xref: /spdk/test/unit/lib/nvmf/rdma.c/rdma_ut.c (revision 0ed85362c8132a2d1927757fbcade66b6660d26a)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright (c) Intel Corporation. All rights reserved.
5  *   Copyright (c) 2019 Mellanox Technologies LTD. All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include "spdk/stdinc.h"
35 #include "spdk_cunit.h"
36 #include "common/lib/test_env.c"
37 #include "common/lib/test_rdma.c"
38 #include "nvmf/rdma.c"
39 #include "nvmf/transport.c"
40 
41 uint64_t g_mr_size;
42 uint64_t g_mr_next_size;
43 struct ibv_mr g_rdma_mr;
44 
45 #define RDMA_UT_UNITS_IN_MAX_IO 16
46 
47 struct spdk_nvmf_transport_opts g_rdma_ut_transport_opts = {
48 	.max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH,
49 	.max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR,
50 	.in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE,
51 	.max_io_size = (SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE * RDMA_UT_UNITS_IN_MAX_IO),
52 	.io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE,
53 	.max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH,
54 	.num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS,
55 };
56 
57 SPDK_LOG_REGISTER_COMPONENT("nvmf", SPDK_LOG_NVMF)
58 DEFINE_STUB(spdk_mem_map_set_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
59 		uint64_t size, uint64_t translation), 0);
60 DEFINE_STUB(spdk_mem_map_clear_translation, int, (struct spdk_mem_map *map, uint64_t vaddr,
61 		uint64_t size), 0);
62 DEFINE_STUB(spdk_mem_map_alloc, struct spdk_mem_map *, (uint64_t default_translation,
63 		const struct spdk_mem_map_ops *ops, void *cb_ctx), NULL);
64 DEFINE_STUB(spdk_nvmf_qpair_disconnect, int, (struct spdk_nvmf_qpair *qpair,
65 		nvmf_qpair_disconnect_cb cb_fn, void *ctx), 0);
66 DEFINE_STUB_V(spdk_mem_map_free, (struct spdk_mem_map **pmap));
67 
68 struct spdk_trace_histories *g_trace_histories;
69 DEFINE_STUB_V(spdk_trace_add_register_fn, (struct spdk_trace_register_fn *reg_fn));
70 DEFINE_STUB_V(spdk_trace_register_object, (uint8_t type, char id_prefix));
71 DEFINE_STUB_V(spdk_trace_register_description, (const char *name,
72 		uint16_t tpoint_id, uint8_t owner_type, uint8_t object_type, uint8_t new_object,
73 		uint8_t arg1_type, const char *arg1_name));
74 DEFINE_STUB_V(_spdk_trace_record, (uint64_t tsc, uint16_t tpoint_id, uint16_t poller_id,
75 				   uint32_t size, uint64_t object_id, uint64_t arg1));
76 
77 DEFINE_STUB_V(spdk_nvmf_ctrlr_data_init, (struct spdk_nvmf_transport_opts *opts,
78 		struct spdk_nvmf_ctrlr_data *cdata));
79 DEFINE_STUB_V(spdk_nvmf_request_exec, (struct spdk_nvmf_request *req));
80 DEFINE_STUB(spdk_nvmf_request_complete, int, (struct spdk_nvmf_request *req), 0);
81 DEFINE_STUB(spdk_nvme_transport_id_compare, int, (const struct spdk_nvme_transport_id *trid1,
82 		const struct spdk_nvme_transport_id *trid2), 0);
83 DEFINE_STUB_V(nvmf_ctrlr_abort_aer, (struct spdk_nvmf_ctrlr *ctrlr));
84 DEFINE_STUB(spdk_nvmf_request_get_dif_ctx, bool, (struct spdk_nvmf_request *req,
85 		struct spdk_dif_ctx *dif_ctx), false);
86 DEFINE_STUB_V(spdk_nvme_trid_populate_transport, (struct spdk_nvme_transport_id *trid,
87 		enum spdk_nvme_transport_type trtype));
88 DEFINE_STUB_V(spdk_nvmf_tgt_new_qpair, (struct spdk_nvmf_tgt *tgt, struct spdk_nvmf_qpair *qpair));
89 DEFINE_STUB(nvmf_ctrlr_abort_request, int, (struct spdk_nvmf_request *req), 0);
90 
91 const char *
92 spdk_nvme_transport_id_trtype_str(enum spdk_nvme_transport_type trtype)
93 {
94 	switch (trtype) {
95 	case SPDK_NVME_TRANSPORT_PCIE:
96 		return "PCIe";
97 	case SPDK_NVME_TRANSPORT_RDMA:
98 		return "RDMA";
99 	case SPDK_NVME_TRANSPORT_FC:
100 		return "FC";
101 	default:
102 		return NULL;
103 	}
104 }
105 
106 int
107 spdk_nvme_transport_id_populate_trstring(struct spdk_nvme_transport_id *trid, const char *trstring)
108 {
109 	int len, i;
110 
111 	if (trstring == NULL) {
112 		return -EINVAL;
113 	}
114 
115 	len = strnlen(trstring, SPDK_NVMF_TRSTRING_MAX_LEN);
116 	if (len == SPDK_NVMF_TRSTRING_MAX_LEN) {
117 		return -EINVAL;
118 	}
119 
120 	/* cast official trstring to uppercase version of input. */
