data_bs * 4; 1055 1056 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 1057 1058 CU_ASSERT(rc == 0); 1059 CU_ASSERT(rdma_req.req.data_from_pool == true); 1060 CU_ASSERT(rdma_req.req.length == data_bs * 4); 1061 CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length); 1062 CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4); 1063 CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000); 1064 CU_ASSERT(rdma_req.data.wr.num_sge == 2); 1065 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 1066 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 1067 CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000); 1068 1069 for (i = 0; i < 2; ++i) { 1070 CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000); 1071 CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs * 2); 1072 } 1073 1074 /* Part 6: simple I/O, one SGL larger than the transport io unit size, io_unit_size is not aligned to md_size, 1075 block size 512 */ 1076 MOCK_SET(spdk_mempool_get, (void *)0x2000); 1077 reset_nvmf_rdma_request(&rdma_req); 1078 spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false, 1079 SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 1080 0, 0, 0, 0, 0); 1081 rdma_req.req.dif_enabled = true; 1082 rtransport.transport.opts.io_unit_size = data_bs * 4; 1083 sgl->keyed.length = data_bs * 6; 1084 1085 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 1086 1087 CU_ASSERT(rc == 0); 1088 CU_ASSERT(rdma_req.req.data_from_pool == true); 1089 CU_ASSERT(rdma_req.req.length == data_bs * 6); 1090 CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length); 1091 CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 6); 1092 CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000); 1093 CU_ASSERT(rdma_req.data.wr.num_sge == 7); 1094 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 1095 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 1096 CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000); 1097 1098 for (i = 0; i < 3; ++i) { 1099 CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size)); 1100 CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs); 1101 CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY); 1102 } 1103 CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size)); 1104 CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488); 1105 CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY); 1106 1107 /* 2nd IO buffer consumed */ 1108 CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000); 1109 CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24); 1110 CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY); 1111 1112 CU_ASSERT(rdma_req.data.wr.sg_list[5].addr == 0x2000 + 24 + md_size); 1113 CU_ASSERT(rdma_req.data.wr.sg_list[5].length == 512); 1114 CU_ASSERT(rdma_req.data.wr.sg_list[5].lkey == RDMA_UT_LKEY); 1115 1116 CU_ASSERT(rdma_req.data.wr.sg_list[6].addr == 0x2000 + 24 + 512 + md_size * 2); 1117 CU_ASSERT(rdma_req.data.wr.sg_list[6].length == 512); 1118 CU_ASSERT(rdma_req.data.wr.sg_list[6].lkey == RDMA_UT_LKEY); 1119 1120 /* Part 7: simple I/O, number of SGL entries exceeds the number of entries 1121 one WR can hold. Additional WR is chained */ 1122 MOCK_SET(spdk_mempool_get, data2_buffer); 1123 aligned_buffer = (void *)((uintptr_t)(data2_buffer + NVMF_DATA_BUFFER_MASK) & 1124 ~NVMF_DATA_BUFFER_MASK); 1125 reset_nvmf_rdma_request(&rdma_req); 1126 spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false, 1127 SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 1128 0, 0, 0, 0, 0); 1129 rdma_req.req.dif_enabled = true; 1130 rtransport.transport.opts.io_unit_size = data_bs * 16; 1131 sgl->keyed.length = data_bs * 16; 1132 1133 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 1134 1135 CU_ASSERT(rc == 0); 1136 CU_ASSERT(rdma_req.req.data_from_pool == true); 1137 CU_ASSERT(rdma_req.req.length == data_bs * 16); 1138 CU_ASSERT(rdma_req.req.iovcnt == 2); 1139 CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length); 1140 CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 16); 1141 CU_ASSERT(rdma_req.req.data == aligned_buffer); 1142 CU_ASSERT(rdma_req.data.wr.num_sge == 16); 1143 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 1144 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 1145 1146 for (i = 0; i < 15; ++i) { 1147 CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)aligned_buffer + i * (data_bs + md_size)); 1148 CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs); 1149 CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY); 1150 } 1151 1152 /* 8192 - (512 + 8) * 15 = 392 */ 1153 CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)aligned_buffer + i * (data_bs + md_size)); 1154 CU_ASSERT(rdma_req.data.wr.sg_list[i].length == 392); 1155 CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY); 1156 1157 /* additional wr from pool */ 1158 CU_ASSERT(rdma_req.data.wr.next == (void *)&data2->wr); 1159 CU_ASSERT(rdma_req.data.wr.next->num_sge == 1); 1160 CU_ASSERT(rdma_req.data.wr.next->next == &rdma_req.rsp.wr); 1161 /* 2nd IO buffer */ 1162 CU_ASSERT(data2->wr.sg_list[0].addr == (uintptr_t)aligned_buffer); 1163 CU_ASSERT(data2->wr.sg_list[0].length == 120); 1164 CU_ASSERT(data2->wr.sg_list[0].lkey == RDMA_UT_LKEY); 1165 1166 /* Part 8: simple I/O, data with metadata do not fit to 1 io_buffer */ 1167 MOCK_SET(spdk_mempool_get, (void *)0x2000); 1168 reset_nvmf_rdma_request(&rdma_req); 1169 spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false, 1170 SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 1171 0, 0, 0, 0, 0); 1172 rdma_req.req.dif_enabled = true; 1173 rtransport.transport.opts.io_unit_size = 516; 1174 sgl->keyed.length = data_bs * 2; 1175 1176 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 1177 1178 CU_ASSERT(rc == 0); 1179 CU_ASSERT(rdma_req.req.data_from_pool == true); 1180 CU_ASSERT(rdma_req.req.length == data_bs * 2); 1181 CU_ASSERT(rdma_req.req.iovcnt == 3); 1182 CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length); 1183 CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 2); 1184 CU_ASSERT(rdma_req.req.data == (void *)0x2000); 1185 CU_ASSERT(rdma_req.data.wr.num_sge == 2); 1186 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE); 1187 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF); 1188 1189 CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000); 1190 CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 512); 1191 CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY); 1192 1193 /* 2nd IO buffer consumed, offset 4 bytes due to part of the metadata 1194 is located at the beginning of that buffer */ 1195 CU_ASSERT(rdma_req.data.wr.sg_list[1].addr == 0x2000 + 4); 1196 CU_ASSERT(rdma_req.data.wr.sg_list[1].length == 512); 1197 CU_ASSERT(rdma_req.data.wr.sg_list[1].lkey == RDMA_UT_LKEY); 1198 1199 /* Test 2: Multi SGL */ 1200 sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT; 1201 sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET; 1202 sgl->address = 0; 1203 rdma_req.recv->buf = (void *)&sgl_desc; 1204 MOCK_SET(spdk_mempool_get, data_buffer); 1205 aligned_buffer = (void *)((uintptr_t)(data_buffer + NVMF_DATA_BUFFER_MASK) & 1206 ~NVMF_DATA_BUFFER_MASK); 1207 1208 /* part 1: 2 segments each with 1 wr. io_unit_size is aligned with data_bs + md_size */ 1209 reset_nvmf_rdma_request(&rdma_req); 1210 spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false, 1211 SPDK_DIF_TYPE1, 1212 SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 0, 0, 0, 0, 0); 1213 rdma_req.req.dif_enabled = true; 1214 rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4; 1215 sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor); 1216 1217 for (i = 0; i < 2; i++) { 1218 sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK; 1219 sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS; 1220 sgl_desc[i].keyed.length = data_bs * 4; 1221 sgl_desc[i].address = 0x4000 + i * data_bs * 4; 1222 sgl_desc[i].keyed.key = 0x44; 1223 } 1224 1225 rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req); 1226 1227 CU_ASSERT(rc == 0); 1228 CU_ASSERT(rdma_req.req.data_from_pool == true); 1229 CU_ASSERT(rdma_req.req.length == data_bs * 4 * 2); 1230 CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length); 1231 CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4 * 2); 1232 CU_ASSERT(rdma_req.data.wr.num_sge == 1); 1233 CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == (uintptr_t)(aligned_buffer)); 1234 CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs * 4); 1235 1236 CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44); 1237 CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000); 1238 CU_ASSERT(rdma_req.data.wr.next == &data->wr); 1239 CU_ASSERT(data->wr.wr.rdma.rkey == 0x44); 1240 CU_ASSERT(data->wr.wr.rdma.remote_addr == 0x4000 + data_bs * 4); 1241 CU_ASSERT(data->wr.num_sge == 1); 1242 CU_ASSERT(data->wr.sg_list[0].addr == (uintptr_t)(aligned_buffer)); 1243 CU_ASSERT(data->wr.sg_list[0].length == data_bs * 4); 1244 1245 CU_ASSERT(data->wr.next == &rdma_req.rsp.wr); 1246 reset_nvmf_rdma_request(&rdma_req); 1247 } 1248 1249 static void 1250 test_nvmf_rdma_opts_init(void) 1251 { 1252 struct spdk_nvmf_transport_opts opts = {}; 1253 1254 nvmf_rdma_opts_init(&opts); 1255 CU_ASSERT(opts.max_queue_depth == SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH); 1256 CU_ASSERT(opts.max_qpairs_per_ctrlr == SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR); 1257 CU_ASSERT(opts.in_capsule_data_size == SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE); 1258 CU_ASSERT(opts.max_io_size == SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE); 1259 CU_ASSERT(opts.io_unit_size == SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE); 1260 CU_ASSERT(opts.max_aq_depth == SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH); 1261 CU_ASSERT(opts.num_shared_buffers == SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS); 1262 CU_ASSERT(opts.buf_cache_size == SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE); 1263 CU_ASSERT(opts.dif_insert_or_strip == SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP); 1264 CU_ASSERT(opts.abort_timeout_sec == SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC); 1265 CU_ASSERT(opts.transport_specific == NULL); 1266 } 1267 1268 static void 1269 test_nvmf_rdma_request_free_data(void) 1270 { 1271 struct