xref: /spdk/test/unit/lib/nvmf/rdma.c/rdma_ut.c (revision 9efad7468f30e1c5f7442823f5a8b17acd1e6a9b)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright (c) Intel Corporation. All rights reserved.
5  *   Copyright (c) 2019, 2021 Mellanox Technologies LTD. All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include "spdk/stdinc.h"
35 #include "spdk_cunit.h"
36 #include "common/lib/test_env.c"
37 #include "common/lib/test_rdma.c"
38 #include "nvmf/rdma.c"
39 #include "nvmf/transport.c"
40 
41 #define RDMA_UT_UNITS_IN_MAX_IO 16
42 
43 struct spdk_nvmf_transport_opts g_rdma_ut_transport_opts = {
44 	.max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH,
45 	.max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR,
46 	.in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE,
47 	.max_io_size = (SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE * RDMA_UT_UNITS_IN_MAX_IO),
48 	.io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE,
49 	.max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH,
50 	.num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS,
51 };
52 
53 SPDK_LOG_REGISTER_COMPONENT(nvmf)
54 DEFINE_STUB(spdk_mem_map_set_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
55 		uint64_t size, uint64_t translation), 0);
56 DEFINE_STUB(spdk_mem_map_clear_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
57 		uint64_t size), 0);
58 DEFINE_STUB(spdk_mem_map_alloc, struct spdk_mem_map *, (uint64_t default_translation,
59 		const struct spdk_mem_map_ops *ops, void *cb_ctx), NULL);
60 DEFINE_STUB(spdk_nvmf_qpair_disconnect, int, (struct spdk_nvmf_qpair *qpair,
61 		nvmf_qpair_disconnect_cb cb_fn, void *ctx), 0);
62 DEFINE_STUB(spdk_nvmf_qpair_get_listen_trid, int,
63 	    (struct spdk_nvmf_qpair *qpair, struct spdk_nvme_transport_id *trid), 0);
64 DEFINE_STUB_V(spdk_mem_map_free, (struct spdk_mem_map **pmap));
65 
66 DEFINE_STUB_V(spdk_nvmf_ctrlr_data_init, (struct spdk_nvmf_transport_opts *opts,
67 		struct spdk_nvmf_ctrlr_data *cdata));
68 DEFINE_STUB_V(spdk_nvmf_request_exec, (struct spdk_nvmf_request *req));
69 DEFINE_STUB(spdk_nvmf_request_complete, int, (struct spdk_nvmf_request *req), 0);
70 DEFINE_STUB(spdk_nvme_transport_id_compare, int, (const struct spdk_nvme_transport_id *trid1,
71 		const struct spdk_nvme_transport_id *trid2), 0);
72 DEFINE_STUB_V(nvmf_ctrlr_abort_aer, (struct spdk_nvmf_ctrlr *ctrlr));
73 DEFINE_STUB(spdk_nvmf_request_get_dif_ctx, bool, (struct spdk_nvmf_request *req,
74 		struct spdk_dif_ctx *dif_ctx), false);
75 DEFINE_STUB_V(spdk_nvme_trid_populate_transport, (struct spdk_nvme_transport_id *trid,
76 		enum spdk_nvme_transport_type trtype));
77 DEFINE_STUB_V(spdk_nvmf_tgt_new_qpair, (struct spdk_nvmf_tgt *tgt, struct spdk_nvmf_qpair *qpair));
78 DEFINE_STUB(nvmf_ctrlr_abort_request, int, (struct spdk_nvmf_request *req), 0);
79 DEFINE_STUB(spdk_nvme_transport_id_adrfam_str, const char *, (enum spdk_nvmf_adrfam adrfam), NULL);
80 
81 int ibv_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
82 		 int attr_mask, struct ibv_qp_init_attr *init_attr)
83 {
84 	if (qp == NULL) {
85 		return -1;
86 	} else {
87 		attr->port_num = 80;
88 
89 		if (qp->state == IBV_QPS_ERR) {
90 			attr->qp_state = 10;
91 		} else {
92 			attr->qp_state = IBV_QPS_INIT;
93 		}
94 
95 		return 0;
96 	}
97 }
98 
99 const char *
100 spdk_nvme_transport_id_trtype_str(enum spdk_nvme_transport_type trtype)
101 {
102 	switch (trtype) {
103 	case SPDK_NVME_TRANSPORT_PCIE:
104 		return "PCIe";
105 	case SPDK_NVME_TRANSPORT_RDMA:
106 		return "RDMA";
107 	case SPDK_NVME_TRANSPORT_FC:
108 		return "FC";
109 	default:
110 		return NULL;
111 	}
112 }
113 
114 int
115 spdk_nvme_transport_id_populate_trstring(struct spdk_nvme_transport_id *trid, const char *trstring)
116 {
117 	int len, i;
118 
119 	if (trstring == NULL) {
120 		return -EINVAL;
121 	}
122 
123 	len = strnlen(trstring, SPDK_NVMF_TRSTRING_MAX_LEN);
124 	if (len == SPDK_NVMF_TRSTRING_MAX_LEN) {
125 		return -EINVAL;
126 	}
127 
128 	/* cast official trstring to uppercase version of input. */
129 	for (i = 0; i < len; i++) {
130 		trid->trstring[i] = toupper(trstring[i]);
131 	}
132 	return 0;
133 }
134 
135 static void reset_nvmf_rdma_request(struct spdk_nvmf_rdma_request *rdma_req)
136 {
137 	int i;
138 
139 	rdma_req->req.length = 0;
140 	rdma_req->req.data_from_pool = false;
141 	rdma_req->req.data = NULL;
142 	rdma_req->data.wr.num_sge = 0;
143 	rdma_req->data.wr.wr.rdma.remote_addr = 0;
144 	rdma_req->data.wr.wr.rdma.rkey = 0;
145 	rdma_req->offset = 0;
146 	memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
147 
148 	for (i = 0; i < SPDK_NVMF_MAX_SGL_ENTRIES; i++) {
149 		rdma_req->req.iov[i].iov_base = 0;
150 		rdma_req->req.iov[i].iov_len = 0;
151 		rdma_req->req.buffers[i] = 0;
152 		rdma_req->data.wr.sg_list[i].addr = 0;
153 		rdma_req->data.wr.sg_list[i].length = 0;
154 		rdma_req->data.wr.sg_list[i].lkey = 0;
155 	}
156 	rdma_req->req.iovcnt = 0;
157 }
158 
159 static void
160 test_spdk_nvmf_rdma_request_parse_sgl(void)
161 {
162 	struct spdk_nvmf_rdma_transport rtransport;
163 	struct spdk_nvmf_rdma_device device;
164 	struct spdk_nvmf_rdma_request rdma_req = {};
165 	struct spdk_nvmf_rdma_recv recv;
166 	struct spdk_nvmf_rdma_poll_group group;
167 	struct spdk_nvmf_rdma_qpair rqpair;
168 	struct spdk_nvmf_rdma_poller poller;
169 	union nvmf_c2h_msg cpl;
170 	union nvmf_h2c_msg cmd;
171 	struct spdk_nvme_sgl_descriptor *sgl;
172 	struct spdk_nvmf_transport_pg_cache_buf bufs[4];
173 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
174 	struct spdk_nvmf_rdma_request_data data;
175 	int rc, i;
176 	uint32_t sgl_length;
177 	uintptr_t aligned_buffer_address;
178 
179 	data.wr.sg_list = data.sgl;
180 	STAILQ_INIT(&group.group.buf_cache);
181 	group.group.buf_cache_size = 0;
182 	group.group.buf_cache_count = 0;
183 	group.group.transport = &rtransport.transport;
184 	poller.group = &group;
185 	rqpair.poller = &poller;
186 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
187 
188 	sgl = &cmd.nvme_cmd.dptr.sgl1;
189 	rdma_req.recv = &recv;
190 	rdma_req.req.cmd = &cmd;
191 	rdma_req.req.rsp = &cpl;
192 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
193 	rdma_req.req.qpair = &rqpair.qpair;
194 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
195 
196 	rtransport.transport.opts = g_rdma_ut_transport_opts;
197 	rtransport.data_wr_pool = NULL;
198 	rtransport.transport.data_buf_pool = NULL;
199 
200 	device.attr.device_cap_flags = 0;
201 	sgl->keyed.key = 0xEEEE;
202 	sgl->address = 0xFFFF;
203 	rdma_req.recv->buf = (void *)0xDDDD;
204 
205 	/* Test 1: sgl type: keyed data block subtype: address */
206 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
207 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
208 
209 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size */
210 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
211 	reset_nvmf_rdma_request(&rdma_req);
212 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
213 
214 	device.map = (void *)0x0;
