xref: /spdk/test/unit/lib/nvmf/rdma.c/rdma_ut.c (revision c164db9ffe3718ad4e4f5bab380ccfa62c2fa672)
1 /*   SPDX-License-Identifier: BSD-3-Clause
2  *   Copyright (C) 2018 Intel Corporation. All rights reserved.
3  *   Copyright (c) 2019, 2021 Mellanox Technologies LTD. All rights reserved.
4  *   Copyright (c) 2023, 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
5  */
6 
7 #include "spdk/stdinc.h"
8 #include "spdk_internal/cunit.h"
9 #include "common/lib/test_env.c"
10 #include "common/lib/test_iobuf.c"
11 #include "common/lib/test_rdma.c"
12 #include "nvmf/rdma.c"
13 #include "nvmf/transport.c"
14 
15 #define RDMA_UT_UNITS_IN_MAX_IO 16
16 
17 struct spdk_nvmf_transport_opts g_rdma_ut_transport_opts = {
18 	.max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH,
19 	.max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR,
20 	.in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE,
21 	.max_io_size = (SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE * RDMA_UT_UNITS_IN_MAX_IO),
22 	.io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE,
23 	.max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH,
24 	.num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS,
25 };
26 
27 SPDK_LOG_REGISTER_COMPONENT(nvmf)
28 DEFINE_STUB(spdk_mem_map_set_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
29 		uint64_t size, uint64_t translation), 0);
30 DEFINE_STUB(spdk_mem_map_clear_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
31 		uint64_t size), 0);
32 DEFINE_STUB(spdk_mem_map_alloc, struct spdk_mem_map *, (uint64_t default_translation,
33 		const struct spdk_mem_map_ops *ops, void *cb_ctx), NULL);
34 DEFINE_STUB(spdk_nvmf_qpair_disconnect, int, (struct spdk_nvmf_qpair *qpair), 0);
35 DEFINE_STUB(spdk_nvmf_qpair_get_listen_trid, int,
36 	    (struct spdk_nvmf_qpair *qpair, struct spdk_nvme_transport_id *trid), 0);
37 DEFINE_STUB_V(spdk_mem_map_free, (struct spdk_mem_map **pmap));
38 
39 DEFINE_STUB_V(spdk_nvmf_request_exec, (struct spdk_nvmf_request *req));
40 DEFINE_STUB(spdk_nvmf_request_complete, int, (struct spdk_nvmf_request *req), 0);
41 DEFINE_STUB(spdk_nvme_transport_id_compare, int, (const struct spdk_nvme_transport_id *trid1,
42 		const struct spdk_nvme_transport_id *trid2), 0);
43 DEFINE_STUB_V(spdk_nvmf_ctrlr_abort_aer, (struct spdk_nvmf_ctrlr *ctrlr));
44 DEFINE_STUB(spdk_nvmf_request_get_dif_ctx, bool, (struct spdk_nvmf_request *req,
45 		struct spdk_dif_ctx *dif_ctx), false);
46 DEFINE_STUB_V(spdk_nvme_trid_populate_transport, (struct spdk_nvme_transport_id *trid,
47 		enum spdk_nvme_transport_type trtype));
48 DEFINE_STUB_V(spdk_nvmf_tgt_new_qpair, (struct spdk_nvmf_tgt *tgt, struct spdk_nvmf_qpair *qpair));
49 DEFINE_STUB(nvmf_ctrlr_abort_request, int, (struct spdk_nvmf_request *req), 0);
50 DEFINE_STUB(spdk_nvme_transport_id_adrfam_str, const char *, (enum spdk_nvmf_adrfam adrfam), NULL);
51 DEFINE_STUB(ibv_dereg_mr, int, (struct ibv_mr *mr), 0);
52 DEFINE_STUB(ibv_resize_cq, int, (struct ibv_cq *cq, int cqe), 0);
53 DEFINE_STUB(spdk_mempool_lookup, struct spdk_mempool *, (const char *name), NULL);
54 DEFINE_STUB(spdk_rdma_cm_id_get_numa_id, int32_t, (struct rdma_cm_id *cm_id), 0);
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->data.wr.num_sge = 0;
136 	rdma_req->data.wr.wr.rdma.remote_addr = 0;
137 	rdma_req->data.wr.wr.rdma.rkey = 0;
138 	rdma_req->offset = 0;
139 	memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
140 
141 	for (i = 0; i < SPDK_NVMF_MAX_SGL_ENTRIES; i++) {
142 		rdma_req->req.iov[i].iov_base = 0;
143 		rdma_req->req.iov[i].iov_len = 0;
144 		rdma_req->data.wr.sg_list[i].addr = 0;
145 		rdma_req->data.wr.sg_list[i].length = 0;
146 		rdma_req->data.wr.sg_list[i].lkey = 0;
147 	}
148 	rdma_req->req.iovcnt = 0;
149 	if (rdma_req->req.stripped_data) {
150 		free(rdma_req->req.stripped_data);
151 		rdma_req->req.stripped_data = NULL;
152 	}
153 }
154 
155 static void
156 test_spdk_nvmf_rdma_request_parse_sgl(void)
157 {
158 	struct spdk_nvmf_rdma_transport rtransport = {};
159 	struct spdk_nvmf_transport_ops ops = {};
160 	struct spdk_nvmf_rdma_device device;
161 	struct spdk_nvmf_rdma_request rdma_req = {};
162 	struct spdk_nvmf_rdma_recv recv;
163 	struct spdk_nvmf_rdma_poll_group group;
164 	struct spdk_nvmf_rdma_qpair rqpair;
165 	struct spdk_nvmf_rdma_poller poller;
166 	union nvmf_c2h_msg cpl;
167 	union nvmf_h2c_msg cmd;
168 	struct spdk_nvme_sgl_descriptor *sgl;
169 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
170 	struct spdk_nvmf_rdma_request_data data;
171 	int rc, i;
172 	uint32_t sgl_length;
173 
174 	data.wr.sg_list = data.sgl;
175 	group.group.transport = &rtransport.transport;
176 	poller.group = &group;
177 	rqpair.poller = &poller;
178 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
179 
180 	sgl = &cmd.nvme_cmd.dptr.sgl1;
181 	rdma_req.recv = &recv;
182 	rdma_req.req.cmd = &cmd;
183 	rdma_req.req.rsp = &cpl;
184 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
185 	rdma_req.req.qpair = &rqpair.qpair;
186 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
187 
188 	rtransport.transport.opts = g_rdma_ut_transport_opts;
189 	rtransport.data_wr_pool = NULL;
190 	rtransport.transport.ops = &ops;
191 
192 	device.attr.device_cap_flags = 0;
193 	sgl->keyed.key = 0xEEEE;
194 	sgl->address = 0xFFFF;
195 	rdma_req.recv->buf = (void *)0xDDDD;
196 
197 	/* Test 1: sgl type: keyed data block subtype: address */
198 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
199 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
200 
201 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size */
202 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
203 	reset_nvmf_rdma_request(&rdma_req);
204 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
205 
206 	device.map = (void *)0x0;
207 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
208 	CU_ASSERT(rc == 0);
209 	CU_ASSERT(rdma_req.req.data_from_pool == true);
210 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
211 	CU_ASSERT((uint64_t)rdma_req.req.iovcnt == 1);
212 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
213 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
214 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
215 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
216 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
217 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
218 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
219 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
220 
221 	/* Part 2: simple I/O, one SGL larger than the transport io unit size (equal to the max io size) */
222 	reset_nvmf_rdma_request(&rdma_req);
223 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
224 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
225 
226 	CU_ASSERT(rc == 0);
227 	CU_ASSERT(rdma_req.req.data_from_pool == true);
228 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO);
229 	CU_ASSERT(rdma_req.data.wr.num_sge == RDMA_UT_UNITS_IN_MAX_IO);
230 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
231 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
232 	for (i = 0; i < RDMA_UT_UNITS_IN_MAX_IO; i++) {
233 		CU_ASSERT((uint64_t)rdma_req.req.iov[i].iov_base == 0x2000);
234 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
235 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
236 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
237 	}
238 
239 	/* Part 3: simple I/O one SGL larger than the transport max io size */
240 	reset_nvmf_rdma_request(&rdma_req);
241 	sgl->keyed.length = rtransport.transport.opts.max_io_size * 2;
