xref: /spdk/test/unit/lib/nvmf/rdma.c/rdma_ut.c (revision 024179d8eeab28c311bf40a9de14bf681e87af18)
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 "common/lib/test_rdma.c"
38 #include "nvmf/rdma.c"
39 #include "nvmf/transport.c"
40 
41 uint64_t g_mr_size;
42 uint64_t g_mr_next_size;
43 struct ibv_mr g_rdma_mr;
44 
45 #define RDMA_UT_UNITS_IN_MAX_IO 16
46 
47 struct spdk_nvmf_transport_opts g_rdma_ut_transport_opts = {
48 	.max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH,
49 	.max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR,
50 	.in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE,
51 	.max_io_size = (SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE * RDMA_UT_UNITS_IN_MAX_IO),
52 	.io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE,
53 	.max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH,
54 	.num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS,
55 };
56 
57 SPDK_LOG_REGISTER_COMPONENT(nvmf)
58 DEFINE_STUB(spdk_mem_map_set_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
59 		uint64_t size, uint64_t translation), 0);
60 DEFINE_STUB(spdk_mem_map_clear_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
61 		uint64_t size), 0);
62 DEFINE_STUB(spdk_mem_map_alloc, struct spdk_mem_map *, (uint64_t default_translation,
63 		const struct spdk_mem_map_ops *ops, void *cb_ctx), NULL);
64 DEFINE_STUB(spdk_nvmf_qpair_disconnect, int, (struct spdk_nvmf_qpair *qpair,
65 		nvmf_qpair_disconnect_cb cb_fn, void *ctx), 0);
66 DEFINE_STUB(spdk_nvmf_qpair_get_listen_trid, int,
67 	    (struct spdk_nvmf_qpair *qpair, struct spdk_nvme_transport_id *trid), 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_ctrlr_data_init, (struct spdk_nvmf_transport_opts *opts,
80 		struct spdk_nvmf_ctrlr_data *cdata));
81 DEFINE_STUB_V(spdk_nvmf_request_exec, (struct spdk_nvmf_request *req));
82 DEFINE_STUB(spdk_nvmf_request_complete, int, (struct spdk_nvmf_request *req), 0);
83 DEFINE_STUB(spdk_nvme_transport_id_compare, int, (const struct spdk_nvme_transport_id *trid1,
84 		const struct spdk_nvme_transport_id *trid2), 0);
85 DEFINE_STUB_V(nvmf_ctrlr_abort_aer, (struct spdk_nvmf_ctrlr *ctrlr));
86 DEFINE_STUB(spdk_nvmf_request_get_dif_ctx, bool, (struct spdk_nvmf_request *req,
87 		struct spdk_dif_ctx *dif_ctx), false);
88 DEFINE_STUB_V(spdk_nvme_trid_populate_transport, (struct spdk_nvme_transport_id *trid,
89 		enum spdk_nvme_transport_type trtype));
90 DEFINE_STUB_V(spdk_nvmf_tgt_new_qpair, (struct spdk_nvmf_tgt *tgt, struct spdk_nvmf_qpair *qpair));
91 DEFINE_STUB(nvmf_ctrlr_abort_request, int, (struct spdk_nvmf_request *req), 0);
92 
93 const char *
94 spdk_nvme_transport_id_trtype_str(enum spdk_nvme_transport_type trtype)
95 {
96 	switch (trtype) {
97 	case SPDK_NVME_TRANSPORT_PCIE:
98 		return "PCIe";
99 	case SPDK_NVME_TRANSPORT_RDMA:
100 		return "RDMA";
101 	case SPDK_NVME_TRANSPORT_FC:
102 		return "FC";
103 	default:
104 		return NULL;
105 	}
106 }
107 
108 int
109 spdk_nvme_transport_id_populate_trstring(struct spdk_nvme_transport_id *trid, const char *trstring)
110 {
111 	int len, i;
112 
113 	if (trstring == NULL) {
114 		return -EINVAL;
115 	}
116 
117 	len = strnlen(trstring, SPDK_NVMF_TRSTRING_MAX_LEN);
118 	if (len == SPDK_NVMF_TRSTRING_MAX_LEN) {
119 		return -EINVAL;
120 	}
121 
122 	/* cast official trstring to uppercase version of input. */
123 	for (i = 0; i < len; i++) {
124 		trid->trstring[i] = toupper(trstring[i]);
125 	}
126 	return 0;
127 }
128 
129 uint64_t
130 spdk_mem_map_translate(const struct spdk_mem_map *map, uint64_t vaddr, uint64_t *size)
131 {
132 	if (g_mr_size != 0) {
133 		*(uint32_t *)size = g_mr_size;
134 		if (g_mr_next_size != 0) {
135 			g_mr_size = g_mr_next_size;
136 		}
137 	}
138 
139 	return (uint64_t)&g_rdma_mr;
140 }
141 
142 static void reset_nvmf_rdma_request(struct spdk_nvmf_rdma_request *rdma_req)
143 {
144 	int i;
145 
146 	rdma_req->req.length = 0;
147 	rdma_req->req.data_from_pool = false;
148 	rdma_req->req.data = NULL;
149 	rdma_req->data.wr.num_sge = 0;
150 	rdma_req->data.wr.wr.rdma.remote_addr = 0;
151 	rdma_req->data.wr.wr.rdma.rkey = 0;
152 	memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
153 
154 	for (i = 0; i < SPDK_NVMF_MAX_SGL_ENTRIES; i++) {
155 		rdma_req->req.iov[i].iov_base = 0;
156 		rdma_req->req.iov[i].iov_len = 0;
157 		rdma_req->req.buffers[i] = 0;
158 		rdma_req->data.wr.sg_list[i].addr = 0;
159 		rdma_req->data.wr.sg_list[i].length = 0;
160 		rdma_req->data.wr.sg_list[i].lkey = 0;
161 	}
162 	rdma_req->req.iovcnt = 0;
163 }
164 
165 static void
166 test_spdk_nvmf_rdma_request_parse_sgl(void)
167 {
168 	struct spdk_nvmf_rdma_transport rtransport;
169 	struct spdk_nvmf_rdma_device device;
170 	struct spdk_nvmf_rdma_request rdma_req = {};
171 	struct spdk_nvmf_rdma_recv recv;
172 	struct spdk_nvmf_rdma_poll_group group;
173 	struct spdk_nvmf_rdma_qpair rqpair;
174 	struct spdk_nvmf_rdma_poller poller;
175 	union nvmf_c2h_msg cpl;
176 	union nvmf_h2c_msg cmd;
177 	struct spdk_nvme_sgl_descriptor *sgl;
178 	struct spdk_nvmf_transport_pg_cache_buf bufs[4];
179 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
180 	struct spdk_nvmf_rdma_request_data data;
181 	struct spdk_nvmf_transport_pg_cache_buf	buffer;
182 	struct spdk_nvmf_transport_pg_cache_buf	*buffer_ptr;
183 	int rc, i;
184 
185 	data.wr.sg_list = data.sgl;
186 	STAILQ_INIT(&group.group.buf_cache);
187 	group.group.buf_cache_size = 0;
188 	group.group.buf_cache_count = 0;
189 	group.group.transport = &rtransport.transport;
190 	STAILQ_INIT(&group.retired_bufs);
191 	poller.group = &group;
192 	rqpair.poller = &poller;
193 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
194 
195 	sgl = &cmd.nvme_cmd.dptr.sgl1;
196 	rdma_req.recv = &recv;
197 	rdma_req.req.cmd = &cmd;
198 	rdma_req.req.rsp = &cpl;
199 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
200 	rdma_req.req.qpair = &rqpair.qpair;
201 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
202 
203 	rtransport.transport.opts = g_rdma_ut_transport_opts;
204 	rtransport.data_wr_pool = NULL;
205 	rtransport.transport.data_buf_pool = NULL;
206 
207 	device.attr.device_cap_flags = 0;
208 	g_rdma_mr.lkey = 0xABCD;
209 	sgl->keyed.key = 0xEEEE;
210 	sgl->address = 0xFFFF;
211 	rdma_req.recv->buf = (void *)0xDDDD;
212 
213 	/* Test 1: sgl type: keyed data block subtype: address */
214 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
215 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
216 
217 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size */
218 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
219 	reset_nvmf_rdma_request(&rdma_req);
220 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
221 
222 	device.map = (void *)0x0;
223 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
224 	CU_ASSERT(rc == 0);