spdk_nvmf_rdma_request rdma_req = {}; 1272 struct spdk_nvmf_rdma_transport rtransport = {}; 1273 struct spdk_nvmf_rdma_request_data *next_request_data = NULL; 1274 1275 MOCK_CLEAR(spdk_mempool_get); 1276 rtransport.data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", 1277 SPDK_NVMF_MAX_SGL_ENTRIES, 1278 sizeof(struct spdk_nvmf_rdma_request_data), 1279 SPDK_MEMPOOL_DEFAULT_CACHE_SIZE, 1280 SPDK_ENV_SOCKET_ID_ANY); 1281 next_request_data = spdk_mempool_get(rtransport.data_wr_pool); 1282 SPDK_CU_ASSERT_FATAL(((struct test_mempool *)rtransport.data_wr_pool)->count == 1283 SPDK_NVMF_MAX_SGL_ENTRIES - 1); 1284 next_request_data->wr.wr_id = 1; 1285 next_request_data->wr.num_sge = 2; 1286 next_request_data->wr.next = NULL; 1287 rdma_req.data.wr.next = &next_request_data->wr; 1288 rdma_req.data.wr.wr_id = 1; 1289 rdma_req.data.wr.num_sge = 2; 1290 1291 nvmf_rdma_request_free_data(&rdma_req, &rtransport); 1292 /* Check if next_request_data put into memory pool */ 1293 CU_ASSERT(((struct test_mempool *)rtransport.data_wr_pool)->count == SPDK_NVMF_MAX_SGL_ENTRIES); 1294 CU_ASSERT(rdma_req.data.wr.num_sge == 0); 1295 1296 spdk_mempool_free(rtransport.data_wr_pool); 1297 } 1298 1299 static void 1300 test_nvmf_rdma_update_ibv_state(void) 1301 { 1302 struct spdk_nvmf_rdma_qpair rqpair = {}; 1303 struct spdk_rdma_qp rdma_qp = {}; 1304 struct ibv_qp qp = {}; 1305 int rc = 0; 1306 1307 rqpair.rdma_qp = &rdma_qp; 1308 1309 /* Case 1: Failed to get updated RDMA queue pair state */ 1310 rqpair.ibv_state = IBV_QPS_INIT; 1311 rqpair.rdma_qp->qp = NULL; 1312 1313 rc = nvmf_rdma_update_ibv_state(&rqpair); 1314 CU_ASSERT(rc == IBV_QPS_ERR + 1); 1315 1316 /* Case 2: Bad state updated */ 1317 rqpair.rdma_qp->qp = &qp; 1318 qp.state = IBV_QPS_ERR; 1319 rc = nvmf_rdma_update_ibv_state(&rqpair); 1320 CU_ASSERT(rqpair.ibv_state == 10); 1321 CU_ASSERT(rc == IBV_QPS_ERR + 1); 1322 1323 /* Case 3: Pass */ 1324 qp.state = IBV_QPS_INIT; 1325 rc = nvmf_rdma_update_ibv_state(&rqpair); 1326 CU_ASSERT(rqpair.ibv_state == IBV_QPS_INIT); 1327 CU_ASSERT(rc == IBV_QPS_INIT); 1328 } 1329 1330 static void 1331 test_nvmf_rdma_resources_create(void) 1332 { 1333 static struct spdk_nvmf_rdma_resources *rdma_resource; 1334 struct spdk_nvmf_rdma_resource_opts opts = {}; 1335 struct spdk_nvmf_rdma_qpair qpair = {}; 1336 struct spdk_nvmf_rdma_recv *recv = NULL; 1337 struct spdk_nvmf_rdma_request *req = NULL; 1338 const int DEPTH = 128; 1339 1340 opts.max_queue_depth = DEPTH; 1341 opts.in_capsule_data_size = 4096; 1342 opts.shared = true; 1343 opts.qpair = &qpair; 1344 1345 rdma_resource = nvmf_rdma_resources_create(&opts); 1346 CU_ASSERT(rdma_resource != NULL); 1347 /* Just check first and last entry */ 1348 recv = &rdma_resource->recvs[0]; 1349 req = &rdma_resource->reqs[0]; 1350 CU_ASSERT(recv->rdma_wr.type == RDMA_WR_TYPE_RECV); 1351 CU_ASSERT((uintptr_t)recv->buf == (uintptr_t)(rdma_resource->bufs)); 1352 CU_ASSERT(recv->sgl[0].addr == (uintptr_t)&rdma_resource->cmds[0]); 1353 CU_ASSERT(recv->sgl[0].length == sizeof(rdma_resource->cmds[0])); 1354 CU_ASSERT(recv->sgl[0].lkey == RDMA_UT_LKEY); 1355 CU_ASSERT(recv->wr.num_sge == 2); 1356 CU_ASSERT(recv->wr.wr_id == (uintptr_t)&rdma_resource->recvs[0].rdma_wr); 1357 CU_ASSERT(recv->wr.sg_list == rdma_resource->recvs[0].sgl); 1358 CU_ASSERT(req->req.rsp == &rdma_resource->cpls[0]); 1359 CU_ASSERT(req->rsp.sgl[0].addr == (uintptr_t)&rdma_resource->cpls[0]); 1360 CU_ASSERT(req->rsp.sgl[0].length == sizeof(rdma_resource->cpls[0])); 1361 CU_ASSERT(req->rsp.sgl[0].lkey == RDMA_UT_LKEY); 1362 CU_ASSERT(req->rsp.rdma_wr.type == RDMA_WR_TYPE_SEND); 1363 CU_ASSERT(req->rsp.wr.wr_id == (uintptr_t)&rdma_resource->reqs[0].rsp.rdma_wr); 1364 CU_ASSERT(req->rsp.wr.next == NULL); 1365 CU_ASSERT(req->rsp.wr.opcode == IBV_WR_SEND); 1366 CU_ASSERT(req->rsp.wr.send_flags == IBV_SEND_SIGNALED); 1367 CU_ASSERT(req->rsp.wr.sg_list == rdma_resource->reqs[0].rsp.sgl); 1368 CU_ASSERT(req->rsp.wr.num_sge == NVMF_DEFAULT_RSP_SGE); 1369 CU_ASSERT(req->data.rdma_wr.type == RDMA_WR_TYPE_DATA); 1370 CU_ASSERT(req->data.wr.wr_id == (uintptr_t)&rdma_resource->reqs[0].data.rdma_wr); 1371 CU_ASSERT(req->data.wr.next == NULL); 1372 CU_ASSERT(req->data.wr.send_flags == IBV_SEND_SIGNALED); 1373 CU_ASSERT(req->data.wr.sg_list == rdma_resource->reqs[0].data.sgl); 1374 CU_ASSERT(req->data.wr.num_sge == SPDK_NVMF_MAX_SGL_ENTRIES); 1375 CU_ASSERT(req->state == RDMA_REQUEST_STATE_FREE); 1376 1377 recv = &rdma_resource->recvs[DEPTH - 1]; 1378 req = &rdma_resource->reqs[DEPTH - 1]; 1379 CU_ASSERT(recv->rdma_wr.type == RDMA_WR_TYPE_RECV); 1380 CU_ASSERT((uintptr_t)recv->buf == (uintptr_t)(rdma_resource->bufs + 1381 (DEPTH - 1) * 4096)); 1382 CU_ASSERT(recv->sgl[0].addr == (uintptr_t)&rdma_resource->cmds[DEPTH - 1]); 1383 CU_ASSERT(recv->sgl[0].length == sizeof(rdma_resource->cmds[DEPTH - 1])); 1384 CU_ASSERT(recv->sgl[0].lkey == RDMA_UT_LKEY); 1385 CU_ASSERT(recv->wr.num_sge == 2); 1386 CU_ASSERT(recv->wr.wr_id == (uintptr_t)&rdma_resource->recvs[DEPTH - 1].rdma_wr); 1387 CU_ASSERT(recv->wr.sg_list == rdma_resource->recvs[DEPTH - 1].sgl); 1388 CU_ASSERT(req->req.rsp == &rdma_resource->cpls[DEPTH - 1]); 1389 CU_ASSERT(req->rsp.sgl[0].addr == (uintptr_t)&rdma_resource->cpls[DEPTH - 1]); 1390 CU_ASSERT(req->rsp.sgl[0].length == sizeof(rdma_resource->cpls[DEPTH - 1])); 1391 CU_ASSERT(req->rsp.sgl[0].lkey == RDMA_UT_LKEY); 1392 CU_ASSERT(req->rsp.rdma_wr.type == RDMA_WR_TYPE_SEND); 1393 CU_ASSERT(req->rsp.wr.wr_id == (uintptr_t) 1394 &req->rsp.rdma_wr); 1395 CU_ASSERT(req->rsp.wr.next == NULL); 1396 CU_ASSERT(req->rsp.wr.opcode == IBV_WR_SEND); 1397 CU_ASSERT(req->rsp.wr.send_flags == IBV_SEND_SIGNALED); 1398 CU_ASSERT(req->rsp.wr.sg_list == rdma_resource->reqs[DEPTH - 1].rsp.sgl); 1399 CU_ASSERT(req->rsp.wr.num_sge == NVMF_DEFAULT_RSP_SGE); 1400 CU_ASSERT(req->data.rdma_wr.type == RDMA_WR_TYPE_DATA); 1401 CU_ASSERT(req->data.wr.wr_id == (uintptr_t) 1402 &req->data.rdma_wr); 1403 CU_ASSERT(req->data.wr.next == NULL); 1404 CU_ASSERT(req->data.wr.send_flags == IBV_SEND_SIGNALED); 1405 CU_ASSERT(req->data.wr.sg_list == rdma_resource->reqs[DEPTH - 1].data.sgl); 1406 CU_ASSERT(req->data.wr.num_sge == SPDK_NVMF_MAX_SGL_ENTRIES); 1407 CU_ASSERT(req->state == RDMA_REQUEST_STATE_FREE); 1408 1409 nvmf_rdma_resources_destroy(rdma_resource); 1410 } 1411 1412 static void 1413 test_nvmf_rdma_qpair_compare(void) 1414 { 1415 struct spdk_nvmf_rdma_qpair rqpair1 = {}, rqpair2 = {}; 1416 1417 rqpair1.qp_num = 0; 1418 rqpair2.qp_num = UINT32_MAX; 1419 1420 CU_ASSERT(nvmf_rdma_qpair_compare(&rqpair1, &rqpair2) < 0); 1421 CU_ASSERT(nvmf_rdma_qpair_compare(&rqpair2, &rqpair1) > 0); 1422 } 1423 1424 static void 1425 test_nvmf_rdma_resize_cq(void) 1426 { 1427 int rc = -1; 1428 int tnum_wr = 0; 1429 int tnum_cqe = 0; 1430 struct spdk_nvmf_rdma_qpair rqpair = {}; 1431 struct spdk_nvmf_rdma_poller rpoller = {}; 1432 struct spdk_nvmf_rdma_device