xref: /spdk/test/unit/lib/nvmf/rdma.c/rdma_ut.c (revision 45a053c5777494f4e8ce4bc1191c9de3920377f7)
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 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,
35 		nvmf_qpair_disconnect_cb cb_fn, void *ctx), 0);
36 DEFINE_STUB(spdk_nvmf_qpair_get_listen_trid, int,
37 	    (struct spdk_nvmf_qpair *qpair, struct spdk_nvme_transport_id *trid), 0);
38 DEFINE_STUB_V(spdk_mem_map_free, (struct spdk_mem_map **pmap));
39 
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(spdk_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->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_rdma_device device;
160 	struct spdk_nvmf_rdma_request rdma_req = {};
161 	struct spdk_nvmf_rdma_recv recv;
162 	struct spdk_nvmf_rdma_poll_group group;
163 	struct spdk_nvmf_rdma_qpair rqpair;
164 	struct spdk_nvmf_rdma_poller poller;
165 	union nvmf_c2h_msg cpl;
166 	union nvmf_h2c_msg cmd;
167 	struct spdk_nvme_sgl_descriptor *sgl;
168 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
169 	struct spdk_nvmf_rdma_request_data data;
170 	int rc, i;
171 	uint32_t sgl_length;
172 
173 	data.wr.sg_list = data.sgl;
174 	group.group.transport = &rtransport.transport;
175 	poller.group = &group;
176 	rqpair.poller = &poller;
177 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
178 
179 	sgl = &cmd.nvme_cmd.dptr.sgl1;
180 	rdma_req.recv = &recv;
181 	rdma_req.req.cmd = &cmd;
182 	rdma_req.req.rsp = &cpl;
183 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
184 	rdma_req.req.qpair = &rqpair.qpair;
185 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
186 
187 	rtransport.transport.opts = g_rdma_ut_transport_opts;
188 	rtransport.data_wr_pool = NULL;
189 
190 	device.attr.device_cap_flags = 0;
191 	sgl->keyed.key = 0xEEEE;
192 	sgl->address = 0xFFFF;
193 	rdma_req.recv->buf = (void *)0xDDDD;
194 
195 	/* Test 1: sgl type: keyed data block subtype: address */
196 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
197 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
198 
199 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size */
200 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
201 	reset_nvmf_rdma_request(&rdma_req);
202 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
203 
204 	device.map = (void *)0x0;
205 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
206 	CU_ASSERT(rc == 0);
207 	CU_ASSERT(rdma_req.req.data_from_pool == true);
208 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
209 	CU_ASSERT((uint64_t)rdma_req.req.iovcnt == 1);
210 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
211 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
212 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
213 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
214 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
215 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
216 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
217 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
218 
219 	/* Part 2: simple I/O, one SGL larger than the transport io unit size (equal to the max io size) */
220 	reset_nvmf_rdma_request(&rdma_req);
221 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
222 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
223 
224 	CU_ASSERT(rc == 0);
225 	CU_ASSERT(rdma_req.req.data_from_pool == true);
226 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO);
227 	CU_ASSERT(rdma_req.data.wr.num_sge == RDMA_UT_UNITS_IN_MAX_IO);
228 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
229 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
230 	for (i = 0; i < RDMA_UT_UNITS_IN_MAX_IO; i++) {
231 		CU_ASSERT((uint64_t)rdma_req.req.iov[i].iov_base == 0x2000);
232 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
233 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
234 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
235 	}
236 
237 	/* Part 3: simple I/O one SGL larger than the transport max io size */
238 	reset_nvmf_rdma_request(&rdma_req);
239 	sgl->keyed.length = rtransport.transport.opts.max_io_size * 2;
240 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
241 
242 	CU_ASSERT(rc == -1);
243 
244 	/* Part 4: Pretend there are no buffer pools */
245 	MOCK_SET(spdk_iobuf_get, NULL);
246 	reset_nvmf_rdma_request(&rdma_req);
247 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
248 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
249 
250 	CU_ASSERT(rc == 0);
251 	CU_ASSERT(rdma_req.req.data_from_pool == false);
252 	CU_ASSERT(rdma_req.req.iovcnt == 0);
253 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
254 	CU_ASSERT(rdma_req.req.iov[0].iov_base == NULL);
255 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0);
256 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 0);
257 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == 0);
258 
259 	rdma_req.recv->buf = (void *)0xDDDD;
260 	/* Test 2: sgl type: keyed data block subtype: offset (in capsule data) */
261 	sgl->generic.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
262 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
263 
264 	/* Part 1: Normal I/O smaller than in capsule data size no offset */
265 	reset_nvmf_rdma_request(&rdma_req);
266 	sgl->address = 0;
267 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
268 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
269 
270 	CU_ASSERT(rc == 0);
271 	CU_ASSERT(rdma_req.req.iovcnt == 1);
272 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)0xDDDD);
273 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.in_capsule_data_size);
274 	CU_ASSERT(rdma_req.req.data_from_pool == false);
275 
276 	/* Part 2: I/O offset + length too large */
277 	reset_nvmf_rdma_request(&rdma_req);
278 	sgl->address = rtransport.transport.opts.in_capsule_data_size;
279 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
280 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
281 
282 	CU_ASSERT(rc == -1);
283 
284 	/* Part 3: I/O too large */
285 	reset_nvmf_rdma_request(&rdma_req);
286 	sgl->address = 0;
287 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size * 2;
288 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
289 
290 	CU_ASSERT(rc == -1);
291 
292 	/* Test 3: Multi SGL */
293 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
294 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
295 	sgl->address = 0;
296 	rdma_req.recv->buf = (void *)&sgl_desc;
297 	MOCK_SET(spdk_iobuf_get, &data);
298 	MOCK_SET(spdk_mempool_get, &data);
299 
300 	/* part 1: 2 segments each with 1 wr. */
301 	reset_nvmf_rdma_request(&rdma_req);
302 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
303 	for (i = 0; i < 2; i++) {
304 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
305 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
306 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size;
307 		sgl_desc[i].address = 0x4000 + i * rtransport.transport.opts.io_unit_size;
308 		sgl_desc[i].keyed.key = 0x44;
309 	}
310 
311 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
312 
313 	CU_ASSERT(rc == 0);
314 	CU_ASSERT(rdma_req.req.data_from_pool == true);
315 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 2);