121 	for (i = 0; i < len; i++) {
122 		trid->trstring[i] = toupper(trstring[i]);
123 	}
124 	return 0;
125 }
126 
127 uint64_t
128 spdk_mem_map_translate(const struct spdk_mem_map *map, uint64_t vaddr, uint64_t *size)
129 {
130 	if (g_mr_size != 0) {
131 		*(uint32_t *)size = g_mr_size;
132 		if (g_mr_next_size != 0) {
133 			g_mr_size = g_mr_next_size;
134 		}
135 	}
136 
137 	return (uint64_t)&g_rdma_mr;
138 }
139 
140 static void reset_nvmf_rdma_request(struct spdk_nvmf_rdma_request *rdma_req)
141 {
142 	int i;
143 
144 	rdma_req->req.length = 0;
145 	rdma_req->req.data_from_pool = false;
146 	rdma_req->req.data = NULL;
147 	rdma_req->data.wr.num_sge = 0;
148 	rdma_req->data.wr.wr.rdma.remote_addr = 0;
149 	rdma_req->data.wr.wr.rdma.rkey = 0;
150 	memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
151 
152 	for (i = 0; i < SPDK_NVMF_MAX_SGL_ENTRIES; i++) {
153 		rdma_req->req.iov[i].iov_base = 0;
154 		rdma_req->req.iov[i].iov_len = 0;
155 		rdma_req->req.buffers[i] = 0;
156 		rdma_req->data.wr.sg_list[i].addr = 0;
157 		rdma_req->data.wr.sg_list[i].length = 0;
158 		rdma_req->data.wr.sg_list[i].lkey = 0;
159 	}
160 	rdma_req->req.iovcnt = 0;
161 }
162 
163 static void
164 test_spdk_nvmf_rdma_request_parse_sgl(void)
165 {
166 	struct spdk_nvmf_rdma_transport rtransport;
167 	struct spdk_nvmf_rdma_device device;
168 	struct spdk_nvmf_rdma_request rdma_req = {};
169 	struct spdk_nvmf_rdma_recv recv;
170 	struct spdk_nvmf_rdma_poll_group group;
171 	struct spdk_nvmf_rdma_qpair rqpair;
172 	struct spdk_nvmf_rdma_poller poller;
173 	union nvmf_c2h_msg cpl;
174 	union nvmf_h2c_msg cmd;
175 	struct spdk_nvme_sgl_descriptor *sgl;
176 	struct spdk_nvmf_transport_pg_cache_buf bufs[4];
177 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
178 	struct spdk_nvmf_rdma_request_data data;
179 	struct spdk_nvmf_transport_pg_cache_buf	buffer;
180 	struct spdk_nvmf_transport_pg_cache_buf	*buffer_ptr;
181 	int rc, i;
182 
183 	data.wr.sg_list = data.sgl;
184 	STAILQ_INIT(&group.group.buf_cache);
185 	group.group.buf_cache_size = 0;
186 	group.group.buf_cache_count = 0;
187 	group.group.transport = &rtransport.transport;
188 	STAILQ_INIT(&group.retired_bufs);
189 	poller.group = &group;
190 	rqpair.poller = &poller;
191 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
192 
193 	sgl = &cmd.nvme_cmd.dptr.sgl1;
194 	rdma_req.recv = &recv;
195 	rdma_req.req.cmd = &cmd;
196 	rdma_req.req.rsp = &cpl;
197 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
198 	rdma_req.req.qpair = &rqpair.qpair;
199 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
200 
201 	rtransport.transport.opts = g_rdma_ut_transport_opts;
202 	rtransport.data_wr_pool = NULL;
203 	rtransport.transport.data_buf_pool = NULL;
204 
205 	device.attr.device_cap_flags = 0;
206 	g_rdma_mr.lkey = 0xABCD;
207 	sgl->keyed.key = 0xEEEE;
208 	sgl->address = 0xFFFF;
209 	rdma_req.recv->buf = (void *)0xDDDD;
210 
211 	/* Test 1: sgl type: keyed data block subtype: address */
212 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
213 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
214 
215 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size */
216 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
217 	reset_nvmf_rdma_request(&rdma_req);
218 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
219 
220 	device.map = (void *)0x0;
221 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
222 	CU_ASSERT(rc == 0);
223 	CU_ASSERT(rdma_req.req.data_from_pool == true);
224 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
225 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
226 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
227 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
228 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
229 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
230 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
231 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
232 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
233 
234 	/* Part 2: simple I/O, one SGL larger than the transport io unit size (equal to the max io size) */