215 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
216 	CU_ASSERT(rc == 0);
217 	CU_ASSERT(rdma_req.req.data_from_pool == true);
218 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
219 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
220 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
221 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
222 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
223 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
224 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
225 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
226 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
227 
228 	/* Part 2: simple I/O, one SGL larger than the transport io unit size (equal to the max io size) */
229 	reset_nvmf_rdma_request(&rdma_req);
230 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
231 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
232 
233 	CU_ASSERT(rc == 0);
234 	CU_ASSERT(rdma_req.req.data_from_pool == true);
235 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO);
236 	CU_ASSERT(rdma_req.data.wr.num_sge == RDMA_UT_UNITS_IN_MAX_IO);
237 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
238 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
239 	for (i = 0; i < RDMA_UT_UNITS_IN_MAX_IO; i++) {
240 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
241 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
242 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
243 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
244 	}
245 
246 	/* Part 3: simple I/O one SGL larger than the transport max io size */
247 	reset_nvmf_rdma_request(&rdma_req);
248 	sgl->keyed.length = rtransport.transport.opts.max_io_size * 2;
249 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
250 
251 	CU_ASSERT(rc == -1);
252 
253 	/* Part 4: Pretend there are no buffer pools */
254 	MOCK_SET(spdk_mempool_get, NULL);
255 	reset_nvmf_rdma_request(&rdma_req);
256 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
257 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
258 
259 	CU_ASSERT(rc == 0);
260 	CU_ASSERT(rdma_req.req.data_from_pool == false);
261 	CU_ASSERT(rdma_req.req.data == NULL);
262 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
263 	CU_ASSERT(rdma_req.req.buffers[0] == NULL);
264 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0);
265 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 0);
266 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == 0);
267 
268 	rdma_req.recv->buf = (void *)0xDDDD;
269 	/* Test 2: sgl type: keyed data block subtype: offset (in capsule data) */
270 	sgl->generic.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
271 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
272 
273 	/* Part 1: Normal I/O smaller than in capsule data size no offset */
274 	reset_nvmf_rdma_request(&rdma_req);
275 	sgl->address = 0;
276 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
277 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
278 
279 	CU_ASSERT(rc == 0);
280 	CU_ASSERT(rdma_req.req.data == (void *)0xDDDD);
281 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.in_capsule_data_size);
282 	CU_ASSERT(rdma_req.req.data_from_pool == false);
283 
284 	/* Part 2: I/O offset + length too large */
285 	reset_nvmf_rdma_request(&rdma_req);
286 	sgl->address = rtransport.transport.opts.in_capsule_data_size;
287 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
288 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
289 
290 	CU_ASSERT(rc == -1);
291 
292 	/* Part 3: I/O too large */
293 	reset_nvmf_rdma_request(&rdma_req);
294 	sgl->address = 0;
295 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size * 2;
296 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
297 
298 	CU_ASSERT(rc == -1);
299 
300 	/* Test 3: Multi SGL */
301 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
302 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
303 	sgl->address = 0;
304 	rdma_req.recv->buf = (void *)&sgl_desc;
305 	MOCK_SET(spdk_mempool_get, &data);
306 
307 	/* part 1: 2 segments each with 1 wr. */
308 	reset_nvmf_rdma_request(&rdma_req);
309 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
310 	for (i = 0; i < 2; i++) {
311 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
312 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
313 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size;
314 		sgl_desc[i].address = 0x4000 + i * rtransport.transport.opts.io_unit_size;
315 		sgl_desc[i].keyed.key = 0x44;
316 	}
317 
318 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
319 
320 	CU_ASSERT(rc == 0);
321 	CU_ASSERT(rdma_req.req.data_from_pool == true);
322 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 2);
323 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
324 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
325 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
326 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
327 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
328 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size);
329 	CU_ASSERT(data.wr.num_sge == 1);
330 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
331 
332 	/* part 2: 2 segments, each with 1 wr containing 8 sge_elements */
333 	reset_nvmf_rdma_request(&rdma_req);
334 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
335 	for (i = 0; i < 2; i++) {
336 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
337 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
338 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size * 8;
339 		sgl_desc[i].address = 0x4000 + i * 8 * rtransport.transport.opts.io_unit_size;
340 		sgl_desc[i].keyed.key = 0x44;
341 	}
342 
343 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
344 
345 	CU_ASSERT(rc == 0);
346 	CU_ASSERT(rdma_req.req.data_from_pool == true);
347 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
348 	CU_ASSERT(rdma_req.req.iovcnt == 16);
349 	CU_ASSERT(rdma_req.data.wr.num_sge == 8);
350 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
351 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
352 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
353 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
354 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 8);
355 	CU_ASSERT(data.wr.num_sge == 8);
356 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
357 
358 	/* part 3: 2 segments, one very large, one very small */
359 	reset_nvmf_rdma_request(&rdma_req);
360 	for (i = 0; i < 2; i++) {
361 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
362 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
363 		sgl_desc[i].keyed.key = 0x44;
364 	}
365 
366 	sgl_desc[0].keyed.length = rtransport.transport.opts.io_unit_size * 15 +
367 				   rtransport.transport.opts.io_unit_size / 2;
368 	sgl_desc[0].address = 0x4000;
369 	sgl_desc[1].keyed.length = rtransport.transport.opts.io_unit_size / 2;
370 	sgl_desc[1].address = 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
