xref: /spdk/test/unit/lib/nvmf/rdma.c/rdma_ut.c (revision 49c6afbf128a130cd17d695170723fa4ff7699d0)
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 int ibv_query_qp(struct ibv_qp *qp, struct ibv_qp_attr *attr,
81 		 int attr_mask, struct ibv_qp_init_attr *init_attr)
82 {
83 	if (qp == NULL) {
84 		return -1;
85 	} else {
86 		attr->port_num = 80;
87 
88 		if (qp->state == IBV_QPS_ERR) {
89 			attr->qp_state = 10;
90 		} else {
91 			attr->qp_state = IBV_QPS_INIT;
92 		}
93 
94 		return 0;
95 	}
96 }
97 
98 const char *
99 spdk_nvme_transport_id_trtype_str(enum spdk_nvme_transport_type trtype)
100 {
101 	switch (trtype) {
102 	case SPDK_NVME_TRANSPORT_PCIE:
103 		return "PCIe";
104 	case SPDK_NVME_TRANSPORT_RDMA:
105 		return "RDMA";
106 	case SPDK_NVME_TRANSPORT_FC:
107 		return "FC";
108 	default:
109 		return NULL;
110 	}
111 }
112 
113 int
114 spdk_nvme_transport_id_populate_trstring(struct spdk_nvme_transport_id *trid, const char *trstring)
115 {
116 	int len, i;
117 
118 	if (trstring == NULL) {
119 		return -EINVAL;
120 	}
121 
122 	len = strnlen(trstring, SPDK_NVMF_TRSTRING_MAX_LEN);
123 	if (len == SPDK_NVMF_TRSTRING_MAX_LEN) {
124 		return -EINVAL;
125 	}
126 
127 	/* cast official trstring to uppercase version of input. */
128 	for (i = 0; i < len; i++) {
129 		trid->trstring[i] = toupper(trstring[i]);
130 	}
131 	return 0;
132 }
133 
134 static void reset_nvmf_rdma_request(struct spdk_nvmf_rdma_request *rdma_req)
135 {
136 	int i;
137 
138 	rdma_req->req.length = 0;
139 	rdma_req->req.data_from_pool = false;
140 	rdma_req->req.data = NULL;
141 	rdma_req->data.wr.num_sge = 0;
142 	rdma_req->data.wr.wr.rdma.remote_addr = 0;
143 	rdma_req->data.wr.wr.rdma.rkey = 0;
144 	rdma_req->offset = 0;
145 	memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
146 
147 	for (i = 0; i < SPDK_NVMF_MAX_SGL_ENTRIES; i++) {
148 		rdma_req->req.iov[i].iov_base = 0;
149 		rdma_req->req.iov[i].iov_len = 0;
150 		rdma_req->req.buffers[i] = 0;
151 		rdma_req->data.wr.sg_list[i].addr = 0;
152 		rdma_req->data.wr.sg_list[i].length = 0;
153 		rdma_req->data.wr.sg_list[i].lkey = 0;
154 	}
155 	rdma_req->req.iovcnt = 0;
156 }
157 
158 static void
159 test_spdk_nvmf_rdma_request_parse_sgl(void)
160 {
161 	struct spdk_nvmf_rdma_transport rtransport;
162 	struct spdk_nvmf_rdma_device device;
163 	struct spdk_nvmf_rdma_request rdma_req = {};
164 	struct spdk_nvmf_rdma_recv recv;
165 	struct spdk_nvmf_rdma_poll_group group;
166 	struct spdk_nvmf_rdma_qpair rqpair;
167 	struct spdk_nvmf_rdma_poller poller;
168 	union nvmf_c2h_msg cpl;
169 	union nvmf_h2c_msg cmd;
170 	struct spdk_nvme_sgl_descriptor *sgl;
171 	struct spdk_nvmf_transport_pg_cache_buf bufs[4];
172 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
173 	struct spdk_nvmf_rdma_request_data data;
174 	int rc, i;
175 	uint32_t sgl_length;
176 	uintptr_t aligned_buffer_address;
177 
178 	data.wr.sg_list = data.sgl;
179 	STAILQ_INIT(&group.group.buf_cache);
180 	group.group.buf_cache_size = 0;
181 	group.group.buf_cache_count = 0;
182 	group.group.transport = &rtransport.transport;
183 	poller.group = &group;
184 	rqpair.poller = &poller;
185 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
186 
187 	sgl = &cmd.nvme_cmd.dptr.sgl1;
188 	rdma_req.recv = &recv;
189 	rdma_req.req.cmd = &cmd;
190 	rdma_req.req.rsp = &cpl;
191 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
192 	rdma_req.req.qpair = &rqpair.qpair;
193 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
194 
195 	rtransport.transport.opts = g_rdma_ut_transport_opts;
196 	rtransport.data_wr_pool = NULL;
197 	rtransport.transport.data_buf_pool = NULL;
198 
199 	device.attr.device_cap_flags = 0;
200 	sgl->keyed.key = 0xEEEE;
201 	sgl->address = 0xFFFF;
202 	rdma_req.recv->buf = (void *)0xDDDD;
203 
204 	/* Test 1: sgl type: keyed data block subtype: address */
205 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
206 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
207 
208 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size */
209 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
210 	reset_nvmf_rdma_request(&rdma_req);
211 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
212 
213 	device.map = (void *)0x0;
214 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
215 	CU_ASSERT(rc == 0);
216 	CU_ASSERT(rdma_req.req.data_from_pool == true);
217 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
218 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
219 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
220 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
221 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
222 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
223 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
224 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
225 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
226 
227 	/* Part 2: simple I/O, one SGL larger than the transport io unit size (equal to the max io size) */
228 	reset_nvmf_rdma_request(&rdma_req);
229 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
230 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
231 
232 	CU_ASSERT(rc == 0);
233 	CU_ASSERT(rdma_req.req.data_from_pool == true);
234 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO);
235 	CU_ASSERT(rdma_req.data.wr.num_sge == RDMA_UT_UNITS_IN_MAX_IO);
236 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
237 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
238 	for (i = 0; i < RDMA_UT_UNITS_IN_MAX_IO; i++) {
239 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
240 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
241 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
242 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
243 	}
244 
245 	/* Part 3: simple I/O one SGL larger than the transport max io size */
246 	reset_nvmf_rdma_request(&rdma_req);
247 	sgl->keyed.length = rtransport.transport.opts.max_io_size * 2;
248 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
249 
250 	CU_ASSERT(rc == -1);
251 
252 	/* Part 4: Pretend there are no buffer pools */
253 	MOCK_SET(spdk_mempool_get, NULL);
254 	reset_nvmf_rdma_request(&rdma_req);
255 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
256 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
257 
258 	CU_ASSERT(rc == 0);
259 	CU_ASSERT(rdma_req.req.data_from_pool == false);
260 	CU_ASSERT(rdma_req.req.data == NULL);
261 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
262 	CU_ASSERT(rdma_req.req.buffers[0] == NULL);
263 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0);
264 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 0);
265 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == 0);
266 
