xref: /spdk/test/unit/lib/nvmf/rdma.c/rdma_ut.c (revision ba8f1a9e5d59b6dfa9fd07c1300891b4e03722d9)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright (c) Intel Corporation. All rights reserved.
5  *   Copyright (c) 2019, 2021 Mellanox Technologies LTD. All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include "spdk/stdinc.h"
35 #include "spdk_cunit.h"
36 #include "common/lib/test_env.c"
37 #include "common/lib/test_rdma.c"
38 #include "nvmf/rdma.c"
39 #include "nvmf/transport.c"
40 
41 #define RDMA_UT_UNITS_IN_MAX_IO 16
42 
43 struct spdk_nvmf_transport_opts g_rdma_ut_transport_opts = {
44 	.max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH,
45 	.max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR,
46 	.in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE,
47 	.max_io_size = (SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE * RDMA_UT_UNITS_IN_MAX_IO),
48 	.io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE,
49 	.max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH,
50 	.num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS,
51 };
52 
53 SPDK_LOG_REGISTER_COMPONENT(nvmf)
54 DEFINE_STUB(spdk_mem_map_set_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
55 		uint64_t size, uint64_t translation), 0);
56 DEFINE_STUB(spdk_mem_map_clear_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
57 		uint64_t size), 0);
58 DEFINE_STUB(spdk_mem_map_alloc, struct spdk_mem_map *, (uint64_t default_translation,
59 		const struct spdk_mem_map_ops *ops, void *cb_ctx), NULL);
60 DEFINE_STUB(spdk_nvmf_qpair_disconnect, int, (struct spdk_nvmf_qpair *qpair,
61 		nvmf_qpair_disconnect_cb cb_fn, void *ctx), 0);
62 DEFINE_STUB(spdk_nvmf_qpair_get_listen_trid, int,
63 	    (struct spdk_nvmf_qpair *qpair, struct spdk_nvme_transport_id *trid), 0);
64 DEFINE_STUB_V(spdk_mem_map_free, (struct spdk_mem_map **pmap));
65 
66 DEFINE_STUB_V(spdk_nvmf_ctrlr_data_init, (struct spdk_nvmf_transport_opts *opts,
67 		struct spdk_nvmf_ctrlr_data *cdata));
68 DEFINE_STUB_V(spdk_nvmf_request_exec, (struct spdk_nvmf_request *req));
69 DEFINE_STUB(spdk_nvmf_request_complete, int, (struct spdk_nvmf_request *req), 0);
70 DEFINE_STUB(spdk_nvme_transport_id_compare, int, (const struct spdk_nvme_transport_id *trid1,
71 		const struct spdk_nvme_transport_id *trid2), 0);
72 DEFINE_STUB_V(nvmf_ctrlr_abort_aer, (struct spdk_nvmf_ctrlr *ctrlr));
73 DEFINE_STUB(spdk_nvmf_request_get_dif_ctx, bool, (struct spdk_nvmf_request *req,
74 		struct spdk_dif_ctx *dif_ctx), false);
75 DEFINE_STUB_V(spdk_nvme_trid_populate_transport, (struct spdk_nvme_transport_id *trid,
76 		enum spdk_nvme_transport_type trtype));
77 DEFINE_STUB_V(spdk_nvmf_tgt_new_qpair, (struct spdk_nvmf_tgt *tgt, struct spdk_nvmf_qpair *qpair));
78 DEFINE_STUB(nvmf_ctrlr_abort_request, int, (struct spdk_nvmf_request *req), 0);
79 DEFINE_STUB(spdk_nvme_transport_id_adrfam_str, const char *, (enum spdk_nvmf_adrfam adrfam), NULL);
80 DEFINE_STUB(ibv_dereg_mr, int, (struct ibv_mr *mr), 0);
81 
82 /* ibv_reg_mr can be a macro, need to undefine it */
83 #ifdef ibv_reg_mr
84 #undef ibv_reg_mr
85 #endif
86 
87 DEFINE_RETURN_MOCK(ibv_reg_mr, struct ibv_mr *);
88 struct ibv_mr *
89 ibv_reg_mr(struct ibv_pd *pd, void *addr, size_t length, int access)
90 {
91 	HANDLE_RETURN_MOCK(ibv_reg_mr);
92 	if (length > 0) {
93 		return &g_rdma_mr;
94 	} else {
95 		return NULL;
96 	}
97 }
98 
99 int ibv_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
100 		 int attr_mask, struct ibv_qp_init_attr *init_attr)
101 {
102 	if (qp == NULL) {
103 		return -1;
104 	} else {
105 		attr->port_num = 80;
106 
107 		if (qp->state == IBV_QPS_ERR) {
108 			attr->qp_state = 10;
109 		} else {
110 			attr->qp_state = IBV_QPS_INIT;
111 		}
112 
113 		return 0;
114 	}
115 }
116 
117 const char *
118 spdk_nvme_transport_id_trtype_str(enum spdk_nvme_transport_type trtype)
119 {
120 	switch (trtype) {
121 	case SPDK_NVME_TRANSPORT_PCIE:
122 		return "PCIe";
123 	case SPDK_NVME_TRANSPORT_RDMA:
124 		return "RDMA";
125 	case SPDK_NVME_TRANSPORT_FC:
126 		return "FC";
127 	default:
128 		return NULL;
129 	}
130 }
131 
132 int
133 spdk_nvme_transport_id_populate_trstring(struct spdk_nvme_transport_id *trid, const char *trstring)
134 {
135 	int len, i;
136 
137 	if (trstring == NULL) {
138 		return -EINVAL;
139 	}
140 
141 	len = strnlen(trstring, SPDK_NVMF_TRSTRING_MAX_LEN);
142 	if (len == SPDK_NVMF_TRSTRING_MAX_LEN) {
143 		return -EINVAL;
144 	}
145 
146 	/* cast official trstring to uppercase version of input. */
147 	for (i = 0; i < len; i++) {
148 		trid->trstring[i] = toupper(trstring[i]);
149 	}
150 	return 0;
151 }
152 
153 static void reset_nvmf_rdma_request(struct spdk_nvmf_rdma_request *rdma_req)
154 {
155 	int i;
156 
157 	rdma_req->req.length = 0;
158 	rdma_req->req.data_from_pool = false;
159 	rdma_req->req.data = NULL;
160 	rdma_req->data.wr.num_sge = 0;
161 	rdma_req->data.wr.wr.rdma.remote_addr = 0;
162 	rdma_req->data.wr.wr.rdma.rkey = 0;
163 	rdma_req->offset = 0;
164 	memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
165 
166 	for (i = 0; i < SPDK_NVMF_MAX_SGL_ENTRIES; i++) {
167 		rdma_req->req.iov[i].iov_base = 0;
168 		rdma_req->req.iov[i].iov_len = 0;
169 		rdma_req->req.buffers[i] = 0;
170 		rdma_req->data.wr.sg_list[i].addr = 0;
171 		rdma_req->data.wr.sg_list[i].length = 0;
172 		rdma_req->data.wr.sg_list[i].lkey = 0;
173 	}
174 	rdma_req->req.iovcnt = 0;
175 	if (rdma_req->req.stripped_data) {
176 		free(rdma_req->req.stripped_data);
177 		rdma_req->req.stripped_data = NULL;
178 	}
179 }
180 
181 static void
182 test_spdk_nvmf_rdma_request_parse_sgl(void)
183 {
184 	struct spdk_nvmf_rdma_transport rtransport;
185 	struct spdk_nvmf_rdma_device device;
186 	struct spdk_nvmf_rdma_request rdma_req = {};
187 	struct spdk_nvmf_rdma_recv recv;
188 	struct spdk_nvmf_rdma_poll_group group;
189 	struct spdk_nvmf_rdma_qpair rqpair;
190 	struct spdk_nvmf_rdma_poller poller;
191 	union nvmf_c2h_msg cpl;
192 	union nvmf_h2c_msg cmd;
193 	struct spdk_nvme_sgl_descriptor *sgl;
194 	struct spdk_nvmf_transport_pg_cache_buf bufs[4];
195 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
196 	struct spdk_nvmf_rdma_request_data data;
197 	int rc, i;
198 	uint32_t sgl_length;
199 	uintptr_t aligned_buffer_address;
200 
201 	data.wr.sg_list = data.sgl;
202 	STAILQ_INIT(&group.group.buf_cache);
203 	group.group.buf_cache_size = 0;
204 	group.group.buf_cache_count = 0;
205 	group.group.transport = &rtransport.transport;
206 	poller.group = &group;
207 	rqpair.poller = &poller;
208 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
209 
210 	sgl = &cmd.nvme_cmd.dptr.sgl1;
211 	rdma_req.recv = &recv;
212 	rdma_req.req.cmd = &cmd;
213 	rdma_req.req.rsp = &cpl;
214 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
215 	rdma_req.req.qpair = &rqpair.qpair;
216 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
217 
218 	rtransport.transport.opts = g_rdma_ut_transport_opts;
219 	rtransport.data_wr_pool = NULL;
220 	rtransport.transport.data_buf_pool = NULL;
221 
222 	device.attr.device_cap_flags = 0;
223 	sgl->keyed.key = 0xEEEE;
224 	sgl->address = 0xFFFF;
225 	rdma_req.recv->buf = (void *)0xDDDD;
226 
227 	/* Test 1: sgl type: keyed data block subtype: address */
228 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