371 			      rtransport.transport.opts.io_unit_size / 2;
372 
373 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
374 
375 	CU_ASSERT(rc == 0);
376 	CU_ASSERT(rdma_req.req.data_from_pool == true);
377 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
378 	CU_ASSERT(rdma_req.req.iovcnt == 16);
379 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
380 	for (i = 0; i < 15; i++) {
381 		CU_ASSERT(rdma_req.data.sgl[i].length == rtransport.transport.opts.io_unit_size);
382 	}
383 	CU_ASSERT(rdma_req.data.sgl[15].length == rtransport.transport.opts.io_unit_size / 2);
384 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
385 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
386 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
387 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
388 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
389 		  rtransport.transport.opts.io_unit_size / 2);
390 	CU_ASSERT(data.sgl[0].length == rtransport.transport.opts.io_unit_size / 2);
391 	CU_ASSERT(data.wr.num_sge == 1);
392 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
393 
394 	/* part 4: 2 SGL descriptors, each length is transport buffer / 2
395 	 * 1 transport buffers should be allocated */
396 	reset_nvmf_rdma_request(&rdma_req);
397 	aligned_buffer_address = ((uintptr_t)(&data) + NVMF_DATA_BUFFER_MASK) & ~NVMF_DATA_BUFFER_MASK;
398 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
399 	sgl_length = rtransport.transport.opts.io_unit_size / 2;
400 	for (i = 0; i < 2; i++) {
401 		sgl_desc[i].keyed.length = sgl_length;
402 		sgl_desc[i].address = 0x4000 + i * sgl_length;
403 	}
404 
405 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
406 
407 	CU_ASSERT(rc == 0);
408 	CU_ASSERT(rdma_req.req.data_from_pool == true);
409 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size);
410 	CU_ASSERT(rdma_req.req.iovcnt == 1);
411 
412 	CU_ASSERT(rdma_req.data.sgl[0].length == sgl_length);
413 	/* We mocked mempool_get to return address of data variable. Mempool is used
414 	 * to get both additional WRs and data buffers, so data points to &data */
415 	CU_ASSERT(rdma_req.data.sgl[0].addr == aligned_buffer_address);
416 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
417 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
418 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
419 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
420 
421 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
422 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + sgl_length);
423 	CU_ASSERT(data.sgl[0].length == sgl_length);
424 	CU_ASSERT(data.sgl[0].addr == aligned_buffer_address + sgl_length);
425 	CU_ASSERT(data.wr.num_sge == 1);
426 
427 	/* Test 4: use PG buffer cache */
428 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
429 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
430 	sgl->address = 0xFFFF;
431 	rdma_req.recv->buf = (void *)0xDDDD;
432 	sgl->keyed.key = 0xEEEE;
433 
434 	for (i = 0; i < 4; i++) {
435 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
436 	}
437 
438 	/* part 1: use the four buffers from the pg cache */
439 	group.group.buf_cache_size = 4;
440 	group.group.buf_cache_count = 4;
441 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
442 	reset_nvmf_rdma_request(&rdma_req);
443 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * 4;
444 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
445 
446 	SPDK_CU_ASSERT_FATAL(rc == 0);
447 	CU_ASSERT(rdma_req.req.data_from_pool == true);
448 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
449 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
450 			~NVMF_DATA_BUFFER_MASK));
451 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
452 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
453 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
454 	CU_ASSERT(group.group.buf_cache_count == 0);
455 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
456 	for (i = 0; i < 4; i++) {
457 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
458 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
459 				~NVMF_DATA_BUFFER_MASK));
460 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
461 	}
462 
463 	/* part 2: now that we have used the buffers from the cache, try again. We should get mempool buffers. */
464 	reset_nvmf_rdma_request(&rdma_req);
465 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
466 
467 	SPDK_CU_ASSERT_FATAL(rc == 0);
468 	CU_ASSERT(rdma_req.req.data_from_pool == true);
469 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
470 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
471 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
472 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
473 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
474 	CU_ASSERT(group.group.buf_cache_count == 0);
475 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
476 	for (i = 0; i < 4; i++) {
477 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
478 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
479 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
480 		CU_ASSERT(group.group.buf_cache_count == 0);
481 	}
482 
483 	/* part 3: half and half */
484 	group.group.buf_cache_count = 2;
485 
486 	for (i = 0; i < 2; i++) {
487 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
488 	}
489 	reset_nvmf_rdma_request(&rdma_req);
490 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
491 
492 	SPDK_CU_ASSERT_FATAL(rc == 0);
493 	CU_ASSERT(rdma_req.req.data_from_pool == true);
494 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
495 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
496 			~NVMF_DATA_BUFFER_MASK));
497 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
498 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
499 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
500 	CU_ASSERT(group.group.buf_cache_count == 0);
501 	for (i = 0; i < 2; i++) {
502 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
503 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
504 				~NVMF_DATA_BUFFER_MASK));
505 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
506 	}
507 	for (i = 2; i < 4; i++) {
508 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
509 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
510 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
511 	}
512 
513 	reset_nvmf_rdma_request(&rdma_req);
514 }
515 
516 static struct spdk_nvmf_rdma_recv *
517 create_recv(struct spdk_nvmf_rdma_qpair *rqpair, enum spdk_nvme_nvm_opcode opc)
518 {
519 	struct spdk_nvmf_rdma_recv *rdma_recv;
520 	union nvmf_h2c_msg *cmd;
521 	struct spdk_nvme_sgl_descriptor *sgl;
522 
523 	rdma_recv = calloc(1, sizeof(*rdma_recv));
524 	rdma_recv->qpair = rqpair;
525 	cmd = calloc(1, sizeof(*cmd));
526 	rdma_recv->sgl[0].addr = (uintptr_t)cmd;
527 	cmd->nvme_cmd.opc = opc;
528 	sgl = &cmd->nvme_cmd.dptr.sgl1;
529 	sgl->keyed.key = 0xEEEE;
530 	sgl->address = 0xFFFF;
531 	sgl->keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