225 	CU_ASSERT(rdma_req.req.data_from_pool == true);
226 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
227 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
228 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
229 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
230 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
231 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
232 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
233 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
234 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
235 
236 	/* Part 2: simple I/O, one SGL larger than the transport io unit size (equal to the max io size) */
237 	reset_nvmf_rdma_request(&rdma_req);
238 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
239 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
240 
241 	CU_ASSERT(rc == 0);
242 	CU_ASSERT(rdma_req.req.data_from_pool == true);
243 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO);
244 	CU_ASSERT(rdma_req.data.wr.num_sge == RDMA_UT_UNITS_IN_MAX_IO);
245 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
246 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
247 	for (i = 0; i < RDMA_UT_UNITS_IN_MAX_IO; i++) {
248 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
249 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
250 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
251 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
252 	}
253 
254 	/* Part 3: simple I/O one SGL larger than the transport max io size */
255 	reset_nvmf_rdma_request(&rdma_req);
256 	sgl->keyed.length = rtransport.transport.opts.max_io_size * 2;
257 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
258 
259 	CU_ASSERT(rc == -1);
260 
261 	/* Part 4: Pretend there are no buffer pools */
262 	MOCK_SET(spdk_mempool_get, NULL);
263 	reset_nvmf_rdma_request(&rdma_req);
264 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
265 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
266 
267 	CU_ASSERT(rc == 0);
268 	CU_ASSERT(rdma_req.req.data_from_pool == false);
269 	CU_ASSERT(rdma_req.req.data == NULL);
270 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
271 	CU_ASSERT(rdma_req.req.buffers[0] == NULL);
272 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0);
273 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 0);
274 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == 0);
275 
276 	rdma_req.recv->buf = (void *)0xDDDD;
277 	/* Test 2: sgl type: keyed data block subtype: offset (in capsule data) */
278 	sgl->generic.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
279 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
280 
281 	/* Part 1: Normal I/O smaller than in capsule data size no offset */
282 	reset_nvmf_rdma_request(&rdma_req);
283 	sgl->address = 0;
284 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
285 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
286 
287 	CU_ASSERT(rc == 0);
288 	CU_ASSERT(rdma_req.req.data == (void *)0xDDDD);
289 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.in_capsule_data_size);
290 	CU_ASSERT(rdma_req.req.data_from_pool == false);
291 
292 	/* Part 2: I/O offset + length too large */
293 	reset_nvmf_rdma_request(&rdma_req);
294 	sgl->address = rtransport.transport.opts.in_capsule_data_size;
295 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
296 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
297 
298 	CU_ASSERT(rc == -1);
299 
300 	/* Part 3: I/O too large */
301 	reset_nvmf_rdma_request(&rdma_req);
302 	sgl->address = 0;
303 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size * 2;
304 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
305 
306 	CU_ASSERT(rc == -1);
307 
308 	/* Test 3: Multi SGL */
309 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
310 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
311 	sgl->address = 0;
312 	rdma_req.recv->buf = (void *)&sgl_desc;
313 	MOCK_SET(spdk_mempool_get, &data);
314 
315 	/* part 1: 2 segments each with 1 wr. */
316 	reset_nvmf_rdma_request(&rdma_req);
317 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
318 	for (i = 0; i < 2; i++) {
319 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
320 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
321 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size;
322 		sgl_desc[i].address = 0x4000 + i * rtransport.transport.opts.io_unit_size;
323 		sgl_desc[i].keyed.key = 0x44;
324 	}
325 
326 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
327 
328 	CU_ASSERT(rc == 0);
329 	CU_ASSERT(rdma_req.req.data_from_pool == true);
330 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 2);
331 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
332 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
333 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
334 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
335 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
336 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size);
337 	CU_ASSERT(data.wr.num_sge == 1);
338 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
339 
340 	/* part 2: 2 segments, each with 1 wr containing 8 sge_elements */
341 	reset_nvmf_rdma_request(&rdma_req);
342 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
343 	for (i = 0; i < 2; i++) {
344 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
345 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
346 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size * 8;
347 		sgl_desc[i].address = 0x4000 + i * 8 * rtransport.transport.opts.io_unit_size;
348 		sgl_desc[i].keyed.key = 0x44;
349 	}
350 
351 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
352 
353 	CU_ASSERT(rc == 0);
354 	CU_ASSERT(rdma_req.req.data_from_pool == true);
355 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
356 	CU_ASSERT(rdma_req.req.iovcnt == 16);
357 	CU_ASSERT(rdma_req.data.wr.num_sge == 8);
358 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
359 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
360 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
361 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
362 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 8);
363 	CU_ASSERT(data.wr.num_sge == 8);
364 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
365 
366 	/* part 3: 2 segments, one very large, one very small */
367 	reset_nvmf_rdma_request(&rdma_req);
368 	for (i = 0; i < 2; i++) {
369 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
370 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
371 		sgl_desc[i].keyed.key = 0x44;
372 	}
373 
374 	sgl_desc[0].keyed.length = rtransport.transport.opts.io_unit_size * 15 +
375 				   rtransport.transport.opts.io_unit_size / 2;
376 	sgl_desc[0].address = 0x4000;