242 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
243 
244 	CU_ASSERT(rc == -1);
245 
246 	/* Part 4: Pretend there are no buffer pools */
247 	MOCK_SET(spdk_iobuf_get, NULL);
248 	reset_nvmf_rdma_request(&rdma_req);
249 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
250 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
251 
252 	CU_ASSERT(rc == 0);
253 	CU_ASSERT(rdma_req.req.data_from_pool == false);
254 	CU_ASSERT(rdma_req.req.iovcnt == 0);
255 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
256 	CU_ASSERT(rdma_req.req.iov[0].iov_base == NULL);
257 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0);
258 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 0);
259 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == 0);
260 
261 	rdma_req.recv->buf = (void *)0xDDDD;
262 	/* Test 2: sgl type: keyed data block subtype: offset (in capsule data) */
263 	sgl->generic.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
264 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
265 
266 	/* Part 1: Normal I/O smaller than in capsule data size no offset */
267 	reset_nvmf_rdma_request(&rdma_req);
268 	sgl->address = 0;
269 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
270 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
271 
272 	CU_ASSERT(rc == 0);
273 	CU_ASSERT(rdma_req.req.iovcnt == 1);
274 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)0xDDDD);
275 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.in_capsule_data_size);
276 	CU_ASSERT(rdma_req.req.data_from_pool == false);
277 
278 	/* Part 2: I/O offset + length too large */
279 	reset_nvmf_rdma_request(&rdma_req);
280 	sgl->address = rtransport.transport.opts.in_capsule_data_size;
281 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
282 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
283 
284 	CU_ASSERT(rc == -1);
285 
286 	/* Part 3: I/O too large */
287 	reset_nvmf_rdma_request(&rdma_req);
288 	sgl->address = 0;
289 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size * 2;
290 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
291 
292 	CU_ASSERT(rc == -1);
293 
294 	/* Test 3: Multi SGL */
295 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
296 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
297 	sgl->address = 0;
298 	rdma_req.recv->buf = (void *)&sgl_desc;
299 	MOCK_SET(spdk_iobuf_get, &data);
300 	MOCK_SET(spdk_mempool_get, &data);
301 
302 	/* part 1: 2 segments each with 1 wr. */
303 	reset_nvmf_rdma_request(&rdma_req);
304 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
305 	for (i = 0; i < 2; i++) {
306 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
307 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
308 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size;
309 		sgl_desc[i].address = 0x4000 + i * rtransport.transport.opts.io_unit_size;
310 		sgl_desc[i].keyed.key = 0x44;
311 	}
312 
313 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
314 
315 	CU_ASSERT(rc == 0);
316 	CU_ASSERT(rdma_req.req.data_from_pool == true);
317 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 2);
318 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
319 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
320 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
321 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
322 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
323 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size);
324 	CU_ASSERT(data.wr.num_sge == 1);
325 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
326 
327 	/* part 2: 2 segments, each with 1 wr containing 8 sge_elements */
328 	reset_nvmf_rdma_request(&rdma_req);
329 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
330 	for (i = 0; i < 2; i++) {
331 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
332 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
333 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size * 8;
334 		sgl_desc[i].address = 0x4000 + i * 8 * rtransport.transport.opts.io_unit_size;
335 		sgl_desc[i].keyed.key = 0x44;
336 	}
337 
338 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
339 
340 	CU_ASSERT(rc == 0);
341 	CU_ASSERT(rdma_req.req.data_from_pool == true);
342 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
343 	CU_ASSERT(rdma_req.req.iovcnt == 16);
344 	CU_ASSERT(rdma_req.data.wr.num_sge == 8);
345 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
346 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
347 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
348 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
349 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 8);
350 	CU_ASSERT(data.wr.num_sge == 8);
351 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
352 
353 	/* part 3: 2 segments, one very large, one very small */
354 	reset_nvmf_rdma_request(&rdma_req);
355 	for (i = 0; i < 2; i++) {
356 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
357 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
358 		sgl_desc[i].keyed.key = 0x44;
359 	}
360 
361 	sgl_desc[0].keyed.length = rtransport.transport.opts.io_unit_size * 15 +
362 				   rtransport.transport.opts.io_unit_size / 2;
363 	sgl_desc[0].address = 0x4000;
364 	sgl_desc[1].keyed.length = rtransport.transport.opts.io_unit_size / 2;
365 	sgl_desc[1].address = 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
366 			      rtransport.transport.opts.io_unit_size / 2;
367 
368 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
369 
370 	CU_ASSERT(rc == 0);
371 	CU_ASSERT(rdma_req.req.data_from_pool == true);
372 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
373 	CU_ASSERT(rdma_req.req.iovcnt == 16);
374 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
375 	for (i = 0; i < 15; i++) {
376 		CU_ASSERT(rdma_req.data.sgl[i].length == rtransport.transport.opts.io_unit_size);
377 	}
378 	CU_ASSERT(rdma_req.data.sgl[15].length == rtransport.transport.opts.io_unit_size / 2);
379 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
380 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
381 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
382 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
383 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
384 		  rtransport.transport.opts.io_unit_size / 2);
385 	CU_ASSERT(data.sgl[0].length == rtransport.transport.opts.io_unit_size / 2);
386 	CU_ASSERT(data.wr.num_sge == 1);
387 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
388 
389 	/* part 4: 2 SGL descriptors, each length is transport buffer / 2
390 	 * 1 transport buffers should be allocated */
391 	reset_nvmf_rdma_request(&rdma_req);
392 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
393 	sgl_length = rtransport.transport.opts.io_unit_size / 2;
394 	for (i = 0; i < 2; i++) {
395 		sgl_desc[i].keyed.length = sgl_length;
396 		sgl_desc[i].address = 0x4000 + i * sgl_length;
397 	}
398 
399 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
400 
401 	CU_ASSERT(rc == 0);
402 	CU_ASSERT(rdma_req.req.data_from_pool == true);
403 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size);
404 	CU_ASSERT(rdma_req.req.iovcnt == 1);
405 
406 	CU_ASSERT(rdma_req.data.sgl[0].length == sgl_length);
407 	/* We mocked mempool_get to return address of data variable. Mempool is used
408 	 * to get both additional WRs and data buffers, so data points to &data */
409 	CU_ASSERT(rdma_req.data.sgl[0].addr == (uint64_t)&data);
410 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
