xref: /spdk/test/unit/lib/nvmf/rdma.c/rdma_ut.c (revision 9889ab2dc80e40dae92dcef361d53dcba722043d)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright (c) Intel Corporation. All rights reserved.
5  *   Copyright (c) 2019 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 "nvmf/rdma.c"
38 #include "nvmf/transport.c"
39 
40 uint64_t g_mr_size;
41 uint64_t g_mr_next_size;
42 struct ibv_mr g_rdma_mr;
43 
44 #define RDMA_UT_UNITS_IN_MAX_IO 16
45 
46 struct spdk_nvmf_transport_opts g_rdma_ut_transport_opts = {
47 	.max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH,
48 	.max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR,
49 	.in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE,
50 	.max_io_size = (SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE * RDMA_UT_UNITS_IN_MAX_IO),
51 	.io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE,
52 	.max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH,
53 	.num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS,
54 };
55 
56 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_tcp;
57 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_fc;
58 
59 SPDK_LOG_REGISTER_COMPONENT("nvmf", SPDK_LOG_NVMF)
60 DEFINE_STUB(spdk_mem_map_set_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
61 		uint64_t size, uint64_t translation), 0);
62 DEFINE_STUB(spdk_mem_map_clear_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
63 		uint64_t size), 0);
64 DEFINE_STUB(spdk_mem_map_alloc, struct spdk_mem_map *, (uint64_t default_translation,
65 		const struct spdk_mem_map_ops *ops, void *cb_ctx), NULL);
66 DEFINE_STUB(spdk_nvmf_qpair_disconnect, int, (struct spdk_nvmf_qpair *qpair,
67 		nvmf_qpair_disconnect_cb cb_fn, void *ctx), 0);
68 DEFINE_STUB_V(spdk_mem_map_free, (struct spdk_mem_map **pmap));
69 
70 struct spdk_trace_histories *g_trace_histories;
71 DEFINE_STUB_V(spdk_trace_add_register_fn, (struct spdk_trace_register_fn *reg_fn));
72 DEFINE_STUB_V(spdk_trace_register_object, (uint8_t type, char id_prefix));
73 DEFINE_STUB_V(spdk_trace_register_description, (const char *name,
74 		uint16_t tpoint_id, uint8_t owner_type, uint8_t object_type, uint8_t new_object,
75 		uint8_t arg1_type, const char *arg1_name));
76 DEFINE_STUB_V(_spdk_trace_record, (uint64_t tsc, uint16_t tpoint_id, uint16_t poller_id,
77 				   uint32_t size, uint64_t object_id, uint64_t arg1));
78 
79 DEFINE_STUB_V(spdk_nvmf_request_exec, (struct spdk_nvmf_request *req));
80 DEFINE_STUB(spdk_nvme_transport_id_compare, int, (const struct spdk_nvme_transport_id *trid1,
81 		const struct spdk_nvme_transport_id *trid2), 0);
82 DEFINE_STUB_V(spdk_nvmf_ctrlr_abort_aer, (struct spdk_nvmf_ctrlr *ctrlr));
83 DEFINE_STUB(spdk_nvmf_request_get_dif_ctx, bool, (struct spdk_nvmf_request *req,
84 		struct spdk_dif_ctx *dif_ctx), false);
85 
86 uint64_t
87 spdk_mem_map_translate(const struct spdk_mem_map *map, uint64_t vaddr, uint64_t *size)
88 {
89 	if (g_mr_size != 0) {
90 		*(uint32_t *)size = g_mr_size;
91 		if (g_mr_next_size != 0) {
92 			g_mr_size = g_mr_next_size;
93 		}
94 	}
95 
96 	return (uint64_t)&g_rdma_mr;
97 }
98 
99 static void reset_nvmf_rdma_request(struct spdk_nvmf_rdma_request *rdma_req)
100 {
101 	int i;
102 
103 	rdma_req->req.length = 0;
104 	rdma_req->req.data_from_pool = false;
105 	rdma_req->req.data = NULL;
106 	rdma_req->data.wr.num_sge = 0;
107 	rdma_req->data.wr.wr.rdma.remote_addr = 0;
108 	rdma_req->data.wr.wr.rdma.rkey = 0;
109 	memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
110 
111 	for (i = 0; i < SPDK_NVMF_MAX_SGL_ENTRIES; i++) {
112 		rdma_req->req.iov[i].iov_base = 0;
113 		rdma_req->req.iov[i].iov_len = 0;
114 		rdma_req->req.buffers[i] = 0;
115 		rdma_req->data.wr.sg_list[i].addr = 0;
116 		rdma_req->data.wr.sg_list[i].length = 0;
117 		rdma_req->data.wr.sg_list[i].lkey = 0;
118 	}
119 	rdma_req->req.iovcnt = 0;
120 }
121 
122 static void
123 test_spdk_nvmf_rdma_request_parse_sgl(void)
124 {
125 	struct spdk_nvmf_rdma_transport rtransport;
126 	struct spdk_nvmf_rdma_device device;
127 	struct spdk_nvmf_rdma_request rdma_req = {};
128 	struct spdk_nvmf_rdma_recv recv;
129 	struct spdk_nvmf_rdma_poll_group group;
130 	struct spdk_nvmf_rdma_qpair rqpair;
131 	struct spdk_nvmf_rdma_poller poller;
132 	union nvmf_c2h_msg cpl;
133 	union nvmf_h2c_msg cmd;
134 	struct spdk_nvme_sgl_descriptor *sgl;
135 	struct spdk_nvmf_transport_pg_cache_buf bufs[4];
136 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
137 	struct spdk_nvmf_rdma_request_data data;
138 	struct spdk_nvmf_transport_pg_cache_buf	buffer;
139 	struct spdk_nvmf_transport_pg_cache_buf	*buffer_ptr;
140 	int rc, i;
141 
142 	data.wr.sg_list = data.sgl;
143 	STAILQ_INIT(&group.group.buf_cache);
144 	group.group.buf_cache_size = 0;
145 	group.group.buf_cache_count = 0;
146 	group.group.transport = &rtransport.transport;
147 	STAILQ_INIT(&group.retired_bufs);
148 	poller.group = &group;
149 	rqpair.poller = &poller;
150 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
151 
152 	sgl = &cmd.nvme_cmd.dptr.sgl1;
153 	rdma_req.recv = &recv;
154 	rdma_req.req.cmd = &cmd;
155 	rdma_req.req.rsp = &cpl;
156 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
157 	rdma_req.req.qpair = &rqpair.qpair;
158 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
159 
160 	rtransport.transport.opts = g_rdma_ut_transport_opts;
161 	rtransport.data_wr_pool = NULL;
162 	rtransport.transport.data_buf_pool = NULL;
163 
164 	device.attr.device_cap_flags = 0;
165 	g_rdma_mr.lkey = 0xABCD;
166 	sgl->keyed.key = 0xEEEE;
167 	sgl->address = 0xFFFF;
168 	rdma_req.recv->buf = (void *)0xDDDD;
169 
170 	/* Test 1: sgl type: keyed data block subtype: address */
171 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
172 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
173 
174 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size */
175 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
176 	reset_nvmf_rdma_request(&rdma_req);
177 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
178 
179 	device.map = (void *)0x0;
180 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
181 	CU_ASSERT(rc == 0);
182 	CU_ASSERT(rdma_req.req.data_from_pool == true);
183 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
184 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
185 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
186 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
187 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
188 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
189 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
190 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
191 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
192 