377 	sgl_desc[1].keyed.length = rtransport.transport.opts.io_unit_size / 2;
378 	sgl_desc[1].address = 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
379 			      rtransport.transport.opts.io_unit_size / 2;
380 
381 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
382 
383 	CU_ASSERT(rc == 0);
384 	CU_ASSERT(rdma_req.req.data_from_pool == true);
385 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
386 	CU_ASSERT(rdma_req.req.iovcnt == 17);
387 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
388 	for (i = 0; i < 15; i++) {
389 		CU_ASSERT(rdma_req.data.sgl[i].length == rtransport.transport.opts.io_unit_size);
390 	}
391 	CU_ASSERT(rdma_req.data.sgl[15].length == rtransport.transport.opts.io_unit_size / 2);
392 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
393 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
394 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
395 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
396 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
397 		  rtransport.transport.opts.io_unit_size / 2);
398 	CU_ASSERT(data.sgl[0].length == rtransport.transport.opts.io_unit_size / 2);
399 	CU_ASSERT(data.wr.num_sge == 1);
400 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
401 
402 	/* Test 4: use PG buffer cache */
403 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
404 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
405 	sgl->address = 0xFFFF;
406 	rdma_req.recv->buf = (void *)0xDDDD;
407 	g_rdma_mr.lkey = 0xABCD;
408 	sgl->keyed.key = 0xEEEE;
409 
410 	for (i = 0; i < 4; i++) {
411 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
412 	}
413 
414 	/* part 1: use the four buffers from the pg cache */
415 	group.group.buf_cache_size = 4;
416 	group.group.buf_cache_count = 4;
417 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
418 	reset_nvmf_rdma_request(&rdma_req);
419 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * 4;
420 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
421 
422 	SPDK_CU_ASSERT_FATAL(rc == 0);
423 	CU_ASSERT(rdma_req.req.data_from_pool == true);
424 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
425 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
426 			~NVMF_DATA_BUFFER_MASK));
427 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
428 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
429 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
430 	CU_ASSERT(group.group.buf_cache_count == 0);
431 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
432 	for (i = 0; i < 4; i++) {
433 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
434 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
435 				~NVMF_DATA_BUFFER_MASK));
436 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
437 	}
438 
439 	/* part 2: now that we have used the buffers from the cache, try again. We should get mempool buffers. */
440 	reset_nvmf_rdma_request(&rdma_req);
441 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
442 
443 	SPDK_CU_ASSERT_FATAL(rc == 0);
444 	CU_ASSERT(rdma_req.req.data_from_pool == true);
445 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
446 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
447 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
448 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
449 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
450 	CU_ASSERT(group.group.buf_cache_count == 0);
451 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
452 	for (i = 0; i < 4; i++) {
453 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
454 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
455 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
456 		CU_ASSERT(group.group.buf_cache_count == 0);
457 	}
458 
459 	/* part 3: half and half */
460 	group.group.buf_cache_count = 2;
461 
462 	for (i = 0; i < 2; i++) {
463 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
464 	}
465 	reset_nvmf_rdma_request(&rdma_req);
466 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
467 
468 	SPDK_CU_ASSERT_FATAL(rc == 0);
469 	CU_ASSERT(rdma_req.req.data_from_pool == true);
470 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
471 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
472 			~NVMF_DATA_BUFFER_MASK));
473 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
474 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
475 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
476 	CU_ASSERT(group.group.buf_cache_count == 0);
477 	for (i = 0; i < 2; i++) {
478 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
479 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
480 				~NVMF_DATA_BUFFER_MASK));
481 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
482 	}
483 	for (i = 2; i < 4; i++) {
484 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
485 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
486 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
487 	}
488 
489 	reset_nvmf_rdma_request(&rdma_req);
490 	/* Test 5 dealing with a buffer split over two Memory Regions */
491 	MOCK_SET(spdk_mempool_get, (void *)&buffer);
492 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
493 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
494 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
495 	g_mr_size = rtransport.transport.opts.io_unit_size / 4;
496 	g_mr_next_size = rtransport.transport.opts.io_unit_size / 2;
497 
498 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
499 	SPDK_CU_ASSERT_FATAL(rc == 0);
500 	CU_ASSERT(rdma_req.req.data_from_pool == true);
501 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
502 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&buffer + NVMF_DATA_BUFFER_MASK) &
503 			~NVMF_DATA_BUFFER_MASK));
504 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
505 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
506 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
507 	CU_ASSERT(rdma_req.req.buffers[0] == &buffer);
508 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == (((uint64_t)&buffer + NVMF_DATA_BUFFER_MASK) &
509 			~NVMF_DATA_BUFFER_MASK));
510 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
511 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
512 	buffer_ptr = STAILQ_FIRST(&group.retired_bufs);
513 	CU_ASSERT(buffer_ptr == &buffer);
514 	STAILQ_REMOVE(&group.retired_bufs, buffer_ptr, spdk_nvmf_transport_pg_cache_buf, link);
515 	CU_ASSERT(STAILQ_EMPTY(&group.retired_bufs));
516 	g_mr_size = 0;
517 	g_mr_next_size = 0;
518 
519 	reset_nvmf_rdma_request(&rdma_req);
520 }
521 
522 static struct spdk_nvmf_rdma_recv *
523 create_recv(struct spdk_nvmf_rdma_qpair *rqpair, enum spdk_nvme_nvm_opcode opc)
524 {
525 	struct spdk_nvmf_rdma_recv *rdma_recv;
526 	union nvmf_h2c_msg *cmd;
