xref: /spdk/test/unit/lib/nvmf/rdma.c/rdma_ut.c (revision e54df32e6be5d73aafd29e9493dc58ec8f0c0cd3)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright (c) Intel Corporation. All rights reserved.
5  *   Copyright (c) 2019, 2021 Mellanox Technologies LTD. All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include "spdk/stdinc.h"
35 #include "spdk_cunit.h"
36 #include "common/lib/test_env.c"
37 #include "common/lib/test_rdma.c"
38 #include "nvmf/rdma.c"
39 #include "nvmf/transport.c"
40 
41 #define RDMA_UT_UNITS_IN_MAX_IO 16
42 
43 struct spdk_nvmf_transport_opts g_rdma_ut_transport_opts = {
44 	.max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH,
45 	.max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR,
46 	.in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE,
47 	.max_io_size = (SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE * RDMA_UT_UNITS_IN_MAX_IO),
48 	.io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE,
49 	.max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH,
50 	.num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS,
51 };
52 
53 SPDK_LOG_REGISTER_COMPONENT(nvmf)
54 DEFINE_STUB(spdk_mem_map_set_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
55 		uint64_t size, uint64_t translation), 0);
56 DEFINE_STUB(spdk_mem_map_clear_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
57 		uint64_t size), 0);
58 DEFINE_STUB(spdk_mem_map_alloc, struct spdk_mem_map *, (uint64_t default_translation,
59 		const struct spdk_mem_map_ops *ops, void *cb_ctx), NULL);
60 DEFINE_STUB(spdk_nvmf_qpair_disconnect, int, (struct spdk_nvmf_qpair *qpair,
61 		nvmf_qpair_disconnect_cb cb_fn, void *ctx), 0);
62 DEFINE_STUB(spdk_nvmf_qpair_get_listen_trid, int,
63 	    (struct spdk_nvmf_qpair *qpair, struct spdk_nvme_transport_id *trid), 0);
64 DEFINE_STUB_V(spdk_mem_map_free, (struct spdk_mem_map **pmap));
65 
66 DEFINE_STUB_V(spdk_nvmf_ctrlr_data_init, (struct spdk_nvmf_transport_opts *opts,
67 		struct spdk_nvmf_ctrlr_data *cdata));
68 DEFINE_STUB_V(spdk_nvmf_request_exec, (struct spdk_nvmf_request *req));
69 DEFINE_STUB(spdk_nvmf_request_complete, int, (struct spdk_nvmf_request *req), 0);
70 DEFINE_STUB(spdk_nvme_transport_id_compare, int, (const struct spdk_nvme_transport_id *trid1,
71 		const struct spdk_nvme_transport_id *trid2), 0);
72 DEFINE_STUB_V(nvmf_ctrlr_abort_aer, (struct spdk_nvmf_ctrlr *ctrlr));
73 DEFINE_STUB(spdk_nvmf_request_get_dif_ctx, bool, (struct spdk_nvmf_request *req,
74 		struct spdk_dif_ctx *dif_ctx), false);
75 DEFINE_STUB_V(spdk_nvme_trid_populate_transport, (struct spdk_nvme_transport_id *trid,
76 		enum spdk_nvme_transport_type trtype));
77 DEFINE_STUB_V(spdk_nvmf_tgt_new_qpair, (struct spdk_nvmf_tgt *tgt, struct spdk_nvmf_qpair *qpair));
78 DEFINE_STUB(nvmf_ctrlr_abort_request, int, (struct spdk_nvmf_request *req), 0);
79 
80 const char *
81 spdk_nvme_transport_id_trtype_str(enum spdk_nvme_transport_type trtype)
82 {
83 	switch (trtype) {
84 	case SPDK_NVME_TRANSPORT_PCIE:
85 		return "PCIe";
86 	case SPDK_NVME_TRANSPORT_RDMA:
87 		return "RDMA";
88 	case SPDK_NVME_TRANSPORT_FC:
89 		return "FC";
90 	default:
91 		return NULL;
92 	}
93 }
94 
95 int
96 spdk_nvme_transport_id_populate_trstring(struct spdk_nvme_transport_id *trid, const char *trstring)
97 {
98 	int len, i;
99 
100 	if (trstring == NULL) {
101 		return -EINVAL;
102 	}
103 
104 	len = strnlen(trstring, SPDK_NVMF_TRSTRING_MAX_LEN);
105 	if (len == SPDK_NVMF_TRSTRING_MAX_LEN) {
106 		return -EINVAL;
107 	}
108 
109 	/* cast official trstring to uppercase version of input. */
110 	for (i = 0; i < len; i++) {
111 		trid->trstring[i] = toupper(trstring[i]);
112 	}
113 	return 0;
114 }
115 
116 static void reset_nvmf_rdma_request(struct spdk_nvmf_rdma_request *rdma_req)
117 {
118 	int i;
119 
120 	rdma_req->req.length = 0;
121 	rdma_req->req.data_from_pool = false;
122 	rdma_req->req.data = NULL;
123 	rdma_req->data.wr.num_sge = 0;
124 	rdma_req->data.wr.wr.rdma.remote_addr = 0;
125 	rdma_req->data.wr.wr.rdma.rkey = 0;
126 	rdma_req->offset = 0;
127 	memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
128 
129 	for (i = 0; i < SPDK_NVMF_MAX_SGL_ENTRIES; i++) {
130 		rdma_req->req.iov[i].iov_base = 0;
131 		rdma_req->req.iov[i].iov_len = 0;
132 		rdma_req->req.buffers[i] = 0;
133 		rdma_req->data.wr.sg_list[i].addr = 0;
134 		rdma_req->data.wr.sg_list[i].length = 0;
135 		rdma_req->data.wr.sg_list[i].lkey = 0;
136 	}
137 	rdma_req->req.iovcnt = 0;
138 }
139 
140 static void
141 test_spdk_nvmf_rdma_request_parse_sgl(void)
142 {
143 	struct spdk_nvmf_rdma_transport rtransport;
144 	struct spdk_nvmf_rdma_device device;
145 	struct spdk_nvmf_rdma_request rdma_req = {};
146 	struct spdk_nvmf_rdma_recv recv;
147 	struct spdk_nvmf_rdma_poll_group group;
148 	struct spdk_nvmf_rdma_qpair rqpair;
149 	struct spdk_nvmf_rdma_poller poller;
150 	union nvmf_c2h_msg cpl;
151 	union nvmf_h2c_msg cmd;
152 	struct spdk_nvme_sgl_descriptor *sgl;
153 	struct spdk_nvmf_transport_pg_cache_buf bufs[4];
154 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
155 	struct spdk_nvmf_rdma_request_data data;
156 	int rc, i;
157 	uint32_t sgl_length;
158 	uintptr_t aligned_buffer_address;
159 
160 	data.wr.sg_list = data.sgl;
161 	STAILQ_INIT(&group.group.buf_cache);
162 	group.group.buf_cache_size = 0;
163 	group.group.buf_cache_count = 0;
164 	group.group.transport = &rtransport.transport;
165 	poller.group = &group;
166 	rqpair.poller = &poller;
167 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
168 
169 	sgl = &cmd.nvme_cmd.dptr.sgl1;
170 	rdma_req.recv = &recv;
171 	rdma_req.req.cmd = &cmd;
172 	rdma_req.req.rsp = &cpl;
173 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
174 	rdma_req.req.qpair = &rqpair.qpair;
175 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
176 
177 	rtransport.transport.opts = g_rdma_ut_transport_opts;
178 	rtransport.data_wr_pool = NULL;
179 	rtransport.transport.data_buf_pool = NULL;
180 
181 	device.attr.device_cap_flags = 0;
182 	sgl->keyed.key = 0xEEEE;
183 	sgl->address = 0xFFFF;
184 	rdma_req.recv->buf = (void *)0xDDDD;
185 
186 	/* Test 1: sgl type: keyed data block subtype: address */
187 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
188 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
189 