235 	reset_nvmf_rdma_request(&rdma_req);
236 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
237 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
238 
239 	CU_ASSERT(rc == 0);
240 	CU_ASSERT(rdma_req.req.data_from_pool == true);
241 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO);
242 	CU_ASSERT(rdma_req.data.wr.num_sge == RDMA_UT_UNITS_IN_MAX_IO);
243 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
244 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
245 	for (i = 0; i < RDMA_UT_UNITS_IN_MAX_IO; i++) {
246 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
247 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
248 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
249 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
250 	}
251 
252 	/* Part 3: simple I/O one SGL larger than the transport max io size */
253 	reset_nvmf_rdma_request(&rdma_req);
254 	sgl->keyed.length = rtransport.transport.opts.max_io_size * 2;
255 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
256 
257 	CU_ASSERT(rc == -1);
258 
259 	/* Part 4: Pretend there are no buffer pools */
260 	MOCK_SET(spdk_mempool_get, NULL);
261 	reset_nvmf_rdma_request(&rdma_req);
262 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * RDMA_UT_UNITS_IN_MAX_IO;
263 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
264 
265 	CU_ASSERT(rc == 0);
266 	CU_ASSERT(rdma_req.req.data_from_pool == false);
267 	CU_ASSERT(rdma_req.req.data == NULL);
268 	CU_ASSERT(rdma_req.data.wr.num_sge == 0);
269 	CU_ASSERT(rdma_req.req.buffers[0] == NULL);
270 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0);
271 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 0);
272 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == 0);
273 
274 	rdma_req.recv->buf = (void *)0xDDDD;
275 	/* Test 2: sgl type: keyed data block subtype: offset (in capsule data) */
276 	sgl->generic.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
277 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
278 
279 	/* Part 1: Normal I/O smaller than in capsule data size no offset */
280 	reset_nvmf_rdma_request(&rdma_req);
281 	sgl->address = 0;
282 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
283 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
284 
285 	CU_ASSERT(rc == 0);
286 	CU_ASSERT(rdma_req.req.data == (void *)0xDDDD);
287 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.in_capsule_data_size);
288 	CU_ASSERT(rdma_req.req.data_from_pool == false);
289 
290 	/* Part 2: I/O offset + length too large */
291 	reset_nvmf_rdma_request(&rdma_req);
292 	sgl->address = rtransport.transport.opts.in_capsule_data_size;
293 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size;
294 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
295 
296 	CU_ASSERT(rc == -1);
297 
298 	/* Part 3: I/O too large */
299 	reset_nvmf_rdma_request(&rdma_req);
300 	sgl->address = 0;
301 	sgl->unkeyed.length = rtransport.transport.opts.in_capsule_data_size * 2;
302 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
303 
304 	CU_ASSERT(rc == -1);
305 
306 	/* Test 3: Multi SGL */
307 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
308 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
309 	sgl->address = 0;
310 	rdma_req.recv->buf = (void *)&sgl_desc;
311 	MOCK_SET(spdk_mempool_get, &data);
312 
313 	/* part 1: 2 segments each with 1 wr. */
314 	reset_nvmf_rdma_request(&rdma_req);
315 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
316 	for (i = 0; i < 2; i++) {
317 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
318 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
319 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size;
320 		sgl_desc[i].address = 0x4000 + i * rtransport.transport.opts.io_unit_size;
321 		sgl_desc[i].keyed.key = 0x44;
322 	}
323 
324 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
325 
326 	CU_ASSERT(rc == 0);
327 	CU_ASSERT(rdma_req.req.data_from_pool == true);
328 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 2);
329 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