527 	struct spdk_nvme_sgl_descriptor *sgl;
528 
529 	rdma_recv = calloc(1, sizeof(*rdma_recv));
530 	rdma_recv->qpair = rqpair;
531 	cmd = calloc(1, sizeof(*cmd));
532 	rdma_recv->sgl[0].addr = (uintptr_t)cmd;
533 	cmd->nvme_cmd.opc = opc;
534 	sgl = &cmd->nvme_cmd.dptr.sgl1;
535 	sgl->keyed.key = 0xEEEE;
536 	sgl->address = 0xFFFF;
537 	sgl->keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
538 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
539 	sgl->keyed.length = 1;
540 
541 	return rdma_recv;
542 }
543 
544 static void
545 free_recv(struct spdk_nvmf_rdma_recv *rdma_recv)
546 {
547 	free((void *)rdma_recv->sgl[0].addr);
548 	free(rdma_recv);
549 }
550 
551 static struct spdk_nvmf_rdma_request *
552 create_req(struct spdk_nvmf_rdma_qpair *rqpair,
553 	   struct spdk_nvmf_rdma_recv *rdma_recv)
554 {
555 	struct spdk_nvmf_rdma_request *rdma_req;
556 	union nvmf_c2h_msg *cpl;
557 
558 	rdma_req = calloc(1, sizeof(*rdma_req));
559 	rdma_req->recv = rdma_recv;
560 	rdma_req->req.qpair = &rqpair->qpair;
561 	rdma_req->state = RDMA_REQUEST_STATE_NEW;
562 	rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr;
563 	rdma_req->data.wr.sg_list = rdma_req->data.sgl;
564 	cpl = calloc(1, sizeof(*cpl));
565 	rdma_req->rsp.sgl[0].addr = (uintptr_t)cpl;
566 	rdma_req->req.rsp = cpl;
567 
568 	return rdma_req;
569 }
570 
571 static void
572 free_req(struct spdk_nvmf_rdma_request *rdma_req)
573 {
574 	free((void *)rdma_req->rsp.sgl[0].addr);
575 	free(rdma_req);
576 }
577 
578 static void
579 qpair_reset(struct spdk_nvmf_rdma_qpair *rqpair,
580 	    struct spdk_nvmf_rdma_poller *poller,
581 	    struct spdk_nvmf_rdma_device *device,
582 	    struct spdk_nvmf_rdma_resources *resources,
583 	    struct spdk_nvmf_transport *transport)
584 {
585 	memset(rqpair, 0, sizeof(*rqpair));
586 	STAILQ_INIT(&rqpair->pending_rdma_write_queue);
587 	STAILQ_INIT(&rqpair->pending_rdma_read_queue);
588 	rqpair->poller = poller;
589 	rqpair->device = device;
590 	rqpair->resources = resources;
591 	rqpair->qpair.qid = 1;
592 	rqpair->ibv_state = IBV_QPS_RTS;
593 	rqpair->qpair.state = SPDK_NVMF_QPAIR_ACTIVE;
594 	rqpair->max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
595 	rqpair->max_send_depth = 16;
596 	rqpair->max_read_depth = 16;
597 	rqpair->qpair.transport = transport;
598 	resources->recvs_to_post.first = resources->recvs_to_post.last = NULL;
599 }
600 
601 static void
602 poller_reset(struct spdk_nvmf_rdma_poller *poller,
603 	     struct spdk_nvmf_rdma_poll_group *group)
604 {
605 	memset(poller, 0, sizeof(*poller));
606 	STAILQ_INIT(&poller->qpairs_pending_recv);
607 	STAILQ_INIT(&poller->qpairs_pending_send);
608 	poller->group = group;
609 }
610 
611 static void
612 test_spdk_nvmf_rdma_request_process(void)
613 {
614 	struct spdk_nvmf_rdma_transport rtransport = {};
615 	struct spdk_nvmf_rdma_poll_group group = {};
616 	struct spdk_nvmf_rdma_poller poller = {};
617 	struct spdk_nvmf_rdma_device device = {};
618 	struct spdk_nvmf_rdma_resources resources = {};
619 	struct spdk_nvmf_rdma_qpair rqpair = {};
620 	struct spdk_nvmf_rdma_recv *rdma_recv;
621 	struct spdk_nvmf_rdma_request *rdma_req;
622 	bool progress;
623 
624 	STAILQ_INIT(&group.group.buf_cache);
625 	STAILQ_INIT(&group.group.pending_buf_queue);
626 	group.group.buf_cache_size = 0;
627 	group.group.buf_cache_count = 0;
628 	poller_reset(&poller, &group);
629 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
630 
631 	rtransport.transport.opts = g_rdma_ut_transport_opts;
632 	rtransport.transport.data_buf_pool = spdk_mempool_create("test_data_pool", 16, 128, 0, 0);
633 	rtransport.data_wr_pool = spdk_mempool_create("test_wr_pool", 128,
634 				  sizeof(struct spdk_nvmf_rdma_request_data),
635 				  0, 0);
636 	MOCK_CLEAR(spdk_mempool_get);
637 
638 	device.attr.device_cap_flags = 0;
639 	device.map = (void *)0x0;
640 	g_rdma_mr.lkey = 0xABCD;
641 
642 	/* Test 1: single SGL READ request */
643 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_READ);
644 	rdma_req = create_req(&rqpair, rdma_recv);
645 	rqpair.current_recv_depth = 1;
646 	/* NEW -> EXECUTING */
647 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
648 	CU_ASSERT(progress == true);
649 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
650 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST);
651 	/* EXECUTED -> TRANSFERRING_C2H */
652 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
653 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
654 	CU_ASSERT(progress == true);
655 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST);
656 	CU_ASSERT(rdma_req->recv == NULL);
657 	CU_ASSERT(resources.recvs_to_post.first == &rdma_recv->wr);
658 	CU_ASSERT(resources.recvs_to_post.last == &rdma_recv->wr);
659 	/* COMPLETED -> FREE */
660 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
661 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
662 	CU_ASSERT(progress == true);
663 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
664 
665 	free_recv(rdma_recv);
666 	free_req(rdma_req);
667 	poller_reset(&poller, &group);
668 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
669 
670 	/* Test 2: single SGL WRITE request */
671 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
672 	rdma_req = create_req(&rqpair, rdma_recv);
673 	rqpair.current_recv_depth = 1;
674 	/* NEW -> TRANSFERRING_H2C */
675 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
676 	CU_ASSERT(progress == true);
677 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
678 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
679 	STAILQ_INIT(&poller.qpairs_pending_send);
680 	/* READY_TO_EXECUTE -> EXECUTING */
681 	rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
682 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
683 	CU_ASSERT(progress == true);
684 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
685 	/* EXECUTED -> COMPLETING */
686 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
687 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
688 	CU_ASSERT(progress == true);
689 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
690 	CU_ASSERT(rdma_req->recv == NULL);
691 	CU_ASSERT(resources.recvs_to_post.first == &rdma_recv->wr);
692 	CU_ASSERT(resources.recvs_to_post.last == &rdma_recv->wr);
693 	/* COMPLETED -> FREE */
694 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
695 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
696 	CU_ASSERT(progress == true);
697 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
698 
699 	free_recv(rdma_recv);
700 	free_req(rdma_req);
701 	poller_reset(&poller, &group);
702 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
703 
704 	/* Test 3: WRITE+WRITE ibv_send batching */
705 	{
706 		struct spdk_nvmf_rdma_recv *recv1, *recv2;
707 		struct spdk_nvmf_rdma_request *req1, *req2;