316 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
317 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
318 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
319 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
320 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
321 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size);
322 	CU_ASSERT(data.wr.num_sge == 1);
323 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
324 
325 	/* part 2: 2 segments, each with 1 wr containing 8 sge_elements */
326 	reset_nvmf_rdma_request(&rdma_req);
327 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
328 	for (i = 0; i < 2; i++) {
329 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
330 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
331 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size * 8;
332 		sgl_desc[i].address = 0x4000 + i * 8 * rtransport.transport.opts.io_unit_size;
333 		sgl_desc[i].keyed.key = 0x44;
334 	}
335 
336 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
337 
338 	CU_ASSERT(rc == 0);
339 	CU_ASSERT(rdma_req.req.data_from_pool == true);
340 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
341 	CU_ASSERT(rdma_req.req.iovcnt == 16);
342 	CU_ASSERT(rdma_req.data.wr.num_sge == 8);
343 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
344 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
345 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
346 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
347 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 8);
348 	CU_ASSERT(data.wr.num_sge == 8);
349 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
350 
351 	/* part 3: 2 segments, one very large, one very small */
352 	reset_nvmf_rdma_request(&rdma_req);
353 	for (i = 0; i < 2; i++) {
354 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
355 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
356 		sgl_desc[i].keyed.key = 0x44;
357 	}
358 
359 	sgl_desc[0].keyed.length = rtransport.transport.opts.io_unit_size * 15 +
360 				   rtransport.transport.opts.io_unit_size / 2;
361 	sgl_desc[0].address = 0x4000;
362 	sgl_desc[1].keyed.length = rtransport.transport.opts.io_unit_size / 2;
363 	sgl_desc[1].address = 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
364 			      rtransport.transport.opts.io_unit_size / 2;
365 
366 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
367 
368 	CU_ASSERT(rc == 0);
369 	CU_ASSERT(rdma_req.req.data_from_pool == true);
370 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
371 	CU_ASSERT(rdma_req.req.iovcnt == 16);
372 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
373 	for (i = 0; i < 15; i++) {
374 		CU_ASSERT(rdma_req.data.sgl[i].length == rtransport.transport.opts.io_unit_size);
375 	}
376 	CU_ASSERT(rdma_req.data.sgl[15].length == rtransport.transport.opts.io_unit_size / 2);
377 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
378 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
379 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
380 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
381 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
382 		  rtransport.transport.opts.io_unit_size / 2);
383 	CU_ASSERT(data.sgl[0].length == rtransport.transport.opts.io_unit_size / 2);
384 	CU_ASSERT(data.wr.num_sge == 1);
385 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
386 
387 	/* part 4: 2 SGL descriptors, each length is transport buffer / 2
388 	 * 1 transport buffers should be allocated */
389 	reset_nvmf_rdma_request(&rdma_req);
390 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
391 	sgl_length = rtransport.transport.opts.io_unit_size / 2;
392 	for (i = 0; i < 2; i++) {
393 		sgl_desc[i].keyed.length = sgl_length;
394 		sgl_desc[i].address = 0x4000 + i * sgl_length;
395 	}
396 
397 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
398 
399 	CU_ASSERT(rc == 0);
400 	CU_ASSERT(rdma_req.req.data_from_pool == true);
401 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size);
402 	CU_ASSERT(rdma_req.req.iovcnt == 1);
403 
404 	CU_ASSERT(rdma_req.data.sgl[0].length == sgl_length);
405 	/* We mocked mempool_get to return address of data variable. Mempool is used
406 	 * to get both additional WRs and data buffers, so data points to &data */
407 	CU_ASSERT(rdma_req.data.sgl[0].addr == (uint64_t)&data);
408 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
409 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
410 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
411 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
412 
413 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
414 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + sgl_length);
415 	CU_ASSERT(data.sgl[0].length == sgl_length);
416 	CU_ASSERT(data.sgl[0].addr == (uint64_t)&data + sgl_length);
417 	CU_ASSERT(data.wr.num_sge == 1);
418 
419 	MOCK_CLEAR(spdk_mempool_get);
420 	MOCK_CLEAR(spdk_iobuf_get);
421 
422 	reset_nvmf_rdma_request(&rdma_req);
423 }
424 
425 static struct spdk_nvmf_rdma_recv *
426 create_recv(struct spdk_nvmf_rdma_qpair *rqpair, enum spdk_nvme_nvm_opcode opc)
427 {
428 	struct spdk_nvmf_rdma_recv *rdma_recv;
429 	union nvmf_h2c_msg *cmd;
430 	struct spdk_nvme_sgl_descriptor *sgl;
431 
432 	rdma_recv = calloc(1, sizeof(*rdma_recv));
433 	rdma_recv->qpair = rqpair;
434 	cmd = calloc(1, sizeof(*cmd));
435 	rdma_recv->sgl[0].addr = (uintptr_t)cmd;
436 	cmd->nvme_cmd.opc = opc;
437 	sgl = &cmd->nvme_cmd.dptr.sgl1;
438 	sgl->keyed.key = 0xEEEE;
439 	sgl->address = 0xFFFF;
440 	sgl->keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
441 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
442 	sgl->keyed.length = 1;
443 
444 	return rdma_recv;
445 }
446 
447 static void
448 free_recv(struct spdk_nvmf_rdma_recv *rdma_recv)
449 {
450 	free((void *)rdma_recv->sgl[0].addr);
451 	free(rdma_recv);
452 }
453 
454 static struct spdk_nvmf_rdma_request *
455 create_req(struct spdk_nvmf_rdma_qpair *rqpair,
456 	   struct spdk_nvmf_rdma_recv *rdma_recv)
457 {
458 	struct spdk_nvmf_rdma_request *rdma_req;
459 	union nvmf_c2h_msg *cpl;
460 
461 	rdma_req = calloc(1, sizeof(*rdma_req));
462 	rdma_req->recv = rdma_recv;
463 	rdma_req->req.qpair = &rqpair->qpair;
464 	rdma_req->state = RDMA_REQUEST_STATE_NEW;
465 	rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr;
466 	rdma_req->data.wr.sg_list = rdma_req->data.sgl;
467 	cpl = calloc(1, sizeof(*cpl));
468 	rdma_req->rsp.sgl[0].addr = (uintptr_t)cpl;
469 	rdma_req->req.rsp = cpl;
470 
471 	return rdma_req;
472 }
473 
474 static void
475 free_req(struct spdk_nvmf_rdma_request *rdma_req)
476 {
477 	free((void *)rdma_req->rsp.sgl[0].addr);
478 	free(rdma_req);
479 }
480 
481 static void
482 qpair_reset(struct spdk_nvmf_rdma_qpair *rqpair,
483 	    struct spdk_nvmf_rdma_poller *poller,
484 	    struct spdk_nvmf_rdma_device *device,
485 	    struct spdk_nvmf_rdma_resources *resources,
486 	    struct spdk_nvmf_transport *transport)
487 {
488 	memset(rqpair, 0, sizeof(*rqpair));
489 	STAILQ_INIT(&rqpair->pending_rdma_write_queue);
490 	STAILQ_INIT(&rqpair->pending_rdma_read_queue);
491 	STAILQ_INIT(&rqpair->pending_rdma_send_queue);
492 	rqpair->poller = poller;