330 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
331 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
332 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
333 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
334 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size);
335 	CU_ASSERT(data.wr.num_sge == 1);
336 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
337 
338 	/* part 2: 2 segments, each with 1 wr containing 8 sge_elements */
339 	reset_nvmf_rdma_request(&rdma_req);
340 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
341 	for (i = 0; i < 2; i++) {
342 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
343 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
344 		sgl_desc[i].keyed.length = rtransport.transport.opts.io_unit_size * 8;
345 		sgl_desc[i].address = 0x4000 + i * 8 * rtransport.transport.opts.io_unit_size;
346 		sgl_desc[i].keyed.key = 0x44;
347 	}
348 
349 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
350 
351 	CU_ASSERT(rc == 0);
352 	CU_ASSERT(rdma_req.req.data_from_pool == true);
353 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
354 	CU_ASSERT(rdma_req.req.iovcnt == 16);
355 	CU_ASSERT(rdma_req.data.wr.num_sge == 8);
356 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
357 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
358 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
359 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
360 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 8);
361 	CU_ASSERT(data.wr.num_sge == 8);
362 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
363 
364 	/* part 3: 2 segments, one very large, one very small */
365 	reset_nvmf_rdma_request(&rdma_req);
366 	for (i = 0; i < 2; i++) {
367 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
368 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
369 		sgl_desc[i].keyed.key = 0x44;
370 	}
371 
372 	sgl_desc[0].keyed.length = rtransport.transport.opts.io_unit_size * 15 +
373 				   rtransport.transport.opts.io_unit_size / 2;
374 	sgl_desc[0].address = 0x4000;
375 	sgl_desc[1].keyed.length = rtransport.transport.opts.io_unit_size / 2;
376 	sgl_desc[1].address = 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
377 			      rtransport.transport.opts.io_unit_size / 2;
378 
379 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
380 
381 	CU_ASSERT(rc == 0);
382 	CU_ASSERT(rdma_req.req.data_from_pool == true);
383 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 16);
384 	CU_ASSERT(rdma_req.req.iovcnt == 17);
385 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
386 	for (i = 0; i < 15; i++) {
387 		CU_ASSERT(rdma_req.data.sgl[i].length == rtransport.transport.opts.io_unit_size);
388 	}
389 	CU_ASSERT(rdma_req.data.sgl[15].length == rtransport.transport.opts.io_unit_size / 2);
390 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
391 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
392 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
393 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
394 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + rtransport.transport.opts.io_unit_size * 15 +
395 		  rtransport.transport.opts.io_unit_size / 2);
396 	CU_ASSERT(data.sgl[0].length == rtransport.transport.opts.io_unit_size / 2);
397 	CU_ASSERT(data.wr.num_sge == 1);
398 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
399 
400 	/* Test 4: use PG buffer cache */
401 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
402 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
403 	sgl->address = 0xFFFF;
404 	rdma_req.recv->buf = (void *)0xDDDD;
405 	g_rdma_mr.lkey = 0xABCD;
406 	sgl->keyed.key = 0xEEEE;
407 
408 	for (i = 0; i < 4; i++) {
409 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
410 	}
411 
412 	/* part 1: use the four buffers from the pg cache */
413 	group.group.buf_cache_size = 4;
414 	group.group.buf_cache_count = 4;
415 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
416 	reset_nvmf_rdma_request(&rdma_req);
417 	sgl->keyed.length = rtransport.transport.opts.io_unit_size * 4;
418 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