190 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size */
191 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
192 	reset_nvmf_rdma_request(&rdma_req);
193 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
194 
195 	device.map = (void *)0x0;
196 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
197 	CU_ASSERT(rc == 0);
198 	CU_ASSERT(rdma_req.req.data_from_pool == true);
199 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
200 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
201 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
202 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
203 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
204 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
205 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
206 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
207 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
208 
209 	/* Part 2: simple I/O, one SGL larger than the transport io unit size (equal to the max io size) */
210 	reset_nvmf_rdma_request(&rdma_req);
211 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
212 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
213 
214 	CU_ASSERT(rc == 0);
215 	CU_ASSERT(rdma_req.req.data_from_pool == true);
216 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO);
217 	CU_ASSERT(rdma_req.data.wr.num_sge == RDMA_UT_UNITS_IN_MAX_IO);
218 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
219 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
220 	for (i = 0; i < RDMA_UT_UNITS_IN_MAX_IO; i++) {
221 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
222 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
223 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
224 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
225 	}
226 
227 	/* Part 3: simple I/O one SGL larger than the transport max io size */
228 	reset_nvmf_rdma_request(&rdma_req);
229 	sgl->keyed.length = rtransport.transport.opts.max_io_size * 2;
230 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
231 
232 	CU_ASSERT(rc == -1);
233 
234 	/* Part 4: Pretend there are no buffer pools */
235 	MOCK_SET(spdk_mempool_get, NULL);
236 	reset_nvmf_rdma_request(&rdma_req);
237 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
238 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
239 
240 	CU_ASSERT(rc == 0);
241 	CU_ASSERT(rdma_req.req.data_from_pool == false);
242 	CU_ASSERT(rdma_req.req.data == NULL);
243 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
244 	CU_ASSERT(rdma_req.req.buffers[0] == NULL);
245 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0);
246 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 0);
247 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == 0);
248 
249 	rdma_req.recv->buf = (void *)0xDDDD;
250 	/* Test 2: sgl type: keyed data block subtype: offset (in capsule data) */
251 	sgl->generic.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
252 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
253 
254 	/* Part 1: Normal I/O smaller than in capsule data size no offset */
255 	reset_nvmf_rdma_request(&rdma_req);
256 	sgl->address = 0;
257 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
258 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
259 
260 	CU_ASSERT(rc == 0);
261 	CU_ASSERT(rdma_req.req.data == (void *)0xDDDD);
262 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.in_capsule_data_size);
263 	CU_ASSERT(rdma_req.req.data_from_pool == false);
264 
265 	/* Part 2: I/O offset + length too large */
266 	reset_nvmf_rdma_request(&rdma_req);
267 	sgl->address = rtransport.transport.opts.in_capsule_data_size;
268 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
269 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
270 
271 	CU_ASSERT(rc == -1);
272 
273 	/* Part 3: I/O too large */
274 	reset_nvmf_rdma_request(&rdma_req);
275 	sgl->address = 0;
276 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size * 2;
277 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
278 
279 	CU_ASSERT(rc == -1);
280 
281 	/* Test 3: Multi SGL */
282 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
283 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
284 	sgl->address = 0;
285 	rdma_req.recv->buf = (void *)&sgl_desc;
286 	MOCK_SET(spdk_mempool_get, &data);
287 
288 	/* part 1: 2 segments each with 1 wr. */
289 	reset_nvmf_rdma_request(&rdma_req);
290 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
291 	for (i = 0; i < 2; i++) {
292 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
293 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
294 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size;
295 		sgl_desc[i].address = 0x4000 + i * rtransport.transport.opts.io_unit_size;
296 		sgl_desc[i].keyed.key = 0x44;
297 	}
298 
299 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
300 
301 	CU_ASSERT(rc == 0);
302 	CU_ASSERT(rdma_req.req.data_from_pool == true);
303 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 2);
304 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
305 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
306 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
307 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
308 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
309 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size);
310 	CU_ASSERT(data.wr.num_sge == 1);
311 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
312 
313 	/* part 2: 2 segments, each with 1 wr containing 8 sge_elements */
314 	reset_nvmf_rdma_request(&rdma_req);
315 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
316 	for (i = 0; i < 2; i++) {
317 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
318 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
319 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size * 8;
320 		sgl_desc[i].address = 0x4000 + i * 8 * rtransport.transport.opts.io_unit_size;
321 		sgl_desc[i].keyed.key = 0x44;
322 	}
323 
324 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
325 
326 	CU_ASSERT(rc == 0);
327 	CU_ASSERT(rdma_req.req.data_from_pool == true);
328 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
329 	CU_ASSERT(rdma_req.req.iovcnt == 16);
330 	CU_ASSERT(rdma_req.data.wr.num_sge == 8);