493 	rqpair->device = device;
494 	rqpair->resources = resources;
495 	rqpair->qpair.qid = 1;
496 	rqpair->ibv_state = IBV_QPS_RTS;
497 	rqpair->qpair.state = SPDK_NVMF_QPAIR_ACTIVE;
498 	rqpair->max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
499 	rqpair->max_send_depth = 16;
500 	rqpair->max_read_depth = 16;
501 	rqpair->qpair.transport = transport;
502 }
503 
504 static void
505 poller_reset(struct spdk_nvmf_rdma_poller *poller,
506 	     struct spdk_nvmf_rdma_poll_group *group)
507 {
508 	memset(poller, 0, sizeof(*poller));
509 	STAILQ_INIT(&poller->qpairs_pending_recv);
510 	STAILQ_INIT(&poller->qpairs_pending_send);
511 	poller->group = group;
512 }
513 
514 static void
515 test_spdk_nvmf_rdma_request_process(void)
516 {
517 	struct spdk_nvmf_rdma_transport rtransport = {};
518 	struct spdk_nvmf_rdma_poll_group group = {};
519 	struct spdk_nvmf_rdma_poller poller = {};
520 	struct spdk_nvmf_rdma_device device = {};
521 	struct spdk_nvmf_rdma_resources resources = {};
522 	struct spdk_nvmf_rdma_qpair rqpair = {};
523 	struct spdk_nvmf_rdma_recv *rdma_recv;
524 	struct spdk_nvmf_rdma_request *rdma_req;
525 	struct spdk_iobuf_channel ch = {};
526 	bool progress;
527 
528 	group.group.buf_cache = &ch;
529 
530 	STAILQ_INIT(&group.group.pending_buf_queue);
531 	poller_reset(&poller, &group);
532 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
533 
534 	rtransport.transport.opts = g_rdma_ut_transport_opts;
535 	rtransport.data_wr_pool = spdk_mempool_create("test_wr_pool", 128,
536 				  sizeof(struct spdk_nvmf_rdma_request_data),
537 				  0, 0);
538 	MOCK_CLEAR(spdk_iobuf_get);
539 
540 	device.attr.device_cap_flags = 0;
541 	device.map = (void *)0x0;
542 
543 	/* Test 1: single SGL READ request */
544 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_READ);
545 	rdma_req = create_req(&rqpair, rdma_recv);
546 	rqpair.current_recv_depth = 1;
547 	/* NEW -> EXECUTING */
548 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
549 	CU_ASSERT(progress == true);
550 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
551 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST);
552 	/* EXECUTED -> TRANSFERRING_C2H */
553 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
554 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
555 	CU_ASSERT(progress == true);
556 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST);
557 	CU_ASSERT(rdma_req->recv == NULL);
558 	/* COMPLETED -> FREE */
559 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
560 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
561 	CU_ASSERT(progress == true);
562 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
563 
564 	free_recv(rdma_recv);
565 	free_req(rdma_req);
566 	poller_reset(&poller, &group);
567 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
568 
569 	/* Test 2: single SGL WRITE request */
570 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
571 	rdma_req = create_req(&rqpair, rdma_recv);
572 	rqpair.current_recv_depth = 1;
573 	/* NEW -> TRANSFERRING_H2C */
574 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
575 	CU_ASSERT(progress == true);
576 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
577 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
578 	STAILQ_INIT(&poller.qpairs_pending_send);
579 	/* READY_TO_EXECUTE -> EXECUTING */
580 	rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
581 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
582 	CU_ASSERT(progress == true);
583 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
584 	/* EXECUTED -> COMPLETING */
585 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
586 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
587 	CU_ASSERT(progress == true);
588 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
589 	CU_ASSERT(rdma_req->recv == NULL);
590 	/* COMPLETED -> FREE */
591 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
592 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
593 	CU_ASSERT(progress == true);
594 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
595 
596 	free_recv(rdma_recv);
597 	free_req(rdma_req);
598 	poller_reset(&poller, &group);
599 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
600 
601 	/* Test 3: WRITE+WRITE ibv_send batching */
602 	{
603 		struct spdk_nvmf_rdma_recv *recv1, *recv2;
604 		struct spdk_nvmf_rdma_request *req1, *req2;
605 		recv1 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
606 		req1 = create_req(&rqpair, recv1);
607 		recv2 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
608 		req2 = create_req(&rqpair, recv2);
609 
610 		/* WRITE 1: NEW -> TRANSFERRING_H2C */
611 		rqpair.current_recv_depth = 1;
612 		nvmf_rdma_request_process(&rtransport, req1);
613 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
614 
615 		/* WRITE 2: NEW -> TRANSFERRING_H2C */
616 		rqpair.current_recv_depth = 2;
617 		nvmf_rdma_request_process(&rtransport, req2);
618 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
619 
620 		STAILQ_INIT(&poller.qpairs_pending_send);
621 
622 		/* WRITE 1 completes before WRITE 2 has finished RDMA reading */
623 		/* WRITE 1: READY_TO_EXECUTE -> EXECUTING */
624 		req1->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
625 		nvmf_rdma_request_process(&rtransport, req1);
626 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_EXECUTING);
627 		/* WRITE 1: EXECUTED -> COMPLETING */
628 		req1->state = RDMA_REQUEST_STATE_EXECUTED;
629 		nvmf_rdma_request_process(&rtransport, req1);
630 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_COMPLETING);
631 		STAILQ_INIT(&poller.qpairs_pending_send);
632 		/* WRITE 1: COMPLETED -> FREE */
633 		req1->state = RDMA_REQUEST_STATE_COMPLETED;
634 		nvmf_rdma_request_process(&rtransport, req1);
635 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_FREE);
636 
637 		/* Now WRITE 2 has finished reading and completes */
638 		/* WRITE 2: COMPLETED -> FREE */
639 		/* WRITE 2: READY_TO_EXECUTE -> EXECUTING */
640 		req2->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
641 		nvmf_rdma_request_process(&rtransport, req2);
642 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_EXECUTING);
643 		/* WRITE 1: EXECUTED -> COMPLETING */
644 		req2->state = RDMA_REQUEST_STATE_EXECUTED;
645 		nvmf_rdma_request_process(&rtransport, req2);
646 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_COMPLETING);
647 		STAILQ_INIT(&poller.qpairs_pending_send);
648 		/* WRITE 1: COMPLETED -> FREE */
649 		req2->state = RDMA_REQUEST_STATE_COMPLETED;
650 		nvmf_rdma_request_process(&rtransport, req2);
651 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_FREE);
652 
653 		free_recv(recv1);
654 		free_req(req1);
655 		free_recv(recv2);
656 		free_req(req2);
657 		poller_reset(&poller, &group);
658 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
659 	}
660 
661 	/* Test 4, invalid command, check xfer type */
662 	{
663 		struct spdk_nvmf_rdma_recv *rdma_recv_inv;
664 		struct spdk_nvmf_rdma_request *rdma_req_inv;
665 		/* construct an opcode that specifies BIDIRECTIONAL transfer */
666 		uint8_t opc = 0x10 | SPDK_NVME_DATA_BIDIRECTIONAL;
667 
668 		rdma_recv_inv = create_recv(&rqpair, opc);