532 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
533 	sgl->keyed.length = 1;
534 
535 	return rdma_recv;
536 }
537 
538 static void
539 free_recv(struct spdk_nvmf_rdma_recv *rdma_recv)
540 {
541 	free((void *)rdma_recv->sgl[0].addr);
542 	free(rdma_recv);
543 }
544 
545 static struct spdk_nvmf_rdma_request *
546 create_req(struct spdk_nvmf_rdma_qpair *rqpair,
547 	   struct spdk_nvmf_rdma_recv *rdma_recv)
548 {
549 	struct spdk_nvmf_rdma_request *rdma_req;
550 	union nvmf_c2h_msg *cpl;
551 
552 	rdma_req = calloc(1, sizeof(*rdma_req));
553 	rdma_req->recv = rdma_recv;
554 	rdma_req->req.qpair = &rqpair->qpair;
555 	rdma_req->state = RDMA_REQUEST_STATE_NEW;
556 	rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr;
557 	rdma_req->data.wr.sg_list = rdma_req->data.sgl;
558 	cpl = calloc(1, sizeof(*cpl));
559 	rdma_req->rsp.sgl[0].addr = (uintptr_t)cpl;
560 	rdma_req->req.rsp = cpl;
561 
562 	return rdma_req;
563 }
564 
565 static void
566 free_req(struct spdk_nvmf_rdma_request *rdma_req)
567 {
568 	free((void *)rdma_req->rsp.sgl[0].addr);
569 	free(rdma_req);
570 }
571 
572 static void
573 qpair_reset(struct spdk_nvmf_rdma_qpair *rqpair,
574 	    struct spdk_nvmf_rdma_poller *poller,
575 	    struct spdk_nvmf_rdma_device *device,
576 	    struct spdk_nvmf_rdma_resources *resources,
577 	    struct spdk_nvmf_transport *transport)
578 {
579 	memset(rqpair, 0, sizeof(*rqpair));
580 	STAILQ_INIT(&rqpair->pending_rdma_write_queue);
581 	STAILQ_INIT(&rqpair->pending_rdma_read_queue);
582 	rqpair->poller = poller;
583 	rqpair->device = device;
584 	rqpair->resources = resources;
585 	rqpair->qpair.qid = 1;
586 	rqpair->ibv_state = IBV_QPS_RTS;
587 	rqpair->qpair.state = SPDK_NVMF_QPAIR_ACTIVE;
588 	rqpair->max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
589 	rqpair->max_send_depth = 16;
590 	rqpair->max_read_depth = 16;
591 	rqpair->qpair.transport = transport;
592 }
593 
594 static void
595 poller_reset(struct spdk_nvmf_rdma_poller *poller,
596 	     struct spdk_nvmf_rdma_poll_group *group)
597 {
598 	memset(poller, 0, sizeof(*poller));
599 	STAILQ_INIT(&poller->qpairs_pending_recv);
600 	STAILQ_INIT(&poller->qpairs_pending_send);
601 	poller->group = group;
602 }
603 
604 static void
605 test_spdk_nvmf_rdma_request_process(void)
606 {
607 	struct spdk_nvmf_rdma_transport rtransport = {};
608 	struct spdk_nvmf_rdma_poll_group group = {};
609 	struct spdk_nvmf_rdma_poller poller = {};
610 	struct spdk_nvmf_rdma_device device = {};
611 	struct spdk_nvmf_rdma_resources resources = {};
612 	struct spdk_nvmf_rdma_qpair rqpair = {};
613 	struct spdk_nvmf_rdma_recv *rdma_recv;
614 	struct spdk_nvmf_rdma_request *rdma_req;
615 	bool progress;
616 
617 	STAILQ_INIT(&group.group.buf_cache);
618 	STAILQ_INIT(&group.group.pending_buf_queue);
619 	group.group.buf_cache_size = 0;
620 	group.group.buf_cache_count = 0;
621 	poller_reset(&poller, &group);
622 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
623 
624 	rtransport.transport.opts = g_rdma_ut_transport_opts;
625 	rtransport.transport.data_buf_pool = spdk_mempool_create("test_data_pool", 16, 128, 0, 0);
626 	rtransport.data_wr_pool = spdk_mempool_create("test_wr_pool", 128,
627 				  sizeof(struct spdk_nvmf_rdma_request_data),
628 				  0, 0);
629 	MOCK_CLEAR(spdk_mempool_get);
630 
631 	device.attr.device_cap_flags = 0;
632 	device.map = (void *)0x0;
633 
634 	/* Test 1: single SGL READ request */
635 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_READ);
636 	rdma_req = create_req(&rqpair, rdma_recv);
637 	rqpair.current_recv_depth = 1;
638 	/* NEW -> EXECUTING */
639 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
640 	CU_ASSERT(progress == true);
641 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
642 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST);
643 	/* EXECUTED -> TRANSFERRING_C2H */
644 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
645 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
646 	CU_ASSERT(progress == true);
647 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST);
648 	CU_ASSERT(rdma_req->recv == NULL);
649 	/* COMPLETED -> FREE */
650 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
651 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
652 	CU_ASSERT(progress == true);
653 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
654 
655 	free_recv(rdma_recv);
656 	free_req(rdma_req);
657 	poller_reset(&poller, &group);
658 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
659 
660 	/* Test 2: single SGL WRITE request */
661 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
662 	rdma_req = create_req(&rqpair, rdma_recv);
663 	rqpair.current_recv_depth = 1;
664 	/* NEW -> TRANSFERRING_H2C */
665 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
666 	CU_ASSERT(progress == true);
667 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
668 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
669 	STAILQ_INIT(&poller.qpairs_pending_send);
670 	/* READY_TO_EXECUTE -> EXECUTING */
671 	rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
672 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
673 	CU_ASSERT(progress == true);
674 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
675 	/* EXECUTED -> COMPLETING */
676 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
677 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
678 	CU_ASSERT(progress == true);
679 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
680 	CU_ASSERT(rdma_req->recv == NULL);
681 	/* COMPLETED -> FREE */
682 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
683 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
684 	CU_ASSERT(progress == true);
685 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
686 
687 	free_recv(rdma_recv);
688 	free_req(rdma_req);
689 	poller_reset(&poller, &group);
690 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
691 
692 	/* Test 3: WRITE+WRITE ibv_send batching */
693 	{
694 		struct spdk_nvmf_rdma_recv *recv1, *recv2;
695 		struct spdk_nvmf_rdma_request *req1, *req2;
696 		recv1 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
697 		req1 = create_req(&rqpair, recv1);
698 		recv2 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
699 		req2 = create_req(&rqpair, recv2);
700 
701 		/* WRITE 1: NEW -> TRANSFERRING_H2C */
702 		rqpair.current_recv_depth = 1;
703 		nvmf_rdma_request_process(&rtransport, req1);
704 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
705 
706 		/* WRITE 2: NEW -> TRANSFERRING_H2C */
707 		rqpair.current_recv_depth = 2;
708 		nvmf_rdma_request_process(&rtransport, req2);
709 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
710 
711 		STAILQ_INIT(&poller.qpairs_pending_send);
712 
713 		/* WRITE 1 completes before WRITE 2 has finished RDMA reading */
714 		/* WRITE 1: READY_TO_EXECUTE -> EXECUTING */
715 		req1->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
716 		nvmf_rdma_request_process(&rtransport, req1);
717 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_EXECUTING);
718 		/* WRITE 1: EXECUTED -> COMPLETING */