419 
420 	SPDK_CU_ASSERT_FATAL(rc == 0);
421 	CU_ASSERT(rdma_req.req.data_from_pool == true);
422 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
423 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
424 			~NVMF_DATA_BUFFER_MASK));
425 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
426 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
427 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
428 	CU_ASSERT(group.group.buf_cache_count == 0);
429 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
430 	for (i = 0; i < 4; i++) {
431 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
432 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
433 				~NVMF_DATA_BUFFER_MASK));
434 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
435 	}
436 
437 	/* part 2: now that we have used the buffers from the cache, try again. We should get mempool buffers. */
438 	reset_nvmf_rdma_request(&rdma_req);
439 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
440 
441 	SPDK_CU_ASSERT_FATAL(rc == 0);
442 	CU_ASSERT(rdma_req.req.data_from_pool == true);
443 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size * 4);
444 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
445 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
446 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
447 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
448 	CU_ASSERT(group.group.buf_cache_count == 0);
449 	CU_ASSERT(STAILQ_EMPTY(&group.group.buf_cache));
450 	for (i = 0; i < 4; i++) {
451 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
452 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
453 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
454 		CU_ASSERT(group.group.buf_cache_count == 0);
455 	}
456 
457 	/* part 3: half and half */
458 	group.group.buf_cache_count = 2;
459 
460 	for (i = 0; i < 2; i++) {
461 		STAILQ_INSERT_TAIL(&group.group.buf_cache, &bufs[i], link);
462 	}
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 == (((uint64_t)&bufs[0] + NVMF_DATA_BUFFER_MASK) &
470 			~NVMF_DATA_BUFFER_MASK));
471 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
472 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
473 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
474 	CU_ASSERT(group.group.buf_cache_count == 0);
475 	for (i = 0; i < 2; i++) {
476 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == (uint64_t)&bufs[i]);
477 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (((uint64_t)&bufs[i] + NVMF_DATA_BUFFER_MASK) &
478 				~NVMF_DATA_BUFFER_MASK));
479 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
480 	}
481 	for (i = 2; i < 4; i++) {
482 		CU_ASSERT((uint64_t)rdma_req.req.buffers[i] == 0x2000);
483 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000);
484 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == rtransport.transport.opts.io_unit_size);
485 	}
486 
487 	reset_nvmf_rdma_request(&rdma_req);
488 	/* Test 5 dealing with a buffer split over two Memory Regions */
489 	MOCK_SET(spdk_mempool_get, (void *)&buffer);
490 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
491 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
492 	sgl->keyed.length = rtransport.transport.opts.io_unit_size / 2;
493 	g_mr_size = rtransport.transport.opts.io_unit_size / 4;
494 	g_mr_next_size = rtransport.transport.opts.io_unit_size / 2;
495 
496 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
497 	SPDK_CU_ASSERT_FATAL(rc == 0);
498 	CU_ASSERT(rdma_req.req.data_from_pool == true);
499 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
500 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&buffer + NVMF_DATA_BUFFER_MASK) &
501 			~NVMF_DATA_BUFFER_MASK));
502 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
503 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
504 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
505 	CU_ASSERT(rdma_req.req.buffers[0] == &buffer);
506 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == (((uint64_t)&buffer + NVMF_DATA_BUFFER_MASK) &
507 			~NVMF_DATA_BUFFER_MASK));
508 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == rtransport.transport.opts.io_unit_size / 2);