669 		rdma_req_inv = create_req(&rqpair, rdma_recv_inv);
670 
671 		/* NEW -> RDMA_REQUEST_STATE_COMPLETING */
672 		rqpair.current_recv_depth = 1;
673 		progress = nvmf_rdma_request_process(&rtransport, rdma_req_inv);
674 		CU_ASSERT(progress == true);
675 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_COMPLETING);
676 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sct == SPDK_NVME_SCT_GENERIC);
677 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sc == SPDK_NVME_SC_INVALID_OPCODE);
678 
679 		/* RDMA_REQUEST_STATE_COMPLETED -> FREE */
680 		rdma_req_inv->state = RDMA_REQUEST_STATE_COMPLETED;
681 		nvmf_rdma_request_process(&rtransport, rdma_req_inv);
682 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_FREE);
683 
684 		free_recv(rdma_recv_inv);
685 		free_req(rdma_req_inv);
686 		poller_reset(&poller, &group);
687 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
688 	}
689 
690 	/* Test 5: Write response waits in queue */
691 	{
692 		rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
693 		rdma_req = create_req(&rqpair, rdma_recv);
694 		rqpair.current_recv_depth = 1;
695 		/* NEW -> TRANSFERRING_H2C */
696 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
697 		CU_ASSERT(progress == true);
698 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
699 		CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
700 		STAILQ_INIT(&poller.qpairs_pending_send);
701 		/* READY_TO_EXECUTE -> EXECUTING */
702 		rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
703 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
704 		CU_ASSERT(progress == true);
705 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
706 		/* EXECUTED -> COMPLETING */
707 		rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
708 		/* Send queue is full */
709 		rqpair.current_send_depth = rqpair.max_send_depth;
710 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
711 		CU_ASSERT(progress == true);
712 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING);
713 		CU_ASSERT(rdma_req == STAILQ_FIRST(&rqpair.pending_rdma_send_queue));
714 
715 		/* Send queue is still full */
716 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
717 		CU_ASSERT(progress == false);
718 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING);
719 		CU_ASSERT(rdma_req == STAILQ_FIRST(&rqpair.pending_rdma_send_queue));
720 
721 		/* Slot is available */
722 		rqpair.current_send_depth = rqpair.max_send_depth - 1;
723 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
724 		CU_ASSERT(progress == true);
725 		CU_ASSERT(STAILQ_EMPTY(&rqpair.pending_rdma_send_queue));
726 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
727 		CU_ASSERT(rdma_req->recv == NULL);
728 		/* COMPLETED -> FREE */
729 		rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
730 		progress = nvmf_rdma_request_process(&rtransport, rdma_req);
731 		CU_ASSERT(progress == true);
732 		CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
733 
734 		free_recv(rdma_recv);
735 		free_req(rdma_req);
736 		poller_reset(&poller, &group);
737 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
738 
739 	}
740 
741 	spdk_mempool_free(rtransport.data_wr_pool);
742 }
743 
744 #define TEST_GROUPS_COUNT 5
745 static void
746 test_nvmf_rdma_get_optimal_poll_group(void)
747 {
748 	struct spdk_nvmf_rdma_transport rtransport = {};
749 	struct spdk_nvmf_transport *transport = &rtransport.transport;
750 	struct spdk_nvmf_rdma_qpair rqpair = {};
751 	struct spdk_nvmf_transport_poll_group *groups[TEST_GROUPS_COUNT];
752 	struct spdk_nvmf_rdma_poll_group *rgroups[TEST_GROUPS_COUNT];
753 	struct spdk_nvmf_transport_poll_group *result;
754 	struct spdk_nvmf_poll_group group = {};
755 	uint32_t i;
756 
757 	rqpair.qpair.transport = transport;
758 	TAILQ_INIT(&rtransport.poll_groups);
759 
760 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
761 		groups[i] = nvmf_rdma_poll_group_create(transport, NULL);
762 		CU_ASSERT(groups[i] != NULL);
763 		groups[i]->group = &group;
764 		rgroups[i] = SPDK_CONTAINEROF(groups[i], struct spdk_nvmf_rdma_poll_group, group);
765 		groups[i]->transport = transport;
766 	}
767 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[0]);
768 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[0]);
769 
770 	/* Emulate connection of %TEST_GROUPS_COUNT% initiators - each creates 1 admin and 1 io qp */
771 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
772 		rqpair.qpair.qid = 0;
773 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
774 		CU_ASSERT(result == groups[i]);
775 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
776 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i]);
777 
778 		rqpair.qpair.qid = 1;
779 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
780 		CU_ASSERT(result == groups[i]);
781 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
782 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
783 	}
784 	/* wrap around, admin/io pg point to the first pg
785 	   Destroy all poll groups except of the last one */
786 	for (i = 0; i < TEST_GROUPS_COUNT - 1; i++) {
787 		nvmf_rdma_poll_group_destroy(groups[i]);
788 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[i + 1]);
789 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i + 1]);
790 	}
791 
792 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
793 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
794 
795 	/* Check that pointers to the next admin/io poll groups are not changed */
796 	rqpair.qpair.qid = 0;
797 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
798 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
799 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
800 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
801 
802 	rqpair.qpair.qid = 1;
803 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
804 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
805 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
806 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
807 
808 	/* Remove the last poll group, check that pointers are NULL */
809 	nvmf_rdma_poll_group_destroy(groups[TEST_GROUPS_COUNT - 1]);
810 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == NULL);
811 	CU_ASSERT(rtransport.conn_sched.next_io_pg == NULL);
812 
813 	/* Request optimal poll group, result must be NULL */
814 	rqpair.qpair.qid = 0;
815 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
816 	CU_ASSERT(result == NULL);
817 
818 	rqpair.qpair.qid = 1;
819 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
820 	CU_ASSERT(result == NULL);
821 }
822 #undef TEST_GROUPS_COUNT
823 
824 static void
825 test_spdk_nvmf_rdma_request_parse_sgl_with_md(void)
826 {
827 	struct spdk_nvmf_rdma_transport rtransport;
828 	struct spdk_nvmf_rdma_device device;
829 	struct spdk_nvmf_rdma_request rdma_req = {};
830 	struct spdk_nvmf_rdma_recv recv;
831 	struct spdk_nvmf_rdma_poll_group group;
832 	struct spdk_nvmf_rdma_qpair rqpair;
833 	struct spdk_nvmf_rdma_poller poller;
834 	union nvmf_c2h_msg cpl;
835 	union nvmf_h2c_msg cmd;
836 	struct spdk_nvme_sgl_descriptor *sgl;
837 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
838 	char data_buffer[8192];