411 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
412 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
413 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
414 
415 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
416 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + sgl_length);
417 	CU_ASSERT(data.sgl[0].length == sgl_length);
418 	CU_ASSERT(data.sgl[0].addr == (uint64_t)&data + sgl_length);
419 	CU_ASSERT(data.wr.num_sge == 1);
420 
421 	MOCK_CLEAR(spdk_mempool_get);
422 	MOCK_CLEAR(spdk_iobuf_get);
423 
424 	reset_nvmf_rdma_request(&rdma_req);
425 }
426 
427 static struct spdk_nvmf_rdma_recv *
428 create_recv(struct spdk_nvmf_rdma_qpair *rqpair, enum spdk_nvme_nvm_opcode opc)
429 {
430 	struct spdk_nvmf_rdma_recv *rdma_recv;
431 	union nvmf_h2c_msg *cmd;
432 	struct spdk_nvme_sgl_descriptor *sgl;
433 
434 	rdma_recv = calloc(1, sizeof(*rdma_recv));
435 	rdma_recv->qpair = rqpair;
436 	cmd = calloc(1, sizeof(*cmd));
437 	rdma_recv->sgl[0].addr = (uintptr_t)cmd;
438 	cmd->nvme_cmd.opc = opc;
439 	sgl = &cmd->nvme_cmd.dptr.sgl1;
440 	sgl->keyed.key = 0xEEEE;
441 	sgl->address = 0xFFFF;
442 	sgl->keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
443 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
444 	sgl->keyed.length = 1;
445 
446 	return rdma_recv;
447 }
448 
449 static void
450 free_recv(struct spdk_nvmf_rdma_recv *rdma_recv)
451 {
452 	free((void *)rdma_recv->sgl[0].addr);
453 	free(rdma_recv);
454 }
455 
456 static struct spdk_nvmf_rdma_request *
457 create_req(struct spdk_nvmf_rdma_qpair *rqpair,
458 	   struct spdk_nvmf_rdma_recv *rdma_recv)
459 {
460 	struct spdk_nvmf_rdma_request *rdma_req;
461 	union nvmf_c2h_msg *cpl;
462 
463 	rdma_req = calloc(1, sizeof(*rdma_req));
464 	rdma_req->recv = rdma_recv;
465 	rdma_req->req.qpair = &rqpair->qpair;
466 	rdma_req->state = RDMA_REQUEST_STATE_NEW;
467 	rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr;
468 	rdma_req->data.wr.sg_list = rdma_req->data.sgl;
469 	cpl = calloc(1, sizeof(*cpl));
470 	rdma_req->rsp.sgl[0].addr = (uintptr_t)cpl;
471 	rdma_req->req.rsp = cpl;
472 
473 	return rdma_req;
474 }
475 
476 static void
477 free_req(struct spdk_nvmf_rdma_request *rdma_req)
478 {
479 	free((void *)rdma_req->rsp.sgl[0].addr);
480 	free(rdma_req);
481 }
482 
483 static void
484 qpair_reset(struct spdk_nvmf_rdma_qpair *rqpair,
485 	    struct spdk_nvmf_rdma_poller *poller,
486 	    struct spdk_nvmf_rdma_device *device,
487 	    struct spdk_nvmf_rdma_resources *resources,
488 	    struct spdk_nvmf_transport *transport)
489 {
490 	memset(rqpair, 0, sizeof(*rqpair));
491 	STAILQ_INIT(&rqpair->pending_rdma_write_queue);
492 	STAILQ_INIT(&rqpair->pending_rdma_read_queue);
493 	STAILQ_INIT(&rqpair->pending_rdma_send_queue);
494 	rqpair->poller = poller;
495 	rqpair->device = device;
496 	rqpair->resources = resources;
497 	rqpair->qpair.qid = 1;
498 	rqpair->qpair.state = SPDK_NVMF_QPAIR_ENABLED;
499 	rqpair->max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
500 	rqpair->max_send_depth = 16;
501 	rqpair->max_read_depth = 16;
502 	rqpair->qpair.transport = transport;
503 }
504 
505 static void
506 poller_reset(struct spdk_nvmf_rdma_poller *poller,
507 	     struct spdk_nvmf_rdma_poll_group *group)
508 {
509 	memset(poller, 0, sizeof(*poller));
510 	STAILQ_INIT(&poller->qpairs_pending_recv);
511 	STAILQ_INIT(&poller->qpairs_pending_send);
512 	poller->group = group;
513 }
514 
515 static void
516 test_spdk_nvmf_rdma_request_process(void)
517 {
518 	struct spdk_nvmf_rdma_transport rtransport = {};
519 	struct spdk_nvmf_transport_ops ops = {};
520 	struct spdk_nvmf_rdma_poll_group group = {};
521 	struct spdk_nvmf_rdma_poller poller = {};
522 	struct spdk_nvmf_rdma_device device = {};
523 	struct spdk_nvmf_rdma_resources resources = {};
524 	struct spdk_nvmf_rdma_qpair rqpair = {};
525 	struct spdk_nvmf_rdma_recv *rdma_recv;
526 	struct spdk_nvmf_rdma_request *rdma_req;
527 	struct spdk_iobuf_channel ch = {};
528 	bool progress;
529 
530 	group.group.buf_cache = &ch;
531 
532 	STAILQ_INIT(&group.group.pending_buf_queue);
533 	poller_reset(&poller, &group);
534 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
535 
536 	rtransport.transport.opts = g_rdma_ut_transport_opts;
537 	rtransport.data_wr_pool = spdk_mempool_create("test_wr_pool", 128,
538 				  sizeof(struct spdk_nvmf_rdma_request_data),
539 				  0, 0);
540 	rtransport.transport.ops = &ops;
541 	MOCK_CLEAR(spdk_iobuf_get);
542 
543 	device.attr.device_cap_flags = 0;
544 	device.map = (void *)0x0;
545 
546 	/* Test 1: single SGL READ request */
547 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_READ);
548 	rdma_req = create_req(&rqpair, rdma_recv);
549 	rqpair.current_recv_depth = 1;
550 	/* NEW -> EXECUTING */
551 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
552 	CU_ASSERT(progress == true);
553 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
554 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST);
555 	/* EXECUTED -> TRANSFERRING_C2H */
556 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
557 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
558 	CU_ASSERT(progress == true);
559 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST);
560 	CU_ASSERT(rdma_req->recv == NULL);
561 	/* COMPLETED -> FREE */
562 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
563 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
564 	CU_ASSERT(progress == true);
565 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
566 
567 	free_recv(rdma_recv);
568 	free_req(rdma_req);
569 	poller_reset(&poller, &group);
570 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
571 
572 	/* Test 2: single SGL WRITE request */
573 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
574 	rdma_req = create_req(&rqpair, rdma_recv);
575 	rqpair.current_recv_depth = 1;
576 	/* NEW -> TRANSFERRING_H2C */
577 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
578 	CU_ASSERT(progress == true);
579 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
580 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
581 	STAILQ_INIT(&poller.qpairs_pending_send);
582 	/* READY_TO_EXECUTE -> EXECUTING */
583 	rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
584 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
585 	CU_ASSERT(progress == true);
586 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
587 	/* EXECUTED -> COMPLETING */
588 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
589 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
590 	CU_ASSERT(progress == true);
591 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
592 	CU_ASSERT(rdma_req->recv == NULL);
593 	/* COMPLETED -> FREE */
594 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
595 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
596 	CU_ASSERT(progress == true);
597 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
598 
599 	free_recv(rdma_recv);
600 	free_req(rdma_req);
601 	poller_reset(&poller, &group);
602 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
603 
604 	/* Test 3: WRITE+WRITE ibv_send batching */
605 	{
606 		struct spdk_nvmf_rdma_recv *recv1, *recv2;
607 		struct spdk_nvmf_rdma_request *req1, *req2;
608 		recv1 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
609 		req1 = create_req(&rqpair, recv1);
610 		recv2 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
611 		req2 = create_req(&rqpair, recv2);
612 
613 		/* WRITE 1: NEW -> TRANSFERRING_H2C */
614 		rqpair.current_recv_depth = 1;
615 		nvmf_rdma_request_process(&rtransport, req1);