rdevice = {}; 1433 struct ibv_context ircontext = {}; 1434 struct ibv_device idevice = {}; 1435 1436 rdevice.context = &ircontext; 1437 rqpair.poller = &rpoller; 1438 ircontext.device = &idevice; 1439 1440 /* Test1: Current capacity support required size. */ 1441 rpoller.required_num_wr = 10; 1442 rpoller.num_cqe = 20; 1443 rqpair.max_queue_depth = 2; 1444 tnum_wr = rpoller.required_num_wr; 1445 tnum_cqe = rpoller.num_cqe; 1446 1447 rc = nvmf_rdma_resize_cq(&rqpair, &rdevice); 1448 CU_ASSERT(rc == 0); 1449 CU_ASSERT(rpoller.required_num_wr == 10 + MAX_WR_PER_QP(rqpair.max_queue_depth)); 1450 CU_ASSERT(rpoller.required_num_wr > tnum_wr); 1451 CU_ASSERT(rpoller.num_cqe == tnum_cqe); 1452 1453 /* Test2: iWARP doesn't support CQ resize. */ 1454 tnum_wr = rpoller.required_num_wr; 1455 tnum_cqe = rpoller.num_cqe; 1456 idevice.transport_type = IBV_TRANSPORT_IWARP; 1457 1458 rc = nvmf_rdma_resize_cq(&rqpair, &rdevice); 1459 CU_ASSERT(rc == -1); 1460 CU_ASSERT(rpoller.required_num_wr == tnum_wr); 1461 CU_ASSERT(rpoller.num_cqe == tnum_cqe); 1462 1463 1464 /* Test3: RDMA CQE requirement exceeds device max_cqe limitation. */ 1465 tnum_wr = rpoller.required_num_wr; 1466 tnum_cqe = rpoller.num_cqe; 1467 idevice.transport_type = IBV_TRANSPORT_UNKNOWN; 1468 rdevice.attr.max_cqe = 3; 1469 1470 rc = nvmf_rdma_resize_cq(&rqpair, &rdevice); 1471 CU_ASSERT(rc == -1); 1472 CU_ASSERT(rpoller.required_num_wr == tnum_wr); 1473 CU_ASSERT(rpoller.num_cqe == tnum_cqe); 1474 1475 /* Test4: RDMA CQ resize failed. */ 1476 tnum_wr = rpoller.required_num_wr; 1477 tnum_cqe = rpoller.num_cqe; 1478 idevice.transport_type = IBV_TRANSPORT_IB; 1479 rdevice.attr.max_cqe = 30; 1480 MOCK_SET(ibv_resize_cq, -1); 1481 1482 rc = nvmf_rdma_resize_cq(&rqpair, &rdevice); 1483 CU_ASSERT(rc == -1); 1484 CU_ASSERT(rpoller.required_num_wr == tnum_wr); 1485 CU_ASSERT(rpoller.num_cqe == tnum_cqe); 1486 1487 /* Test5: RDMA CQ resize success. rsize = MIN(MAX(num_cqe * 2, required_num_wr), device->attr.max_cqe). */ 1488 tnum_wr = rpoller.required_num_wr; 1489 tnum_cqe = rpoller.num_cqe; 1490 MOCK_SET(ibv_resize_cq, 0); 1491 1492 rc = nvmf_rdma_resize_cq(&rqpair, &rdevice); 1493 CU_ASSERT(rc == 0); 1494 CU_ASSERT(rpoller.num_cqe = 30); 1495 CU_ASSERT(rpoller.required_num_wr == 18 + MAX_WR_PER_QP(rqpair.max_queue_depth)); 1496 CU_ASSERT(rpoller.required_num_wr > tnum_wr); 1497 CU_ASSERT(rpoller.num_cqe > tnum_cqe); 1498 } 1499 1500 int 1501 main(int argc, char **argv) 1502 { 1503 CU_pSuite suite = NULL; 1504 unsigned int num_failures; 1505 1506 CU_set_error_action(CUEA_ABORT); 1507 CU_initialize_registry(); 1508 1509 suite = CU_add_suite("nvmf", NULL, NULL); 1510 1511 CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl); 1512 CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_process); 1513 CU_ADD_TEST(suite, test_nvmf_rdma_get_optimal_poll_group); 1514 CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl_with_md); 1515 CU_ADD_TEST(suite, test_nvmf_rdma_opts_init); 1516 CU_ADD_TEST(suite, test_nvmf_rdma_request_free_data); 1517 CU_ADD_TEST(suite, test_nvmf_rdma_update_ibv_state); 1518 CU_ADD_TEST(suite, test_nvmf_rdma_resources_create); 1519 CU_ADD_TEST(suite, test_nvmf_rdma_qpair_compare); 1520 CU_ADD_TEST(suite, test_nvmf_rdma_resize_cq); 1521 1522 CU_basic_set_mode(CU_BRM_VERBOSE); 1523 CU_basic_run_tests(); 1524 num_failures = CU_get_number_of_failures(); 1525 CU_cleanup_registry(); 1526 return num_failures; 1527 } 1528