839 	struct spdk_nvmf_rdma_request_data *data = (struct spdk_nvmf_rdma_request_data *)data_buffer;
840 	char data2_buffer[8192];
841 	struct spdk_nvmf_rdma_request_data *data2 = (struct spdk_nvmf_rdma_request_data *)data2_buffer;
842 	const uint32_t data_bs = 512;
843 	const uint32_t md_size = 8;
844 	int rc, i;
845 	struct spdk_dif_ctx_init_ext_opts dif_opts;
846 
847 	MOCK_CLEAR(spdk_mempool_get);
848 	MOCK_CLEAR(spdk_iobuf_get);
849 
850 	data->wr.sg_list = data->sgl;
851 	group.group.transport = &rtransport.transport;
852 	poller.group = &group;
853 	rqpair.poller = &poller;
854 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
855 
856 	sgl = &cmd.nvme_cmd.dptr.sgl1;
857 	rdma_req.recv = &recv;
858 	rdma_req.req.cmd = &cmd;
859 	rdma_req.req.rsp = &cpl;
860 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
861 	rdma_req.req.qpair = &rqpair.qpair;
862 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
863 
864 	rtransport.transport.opts = g_rdma_ut_transport_opts;
865 	rtransport.data_wr_pool = NULL;
866 
867 	device.attr.device_cap_flags = 0;
868 	device.map = NULL;
869 	sgl->keyed.key = 0xEEEE;
870 	sgl->address = 0xFFFF;
871 	rdma_req.recv->buf = (void *)0xDDDD;
872 
873 	/* Test 1: sgl type: keyed data block subtype: address */
874 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
875 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
876 
877 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size, block size 512 */
878 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
879 	reset_nvmf_rdma_request(&rdma_req);
880 	dif_opts.size = SPDK_SIZEOF(&dif_opts, dif_pi_format);
881 	dif_opts.dif_pi_format = SPDK_DIF_PI_FORMAT_16;
882 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
883 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
884 			  0, 0, 0, 0, 0, &dif_opts);
885 	rdma_req.req.dif_enabled = true;
886 	rtransport.transport.opts.io_unit_size = data_bs * 8;
887 	rdma_req.req.qpair->transport = &rtransport.transport;
888 	sgl->keyed.length = data_bs * 4;
889 
890 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
891 
892 	CU_ASSERT(rc == 0);
893 	CU_ASSERT(rdma_req.req.data_from_pool == true);
894 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
895 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
896 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
897 	CU_ASSERT(rdma_req.req.iovcnt == 1);
898 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
899 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
900 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
901 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
902 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
903 
904 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
905 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rdma_req.req.length);
906 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
907 
908 	/* Part 2: simple I/O, one SGL equal to io unit size, io_unit_size is not aligned with md_size,
909 		block size 512 */
910 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
911 	reset_nvmf_rdma_request(&rdma_req);
912 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
913 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
914 			  0, 0, 0, 0, 0, &dif_opts);
915 	rdma_req.req.dif_enabled = true;
916 	rtransport.transport.opts.io_unit_size = data_bs * 4;
917 	sgl->keyed.length = data_bs * 4;
918 
919 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
920 
921 	CU_ASSERT(rc == 0);
922 	CU_ASSERT(rdma_req.req.data_from_pool == true);
923 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
924 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
925 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
926 	CU_ASSERT(rdma_req.req.iovcnt == 2);
927 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
928 	CU_ASSERT(rdma_req.data.wr.num_sge == 5);
929 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
930 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
931 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
932 
933 	for (i = 0; i < 3; ++i) {
934 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
935 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
936 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
937 	}
938 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
939 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
940 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
941 
942 	/* 2nd buffer consumed */
943 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
944 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
945 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
946 
947 	/* Part 3: simple I/O, one SGL equal io unit size, io_unit_size is equal to block size 512 bytes */
948 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
949 	reset_nvmf_rdma_request(&rdma_req);
950 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
951 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
952 			  0, 0, 0, 0, 0, &dif_opts);
953 	rdma_req.req.dif_enabled = true;
954 	rtransport.transport.opts.io_unit_size = data_bs;
955 	sgl->keyed.length = data_bs;
956 
957 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
958 
959 	CU_ASSERT(rc == 0);
960 	CU_ASSERT(rdma_req.req.data_from_pool == true);
961 	CU_ASSERT(rdma_req.req.length == data_bs);
962 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
963 	CU_ASSERT(rdma_req.req.dif.elba_length == data_bs + md_size);
964 	CU_ASSERT(rdma_req.req.iovcnt == 2);
965 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
966 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
967 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
968 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
969 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
970 
971 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
972 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs);
973 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
974 
975 	CU_ASSERT(rdma_req.req.iovcnt == 2);
976 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)((unsigned long)0x2000));
977 	CU_ASSERT(rdma_req.req.iov[0].iov_len == data_bs);
978 	/* 2nd buffer consumed for metadata */
979 	CU_ASSERT(rdma_req.req.iov[1].iov_base == (void *)((unsigned long)0x2000));
980 	CU_ASSERT(rdma_req.req.iov[1].iov_len == md_size);
981 
982 	/* Part 4: simple I/O, one SGL equal io unit size, io_unit_size is aligned with md_size,
983 	   block size 512 */
984 	MOCK_SET(spdk_iobuf_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, &dif_opts);
989 	rdma_req.req.dif_enabled = true;
990 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
991 	sgl->keyed.length = data_bs * 4;
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 * 4);
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) * 4);
1000 	CU_ASSERT(rdma_req.req.iovcnt == 1);
1001 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1002 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1003 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1004 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1005 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1006 
1007 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1008 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rdma_req.req.length);