708 		recv1 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
709 		req1 = create_req(&rqpair, recv1);
710 		recv2 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
711 		req2 = create_req(&rqpair, recv2);
712 
713 		/* WRITE 1: NEW -> TRANSFERRING_H2C */
714 		rqpair.current_recv_depth = 1;
715 		nvmf_rdma_request_process(&rtransport, req1);
716 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
717 
718 		/* WRITE 2: NEW -> TRANSFERRING_H2C */
719 		rqpair.current_recv_depth = 2;
720 		nvmf_rdma_request_process(&rtransport, req2);
721 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
722 
723 		STAILQ_INIT(&poller.qpairs_pending_send);
724 
725 		/* WRITE 1 completes before WRITE 2 has finished RDMA reading */
726 		/* WRITE 1: READY_TO_EXECUTE -> EXECUTING */
727 		req1->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
728 		nvmf_rdma_request_process(&rtransport, req1);
729 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_EXECUTING);
730 		/* WRITE 1: EXECUTED -> COMPLETING */
731 		req1->state = RDMA_REQUEST_STATE_EXECUTED;
732 		nvmf_rdma_request_process(&rtransport, req1);
733 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_COMPLETING);
734 		STAILQ_INIT(&poller.qpairs_pending_send);
735 		/* WRITE 1: COMPLETED -> FREE */
736 		req1->state = RDMA_REQUEST_STATE_COMPLETED;
737 		nvmf_rdma_request_process(&rtransport, req1);
738 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_FREE);
739 
740 		/* Now WRITE 2 has finished reading and completes */
741 		/* WRITE 2: COMPLETED -> FREE */
742 		/* WRITE 2: READY_TO_EXECUTE -> EXECUTING */
743 		req2->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
744 		nvmf_rdma_request_process(&rtransport, req2);
745 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_EXECUTING);
746 		/* WRITE 1: EXECUTED -> COMPLETING */
747 		req2->state = RDMA_REQUEST_STATE_EXECUTED;
748 		nvmf_rdma_request_process(&rtransport, req2);
749 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_COMPLETING);
750 		STAILQ_INIT(&poller.qpairs_pending_send);
751 		/* WRITE 1: COMPLETED -> FREE */
752 		req2->state = RDMA_REQUEST_STATE_COMPLETED;
753 		nvmf_rdma_request_process(&rtransport, req2);
754 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_FREE);
755 
756 		free_recv(recv1);
757 		free_req(req1);
758 		free_recv(recv2);
759 		free_req(req2);
760 		poller_reset(&poller, &group);
761 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
762 	}
763 
764 	/* Test 4, invalid command, check xfer type */
765 	{
766 		struct spdk_nvmf_rdma_recv *rdma_recv_inv;
767 		struct spdk_nvmf_rdma_request *rdma_req_inv;
768 		/* construct an opcode that specifies BIDIRECTIONAL transfer */
769 		uint8_t opc = 0x10 | SPDK_NVME_DATA_BIDIRECTIONAL;
770 
771 		rdma_recv_inv = create_recv(&rqpair, opc);
772 		rdma_req_inv = create_req(&rqpair, rdma_recv_inv);
773 
774 		/* NEW -> RDMA_REQUEST_STATE_COMPLETING */
775 		rqpair.current_recv_depth = 1;
776 		progress = nvmf_rdma_request_process(&rtransport, rdma_req_inv);
777 		CU_ASSERT(progress == true);
778 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_COMPLETING);
779 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sct == SPDK_NVME_SCT_GENERIC);
780 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sc == SPDK_NVME_SC_INVALID_OPCODE);
781 
782 		/* RDMA_REQUEST_STATE_COMPLETED -> FREE */
783 		rdma_req_inv->state = RDMA_REQUEST_STATE_COMPLETED;
784 		nvmf_rdma_request_process(&rtransport, rdma_req_inv);
785 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_FREE);
786 
787 		free_recv(rdma_recv_inv);
788 		free_req(rdma_req_inv);
789 		poller_reset(&poller, &group);
790 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
791 	}
792 
793 	spdk_mempool_free(rtransport.transport.data_buf_pool);
794 	spdk_mempool_free(rtransport.data_wr_pool);
795 }
796 
797 #define TEST_GROUPS_COUNT 5
798 static void
799 test_nvmf_rdma_get_optimal_poll_group(void)
800 {
801 	struct spdk_nvmf_rdma_transport rtransport = {};
802 	struct spdk_nvmf_transport *transport = &rtransport.transport;
803 	struct spdk_nvmf_rdma_qpair rqpair = {};
804 	struct spdk_nvmf_transport_poll_group *groups[TEST_GROUPS_COUNT];
805 	struct spdk_nvmf_rdma_poll_group *rgroups[TEST_GROUPS_COUNT];
806 	struct spdk_nvmf_transport_poll_group *result;
807 	uint32_t i;
808 
809 	rqpair.qpair.transport = transport;
810 	pthread_mutex_init(&rtransport.lock, NULL);
811 	TAILQ_INIT(&rtransport.poll_groups);
812 
813 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
814 		groups[i] = nvmf_rdma_poll_group_create(transport);
815 		CU_ASSERT(groups[i] != NULL);
816 		rgroups[i] = SPDK_CONTAINEROF(groups[i], struct spdk_nvmf_rdma_poll_group, group);
817 		groups[i]->transport = transport;
818 	}
819 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[0]);
820 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[0]);
821 
822 	/* Emulate connection of %TEST_GROUPS_COUNT% initiators - each creates 1 admin and 1 io qp */
823 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
824 		rqpair.qpair.qid = 0;
825 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
826 		CU_ASSERT(result == groups[i]);
827 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
828 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i]);
829 
830 		rqpair.qpair.qid = 1;
831 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
832 		CU_ASSERT(result == groups[i]);
833 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
834 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
835 	}
836 	/* wrap around, admin/io pg point to the first pg
837 	   Destroy all poll groups except of the last one */
838 	for (i = 0; i < TEST_GROUPS_COUNT - 1; i++) {
839 		nvmf_rdma_poll_group_destroy(groups[i]);
840 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[i + 1]);
841 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i + 1]);
842 	}
843 
844 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
845 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
846 
847 	/* Check that pointers to the next admin/io poll groups are not changed */
848 	rqpair.qpair.qid = 0;
849 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
850 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
851 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
852 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
853 
854 	rqpair.qpair.qid = 1;
855 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
856 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
857 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
858 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
859 
860 	/* Remove the last poll group, check that pointers are NULL */
861 	nvmf_rdma_poll_group_destroy(groups[TEST_GROUPS_COUNT - 1]);