331 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
332 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
333 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
334 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
335 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 8);
336 	CU_ASSERT(data.wr.num_sge == 8);
337 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
338 
339 	/* part 3: 2 segments, one very large, one very small */
340 	reset_nvmf_rdma_request(&rdma_req);
341 	for (i = 0; i < 2; i++) {
342 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
343 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
344 		sgl_desc[i].keyed.key = 0x44;
345 	}
346 
347 	sgl_desc[0].keyed.length = rtransport.transport.opts.io_unit_size * 15 +
348 				   rtransport.transport.opts.io_unit_size / 2;
349 	sgl_desc[0].address = 0x4000;
350 	sgl_desc[1].keyed.length = rtransport.transport.opts.io_unit_size / 2;
351 	sgl_desc[1].address = 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
352 			      rtransport.transport.opts.io_unit_size / 2;
353 
354 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
355 
356 	CU_ASSERT(rc == 0);
357 	CU_ASSERT(rdma_req.req.data_from_pool == true);
358 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
359 	CU_ASSERT(rdma_req.req.iovcnt == 16);
360 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
361 	for (i = 0; i < 15; i++) {
362 		CU_ASSERT(rdma_req.data.sgl[i].length == rtransport.transport.opts.io_unit_size);
363 	}
364 	CU_ASSERT(rdma_req.data.sgl[15].length == rtransport.transport.opts.io_unit_size / 2);
365 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
366 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
367 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
368 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
369 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
370 		  rtransport.transport.opts.io_unit_size / 2);
371 	CU_ASSERT(data.sgl[0].length == rtransport.transport.opts.io_unit_size / 2);
372 	CU_ASSERT(data.wr.num_sge == 1);
373 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
374 
375 	/* part 4: 2 SGL descriptors, each length is transport buffer / 2
376 	 * 1 transport buffers should be allocated */
377 	reset_nvmf_rdma_request(&rdma_req);
378 	aligned_buffer_address = ((uintptr_t)(&data) + NVMF_DATA_BUFFER_MASK) & ~NVMF_DATA_BUFFER_MASK;
379 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
380 	sgl_length = rtransport.transport.opts.io_unit_size / 2;
381 	for (i = 0; i < 2; i++) {
382 		sgl_desc[i].keyed.length = sgl_length;
383 		sgl_desc[i].address = 0x4000 + i * sgl_length;
384 	}
385 
386 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
387 
388 	CU_ASSERT(rc == 0);
389 	CU_ASSERT(rdma_req.req.data_from_pool == true);
390 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size);
391 	CU_ASSERT(rdma_req.req.iovcnt == 1);
392 
393 	CU_ASSERT(rdma_req.data.sgl[0].length == sgl_length);
394 	/* We mocked mempool_get to return address of data variable. Mempool is used
395 	 * to get both additional WRs and data buffers, so data points to &data */
396 	CU_ASSERT(rdma_req.data.sgl[0].addr == aligned_buffer_address);
397 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
398 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
399 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
400 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
401 
402 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
403 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + sgl_length);
404 	CU_ASSERT(data.sgl[0].length == sgl_length);
405 	CU_ASSERT(data.sgl[0].addr == aligned_buffer_address + sgl_length);
406 	CU_ASSERT(data.wr.num_sge == 1);
407 
408 	/* Test 4: use PG buffer cache */
409 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
410 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
411 	sgl->address = 0xFFFF;
412 	rdma_req.recv->buf = (void *)0xDDDD;
413 	sgl->keyed.key = 0xEEEE;
414 
415 	for (i = 0; i < 4; i++) {
416 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
417 	}
418 
419 	/* part 1: use the four buffers from the pg cache */
420 	group.group.buf_cache_size = 4;
421 	group.group.buf_cache_count = 4;
422 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
423 	reset_nvmf_rdma_request(&rdma_req);
424 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * 4;
425 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
426 
427 	SPDK_CU_ASSERT_FATAL(rc == 0);
428 	CU_ASSERT(rdma_req.req.data_from_pool == true);
429 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
430 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
431 			~NVMF_DATA_BUFFER_MASK));
432 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
433 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
434 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
435 	CU_ASSERT(group.group.buf_cache_count == 0);
436 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
437 	for (i = 0; i < 4; i++) {
438 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
439 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
440 				~NVMF_DATA_BUFFER_MASK));
441 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
442 	}
443 
444 	/* part 2: now that we have used the buffers from the cache, try again. We should get mempool buffers. */
445 	reset_nvmf_rdma_request(&rdma_req);
446 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
447 
448 	SPDK_CU_ASSERT_FATAL(rc == 0);
449 	CU_ASSERT(rdma_req.req.data_from_pool == true);
450 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
451 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
452 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
453 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
454 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
455 	CU_ASSERT(group.group.buf_cache_count == 0);
456 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
457 	for (i = 0; i < 4; i++) {
458 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
459 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
460 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
461 		CU_ASSERT(group.group.buf_cache_count == 0);
462 	}
463 
464 	/* part 3: half and half */
465 	group.group.buf_cache_count = 2;
466 
467 	for (i = 0; i < 2; i++) {