229 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
230 
231 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size */
232 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
233 	reset_nvmf_rdma_request(&rdma_req);
234 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
235 
236 	device.map = (void *)0x0;
237 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
238 	CU_ASSERT(rc == 0);
239 	CU_ASSERT(rdma_req.req.data_from_pool == true);
240 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
241 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
242 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
243 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
244 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
245 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
246 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
247 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
248 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
249 
250 	/* Part 2: simple I/O, one SGL larger than the transport io unit size (equal to the max io size) */
251 	reset_nvmf_rdma_request(&rdma_req);
252 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
253 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
254 
255 	CU_ASSERT(rc == 0);
256 	CU_ASSERT(rdma_req.req.data_from_pool == true);
257 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO);
258 	CU_ASSERT(rdma_req.data.wr.num_sge == RDMA_UT_UNITS_IN_MAX_IO);
259 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
260 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
261 	for (i = 0; i < RDMA_UT_UNITS_IN_MAX_IO; i++) {
262 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
263 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
264 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
265 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
266 	}
267 
268 	/* Part 3: simple I/O one SGL larger than the transport max io size */
269 	reset_nvmf_rdma_request(&rdma_req);
270 	sgl->keyed.length = rtransport.transport.opts.max_io_size * 2;
271 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
272 
273 	CU_ASSERT(rc == -1);
274 
275 	/* Part 4: Pretend there are no buffer pools */
276 	MOCK_SET(spdk_mempool_get, NULL);
277 	reset_nvmf_rdma_request(&rdma_req);
278 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
279 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
280 
281 	CU_ASSERT(rc == 0);
282 	CU_ASSERT(rdma_req.req.data_from_pool == false);
283 	CU_ASSERT(rdma_req.req.data == NULL);
284 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
285 	CU_ASSERT(rdma_req.req.buffers[0] == NULL);
286 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0);
287 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 0);
288 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == 0);
289 
290 	rdma_req.recv->buf = (void *)0xDDDD;
291 	/* Test 2: sgl type: keyed data block subtype: offset (in capsule data) */
292 	sgl->generic.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
293 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
294 
295 	/* Part 1: Normal I/O smaller than in capsule data size no offset */
296 	reset_nvmf_rdma_request(&rdma_req);
297 	sgl->address = 0;
298 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
299 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
300 
301 	CU_ASSERT(rc == 0);
302 	CU_ASSERT(rdma_req.req.data == (void *)0xDDDD);
303 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.in_capsule_data_size);
304 	CU_ASSERT(rdma_req.req.data_from_pool == false);
305 
306 	/* Part 2: I/O offset + length too large */
307 	reset_nvmf_rdma_request(&rdma_req);
308 	sgl->address = rtransport.transport.opts.in_capsule_data_size;
309 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
310 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
311 
312 	CU_ASSERT(rc == -1);
313 
314 	/* Part 3: I/O too large */
315 	reset_nvmf_rdma_request(&rdma_req);
316 	sgl->address = 0;
317 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size * 2;
318 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
319 
320 	CU_ASSERT(rc == -1);
321 
322 	/* Test 3: Multi SGL */
323 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
324 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
325 	sgl->address = 0;
326 	rdma_req.recv->buf = (void *)&sgl_desc;
327 	MOCK_SET(spdk_mempool_get, &data);
328 
329 	/* part 1: 2 segments each with 1 wr. */
330 	reset_nvmf_rdma_request(&rdma_req);
331 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
332 	for (i = 0; i < 2; i++) {
333 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
334 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
335 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size;
336 		sgl_desc[i].address = 0x4000 + i * rtransport.transport.opts.io_unit_size;
337 		sgl_desc[i].keyed.key = 0x44;
338 	}
339 
340 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
341 
342 	CU_ASSERT(rc == 0);
343 	CU_ASSERT(rdma_req.req.data_from_pool == true);
344 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 2);
345 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
346 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
347 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
348 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
349 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
350 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size);
351 	CU_ASSERT(data.wr.num_sge == 1);
352 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
353 
354 	/* part 2: 2 segments, each with 1 wr containing 8 sge_elements */
355 	reset_nvmf_rdma_request(&rdma_req);
356 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
357 	for (i = 0; i < 2; i++) {
358 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
359 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
360 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size * 8;
361 		sgl_desc[i].address = 0x4000 + i * 8 * rtransport.transport.opts.io_unit_size;
362 		sgl_desc[i].keyed.key = 0x44;
363 	}
364 
365 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
366 
367 	CU_ASSERT(rc == 0);
368 	CU_ASSERT(rdma_req.req.data_from_pool == true);
369 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
370 	CU_ASSERT(rdma_req.req.iovcnt == 16);
371 	CU_ASSERT(rdma_req.data.wr.num_sge == 8);
372 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
373 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
374 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
375 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
376 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 8);
377 	CU_ASSERT(data.wr.num_sge == 8);
378 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
379 
380 	/* part 3: 2 segments, one very large, one very small */
381 	reset_nvmf_rdma_request(&rdma_req);
382 	for (i = 0; i < 2; i++) {
383 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
384 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
385 		sgl_desc[i].keyed.key = 0x44;
386 	}
387 
388 	sgl_desc[0].keyed.length = rtransport.transport.opts.io_unit_size * 15 +
389 				   rtransport.transport.opts.io_unit_size / 2;
390 	sgl_desc[0].address = 0x4000;
391 	sgl_desc[1].keyed.length = rtransport.transport.opts.io_unit_size / 2;