862 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == NULL);
863 	CU_ASSERT(rtransport.conn_sched.next_io_pg == NULL);
864 
865 	/* Request optimal poll group, result must be NULL */
866 	rqpair.qpair.qid = 0;
867 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
868 	CU_ASSERT(result == NULL);
869 
870 	rqpair.qpair.qid = 1;
871 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
872 	CU_ASSERT(result == NULL);
873 
874 	pthread_mutex_destroy(&rtransport.lock);
875 }
876 #undef TEST_GROUPS_COUNT
877 
878 static void
879 test_spdk_nvmf_rdma_request_parse_sgl_with_md(void)
880 {
881 	struct spdk_nvmf_rdma_transport rtransport;
882 	struct spdk_nvmf_rdma_device device;
883 	struct spdk_nvmf_rdma_request rdma_req = {};
884 	struct spdk_nvmf_rdma_recv recv;
885 	struct spdk_nvmf_rdma_poll_group group;
886 	struct spdk_nvmf_rdma_qpair rqpair;
887 	struct spdk_nvmf_rdma_poller poller;
888 	union nvmf_c2h_msg cpl;
889 	union nvmf_h2c_msg cmd;
890 	struct spdk_nvme_sgl_descriptor *sgl;
891 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
892 	struct spdk_nvmf_rdma_request_data data;
893 	char data2_buffer[8192];
894 	struct spdk_nvmf_rdma_request_data *data2 = (struct spdk_nvmf_rdma_request_data *)data2_buffer;
895 	struct spdk_nvmf_transport_pg_cache_buf	buffer;
896 	struct spdk_nvmf_transport_pg_cache_buf	*buffer_ptr;
897 	const uint32_t data_bs = 512;
898 	const uint32_t md_size = 8;
899 	int rc, i;
900 	void *aligned_buffer;
901 
902 	data.wr.sg_list = data.sgl;
903 	STAILQ_INIT(&group.group.buf_cache);
904 	group.group.buf_cache_size = 0;
905 	group.group.buf_cache_count = 0;
906 	group.group.transport = &rtransport.transport;
907 	STAILQ_INIT(&group.retired_bufs);
908 	poller.group = &group;
909 	rqpair.poller = &poller;
910 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
911 
912 	sgl = &cmd.nvme_cmd.dptr.sgl1;
913 	rdma_req.recv = &recv;
914 	rdma_req.req.cmd = &cmd;
915 	rdma_req.req.rsp = &cpl;
916 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
917 	rdma_req.req.qpair = &rqpair.qpair;
918 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
919 
920 	rtransport.transport.opts = g_rdma_ut_transport_opts;
921 	rtransport.data_wr_pool = NULL;
922 	rtransport.transport.data_buf_pool = NULL;
923 
924 	device.attr.device_cap_flags = 0;
925 	device.map = NULL;
926 	g_rdma_mr.lkey = 0xABCD;
927 	sgl->keyed.key = 0xEEEE;
928 	sgl->address = 0xFFFF;
929 	rdma_req.recv->buf = (void *)0xDDDD;
930 
931 	/* Test 1: sgl type: keyed data block subtype: address */
932 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
933 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
934 
935 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size, block size 512 */
936 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
937 	reset_nvmf_rdma_request(&rdma_req);
938 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
939 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
940 			  0, 0, 0, 0, 0);
941 	rdma_req.req.dif.dif_insert_or_strip = true;
942 	rtransport.transport.opts.io_unit_size = data_bs * 8;
943 	sgl->keyed.length = data_bs * 4;
944 
945 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
946 
947 	CU_ASSERT(rc == 0);
948 	CU_ASSERT(rdma_req.req.data_from_pool == true);
949 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
950 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
951 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
952 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
953 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
954 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
955 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
956 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
957 
958 	for (i = 0; i < 4; ++i) {
959 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
960 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
961 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
962 	}
963 
964 	/* Part 2: simple I/O, one SGL equal to io unit size, io_unit_size is not aligned with md_size,
965 		block size 512 */
966 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
967 	reset_nvmf_rdma_request(&rdma_req);
968 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
969 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
970 			  0, 0, 0, 0, 0);
971 	rdma_req.req.dif.dif_insert_or_strip = true;
972 	rtransport.transport.opts.io_unit_size = data_bs * 4;
973 	sgl->keyed.length = data_bs * 4;
974 
975 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
976 
977 	CU_ASSERT(rc == 0);
978 	CU_ASSERT(rdma_req.req.data_from_pool == true);
979 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
980 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
981 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
982 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
983 	CU_ASSERT(rdma_req.data.wr.num_sge == 5);
984 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
985 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
986 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
987 
988 	for (i = 0; i < 3; ++i) {
989 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
990 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
991 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
992 	}
993 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
994 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
995 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == g_rdma_mr.lkey);
996 
997 	/* 2nd buffer consumed */
998 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
999 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
1000 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == g_rdma_mr.lkey);
1001 
1002 	/* Part 3: simple I/O, one SGL equal io unit size, io_unit_size is equal to block size 512 bytes */
1003 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1004 	reset_nvmf_rdma_request(&rdma_req);
1005 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1006 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1007 			  0, 0, 0, 0, 0);
1008 	rdma_req.req.dif.dif_insert_or_strip = true;
1009 	rtransport.transport.opts.io_unit_size = data_bs;
1010 	sgl->keyed.length = data_bs;
1011 
1012 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1013 
1014 	CU_ASSERT(rc == 0);
1015 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1016 	CU_ASSERT(rdma_req.req.length == data_bs);
1017 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1018 	CU_ASSERT(rdma_req.req.dif.elba_length == data_bs + md_size);
1019 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1020 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1021 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1022 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1023 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1024 