719 		req1->state = RDMA_REQUEST_STATE_EXECUTED;
720 		nvmf_rdma_request_process(&rtransport, req1);
721 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_COMPLETING);
722 		STAILQ_INIT(&poller.qpairs_pending_send);
723 		/* WRITE 1: COMPLETED -> FREE */
724 		req1->state = RDMA_REQUEST_STATE_COMPLETED;
725 		nvmf_rdma_request_process(&rtransport, req1);
726 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_FREE);
727 
728 		/* Now WRITE 2 has finished reading and completes */
729 		/* WRITE 2: COMPLETED -> FREE */
730 		/* WRITE 2: READY_TO_EXECUTE -> EXECUTING */
731 		req2->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
732 		nvmf_rdma_request_process(&rtransport, req2);
733 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_EXECUTING);
734 		/* WRITE 1: EXECUTED -> COMPLETING */
735 		req2->state = RDMA_REQUEST_STATE_EXECUTED;
736 		nvmf_rdma_request_process(&rtransport, req2);
737 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_COMPLETING);
738 		STAILQ_INIT(&poller.qpairs_pending_send);
739 		/* WRITE 1: COMPLETED -> FREE */
740 		req2->state = RDMA_REQUEST_STATE_COMPLETED;
741 		nvmf_rdma_request_process(&rtransport, req2);
742 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_FREE);
743 
744 		free_recv(recv1);
745 		free_req(req1);
746 		free_recv(recv2);
747 		free_req(req2);
748 		poller_reset(&poller, &group);
749 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
750 	}
751 
752 	/* Test 4, invalid command, check xfer type */
753 	{
754 		struct spdk_nvmf_rdma_recv *rdma_recv_inv;
755 		struct spdk_nvmf_rdma_request *rdma_req_inv;
756 		/* construct an opcode that specifies BIDIRECTIONAL transfer */
757 		uint8_t opc = 0x10 | SPDK_NVME_DATA_BIDIRECTIONAL;
758 
759 		rdma_recv_inv = create_recv(&rqpair, opc);
760 		rdma_req_inv = create_req(&rqpair, rdma_recv_inv);
761 
762 		/* NEW -> RDMA_REQUEST_STATE_COMPLETING */
763 		rqpair.current_recv_depth = 1;
764 		progress = nvmf_rdma_request_process(&rtransport, rdma_req_inv);
765 		CU_ASSERT(progress == true);
766 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_COMPLETING);
767 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sct == SPDK_NVME_SCT_GENERIC);
768 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sc == SPDK_NVME_SC_INVALID_OPCODE);
769 
770 		/* RDMA_REQUEST_STATE_COMPLETED -> FREE */
771 		rdma_req_inv->state = RDMA_REQUEST_STATE_COMPLETED;
772 		nvmf_rdma_request_process(&rtransport, rdma_req_inv);
773 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_FREE);
774 
775 		free_recv(rdma_recv_inv);
776 		free_req(rdma_req_inv);
777 		poller_reset(&poller, &group);
778 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
779 	}
780 
781 	spdk_mempool_free(rtransport.transport.data_buf_pool);
782 	spdk_mempool_free(rtransport.data_wr_pool);
783 }
784 
785 #define TEST_GROUPS_COUNT 5
786 static void
787 test_nvmf_rdma_get_optimal_poll_group(void)
788 {
789 	struct spdk_nvmf_rdma_transport rtransport = {};
790 	struct spdk_nvmf_transport *transport = &rtransport.transport;
791 	struct spdk_nvmf_rdma_qpair rqpair = {};
792 	struct spdk_nvmf_transport_poll_group *groups[TEST_GROUPS_COUNT];
793 	struct spdk_nvmf_rdma_poll_group *rgroups[TEST_GROUPS_COUNT];
794 	struct spdk_nvmf_transport_poll_group *result;
795 	uint32_t i;
796 
797 	rqpair.qpair.transport = transport;
798 	pthread_mutex_init(&rtransport.lock, NULL);
799 	TAILQ_INIT(&rtransport.poll_groups);
800 
801 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
802 		groups[i] = nvmf_rdma_poll_group_create(transport);
803 		CU_ASSERT(groups[i] != NULL);
804 		rgroups[i] = SPDK_CONTAINEROF(groups[i], struct spdk_nvmf_rdma_poll_group, group);
805 		groups[i]->transport = transport;
806 	}
807 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[0]);
808 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[0]);
809 
810 	/* Emulate connection of %TEST_GROUPS_COUNT% initiators - each creates 1 admin and 1 io qp */
811 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
812 		rqpair.qpair.qid = 0;
813 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
814 		CU_ASSERT(result == groups[i]);
815 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
816 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i]);
817 
818 		rqpair.qpair.qid = 1;
819 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
820 		CU_ASSERT(result == groups[i]);
821 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
822 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
823 	}
824 	/* wrap around, admin/io pg point to the first pg
825 	   Destroy all poll groups except of the last one */
826 	for (i = 0; i < TEST_GROUPS_COUNT - 1; i++) {
827 		nvmf_rdma_poll_group_destroy(groups[i]);
828 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[i + 1]);
829 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i + 1]);
830 	}
831 
832 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
833 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
834 
835 	/* Check that pointers to the next admin/io poll groups are not changed */
836 	rqpair.qpair.qid = 0;
837 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
838 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
839 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
840 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
841 
842 	rqpair.qpair.qid = 1;
843 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
844 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
845 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
846 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
847 
848 	/* Remove the last poll group, check that pointers are NULL */
849 	nvmf_rdma_poll_group_destroy(groups[TEST_GROUPS_COUNT - 1]);
850 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == NULL);
851 	CU_ASSERT(rtransport.conn_sched.next_io_pg == NULL);
852 
853 	/* Request optimal poll group, result must be NULL */
854 	rqpair.qpair.qid = 0;
855 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
856 	CU_ASSERT(result == NULL);
857 
858 	rqpair.qpair.qid = 1;
859 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
860 	CU_ASSERT(result == NULL);
861 
862 	pthread_mutex_destroy(&rtransport.lock);
863 }
864 #undef TEST_GROUPS_COUNT
865 
866 static void
867 test_spdk_nvmf_rdma_request_parse_sgl_with_md(void)
868 {
869 	struct spdk_nvmf_rdma_transport rtransport;
870 	struct spdk_nvmf_rdma_device device;
871 	struct spdk_nvmf_rdma_request rdma_req = {};
872 	struct spdk_nvmf_rdma_recv recv;
873 	struct spdk_nvmf_rdma_poll_group group;
874 	struct spdk_nvmf_rdma_qpair rqpair;
875 	struct spdk_nvmf_rdma_poller poller;
876 	union nvmf_c2h_msg cpl;
877 	union nvmf_h2c_msg cmd;
878 	struct spdk_nvme_sgl_descriptor *sgl;
879 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
880 	char data_buffer[8192];
881 	struct spdk_nvmf_rdma_request_data *data = (struct spdk_nvmf_rdma_request_data *)data_buffer;
882 	char data2_buffer[8192];
883 	struct spdk_nvmf_rdma_request_data *data2 = (struct spdk_nvmf_rdma_request_data *)data2_buffer;