392 	sgl_desc[1].address = 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
393 			      rtransport.transport.opts.io_unit_size / 2;
394 
395 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
396 
397 	CU_ASSERT(rc == 0);
398 	CU_ASSERT(rdma_req.req.data_from_pool == true);
399 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
400 	CU_ASSERT(rdma_req.req.iovcnt == 16);
401 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
402 	for (i = 0; i < 15; i++) {
403 		CU_ASSERT(rdma_req.data.sgl[i].length == rtransport.transport.opts.io_unit_size);
404 	}
405 	CU_ASSERT(rdma_req.data.sgl[15].length == rtransport.transport.opts.io_unit_size / 2);
406 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
407 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
408 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
409 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
410 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
411 		  rtransport.transport.opts.io_unit_size / 2);
412 	CU_ASSERT(data.sgl[0].length == rtransport.transport.opts.io_unit_size / 2);
413 	CU_ASSERT(data.wr.num_sge == 1);
414 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
415 
416 	/* part 4: 2 SGL descriptors, each length is transport buffer / 2
417 	 * 1 transport buffers should be allocated */
418 	reset_nvmf_rdma_request(&rdma_req);
419 	aligned_buffer_address = ((uintptr_t)(&data) + NVMF_DATA_BUFFER_MASK) & ~NVMF_DATA_BUFFER_MASK;
420 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
421 	sgl_length = rtransport.transport.opts.io_unit_size / 2;
422 	for (i = 0; i < 2; i++) {
423 		sgl_desc[i].keyed.length = sgl_length;
424 		sgl_desc[i].address = 0x4000 + i * sgl_length;
425 	}
426 
427 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
428 
429 	CU_ASSERT(rc == 0);
430 	CU_ASSERT(rdma_req.req.data_from_pool == true);
431 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size);
432 	CU_ASSERT(rdma_req.req.iovcnt == 1);
433 
434 	CU_ASSERT(rdma_req.data.sgl[0].length == sgl_length);
435 	/* We mocked mempool_get to return address of data variable. Mempool is used
436 	 * to get both additional WRs and data buffers, so data points to &data */
437 	CU_ASSERT(rdma_req.data.sgl[0].addr == aligned_buffer_address);
438 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
439 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
440 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
441 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
442 
443 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
444 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + sgl_length);
445 	CU_ASSERT(data.sgl[0].length == sgl_length);
446 	CU_ASSERT(data.sgl[0].addr == aligned_buffer_address + sgl_length);
447 	CU_ASSERT(data.wr.num_sge == 1);
448 
449 	/* Test 4: use PG buffer cache */
450 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
451 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
452 	sgl->address = 0xFFFF;
453 	rdma_req.recv->buf = (void *)0xDDDD;
454 	sgl->keyed.key = 0xEEEE;
455 
456 	for (i = 0; i < 4; i++) {
457 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
458 	}
459 
460 	/* part 1: use the four buffers from the pg cache */
461 	group.group.buf_cache_size = 4;
462 	group.group.buf_cache_count = 4;
463 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
464 	reset_nvmf_rdma_request(&rdma_req);
465 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * 4;
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 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
478 	for (i = 0; i < 4; i++) {
479 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
480 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
481 				~NVMF_DATA_BUFFER_MASK));
482 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
483 	}
484 
485 	/* part 2: now that we have used the buffers from the cache, try again. We should get mempool buffers. */
486 	reset_nvmf_rdma_request(&rdma_req);
487 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
488 
489 	SPDK_CU_ASSERT_FATAL(rc == 0);
490 	CU_ASSERT(rdma_req.req.data_from_pool == true);
491 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
492 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
493 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
494 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
495 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
496 	CU_ASSERT(group.group.buf_cache_count == 0);
497 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
498 	for (i = 0; i < 4; i++) {
499 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
500 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
501 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
502 		CU_ASSERT(group.group.buf_cache_count == 0);
503 	}
504 
505 	/* part 3: half and half */
506 	group.group.buf_cache_count = 2;
507 
508 	for (i = 0; i < 2; i++) {
509 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
510 	}
511 	reset_nvmf_rdma_request(&rdma_req);
512 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
513 
514 	SPDK_CU_ASSERT_FATAL(rc == 0);
515 	CU_ASSERT(rdma_req.req.data_from_pool == true);
516 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
517 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
518 			~NVMF_DATA_BUFFER_MASK));
519 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
520 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
521 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
522 	CU_ASSERT(group.group.buf_cache_count == 0);
523 	for (i = 0; i < 2; i++) {
524 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
525 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
526 				~NVMF_DATA_BUFFER_MASK));
527 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
528 	}
529 	for (i = 2; i < 4; i++) {
530 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
531 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
532 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
533 	}
534 
535 	reset_nvmf_rdma_request(&rdma_req);
536 }
537 
538 static struct spdk_nvmf_rdma_recv *
539 create_recv(struct spdk_nvmf_rdma_qpair *rqpair, enum spdk_nvme_nvm_opcode opc)
540 {
541 	struct spdk_nvmf_rdma_recv *rdma_recv;
542 	union nvmf_h2c_msg *cmd;
543 	struct spdk_nvme_sgl_descriptor *sgl;
544 
545 	rdma_recv = calloc(1, sizeof(*rdma_recv));
546 	rdma_recv->qpair = rqpair;
547 	cmd = calloc(1, sizeof(*cmd));
548 	rdma_recv->sgl[0].addr = (uintptr_t)cmd;
549 	cmd->nvme_cmd.opc = opc;
550 	sgl = &cmd->nvme_cmd.dptr.sgl1;
551 	sgl->keyed.key = 0xEEEE;
552 	sgl->address = 0xFFFF;
553 	sgl->keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
554 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
555 	sgl->keyed.length = 1;
556 
557 	return rdma_recv;
558 }
559 
560 static void
561 free_recv(struct spdk_nvmf_rdma_recv *rdma_recv)
562 {
563 	free((void *)rdma_recv->sgl[0].addr);
564 	free(rdma_recv);
565 }
566 
567 static struct spdk_nvmf_rdma_request *
568 create_req(struct spdk_nvmf_rdma_qpair *rqpair,
569 	   struct spdk_nvmf_rdma_recv *rdma_recv)
570 {
571 	struct spdk_nvmf_rdma_request *rdma_req;
572 	union nvmf_c2h_msg *cpl;
573 
574 	rdma_req = calloc(1, sizeof(*rdma_req));
575 	rdma_req->recv = rdma_recv;
576 	rdma_req->req.qpair = &rqpair->qpair;
577 	rdma_req->state = RDMA_REQUEST_STATE_NEW;