1025 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1026 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs);
1027 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
1028 
1029 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1030 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)((unsigned long)0x2000));
1031 	CU_ASSERT(rdma_req.req.iov[0].iov_len == data_bs);
1032 	/* 2nd buffer consumed for metadata */
1033 	CU_ASSERT(rdma_req.req.iov[1].iov_base == (void *)((unsigned long)0x2000));
1034 	CU_ASSERT(rdma_req.req.iov[1].iov_len == md_size);
1035 
1036 	/* Part 4: simple I/O, one SGL equal io unit size, io_unit_size is aligned with md_size,
1037 	   block size 512 */
1038 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1039 	reset_nvmf_rdma_request(&rdma_req);
1040 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1041 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1042 			  0, 0, 0, 0, 0);
1043 	rdma_req.req.dif.dif_insert_or_strip = true;
1044 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1045 	sgl->keyed.length = data_bs * 4;
1046 
1047 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1048 
1049 	CU_ASSERT(rc == 0);
1050 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1051 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1052 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1053 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1054 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1055 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1056 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1057 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1058 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1059 
1060 	for (i = 0; i < 4; ++i) {
1061 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1062 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1063 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
1064 	}
1065 
1066 	/* Part 5: simple I/O, one SGL equal to 2x io unit size, io_unit_size is aligned with md_size,
1067 	   block size 512 */
1068 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1069 	reset_nvmf_rdma_request(&rdma_req);
1070 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1071 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1072 			  0, 0, 0, 0, 0);
1073 	rdma_req.req.dif.dif_insert_or_strip = true;
1074 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 2;
1075 	sgl->keyed.length = data_bs * 4;
1076 
1077 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1078 
1079 	CU_ASSERT(rc == 0);
1080 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1081 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1082 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1083 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1084 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1085 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1086 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1087 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1088 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1089 
1090 	for (i = 0; i < 2; ++i) {
1091 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1092 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1093 	}
1094 	for (i = 0; i < 2; ++i) {
1095 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].addr == 0x2000 + i * (data_bs + md_size));
1096 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].length == data_bs);
1097 	}
1098 
1099 	/* Part 6: simple I/O, one SGL larger than the transport io unit size, io_unit_size is not aligned to md_size,
1100 	   block size 512 */
1101 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1102 	reset_nvmf_rdma_request(&rdma_req);
1103 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1104 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1105 			  0, 0, 0, 0, 0);
1106 	rdma_req.req.dif.dif_insert_or_strip = true;
1107 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1108 	sgl->keyed.length = data_bs * 6;
1109 
1110 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1111 
1112 	CU_ASSERT(rc == 0);
1113 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1114 	CU_ASSERT(rdma_req.req.length == data_bs * 6);
1115 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1116 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 6);
1117 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1118 	CU_ASSERT(rdma_req.data.wr.num_sge == 7);
1119 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1120 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1121 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1122 
1123 	for (i = 0; i < 3; ++i) {
1124 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1125 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1126 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
1127 	}
1128 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
1129 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
1130 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == g_rdma_mr.lkey);
1131 
1132 	/* 2nd IO buffer consumed */
1133 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
1134 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
1135 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == g_rdma_mr.lkey);
1136 
1137 	CU_ASSERT(rdma_req.data.wr.sg_list[5].addr == 0x2000 + 24 + md_size);
1138 	CU_ASSERT(rdma_req.data.wr.sg_list[5].length == 512);
1139 	CU_ASSERT(rdma_req.data.wr.sg_list[5].lkey == g_rdma_mr.lkey);
1140 
1141 	CU_ASSERT(rdma_req.data.wr.sg_list[6].addr == 0x2000 + 24 + 512 + md_size * 2);
1142 	CU_ASSERT(rdma_req.data.wr.sg_list[6].length == 512);
1143 	CU_ASSERT(rdma_req.data.wr.sg_list[6].lkey == g_rdma_mr.lkey);
1144 
1145 	/* Part 7: simple I/O, number of SGL entries exceeds the number of entries
1146 	   one WR can hold. Additional WR is chained */
1147 	MOCK_SET(spdk_mempool_get, data2_buffer);
1148 	aligned_buffer = (void *)((uintptr_t)(data2_buffer + NVMF_DATA_BUFFER_MASK) &
1149 				  ~NVMF_DATA_BUFFER_MASK);
1150 	reset_nvmf_rdma_request(&rdma_req);
1151 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1152 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1153 			  0, 0, 0, 0, 0);
1154 	rdma_req.req.dif.dif_insert_or_strip = true;
1155 	rtransport.transport.opts.io_unit_size = data_bs * 16;
1156 	sgl->keyed.length = data_bs * 16;
1157 
1158 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1159 
1160 	CU_ASSERT(rc == 0);
1161 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1162 	CU_ASSERT(rdma_req.req.length == data_bs * 16);
1163 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1164 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1165 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 16);
1166 	CU_ASSERT(rdma_req.req.data == aligned_buffer);