267 	rdma_req.recv->buf = (void *)0xDDDD;
268 	/* Test 2: sgl type: keyed data block subtype: offset (in capsule data) */
269 	sgl->generic.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
270 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
271 
272 	/* Part 1: Normal I/O smaller than in capsule data size no offset */
273 	reset_nvmf_rdma_request(&rdma_req);
274 	sgl->address = 0;
275 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
276 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
277 
278 	CU_ASSERT(rc == 0);
279 	CU_ASSERT(rdma_req.req.data == (void *)0xDDDD);
280 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.in_capsule_data_size);
281 	CU_ASSERT(rdma_req.req.data_from_pool == false);
282 
283 	/* Part 2: I/O offset + length too large */
284 	reset_nvmf_rdma_request(&rdma_req);
285 	sgl->address = rtransport.transport.opts.in_capsule_data_size;
286 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
287 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
288 
289 	CU_ASSERT(rc == -1);
290 
291 	/* Part 3: I/O too large */
292 	reset_nvmf_rdma_request(&rdma_req);
293 	sgl->address = 0;
294 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size * 2;
295 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
296 
297 	CU_ASSERT(rc == -1);
298 
299 	/* Test 3: Multi SGL */
300 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
301 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
302 	sgl->address = 0;
303 	rdma_req.recv->buf = (void *)&sgl_desc;
304 	MOCK_SET(spdk_mempool_get, &data);
305 
306 	/* part 1: 2 segments each with 1 wr. */
307 	reset_nvmf_rdma_request(&rdma_req);
308 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
309 	for (i = 0; i < 2; i++) {
310 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
311 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
312 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size;
313 		sgl_desc[i].address = 0x4000 + i * rtransport.transport.opts.io_unit_size;
314 		sgl_desc[i].keyed.key = 0x44;
315 	}
316 
317 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
318 
319 	CU_ASSERT(rc == 0);
320 	CU_ASSERT(rdma_req.req.data_from_pool == true);
321 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 2);
322 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
323 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
324 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
325 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
326 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
327 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size);
328 	CU_ASSERT(data.wr.num_sge == 1);
329 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
330 
331 	/* part 2: 2 segments, each with 1 wr containing 8 sge_elements */
332 	reset_nvmf_rdma_request(&rdma_req);
333 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
334 	for (i = 0; i < 2; i++) {
335 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
336 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
337 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size * 8;
338 		sgl_desc[i].address = 0x4000 + i * 8 * rtransport.transport.opts.io_unit_size;
339 		sgl_desc[i].keyed.key = 0x44;
340 	}
341 
342 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
343 
344 	CU_ASSERT(rc == 0);
345 	CU_ASSERT(rdma_req.req.data_from_pool == true);
346 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
347 	CU_ASSERT(rdma_req.req.iovcnt == 16);
348 	CU_ASSERT(rdma_req.data.wr.num_sge == 8);
349 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
350 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
351 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
352 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
353 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 8);
354 	CU_ASSERT(data.wr.num_sge == 8);
355 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
356 
357 	/* part 3: 2 segments, one very large, one very small */
358 	reset_nvmf_rdma_request(&rdma_req);
359 	for (i = 0; i < 2; i++) {
360 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
361 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
362 		sgl_desc[i].keyed.key = 0x44;
363 	}
364 
365 	sgl_desc[0].keyed.length = rtransport.transport.opts.io_unit_size * 15 +
366 				   rtransport.transport.opts.io_unit_size / 2;
367 	sgl_desc[0].address = 0x4000;
368 	sgl_desc[1].keyed.length = rtransport.transport.opts.io_unit_size / 2;
369 	sgl_desc[1].address = 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
370 			      rtransport.transport.opts.io_unit_size / 2;
371 
372 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
373 
374 	CU_ASSERT(rc == 0);
375 	CU_ASSERT(rdma_req.req.data_from_pool == true);
376 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
377 	CU_ASSERT(rdma_req.req.iovcnt == 16);
378 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
379 	for (i = 0; i < 15; i++) {
380 		CU_ASSERT(rdma_req.data.sgl[i].length == rtransport.transport.opts.io_unit_size);
381 	}
382 	CU_ASSERT(rdma_req.data.sgl[15].length == rtransport.transport.opts.io_unit_size / 2);
383 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
384 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
385 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
386 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
387 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
388 		  rtransport.transport.opts.io_unit_size / 2);
389 	CU_ASSERT(data.sgl[0].length == rtransport.transport.opts.io_unit_size / 2);
390 	CU_ASSERT(data.wr.num_sge == 1);
391 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
392 
393 	/* part 4: 2 SGL descriptors, each length is transport buffer / 2
394 	 * 1 transport buffers should be allocated */
395 	reset_nvmf_rdma_request(&rdma_req);
396 	aligned_buffer_address = ((uintptr_t)(&data) + NVMF_DATA_BUFFER_MASK) & ~NVMF_DATA_BUFFER_MASK;
397 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
398 	sgl_length = rtransport.transport.opts.io_unit_size / 2;
399 	for (i = 0; i < 2; i++) {
400 		sgl_desc[i].keyed.length = sgl_length;
401 		sgl_desc[i].address = 0x4000 + i * sgl_length;
402 	}
403 
404 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
405 
406 	CU_ASSERT(rc == 0);
407 	CU_ASSERT(rdma_req.req.data_from_pool == true);
408 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size);
409 	CU_ASSERT(rdma_req.req.iovcnt == 1);
410 
411 	CU_ASSERT(rdma_req.data.sgl[0].length == sgl_length);
412 	/* We mocked mempool_get to return address of data variable. Mempool is used
413 	 * to get both additional WRs and data buffers, so data points to &data */
414 	CU_ASSERT(rdma_req.data.sgl[0].addr == aligned_buffer_address);
415 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