1009 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1010 
1011 	/* Part 5: simple I/O, one SGL equal to 2x io unit size, io_unit_size is aligned with md_size,
1012 	   block size 512 */
1013 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
1014 	reset_nvmf_rdma_request(&rdma_req);
1015 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1016 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1017 			  0, 0, 0, 0, 0, &dif_opts);
1018 	rdma_req.req.dif_enabled = true;
1019 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 2;
1020 	sgl->keyed.length = data_bs * 4;
1021 
1022 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1023 
1024 	CU_ASSERT(rc == 0);
1025 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1026 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1027 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1028 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1029 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1030 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1031 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1032 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1033 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1034 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1035 
1036 	for (i = 0; i < 2; ++i) {
1037 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
1038 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs * 2);
1039 	}
1040 
1041 	/* Part 6: simple I/O, one SGL larger than the transport io unit size, io_unit_size is not aligned to md_size,
1042 	   block size 512 */
1043 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
1044 	reset_nvmf_rdma_request(&rdma_req);
1045 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1046 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1047 			  0, 0, 0, 0, 0, &dif_opts);
1048 	rdma_req.req.dif_enabled = true;
1049 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1050 	sgl->keyed.length = data_bs * 6;
1051 
1052 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1053 
1054 	CU_ASSERT(rc == 0);
1055 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1056 	CU_ASSERT(rdma_req.req.length == data_bs * 6);
1057 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1058 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 6);
1059 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1060 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1061 	CU_ASSERT(rdma_req.data.wr.num_sge == 7);
1062 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1063 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1064 	CU_ASSERT((uint64_t)rdma_req.req.iov[0].iov_base == 0x2000);
1065 
1066 	for (i = 0; i < 3; ++i) {
1067 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1068 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1069 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1070 	}
1071 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
1072 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
1073 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
1074 
1075 	/* 2nd IO buffer consumed */
1076 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
1077 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
1078 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
1079 
1080 	CU_ASSERT(rdma_req.data.wr.sg_list[5].addr == 0x2000 + 24 + md_size);
1081 	CU_ASSERT(rdma_req.data.wr.sg_list[5].length == 512);
1082 	CU_ASSERT(rdma_req.data.wr.sg_list[5].lkey == RDMA_UT_LKEY);
1083 
1084 	CU_ASSERT(rdma_req.data.wr.sg_list[6].addr == 0x2000 + 24 + 512 + md_size * 2);
1085 	CU_ASSERT(rdma_req.data.wr.sg_list[6].length == 512);
1086 	CU_ASSERT(rdma_req.data.wr.sg_list[6].lkey == RDMA_UT_LKEY);
1087 
1088 	/* Part 7: simple I/O, number of SGL entries exceeds the number of entries
1089 	   one WR can hold. Additional WR is chained */
1090 	MOCK_SET(spdk_iobuf_get, data2_buffer);
1091 	MOCK_SET(spdk_mempool_get, data2_buffer);
1092 	reset_nvmf_rdma_request(&rdma_req);
1093 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1094 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1095 			  0, 0, 0, 0, 0, &dif_opts);
1096 	rdma_req.req.dif_enabled = true;
1097 	rtransport.transport.opts.io_unit_size = data_bs * 16;
1098 	sgl->keyed.length = data_bs * 16;
1099 
1100 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1101 
1102 	CU_ASSERT(rc == 0);
1103 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1104 	CU_ASSERT(rdma_req.req.length == data_bs * 16);
1105 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1106 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1107 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 16);
1108 	CU_ASSERT(rdma_req.req.iov[0].iov_base == data2_buffer);
1109 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
1110 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1111 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1112 
1113 	for (i = 0; i < 15; ++i) {
1114 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)data2_buffer + i * (data_bs + md_size));
1115 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1116 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1117 	}
1118 
1119 	/* 8192 - (512 + 8) * 15 = 392 */
1120 	CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)data2_buffer + i * (data_bs + md_size));
1121 	CU_ASSERT(rdma_req.data.wr.sg_list[i].length == 392);
1122 	CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1123 
1124 	/* additional wr from pool */
1125 	CU_ASSERT(rdma_req.data.wr.next == (void *)&data2->wr);
1126 	CU_ASSERT(rdma_req.data.wr.next->num_sge == 1);
1127 	CU_ASSERT(rdma_req.data.wr.next->next == &rdma_req.rsp.wr);
1128 	/* 2nd IO buffer */
1129 	CU_ASSERT(data2->wr.sg_list[0].addr == (uintptr_t)data2_buffer);
1130 	CU_ASSERT(data2->wr.sg_list[0].length == 120);
1131 	CU_ASSERT(data2->wr.sg_list[0].lkey == RDMA_UT_LKEY);
1132 
1133 	/* Part 8: simple I/O, data with metadata do not fit to 1 io_buffer */
1134 	MOCK_SET(spdk_iobuf_get, (void *)0x2000);
1135 	reset_nvmf_rdma_request(&rdma_req);
1136 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1137 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1138 			  0, 0, 0, 0, 0, &dif_opts);
1139 	rdma_req.req.dif_enabled = true;
1140 	rtransport.transport.opts.io_unit_size = 516;
1141 	sgl->keyed.length = data_bs * 2;
1142 
1143 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1144 
1145 	CU_ASSERT(rc == 0);
1146 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1147 	CU_ASSERT(rdma_req.req.length == data_bs * 2);
1148 	CU_ASSERT(rdma_req.req.iovcnt == 3);
1149 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1150 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 2);
1151 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)0x2000);
1152 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1153 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1154 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1155 
1156 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1157 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 512);
1158 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1159 