616 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
617 
618 		/* WRITE 2: NEW -> TRANSFERRING_H2C */
619 		rqpair.current_recv_depth = 2;
620 		nvmf_rdma_request_process(&rtransport, req2);
621 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
622 
623 		STAILQ_INIT(&poller.qpairs_pending_send);
624 
625 		/* WRITE 1 completes before WRITE 2 has finished RDMA reading */
626 		/* WRITE 1: READY_TO_EXECUTE -> EXECUTING */
627 		req1->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
628 		nvmf_rdma_request_process(&rtransport, req1);
629 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_EXECUTING);
630 		/* WRITE 1: EXECUTED -> COMPLETING */
631 		req1->state = RDMA_REQUEST_STATE_EXECUTED;
632 		nvmf_rdma_request_process(&rtransport, req1);
633 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_COMPLETING);
634 		STAILQ_INIT(&poller.qpairs_pending_send);
635 		/* WRITE 1: COMPLETED -> FREE */
636 		req1->state = RDMA_REQUEST_STATE_COMPLETED;
637 		nvmf_rdma_request_process(&rtransport, req1);
638 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_FREE);
639 
640 		/* Now WRITE 2 has finished reading and completes */
641 		/* WRITE 2: COMPLETED -> FREE */
642 		/* WRITE 2: READY_TO_EXECUTE -> EXECUTING */
643 		req2->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
644 		nvmf_rdma_request_process(&rtransport, req2);
645 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_EXECUTING);
646 		/* WRITE 1: EXECUTED -> COMPLETING */
647 		req2->state = RDMA_REQUEST_STATE_EXECUTED;
648 		nvmf_rdma_request_process(&rtransport, req2);
649 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_COMPLETING);
650 		STAILQ_INIT(&poller.qpairs_pending_send);
651 		/* WRITE 1: COMPLETED -> FREE */
652 		req2->state = RDMA_REQUEST_STATE_COMPLETED;
653 		nvmf_rdma_request_process(&rtransport, req2);
654 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_FREE);
655 
656 		free_recv(recv1);
657 		free_req(req1);
658 		free_recv(recv2);
659 		free_req(req2);
660 		poller_reset(&poller, &group);
661 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
662 	}
663 
664 	/* Test 4, invalid command, check xfer type */
665 	{
666 		struct spdk_nvmf_rdma_recv *rdma_recv_inv;
667 		struct spdk_nvmf_rdma_request *rdma_req_inv;
668 		/* construct an opcode that specifies BIDIRECTIONAL transfer */
669 		uint8_t opc = 0x10 | SPDK_NVME_DATA_BIDIRECTIONAL;
670 
671 		rdma_recv_inv = create_recv(&rqpair, opc);
672 		rdma_req_inv = create_req(&rqpair, rdma_recv_inv);
673 
674 		/* NEW -> RDMA_REQUEST_STATE_COMPLETING */
675 		rqpair.current_recv_depth = 1;
676 		progress = nvmf_rdma_request_process(&rtransport, rdma_req_inv);
677 		CU_ASSERT(progress == true);
678 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_COMPLETING);
679 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sct == SPDK_NVME_SCT_GENERIC);
680 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sc == SPDK_NVME_SC_INVALID_OPCODE);
681 
682 		/* RDMA_REQUEST_STATE_COMPLETED -> FREE */
683 		rdma_req_inv->state = RDMA_REQUEST_STATE_COMPLETED;
684 		nvmf_rdma_request_process(&rtransport, rdma_req_inv);
685 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_FREE);
686 
687 		free_recv(rdma_recv_inv);
688 		free_req(rdma_req_inv);
689 		poller_reset(&poller, &group);
690 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
691 	}
692 
693 	/* Test 5: Write response waits in queue */
694 	{
695 		rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
696 		rdma_req = create_req(&rqpair, rdma_recv);
697 		rqpair.current_recv_depth = 1;
698 		/* NEW -> TRANSFERRING_H2C */
699 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
700 		CU_ASSERT(progress == true);
701 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
702 		CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
703 		STAILQ_INIT(&poller.qpairs_pending_send);
704 		/* READY_TO_EXECUTE -> EXECUTING */
705 		rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
706 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
707 		CU_ASSERT(progress == true);
708 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
709 		/* EXECUTED -> COMPLETING */
710 		rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
711 		/* Send queue is full */
712 		rqpair.current_send_depth = rqpair.max_send_depth;
713 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
714 		CU_ASSERT(progress == true);
715 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING);
716 		CU_ASSERT(rdma_req == STAILQ_FIRST(&rqpair.pending_rdma_send_queue));
717 
718 		/* Send queue is still full */
719 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
720 		CU_ASSERT(progress == false);
721 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING);
722 		CU_ASSERT(rdma_req == STAILQ_FIRST(&rqpair.pending_rdma_send_queue));
723 
724 		/* Slot is available */
725 		rqpair.current_send_depth = rqpair.max_send_depth - 1;
726 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
727 		CU_ASSERT(progress == true);
728 		CU_ASSERT(STAILQ_EMPTY(&rqpair.pending_rdma_send_queue));
729 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
730 		CU_ASSERT(rdma_req->recv == NULL);
731 		/* COMPLETED -> FREE */
732 		rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
733 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
734 		CU_ASSERT(progress == true);
735 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
736 
737 		free_recv(rdma_recv);
738 		free_req(rdma_req);
739 		poller_reset(&poller, &group);
740 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
741 
742 	}
743 
744 	spdk_mempool_free(rtransport.data_wr_pool);
745 }
746 
747 #define TEST_GROUPS_COUNT 5
748 static void
749 test_nvmf_rdma_get_optimal_poll_group(void)
750 {
751 	struct spdk_nvmf_rdma_transport rtransport = {};
752 	struct spdk_nvmf_transport *transport = &rtransport.transport;
753 	struct spdk_nvmf_rdma_qpair rqpair = {};
754 	struct spdk_nvmf_transport_poll_group *groups[TEST_GROUPS_COUNT];
755 	struct spdk_nvmf_rdma_poll_group *rgroups[TEST_GROUPS_COUNT];
756 	struct spdk_nvmf_transport_poll_group *result;
757 	struct spdk_nvmf_poll_group group = {};
758 	uint32_t i;
759 
760 	rqpair.qpair.transport = transport;
761 	TAILQ_INIT(&rtransport.poll_groups);
762 
763 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
764 		groups[i] = nvmf_rdma_poll_group_create(transport, NULL);
765 		CU_ASSERT(groups[i] != NULL);
766 		groups[i]->group = &group;
767 		rgroups[i] = SPDK_CONTAINEROF(groups[i], struct spdk_nvmf_rdma_poll_group, group);
768 		groups[i]->transport = transport;
769 	}
770 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[0]);
771 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[0]);
772 
773 	/* Emulate connection of %TEST_GROUPS_COUNT% initiators - each creates 1 admin and 1 io qp */
774 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
775 		rqpair.qpair.qid = 0;
776 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
777 		CU_ASSERT(result == groups[i]);
778 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
779 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i]);
780 
781 		rqpair.qpair.qid = 1;
782 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
783 		CU_ASSERT(result == groups[i]);
784 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
785 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
786 	}
787 	/* wrap around, admin/io pg point to the first pg
788 	   Destroy all poll groups except of the last one */
789 	for (i = 0; i < TEST_GROUPS_COUNT - 1; i++) {
790 		nvmf_rdma_poll_group_destroy(groups[i]);