884 	const uint32_t data_bs = 512;
885 	const uint32_t md_size = 8;
886 	int rc, i;
887 	void *aligned_buffer;
888 
889 	data->wr.sg_list = data->sgl;
890 	STAILQ_INIT(&group.group.buf_cache);
891 	group.group.buf_cache_size = 0;
892 	group.group.buf_cache_count = 0;
893 	group.group.transport = &rtransport.transport;
894 	poller.group = &group;
895 	rqpair.poller = &poller;
896 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
897 
898 	sgl = &cmd.nvme_cmd.dptr.sgl1;
899 	rdma_req.recv = &recv;
900 	rdma_req.req.cmd = &cmd;
901 	rdma_req.req.rsp = &cpl;
902 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
903 	rdma_req.req.qpair = &rqpair.qpair;
904 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
905 
906 	rtransport.transport.opts = g_rdma_ut_transport_opts;
907 	rtransport.data_wr_pool = NULL;
908 	rtransport.transport.data_buf_pool = NULL;
909 
910 	device.attr.device_cap_flags = 0;
911 	device.map = NULL;
912 	sgl->keyed.key = 0xEEEE;
913 	sgl->address = 0xFFFF;
914 	rdma_req.recv->buf = (void *)0xDDDD;
915 
916 	/* Test 1: sgl type: keyed data block subtype: address */
917 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
918 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
919 
920 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size, block size 512 */
921 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
922 	reset_nvmf_rdma_request(&rdma_req);
923 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
924 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
925 			  0, 0, 0, 0, 0);
926 	rdma_req.req.dif_enabled = true;
927 	rtransport.transport.opts.io_unit_size = data_bs * 8;
928 	sgl->keyed.length = data_bs * 4;
929 
930 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
931 
932 	CU_ASSERT(rc == 0);
933 	CU_ASSERT(rdma_req.req.data_from_pool == true);
934 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
935 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
936 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
937 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
938 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
939 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
940 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
941 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
942 
943 	for (i = 0; i < 4; ++i) {
944 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
945 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
946 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
947 	}
948 
949 	/* Part 2: simple I/O, one SGL equal to io unit size, io_unit_size is not aligned with md_size,
950 		block size 512 */
951 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
952 	reset_nvmf_rdma_request(&rdma_req);
953 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
954 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
955 			  0, 0, 0, 0, 0);
956 	rdma_req.req.dif_enabled = true;
957 	rtransport.transport.opts.io_unit_size = data_bs * 4;
958 	sgl->keyed.length = data_bs * 4;
959 
960 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
961 
962 	CU_ASSERT(rc == 0);
963 	CU_ASSERT(rdma_req.req.data_from_pool == true);
964 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
965 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
966 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
967 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
968 	CU_ASSERT(rdma_req.data.wr.num_sge == 5);
969 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
970 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
971 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
972 
973 	for (i = 0; i < 3; ++i) {
974 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
975 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
976 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
977 	}
978 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
979 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
980 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
981 
982 	/* 2nd buffer consumed */
983 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
984 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
985 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
986 
987 	/* Part 3: simple I/O, one SGL equal io unit size, io_unit_size is equal to block size 512 bytes */
988 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
989 	reset_nvmf_rdma_request(&rdma_req);
990 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
991 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
992 			  0, 0, 0, 0, 0);
993 	rdma_req.req.dif_enabled = true;
994 	rtransport.transport.opts.io_unit_size = data_bs;
995 	sgl->keyed.length = data_bs;
996 
997 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
998 
999 	CU_ASSERT(rc == 0);
1000 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1001 	CU_ASSERT(rdma_req.req.length == data_bs);
1002 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1003 	CU_ASSERT(rdma_req.req.dif.elba_length == data_bs + md_size);
1004 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1005 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1006 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1007 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1008 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1009 
1010 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1011 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs);
1012 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1013 
1014 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1015 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)((unsigned long)0x2000));
1016 	CU_ASSERT(rdma_req.req.iov[0].iov_len == data_bs);
1017 	/* 2nd buffer consumed for metadata */
1018 	CU_ASSERT(rdma_req.req.iov[1].iov_base == (void *)((unsigned long)0x2000));
1019 	CU_ASSERT(rdma_req.req.iov[1].iov_len == md_size);
1020 
1021 	/* Part 4: simple I/O, one SGL equal io unit size, io_unit_size is aligned with md_size,
1022 	   block size 512 */
1023 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1024 	reset_nvmf_rdma_request(&rdma_req);
1025 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1026 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1027 			  0, 0, 0, 0, 0);
1028 	rdma_req.req.dif_enabled = true;
1029 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1030 	sgl->keyed.length = data_bs * 4;
1031 
1032 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1033 
1034 	CU_ASSERT(rc == 0);
1035 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1036 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1037 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1038 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1039 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1040 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1041 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1042 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1043 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1044 