468 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
469 	}
470 	reset_nvmf_rdma_request(&rdma_req);
471 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
472 
473 	SPDK_CU_ASSERT_FATAL(rc == 0);
474 	CU_ASSERT(rdma_req.req.data_from_pool == true);
475 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
476 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
477 			~NVMF_DATA_BUFFER_MASK));
478 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
479 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
480 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
481 	CU_ASSERT(group.group.buf_cache_count == 0);
482 	for (i = 0; i < 2; i++) {
483 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
484 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
485 				~NVMF_DATA_BUFFER_MASK));
486 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
487 	}
488 	for (i = 2; i < 4; i++) {
489 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
490 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
491 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
492 	}
493 
494 	reset_nvmf_rdma_request(&rdma_req);
495 }
496 
497 static struct spdk_nvmf_rdma_recv *
498 create_recv(struct spdk_nvmf_rdma_qpair *rqpair, enum spdk_nvme_nvm_opcode opc)
499 {
500 	struct spdk_nvmf_rdma_recv *rdma_recv;
501 	union nvmf_h2c_msg *cmd;
502 	struct spdk_nvme_sgl_descriptor *sgl;
503 
504 	rdma_recv = calloc(1, sizeof(*rdma_recv));
505 	rdma_recv->qpair = rqpair;
506 	cmd = calloc(1, sizeof(*cmd));
507 	rdma_recv->sgl[0].addr = (uintptr_t)cmd;
508 	cmd->nvme_cmd.opc = opc;
509 	sgl = &cmd->nvme_cmd.dptr.sgl1;
510 	sgl->keyed.key = 0xEEEE;
511 	sgl->address = 0xFFFF;
512 	sgl->keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
513 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
514 	sgl->keyed.length = 1;
515 
516 	return rdma_recv;
517 }
518 
519 static void
520 free_recv(struct spdk_nvmf_rdma_recv *rdma_recv)
521 {
522 	free((void *)rdma_recv->sgl[0].addr);
523 	free(rdma_recv);
524 }
525 
526 static struct spdk_nvmf_rdma_request *
527 create_req(struct spdk_nvmf_rdma_qpair *rqpair,
528 	   struct spdk_nvmf_rdma_recv *rdma_recv)
529 {
530 	struct spdk_nvmf_rdma_request *rdma_req;
531 	union nvmf_c2h_msg *cpl;
532 
533 	rdma_req = calloc(1, sizeof(*rdma_req));
534 	rdma_req->recv = rdma_recv;
535 	rdma_req->req.qpair = &rqpair->qpair;
536 	rdma_req->state = RDMA_REQUEST_STATE_NEW;
537 	rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr;
538 	rdma_req->data.wr.sg_list = rdma_req->data.sgl;
539 	cpl = calloc(1, sizeof(*cpl));
540 	rdma_req->rsp.sgl[0].addr = (uintptr_t)cpl;
541 	rdma_req->req.rsp = cpl;
542 
543 	return rdma_req;
544 }
545 
546 static void
547 free_req(struct spdk_nvmf_rdma_request *rdma_req)
548 {
549 	free((void *)rdma_req->rsp.sgl[0].addr);
550 	free(rdma_req);
551 }
552 
553 static void
554 qpair_reset(struct spdk_nvmf_rdma_qpair *rqpair,
555 	    struct spdk_nvmf_rdma_poller *poller,
556 	    struct spdk_nvmf_rdma_device *device,
557 	    struct spdk_nvmf_rdma_resources *resources,
558 	    struct spdk_nvmf_transport *transport)
559 {
560 	memset(rqpair, 0, sizeof(*rqpair));
561 	STAILQ_INIT(&rqpair->pending_rdma_write_queue);
562 	STAILQ_INIT(&rqpair->pending_rdma_read_queue);
563 	rqpair->poller = poller;
564 	rqpair->device = device;
565 	rqpair->resources = resources;
566 	rqpair->qpair.qid = 1;
567 	rqpair->ibv_state = IBV_QPS_RTS;
568 	rqpair->qpair.state = SPDK_NVMF_QPAIR_ACTIVE;
569 	rqpair->max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
570 	rqpair->max_send_depth = 16;
571 	rqpair->max_read_depth = 16;
572 	rqpair->qpair.transport = transport;
573 }
574 
575 static void
576 poller_reset(struct spdk_nvmf_rdma_poller *poller,
577 	     struct spdk_nvmf_rdma_poll_group *group)
578 {
579 	memset(poller, 0, sizeof(*poller));
580 	STAILQ_INIT(&poller->qpairs_pending_recv);
581 	STAILQ_INIT(&poller->qpairs_pending_send);
582 	poller->group = group;
583 }
584 
585 static void
586 test_spdk_nvmf_rdma_request_process(void)
587 {
588 	struct spdk_nvmf_rdma_transport rtransport = {};
589 	struct spdk_nvmf_rdma_poll_group group = {};
590 	struct spdk_nvmf_rdma_poller poller = {};
591 	struct spdk_nvmf_rdma_device device = {};
592 	struct spdk_nvmf_rdma_resources resources = {};
593 	struct spdk_nvmf_rdma_qpair rqpair = {};
594 	struct spdk_nvmf_rdma_recv *rdma_recv;
595 	struct spdk_nvmf_rdma_request *rdma_req;
596 	bool progress;
597 
598 	STAILQ_INIT(&group.group.buf_cache);
599 	STAILQ_INIT(&group.group.pending_buf_queue);
600 	group.group.buf_cache_size = 0;
601 	group.group.buf_cache_count = 0;
602 	poller_reset(&poller, &group);
603 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
604 
605 	rtransport.transport.opts = g_rdma_ut_transport_opts;
606 	rtransport.transport.data_buf_pool = spdk_mempool_create("test_data_pool", 16, 128, 0, 0);
607 	rtransport.data_wr_pool = spdk_mempool_create("test_wr_pool", 128,
608 				  sizeof(struct spdk_nvmf_rdma_request_data),
609 				  0, 0);
610 	MOCK_CLEAR(spdk_mempool_get);
611 
612 	device.attr.device_cap_flags = 0;
613 	device.map = (void *)0x0;
614 
615 	/* Test 1: single SGL READ request */
616 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_READ);
617 	rdma_req = create_req(&rqpair, rdma_recv);
618 	rqpair.current_recv_depth = 1;
619 	/* NEW -> EXECUTING */
620 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
621 	CU_ASSERT(progress == true);
622 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
623 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST);
624 	/* EXECUTED -> TRANSFERRING_C2H */
625 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
626 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
627 	CU_ASSERT(progress == true);
628 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST);
629 	CU_ASSERT(rdma_req->recv == NULL);
630 	/* COMPLETED -> FREE */
631 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
632 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
633 	CU_ASSERT(progress == true);
634 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
635 
636 	free_recv(rdma_recv);
637 	free_req(rdma_req);
638 	poller_reset(&poller, &group);
639 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
640 
641 	/* Test 2: single SGL WRITE request */