1160 	/* 2nd IO buffer consumed, offset 4 bytes due to part of the metadata
1161 	  is located at the beginning of that buffer */
1162 	CU_ASSERT(rdma_req.data.wr.sg_list[1].addr == 0x2000 + 4);
1163 	CU_ASSERT(rdma_req.data.wr.sg_list[1].length == 512);
1164 	CU_ASSERT(rdma_req.data.wr.sg_list[1].lkey == RDMA_UT_LKEY);
1165 
1166 	/* Test 2: Multi SGL */
1167 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1168 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1169 	sgl->address = 0;
1170 	rdma_req.recv->buf = (void *)&sgl_desc;
1171 	MOCK_SET(spdk_mempool_get, data_buffer);
1172 	MOCK_SET(spdk_iobuf_get, data_buffer);
1173 
1174 	/* part 1: 2 segments each with 1 wr. io_unit_size is aligned with data_bs + md_size */
1175 	reset_nvmf_rdma_request(&rdma_req);
1176 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1177 			  SPDK_DIF_TYPE1,
1178 			  SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1179 			  0, 0, 0, 0, 0, &dif_opts);
1180 	rdma_req.req.dif_enabled = true;
1181 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1182 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
1183 
1184 	for (i = 0; i < 2; i++) {
1185 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1186 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1187 		sgl_desc[i].keyed.length = data_bs * 4;
1188 		sgl_desc[i].address = 0x4000 + i * data_bs * 4;
1189 		sgl_desc[i].keyed.key = 0x44;
1190 	}
1191 
1192 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1193 
1194 	CU_ASSERT(rc == 0);
1195 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1196 	CU_ASSERT(rdma_req.req.length == data_bs * 4 * 2);
1197 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1198 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4 * 2);
1199 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1200 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == (uintptr_t)(data_buffer));
1201 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs * 4);
1202 
1203 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
1204 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
1205 	CU_ASSERT(rdma_req.data.wr.next == &data->wr);
1206 	CU_ASSERT(data->wr.wr.rdma.rkey == 0x44);
1207 	CU_ASSERT(data->wr.wr.rdma.remote_addr == 0x4000 + data_bs * 4);
1208 	CU_ASSERT(data->wr.num_sge == 1);
1209 	CU_ASSERT(data->wr.sg_list[0].addr == (uintptr_t)(data_buffer));
1210 	CU_ASSERT(data->wr.sg_list[0].length == data_bs * 4);
1211 
1212 	CU_ASSERT(data->wr.next == &rdma_req.rsp.wr);
1213 	reset_nvmf_rdma_request(&rdma_req);
1214 }
1215 
1216 static void
1217 test_nvmf_rdma_opts_init(void)
1218 {
1219 	struct spdk_nvmf_transport_opts	opts = {};
1220 
1221 	nvmf_rdma_opts_init(&opts);
1222 	CU_ASSERT(opts.max_queue_depth == SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH);
1223 	CU_ASSERT(opts.max_qpairs_per_ctrlr ==	SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR);
1224 	CU_ASSERT(opts.in_capsule_data_size ==	SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE);
1225 	CU_ASSERT(opts.max_io_size == SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE);
1226 	CU_ASSERT(opts.io_unit_size == SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
1227 	CU_ASSERT(opts.max_aq_depth == SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH);
1228 	CU_ASSERT(opts.num_shared_buffers == SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS);
1229 	CU_ASSERT(opts.buf_cache_size == SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE);
1230 	CU_ASSERT(opts.dif_insert_or_strip == SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP);
1231 	CU_ASSERT(opts.abort_timeout_sec == SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC);
1232 	CU_ASSERT(opts.transport_specific == NULL);
1233 }
1234 
1235 static void
1236 test_nvmf_rdma_request_free_data(void)
1237 {
1238 	struct spdk_nvmf_rdma_request rdma_req = {};
1239 	struct spdk_nvmf_rdma_transport rtransport = {};
1240 	struct spdk_nvmf_rdma_request_data *next_request_data = NULL;
1241 
1242 	MOCK_CLEAR(spdk_mempool_get);
1243 	rtransport.data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data",
1244 				  SPDK_NVMF_MAX_SGL_ENTRIES,
1245 				  sizeof(struct spdk_nvmf_rdma_request_data),
1246 				  SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
1247 				  SPDK_ENV_SOCKET_ID_ANY);
1248 	next_request_data = spdk_mempool_get(rtransport.data_wr_pool);
1249 	SPDK_CU_ASSERT_FATAL(((struct test_mempool *)rtransport.data_wr_pool)->count ==
1250 			     SPDK_NVMF_MAX_SGL_ENTRIES - 1);
1251 	next_request_data->wr.wr_id = (uint64_t)&rdma_req.data_wr;
1252 	next_request_data->wr.num_sge = 2;
1253 	next_request_data->wr.next = NULL;
1254 	rdma_req.data.wr.next = &next_request_data->wr;
1255 	rdma_req.data.wr.wr_id = (uint64_t)&rdma_req.data_wr;
1256 	rdma_req.data.wr.num_sge = 2;
1257 	rdma_req.transfer_wr = &rdma_req.data.wr;
1258 
1259 	nvmf_rdma_request_free_data(&rdma_req, &rtransport);
1260 	/* Check if next_request_data put into memory pool */
1261 	CU_ASSERT(((struct test_mempool *)rtransport.data_wr_pool)->count == SPDK_NVMF_MAX_SGL_ENTRIES);
1262 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
1263 
1264 	spdk_mempool_free(rtransport.data_wr_pool);
1265 }
1266 
1267 static void
1268 test_nvmf_rdma_update_ibv_state(void)
1269 {
1270 	struct spdk_nvmf_rdma_qpair rqpair = {};
1271 	struct spdk_rdma_qp rdma_qp = {};
1272 	struct ibv_qp qp = {};
1273 	int rc = 0;
1274 
1275 	rqpair.rdma_qp = &rdma_qp;
1276 
1277 	/* Case 1: Failed to get updated RDMA queue pair state */
1278 	rqpair.ibv_state = IBV_QPS_INIT;
1279 	rqpair.rdma_qp->qp = NULL;
1280 
1281 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1282 	CU_ASSERT(rc == IBV_QPS_ERR + 1);
1283 
1284 	/* Case 2: Bad state updated */
1285 	rqpair.rdma_qp->qp = &qp;
1286 	qp.state = IBV_QPS_ERR;
1287 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1288 	CU_ASSERT(rqpair.ibv_state == 10);
1289 	CU_ASSERT(rc == IBV_QPS_ERR + 1);
1290 
1291 	/* Case 3: Pass */
1292 	qp.state = IBV_QPS_INIT;
1293 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1294 	CU_ASSERT(rqpair.ibv_state == IBV_QPS_INIT);
1295 	CU_ASSERT(rc == IBV_QPS_INIT);
1296 }
1297 
1298 static void
1299 test_nvmf_rdma_resources_create(void)
1300 {
1301 	static struct spdk_nvmf_rdma_resources *rdma_resource;
1302 	struct spdk_nvmf_rdma_resource_opts opts = {};
1303 	struct spdk_nvmf_rdma_qpair qpair = {};
1304 	struct spdk_nvmf_rdma_recv *recv = NULL;
1305 	struct spdk_nvmf_rdma_request *req = NULL;
1306 	const int DEPTH = 128;
1307 
1308 	opts.max_queue_depth = DEPTH;
1309 	opts.in_capsule_data_size = 4096;
1310 	opts.shared = true;
1311 	opts.qpair = &qpair;
1312 
1313 	rdma_resource = nvmf_rdma_resources_create(&opts);
1314 	CU_ASSERT(rdma_resource != NULL);
1315 	/* Just check first and last entry */
1316 	recv = &rdma_resource->recvs[0];
1317 	req = &rdma_resource->reqs[0];
1318 	CU_ASSERT(recv->rdma_wr.type == RDMA_WR_TYPE_RECV);
1319 	CU_ASSERT((uintptr_t)recv->buf == (uintptr_t)(rdma_resource->bufs));
1320 	CU_ASSERT(recv->sgl[0].addr == (uintptr_t)&rdma_resource->cmds[0]);
1321 	CU_ASSERT(recv->sgl[0].length == sizeof(rdma_resource->cmds[0]));
1322 	CU_ASSERT(recv->sgl[0].lkey == RDMA_UT_LKEY);
1323 	CU_ASSERT(recv->wr.num_sge == 2);