578 	rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr;
579 	rdma_req->data.wr.sg_list = rdma_req->data.sgl;
580 	cpl = calloc(1, sizeof(*cpl));
581 	rdma_req->rsp.sgl[0].addr = (uintptr_t)cpl;
582 	rdma_req->req.rsp = cpl;
583 
584 	return rdma_req;
585 }
586 
587 static void
588 free_req(struct spdk_nvmf_rdma_request *rdma_req)
589 {
590 	free((void *)rdma_req->rsp.sgl[0].addr);
591 	free(rdma_req);
592 }
593 
594 static void
595 qpair_reset(struct spdk_nvmf_rdma_qpair *rqpair,
596 	    struct spdk_nvmf_rdma_poller *poller,
597 	    struct spdk_nvmf_rdma_device *device,
598 	    struct spdk_nvmf_rdma_resources *resources,
599 	    struct spdk_nvmf_transport *transport)
600 {
601 	memset(rqpair, 0, sizeof(*rqpair));
602 	STAILQ_INIT(&rqpair->pending_rdma_write_queue);
603 	STAILQ_INIT(&rqpair->pending_rdma_read_queue);
604 	rqpair->poller = poller;
605 	rqpair->device = device;
606 	rqpair->resources = resources;
607 	rqpair->qpair.qid = 1;
608 	rqpair->ibv_state = IBV_QPS_RTS;
609 	rqpair->qpair.state = SPDK_NVMF_QPAIR_ACTIVE;
610 	rqpair->max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
611 	rqpair->max_send_depth = 16;
612 	rqpair->max_read_depth = 16;
613 	rqpair->qpair.transport = transport;
614 }
615 
616 static void
617 poller_reset(struct spdk_nvmf_rdma_poller *poller,
618 	     struct spdk_nvmf_rdma_poll_group *group)
619 {
620 	memset(poller, 0, sizeof(*poller));
621 	STAILQ_INIT(&poller->qpairs_pending_recv);
622 	STAILQ_INIT(&poller->qpairs_pending_send);
623 	poller->group = group;
624 }
625 
626 static void
627 test_spdk_nvmf_rdma_request_process(void)
628 {
629 	struct spdk_nvmf_rdma_transport rtransport = {};
630 	struct spdk_nvmf_rdma_poll_group group = {};
631 	struct spdk_nvmf_rdma_poller poller = {};
632 	struct spdk_nvmf_rdma_device device = {};
633 	struct spdk_nvmf_rdma_resources resources = {};
634 	struct spdk_nvmf_rdma_qpair rqpair = {};
635 	struct spdk_nvmf_rdma_recv *rdma_recv;
636 	struct spdk_nvmf_rdma_request *rdma_req;
637 	bool progress;
638 
639 	STAILQ_INIT(&group.group.buf_cache);
640 	STAILQ_INIT(&group.group.pending_buf_queue);
641 	group.group.buf_cache_size = 0;
642 	group.group.buf_cache_count = 0;
643 	poller_reset(&poller, &group);
644 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
645 
646 	rtransport.transport.opts = g_rdma_ut_transport_opts;
647 	rtransport.transport.data_buf_pool = spdk_mempool_create("test_data_pool", 16, 128, 0, 0);
648 	rtransport.data_wr_pool = spdk_mempool_create("test_wr_pool", 128,
649 				  sizeof(struct spdk_nvmf_rdma_request_data),
650 				  0, 0);
651 	MOCK_CLEAR(spdk_mempool_get);
652 
653 	device.attr.device_cap_flags = 0;
654 	device.map = (void *)0x0;
655 
656 	/* Test 1: single SGL READ request */
657 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_READ);
658 	rdma_req = create_req(&rqpair, rdma_recv);
659 	rqpair.current_recv_depth = 1;
660 	/* NEW -> EXECUTING */
661 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
662 	CU_ASSERT(progress == true);
663 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
664 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST);
665 	/* EXECUTED -> TRANSFERRING_C2H */
666 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
667 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
668 	CU_ASSERT(progress == true);
669 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST);
670 	CU_ASSERT(rdma_req->recv == NULL);
671 	/* COMPLETED -> FREE */
672 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
673 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
674 	CU_ASSERT(progress == true);
675 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
676 
677 	free_recv(rdma_recv);
678 	free_req(rdma_req);
679 	poller_reset(&poller, &group);
680 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
681 
682 	/* Test 2: single SGL WRITE request */
683 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
684 	rdma_req = create_req(&rqpair, rdma_recv);
685 	rqpair.current_recv_depth = 1;
686 	/* NEW -> TRANSFERRING_H2C */
687 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
688 	CU_ASSERT(progress == true);
689 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
690 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
691 	STAILQ_INIT(&poller.qpairs_pending_send);
692 	/* READY_TO_EXECUTE -> EXECUTING */
693 	rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
694 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
695 	CU_ASSERT(progress == true);
696 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
697 	/* EXECUTED -> COMPLETING */
698 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
699 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
700 	CU_ASSERT(progress == true);
701 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
702 	CU_ASSERT(rdma_req->recv == NULL);
703 	/* COMPLETED -> FREE */
704 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
705 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
706 	CU_ASSERT(progress == true);
707 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
708 
709 	free_recv(rdma_recv);
710 	free_req(rdma_req);
711 	poller_reset(&poller, &group);
712 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
713 
714 	/* Test 3: WRITE+WRITE ibv_send batching */
715 	{
716 		struct spdk_nvmf_rdma_recv *recv1, *recv2;
717 		struct spdk_nvmf_rdma_request *req1, *req2;
718 		recv1 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
719 		req1 = create_req(&rqpair, recv1);
720 		recv2 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
721 		req2 = create_req(&rqpair, recv2);
722 
723 		/* WRITE 1: NEW -> TRANSFERRING_H2C */
724 		rqpair.current_recv_depth = 1;
725 		nvmf_rdma_request_process(&rtransport, req1);
726 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
727 
728 		/* WRITE 2: NEW -> TRANSFERRING_H2C */
729 		rqpair.current_recv_depth = 2;
730 		nvmf_rdma_request_process(&rtransport, req2);
731 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
732 
733 		STAILQ_INIT(&poller.qpairs_pending_send);
734 
735 		/* WRITE 1 completes before WRITE 2 has finished RDMA reading */
736 		/* WRITE 1: READY_TO_EXECUTE -> EXECUTING */
737 		req1->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
738 		nvmf_rdma_request_process(&rtransport, req1);
739 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_EXECUTING);
740 		/* WRITE 1: EXECUTED -> COMPLETING */
741 		req1->state = RDMA_REQUEST_STATE_EXECUTED;
742 		nvmf_rdma_request_process(&rtransport, req1);
743 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_COMPLETING);
744 		STAILQ_INIT(&poller.qpairs_pending_send);
745 		/* WRITE 1: COMPLETED -> FREE */
746 		req1->state = RDMA_REQUEST_STATE_COMPLETED;
747 		nvmf_rdma_request_process(&rtransport, req1);
748 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_FREE);
749 
750 		/* Now WRITE 2 has finished reading and completes */
751 		/* WRITE 2: COMPLETED -> FREE */
752 		/* WRITE 2: READY_TO_EXECUTE -> EXECUTING */
753 		req2->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
754 		nvmf_rdma_request_process(&rtransport, req2);
755 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_EXECUTING);
756 		/* WRITE 1: EXECUTED -> COMPLETING */
757 		req2->state = RDMA_REQUEST_STATE_EXECUTED;
758 		nvmf_rdma_request_process(&rtransport, req2);