193 	/* Part 2: simple I/O, one SGL larger than the transport io unit size (equal to the max io size) */
194 	reset_nvmf_rdma_request(&rdma_req);
195 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
196 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
197 
198 	CU_ASSERT(rc == 0);
199 	CU_ASSERT(rdma_req.req.data_from_pool == true);
200 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO);
201 	CU_ASSERT(rdma_req.data.wr.num_sge == RDMA_UT_UNITS_IN_MAX_IO);
202 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
203 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
204 	for (i = 0; i < RDMA_UT_UNITS_IN_MAX_IO; i++) {
205 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
206 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
207 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
208 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
209 	}
210 
211 	/* Part 3: simple I/O one SGL larger than the transport max io size */
212 	reset_nvmf_rdma_request(&rdma_req);
213 	sgl->keyed.length = rtransport.transport.opts.max_io_size * 2;
214 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
215 
216 	CU_ASSERT(rc == -1);
217 
218 	/* Part 4: Pretend there are no buffer pools */
219 	MOCK_SET(spdk_mempool_get, NULL);
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 = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
223 
224 	CU_ASSERT(rc == 0);
225 	CU_ASSERT(rdma_req.req.data_from_pool == false);
226 	CU_ASSERT(rdma_req.req.data == NULL);
227 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
228 	CU_ASSERT(rdma_req.req.buffers[0] == NULL);
229 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0);
230 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 0);
231 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == 0);
232 
233 	rdma_req.recv->buf = (void *)0xDDDD;
234 	/* Test 2: sgl type: keyed data block subtype: offset (in capsule data) */
235 	sgl->generic.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
236 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
237 
238 	/* Part 1: Normal I/O smaller than in capsule data size no offset */
239 	reset_nvmf_rdma_request(&rdma_req);
240 	sgl->address = 0;
241 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
242 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
243 
244 	CU_ASSERT(rc == 0);
245 	CU_ASSERT(rdma_req.req.data == (void *)0xDDDD);
246 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.in_capsule_data_size);
247 	CU_ASSERT(rdma_req.req.data_from_pool == false);
248 
249 	/* Part 2: I/O offset + length too large */
250 	reset_nvmf_rdma_request(&rdma_req);
251 	sgl->address = rtransport.transport.opts.in_capsule_data_size;
252 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
253 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
254 
255 	CU_ASSERT(rc == -1);
256 
257 	/* Part 3: I/O too large */
258 	reset_nvmf_rdma_request(&rdma_req);
259 	sgl->address = 0;
260 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size * 2;
261 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
262 
263 	CU_ASSERT(rc == -1);
264 
265 	/* Test 3: Multi SGL */
266 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
267 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
268 	sgl->address = 0;
269 	rdma_req.recv->buf = (void *)&sgl_desc;
270 	MOCK_SET(spdk_mempool_get, &data);
271 
272 	/* part 1: 2 segments each with 1 wr. */
273 	reset_nvmf_rdma_request(&rdma_req);
274 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
275 	for (i = 0; i < 2; i++) {
276 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
277 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
278 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size;
279 		sgl_desc[i].address = 0x4000 + i * rtransport.transport.opts.io_unit_size;
280 		sgl_desc[i].keyed.key = 0x44;
281 	}
282 
283 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
284 
285 	CU_ASSERT(rc == 0);
286 	CU_ASSERT(rdma_req.req.data_from_pool == true);
287 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 2);
288 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
289 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
290 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
291 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
292 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
293 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size);
294 	CU_ASSERT(data.wr.num_sge == 1);
295 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
296 
297 	/* part 2: 2 segments, each with 1 wr containing 8 sge_elements */
298 	reset_nvmf_rdma_request(&rdma_req);
299 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
300 	for (i = 0; i < 2; i++) {
301 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
302 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
303 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size * 8;
304 		sgl_desc[i].address = 0x4000 + i * 8 * rtransport.transport.opts.io_unit_size;
305 		sgl_desc[i].keyed.key = 0x44;
306 	}
307 
308 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
309 
310 	CU_ASSERT(rc == 0);
311 	CU_ASSERT(rdma_req.req.data_from_pool == true);
312 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
313 	CU_ASSERT(rdma_req.req.iovcnt == 16);
314 	CU_ASSERT(rdma_req.data.wr.num_sge == 8);
315 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
316 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
317 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
318 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
319 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 8);
320 	CU_ASSERT(data.wr.num_sge == 8);
321 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
322 
323 	/* part 3: 2 segments, one very large, one very small */
324 	reset_nvmf_rdma_request(&rdma_req);
325 	for (i = 0; i < 2; i++) {
326 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
327 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
328 		sgl_desc[i].keyed.key = 0x44;
329 	}
330 
331 	sgl_desc[0].keyed.length = rtransport.transport.opts.io_unit_size * 15 +
332 				   rtransport.transport.opts.io_unit_size / 2;
333 	sgl_desc[0].address = 0x4000;
334 	sgl_desc[1].keyed.length = rtransport.transport.opts.io_unit_size / 2;
335 	sgl_desc[1].address = 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
336 			      rtransport.transport.opts.io_unit_size / 2;
337 
338 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
339 
340 	CU_ASSERT(rc == 0);
341 	CU_ASSERT(rdma_req.req.data_from_pool == true);
342 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
343 	CU_ASSERT(rdma_req.req.iovcnt == 17);
344 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
345 	for (i = 0; i < 15; i++) {
346 		CU_ASSERT(rdma_req.data.sgl[i].length == rtransport.transport.opts.io_unit_size);