416 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
417 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
418 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
419 
420 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
421 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + sgl_length);
422 	CU_ASSERT(data.sgl[0].length == sgl_length);
423 	CU_ASSERT(data.sgl[0].addr == aligned_buffer_address + sgl_length);
424 	CU_ASSERT(data.wr.num_sge == 1);
425 
426 	/* Test 4: use PG buffer cache */
427 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
428 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
429 	sgl->address = 0xFFFF;
430 	rdma_req.recv->buf = (void *)0xDDDD;
431 	sgl->keyed.key = 0xEEEE;
432 
433 	for (i = 0; i < 4; i++) {
434 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
435 	}
436 
437 	/* part 1: use the four buffers from the pg cache */
438 	group.group.buf_cache_size = 4;
439 	group.group.buf_cache_count = 4;
440 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
441 	reset_nvmf_rdma_request(&rdma_req);
442 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * 4;
443 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
444 
445 	SPDK_CU_ASSERT_FATAL(rc == 0);
446 	CU_ASSERT(rdma_req.req.data_from_pool == true);
447 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
448 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
449 			~NVMF_DATA_BUFFER_MASK));
450 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
451 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
452 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
453 	CU_ASSERT(group.group.buf_cache_count == 0);
454 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
455 	for (i = 0; i < 4; i++) {
456 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
457 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
458 				~NVMF_DATA_BUFFER_MASK));
459 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
460 	}
461 
462 	/* part 2: now that we have used the buffers from the cache, try again. We should get mempool buffers. */
463 	reset_nvmf_rdma_request(&rdma_req);
464 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
465 
466 	SPDK_CU_ASSERT_FATAL(rc == 0);
467 	CU_ASSERT(rdma_req.req.data_from_pool == true);
468 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
469 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
470 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
471 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
472 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
473 	CU_ASSERT(group.group.buf_cache_count == 0);
474 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
475 	for (i = 0; i < 4; i++) {
476 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
477 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
478 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
479 		CU_ASSERT(group.group.buf_cache_count == 0);
480 	}
481 
482 	/* part 3: half and half */
483 	group.group.buf_cache_count = 2;
484 
485 	for (i = 0; i < 2; i++) {
486 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
487 	}
488 	reset_nvmf_rdma_request(&rdma_req);
489 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
490 
491 	SPDK_CU_ASSERT_FATAL(rc == 0);
492 	CU_ASSERT(rdma_req.req.data_from_pool == true);
493 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
494 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
495 			~NVMF_DATA_BUFFER_MASK));
496 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
497 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
498 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
499 	CU_ASSERT(group.group.buf_cache_count == 0);
500 	for (i = 0; i < 2; i++) {
501 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
502 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
503 				~NVMF_DATA_BUFFER_MASK));
504 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
505 	}
506 	for (i = 2; i < 4; i++) {
507 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
508 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
509 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
510 	}
511 
512 	reset_nvmf_rdma_request(&rdma_req);
513 }
514 
515 static struct spdk_nvmf_rdma_recv *
516 create_recv(struct spdk_nvmf_rdma_qpair *rqpair, enum spdk_nvme_nvm_opcode opc)
517 {
518 	struct spdk_nvmf_rdma_recv *rdma_recv;
519 	union nvmf_h2c_msg *cmd;
520 	struct spdk_nvme_sgl_descriptor *sgl;
521 
522 	rdma_recv = calloc(1, sizeof(*rdma_recv));
523 	rdma_recv->qpair = rqpair;
524 	cmd = calloc(1, sizeof(*cmd));
525 	rdma_recv->sgl[0].addr = (uintptr_t)cmd;
526 	cmd->nvme_cmd.opc = opc;
527 	sgl = &cmd->nvme_cmd.dptr.sgl1;
528 	sgl->keyed.key = 0xEEEE;
529 	sgl->address = 0xFFFF;
530 	sgl->keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
531 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
532 	sgl->keyed.length = 1;
533 
534 	return rdma_recv;
535 }
536 
537 static void
538 free_recv(struct spdk_nvmf_rdma_recv *rdma_recv)
539 {
540 	free((void *)rdma_recv->sgl[0].addr);
541 	free(rdma_recv);
542 }
543 
544 static struct spdk_nvmf_rdma_request *
545 create_req(struct spdk_nvmf_rdma_qpair *rqpair,
546 	   struct spdk_nvmf_rdma_recv *rdma_recv)
547 {
548 	struct spdk_nvmf_rdma_request *rdma_req;
549 	union nvmf_c2h_msg *cpl;
550 
551 	rdma_req = calloc(1, sizeof(*rdma_req));
552 	rdma_req->recv = rdma_recv;
553 	rdma_req->req.qpair = &rqpair->qpair;
554 	rdma_req->state = RDMA_REQUEST_STATE_NEW;
555 	rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr;
556 	rdma_req->data.wr.sg_list = rdma_req->data.sgl;
557 	cpl = calloc(1, sizeof(*cpl));
558 	rdma_req->rsp.sgl[0].addr = (uintptr_t)cpl;
559 	rdma_req->req.rsp = cpl;
560 
561 	return rdma_req;
562 }
563 
564 static void
565 free_req(struct spdk_nvmf_rdma_request *rdma_req)
566 {
567 	free((void *)rdma_req->rsp.sgl[0].addr);
568 	free(rdma_req);
569 }
570 
571 static void
572 qpair_reset(struct spdk_nvmf_rdma_qpair *rqpair,
573 	    struct spdk_nvmf_rdma_poller *poller,
574 	    struct spdk_nvmf_rdma_device *device,
575 	    struct spdk_nvmf_rdma_resources *resources,
576 	    struct spdk_nvmf_transport *transport)
577 {
578 	memset(rqpair, 0, sizeof(*rqpair));
579 	STAILQ_INIT(&rqpair->pending_rdma_write_queue);
580 	STAILQ_INIT(&rqpair->pending_rdma_read_queue);
581 	rqpair->poller = poller;
582 	rqpair->device = device;
583 	rqpair->resources = resources;
584 	rqpair->qpair.qid = 1;
585 	rqpair->ibv_state = IBV_QPS_RTS;
586 	rqpair->qpair.state = SPDK_NVMF_QPAIR_ACTIVE;
587 	rqpair->max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
588 	rqpair->max_send_depth = 16;
589 	rqpair->max_read_depth = 16;
590 	rqpair->qpair.transport = transport;
591 }
592 
593 static void
594 poller_reset(struct spdk_nvmf_rdma_poller *poller,