759 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_COMPLETING);
760 		STAILQ_INIT(&poller.qpairs_pending_send);
761 		/* WRITE 1: COMPLETED -> FREE */
762 		req2->state = RDMA_REQUEST_STATE_COMPLETED;
763 		nvmf_rdma_request_process(&rtransport, req2);
764 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_FREE);
765 
766 		free_recv(recv1);
767 		free_req(req1);
768 		free_recv(recv2);
769 		free_req(req2);
770 		poller_reset(&poller, &group);
771 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
772 	}
773 
774 	/* Test 4, invalid command, check xfer type */
775 	{
776 		struct spdk_nvmf_rdma_recv *rdma_recv_inv;
777 		struct spdk_nvmf_rdma_request *rdma_req_inv;
778 		/* construct an opcode that specifies BIDIRECTIONAL transfer */
779 		uint8_t opc = 0x10 | SPDK_NVME_DATA_BIDIRECTIONAL;
780 
781 		rdma_recv_inv = create_recv(&rqpair, opc);
782 		rdma_req_inv = create_req(&rqpair, rdma_recv_inv);
783 
784 		/* NEW -> RDMA_REQUEST_STATE_COMPLETING */
785 		rqpair.current_recv_depth = 1;
786 		progress = nvmf_rdma_request_process(&rtransport, rdma_req_inv);
787 		CU_ASSERT(progress == true);
788 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_COMPLETING);
789 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sct == SPDK_NVME_SCT_GENERIC);
790 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sc == SPDK_NVME_SC_INVALID_OPCODE);
791 
792 		/* RDMA_REQUEST_STATE_COMPLETED -> FREE */
793 		rdma_req_inv->state = RDMA_REQUEST_STATE_COMPLETED;
794 		nvmf_rdma_request_process(&rtransport, rdma_req_inv);
795 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_FREE);
796 
797 		free_recv(rdma_recv_inv);
798 		free_req(rdma_req_inv);
799 		poller_reset(&poller, &group);
800 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
801 	}
802 
803 	spdk_mempool_free(rtransport.transport.data_buf_pool);
804 	spdk_mempool_free(rtransport.data_wr_pool);
805 }
806 
807 #define TEST_GROUPS_COUNT 5
808 static void
809 test_nvmf_rdma_get_optimal_poll_group(void)
810 {
811 	struct spdk_nvmf_rdma_transport rtransport = {};
812 	struct spdk_nvmf_transport *transport = &rtransport.transport;
813 	struct spdk_nvmf_rdma_qpair rqpair = {};
814 	struct spdk_nvmf_transport_poll_group *groups[TEST_GROUPS_COUNT];
815 	struct spdk_nvmf_rdma_poll_group *rgroups[TEST_GROUPS_COUNT];
816 	struct spdk_nvmf_transport_poll_group *result;
817 	uint32_t i;
818 
819 	rqpair.qpair.transport = transport;
820 	pthread_mutex_init(&rtransport.lock, NULL);
821 	TAILQ_INIT(&rtransport.poll_groups);
822 
823 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
824 		groups[i] = nvmf_rdma_poll_group_create(transport, NULL);
825 		CU_ASSERT(groups[i] != NULL);
826 		rgroups[i] = SPDK_CONTAINEROF(groups[i], struct spdk_nvmf_rdma_poll_group, group);
827 		groups[i]->transport = transport;
828 	}
829 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[0]);
830 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[0]);
831 
832 	/* Emulate connection of %TEST_GROUPS_COUNT% initiators - each creates 1 admin and 1 io qp */
833 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
834 		rqpair.qpair.qid = 0;
835 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
836 		CU_ASSERT(result == groups[i]);
837 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
838 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i]);
839 
840 		rqpair.qpair.qid = 1;
841 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
842 		CU_ASSERT(result == groups[i]);
843 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
844 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
845 	}
846 	/* wrap around, admin/io pg point to the first pg
847 	   Destroy all poll groups except of the last one */
848 	for (i = 0; i < TEST_GROUPS_COUNT - 1; i++) {
849 		nvmf_rdma_poll_group_destroy(groups[i]);
850 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[i + 1]);
851 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i + 1]);
852 	}
853 
854 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
855 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
856 
857 	/* Check that pointers to the next admin/io poll groups are not changed */
858 	rqpair.qpair.qid = 0;
859 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
860 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
861 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
862 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
863 
864 	rqpair.qpair.qid = 1;
865 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
866 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
867 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
868 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
869 
870 	/* Remove the last poll group, check that pointers are NULL */
871 	nvmf_rdma_poll_group_destroy(groups[TEST_GROUPS_COUNT - 1]);
872 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == NULL);
873 	CU_ASSERT(rtransport.conn_sched.next_io_pg == NULL);
874 
875 	/* Request optimal poll group, result must be NULL */
876 	rqpair.qpair.qid = 0;
877 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
878 	CU_ASSERT(result == NULL);
879 
880 	rqpair.qpair.qid = 1;
881 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
882 	CU_ASSERT(result == NULL);
883 
884 	pthread_mutex_destroy(&rtransport.lock);
885 }
886 #undef TEST_GROUPS_COUNT
887 
888 static void
889 test_spdk_nvmf_rdma_request_parse_sgl_with_md(void)
890 {
891 	struct spdk_nvmf_rdma_transport rtransport;
892 	struct spdk_nvmf_rdma_device device;
893 	struct spdk_nvmf_rdma_request rdma_req = {};
894 	struct spdk_nvmf_rdma_recv recv;
895 	struct spdk_nvmf_rdma_poll_group group;
896 	struct spdk_nvmf_rdma_qpair rqpair;
897 	struct spdk_nvmf_rdma_poller poller;
898 	union nvmf_c2h_msg cpl;
899 	union nvmf_h2c_msg cmd;
900 	struct spdk_nvme_sgl_descriptor *sgl;
901 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
902 	char data_buffer[8192];
903 	struct spdk_nvmf_rdma_request_data *data = (struct spdk_nvmf_rdma_request_data *)data_buffer;
904 	char data2_buffer[8192];
905 	struct spdk_nvmf_rdma_request_data *data2 = (struct spdk_nvmf_rdma_request_data *)data2_buffer;
906 	const uint32_t data_bs = 512;
907 	const uint32_t md_size = 8;
908 	int rc, i;
909 	void *aligned_buffer;
910 
911 	data->wr.sg_list = data->sgl;
912 	STAILQ_INIT(&group.group.buf_cache);
913 	group.group.buf_cache_size = 0;
914 	group.group.buf_cache_count = 0;
915 	group.group.transport = &rtransport.transport;
916 	poller.group = &group;
917 	rqpair.poller = &poller;
918 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
919 
920 	sgl = &cmd.nvme_cmd.dptr.sgl1;
921 	rdma_req.recv = &recv;
922 	rdma_req.req.cmd = &cmd;
923 	rdma_req.req.rsp = &cpl;
924 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
925 	rdma_req.req.qpair = &rqpair.qpair;
926 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
927 
928 	rtransport.transport.opts = g_rdma_ut_transport_opts;
929 	rtransport.data_wr_pool = NULL;
930 	rtransport.transport.data_buf_pool = NULL;
931 
932 	device.attr.device_cap_flags = 0;
933 	device.map = NULL;
934 	sgl->keyed.key = 0xEEEE;
935 	sgl->address = 0xFFFF;
936 	rdma_req.recv->buf = (void *)0xDDDD;
937 
938 	/* Test 1: sgl type: keyed data block subtype: address */
939 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