347 	}
348 	CU_ASSERT(rdma_req.data.sgl[15].length == rtransport.transport.opts.io_unit_size / 2);
349 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
350 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
351 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
352 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
353 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
354 		  rtransport.transport.opts.io_unit_size / 2);
355 	CU_ASSERT(data.sgl[0].length == rtransport.transport.opts.io_unit_size / 2);
356 	CU_ASSERT(data.wr.num_sge == 1);
357 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
358 
359 	/* Test 4: use PG buffer cache */
360 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
361 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
362 	sgl->address = 0xFFFF;
363 	rdma_req.recv->buf = (void *)0xDDDD;
364 	g_rdma_mr.lkey = 0xABCD;
365 	sgl->keyed.key = 0xEEEE;
366 
367 	for (i = 0; i < 4; i++) {
368 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
369 	}
370 
371 	/* part 1: use the four buffers from the pg cache */
372 	group.group.buf_cache_size = 4;
373 	group.group.buf_cache_count = 4;
374 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
375 	reset_nvmf_rdma_request(&rdma_req);
376 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * 4;
377 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
378 
379 	SPDK_CU_ASSERT_FATAL(rc == 0);
380 	CU_ASSERT(rdma_req.req.data_from_pool == true);
381 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
382 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
383 			~NVMF_DATA_BUFFER_MASK));
384 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
385 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
386 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
387 	CU_ASSERT(group.group.buf_cache_count == 0);
388 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
389 	for (i = 0; i < 4; i++) {
390 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
391 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
392 				~NVMF_DATA_BUFFER_MASK));
393 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
394 	}
395 
396 	/* part 2: now that we have used the buffers from the cache, try again. We should get mempool buffers. */
397 	reset_nvmf_rdma_request(&rdma_req);
398 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
399 
400 	SPDK_CU_ASSERT_FATAL(rc == 0);
401 	CU_ASSERT(rdma_req.req.data_from_pool == true);
402 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
403 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
404 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
405 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
406 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
407 	CU_ASSERT(group.group.buf_cache_count == 0);
408 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
409 	for (i = 0; i < 4; i++) {
410 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
411 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
412 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
413 		CU_ASSERT(group.group.buf_cache_count == 0);
414 	}
415 
416 	/* part 3: half and half */
417 	group.group.buf_cache_count = 2;
418 
419 	for (i = 0; i < 2; i++) {
420 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
421 	}
422 	reset_nvmf_rdma_request(&rdma_req);
423 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
424 
425 	SPDK_CU_ASSERT_FATAL(rc == 0);
426 	CU_ASSERT(rdma_req.req.data_from_pool == true);
427 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
428 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
429 			~NVMF_DATA_BUFFER_MASK));
430 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
431 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
432 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
433 	CU_ASSERT(group.group.buf_cache_count == 0);
434 	for (i = 0; i < 2; i++) {
435 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
436 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
437 				~NVMF_DATA_BUFFER_MASK));
438 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
439 	}
440 	for (i = 2; i < 4; i++) {
441 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
442 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
443 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
444 	}
445 
446 	reset_nvmf_rdma_request(&rdma_req);
447 	/* Test 5 dealing with a buffer split over two Memory Regions */
448 	MOCK_SET(spdk_mempool_get, (void *)&buffer);
449 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
450 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
451 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
452 	g_mr_size = rtransport.transport.opts.io_unit_size / 4;
453 	g_mr_next_size = rtransport.transport.opts.io_unit_size / 2;
454 
455 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
456 	SPDK_CU_ASSERT_FATAL(rc == 0);
457 	CU_ASSERT(rdma_req.req.data_from_pool == true);
458 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
459 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&buffer + NVMF_DATA_BUFFER_MASK) &
460 			~NVMF_DATA_BUFFER_MASK));
461 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
462 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
463 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
464 	CU_ASSERT(rdma_req.req.buffers[0] == &buffer);
465 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == (((uint64_t)&buffer + NVMF_DATA_BUFFER_MASK) &
466 			~NVMF_DATA_BUFFER_MASK));
467 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
468 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
469 	buffer_ptr = STAILQ_FIRST(&group.retired_bufs);
470 	CU_ASSERT(buffer_ptr == &buffer);
471 	STAILQ_REMOVE(&group.retired_bufs, buffer_ptr, spdk_nvmf_transport_pg_cache_buf, link);
472 	CU_ASSERT(STAILQ_EMPTY(&group.retired_bufs));
473 	g_mr_size = 0;
474 	g_mr_next_size = 0;
475 
476 	reset_nvmf_rdma_request(&rdma_req);
477 }
478 
479 static struct spdk_nvmf_rdma_recv *
480 create_recv(struct spdk_nvmf_rdma_qpair *rqpair, enum spdk_nvme_nvm_opcode opc)
481 {
482 	struct spdk_nvmf_rdma_recv *rdma_recv;
483 	union nvmf_h2c_msg *cmd;
484 	struct spdk_nvme_sgl_descriptor *sgl;
485 
486 	rdma_recv = calloc(1, sizeof(*rdma_recv));
487 	rdma_recv->qpair = rqpair;
488 	cmd = calloc(1, sizeof(*cmd));
489 	rdma_recv->sgl[0].addr = (uintptr_t)cmd;
490 	cmd->nvme_cmd.opc = opc;
491 	sgl = &cmd->nvme_cmd.dptr.sgl1;
492 	sgl->keyed.key = 0xEEEE;
493 	sgl->address = 0xFFFF;
494 	sgl->keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
495 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
496 	sgl->keyed.length = 1;
497 
498 	return rdma_recv;
499 }
500 
501 static void