791 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[i + 1]);
792 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i + 1]);
793 	}
794 
795 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
796 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
797 
798 	/* Check that pointers to the next admin/io poll groups are not changed */
799 	rqpair.qpair.qid = 0;
800 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
801 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
802 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
803 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
804 
805 	rqpair.qpair.qid = 1;
806 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
807 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
808 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
809 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
810 
811 	/* Remove the last poll group, check that pointers are NULL */
812 	nvmf_rdma_poll_group_destroy(groups[TEST_GROUPS_COUNT - 1]);
813 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == NULL);
814 	CU_ASSERT(rtransport.conn_sched.next_io_pg == NULL);
815 
816 	/* Request optimal poll group, result must be NULL */
817 	rqpair.qpair.qid = 0;
818 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
819 	CU_ASSERT(result == NULL);
820 
821 	rqpair.qpair.qid = 1;
822 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
823 	CU_ASSERT(result == NULL);
824 }
825 #undef TEST_GROUPS_COUNT
826 
827 static void
828 test_spdk_nvmf_rdma_request_parse_sgl_with_md(void)
829 {
830 	struct spdk_nvmf_rdma_transport rtransport = {};
831 	struct spdk_nvmf_transport_ops ops = {};
832 	struct spdk_nvmf_rdma_device device;
833 	struct spdk_nvmf_rdma_request rdma_req = {};
834 	struct spdk_nvmf_rdma_recv recv;
835 	struct spdk_nvmf_rdma_poll_group group;
836 	struct spdk_nvmf_rdma_qpair rqpair;
837 	struct spdk_nvmf_rdma_poller poller;
838 	union nvmf_c2h_msg cpl;
839 	union nvmf_h2c_msg cmd;
840 	struct spdk_nvme_sgl_descriptor *sgl;
841 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
842 	char data_buffer[8192];
843 	struct spdk_nvmf_rdma_request_data *data = (struct spdk_nvmf_rdma_request_data *)data_buffer;
844 	char data2_buffer[8192];
845 	struct spdk_nvmf_rdma_request_data *data2 = (struct spdk_nvmf_rdma_request_data *)data2_buffer;
846 	const uint32_t data_bs = 512;
847 	const uint32_t md_size = 8;
848 	int rc, i;
849 	struct spdk_dif_ctx_init_ext_opts dif_opts;
850 
851 	MOCK_CLEAR(spdk_mempool_get);
852 	MOCK_CLEAR(spdk_iobuf_get);
853 
854 	data->wr.sg_list = data->sgl;
855 	group.group.transport = &rtransport.transport;
856 	poller.group = &group;
857 	rqpair.poller = &poller;
858 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
859 
860 	sgl = &cmd.nvme_cmd.dptr.sgl1;
861 	rdma_req.recv = &recv;
862 	rdma_req.req.cmd = &cmd;
863 	rdma_req.req.rsp = &cpl;
864 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
865 	rdma_req.req.qpair = &rqpair.qpair;
866 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
867 
868 	rtransport.transport.opts = g_rdma_ut_transport_opts;
869 	rtransport.data_wr_pool = NULL;
870 	rtransport.transport.ops = &ops;
871 
872 	device.attr.device_cap_flags = 0;
873 	device.map = NULL;
874 	sgl->keyed.key = 0xEEEE;
875 	sgl->address = 0xFFFF;
876 	rdma_req.recv->buf = (void *)0xDDDD;
877 
878 	/* Test 1: sgl type: keyed data block subtype: address */
879 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
880 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
881 
882 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size, block size 512 */
883 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
884 	reset_nvmf_rdma_request(&rdma_req);
885 	dif_opts.size = SPDK_SIZEOF(&dif_opts, dif_pi_format);
886 	dif_opts.dif_pi_format = SPDK_DIF_PI_FORMAT_16;
887 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
888 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
889 			  0, 0, 0, 0, 0, &dif_opts);
890 	rdma_req.req.dif_enabled = true;
891 	rtransport.transport.opts.io_unit_size = data_bs * 8;
892 	rdma_req.req.qpair->transport = &rtransport.transport;
893 	sgl->keyed.length = data_bs * 4;
894 
895 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
896 
897 	CU_ASSERT(rc == 0);
898 	CU_ASSERT(rdma_req.req.data_from_pool == true);
899 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
900 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
901 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
902 	CU_ASSERT(rdma_req.req.iovcnt == 1);
903 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
904 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
905 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
906 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
907 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
908 
909 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
910 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rdma_req.req.length);
911 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
912 
913 	/* Part 2: simple I/O, one SGL equal to io unit size, io_unit_size is not aligned with md_size,
914 		block size 512 */
915 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
916 	reset_nvmf_rdma_request(&rdma_req);
917 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
918 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
919 			  0, 0, 0, 0, 0, &dif_opts);
920 	rdma_req.req.dif_enabled = true;
921 	rtransport.transport.opts.io_unit_size = data_bs * 4;
922 	sgl->keyed.length = data_bs * 4;
923 
924 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
925 
926 	CU_ASSERT(rc == 0);
927 	CU_ASSERT(rdma_req.req.data_from_pool == true);
928 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
929 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
930 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
931 	CU_ASSERT(rdma_req.req.iovcnt == 2);
932 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
933 	CU_ASSERT(rdma_req.data.wr.num_sge == 5);
934 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
935 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
936 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
937 
938 	for (i = 0; i < 3; ++i) {
939 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
940 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
941 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
942 	}
943 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
944 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
945 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
946 
947 	/* 2nd buffer consumed */
948 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
949 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
950 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
951 
952 	/* Part 3: simple I/O, one SGL equal io unit size, io_unit_size is equal to block size 512 bytes */
953 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
954 	reset_nvmf_rdma_request(&rdma_req);
955 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
956 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
957 			  0, 0, 0, 0, 0, &dif_opts);
958 	rdma_req.req.dif_enabled = true;
959 	rtransport.transport.opts.io_unit_size = data_bs;
960 	sgl->keyed.length = data_bs;
961 
962 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
963 
964 	CU_ASSERT(rc == 0);
965 	CU_ASSERT(rdma_req.req.data_from_pool == true);
966 	CU_ASSERT(rdma_req.req.length == data_bs);
967 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
968 	CU_ASSERT(rdma_req.req.dif.elba_length == data_bs + md_size);