940 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
941 
942 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size, block size 512 */
943 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
944 	reset_nvmf_rdma_request(&rdma_req);
945 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
946 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
947 			  0, 0, 0, 0, 0);
948 	rdma_req.req.dif_enabled = true;
949 	rtransport.transport.opts.io_unit_size = data_bs * 8;
950 	rdma_req.req.qpair->transport = &rtransport.transport;
951 	sgl->keyed.length = data_bs * 4;
952 
953 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
954 
955 	CU_ASSERT(rc == 0);
956 	CU_ASSERT(rdma_req.req.data_from_pool == true);
957 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
958 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
959 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
960 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
961 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
962 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
963 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
964 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
965 
966 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
967 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rdma_req.req.length);
968 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
969 
970 	/* Part 2: simple I/O, one SGL equal to io unit size, io_unit_size is not aligned with md_size,
971 		block size 512 */
972 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
973 	reset_nvmf_rdma_request(&rdma_req);
974 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
975 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
976 			  0, 0, 0, 0, 0);
977 	rdma_req.req.dif_enabled = true;
978 	rtransport.transport.opts.io_unit_size = data_bs * 4;
979 	sgl->keyed.length = data_bs * 4;
980 
981 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
982 
983 	CU_ASSERT(rc == 0);
984 	CU_ASSERT(rdma_req.req.data_from_pool == true);
985 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
986 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
987 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
988 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
989 	CU_ASSERT(rdma_req.data.wr.num_sge == 5);
990 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
991 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
992 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
993 
994 	for (i = 0; i < 3; ++i) {
995 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
996 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
997 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
998 	}
999 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
1000 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
1001 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
1002 
1003 	/* 2nd buffer consumed */
1004 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
1005 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
1006 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
1007 
1008 	/* Part 3: simple I/O, one SGL equal io unit size, io_unit_size is equal to block size 512 bytes */
1009 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1010 	reset_nvmf_rdma_request(&rdma_req);
1011 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1012 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1013 			  0, 0, 0, 0, 0);
1014 	rdma_req.req.dif_enabled = true;
1015 	rtransport.transport.opts.io_unit_size = data_bs;
1016 	sgl->keyed.length = data_bs;
1017 
1018 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1019 
1020 	CU_ASSERT(rc == 0);
1021 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1022 	CU_ASSERT(rdma_req.req.length == data_bs);
1023 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1024 	CU_ASSERT(rdma_req.req.dif.elba_length == data_bs + md_size);
1025 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1026 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1027 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1028 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1029 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1030 
1031 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1032 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs);
1033 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1034 
1035 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1036 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)((unsigned long)0x2000));
1037 	CU_ASSERT(rdma_req.req.iov[0].iov_len == data_bs);
1038 	/* 2nd buffer consumed for metadata */
1039 	CU_ASSERT(rdma_req.req.iov[1].iov_base == (void *)((unsigned long)0x2000));
1040 	CU_ASSERT(rdma_req.req.iov[1].iov_len == md_size);
1041 
1042 	/* Part 4: simple I/O, one SGL equal io unit size, io_unit_size is aligned with md_size,
1043 	   block size 512 */
1044 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1045 	reset_nvmf_rdma_request(&rdma_req);
1046 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1047 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1048 			  0, 0, 0, 0, 0);
1049 	rdma_req.req.dif_enabled = true;
1050 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1051 	sgl->keyed.length = data_bs * 4;
1052 
1053 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1054 
1055 	CU_ASSERT(rc == 0);
1056 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1057 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1058 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1059 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1060 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1061 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1062 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1063 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1064 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1065 
1066 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1067 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rdma_req.req.length);
1068 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1069 
1070 	/* Part 5: simple I/O, one SGL equal to 2x io unit size, io_unit_size is aligned with md_size,
1071 	   block size 512 */
1072 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1073 	reset_nvmf_rdma_request(&rdma_req);
1074 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1075 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1076 			  0, 0, 0, 0, 0);
1077 	rdma_req.req.dif_enabled = true;
1078 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 2;
1079 	sgl->keyed.length = data_bs * 4;
1080 
1081 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1082 
1083 	CU_ASSERT(rc == 0);
1084 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1085 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1086 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1087 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1088 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1089 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1090 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1091 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1092 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1093 
1094 	for (i = 0; i < 2; ++i) {