502 free_recv(struct spdk_nvmf_rdma_recv *rdma_recv)
503 {
504 	free((void *)rdma_recv->sgl[0].addr);
505 	free(rdma_recv);
506 }
507 
508 static struct spdk_nvmf_rdma_request *
509 create_req(struct spdk_nvmf_rdma_qpair *rqpair,
510 	   struct spdk_nvmf_rdma_recv *rdma_recv)
511 {
512 	struct spdk_nvmf_rdma_request *rdma_req;
513 	union nvmf_c2h_msg *cpl;
514 
515 	rdma_req = calloc(1, sizeof(*rdma_req));
516 	rdma_req->recv = rdma_recv;
517 	rdma_req->req.qpair = &rqpair->qpair;
518 	rdma_req->state = RDMA_REQUEST_STATE_NEW;
519 	rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr;
520 	rdma_req->data.wr.sg_list = rdma_req->data.sgl;
521 	cpl = calloc(1, sizeof(*cpl));
522 	rdma_req->rsp.sgl[0].addr = (uintptr_t)cpl;
523 	rdma_req->req.rsp = cpl;
524 
525 	return rdma_req;
526 }
527 
528 static void
529 free_req(struct spdk_nvmf_rdma_request *rdma_req)
530 {
531 	free((void *)rdma_req->rsp.sgl[0].addr);
532 	free(rdma_req);
533 }
534 
535 static void
536 qpair_reset(struct spdk_nvmf_rdma_qpair *rqpair,
537 	    struct spdk_nvmf_rdma_poller *poller,
538 	    struct spdk_nvmf_rdma_port *port,
539 	    struct spdk_nvmf_rdma_resources *resources)
540 {
541 	memset(rqpair, 0, sizeof(*rqpair));
542 	STAILQ_INIT(&rqpair->pending_rdma_write_queue);
543 	STAILQ_INIT(&rqpair->pending_rdma_read_queue);
544 	rqpair->poller = poller;
545 	rqpair->port = port;
546 	rqpair->resources = resources;
547 	rqpair->qpair.qid = 1;
548 	rqpair->ibv_state = IBV_QPS_RTS;
549 	rqpair->qpair.state = SPDK_NVMF_QPAIR_ACTIVE;
550 	rqpair->max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
551 	rqpair->max_send_depth = 16;
552 	rqpair->max_read_depth = 16;
553 	resources->recvs_to_post.first = resources->recvs_to_post.last = NULL;
554 }
555 
556 static void
557 poller_reset(struct spdk_nvmf_rdma_poller *poller,
558 	     struct spdk_nvmf_rdma_poll_group *group)
559 {
560 	memset(poller, 0, sizeof(*poller));
561 	STAILQ_INIT(&poller->qpairs_pending_recv);
562 	STAILQ_INIT(&poller->qpairs_pending_send);
563 	poller->group = group;
564 }
565 
566 static void
567 test_spdk_nvmf_rdma_request_process(void)
568 {
569 	struct spdk_nvmf_rdma_transport rtransport = {};
570 	struct spdk_nvmf_rdma_poll_group group = {};
571 	struct spdk_nvmf_rdma_poller poller = {};
572 	struct spdk_nvmf_rdma_port port = {};
573 	struct spdk_nvmf_rdma_device device = {};
574 	struct spdk_nvmf_rdma_resources resources = {};
575 	struct spdk_nvmf_rdma_qpair rqpair = {};
576 	struct spdk_nvmf_rdma_recv *rdma_recv;
577 	struct spdk_nvmf_rdma_request *rdma_req;
578 	bool progress;
579 
580 	STAILQ_INIT(&group.group.buf_cache);
581 	STAILQ_INIT(&group.group.pending_buf_queue);
582 	group.group.buf_cache_size = 0;
583 	group.group.buf_cache_count = 0;
584 	port.device = &device;
585 	poller_reset(&poller, &group);
586 	qpair_reset(&rqpair, &poller, &port, &resources);
587 
588 	rtransport.transport.opts = g_rdma_ut_transport_opts;
589 	rtransport.transport.data_buf_pool = spdk_mempool_create("test_data_pool", 16, 128, 0, 0);
590 	rtransport.data_wr_pool = spdk_mempool_create("test_wr_pool", 128,
591 				  sizeof(struct spdk_nvmf_rdma_request_data),
592 				  0, 0);
593 	MOCK_CLEAR(spdk_mempool_get);
594 
595 	device.attr.device_cap_flags = 0;
596 	device.map = (void *)0x0;
597 	g_rdma_mr.lkey = 0xABCD;
598 
599 	/* Test 1: single SGL READ request */
600 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_READ);
601 	rdma_req = create_req(&rqpair, rdma_recv);
602 	rqpair.current_recv_depth = 1;
603 	/* NEW -> EXECUTING */
604 	progress = spdk_nvmf_rdma_request_process(&rtransport, rdma_req);
605 	CU_ASSERT(progress == true);
606 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
607 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST);
608 	/* EXECUTED -> TRANSFERRING_C2H */
609 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
610 	progress = spdk_nvmf_rdma_request_process(&rtransport, rdma_req);
611 	CU_ASSERT(progress == true);
612 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST);
613 	CU_ASSERT(rdma_req->recv == NULL);
614 	CU_ASSERT(rqpair.sends_to_post.first == &rdma_req->data.wr);
615 	CU_ASSERT(rqpair.sends_to_post.last == &rdma_req->rsp.wr);
616 	CU_ASSERT(resources.recvs_to_post.first == &rdma_recv->wr);
617 	CU_ASSERT(resources.recvs_to_post.last == &rdma_recv->wr);
618 	/* COMPLETED -> FREE */
619 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
620 	progress = spdk_nvmf_rdma_request_process(&rtransport, rdma_req);
621 	CU_ASSERT(progress == true);
622 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
623 
624 	free_recv(rdma_recv);
625 	free_req(rdma_req);
626 	poller_reset(&poller, &group);
627 	qpair_reset(&rqpair, &poller, &port, &resources);
628 
629 	/* Test 2: single SGL WRITE request */
630 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
631 	rdma_req = create_req(&rqpair, rdma_recv);
632 	rqpair.current_recv_depth = 1;
633 	/* NEW -> TRANSFERRING_H2C */
634 	progress = spdk_nvmf_rdma_request_process(&rtransport, rdma_req);
635 	CU_ASSERT(progress == true);
636 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
637 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
638 	CU_ASSERT(rqpair.sends_to_post.first == &rdma_req->data.wr);
639 	CU_ASSERT(rqpair.sends_to_post.last == &rdma_req->data.wr);
640 	rqpair.sends_to_post.first = rqpair.sends_to_post.last = NULL;
641 	STAILQ_INIT(&poller.qpairs_pending_send);
642 	/* READY_TO_EXECUTE -> EXECUTING */
643 	rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
644 	progress = spdk_nvmf_rdma_request_process(&rtransport, rdma_req);
645 	CU_ASSERT(progress == true);
646 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
647 	/* EXECUTED -> COMPLETING */
648 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
649 	progress = spdk_nvmf_rdma_request_process(&rtransport, rdma_req);
650 	CU_ASSERT(progress == true);
651 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
652 	CU_ASSERT(rdma_req->recv == NULL);
653 	CU_ASSERT(rqpair.sends_to_post.first == &rdma_req->rsp.wr);
654 	CU_ASSERT(rqpair.sends_to_post.last == &rdma_req->rsp.wr);
655 	CU_ASSERT(resources.recvs_to_post.first == &rdma_recv->wr);
656 	CU_ASSERT(resources.recvs_to_post.last == &rdma_recv->wr);
657 	/* COMPLETED -> FREE */
658 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
659 	progress = spdk_nvmf_rdma_request_process(&rtransport, rdma_req);
660 	CU_ASSERT(progress == true);
661 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
662 
663 	free_recv(rdma_recv);
664 	free_req(rdma_req);
665 	poller_reset(&poller, &group);
666 	qpair_reset(&rqpair, &poller, &port, &resources);
667 
668 	/* Test 3: WRITE+WRITE ibv_send batching */
669 	{
670 		struct spdk_nvmf_rdma_recv *recv1, *recv2;