969 	CU_ASSERT(rdma_req.req.iovcnt == 2);
970 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
971 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
972 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
973 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
974 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
975 
976 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
977 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs);
978 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
979 
980 	CU_ASSERT(rdma_req.req.iovcnt == 2);
981 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)((unsigned long)0x2000));
982 	CU_ASSERT(rdma_req.req.iov[0].iov_len == data_bs);
983 	/* 2nd buffer consumed for metadata */
984 	CU_ASSERT(rdma_req.req.iov[1].iov_base == (void *)((unsigned long)0x2000));
985 	CU_ASSERT(rdma_req.req.iov[1].iov_len == md_size);
986 
987 	/* Part 4: simple I/O, one SGL equal io unit size, io_unit_size is aligned with md_size,
988 	   block size 512 */
989 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
990 	reset_nvmf_rdma_request(&rdma_req);
991 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
992 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
993 			  0, 0, 0, 0, 0, &dif_opts);
994 	rdma_req.req.dif_enabled = true;
995 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
996 	sgl->keyed.length = data_bs * 4;
997 
998 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
999 
1000 	CU_ASSERT(rc == 0);
1001 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1002 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1003 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1004 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1005 	CU_ASSERT(rdma_req.req.iovcnt == 1);
1006 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1007 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1008 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1009 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1010 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1011 
1012 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1013 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rdma_req.req.length);
1014 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1015 
1016 	/* Part 5: simple I/O, one SGL equal to 2x io unit size, io_unit_size is aligned with md_size,
1017 	   block size 512 */
1018 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
1019 	reset_nvmf_rdma_request(&rdma_req);
1020 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1021 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1022 			  0, 0, 0, 0, 0, &dif_opts);
1023 	rdma_req.req.dif_enabled = true;
1024 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 2;
1025 	sgl->keyed.length = data_bs * 4;
1026 
1027 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1028 
1029 	CU_ASSERT(rc == 0);
1030 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1031 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1032 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1033 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1034 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1035 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1036 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
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.iov[0].iov_base == 0x2000);
1040 
1041 	for (i = 0; i < 2; ++i) {
1042 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
1043 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs * 2);
1044 	}
1045 
1046 	/* Part 6: simple I/O, one SGL larger than the transport io unit size, io_unit_size is not aligned to md_size,
1047 	   block size 512 */
1048 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
1049 	reset_nvmf_rdma_request(&rdma_req);
1050 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1051 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1052 			  0, 0, 0, 0, 0, &dif_opts);
1053 	rdma_req.req.dif_enabled = true;
1054 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1055 	sgl->keyed.length = data_bs * 6;
1056 
1057 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1058 
1059 	CU_ASSERT(rc == 0);
1060 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1061 	CU_ASSERT(rdma_req.req.length == data_bs * 6);
1062 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1063 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 6);
1064 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1065 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1066 	CU_ASSERT(rdma_req.data.wr.num_sge == 7);
1067 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1068 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1069 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1070 
1071 	for (i = 0; i < 3; ++i) {
1072 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1073 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1074 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1075 	}
1076 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
1077 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
1078 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
1079 
1080 	/* 2nd IO buffer consumed */
1081 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
1082 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
1083 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
1084 
1085 	CU_ASSERT(rdma_req.data.wr.sg_list[5].addr == 0x2000 + 24 + md_size);
1086 	CU_ASSERT(rdma_req.data.wr.sg_list[5].length == 512);
1087 	CU_ASSERT(rdma_req.data.wr.sg_list[5].lkey == RDMA_UT_LKEY);
1088 
1089 	CU_ASSERT(rdma_req.data.wr.sg_list[6].addr == 0x2000 + 24 + 512 + md_size * 2);
1090 	CU_ASSERT(rdma_req.data.wr.sg_list[6].length == 512);
1091 	CU_ASSERT(rdma_req.data.wr.sg_list[6].lkey == RDMA_UT_LKEY);
1092 
1093 	/* Part 7: simple I/O, number of SGL entries exceeds the number of entries
1094 	   one WR can hold. Additional WR is chained */
1095 	MOCK_SET(spdk_iobuf_get, data2_buffer);
1096 	MOCK_SET(spdk_mempool_get, data2_buffer);
1097 	reset_nvmf_rdma_request(&rdma_req);
1098 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1099 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1100 			  0, 0, 0, 0, 0, &dif_opts);
1101 	rdma_req.req.dif_enabled = true;
1102 	rtransport.transport.opts.io_unit_size = data_bs * 16;
1103 	sgl->keyed.length = data_bs * 16;
1104 
1105 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1106 
1107 	CU_ASSERT(rc == 0);
1108 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1109 	CU_ASSERT(rdma_req.req.length == data_bs * 16);
1110 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1111 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1112 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 16);
1113 	CU_ASSERT(rdma_req.req.iov[0].iov_base == data2_buffer);
1114 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
1115 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1116 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1117 
1118 	for (i = 0; i < 15; ++i) {
1119 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)data2_buffer + i * (data_bs + md_size));
1120 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1121 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1122 	}
1123 
1124 	/* 8192 - (512 + 8) * 15 = 392 */
1125 	CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)data2_buffer + i * (data_bs + md_size));
1126 	CU_ASSERT(rdma_req.data.wr.sg_list[i].length == 392);
1127 	CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1128 