595 	     struct spdk_nvmf_rdma_poll_group *group)
596 {
597 	memset(poller, 0, sizeof(*poller));
598 	STAILQ_INIT(&poller->qpairs_pending_recv);
599 	STAILQ_INIT(&poller->qpairs_pending_send);
600 	poller->group = group;
601 }
602 
603 static void
604 test_spdk_nvmf_rdma_request_process(void)
605 {
606 	struct spdk_nvmf_rdma_transport rtransport = {};
607 	struct spdk_nvmf_rdma_poll_group group = {};
608 	struct spdk_nvmf_rdma_poller poller = {};
609 	struct spdk_nvmf_rdma_device device = {};
610 	struct spdk_nvmf_rdma_resources resources = {};
611 	struct spdk_nvmf_rdma_qpair rqpair = {};
612 	struct spdk_nvmf_rdma_recv *rdma_recv;
613 	struct spdk_nvmf_rdma_request *rdma_req;
614 	bool progress;
615 
616 	STAILQ_INIT(&group.group.buf_cache);
617 	STAILQ_INIT(&group.group.pending_buf_queue);
618 	group.group.buf_cache_size = 0;
619 	group.group.buf_cache_count = 0;
620 	poller_reset(&poller, &group);
621 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
622 
623 	rtransport.transport.opts = g_rdma_ut_transport_opts;
624 	rtransport.transport.data_buf_pool = spdk_mempool_create("test_data_pool", 16, 128, 0, 0);
625 	rtransport.data_wr_pool = spdk_mempool_create("test_wr_pool", 128,
626 				  sizeof(struct spdk_nvmf_rdma_request_data),
627 				  0, 0);
628 	MOCK_CLEAR(spdk_mempool_get);
629 
630 	device.attr.device_cap_flags = 0;
631 	device.map = (void *)0x0;
632 
633 	/* Test 1: single SGL READ request */
634 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_READ);
635 	rdma_req = create_req(&rqpair, rdma_recv);
636 	rqpair.current_recv_depth = 1;
637 	/* NEW -> EXECUTING */
638 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
639 	CU_ASSERT(progress == true);
640 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
641 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST);
642 	/* EXECUTED -> TRANSFERRING_C2H */
643 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
644 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
645 	CU_ASSERT(progress == true);
646 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST);
647 	CU_ASSERT(rdma_req->recv == NULL);
648 	/* COMPLETED -> FREE */
649 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
650 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
651 	CU_ASSERT(progress == true);
652 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
653 
654 	free_recv(rdma_recv);
655 	free_req(rdma_req);
656 	poller_reset(&poller, &group);
657 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
658 
659 	/* Test 2: single SGL WRITE request */
660 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
661 	rdma_req = create_req(&rqpair, rdma_recv);
662 	rqpair.current_recv_depth = 1;
663 	/* NEW -> TRANSFERRING_H2C */
664 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
665 	CU_ASSERT(progress == true);
666 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
667 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
668 	STAILQ_INIT(&poller.qpairs_pending_send);
669 	/* READY_TO_EXECUTE -> EXECUTING */
670 	rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
671 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
672 	CU_ASSERT(progress == true);
673 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
674 	/* EXECUTED -> COMPLETING */
675 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
676 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
677 	CU_ASSERT(progress == true);
678 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
679 	CU_ASSERT(rdma_req->recv == NULL);
680 	/* COMPLETED -> FREE */
681 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
682 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
683 	CU_ASSERT(progress == true);
684 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
685 
686 	free_recv(rdma_recv);
687 	free_req(rdma_req);
688 	poller_reset(&poller, &group);
689 	qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
690 
691 	/* Test 3: WRITE+WRITE ibv_send batching */
692 	{
693 		struct spdk_nvmf_rdma_recv *recv1, *recv2;
694 		struct spdk_nvmf_rdma_request *req1, *req2;
695 		recv1 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
696 		req1 = create_req(&rqpair, recv1);
697 		recv2 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
698 		req2 = create_req(&rqpair, recv2);
699 
700 		/* WRITE 1: NEW -> TRANSFERRING_H2C */
701 		rqpair.current_recv_depth = 1;
702 		nvmf_rdma_request_process(&rtransport, req1);
703 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
704 
705 		/* WRITE 2: NEW -> TRANSFERRING_H2C */
706 		rqpair.current_recv_depth = 2;
707 		nvmf_rdma_request_process(&rtransport, req2);
708 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
709 
710 		STAILQ_INIT(&poller.qpairs_pending_send);
711 
712 		/* WRITE 1 completes before WRITE 2 has finished RDMA reading */
713 		/* WRITE 1: READY_TO_EXECUTE -> EXECUTING */
714 		req1->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
715 		nvmf_rdma_request_process(&rtransport, req1);
716 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_EXECUTING);
717 		/* WRITE 1: EXECUTED -> COMPLETING */
718 		req1->state = RDMA_REQUEST_STATE_EXECUTED;
719 		nvmf_rdma_request_process(&rtransport, req1);
720 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_COMPLETING);
721 		STAILQ_INIT(&poller.qpairs_pending_send);
722 		/* WRITE 1: COMPLETED -> FREE */
723 		req1->state = RDMA_REQUEST_STATE_COMPLETED;
724 		nvmf_rdma_request_process(&rtransport, req1);
725 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_FREE);
726 
727 		/* Now WRITE 2 has finished reading and completes */
728 		/* WRITE 2: COMPLETED -> FREE */
729 		/* WRITE 2: READY_TO_EXECUTE -> EXECUTING */
730 		req2->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
731 		nvmf_rdma_request_process(&rtransport, req2);
732 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_EXECUTING);
733 		/* WRITE 1: EXECUTED -> COMPLETING */
734 		req2->state = RDMA_REQUEST_STATE_EXECUTED;
735 		nvmf_rdma_request_process(&rtransport, req2);
736 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_COMPLETING);
737 		STAILQ_INIT(&poller.qpairs_pending_send);
738 		/* WRITE 1: COMPLETED -> FREE */
739 		req2->state = RDMA_REQUEST_STATE_COMPLETED;
740 		nvmf_rdma_request_process(&rtransport, req2);
741 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_FREE);
742 
743 		free_recv(recv1);
744 		free_req(req1);
745 		free_recv(recv2);
746 		free_req(req2);
747 		poller_reset(&poller, &group);
748 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
749 	}
750 
751 	/* Test 4, invalid command, check xfer type */
752 	{
753 		struct spdk_nvmf_rdma_recv *rdma_recv_inv;
754 		struct spdk_nvmf_rdma_request *rdma_req_inv;
755 		/* construct an opcode that specifies BIDIRECTIONAL transfer */