671 		struct spdk_nvmf_rdma_request *req1, *req2;
672 		recv1 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
673 		req1 = create_req(&rqpair, recv1);
674 		recv2 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
675 		req2 = create_req(&rqpair, recv2);
676 
677 		/* WRITE 1: NEW -> TRANSFERRING_H2C */
678 		rqpair.current_recv_depth = 1;
679 		spdk_nvmf_rdma_request_process(&rtransport, req1);
680 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
681 		/* WRITE 1 is the first in batching list */
682 		CU_ASSERT(rqpair.sends_to_post.first == &req1->data.wr);
683 		CU_ASSERT(rqpair.sends_to_post.last == &req1->data.wr);
684 
685 		/* WRITE 2: NEW -> TRANSFERRING_H2C */
686 		rqpair.current_recv_depth = 2;
687 		spdk_nvmf_rdma_request_process(&rtransport, req2);
688 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
689 		/* WRITE 2 is now also in the batching list */
690 		CU_ASSERT(rqpair.sends_to_post.first->next == &req2->data.wr);
691 		CU_ASSERT(rqpair.sends_to_post.last == &req2->data.wr);
692 
693 		/* Send everything */
694 		rqpair.sends_to_post.first = rqpair.sends_to_post.last = NULL;
695 		STAILQ_INIT(&poller.qpairs_pending_send);
696 
697 		/* WRITE 1 completes before WRITE 2 has finished RDMA reading */
698 		/* WRITE 1: READY_TO_EXECUTE -> EXECUTING */
699 		req1->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
700 		spdk_nvmf_rdma_request_process(&rtransport, req1);
701 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_EXECUTING);
702 		/* WRITE 1: EXECUTED -> COMPLETING */
703 		req1->state = RDMA_REQUEST_STATE_EXECUTED;
704 		spdk_nvmf_rdma_request_process(&rtransport, req1);
705 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_COMPLETING);
706 		CU_ASSERT(rqpair.sends_to_post.first == &req1->rsp.wr);
707 		CU_ASSERT(rqpair.sends_to_post.last == &req1->rsp.wr);
708 		rqpair.sends_to_post.first = rqpair.sends_to_post.last = NULL;
709 		STAILQ_INIT(&poller.qpairs_pending_send);
710 		/* WRITE 1: COMPLETED -> FREE */
711 		req1->state = RDMA_REQUEST_STATE_COMPLETED;
712 		spdk_nvmf_rdma_request_process(&rtransport, req1);
713 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_FREE);
714 
715 		/* Now WRITE 2 has finished reading and completes */
716 		/* WRITE 2: COMPLETED -> FREE */
717 		/* WRITE 2: READY_TO_EXECUTE -> EXECUTING */
718 		req2->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
719 		spdk_nvmf_rdma_request_process(&rtransport, req2);
720 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_EXECUTING);
721 		/* WRITE 1: EXECUTED -> COMPLETING */
722 		req2->state = RDMA_REQUEST_STATE_EXECUTED;
723 		spdk_nvmf_rdma_request_process(&rtransport, req2);
724 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_COMPLETING);
725 		CU_ASSERT(rqpair.sends_to_post.first == &req2->rsp.wr);
726 		CU_ASSERT(rqpair.sends_to_post.last == &req2->rsp.wr);
727 		rqpair.sends_to_post.first = rqpair.sends_to_post.last = NULL;
728 		STAILQ_INIT(&poller.qpairs_pending_send);
729 		/* WRITE 1: COMPLETED -> FREE */
730 		req2->state = RDMA_REQUEST_STATE_COMPLETED;
731 		spdk_nvmf_rdma_request_process(&rtransport, req2);
732 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_FREE);
733 
734 		free_recv(recv1);
735 		free_req(req1);
736 		free_recv(recv2);
737 		free_req(req2);
738 		poller_reset(&poller, &group);
739 		qpair_reset(&rqpair, &poller, &port, &resources);
740 	}
741 
742 	spdk_mempool_free(rtransport.transport.data_buf_pool);
743 	spdk_mempool_free(rtransport.data_wr_pool);
744 }
745 
746 #define TEST_GROUPS_COUNT 5
747 static void
748 test_spdk_nvmf_rdma_get_optimal_poll_group(void)
749 {
750 	struct spdk_nvmf_rdma_transport rtransport = {};
751 	struct spdk_nvmf_transport *transport = &rtransport.transport;
752 	struct spdk_nvmf_rdma_qpair rqpair = {};
753 	struct spdk_nvmf_transport_poll_group *groups[TEST_GROUPS_COUNT];
754 	struct spdk_nvmf_rdma_poll_group *rgroups[TEST_GROUPS_COUNT];
755 	struct spdk_nvmf_transport_poll_group *result;
756 	uint32_t i;
757 
758 	rqpair.qpair.transport = transport;
759 	pthread_mutex_init(&rtransport.lock, NULL);
760 	TAILQ_INIT(&rtransport.poll_groups);
761 
762 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
763 		groups[i] = spdk_nvmf_rdma_poll_group_create(transport);
764 		CU_ASSERT(groups[i] != NULL);
765 		rgroups[i] = SPDK_CONTAINEROF(groups[i], struct spdk_nvmf_rdma_poll_group, group);
766 		groups[i]->transport = transport;
767 	}
768 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[0]);
769 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[0]);
770 
771 	/* Emulate connection of %TEST_GROUPS_COUNT% initiators - each creates 1 admin and 1 io qp */
772 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
773 		rqpair.qpair.qid = 0;
774 		result = spdk_nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
775 		CU_ASSERT(result == groups[i]);
776 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
777 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i]);
778 
779 		rqpair.qpair.qid = 1;
780 		result = spdk_nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
781 		CU_ASSERT(result == groups[i]);
782 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
783 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
784 	}
785 	/* wrap around, admin/io pg point to the first pg
786 	   Destroy all poll groups except of the last one */
787 	for (i = 0; i < TEST_GROUPS_COUNT - 1; i++) {
788 		spdk_nvmf_rdma_poll_group_destroy(groups[i]);
789 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[i + 1]);
790 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i + 1]);
791 	}
792 
793 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
794 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
795 
796 	/* Check that pointers to the next admin/io poll groups are not changed */
797 	rqpair.qpair.qid = 0;
798 	result = spdk_nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
799 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
800 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
801 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
802 
803 	rqpair.qpair.qid = 1;
804 	result = spdk_nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
805 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
806 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
807 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
808 
809 	/* Remove the last poll group, check that pointers are NULL */
810 	spdk_nvmf_rdma_poll_group_destroy(groups[TEST_GROUPS_COUNT - 1]);
811 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == NULL);
812 	CU_ASSERT(rtransport.conn_sched.next_io_pg == NULL);
813 
814 	/* Request optimal poll group, result must be NULL */
815 	rqpair.qpair.qid = 0;