1095 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
1096 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs * 2);
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_enabled = 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 == RDMA_UT_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 == RDMA_UT_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 == RDMA_UT_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 == RDMA_UT_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 == RDMA_UT_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_enabled = 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 
1171 	for (i = 0; i < 15; ++i) {
1172 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)aligned_buffer + i * (data_bs + md_size));
1173 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1174 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1175 	}
1176 
1177 	/* 8192 - (512 + 8) * 15 = 392 */
1178 	CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)aligned_buffer + i * (data_bs + md_size));
1179 	CU_ASSERT(rdma_req.data.wr.sg_list[i].length == 392);
1180 	CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1181 
1182 	/* additional wr from pool */
1183 	CU_ASSERT(rdma_req.data.wr.next == (void *)&data2->wr);
1184 	CU_ASSERT(rdma_req.data.wr.next->num_sge == 1);
1185 	CU_ASSERT(rdma_req.data.wr.next->next == &rdma_req.rsp.wr);
1186 	/* 2nd IO buffer */
1187 	CU_ASSERT(data2->wr.sg_list[0].addr == (uintptr_t)aligned_buffer);
1188 	CU_ASSERT(data2->wr.sg_list[0].length == 120);
1189 	CU_ASSERT(data2->wr.sg_list[0].lkey == RDMA_UT_LKEY);
1190 
1191 	/* Part 8: simple I/O, data with metadata do not fit to 1 io_buffer */
1192 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1193 	reset_nvmf_rdma_request(&rdma_req);
1194 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1195 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1196 			  0, 0, 0, 0, 0);
1197 	rdma_req.req.dif_enabled = true;
1198 	rtransport.transport.opts.io_unit_size = 516;
1199 	sgl->keyed.length = data_bs * 2;
1200 
1201 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1202 
1203 	CU_ASSERT(rc == 0);
1204 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1205 	CU_ASSERT(rdma_req.req.length == data_bs * 2);
1206 	CU_ASSERT(rdma_req.req.iovcnt == 3);
1207 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1208 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 2);
1209 	CU_ASSERT(rdma_req.req.data == (void *)0x2000);
1210 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1211 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1212 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1213 
1214 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1215 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 512);
1216 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1217 
1218 	/* 2nd IO buffer consumed, offset 4 bytes due to part of the metadata
1219 	  is located at the beginning of that buffer */
1220 	CU_ASSERT(rdma_req.data.wr.sg_list[1].addr == 0x2000 + 4);
1221 	CU_ASSERT(rdma_req.data.wr.sg_list[1].length == 512);
1222 	CU_ASSERT(rdma_req.data.wr.sg_list[1].lkey == RDMA_UT_LKEY);
1223 
1224 	/* Test 2: Multi SGL */
1225 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1226 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1227 	sgl->address = 0;
1228 	rdma_req.recv->buf = (void *)&sgl_desc;
1229 	MOCK_SET(spdk_mempool_get, data_buffer);
1230 	aligned_buffer = (void *)((uintptr_t)(data_buffer + NVMF_DATA_BUFFER_MASK) &
1231 				  ~NVMF_DATA_BUFFER_MASK);
1232 
1233 	/* part 1: 2 segments each with 1 wr. io_unit_size is aligned with data_bs + md_size */
1234 	reset_nvmf_rdma_request(&rdma_req);
1235 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1236 			  SPDK_DIF_TYPE1,
1237 			  SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 0, 0, 0, 0, 0);
1238 	rdma_req.req.dif_enabled = true;
1239 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1240 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
1241 
1242 	for (i = 0; i < 2; i++) {
1243 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1244 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1245 		sgl_desc[i].keyed.length = data_bs * 4;
1246 		sgl_desc[i].address = 0x4000 + i * data_bs * 4;
1247 		sgl_desc[i].keyed.key = 0x44;
1248 	}
1249 
1250 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1251 
1252 	CU_ASSERT(rc == 0);
1253 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1254 	CU_ASSERT(rdma_req.req.length == data_bs * 4 * 2);
1255 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1256 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4 * 2);
1257 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1258 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == (uintptr_t)(aligned_buffer));
1259 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs * 4);
1260 
1261 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
1262 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
1263 	CU_ASSERT(rdma_req.data.wr.next == &data->wr);
1264 	CU_ASSERT(data->wr.wr.rdma.rkey == 0x44);
1265 	CU_ASSERT(data->wr.wr.rdma.remote_addr == 0x4000 + data_bs * 4);
1266 	CU_ASSERT(data->wr.num_sge == 1);
1267 	CU_ASSERT(data->wr.sg_list[0].addr == (uintptr_t)(aligned_buffer));
1268 	CU_ASSERT(data->wr.sg_list[0].length == data_bs * 4);
1269 
1270 	CU_ASSERT(data->wr.next == &rdma_req.rsp.wr);
1271 	reset_nvmf_rdma_request(&rdma_req);
1272 }
1273 
1274 static void
1275 test_nvmf_rdma_opts_init(void)
1276 {
1277 	struct spdk_nvmf_transport_opts	opts = {};
1278 
1279 	nvmf_rdma_opts_init(&opts);
1280 	CU_ASSERT(opts.max_queue_depth == SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH);
1281 	CU_ASSERT(opts.max_qpairs_per_ctrlr ==	SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR);
1282 	CU_ASSERT(opts.in_capsule_data_size ==	SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE);
1283 	CU_ASSERT(opts.max_io_size == SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE);
1284 	CU_ASSERT(opts.io_unit_size == SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
1285 	CU_ASSERT(opts.max_aq_depth == SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH);
1286 	CU_ASSERT(opts.num_shared_buffers == SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS);
1287 	CU_ASSERT(opts.buf_cache_size == SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE);
1288 	CU_ASSERT(opts.dif_insert_or_strip == SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP);
1289 	CU_ASSERT(opts.abort_timeout_sec == SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC);
1290 	CU_ASSERT(opts.transport_specific == NULL);
1291 }
1292 
1293 static void
1294 test_nvmf_rdma_request_free_data(void)
1295 {
1296 	struct spdk_nvmf_rdma_request rdma_req = {};
1297 	struct spdk_nvmf_rdma_transport rtransport = {};
1298 	struct spdk_nvmf_rdma_request_data *next_request_data = NULL;
1299 
1300 	MOCK_CLEAR(spdk_mempool_get);
1301 	rtransport.data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data",
1302 				  SPDK_NVMF_MAX_SGL_ENTRIES,
1303 				  sizeof(struct spdk_nvmf_rdma_request_data),
1304 				  SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
1305 				  SPDK_ENV_SOCKET_ID_ANY);
1306 	next_request_data = spdk_mempool_get(rtransport.data_wr_pool);