756 		uint8_t opc = 0x10 | SPDK_NVME_DATA_BIDIRECTIONAL;
757 
758 		rdma_recv_inv = create_recv(&rqpair, opc);
759 		rdma_req_inv = create_req(&rqpair, rdma_recv_inv);
760 
761 		/* NEW -> RDMA_REQUEST_STATE_COMPLETING */
762 		rqpair.current_recv_depth = 1;
763 		progress = nvmf_rdma_request_process(&rtransport, rdma_req_inv);
764 		CU_ASSERT(progress == true);
765 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_COMPLETING);
766 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sct == SPDK_NVME_SCT_GENERIC);
767 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sc == SPDK_NVME_SC_INVALID_OPCODE);
768 
769 		/* RDMA_REQUEST_STATE_COMPLETED -> FREE */
770 		rdma_req_inv->state = RDMA_REQUEST_STATE_COMPLETED;
771 		nvmf_rdma_request_process(&rtransport, rdma_req_inv);
772 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_FREE);
773 
774 		free_recv(rdma_recv_inv);
775 		free_req(rdma_req_inv);
776 		poller_reset(&poller, &group);
777 		qpair_reset(&rqpair, &poller, &device, &resources, &rtransport.transport);
778 	}
779 
780 	spdk_mempool_free(rtransport.transport.data_buf_pool);
781 	spdk_mempool_free(rtransport.data_wr_pool);
782 }
783 
784 #define TEST_GROUPS_COUNT 5
785 static void
786 test_nvmf_rdma_get_optimal_poll_group(void)
787 {
788 	struct spdk_nvmf_rdma_transport rtransport = {};
789 	struct spdk_nvmf_transport *transport = &rtransport.transport;
790 	struct spdk_nvmf_rdma_qpair rqpair = {};
791 	struct spdk_nvmf_transport_poll_group *groups[TEST_GROUPS_COUNT];
792 	struct spdk_nvmf_rdma_poll_group *rgroups[TEST_GROUPS_COUNT];
793 	struct spdk_nvmf_transport_poll_group *result;
794 	uint32_t i;
795 
796 	rqpair.qpair.transport = transport;
797 	pthread_mutex_init(&rtransport.lock, NULL);
798 	TAILQ_INIT(&rtransport.poll_groups);
799 
800 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
801 		groups[i] = nvmf_rdma_poll_group_create(transport);
802 		CU_ASSERT(groups[i] != NULL);
803 		rgroups[i] = SPDK_CONTAINEROF(groups[i], struct spdk_nvmf_rdma_poll_group, group);
804 		groups[i]->transport = transport;
805 	}
806 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[0]);
807 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[0]);
808 
809 	/* Emulate connection of %TEST_GROUPS_COUNT% initiators - each creates 1 admin and 1 io qp */
810 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
811 		rqpair.qpair.qid = 0;
812 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
813 		CU_ASSERT(result == groups[i]);
814 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
815 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i]);
816 
817 		rqpair.qpair.qid = 1;
818 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
819 		CU_ASSERT(result == groups[i]);
820 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
821 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
822 	}
823 	/* wrap around, admin/io pg point to the first pg
824 	   Destroy all poll groups except of the last one */
825 	for (i = 0; i < TEST_GROUPS_COUNT - 1; i++) {
826 		nvmf_rdma_poll_group_destroy(groups[i]);
827 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[i + 1]);
828 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i + 1]);
829 	}
830 
831 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
832 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
833 
834 	/* Check that pointers to the next admin/io poll groups are not changed */
835 	rqpair.qpair.qid = 0;
836 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
837 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
838 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
839 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
840 
841 	rqpair.qpair.qid = 1;
842 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
843 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
844 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
845 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
846 
847 	/* Remove the last poll group, check that pointers are NULL */
848 	nvmf_rdma_poll_group_destroy(groups[TEST_GROUPS_COUNT - 1]);
849 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == NULL);
850 	CU_ASSERT(rtransport.conn_sched.next_io_pg == NULL);
851 
852 	/* Request optimal poll group, result must be NULL */
853 	rqpair.qpair.qid = 0;
854 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
855 	CU_ASSERT(result == NULL);
856 
857 	rqpair.qpair.qid = 1;
858 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
859 	CU_ASSERT(result == NULL);
860 
861 	pthread_mutex_destroy(&rtransport.lock);
862 }
863 #undef TEST_GROUPS_COUNT
864 
865 static void
866 test_spdk_nvmf_rdma_request_parse_sgl_with_md(void)
867 {
868 	struct spdk_nvmf_rdma_transport rtransport;
869 	struct spdk_nvmf_rdma_device device;
870 	struct spdk_nvmf_rdma_request rdma_req = {};
871 	struct spdk_nvmf_rdma_recv recv;
872 	struct spdk_nvmf_rdma_poll_group group;
873 	struct spdk_nvmf_rdma_qpair rqpair;
874 	struct spdk_nvmf_rdma_poller poller;
875 	union nvmf_c2h_msg cpl;
876 	union nvmf_h2c_msg cmd;
877 	struct spdk_nvme_sgl_descriptor *sgl;
878 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
879 	char data_buffer[8192];
880 	struct spdk_nvmf_rdma_request_data *data = (struct spdk_nvmf_rdma_request_data *)data_buffer;
881 	char data2_buffer[8192];
882 	struct spdk_nvmf_rdma_request_data *data2 = (struct spdk_nvmf_rdma_request_data *)data2_buffer;
883 	const uint32_t data_bs = 512;
884 	const uint32_t md_size = 8;
885 	int rc, i;
886 	void *aligned_buffer;
887 
888 	data->wr.sg_list = data->sgl;
889 	STAILQ_INIT(&group.group.buf_cache);
890 	group.group.buf_cache_size = 0;
891 	group.group.buf_cache_count = 0;
892 	group.group.transport = &rtransport.transport;
893 	poller.group = &group;
894 	rqpair.poller = &poller;
895 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
896 
897 	sgl = &cmd.nvme_cmd.dptr.sgl1;
898 	rdma_req.recv = &recv;
899 	rdma_req.req.cmd = &cmd;
900 	rdma_req.req.rsp = &cpl;
901 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
902 	rdma_req.req.qpair = &rqpair.qpair;
903 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
904 
905 	rtransport.transport.opts = g_rdma_ut_transport_opts;
906 	rtransport.data_wr_pool = NULL;
907 	rtransport.transport.data_buf_pool = NULL;
908 
909 	device.attr.device_cap_flags = 0;
910 	device.map = NULL;
911 	sgl->keyed.key = 0xEEEE;
912 	sgl->address = 0xFFFF;
913 	rdma_req.recv->buf = (void *)0xDDDD;
914 
915 	/* Test 1: sgl type: keyed data block subtype: address */
916 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
917 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
918 
919 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size, block size 512 */