1045 	for (i = 0; i < 4; ++i) {
1046 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1047 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1048 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1049 	}
1050 
1051 	/* Part 5: simple I/O, one SGL equal to 2x io unit size, io_unit_size is aligned with md_size,
1052 	   block size 512 */
1053 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1054 	reset_nvmf_rdma_request(&rdma_req);
1055 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1056 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1057 			  0, 0, 0, 0, 0);
1058 	rdma_req.req.dif_enabled = true;
1059 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 2;
1060 	sgl->keyed.length = data_bs * 4;
1061 
1062 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1063 
1064 	CU_ASSERT(rc == 0);
1065 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1066 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1067 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1068 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1069 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1070 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1071 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1072 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1073 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1074 
1075 	for (i = 0; i < 2; ++i) {
1076 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1077 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1078 	}
1079 	for (i = 0; i < 2; ++i) {
1080 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].addr == 0x2000 + i * (data_bs + md_size));
1081 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].length == data_bs);
1082 	}
1083 
1084 	/* Part 6: simple I/O, one SGL larger than the transport io unit size, io_unit_size is not aligned to md_size,
1085 	   block size 512 */
1086 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1087 	reset_nvmf_rdma_request(&rdma_req);
1088 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1089 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1090 			  0, 0, 0, 0, 0);
1091 	rdma_req.req.dif_enabled = true;
1092 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1093 	sgl->keyed.length = data_bs * 6;
1094 
1095 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1096 
1097 	CU_ASSERT(rc == 0);
1098 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1099 	CU_ASSERT(rdma_req.req.length == data_bs * 6);
1100 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1101 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 6);
1102 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1103 	CU_ASSERT(rdma_req.data.wr.num_sge == 7);
1104 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1105 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1106 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1107 
1108 	for (i = 0; i < 3; ++i) {
1109 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1110 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1111 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1112 	}
1113 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
1114 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
1115 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
1116 
1117 	/* 2nd IO buffer consumed */
1118 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
1119 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
1120 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
1121 
1122 	CU_ASSERT(rdma_req.data.wr.sg_list[5].addr == 0x2000 + 24 + md_size);
1123 	CU_ASSERT(rdma_req.data.wr.sg_list[5].length == 512);
1124 	CU_ASSERT(rdma_req.data.wr.sg_list[5].lkey == RDMA_UT_LKEY);
1125 
1126 	CU_ASSERT(rdma_req.data.wr.sg_list[6].addr == 0x2000 + 24 + 512 + md_size * 2);
1127 	CU_ASSERT(rdma_req.data.wr.sg_list[6].length == 512);
1128 	CU_ASSERT(rdma_req.data.wr.sg_list[6].lkey == RDMA_UT_LKEY);
1129 
1130 	/* Part 7: simple I/O, number of SGL entries exceeds the number of entries
1131 	   one WR can hold. Additional WR is chained */
1132 	MOCK_SET(spdk_mempool_get, data2_buffer);
1133 	aligned_buffer = (void *)((uintptr_t)(data2_buffer + NVMF_DATA_BUFFER_MASK) &
1134 				  ~NVMF_DATA_BUFFER_MASK);
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);
1139 	rdma_req.req.dif_enabled = true;
1140 	rtransport.transport.opts.io_unit_size = data_bs * 16;
1141 	sgl->keyed.length = data_bs * 16;
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 * 16);
1148 	CU_ASSERT(rdma_req.req.iovcnt == 2);
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) * 16);
1151 	CU_ASSERT(rdma_req.req.data == aligned_buffer);
1152 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
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 	for (i = 0; i < 15; ++i) {
1157 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)aligned_buffer + i * (data_bs + md_size));
1158 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1159 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1160 	}
1161 
1162 	/* 8192 - (512 + 8) * 15 = 392 */
1163 	CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)aligned_buffer + i * (data_bs + md_size));
1164 	CU_ASSERT(rdma_req.data.wr.sg_list[i].length == 392);
1165 	CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1166 
1167 	/* additional wr from pool */
1168 	CU_ASSERT(rdma_req.data.wr.next == (void *)&data2->wr);
1169 	CU_ASSERT(rdma_req.data.wr.next->num_sge == 1);
1170 	CU_ASSERT(rdma_req.data.wr.next->next == &rdma_req.rsp.wr);
1171 	/* 2nd IO buffer */
1172 	CU_ASSERT(data2->wr.sg_list[0].addr == (uintptr_t)aligned_buffer);
1173 	CU_ASSERT(data2->wr.sg_list[0].length == 120);
1174 	CU_ASSERT(data2->wr.sg_list[0].lkey == RDMA_UT_LKEY);
1175 
1176 	/* Part 8: simple I/O, data with metadata do not fit to 1 io_buffer */
1177 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1178 	reset_nvmf_rdma_request(&rdma_req);
1179 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1180 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1181 			  0, 0, 0, 0, 0);
1182 	rdma_req.req.dif_enabled = true;
1183 	rtransport.transport.opts.io_unit_size = 516;
1184 	sgl->keyed.length = data_bs * 2;
1185 
1186 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1187 
1188 	CU_ASSERT(rc == 0);
1189 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1190 	CU_ASSERT(rdma_req.req.length == data_bs * 2);
1191 	CU_ASSERT(rdma_req.req.iovcnt == 3);
1192 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1193 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 2);
1194 	CU_ASSERT(rdma_req.req.data == (void *)0x2000);
1195 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1196 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1197 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1198 
1199 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1200 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 512);
1201 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1202 
1203 	/* 2nd IO buffer consumed, offset 4 bytes due to part of the metadata