816 	result = spdk_nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
817 	CU_ASSERT(result == NULL);
818 
819 	rqpair.qpair.qid = 1;
820 	result = spdk_nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
821 	CU_ASSERT(result == NULL);
822 
823 	pthread_mutex_destroy(&rtransport.lock);
824 }
825 #undef TEST_GROUPS_COUNT
826 
827 static void
828 test_spdk_nvmf_rdma_request_parse_sgl_with_md(void)
829 {
830 	struct spdk_nvmf_rdma_transport rtransport;
831 	struct spdk_nvmf_rdma_device device;
832 	struct spdk_nvmf_rdma_request rdma_req = {};
833 	struct spdk_nvmf_rdma_recv recv;
834 	struct spdk_nvmf_rdma_poll_group group;
835 	struct spdk_nvmf_rdma_qpair rqpair;
836 	struct spdk_nvmf_rdma_poller poller;
837 	union nvmf_c2h_msg cpl;
838 	union nvmf_h2c_msg cmd;
839 	struct spdk_nvme_sgl_descriptor *sgl;
840 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
841 	struct spdk_nvmf_rdma_request_data data;
842 	struct spdk_nvmf_transport_pg_cache_buf	buffer;
843 	struct spdk_nvmf_transport_pg_cache_buf	*buffer_ptr;
844 	const uint32_t data_bs = 512;
845 	const uint32_t md_size = 8;
846 	int rc, i;
847 	void *aligned_buffer;
848 
849 	data.wr.sg_list = data.sgl;
850 	STAILQ_INIT(&group.group.buf_cache);
851 	group.group.buf_cache_size = 0;
852 	group.group.buf_cache_count = 0;
853 	group.group.transport = &rtransport.transport;
854 	STAILQ_INIT(&group.retired_bufs);
855 	poller.group = &group;
856 	rqpair.poller = &poller;
857 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
858 
859 	sgl = &cmd.nvme_cmd.dptr.sgl1;
860 	rdma_req.recv = &recv;
861 	rdma_req.req.cmd = &cmd;
862 	rdma_req.req.rsp = &cpl;
863 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
864 	rdma_req.req.qpair = &rqpair.qpair;
865 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
866 
867 	rtransport.transport.opts = g_rdma_ut_transport_opts;
868 	rtransport.data_wr_pool = NULL;
869 	rtransport.transport.data_buf_pool = NULL;
870 
871 	device.attr.device_cap_flags = 0;
872 	device.map = NULL;
873 	g_rdma_mr.lkey = 0xABCD;
874 	sgl->keyed.key = 0xEEEE;
875 	sgl->address = 0xFFFF;
876 	rdma_req.recv->buf = (void *)0xDDDD;
877 
878 	/* Test 1: sgl type: keyed data block subtype: address */
879 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
880 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
881 
882 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size, block size 512 */
883 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
884 	reset_nvmf_rdma_request(&rdma_req);
885 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
886 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
887 			  0, 0, 0, 0, 0);
888 	rdma_req.req.dif.dif_insert_or_strip = true;
889 	rtransport.transport.opts.io_unit_size = data_bs * 8;
890 	sgl->keyed.length = data_bs * 4;
891 
892 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
893 
894 	CU_ASSERT(rc == 0);
895 	CU_ASSERT(rdma_req.req.data_from_pool == true);
896 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
897 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
898 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
899 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
900 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
901 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
902 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
903 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
904 
905 	for (i = 0; i < 4; ++i) {
906 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
907 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
908 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
909 	}
910 
911 	/* Part 2: simple I/O, one SGL equal to io unit size, io_unit_size is not aligned with md_size,
912 		block size 512 */
913 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
914 	reset_nvmf_rdma_request(&rdma_req);
915 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
916 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
917 			  0, 0, 0, 0, 0);
918 	rdma_req.req.dif.dif_insert_or_strip = true;
919 	rtransport.transport.opts.io_unit_size = data_bs * 4;
920 	sgl->keyed.length = data_bs * 4;
921 
922 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
923 
924 	CU_ASSERT(rc == 0);
925 	CU_ASSERT(rdma_req.req.data_from_pool == true);
926 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
927 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
928 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
929 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
930 	CU_ASSERT(rdma_req.data.wr.num_sge == 5);
931 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
932 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
933 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
934 
935 	for (i = 0; i < 3; ++i) {
936 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
937 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
938 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
939 	}
940 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
941 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
942 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == g_rdma_mr.lkey);
943 
944 	/* 2nd buffer consumed */
945 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
946 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
947 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == g_rdma_mr.lkey);
948 
949 	/* Part 3: simple I/O, one SGL equal io unit size, io_unit_size is equal to block size 512 bytes */
950 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
951 	reset_nvmf_rdma_request(&rdma_req);
952 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
953 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
954 			  0, 0, 0, 0, 0);
955 	rdma_req.req.dif.dif_insert_or_strip = true;
956 	rtransport.transport.opts.io_unit_size = data_bs;
957 	sgl->keyed.length = data_bs;
958 
959 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
960 
961 	CU_ASSERT(rc == 0);
962 	CU_ASSERT(rdma_req.req.data_from_pool == true);
963 	CU_ASSERT(rdma_req.req.length == data_bs);
964 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
965 	CU_ASSERT(rdma_req.req.dif.elba_length == data_bs + md_size);
966 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
967 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
968 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
969 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
970 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
971 
972 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
973 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs);
974 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
975 
976 	CU_ASSERT(rdma_req.req.iovcnt == 2);
977 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)((unsigned long)0x2000));