642 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
643 	rdma_req = create_req(&rqpair, rdma_recv);
644 	rqpair.current_recv_depth = 1;
645 	/* NEW -> TRANSFERRING_H2C */
646 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
647 	CU_ASSERT(progress == true);
648 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
649 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
650 	STAILQ_INIT(&poller.qpairs_pending_send);
651 	/* READY_TO_EXECUTE -> EXECUTING */
652 	rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
653 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
654 	CU_ASSERT(progress == true);
655 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
656 	/* EXECUTED -> COMPLETING */
657 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
658 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
659 	CU_ASSERT(progress == true);
660 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
661 	CU_ASSERT(rdma_req->recv == NULL);
662 	/* COMPLETED -> FREE */
663 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
664 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
665 	CU_ASSERT(progress == true);
666 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
667 
668 	free_recv(rdma_recv);
669 	free_req(rdma_req);
670 	poller_reset(&poller, &group);
671 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
672 
673 	/* Test 3: WRITE+WRITE ibv_send batching */
674 	{
675 		struct spdk_nvmf_rdma_recv *recv1, *recv2;
676 		struct spdk_nvmf_rdma_request *req1, *req2;
677 		recv1 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
678 		req1 = create_req(&rqpair, recv1);
679 		recv2 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
680 		req2 = create_req(&rqpair, recv2);
681 
682 		/* WRITE 1: NEW -> TRANSFERRING_H2C */
683 		rqpair.current_recv_depth = 1;
684 		nvmf_rdma_request_process(&rtransport, req1);
685 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
686 
687 		/* WRITE 2: NEW -> TRANSFERRING_H2C */
688 		rqpair.current_recv_depth = 2;
689 		nvmf_rdma_request_process(&rtransport, req2);
690 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
691 
692 		STAILQ_INIT(&poller.qpairs_pending_send);
693 
694 		/* WRITE 1 completes before WRITE 2 has finished RDMA reading */
695 		/* WRITE 1: READY_TO_EXECUTE -> EXECUTING */
696 		req1->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
697 		nvmf_rdma_request_process(&rtransport, req1);
698 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_EXECUTING);
699 		/* WRITE 1: EXECUTED -> COMPLETING */
700 		req1->state = RDMA_REQUEST_STATE_EXECUTED;
701 		nvmf_rdma_request_process(&rtransport, req1);
702 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_COMPLETING);
703 		STAILQ_INIT(&poller.qpairs_pending_send);
704 		/* WRITE 1: COMPLETED -> FREE */
705 		req1->state = RDMA_REQUEST_STATE_COMPLETED;
706 		nvmf_rdma_request_process(&rtransport, req1);
707 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_FREE);
708 
709 		/* Now WRITE 2 has finished reading and completes */
710 		/* WRITE 2: COMPLETED -> FREE */
711 		/* WRITE 2: READY_TO_EXECUTE -> EXECUTING */
712 		req2->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
713 		nvmf_rdma_request_process(&rtransport, req2);
714 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_EXECUTING);
715 		/* WRITE 1: EXECUTED -> COMPLETING */
716 		req2->state = RDMA_REQUEST_STATE_EXECUTED;
717 		nvmf_rdma_request_process(&rtransport, req2);
718 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_COMPLETING);
719 		STAILQ_INIT(&poller.qpairs_pending_send);
720 		/* WRITE 1: COMPLETED -> FREE */
721 		req2->state = RDMA_REQUEST_STATE_COMPLETED;
722 		nvmf_rdma_request_process(&rtransport, req2);
723 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_FREE);
724 
725 		free_recv(recv1);
726 		free_req(req1);
727 		free_recv(recv2);
728 		free_req(req2);
729 		poller_reset(&poller, &group);
730 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
731 	}
732 
733 	/* Test 4, invalid command, check xfer type */
734 	{
735 		struct spdk_nvmf_rdma_recv *rdma_recv_inv;
736 		struct spdk_nvmf_rdma_request *rdma_req_inv;
737 		/* construct an opcode that specifies BIDIRECTIONAL transfer */
738 		uint8_t opc = 0x10 | SPDK_NVME_DATA_BIDIRECTIONAL;
739 
740 		rdma_recv_inv = create_recv(&rqpair, opc);
741 		rdma_req_inv = create_req(&rqpair, rdma_recv_inv);
742 
743 		/* NEW -> RDMA_REQUEST_STATE_COMPLETING */
744 		rqpair.current_recv_depth = 1;
745 		progress = nvmf_rdma_request_process(&rtransport, rdma_req_inv);
746 		CU_ASSERT(progress == true);
747 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_COMPLETING);
748 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sct == SPDK_NVME_SCT_GENERIC);
749 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sc == SPDK_NVME_SC_INVALID_OPCODE);
750 
751 		/* RDMA_REQUEST_STATE_COMPLETED -> FREE */
752 		rdma_req_inv->state = RDMA_REQUEST_STATE_COMPLETED;
753 		nvmf_rdma_request_process(&rtransport, rdma_req_inv);
754 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_FREE);
755 
756 		free_recv(rdma_recv_inv);
757 		free_req(rdma_req_inv);
758 		poller_reset(&poller, &group);
759 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
760 	}
761 
762 	spdk_mempool_free(rtransport.transport.data_buf_pool);
763 	spdk_mempool_free(rtransport.data_wr_pool);
764 }
765 
766 #define TEST_GROUPS_COUNT 5
767 static void
768 test_nvmf_rdma_get_optimal_poll_group(void)
769 {
770 	struct spdk_nvmf_rdma_transport rtransport = {};
771 	struct spdk_nvmf_transport *transport = &rtransport.transport;
772 	struct spdk_nvmf_rdma_qpair rqpair = {};
773 	struct spdk_nvmf_transport_poll_group *groups[TEST_GROUPS_COUNT];
774 	struct spdk_nvmf_rdma_poll_group *rgroups[TEST_GROUPS_COUNT];
775 	struct spdk_nvmf_transport_poll_group *result;
776 	uint32_t i;
777 
778 	rqpair.qpair.transport = transport;
779 	pthread_mutex_init(&rtransport.lock, NULL);
780 	TAILQ_INIT(&rtransport.poll_groups);
781 
782 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
783 		groups[i] = nvmf_rdma_poll_group_create(transport);
784 		CU_ASSERT(groups[i] != NULL);
785 		rgroups[i] = SPDK_CONTAINEROF(groups[i], struct spdk_nvmf_rdma_poll_group, group);
786 		groups[i]->transport = transport;
787 	}
788 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[0]);