1129 	/* additional wr from pool */
1130 	CU_ASSERT(rdma_req.data.wr.next == (void *)&data2->wr);
1131 	CU_ASSERT(rdma_req.data.wr.next->num_sge == 1);
1132 	CU_ASSERT(rdma_req.data.wr.next->next == &rdma_req.rsp.wr);
1133 	/* 2nd IO buffer */
1134 	CU_ASSERT(data2->wr.sg_list[0].addr == (uintptr_t)data2_buffer);
1135 	CU_ASSERT(data2->wr.sg_list[0].length == 120);
1136 	CU_ASSERT(data2->wr.sg_list[0].lkey == RDMA_UT_LKEY);
1137 
1138 	/* Part 8: simple I/O, data with metadata do not fit to 1 io_buffer */
1139 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
1140 	reset_nvmf_rdma_request(&rdma_req);
1141 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1142 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1143 			  0, 0, 0, 0, 0, &dif_opts);
1144 	rdma_req.req.dif_enabled = true;
1145 	rtransport.transport.opts.io_unit_size = 516;
1146 	sgl->keyed.length = data_bs * 2;
1147 
1148 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1149 
1150 	CU_ASSERT(rc == 0);
1151 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1152 	CU_ASSERT(rdma_req.req.length == data_bs * 2);
1153 	CU_ASSERT(rdma_req.req.iovcnt == 3);
1154 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1155 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 2);
1156 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)0x2000);
1157 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1158 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1159 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1160 
1161 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1162 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 512);
1163 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1164 
1165 	/* 2nd IO buffer consumed, offset 4 bytes due to part of the metadata
1166 	  is located at the beginning of that buffer */
1167 	CU_ASSERT(rdma_req.data.wr.sg_list[1].addr == 0x2000 + 4);
1168 	CU_ASSERT(rdma_req.data.wr.sg_list[1].length == 512);
1169 	CU_ASSERT(rdma_req.data.wr.sg_list[1].lkey == RDMA_UT_LKEY);
1170 
1171 	/* Test 2: Multi SGL */
1172 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1173 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1174 	sgl->address = 0;
1175 	rdma_req.recv->buf = (void *)&sgl_desc;
1176 	MOCK_SET(spdk_mempool_get, data_buffer);
1177 	MOCK_SET(spdk_iobuf_get, data_buffer);
1178 
1179 	/* part 1: 2 segments each with 1 wr. io_unit_size is aligned with data_bs + md_size */
1180 	reset_nvmf_rdma_request(&rdma_req);
1181 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1182 			  SPDK_DIF_TYPE1,
1183 			  SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1184 			  0, 0, 0, 0, 0, &dif_opts);
1185 	rdma_req.req.dif_enabled = true;
1186 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1187 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
1188 
1189 	for (i = 0; i < 2; i++) {
1190 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1191 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1192 		sgl_desc[i].keyed.length = data_bs * 4;
1193 		sgl_desc[i].address = 0x4000 + i * data_bs * 4;
1194 		sgl_desc[i].keyed.key = 0x44;
1195 	}
1196 
1197 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1198 
1199 	CU_ASSERT(rc == 0);
1200 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1201 	CU_ASSERT(rdma_req.req.length == data_bs * 4 * 2);
1202 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1203 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4 * 2);
1204 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1205 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == (uintptr_t)(data_buffer));
1206 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs * 4);
1207 
1208 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
1209 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
1210 	CU_ASSERT(rdma_req.data.wr.next == &data->wr);
1211 	CU_ASSERT(data->wr.wr.rdma.rkey == 0x44);
1212 	CU_ASSERT(data->wr.wr.rdma.remote_addr == 0x4000 + data_bs * 4);
1213 	CU_ASSERT(data->wr.num_sge == 1);
1214 	CU_ASSERT(data->wr.sg_list[0].addr == (uintptr_t)(data_buffer));
1215 	CU_ASSERT(data->wr.sg_list[0].length == data_bs * 4);
1216 
1217 	CU_ASSERT(data->wr.next == &rdma_req.rsp.wr);
1218 	reset_nvmf_rdma_request(&rdma_req);
1219 }
1220 
1221 static void
1222 test_nvmf_rdma_opts_init(void)
1223 {
1224 	struct spdk_nvmf_transport_opts	opts = {};
1225 
1226 	nvmf_rdma_opts_init(&opts);
1227 	CU_ASSERT(opts.max_queue_depth == SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH);
1228 	CU_ASSERT(opts.max_qpairs_per_ctrlr ==	SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR);
1229 	CU_ASSERT(opts.in_capsule_data_size ==	SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE);
1230 	CU_ASSERT(opts.max_io_size == SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE);
1231 	CU_ASSERT(opts.io_unit_size == SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
1232 	CU_ASSERT(opts.max_aq_depth == SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH);
1233 	CU_ASSERT(opts.num_shared_buffers == SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS);
1234 	CU_ASSERT(opts.buf_cache_size == SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE);
1235 	CU_ASSERT(opts.dif_insert_or_strip == SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP);
1236 	CU_ASSERT(opts.abort_timeout_sec == SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC);
1237 	CU_ASSERT(opts.transport_specific == NULL);
1238 }
1239 
1240 static void
1241 test_nvmf_rdma_request_free_data(void)
1242 {
1243 	struct spdk_nvmf_rdma_request rdma_req = {};
1244 	struct spdk_nvmf_rdma_transport rtransport = {};
1245 	struct spdk_nvmf_rdma_request_data *next_request_data = NULL;
1246 
1247 	MOCK_CLEAR(spdk_mempool_get);
1248 	rtransport.data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data",
1249 				  SPDK_NVMF_MAX_SGL_ENTRIES,
1250 				  sizeof(struct spdk_nvmf_rdma_request_data),
1251 				  SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
1252 				  SPDK_ENV_NUMA_ID_ANY);
1253 	next_request_data = spdk_mempool_get(rtransport.data_wr_pool);
1254 	SPDK_CU_ASSERT_FATAL(((struct test_mempool *)rtransport.data_wr_pool)->count ==
1255 			     SPDK_NVMF_MAX_SGL_ENTRIES - 1);
1256 	next_request_data->wr.wr_id = (uint64_t)&rdma_req.data_wr;
1257 	next_request_data->wr.num_sge = 2;
1258 	next_request_data->wr.next = NULL;
1259 	rdma_req.data.wr.next = &next_request_data->wr;
1260 	rdma_req.data.wr.wr_id = (uint64_t)&rdma_req.data_wr;
1261 	rdma_req.data.wr.num_sge = 2;
1262 	rdma_req.transfer_wr = &rdma_req.data.wr;
1263 
1264 	nvmf_rdma_request_free_data(&rdma_req, &rtransport);
1265 	/* Check if next_request_data put into memory pool */
1266 	CU_ASSERT(((struct test_mempool *)rtransport.data_wr_pool)->count == SPDK_NVMF_MAX_SGL_ENTRIES);
1267 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
1268 
1269 	spdk_mempool_free(rtransport.data_wr_pool);
1270 }
1271 
1272 static void
1273 test_nvmf_rdma_resources_create(void)
1274 {
1275 	static struct spdk_nvmf_rdma_resources *rdma_resource;
1276 	struct spdk_nvmf_rdma_resource_opts opts = {};
1277 	struct spdk_nvmf_rdma_qpair qpair = {};
1278 	struct spdk_nvmf_rdma_recv *recv = NULL;
1279 	struct spdk_nvmf_rdma_request *req = NULL;
1280 	const int DEPTH = 128;
1281 
1282 	opts.max_queue_depth = DEPTH;
1283 	opts.in_capsule_data_size = 4096;
1284 	opts.shared = true;
1285 	opts.qpair = &qpair;
1286 
1287 	rdma_resource = nvmf_rdma_resources_create(&opts);
1288 	CU_ASSERT(rdma_resource != NULL);
1289 	/* Just check first and last entry */
1290 	recv = &rdma_resource->recvs[0];