1307 	SPDK_CU_ASSERT_FATAL(((struct test_mempool *)rtransport.data_wr_pool)->count ==
1308 			     SPDK_NVMF_MAX_SGL_ENTRIES - 1);
1309 	next_request_data->wr.wr_id = 1;
1310 	next_request_data->wr.num_sge = 2;
1311 	next_request_data->wr.next = NULL;
1312 	rdma_req.data.wr.next = &next_request_data->wr;
1313 	rdma_req.data.wr.wr_id = 1;
1314 	rdma_req.data.wr.num_sge = 2;
1315 
1316 	nvmf_rdma_request_free_data(&rdma_req, &rtransport);
1317 	/* Check if next_request_data put into memory pool */
1318 	CU_ASSERT(((struct test_mempool *)rtransport.data_wr_pool)->count == SPDK_NVMF_MAX_SGL_ENTRIES);
1319 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
1320 
1321 	spdk_mempool_free(rtransport.data_wr_pool);
1322 }
1323 
1324 static void
1325 test_nvmf_rdma_update_ibv_state(void)
1326 {
1327 	struct spdk_nvmf_rdma_qpair rqpair = {};
1328 	struct spdk_rdma_qp rdma_qp = {};
1329 	struct ibv_qp qp = {};
1330 	int rc = 0;
1331 
1332 	rqpair.rdma_qp = &rdma_qp;
1333 
1334 	/* Case 1: Failed to get updated RDMA queue pair state */
1335 	rqpair.ibv_state = IBV_QPS_INIT;
1336 	rqpair.rdma_qp->qp = NULL;
1337 
1338 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1339 	CU_ASSERT(rc == IBV_QPS_ERR + 1);
1340 
1341 	/* Case 2: Bad state updated */
1342 	rqpair.rdma_qp->qp = &qp;
1343 	qp.state = IBV_QPS_ERR;
1344 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1345 	CU_ASSERT(rqpair.ibv_state == 10);
1346 	CU_ASSERT(rc == IBV_QPS_ERR + 1);
1347 
1348 	/* Case 3: Pass */
1349 	qp.state = IBV_QPS_INIT;
1350 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1351 	CU_ASSERT(rqpair.ibv_state == IBV_QPS_INIT);
1352 	CU_ASSERT(rc == IBV_QPS_INIT);
1353 }
1354 
1355 static void
1356 test_nvmf_rdma_resources_create(void)
1357 {
1358 	static struct spdk_nvmf_rdma_resources *rdma_resource;
1359 	struct spdk_nvmf_rdma_resource_opts opts = {};
1360 	struct spdk_nvmf_rdma_qpair qpair = {};
1361 	struct spdk_nvmf_rdma_recv *recv = NULL;
1362 	struct spdk_nvmf_rdma_request *req = NULL;
1363 	const int DEPTH = 128;
1364 
1365 	opts.max_queue_depth = DEPTH;
1366 	opts.in_capsule_data_size = 4096;
1367 	opts.shared = true;
1368 	opts.qpair = &qpair;
1369 
1370 	rdma_resource = nvmf_rdma_resources_create(&opts);
1371 	CU_ASSERT(rdma_resource != NULL);
1372 	/* Just check first and last entry */
1373 	recv = &rdma_resource->recvs[0];
1374 	req = &rdma_resource->reqs[0];
1375 	CU_ASSERT(recv->rdma_wr.type == RDMA_WR_TYPE_RECV);
1376 	CU_ASSERT((uintptr_t)recv->buf == (uintptr_t)(rdma_resource->bufs));
1377 	CU_ASSERT(recv->sgl[0].addr == (uintptr_t)&rdma_resource->cmds[0]);
1378 	CU_ASSERT(recv->sgl[0].length == sizeof(rdma_resource->cmds[0]));
1379 	CU_ASSERT(recv->sgl[0].lkey == RDMA_UT_LKEY);
1380 	CU_ASSERT(recv->wr.num_sge == 2);
1381 	CU_ASSERT(recv->wr.wr_id == (uintptr_t)&rdma_resource->recvs[0].rdma_wr);
1382 	CU_ASSERT(recv->wr.sg_list == rdma_resource->recvs[0].sgl);
1383 	CU_ASSERT(req->req.rsp == &rdma_resource->cpls[0]);
1384 	CU_ASSERT(req->rsp.sgl[0].addr == (uintptr_t)&rdma_resource->cpls[0]);
1385 	CU_ASSERT(req->rsp.sgl[0].length == sizeof(rdma_resource->cpls[0]));
1386 	CU_ASSERT(req->rsp.sgl[0].lkey == RDMA_UT_LKEY);
1387 	CU_ASSERT(req->rsp.rdma_wr.type == RDMA_WR_TYPE_SEND);
1388 	CU_ASSERT(req->rsp.wr.wr_id == (uintptr_t)&rdma_resource->reqs[0].rsp.rdma_wr);
1389 	CU_ASSERT(req->rsp.wr.next == NULL);
1390 	CU_ASSERT(req->rsp.wr.opcode == IBV_WR_SEND);
1391 	CU_ASSERT(req->rsp.wr.send_flags == IBV_SEND_SIGNALED);
1392 	CU_ASSERT(req->rsp.wr.sg_list == rdma_resource->reqs[0].rsp.sgl);
1393 	CU_ASSERT(req->rsp.wr.num_sge == NVMF_DEFAULT_RSP_SGE);
1394 	CU_ASSERT(req->data.rdma_wr.type == RDMA_WR_TYPE_DATA);
1395 	CU_ASSERT(req->data.wr.wr_id == (uintptr_t)&rdma_resource->reqs[0].data.rdma_wr);
1396 	CU_ASSERT(req->data.wr.next == NULL);
1397 	CU_ASSERT(req->data.wr.send_flags == IBV_SEND_SIGNALED);
1398 	CU_ASSERT(req->data.wr.sg_list == rdma_resource->reqs[0].data.sgl);
1399 	CU_ASSERT(req->data.wr.num_sge == SPDK_NVMF_MAX_SGL_ENTRIES);
1400 	CU_ASSERT(req->state == RDMA_REQUEST_STATE_FREE);
1401 
1402 	recv = &rdma_resource->recvs[DEPTH - 1];
1403 	req = &rdma_resource->reqs[DEPTH - 1];
1404 	CU_ASSERT(recv->rdma_wr.type == RDMA_WR_TYPE_RECV);
1405 	CU_ASSERT((uintptr_t)recv->buf == (uintptr_t)(rdma_resource->bufs +
1406 			(DEPTH - 1) * 4096));
1407 	CU_ASSERT(recv->sgl[0].addr == (uintptr_t)&rdma_resource->cmds[DEPTH - 1]);
1408 	CU_ASSERT(recv->sgl[0].length == sizeof(rdma_resource->cmds[DEPTH - 1]));
1409 	CU_ASSERT(recv->sgl[0].lkey == RDMA_UT_LKEY);
1410 	CU_ASSERT(recv->wr.num_sge == 2);
1411 	CU_ASSERT(recv->wr.wr_id == (uintptr_t)&rdma_resource->recvs[DEPTH - 1].rdma_wr);
1412 	CU_ASSERT(recv->wr.sg_list == rdma_resource->recvs[DEPTH - 1].sgl);
1413 	CU_ASSERT(req->req.rsp == &rdma_resource->cpls[DEPTH - 1]);
1414 	CU_ASSERT(req->rsp.sgl[0].addr == (uintptr_t)&rdma_resource->cpls[DEPTH - 1]);
1415 	CU_ASSERT(req->rsp.sgl[0].length == sizeof(rdma_resource->cpls[DEPTH - 1]));
1416 	CU_ASSERT(req->rsp.sgl[0].lkey == RDMA_UT_LKEY);
1417 	CU_ASSERT(req->rsp.rdma_wr.type == RDMA_WR_TYPE_SEND);
1418 	CU_ASSERT(req->rsp.wr.wr_id == (uintptr_t)
1419 		  &req->rsp.rdma_wr);
1420 	CU_ASSERT(req->rsp.wr.next == NULL);
1421 	CU_ASSERT(req->rsp.wr.opcode == IBV_WR_SEND);
1422 	CU_ASSERT(req->rsp.wr.send_flags == IBV_SEND_SIGNALED);
1423 	CU_ASSERT(req->rsp.wr.sg_list == rdma_resource->reqs[DEPTH - 1].rsp.sgl);
1424 	CU_ASSERT(req->rsp.wr.num_sge == NVMF_DEFAULT_RSP_SGE);
1425 	CU_ASSERT(req->data.rdma_wr.type == RDMA_WR_TYPE_DATA);
1426 	CU_ASSERT(req->data.wr.wr_id == (uintptr_t)
1427 		  &req->data.rdma_wr);
1428 	CU_ASSERT(req->data.wr.next == NULL);
1429 	CU_ASSERT(req->data.wr.send_flags == IBV_SEND_SIGNALED);
1430 	CU_ASSERT(req->data.wr.sg_list == rdma_resource->reqs[DEPTH - 1].data.sgl);
1431 	CU_ASSERT(req->data.wr.num_sge == SPDK_NVMF_MAX_SGL_ENTRIES);
1432 	CU_ASSERT(req->state == RDMA_REQUEST_STATE_FREE);
1433 
1434 	nvmf_rdma_resources_destroy(rdma_resource);
1435 }
1436 
1437 static void
1438 test_nvmf_rdma_qpair_compare(void)
1439 {
1440 	struct spdk_nvmf_rdma_qpair rqpair1 = {}, rqpair2 = {};
1441 
1442 	rqpair1.qp_num = 0;
1443 	rqpair2.qp_num = UINT32_MAX;
1444 
1445 	CU_ASSERT(nvmf_rdma_qpair_compare(&rqpair1, &rqpair2) < 0);
1446 	CU_ASSERT(nvmf_rdma_qpair_compare(&rqpair2, &rqpair1) > 0);
1447 }
1448 
1449 int main(int argc, char **argv)
1450 {
1451 	CU_pSuite	suite = NULL;
1452 	unsigned int	num_failures;
1453 
1454 	CU_set_error_action(CUEA_ABORT);
1455 	CU_initialize_registry();
1456 
1457 	suite = CU_add_suite("nvmf", NULL, NULL);
1458 
1459 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl);
1460 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_process);
1461 	CU_ADD_TEST(suite, test_nvmf_rdma_get_optimal_poll_group);
1462 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl_with_md);
1463 	CU_ADD_TEST(suite, test_nvmf_rdma_opts_init);
1464 	CU_ADD_TEST(suite, test_nvmf_rdma_request_free_data);
1465 	CU_ADD_TEST(suite, test_nvmf_rdma_update_ibv_state);
1466 	CU_ADD_TEST(suite, test_nvmf_rdma_resources_create);
1467 	CU_ADD_TEST(suite, test_nvmf_rdma_qpair_compare);
1468 
1469 	CU_basic_set_mode(CU_BRM_VERBOSE);
1470 	CU_basic_run_tests();
1471 	num_failures = CU_get_number_of_failures();
1472 	CU_cleanup_registry();
1473 	return num_failures;
1474 }
1475