920 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
921 	reset_nvmf_rdma_request(&rdma_req);
922 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
923 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
924 			  0, 0, 0, 0, 0);
925 	rdma_req.req.dif_enabled = true;
926 	rtransport.transport.opts.io_unit_size = data_bs * 8;
927 	sgl->keyed.length = data_bs * 4;
928 
929 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
930 
931 	CU_ASSERT(rc == 0);
932 	CU_ASSERT(rdma_req.req.data_from_pool == true);
933 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
934 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
935 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
936 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
937 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
938 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
939 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
940 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
941 
942 	for (i = 0; i < 4; ++i) {
943 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
944 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
945 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
946 	}
947 
948 	/* Part 2: simple I/O, one SGL equal to io unit size, io_unit_size is not aligned with md_size,
949 		block size 512 */
950 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
951 	reset_nvmf_rdma_request(&rdma_req);
952 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
953 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
954 			  0, 0, 0, 0, 0);
955 	rdma_req.req.dif_enabled = true;
956 	rtransport.transport.opts.io_unit_size = data_bs * 4;
957 	sgl->keyed.length = data_bs * 4;
958 
959 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
960 
961 	CU_ASSERT(rc == 0);
962 	CU_ASSERT(rdma_req.req.data_from_pool == true);
963 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
964 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
965 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
966 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
967 	CU_ASSERT(rdma_req.data.wr.num_sge == 5);
968 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
969 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
970 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
971 
972 	for (i = 0; i < 3; ++i) {
973 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
974 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
975 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
976 	}
977 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
978 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
979 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
980 
981 	/* 2nd buffer consumed */
982 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
983 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
984 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
985 
986 	/* Part 3: simple I/O, one SGL equal io unit size, io_unit_size is equal to block size 512 bytes */
987 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
988 	reset_nvmf_rdma_request(&rdma_req);
989 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
990 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
991 			  0, 0, 0, 0, 0);
992 	rdma_req.req.dif_enabled = true;
993 	rtransport.transport.opts.io_unit_size = data_bs;
994 	sgl->keyed.length = data_bs;
995 
996 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
997 
998 	CU_ASSERT(rc == 0);
999 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1000 	CU_ASSERT(rdma_req.req.length == data_bs);
1001 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1002 	CU_ASSERT(rdma_req.req.dif.elba_length == data_bs + md_size);
1003 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1004 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1005 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1006 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1007 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1008 
1009 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1010 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs);
1011 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1012 
1013 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1014 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)((unsigned long)0x2000));
1015 	CU_ASSERT(rdma_req.req.iov[0].iov_len == data_bs);
1016 	/* 2nd buffer consumed for metadata */
1017 	CU_ASSERT(rdma_req.req.iov[1].iov_base == (void *)((unsigned long)0x2000));
1018 	CU_ASSERT(rdma_req.req.iov[1].iov_len == md_size);
1019 
1020 	/* Part 4: simple I/O, one SGL equal io unit size, io_unit_size is aligned with md_size,
1021 	   block size 512 */
1022 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1023 	reset_nvmf_rdma_request(&rdma_req);
1024 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1025 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1026 			  0, 0, 0, 0, 0);
1027 	rdma_req.req.dif_enabled = true;
1028 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1029 	sgl->keyed.length = data_bs * 4;
1030 
1031 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1032 
1033 	CU_ASSERT(rc == 0);
1034 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1035 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1036 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1037 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1038 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1039 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1040 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1041 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1042 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1043 
1044 	for (i = 0; i < 4; ++i) {
1045 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1046 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1047 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1048 	}
1049 
1050 	/* Part 5: simple I/O, one SGL equal to 2x io unit size, io_unit_size is aligned with md_size,
1051 	   block size 512 */
1052 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1053 	reset_nvmf_rdma_request(&rdma_req);
1054 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1055 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1056 			  0, 0, 0, 0, 0);
1057 	rdma_req.req.dif_enabled = true;
1058 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 2;
1059 	sgl->keyed.length = data_bs * 4;
1060 
1061 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1062 
1063 	CU_ASSERT(rc == 0);