1291 	req = &rdma_resource->reqs[0];
1292 	CU_ASSERT(recv->rdma_wr.type == RDMA_WR_TYPE_RECV);
1293 	CU_ASSERT((uintptr_t)recv->buf == (uintptr_t)(rdma_resource->bufs));
1294 	CU_ASSERT(recv->sgl[0].addr == (uintptr_t)&rdma_resource->cmds[0]);
1295 	CU_ASSERT(recv->sgl[0].length == sizeof(rdma_resource->cmds[0]));
1296 	CU_ASSERT(recv->sgl[0].lkey == RDMA_UT_LKEY);
1297 	CU_ASSERT(recv->wr.num_sge == 2);
1298 	CU_ASSERT(recv->wr.wr_id == (uintptr_t)&rdma_resource->recvs[0].rdma_wr);
1299 	CU_ASSERT(recv->wr.sg_list == rdma_resource->recvs[0].sgl);
1300 	CU_ASSERT(req->req.rsp == &rdma_resource->cpls[0]);
1301 	CU_ASSERT(req->rsp.sgl[0].addr == (uintptr_t)&rdma_resource->cpls[0]);
1302 	CU_ASSERT(req->rsp.sgl[0].length == sizeof(rdma_resource->cpls[0]));
1303 	CU_ASSERT(req->rsp.sgl[0].lkey == RDMA_UT_LKEY);
1304 	CU_ASSERT(req->rsp_wr.type == RDMA_WR_TYPE_SEND);
1305 	CU_ASSERT(req->rsp.wr.wr_id == (uintptr_t)&rdma_resource->reqs[0].rsp_wr);
1306 	CU_ASSERT(req->rsp.wr.next == NULL);
1307 	CU_ASSERT(req->rsp.wr.opcode == IBV_WR_SEND);
1308 	CU_ASSERT(req->rsp.wr.send_flags == IBV_SEND_SIGNALED);
1309 	CU_ASSERT(req->rsp.wr.sg_list == rdma_resource->reqs[0].rsp.sgl);
1310 	CU_ASSERT(req->rsp.wr.num_sge == NVMF_DEFAULT_RSP_SGE);
1311 	CU_ASSERT(req->data_wr.type == RDMA_WR_TYPE_DATA);
1312 	CU_ASSERT(req->data.wr.wr_id == (uintptr_t)&rdma_resource->reqs[0].data_wr);
1313 	CU_ASSERT(req->data.wr.next == NULL);
1314 	CU_ASSERT(req->data.wr.send_flags == IBV_SEND_SIGNALED);
1315 	CU_ASSERT(req->data.wr.sg_list == rdma_resource->reqs[0].data.sgl);
1316 	CU_ASSERT(req->data.wr.num_sge == SPDK_NVMF_MAX_SGL_ENTRIES);
1317 	CU_ASSERT(req->state == RDMA_REQUEST_STATE_FREE);
1318 
1319 	recv = &rdma_resource->recvs[DEPTH - 1];
1320 	req = &rdma_resource->reqs[DEPTH - 1];
1321 	CU_ASSERT(recv->rdma_wr.type == RDMA_WR_TYPE_RECV);
1322 	CU_ASSERT((uintptr_t)recv->buf == (uintptr_t)(rdma_resource->bufs +
1323 			(DEPTH - 1) * 4096));
1324 	CU_ASSERT(recv->sgl[0].addr == (uintptr_t)&rdma_resource->cmds[DEPTH - 1]);
1325 	CU_ASSERT(recv->sgl[0].length == sizeof(rdma_resource->cmds[DEPTH - 1]));
1326 	CU_ASSERT(recv->sgl[0].lkey == RDMA_UT_LKEY);
1327 	CU_ASSERT(recv->wr.num_sge == 2);
1328 	CU_ASSERT(recv->wr.wr_id == (uintptr_t)&rdma_resource->recvs[DEPTH - 1].rdma_wr);
1329 	CU_ASSERT(recv->wr.sg_list == rdma_resource->recvs[DEPTH - 1].sgl);
1330 	CU_ASSERT(req->req.rsp == &rdma_resource->cpls[DEPTH - 1]);
1331 	CU_ASSERT(req->rsp.sgl[0].addr == (uintptr_t)&rdma_resource->cpls[DEPTH - 1]);
1332 	CU_ASSERT(req->rsp.sgl[0].length == sizeof(rdma_resource->cpls[DEPTH - 1]));
1333 	CU_ASSERT(req->rsp.sgl[0].lkey == RDMA_UT_LKEY);
1334 	CU_ASSERT(req->rsp_wr.type == RDMA_WR_TYPE_SEND);
1335 	CU_ASSERT(req->rsp.wr.wr_id == (uintptr_t)&req->rsp_wr);
1336 	CU_ASSERT(req->rsp.wr.next == NULL);
1337 	CU_ASSERT(req->rsp.wr.opcode == IBV_WR_SEND);
1338 	CU_ASSERT(req->rsp.wr.send_flags == IBV_SEND_SIGNALED);
1339 	CU_ASSERT(req->rsp.wr.sg_list == rdma_resource->reqs[DEPTH - 1].rsp.sgl);
1340 	CU_ASSERT(req->rsp.wr.num_sge == NVMF_DEFAULT_RSP_SGE);
1341 	CU_ASSERT(req->data_wr.type == RDMA_WR_TYPE_DATA);
1342 	CU_ASSERT(req->data.wr.wr_id == (uintptr_t)&req->data_wr);
1343 	CU_ASSERT(req->data.wr.next == NULL);
1344 	CU_ASSERT(req->data.wr.send_flags == IBV_SEND_SIGNALED);
1345 	CU_ASSERT(req->data.wr.sg_list == rdma_resource->reqs[DEPTH - 1].data.sgl);
1346 	CU_ASSERT(req->data.wr.num_sge == SPDK_NVMF_MAX_SGL_ENTRIES);
1347 	CU_ASSERT(req->state == RDMA_REQUEST_STATE_FREE);
1348 
1349 	nvmf_rdma_resources_destroy(rdma_resource);
1350 }
1351 
1352 static void
1353 test_nvmf_rdma_qpair_compare(void)
1354 {
1355 	struct spdk_nvmf_rdma_qpair rqpair1 = {}, rqpair2 = {};
1356 
1357 	rqpair1.qp_num = 0;
1358 	rqpair2.qp_num = UINT32_MAX;
1359 
1360 	CU_ASSERT(nvmf_rdma_qpair_compare(&rqpair1, &rqpair2) < 0);
1361 	CU_ASSERT(nvmf_rdma_qpair_compare(&rqpair2, &rqpair1) > 0);
1362 }
1363 
1364 static void
1365 test_nvmf_rdma_resize_cq(void)
1366 {
1367 	int rc = -1;
1368 	int tnum_wr = 0;
1369 	int tnum_cqe = 0;
1370 	struct spdk_nvmf_rdma_qpair rqpair = {};
1371 	struct spdk_nvmf_rdma_poller rpoller = {};
1372 	struct spdk_nvmf_rdma_device rdevice = {};
1373 	struct ibv_context ircontext = {};
1374 	struct ibv_device idevice = {};
1375 
1376 	rdevice.context = &ircontext;
1377 	rqpair.poller = &rpoller;
1378 	ircontext.device = &idevice;
1379 
1380 	/* Test1: Current capacity support required size. */
1381 	rpoller.required_num_wr = 10;
1382 	rpoller.num_cqe = 20;
1383 	rqpair.max_queue_depth = 2;
1384 	tnum_wr = rpoller.required_num_wr;
1385 	tnum_cqe = rpoller.num_cqe;
1386 
1387 	rc = nvmf_rdma_resize_cq(&rqpair, &rdevice);
1388 	CU_ASSERT(rc == 0);
1389 	CU_ASSERT(rpoller.required_num_wr == 10 + MAX_WR_PER_QP(rqpair.max_queue_depth));
1390 	CU_ASSERT(rpoller.required_num_wr > tnum_wr);
1391 	CU_ASSERT(rpoller.num_cqe == tnum_cqe);
1392 
1393 	/* Test2: iWARP doesn't support CQ resize. */
1394 	tnum_wr = rpoller.required_num_wr;
1395 	tnum_cqe = rpoller.num_cqe;
1396 	idevice.transport_type = IBV_TRANSPORT_IWARP;
1397 
1398 	rc = nvmf_rdma_resize_cq(&rqpair, &rdevice);
1399 	CU_ASSERT(rc == -1);
1400 	CU_ASSERT(rpoller.required_num_wr == tnum_wr);
1401 	CU_ASSERT(rpoller.num_cqe == tnum_cqe);
1402 
1403 
1404 	/* Test3: RDMA CQE requirement exceeds device max_cqe limitation. */
1405 	tnum_wr = rpoller.required_num_wr;
1406 	tnum_cqe = rpoller.num_cqe;
1407 	idevice.transport_type = IBV_TRANSPORT_UNKNOWN;
1408 	rdevice.attr.max_cqe = 3;
1409 
1410 	rc = nvmf_rdma_resize_cq(&rqpair, &rdevice);
1411 	CU_ASSERT(rc == -1);
1412 	CU_ASSERT(rpoller.required_num_wr == tnum_wr);
1413 	CU_ASSERT(rpoller.num_cqe == tnum_cqe);
1414 
1415 	/* Test4: RDMA CQ resize failed. */
1416 	tnum_wr = rpoller.required_num_wr;
1417 	tnum_cqe = rpoller.num_cqe;
1418 	idevice.transport_type = IBV_TRANSPORT_IB;
1419 	rdevice.attr.max_cqe = 30;
1420 	MOCK_SET(ibv_resize_cq, -1);
1421 
1422 	rc = nvmf_rdma_resize_cq(&rqpair, &rdevice);
1423 	CU_ASSERT(rc == -1);
1424 	CU_ASSERT(rpoller.required_num_wr == tnum_wr);
1425 	CU_ASSERT(rpoller.num_cqe == tnum_cqe);
1426 
1427 	/* Test5: RDMA CQ resize success. rsize = MIN(MAX(num_cqe * 2, required_num_wr), device->attr.max_cqe). */
1428 	tnum_wr = rpoller.required_num_wr;
1429 	tnum_cqe = rpoller.num_cqe;
1430 	MOCK_SET(ibv_resize_cq, 0);
1431 
1432 	rc = nvmf_rdma_resize_cq(&rqpair, &rdevice);
1433 	CU_ASSERT(rc == 0);
1434 	CU_ASSERT(rpoller.num_cqe == 30);
1435 	CU_ASSERT(rpoller.required_num_wr == 18 + MAX_WR_PER_QP(rqpair.max_queue_depth));
1436 	CU_ASSERT(rpoller.required_num_wr > tnum_wr);
1437 	CU_ASSERT(rpoller.num_cqe > tnum_cqe);
1438 }
1439 
1440 int
1441 main(int argc, char **argv)
1442 {
1443 	CU_pSuite	suite = NULL;
1444 	unsigned int	num_failures;
1445 
1446 	CU_initialize_registry();
1447 
1448 	suite = CU_add_suite("nvmf", NULL, NULL);
1449 
1450 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl);
1451 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_process);
1452 	CU_ADD_TEST(suite, test_nvmf_rdma_get_optimal_poll_group);
1453 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl_with_md);
1454 	CU_ADD_TEST(suite, test_nvmf_rdma_opts_init);
1455 	CU_ADD_TEST(suite, test_nvmf_rdma_request_free_data);
1456 	CU_ADD_TEST(suite, test_nvmf_rdma_resources_create);
1457 	CU_ADD_TEST(suite, test_nvmf_rdma_qpair_compare);
1458 	CU_ADD_TEST(suite, test_nvmf_rdma_resize_cq);
1459 
1460 	num_failures = spdk_ut_run_tests(argc, argv, NULL);
1461 	CU_cleanup_registry();
1462 	return num_failures;
1463 }
1464