789 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[0]);
790 
791 	/* Emulate connection of %TEST_GROUPS_COUNT% initiators - each creates 1 admin and 1 io qp */
792 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
793 		rqpair.qpair.qid = 0;
794 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
795 		CU_ASSERT(result == groups[i]);
796 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
797 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i]);
798 
799 		rqpair.qpair.qid = 1;
800 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
801 		CU_ASSERT(result == groups[i]);
802 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
803 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
804 	}
805 	/* wrap around, admin/io pg point to the first pg
806 	   Destroy all poll groups except of the last one */
807 	for (i = 0; i < TEST_GROUPS_COUNT - 1; i++) {
808 		nvmf_rdma_poll_group_destroy(groups[i]);
809 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[i + 1]);
810 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i + 1]);
811 	}
812 
813 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
814 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
815 
816 	/* Check that pointers to the next admin/io poll groups are not changed */
817 	rqpair.qpair.qid = 0;
818 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
819 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
820 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
821 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
822 
823 	rqpair.qpair.qid = 1;
824 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
825 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
826 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
827 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
828 
829 	/* Remove the last poll group, check that pointers are NULL */
830 	nvmf_rdma_poll_group_destroy(groups[TEST_GROUPS_COUNT - 1]);
831 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == NULL);
832 	CU_ASSERT(rtransport.conn_sched.next_io_pg == NULL);
833 
834 	/* Request optimal poll group, result must be NULL */
835 	rqpair.qpair.qid = 0;
836 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
837 	CU_ASSERT(result == NULL);
838 
839 	rqpair.qpair.qid = 1;
840 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
841 	CU_ASSERT(result == NULL);
842 
843 	pthread_mutex_destroy(&rtransport.lock);
844 }
845 #undef TEST_GROUPS_COUNT
846 
847 static void
848 test_spdk_nvmf_rdma_request_parse_sgl_with_md(void)
849 {
850 	struct spdk_nvmf_rdma_transport rtransport;
851 	struct spdk_nvmf_rdma_device device;
852 	struct spdk_nvmf_rdma_request rdma_req = {};
853 	struct spdk_nvmf_rdma_recv recv;
854 	struct spdk_nvmf_rdma_poll_group group;
855 	struct spdk_nvmf_rdma_qpair rqpair;
856 	struct spdk_nvmf_rdma_poller poller;
857 	union nvmf_c2h_msg cpl;
858 	union nvmf_h2c_msg cmd;
859 	struct spdk_nvme_sgl_descriptor *sgl;
860 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
861 	char data_buffer[8192];
862 	struct spdk_nvmf_rdma_request_data *data = (struct spdk_nvmf_rdma_request_data *)data_buffer;
863 	char data2_buffer[8192];
864 	struct spdk_nvmf_rdma_request_data *data2 = (struct spdk_nvmf_rdma_request_data *)data2_buffer;
865 	const uint32_t data_bs = 512;
866 	const uint32_t md_size = 8;
867 	int rc, i;
868 	void *aligned_buffer;
869 
870 	data->wr.sg_list = data->sgl;
871 	STAILQ_INIT(&group.group.buf_cache);
872 	group.group.buf_cache_size = 0;
873 	group.group.buf_cache_count = 0;
874 	group.group.transport = &rtransport.transport;
875 	poller.group = &group;
876 	rqpair.poller = &poller;
877 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
878 
879 	sgl = &cmd.nvme_cmd.dptr.sgl1;
880 	rdma_req.recv = &recv;
881 	rdma_req.req.cmd = &cmd;
882 	rdma_req.req.rsp = &cpl;
883 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
884 	rdma_req.req.qpair = &rqpair.qpair;
885 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
886 
887 	rtransport.transport.opts = g_rdma_ut_transport_opts;
888 	rtransport.data_wr_pool = NULL;
889 	rtransport.transport.data_buf_pool = NULL;
890 
891 	device.attr.device_cap_flags = 0;
892 	device.map = NULL;
893 	sgl->keyed.key = 0xEEEE;
894 	sgl->address = 0xFFFF;
895 	rdma_req.recv->buf = (void *)0xDDDD;
896 
897 	/* Test 1: sgl type: keyed data block subtype: address */
898 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
899 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
900 
901 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size, block size 512 */
902 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
903 	reset_nvmf_rdma_request(&rdma_req);
904 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
905 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
906 			  0, 0, 0, 0, 0);
907 	rdma_req.req.dif_enabled = true;
908 	rtransport.transport.opts.io_unit_size = data_bs * 8;
909 	sgl->keyed.length = data_bs * 4;
910 
911 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
912 
913 	CU_ASSERT(rc == 0);
914 	CU_ASSERT(rdma_req.req.data_from_pool == true);
915 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
916 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
917 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
918 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
919 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
920 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
921 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
922 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
923 
924 	for (i = 0; i < 4; ++i) {
925 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
926 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
927 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
928 	}
929 
930 	/* Part 2: simple I/O, one SGL equal to io unit size, io_unit_size is not aligned with md_size,
931 		block size 512 */
932 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
933 	reset_nvmf_rdma_request(&rdma_req);
934 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
935 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
936 			  0, 0, 0, 0, 0);
937 	rdma_req.req.dif_enabled = true;
938 	rtransport.transport.opts.io_unit_size = data_bs * 4;