1064 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1065 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1066 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1067 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1068 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1069 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1070 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1071 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1072 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1073 
1074 	for (i = 0; i < 2; ++i) {
1075 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1076 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1077 	}
1078 	for (i = 0; i < 2; ++i) {
1079 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].addr == 0x2000 + i * (data_bs + md_size));
1080 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].length == data_bs);
1081 	}
1082 
1083 	/* Part 6: simple I/O, one SGL larger than the transport io unit size, io_unit_size is not aligned to md_size,
1084 	   block size 512 */
1085 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1086 	reset_nvmf_rdma_request(&rdma_req);
1087 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1088 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1089 			  0, 0, 0, 0, 0);
1090 	rdma_req.req.dif_enabled = true;
1091 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1092 	sgl->keyed.length = data_bs * 6;
1093 
1094 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1095 
1096 	CU_ASSERT(rc == 0);
1097 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1098 	CU_ASSERT(rdma_req.req.length == data_bs * 6);
1099 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1100 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 6);
1101 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1102 	CU_ASSERT(rdma_req.data.wr.num_sge == 7);
1103 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1104 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1105 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1106 
1107 	for (i = 0; i < 3; ++i) {
1108 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1109 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1110 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1111 	}
1112 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
1113 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
1114 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == RDMA_UT_LKEY);
1115 
1116 	/* 2nd IO buffer consumed */
1117 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
1118 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
1119 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == RDMA_UT_LKEY);
1120 
1121 	CU_ASSERT(rdma_req.data.wr.sg_list[5].addr == 0x2000 + 24 + md_size);
1122 	CU_ASSERT(rdma_req.data.wr.sg_list[5].length == 512);
1123 	CU_ASSERT(rdma_req.data.wr.sg_list[5].lkey == RDMA_UT_LKEY);
1124 
1125 	CU_ASSERT(rdma_req.data.wr.sg_list[6].addr == 0x2000 + 24 + 512 + md_size * 2);
1126 	CU_ASSERT(rdma_req.data.wr.sg_list[6].length == 512);
1127 	CU_ASSERT(rdma_req.data.wr.sg_list[6].lkey == RDMA_UT_LKEY);
1128 
1129 	/* Part 7: simple I/O, number of SGL entries exceeds the number of entries
1130 	   one WR can hold. Additional WR is chained */
1131 	MOCK_SET(spdk_mempool_get, data2_buffer);
1132 	aligned_buffer = (void *)((uintptr_t)(data2_buffer + NVMF_DATA_BUFFER_MASK) &
1133 				  ~NVMF_DATA_BUFFER_MASK);
1134 	reset_nvmf_rdma_request(&rdma_req);
1135 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1136 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1137 			  0, 0, 0, 0, 0);
1138 	rdma_req.req.dif_enabled = true;
1139 	rtransport.transport.opts.io_unit_size = data_bs * 16;
1140 	sgl->keyed.length = data_bs * 16;
1141 
1142 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1143 
1144 	CU_ASSERT(rc == 0);
1145 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1146 	CU_ASSERT(rdma_req.req.length == data_bs * 16);
1147 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1148 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1149 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 16);
1150 	CU_ASSERT(rdma_req.req.data == aligned_buffer);
1151 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
1152 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1153 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1154 
1155 	for (i = 0; i < 15; ++i) {
1156 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)aligned_buffer + i * (data_bs + md_size));
1157 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1158 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1159 	}
1160 
1161 	/* 8192 - (512 + 8) * 15 = 392 */
1162 	CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)aligned_buffer + i * (data_bs + md_size));
1163 	CU_ASSERT(rdma_req.data.wr.sg_list[i].length == 392);
1164 	CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == RDMA_UT_LKEY);
1165 
1166 	/* additional wr from pool */
1167 	CU_ASSERT(rdma_req.data.wr.next == (void *)&data2->wr);
1168 	CU_ASSERT(rdma_req.data.wr.next->num_sge == 1);
1169 	CU_ASSERT(rdma_req.data.wr.next->next == &rdma_req.rsp.wr);
1170 	/* 2nd IO buffer */
1171 	CU_ASSERT(data2->wr.sg_list[0].addr == (uintptr_t)aligned_buffer);
1172 	CU_ASSERT(data2->wr.sg_list[0].length == 120);
1173 	CU_ASSERT(data2->wr.sg_list[0].lkey == RDMA_UT_LKEY);
1174 
1175 	/* Part 8: simple I/O, data with metadata do not fit to 1 io_buffer */
1176 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1177 	reset_nvmf_rdma_request(&rdma_req);
1178 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1179 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1180 			  0, 0, 0, 0, 0);
1181 	rdma_req.req.dif_enabled = true;
1182 	rtransport.transport.opts.io_unit_size = 516;
1183 	sgl->keyed.length = data_bs * 2;
1184 
1185 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1186 
1187 	CU_ASSERT(rc == 0);
1188 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1189 	CU_ASSERT(rdma_req.req.length == data_bs * 2);
1190 	CU_ASSERT(rdma_req.req.iovcnt == 3);
1191 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1192 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 2);
1193 	CU_ASSERT(rdma_req.req.data == (void *)0x2000);
1194 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1195 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1196 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1197 
1198 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1199 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 512);
1200 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == RDMA_UT_LKEY);
1201 
1202 	/* 2nd IO buffer consumed, offset 4 bytes due to part of the metadata
1203 	  is located at the beginning of that buffer */
1204 	CU_ASSERT(rdma_req.data.wr.sg_list[1].addr == 0x2000 + 4);
1205 	CU_ASSERT(rdma_req.data.wr.sg_list[1].length == 512);
1206 	CU_ASSERT(rdma_req.data.wr.sg_list[1].lkey == RDMA_UT_LKEY);
1207 
1208 	/* Test 2: Multi SGL */