1204 	  is located at the beginning of that buffer */
1205 	CU_ASSERT(rdma_req.data.wr.sg_list[1].addr == 0x2000 + 4);
1206 	CU_ASSERT(rdma_req.data.wr.sg_list[1].length == 512);
1207 	CU_ASSERT(rdma_req.data.wr.sg_list[1].lkey == RDMA_UT_LKEY);
1208 
1209 	/* Test 2: Multi SGL */
1210 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1211 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1212 	sgl->address = 0;
1213 	rdma_req.recv->buf = (void *)&sgl_desc;
1214 	MOCK_SET(spdk_mempool_get, data_buffer);
1215 	aligned_buffer = (void *)((uintptr_t)(data_buffer + NVMF_DATA_BUFFER_MASK) &
1216 				  ~NVMF_DATA_BUFFER_MASK);
1217 
1218 	/* part 1: 2 segments each with 1 wr. io_unit_size is aligned with data_bs + md_size */
1219 	reset_nvmf_rdma_request(&rdma_req);
1220 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1221 			  SPDK_DIF_TYPE1,
1222 			  SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 0, 0, 0, 0, 0);
1223 	rdma_req.req.dif_enabled = true;
1224 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1225 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
1226 
1227 	for (i = 0; i < 2; i++) {
1228 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1229 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1230 		sgl_desc[i].keyed.length = data_bs * 4;
1231 		sgl_desc[i].address = 0x4000 + i * data_bs * 4;
1232 		sgl_desc[i].keyed.key = 0x44;
1233 	}
1234 
1235 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1236 
1237 	CU_ASSERT(rc == 0);
1238 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1239 	CU_ASSERT(rdma_req.req.length == data_bs * 4 * 2);
1240 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1241 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4 * 2);
1242 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1243 	for (i = 0; i < 4; ++i) {
1244 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)((unsigned char *)aligned_buffer) + i *
1245 			  (data_bs + md_size));
1246 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1247 	}
1248 
1249 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
1250 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
1251 	CU_ASSERT(rdma_req.data.wr.next == &data->wr);
1252 	CU_ASSERT(data->wr.wr.rdma.rkey == 0x44);
1253 	CU_ASSERT(data->wr.wr.rdma.remote_addr == 0x4000 + data_bs * 4);
1254 	CU_ASSERT(data->wr.num_sge == 4);
1255 	for (i = 0; i < 4; ++i) {
1256 		CU_ASSERT(data->wr.sg_list[i].addr == (uintptr_t)((unsigned char *)aligned_buffer) + i *
1257 			  (data_bs + md_size));
1258 		CU_ASSERT(data->wr.sg_list[i].length == data_bs);
1259 	}
1260 
1261 	CU_ASSERT(data->wr.next == &rdma_req.rsp.wr);
1262 }
1263 
1264 static void
1265 test_nvmf_rdma_opts_init(void)
1266 {
1267 	struct spdk_nvmf_transport_opts	opts = {};
1268 
1269 	nvmf_rdma_opts_init(&opts);
1270 	CU_ASSERT(opts.max_queue_depth == SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH);
1271 	CU_ASSERT(opts.max_qpairs_per_ctrlr ==	SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR);
1272 	CU_ASSERT(opts.in_capsule_data_size ==	SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE);
1273 	CU_ASSERT(opts.max_io_size == SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE);
1274 	CU_ASSERT(opts.io_unit_size == SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
1275 	CU_ASSERT(opts.max_aq_depth == SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH);
1276 	CU_ASSERT(opts.num_shared_buffers == SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS);
1277 	CU_ASSERT(opts.buf_cache_size == SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE);
1278 	CU_ASSERT(opts.dif_insert_or_strip == SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP);
1279 	CU_ASSERT(opts.abort_timeout_sec == SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC);
1280 	CU_ASSERT(opts.transport_specific == NULL);
1281 }
1282 
1283 static void
1284 test_nvmf_rdma_request_free_data(void)
1285 {
1286 	struct spdk_nvmf_rdma_request rdma_req = {};
1287 	struct spdk_nvmf_rdma_transport rtransport = {};
1288 	struct spdk_nvmf_rdma_request_data *next_request_data = NULL;
1289 
1290 	MOCK_CLEAR(spdk_mempool_get);
1291 	rtransport.data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data",
1292 				  SPDK_NVMF_MAX_SGL_ENTRIES,
1293 				  sizeof(struct spdk_nvmf_rdma_request_data),
1294 				  SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
1295 				  SPDK_ENV_SOCKET_ID_ANY);
1296 	next_request_data = spdk_mempool_get(rtransport.data_wr_pool);
1297 	SPDK_CU_ASSERT_FATAL(((struct test_mempool *)rtransport.data_wr_pool)->count ==
1298 			     SPDK_NVMF_MAX_SGL_ENTRIES - 1);
1299 	next_request_data->wr.wr_id = 1;
1300 	next_request_data->wr.num_sge = 2;
1301 	next_request_data->wr.next = NULL;
1302 	rdma_req.data.wr.next = &next_request_data->wr;
1303 	rdma_req.data.wr.wr_id = 1;
1304 	rdma_req.data.wr.num_sge = 2;
1305 
1306 	nvmf_rdma_request_free_data(&rdma_req, &rtransport);
1307 	/* Check if next_request_data put into memory pool */
1308 	CU_ASSERT(((struct test_mempool *)rtransport.data_wr_pool)->count == SPDK_NVMF_MAX_SGL_ENTRIES);
1309 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
1310 
1311 	spdk_mempool_free(rtransport.data_wr_pool);
1312 }
1313 
1314 static void
1315 test_nvmf_rdma_update_ibv_state(void)
1316 {
1317 	struct spdk_nvmf_rdma_qpair rqpair = {};
1318 	struct spdk_rdma_qp rdma_qp = {};
1319 	struct ibv_qp qp = {};
1320 	int rc = 0;
1321 
1322 	rqpair.rdma_qp = &rdma_qp;
1323 
1324 	/* Case 1: Failed to get updated RDMA queue pair state */
1325 	rqpair.ibv_state = IBV_QPS_INIT;
1326 	rqpair.rdma_qp->qp = NULL;
1327 
1328 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1329 	CU_ASSERT(rc == IBV_QPS_ERR + 1);
1330 
1331 	/* Case 2: Bad state updated */
1332 	rqpair.rdma_qp->qp = &qp;
1333 	qp.state = IBV_QPS_ERR;
1334 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1335 	CU_ASSERT(rqpair.ibv_state == 10);
1336 	CU_ASSERT(rc == IBV_QPS_ERR + 1);
1337 
1338 	/* Case 3: Pass */
1339 	qp.state = IBV_QPS_INIT;
1340 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1341 	CU_ASSERT(rqpair.ibv_state == IBV_QPS_INIT);
1342 	CU_ASSERT(rc == IBV_QPS_INIT);
1343 }
1344 
1345 int main(int argc, char **argv)
1346 {
1347 	CU_pSuite	suite = NULL;
1348 	unsigned int	num_failures;
1349 
1350 	CU_set_error_action(CUEA_ABORT);
1351 	CU_initialize_registry();
1352 
1353 	suite = CU_add_suite("nvmf", NULL, NULL);
1354 
1355 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl);
1356 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_process);
1357 	CU_ADD_TEST(suite, test_nvmf_rdma_get_optimal_poll_group);
1358 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl_with_md);
1359 	CU_ADD_TEST(suite, test_nvmf_rdma_opts_init);
1360 	CU_ADD_TEST(suite, test_nvmf_rdma_request_free_data);
1361 	CU_ADD_TEST(suite, test_nvmf_rdma_update_ibv_state);
1362 
1363 	CU_basic_set_mode(CU_BRM_VERBOSE);
1364 	CU_basic_run_tests();
1365 	num_failures = CU_get_number_of_failures();
1366 	CU_cleanup_registry();
1367 	return num_failures;
1368 }
1369