939 	sgl->keyed.length = data_bs * 4;
940 
941 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
942 
943 	CU_ASSERT(rc == 0);
944 	CU_ASSERT(rdma_req.req.data_from_pool == true);
945 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
946 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
947 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
948 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
949 	CU_ASSERT(rdma_req.data.wr.num_sge == 5);
950 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
951 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
952 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
953 
954 	for (i = 0; i < 3; ++i) {
955 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
956 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
957 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
958 	}
959 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
960 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
961 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
962 
963 	/* 2nd buffer consumed */
964 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
965 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
966 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
967 
968 	/* Part 3: simple I/O, one SGL equal io unit size, io_unit_size is equal to block size 512 bytes */
969 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
970 	reset_nvmf_rdma_request(&rdma_req);
971 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
972 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
973 			  0, 0, 0, 0, 0);
974 	rdma_req.req.dif_enabled = true;
975 	rtransport.transport.opts.io_unit_size = data_bs;
976 	sgl->keyed.length = data_bs;
977 
978 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
979 
980 	CU_ASSERT(rc == 0);
981 	CU_ASSERT(rdma_req.req.data_from_pool == true);
982 	CU_ASSERT(rdma_req.req.length == data_bs);
983 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
984 	CU_ASSERT(rdma_req.req.dif.elba_length == data_bs + md_size);
985 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
986 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
987 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
988 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
989 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
990 
991 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
992 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs);
993 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
994 
995 	CU_ASSERT(rdma_req.req.iovcnt == 2);
996 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)((unsigned long)0x2000));
997 	CU_ASSERT(rdma_req.req.iov[0].iov_len == data_bs);
998 	/* 2nd buffer consumed for metadata */
999 	CU_ASSERT(rdma_req.req.iov[1].iov_base == (void *)((unsigned long)0x2000));
1000 	CU_ASSERT(rdma_req.req.iov[1].iov_len == md_size);
1001 
1002 	/* Part 4: simple I/O, one SGL equal io unit size, io_unit_size is aligned with md_size,
1003 	   block size 512 */
1004 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1005 	reset_nvmf_rdma_request(&rdma_req);
1006 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1007 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1008 			  0, 0, 0, 0, 0);
1009 	rdma_req.req.dif_enabled = true;
1010 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1011 	sgl->keyed.length = data_bs * 4;
1012 
1013 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1014 
1015 	CU_ASSERT(rc == 0);
1016 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1017 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1018 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1019 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1020 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1021 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1022 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1023 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1024 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1025 
1026 	for (i = 0; i < 4; ++i) {
1027 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1028 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1029 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1030 	}
1031 
1032 	/* Part 5: simple I/O, one SGL equal to 2x io unit size, io_unit_size is aligned with md_size,
1033 	   block size 512 */
1034 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1035 	reset_nvmf_rdma_request(&rdma_req);
1036 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1037 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1038 			  0, 0, 0, 0, 0);
1039 	rdma_req.req.dif_enabled = true;
1040 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 2;
1041 	sgl->keyed.length = data_bs * 4;
1042 
1043 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1044 
1045 	CU_ASSERT(rc == 0);
1046 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1047 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1048 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1049 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1050 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1051 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1052 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1053 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1054 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1055 
1056 	for (i = 0; i < 2; ++i) {
1057 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1058 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1059 	}
1060 	for (i = 0; i < 2; ++i) {
1061 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].addr == 0x2000 + i * (data_bs + md_size));
1062 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].length == data_bs);
1063 	}
1064 
1065 	/* Part 6: simple I/O, one SGL larger than the transport io unit size, io_unit_size is not aligned to md_size,
1066 	   block size 512 */
1067 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1068 	reset_nvmf_rdma_request(&rdma_req);
1069 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1070 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1071 			  0, 0, 0, 0, 0);
1072 	rdma_req.req.dif_enabled = true;
1073 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1074 	sgl->keyed.length = data_bs * 6;
1075 