1324 	CU_ASSERT(recv->wr.wr_id == (uintptr_t)&rdma_resource->recvs[0].rdma_wr);
1325 	CU_ASSERT(recv->wr.sg_list == rdma_resource->recvs[0].sgl);
1326 	CU_ASSERT(req->req.rsp == &rdma_resource->cpls[0]);
1327 	CU_ASSERT(req->rsp.sgl[0].addr == (uintptr_t)&rdma_resource->cpls[0]);
1328 	CU_ASSERT(req->rsp.sgl[0].length == sizeof(rdma_resource->cpls[0]));
1329 	CU_ASSERT(req->rsp.sgl[0].lkey == RDMA_UT_LKEY);
1330 	CU_ASSERT(req->rsp_wr.type == RDMA_WR_TYPE_SEND);
1331 	CU_ASSERT(req->rsp.wr.wr_id == (uintptr_t)&rdma_resource->reqs[0].rsp_wr);
1332 	CU_ASSERT(req->rsp.wr.next == NULL);
1333 	CU_ASSERT(req->rsp.wr.opcode == IBV_WR_SEND);
1334 	CU_ASSERT(req->rsp.wr.send_flags == IBV_SEND_SIGNALED);
1335 	CU_ASSERT(req->rsp.wr.sg_list == rdma_resource->reqs[0].rsp.sgl);
1336 	CU_ASSERT(req->rsp.wr.num_sge == NVMF_DEFAULT_RSP_SGE);
1337 	CU_ASSERT(req->data_wr.type == RDMA_WR_TYPE_DATA);
1338 	CU_ASSERT(req->data.wr.wr_id == (uintptr_t)&rdma_resource->reqs[0].data_wr);
1339 	CU_ASSERT(req->data.wr.next == NULL);
1340 	CU_ASSERT(req->data.wr.send_flags == IBV_SEND_SIGNALED);
1341 	CU_ASSERT(req->data.wr.sg_list == rdma_resource->reqs[0].data.sgl);
1342 	CU_ASSERT(req->data.wr.num_sge == SPDK_NVMF_MAX_SGL_ENTRIES);
1343 	CU_ASSERT(req->state == RDMA_REQUEST_STATE_FREE);
1344 
1345 	recv = &rdma_resource->recvs[DEPTH - 1];
1346 	req = &rdma_resource->reqs[DEPTH - 1];
1347 	CU_ASSERT(recv->rdma_wr.type == RDMA_WR_TYPE_RECV);
1348 	CU_ASSERT((uintptr_t)recv->buf == (uintptr_t)(rdma_resource->bufs +
1349 			(DEPTH - 1) * 4096));
1350 	CU_ASSERT(recv->sgl[0].addr == (uintptr_t)&rdma_resource->cmds[DEPTH - 1]);
1351 	CU_ASSERT(recv->sgl[0].length == sizeof(rdma_resource->cmds[DEPTH - 1]));
1352 	CU_ASSERT(recv->sgl[0].lkey == RDMA_UT_LKEY);
1353 	CU_ASSERT(recv->wr.num_sge == 2);
1354 	CU_ASSERT(recv->wr.wr_id == (uintptr_t)&rdma_resource->recvs[DEPTH - 1].rdma_wr);
1355 	CU_ASSERT(recv->wr.sg_list == rdma_resource->recvs[DEPTH - 1].sgl);
1356 	CU_ASSERT(req->req.rsp == &rdma_resource->cpls[DEPTH - 1]);
1357 	CU_ASSERT(req->rsp.sgl[0].addr == (uintptr_t)&rdma_resource->cpls[DEPTH - 1]);
1358 	CU_ASSERT(req->rsp.sgl[0].length == sizeof(rdma_resource->cpls[DEPTH - 1]));
1359 	CU_ASSERT(req->rsp.sgl[0].lkey == RDMA_UT_LKEY);
1360 	CU_ASSERT(req->rsp_wr.type == RDMA_WR_TYPE_SEND);
1361 	CU_ASSERT(req->rsp.wr.wr_id == (uintptr_t)&req->rsp_wr);
1362 	CU_ASSERT(req->rsp.wr.next == NULL);
1363 	CU_ASSERT(req->rsp.wr.opcode == IBV_WR_SEND);
1364 	CU_ASSERT(req->rsp.wr.send_flags == IBV_SEND_SIGNALED);
1365 	CU_ASSERT(req->rsp.wr.sg_list == rdma_resource->reqs[DEPTH - 1].rsp.sgl);
1366 	CU_ASSERT(req->rsp.wr.num_sge == NVMF_DEFAULT_RSP_SGE);
1367 	CU_ASSERT(req->data_wr.type == RDMA_WR_TYPE_DATA);
1368 	CU_ASSERT(req->data.wr.wr_id == (uintptr_t)&req->data_wr);
1369 	CU_ASSERT(req->data.wr.next == NULL);
1370 	CU_ASSERT(req->data.wr.send_flags == IBV_SEND_SIGNALED);
1371 	CU_ASSERT(req->data.wr.sg_list == rdma_resource->reqs[DEPTH - 1].data.sgl);
1372 	CU_ASSERT(req->data.wr.num_sge == SPDK_NVMF_MAX_SGL_ENTRIES);
1373 	CU_ASSERT(req->state == RDMA_REQUEST_STATE_FREE);
1374 
1375 	nvmf_rdma_resources_destroy(rdma_resource);
1376 }
1377 
1378 static void
1379 test_nvmf_rdma_qpair_compare(void)
1380 {
1381 	struct spdk_nvmf_rdma_qpair rqpair1 = {}, rqpair2 = {};
1382 
1383 	rqpair1.qp_num = 0;
1384 	rqpair2.qp_num = UINT32_MAX;
1385 
1386 	CU_ASSERT(nvmf_rdma_qpair_compare(&rqpair1, &rqpair2) < 0);
1387 	CU_ASSERT(nvmf_rdma_qpair_compare(&rqpair2, &rqpair1) > 0);
1388 }
1389 
1390 static void
1391 test_nvmf_rdma_resize_cq(void)
1392 {
1393 	int rc = -1;
1394 	int tnum_wr = 0;
1395 	int tnum_cqe = 0;
1396 	struct spdk_nvmf_rdma_qpair rqpair = {};
1397 	struct spdk_nvmf_rdma_poller rpoller = {};
1398 	struct spdk_nvmf_rdma_device rdevice = {};
1399 	struct ibv_context ircontext = {};
1400 	struct ibv_device idevice = {};
1401 
1402 	rdevice.context = &ircontext;
1403 	rqpair.poller = &rpoller;
1404 	ircontext.device = &idevice;
1405 
1406 	/* Test1: Current capacity support required size. */
1407 	rpoller.required_num_wr = 10;
1408 	rpoller.num_cqe = 20;
1409 	rqpair.max_queue_depth = 2;
1410 	tnum_wr = rpoller.required_num_wr;
1411 	tnum_cqe = rpoller.num_cqe;
1412 
1413 	rc = nvmf_rdma_resize_cq(&rqpair, &rdevice);
1414 	CU_ASSERT(rc == 0);
1415 	CU_ASSERT(rpoller.required_num_wr == 10 + MAX_WR_PER_QP(rqpair.max_queue_depth));
1416 	CU_ASSERT(rpoller.required_num_wr > tnum_wr);
1417 	CU_ASSERT(rpoller.num_cqe == tnum_cqe);
1418 
1419 	/* Test2: iWARP doesn't support CQ resize. */
1420 	tnum_wr = rpoller.required_num_wr;
1421 	tnum_cqe = rpoller.num_cqe;
1422 	idevice.transport_type = IBV_TRANSPORT_IWARP;
1423 
1424 	rc = nvmf_rdma_resize_cq(&rqpair, &rdevice);
1425 	CU_ASSERT(rc == -1);
1426 	CU_ASSERT(rpoller.required_num_wr == tnum_wr);
1427 	CU_ASSERT(rpoller.num_cqe == tnum_cqe);
1428 
1429 
1430 	/* Test3: RDMA CQE requirement exceeds device max_cqe limitation. */
1431 	tnum_wr = rpoller.required_num_wr;
1432 	tnum_cqe = rpoller.num_cqe;
1433 	idevice.transport_type = IBV_TRANSPORT_UNKNOWN;
1434 	rdevice.attr.max_cqe = 3;
1435 
1436 	rc = nvmf_rdma_resize_cq(&rqpair, &rdevice);
1437 	CU_ASSERT(rc == -1);
1438 	CU_ASSERT(rpoller.required_num_wr == tnum_wr);
1439 	CU_ASSERT(rpoller.num_cqe == tnum_cqe);
1440 
1441 	/* Test4: RDMA CQ resize failed. */
1442 	tnum_wr = rpoller.required_num_wr;
1443 	tnum_cqe = rpoller.num_cqe;
1444 	idevice.transport_type = IBV_TRANSPORT_IB;
1445 	rdevice.attr.max_cqe = 30;
1446 	MOCK_SET(ibv_resize_cq, -1);
1447 
1448 	rc = nvmf_rdma_resize_cq(&rqpair, &rdevice);
1449 	CU_ASSERT(rc == -1);
1450 	CU_ASSERT(rpoller.required_num_wr == tnum_wr);
1451 	CU_ASSERT(rpoller.num_cqe == tnum_cqe);
1452 
1453 	/* Test5: RDMA CQ resize success. rsize = MIN(MAX(num_cqe * 2, required_num_wr), device->attr.max_cqe). */
1454 	tnum_wr = rpoller.required_num_wr;
1455 	tnum_cqe = rpoller.num_cqe;
1456 	MOCK_SET(ibv_resize_cq, 0);
1457 
1458 	rc = nvmf_rdma_resize_cq(&rqpair, &rdevice);
1459 	CU_ASSERT(rc == 0);
1460 	CU_ASSERT(rpoller.num_cqe = 30);
1461 	CU_ASSERT(rpoller.required_num_wr == 18 + MAX_WR_PER_QP(rqpair.max_queue_depth));
1462 	CU_ASSERT(rpoller.required_num_wr > tnum_wr);
1463 	CU_ASSERT(rpoller.num_cqe > tnum_cqe);
1464 }
1465 
1466 int
1467 main(int argc, char **argv)
1468 {
1469 	CU_pSuite	suite = NULL;
1470 	unsigned int	num_failures;
1471 
1472 	CU_initialize_registry();
1473 
1474 	suite = CU_add_suite("nvmf", NULL, NULL);
1475 
1476 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl);
1477 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_process);
1478 	CU_ADD_TEST(suite, test_nvmf_rdma_get_optimal_poll_group);
1479 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl_with_md);
1480 	CU_ADD_TEST(suite, test_nvmf_rdma_opts_init);
1481 	CU_ADD_TEST(suite, test_nvmf_rdma_request_free_data);
1482 	CU_ADD_TEST(suite, test_nvmf_rdma_update_ibv_state);
1483 	CU_ADD_TEST(suite, test_nvmf_rdma_resources_create);
1484 	CU_ADD_TEST(suite, test_nvmf_rdma_qpair_compare);
1485 	CU_ADD_TEST(suite, test_nvmf_rdma_resize_cq);
1486 
1487 	num_failures = spdk_ut_run_tests(argc, argv, NULL);
1488 	CU_cleanup_registry();
1489 	return num_failures;
1490 }
1491