1209 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1210 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1211 	sgl->address = 0;
1212 	rdma_req.recv->buf = (void *)&sgl_desc;
1213 	MOCK_SET(spdk_mempool_get, data_buffer);
1214 	aligned_buffer = (void *)((uintptr_t)(data_buffer + NVMF_DATA_BUFFER_MASK) &
1215 				  ~NVMF_DATA_BUFFER_MASK);
1216 
1217 	/* part 1: 2 segments each with 1 wr. io_unit_size is aligned with data_bs + md_size */
1218 	reset_nvmf_rdma_request(&rdma_req);
1219 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1220 			  SPDK_DIF_TYPE1,
1221 			  SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 0, 0, 0, 0, 0);
1222 	rdma_req.req.dif_enabled = true;
1223 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1224 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
1225 
1226 	for (i = 0; i < 2; i++) {
1227 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1228 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1229 		sgl_desc[i].keyed.length = data_bs * 4;
1230 		sgl_desc[i].address = 0x4000 + i * data_bs * 4;
1231 		sgl_desc[i].keyed.key = 0x44;
1232 	}
1233 
1234 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1235 
1236 	CU_ASSERT(rc == 0);
1237 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1238 	CU_ASSERT(rdma_req.req.length == data_bs * 4 * 2);
1239 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1240 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4 * 2);
1241 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1242 	for (i = 0; i < 4; ++i) {
1243 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)((unsigned char *)aligned_buffer) + i *
1244 			  (data_bs + md_size));
1245 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1246 	}
1247 
1248 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
1249 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
1250 	CU_ASSERT(rdma_req.data.wr.next == &data->wr);
1251 	CU_ASSERT(data->wr.wr.rdma.rkey == 0x44);
1252 	CU_ASSERT(data->wr.wr.rdma.remote_addr == 0x4000 + data_bs * 4);
1253 	CU_ASSERT(data->wr.num_sge == 4);
1254 	for (i = 0; i < 4; ++i) {
1255 		CU_ASSERT(data->wr.sg_list[i].addr == (uintptr_t)((unsigned char *)aligned_buffer) + i *
1256 			  (data_bs + md_size));
1257 		CU_ASSERT(data->wr.sg_list[i].length == data_bs);
1258 	}
1259 
1260 	CU_ASSERT(data->wr.next == &rdma_req.rsp.wr);
1261 }
1262 
1263 static void
1264 test_nvmf_rdma_opts_init(void)
1265 {
1266 	struct spdk_nvmf_transport_opts	opts = {};
1267 
1268 	nvmf_rdma_opts_init(&opts);
1269 	CU_ASSERT(opts.max_queue_depth == SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH);
1270 	CU_ASSERT(opts.max_qpairs_per_ctrlr ==	SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR);
1271 	CU_ASSERT(opts.in_capsule_data_size ==	SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE);
1272 	CU_ASSERT(opts.max_io_size == SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE);
1273 	CU_ASSERT(opts.io_unit_size == SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
1274 	CU_ASSERT(opts.max_aq_depth == SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH);
1275 	CU_ASSERT(opts.num_shared_buffers == SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS);
1276 	CU_ASSERT(opts.buf_cache_size == SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE);
1277 	CU_ASSERT(opts.dif_insert_or_strip == SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP);
1278 	CU_ASSERT(opts.abort_timeout_sec == SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC);
1279 	CU_ASSERT(opts.transport_specific == NULL);
1280 }
1281 
1282 static void
1283 test_nvmf_rdma_request_free_data(void)
1284 {
1285 	struct spdk_nvmf_rdma_request rdma_req = {};
1286 	struct spdk_nvmf_rdma_transport rtransport = {};
1287 	struct spdk_nvmf_rdma_request_data *next_request_data = NULL;
1288 
1289 	MOCK_CLEAR(spdk_mempool_get);
1290 	rtransport.data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data",
1291 				  SPDK_NVMF_MAX_SGL_ENTRIES,
1292 				  sizeof(struct spdk_nvmf_rdma_request_data),
1293 				  SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
1294 				  SPDK_ENV_SOCKET_ID_ANY);
1295 	next_request_data = spdk_mempool_get(rtransport.data_wr_pool);
1296 	SPDK_CU_ASSERT_FATAL(((struct test_mempool *)rtransport.data_wr_pool)->count ==
1297 			     SPDK_NVMF_MAX_SGL_ENTRIES - 1);
1298 	next_request_data->wr.wr_id = 1;
1299 	next_request_data->wr.num_sge = 2;
1300 	next_request_data->wr.next = NULL;
1301 	rdma_req.data.wr.next = &next_request_data->wr;
1302 	rdma_req.data.wr.wr_id = 1;
1303 	rdma_req.data.wr.num_sge = 2;
1304 
1305 	nvmf_rdma_request_free_data(&rdma_req, &rtransport);
1306 	/* Check if next_request_data put into memory pool */
1307 	CU_ASSERT(((struct test_mempool *)rtransport.data_wr_pool)->count == SPDK_NVMF_MAX_SGL_ENTRIES);
1308 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
1309 
1310 	spdk_mempool_free(rtransport.data_wr_pool);
1311 }
1312 
1313 static void
1314 test_nvmf_rdma_update_ibv_state(void)
1315 {
1316 	struct spdk_nvmf_rdma_qpair rqpair = {};
1317 	struct spdk_rdma_qp rdma_qp = {};
1318 	struct ibv_qp qp = {};
1319 	int rc = 0;
1320 
1321 	rqpair.rdma_qp = &rdma_qp;
1322 
1323 	/* Case 1: Failed to get updated RDMA queue pair state */
1324 	rqpair.ibv_state = IBV_QPS_INIT;
1325 	rqpair.rdma_qp->qp = NULL;
1326 
1327 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1328 	CU_ASSERT(rc == IBV_QPS_ERR + 1);
1329 
1330 	/* Case 2: Bad state updated */
1331 	rqpair.rdma_qp->qp = &qp;
1332 	qp.state = IBV_QPS_ERR;
1333 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1334 	CU_ASSERT(rqpair.ibv_state == 10);
1335 	CU_ASSERT(rc == IBV_QPS_ERR + 1);
1336 
1337 	/* Case 3: Pass */
1338 	qp.state = IBV_QPS_INIT;
1339 	rc = nvmf_rdma_update_ibv_state(&rqpair);
1340 	CU_ASSERT(rqpair.ibv_state == IBV_QPS_INIT);
1341 	CU_ASSERT(rc == IBV_QPS_INIT);
1342 }
1343 
1344 int main(int argc, char **argv)
1345 {
1346 	CU_pSuite	suite = NULL;
1347 	unsigned int	num_failures;
1348 
1349 	CU_set_error_action(CUEA_ABORT);
1350 	CU_initialize_registry();
1351 
1352 	suite = CU_add_suite("nvmf", NULL, NULL);
1353 
1354 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl);
1355 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_process);
1356 	CU_ADD_TEST(suite, test_nvmf_rdma_get_optimal_poll_group);
1357 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl_with_md);
1358 	CU_ADD_TEST(suite, test_nvmf_rdma_opts_init);
1359 	CU_ADD_TEST(suite, test_nvmf_rdma_request_free_data);
1360 	CU_ADD_TEST(suite, test_nvmf_rdma_update_ibv_state);
1361 
1362 	CU_basic_set_mode(CU_BRM_VERBOSE);
1363 	CU_basic_run_tests();
1364 	num_failures = CU_get_number_of_failures();
1365 	CU_cleanup_registry();
1366 	return num_failures;
1367 }
1368