1167 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
1168 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1169 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1170 	/* additional wr from pool */
1171 	CU_ASSERT(rdma_req.data.wr.next == (void *)&data2->wr);
1172 	CU_ASSERT(rdma_req.data.wr.next->num_sge == 1);
1173 	CU_ASSERT(rdma_req.data.wr.next->next == &rdma_req.rsp.wr);
1174 
1175 	/* Part 8: simple I/O, data with metadata do not fit to 1 io_buffer */
1176 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1177 	reset_nvmf_rdma_request(&rdma_req);
1178 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1179 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1180 			  0, 0, 0, 0, 0);
1181 	rdma_req.req.dif.dif_insert_or_strip = true;
1182 	rtransport.transport.opts.io_unit_size = 516;
1183 	sgl->keyed.length = data_bs * 2;
1184 
1185 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1186 
1187 	CU_ASSERT(rc == 0);
1188 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1189 	CU_ASSERT(rdma_req.req.length == data_bs * 2);
1190 	CU_ASSERT(rdma_req.req.iovcnt == 3);
1191 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1192 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 2);
1193 	CU_ASSERT(rdma_req.req.data == (void *)0x2000);
1194 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1195 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1196 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1197 
1198 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1199 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 512);
1200 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
1201 
1202 	/* 2nd IO buffer consumed, offset 4 bytes due to part of the metadata
1203 	  is located at the beginning of that buffer */
1204 	CU_ASSERT(rdma_req.data.wr.sg_list[1].addr == 0x2000 + 4);
1205 	CU_ASSERT(rdma_req.data.wr.sg_list[1].length == 512);
1206 	CU_ASSERT(rdma_req.data.wr.sg_list[1].lkey == g_rdma_mr.lkey);
1207 
1208 	/* Test 9 dealing with a buffer split over two Memory Regions */
1209 	MOCK_SET(spdk_mempool_get, (void *)&buffer);
1210 	reset_nvmf_rdma_request(&rdma_req);
1211 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1212 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1213 			  0, 0, 0, 0, 0);
1214 	rdma_req.req.dif.dif_insert_or_strip = true;
1215 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1216 	sgl->keyed.length = data_bs * 2;
1217 	g_mr_size = data_bs;
1218 	g_mr_next_size = rtransport.transport.opts.io_unit_size;
1219 
1220 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1221 	SPDK_CU_ASSERT_FATAL(rc == 0);
1222 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1223 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
1224 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&buffer + NVMF_DATA_BUFFER_MASK) &
1225 			~NVMF_DATA_BUFFER_MASK));
1226 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1227 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1228 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1229 	CU_ASSERT(rdma_req.req.buffers[0] == &buffer);
1230 	for (i = 0; i < 2; i++) {
1231 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uint64_t)rdma_req.req.data + i *
1232 			  (data_bs + md_size));
1233 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1234 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
1235 	}
1236 	buffer_ptr = STAILQ_FIRST(&group.retired_bufs);
1237 	CU_ASSERT(buffer_ptr == &buffer);
1238 	STAILQ_REMOVE(&group.retired_bufs, buffer_ptr, spdk_nvmf_transport_pg_cache_buf, link);
1239 	CU_ASSERT(STAILQ_EMPTY(&group.retired_bufs));
1240 	g_mr_size = 0;
1241 	g_mr_next_size = 0;
1242 
1243 	/* Test 2: Multi SGL */
1244 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1245 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1246 	sgl->address = 0;
1247 	rdma_req.recv->buf = (void *)&sgl_desc;
1248 	MOCK_SET(spdk_mempool_get, &data);
1249 	aligned_buffer = (void *)((uintptr_t)((char *)&data + NVMF_DATA_BUFFER_MASK) &
1250 				  ~NVMF_DATA_BUFFER_MASK);
1251 
1252 	/* part 1: 2 segments each with 1 wr. io_unit_size is aligned with data_bs + md_size */
1253 	reset_nvmf_rdma_request(&rdma_req);
1254 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1255 			  SPDK_DIF_TYPE1,
1256 			  SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 0, 0, 0, 0, 0);
1257 	rdma_req.req.dif.dif_insert_or_strip = true;
1258 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1259 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
1260 
1261 	for (i = 0; i < 2; i++) {
1262 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1263 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1264 		sgl_desc[i].keyed.length = data_bs * 4;
1265 		sgl_desc[i].address = 0x4000 + i * data_bs * 4;
1266 		sgl_desc[i].keyed.key = 0x44;
1267 	}
1268 
1269 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1270 
1271 	CU_ASSERT(rc == 0);
1272 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1273 	CU_ASSERT(rdma_req.req.length == data_bs * 4 * 2);
1274 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1275 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4 * 2);
1276 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1277 	for (i = 0; i < 4; ++i) {
1278 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)((unsigned char *)aligned_buffer) + i *
1279 			  (data_bs + md_size));
1280 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1281 	}
1282 
1283 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
1284 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
1285 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
1286 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
1287 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + data_bs * 4);
1288 	CU_ASSERT(data.wr.num_sge == 4);
1289 	for (i = 0; i < 4; ++i) {
1290 		CU_ASSERT(data.wr.sg_list[i].addr == (uintptr_t)((unsigned char *)aligned_buffer) + i *
1291 			  (data_bs + md_size));
1292 		CU_ASSERT(data.wr.sg_list[i].length == data_bs);
1293 	}
1294 
1295 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
1296 }
1297 
1298 int main(int argc, char **argv)
1299 {
1300 	CU_pSuite	suite = NULL;
1301 	unsigned int	num_failures;
1302 
1303 	CU_set_error_action(CUEA_ABORT);
1304 	CU_initialize_registry();
1305 
1306 	suite = CU_add_suite("nvmf", NULL, NULL);
1307 
1308 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl);
1309 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_process);
1310 	CU_ADD_TEST(suite, test_nvmf_rdma_get_optimal_poll_group);
1311 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl_with_md);
1312 
1313 	CU_basic_set_mode(CU_BRM_VERBOSE);
1314 	CU_basic_run_tests();
1315 	num_failures = CU_get_number_of_failures();
1316 	CU_cleanup_registry();
1317 	return num_failures;
1318 }
1319