978 	CU_ASSERT(rdma_req.req.iov[0].iov_len == data_bs);
979 	/* 2nd buffer consumed for metadata */
980 	CU_ASSERT(rdma_req.req.iov[1].iov_base == (void *)((unsigned long)0x2000));
981 	CU_ASSERT(rdma_req.req.iov[1].iov_len == md_size);
982 
983 	/* Part 4: simple I/O, one SGL equal io unit size, io_unit_size is aligned with md_size,
984 	   block size 512 */
985 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
986 	reset_nvmf_rdma_request(&rdma_req);
987 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
988 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
989 			  0, 0, 0, 0, 0);
990 	rdma_req.req.dif.dif_insert_or_strip = true;
991 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
992 	sgl->keyed.length = data_bs * 4;
993 
994 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
995 
996 	CU_ASSERT(rc == 0);
997 	CU_ASSERT(rdma_req.req.data_from_pool == true);
998 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
999 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1000 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1001 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1002 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
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.buffers[0] == 0x2000);
1006 
1007 	for (i = 0; i < 4; ++i) {
1008 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1009 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1010 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
1011 	}
1012 
1013 	/* Part 5: simple I/O, one SGL equal to 2x io unit size, io_unit_size is aligned with md_size,
1014 	   block size 512 */
1015 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1016 	reset_nvmf_rdma_request(&rdma_req);
1017 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1018 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1019 			  0, 0, 0, 0, 0);
1020 	rdma_req.req.dif.dif_insert_or_strip = true;
1021 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 2;
1022 	sgl->keyed.length = data_bs * 4;
1023 
1024 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1025 
1026 	CU_ASSERT(rc == 0);
1027 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1028 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1029 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1030 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1031 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1032 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1033 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1034 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1035 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1036 
1037 	for (i = 0; i < 2; ++i) {
1038 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1039 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1040 	}
1041 	for (i = 0; i < 2; ++i) {
1042 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].addr == 0x2000 + i * (data_bs + md_size));
1043 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].length == data_bs);
1044 	}
1045 
1046 	/* Part 6: simple I/O, one SGL larger than the transport io unit size, io_unit_size is not aligned to md_size,
1047 	   block size 512 */
1048 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1049 	reset_nvmf_rdma_request(&rdma_req);
1050 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1051 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1052 			  0, 0, 0, 0, 0);
1053 	rdma_req.req.dif.dif_insert_or_strip = true;
1054 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1055 	sgl->keyed.length = data_bs * 6;
1056 
1057 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1058 
1059 	CU_ASSERT(rc == 0);
1060 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1061 	CU_ASSERT(rdma_req.req.length == data_bs * 6);
1062 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1063 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 6);
1064 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1065 	CU_ASSERT(rdma_req.data.wr.num_sge == 7);
1066 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1067 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1068 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1069 
1070 	for (i = 0; i < 3; ++i) {
1071 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1072 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1073 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
1074 	}
1075 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
1076 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
1077 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == g_rdma_mr.lkey);
1078 
1079 	/* 2nd IO buffer consumed */
1080 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
1081 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
1082 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == g_rdma_mr.lkey);
1083 
1084 	CU_ASSERT(rdma_req.data.wr.sg_list[5].addr == 0x2000 + 24 + md_size);
1085 	CU_ASSERT(rdma_req.data.wr.sg_list[5].length == 512);
1086 	CU_ASSERT(rdma_req.data.wr.sg_list[5].lkey == g_rdma_mr.lkey);
1087 
1088 	CU_ASSERT(rdma_req.data.wr.sg_list[6].addr == 0x2000 + 24 + 512 + md_size * 2);
1089 	CU_ASSERT(rdma_req.data.wr.sg_list[6].length == 512);
1090 	CU_ASSERT(rdma_req.data.wr.sg_list[6].lkey == g_rdma_mr.lkey);
1091 
1092 	/* Part 7: simple I/O, number of SGL entries exceeds the number of entries
1093 	   one WR can hold. Additional WR is chained */
1094 	MOCK_SET(spdk_mempool_get, &data);
1095 	aligned_buffer = (void *)((uintptr_t)((char *)&data + NVMF_DATA_BUFFER_MASK) &
1096 				  ~NVMF_DATA_BUFFER_MASK);
1097 	reset_nvmf_rdma_request(&rdma_req);
1098 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1099 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1100 			  0, 0, 0, 0, 0);
1101 	rdma_req.req.dif.dif_insert_or_strip = true;
1102 	rtransport.transport.opts.io_unit_size = data_bs * 16;
1103 	sgl->keyed.length = data_bs * 16;
1104 
1105 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1106 
1107 	CU_ASSERT(rc == 0);
1108 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1109 	CU_ASSERT(rdma_req.req.length == data_bs * 16);
1110 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1111 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1112 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 16);
1113 	CU_ASSERT(rdma_req.req.data == aligned_buffer);
1114 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
1115 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1116 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1117 	/* additional wr from pool */
1118 	CU_ASSERT(rdma_req.data.wr.next == (void *)&data.wr);
1119 	CU_ASSERT(rdma_req.data.wr.next->num_sge == 1);
1120 	CU_ASSERT(rdma_req.data.wr.next->next == &rdma_req.rsp.wr);