1076 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1077 
1078 	CU_ASSERT(rc == 0);
1079 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1080 	CU_ASSERT(rdma_req.req.length == data_bs * 6);
1081 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1082 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 6);
1083 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1084 	CU_ASSERT(rdma_req.data.wr.num_sge == 7);
1085 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1086 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1087 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1088 
1089 	for (i = 0; i < 3; ++i) {
1090 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1091 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1092 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1093 	}
1094 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
1095 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
1096 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
1097 
1098 	/* 2nd IO buffer consumed */
1099 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
1100 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
1101 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
1102 
1103 	CU_ASSERT(rdma_req.data.wr.sg_list[5].addr == 0x2000 + 24 + md_size);
1104 	CU_ASSERT(rdma_req.data.wr.sg_list[5].length == 512);
1105 	CU_ASSERT(rdma_req.data.wr.sg_list[5].lkey == RDMA_UT_LKEY);
1106 
1107 	CU_ASSERT(rdma_req.data.wr.sg_list[6].addr == 0x2000 + 24 + 512 + md_size * 2);
1108 	CU_ASSERT(rdma_req.data.wr.sg_list[6].length == 512);
1109 	CU_ASSERT(rdma_req.data.wr.sg_list[6].lkey == RDMA_UT_LKEY);
1110 
1111 	/* Part 7: simple I/O, number of SGL entries exceeds the number of entries
1112 	   one WR can hold. Additional WR is chained */
1113 	MOCK_SET(spdk_mempool_get, data2_buffer);
1114 	aligned_buffer = (void *)((uintptr_t)(data2_buffer + NVMF_DATA_BUFFER_MASK) &
1115 				  ~NVMF_DATA_BUFFER_MASK);
1116 	reset_nvmf_rdma_request(&rdma_req);
1117 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1118 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1119 			  0, 0, 0, 0, 0);
1120 	rdma_req.req.dif_enabled = true;
1121 	rtransport.transport.opts.io_unit_size = data_bs * 16;
1122 	sgl->keyed.length = data_bs * 16;
1123 
1124 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1125 
1126 	CU_ASSERT(rc == 0);
1127 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1128 	CU_ASSERT(rdma_req.req.length == data_bs * 16);
1129 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1130 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1131 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 16);
1132 	CU_ASSERT(rdma_req.req.data == aligned_buffer);
1133 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
1134 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1135 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1136 
1137 	for (i = 0; i < 15; ++i) {
1138 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)aligned_buffer + i * (data_bs + md_size));
1139 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1140 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1141 	}
1142 
1143 	/* 8192 - (512 + 8) * 15 = 392 */
1144 	CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)aligned_buffer + i * (data_bs + md_size));
1145 	CU_ASSERT(rdma_req.data.wr.sg_list[i].length == 392);
1146 	CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1147 
1148 	/* additional wr from pool */
1149 	CU_ASSERT(rdma_req.data.wr.next == (void *)&data2->wr);
1150 	CU_ASSERT(rdma_req.data.wr.next->num_sge == 1);
1151 	CU_ASSERT(rdma_req.data.wr.next->next == &rdma_req.rsp.wr);
1152 	/* 2nd IO buffer */
1153 	CU_ASSERT(data2->wr.sg_list[0].addr == (uintptr_t)aligned_buffer);
1154 	CU_ASSERT(data2->wr.sg_list[0].length == 120);
1155 	CU_ASSERT(data2->wr.sg_list[0].lkey == RDMA_UT_LKEY);
1156 
1157 	/* Part 8: simple I/O, data with metadata do not fit to 1 io_buffer */
1158 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1159 	reset_nvmf_rdma_request(&rdma_req);
1160 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1161 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1162 			  0, 0, 0, 0, 0);
1163 	rdma_req.req.dif_enabled = true;
1164 	rtransport.transport.opts.io_unit_size = 516;
1165 	sgl->keyed.length = data_bs * 2;
1166 
1167 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1168 
1169 	CU_ASSERT(rc == 0);
1170 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1171 	CU_ASSERT(rdma_req.req.length == data_bs * 2);
1172 	CU_ASSERT(rdma_req.req.iovcnt == 3);
1173 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1174 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 2);
1175 	CU_ASSERT(rdma_req.req.data == (void *)0x2000);
1176 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1177 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1178 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1179 
1180 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1181 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 512);
1182 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1183 
1184 	/* 2nd IO buffer consumed, offset 4 bytes due to part of the metadata
1185 	  is located at the beginning of that buffer */
1186 	CU_ASSERT(rdma_req.data.wr.sg_list[1].addr == 0x2000 + 4);
1187 	CU_ASSERT(rdma_req.data.wr.sg_list[1].length == 512);
1188 	CU_ASSERT(rdma_req.data.wr.sg_list[1].lkey == RDMA_UT_LKEY);
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_buffer);
1196 	aligned_buffer = (void *)((uintptr_t)(data_buffer + 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_enabled = 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 = 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 	CU_set_error_action(CUEA_ABORT);
1251 	CU_initialize_registry();
1252 
1253 	suite = CU_add_suite("nvmf", NULL, NULL);
1254 
1255 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl);
1256 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_process);
1257 	CU_ADD_TEST(suite, test_nvmf_rdma_get_optimal_poll_group);
1258 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl_with_md);
1259 
1260 	CU_basic_set_mode(CU_BRM_VERBOSE);
1261 	CU_basic_run_tests();
1262 	num_failures = CU_get_number_of_failures();
1263 	CU_cleanup_registry();
1264 	return num_failures;
1265 }
1266