1121 
1122 	/* Part 8: simple I/O, data with metadata do not fit to 1 io_buffer */
1123 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1124 	reset_nvmf_rdma_request(&rdma_req);
1125 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1126 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1127 			  0, 0, 0, 0, 0);
1128 	rdma_req.req.dif.dif_insert_or_strip = true;
1129 	rtransport.transport.opts.io_unit_size = 516;
1130 	sgl->keyed.length = data_bs * 2;
1131 
1132 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1133 
1134 	CU_ASSERT(rc == 0);
1135 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1136 	CU_ASSERT(rdma_req.req.length == data_bs * 2);
1137 	CU_ASSERT(rdma_req.req.iovcnt == 3);
1138 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1139 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 2);
1140 	CU_ASSERT(rdma_req.req.data == (void *)0x2000);
1141 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1142 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1143 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1144 
1145 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1146 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 512);
1147 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
1148 
1149 	/* 2nd IO buffer consumed, offset 4 bytes due to part of the metadata
1150 	  is located at the beginning of that buffer */
1151 	CU_ASSERT(rdma_req.data.wr.sg_list[1].addr == 0x2000 + 4);
1152 	CU_ASSERT(rdma_req.data.wr.sg_list[1].length == 512);
1153 	CU_ASSERT(rdma_req.data.wr.sg_list[1].lkey == g_rdma_mr.lkey);
1154 
1155 	/* Test 9 dealing with a buffer split over two Memory Regions */
1156 	MOCK_SET(spdk_mempool_get, (void *)&buffer);
1157 	reset_nvmf_rdma_request(&rdma_req);
1158 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1159 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1160 			  0, 0, 0, 0, 0);
1161 	rdma_req.req.dif.dif_insert_or_strip = true;
1162 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1163 	sgl->keyed.length = data_bs * 2;
1164 	g_mr_size = data_bs;
1165 	g_mr_next_size = rtransport.transport.opts.io_unit_size;
1166 
1167 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1168 	SPDK_CU_ASSERT_FATAL(rc == 0);
1169 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1170 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
1171 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&buffer + NVMF_DATA_BUFFER_MASK) &
1172 			~NVMF_DATA_BUFFER_MASK));
1173 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1174 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1175 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1176 	CU_ASSERT(rdma_req.req.buffers[0] == &buffer);
1177 	for (i = 0; i < 2; i++) {
1178 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uint64_t)rdma_req.req.data + i *
1179 			  (data_bs + md_size));
1180 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1181 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
1182 	}
1183 	buffer_ptr = STAILQ_FIRST(&group.retired_bufs);
1184 	CU_ASSERT(buffer_ptr == &buffer);
1185 	STAILQ_REMOVE(&group.retired_bufs, buffer_ptr, spdk_nvmf_transport_pg_cache_buf, link);
1186 	CU_ASSERT(STAILQ_EMPTY(&group.retired_bufs));
1187 	g_mr_size = 0;
1188 	g_mr_next_size = 0;
1189 
1190 	/* Test 2: Multi SGL */
1191 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1192 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1193 	sgl->address = 0;
1194 	rdma_req.recv->buf = (void *)&sgl_desc;
1195 	MOCK_SET(spdk_mempool_get, &data);
1196 	aligned_buffer = (void *)((uintptr_t)((char *)&data + NVMF_DATA_BUFFER_MASK) &
1197 				  ~NVMF_DATA_BUFFER_MASK);
1198 
1199 	/* part 1: 2 segments each with 1 wr. io_unit_size is aligned with data_bs + md_size */
1200 	reset_nvmf_rdma_request(&rdma_req);
1201 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1202 			  SPDK_DIF_TYPE1,
1203 			  SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 0, 0, 0, 0, 0);
1204 	rdma_req.req.dif.dif_insert_or_strip = true;
1205 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1206 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
1207 
1208 	for (i = 0; i < 2; i++) {
1209 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1210 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1211 		sgl_desc[i].keyed.length = data_bs * 4;
1212 		sgl_desc[i].address = 0x4000 + i * data_bs * 4;
1213 		sgl_desc[i].keyed.key = 0x44;
1214 	}
1215 
1216 	rc = spdk_nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1217 
1218 	CU_ASSERT(rc == 0);
1219 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1220 	CU_ASSERT(rdma_req.req.length == data_bs * 4 * 2);
1221 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1222 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4 * 2);
1223 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1224 	for (i = 0; i < 4; ++i) {
1225 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)((unsigned char *)aligned_buffer) + i *
1226 			  (data_bs + md_size));
1227 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1228 	}
1229 
1230 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
1231 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
1232 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
1233 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
1234 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + data_bs * 4);
1235 	CU_ASSERT(data.wr.num_sge == 4);
1236 	for (i = 0; i < 4; ++i) {
1237 		CU_ASSERT(data.wr.sg_list[i].addr == (uintptr_t)((unsigned char *)aligned_buffer) + i *
1238 			  (data_bs + md_size));
1239 		CU_ASSERT(data.wr.sg_list[i].length == data_bs);
1240 	}
1241 
1242 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
1243 }
1244 
1245 int main(int argc, char **argv)
1246 {
1247 	CU_pSuite	suite = NULL;
1248 	unsigned int	num_failures;
1249 
1250 	if (CU_initialize_registry() != CUE_SUCCESS) {
1251 		return CU_get_error();
1252 	}
1253 
1254 	suite = CU_add_suite("nvmf", NULL, NULL);
1255 	if (suite == NULL) {
1256 		CU_cleanup_registry();
1257 		return CU_get_error();
1258 	}
1259 
1260 	if (!CU_add_test(suite, "test_parse_sgl", test_spdk_nvmf_rdma_request_parse_sgl) ||
1261 	    !CU_add_test(suite, "test_request_process", test_spdk_nvmf_rdma_request_process) ||
1262 	    !CU_add_test(suite, "test_optimal_pg", test_spdk_nvmf_rdma_get_optimal_poll_group) ||
1263 	    !CU_add_test(suite, "test_parse_sgl_with_md", test_spdk_nvmf_rdma_request_parse_sgl_with_md)) {
1264 		CU_cleanup_registry();
1265 		return CU_get_error();
1266 	}
1267 
1268 	CU_basic_set_mode(CU_BRM_VERBOSE);
1269 	CU_basic_run_tests();
1270 	num_failures = CU_get_number_of_failures();
1271 	CU_cleanup_registry();
1272 	return num_failures;
1273 }
1274