509 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
510 	buffer_ptr = STAILQ_FIRST(&group.retired_bufs);
511 	CU_ASSERT(buffer_ptr == &buffer);
512 	STAILQ_REMOVE(&group.retired_bufs, buffer_ptr, spdk_nvmf_transport_pg_cache_buf, link);
513 	CU_ASSERT(STAILQ_EMPTY(&group.retired_bufs));
514 	g_mr_size = 0;
515 	g_mr_next_size = 0;
516 
517 	reset_nvmf_rdma_request(&rdma_req);
518 }
519 
520 static struct spdk_nvmf_rdma_recv *
521 create_recv(struct spdk_nvmf_rdma_qpair *rqpair, enum spdk_nvme_nvm_opcode opc)
522 {
523 	struct spdk_nvmf_rdma_recv *rdma_recv;
524 	union nvmf_h2c_msg *cmd;
525 	struct spdk_nvme_sgl_descriptor *sgl;
526 
527 	rdma_recv = calloc(1, sizeof(*rdma_recv));
528 	rdma_recv->qpair = rqpair;
529 	cmd = calloc(1, sizeof(*cmd));
530 	rdma_recv->sgl[0].addr = (uintptr_t)cmd;
531 	cmd->nvme_cmd.opc = opc;
532 	sgl = &cmd->nvme_cmd.dptr.sgl1;
533 	sgl->keyed.key = 0xEEEE;
534 	sgl->address = 0xFFFF;
535 	sgl->keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
536 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
537 	sgl->keyed.length = 1;
538 
539 	return rdma_recv;
540 }
541 
542 static void
543 free_recv(struct spdk_nvmf_rdma_recv *rdma_recv)
544 {
545 	free((void *)rdma_recv->sgl[0].addr);
546 	free(rdma_recv);
547 }
548 
549 static struct spdk_nvmf_rdma_request *
550 create_req(struct spdk_nvmf_rdma_qpair *rqpair,
551 	   struct spdk_nvmf_rdma_recv *rdma_recv)
552 {
553 	struct spdk_nvmf_rdma_request *rdma_req;
554 	union nvmf_c2h_msg *cpl;
555 
556 	rdma_req = calloc(1, sizeof(*rdma_req));
557 	rdma_req->recv = rdma_recv;
558 	rdma_req->req.qpair = &rqpair->qpair;
559 	rdma_req->state = RDMA_REQUEST_STATE_NEW;
560 	rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr;
561 	rdma_req->data.wr.sg_list = rdma_req->data.sgl;
562 	cpl = calloc(1, sizeof(*cpl));
563 	rdma_req->rsp.sgl[0].addr = (uintptr_t)cpl;
564 	rdma_req->req.rsp = cpl;
565 
566 	return rdma_req;
567 }
568 
569 static void
570 free_req(struct spdk_nvmf_rdma_request *rdma_req)
571 {
572 	free((void *)rdma_req->rsp.sgl[0].addr);
573 	free(rdma_req);
574 }
575 
576 static void
577 qpair_reset(struct spdk_nvmf_rdma_qpair *rqpair,
578 	    struct spdk_nvmf_rdma_poller *poller,
579 	    struct spdk_nvmf_rdma_device *device,
580 	    struct spdk_nvmf_rdma_resources *resources)
581 {
582 	memset(rqpair, 0, sizeof(*rqpair));
583 	STAILQ_INIT(&rqpair->pending_rdma_write_queue);
584 	STAILQ_INIT(&rqpair->pending_rdma_read_queue);
585 	rqpair->poller = poller;
586 	rqpair->device = device;
587 	rqpair->resources = resources;
588 	rqpair->qpair.qid = 1;
589 	rqpair->ibv_state = IBV_QPS_RTS;
590 	rqpair->qpair.state = SPDK_NVMF_QPAIR_ACTIVE;
591 	rqpair->max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
592 	rqpair->max_send_depth = 16;
593 	rqpair->max_read_depth = 16;
594 	resources->recvs_to_post.first = resources->recvs_to_post.last = NULL;
595 }
596 
597 static void
598 poller_reset(struct spdk_nvmf_rdma_poller *poller,
599 	     struct spdk_nvmf_rdma_poll_group *group)
600 {
601 	memset(poller, 0, sizeof(*poller));
602 	STAILQ_INIT(&poller->qpairs_pending_recv);
603 	STAILQ_INIT(&poller->qpairs_pending_send);
604 	poller->group = group;
605 }
606 
607 static void
608 test_spdk_nvmf_rdma_request_process(void)
609 {
610 	struct spdk_nvmf_rdma_transport rtransport = {};
611 	struct spdk_nvmf_rdma_poll_group group = {};
612 	struct spdk_nvmf_rdma_poller poller = {};
613 	struct spdk_nvmf_rdma_device device = {};
614 	struct spdk_nvmf_rdma_resources resources = {};
615 	struct spdk_nvmf_rdma_qpair rqpair = {};
616 	struct spdk_nvmf_rdma_recv *rdma_recv;
617 	struct spdk_nvmf_rdma_request *rdma_req;
618 	bool progress;
619 
620 	STAILQ_INIT(&group.group.buf_cache);
621 	STAILQ_INIT(&group.group.pending_buf_queue);
622 	group.group.buf_cache_size = 0;
623 	group.group.buf_cache_count = 0;
624 	poller_reset(&poller, &group);
625 	qpair_reset(&rqpair, &poller, &device, &resources);
626 
627 	rtransport.transport.opts = g_rdma_ut_transport_opts;
628 	rtransport.transport.data_buf_pool = spdk_mempool_create("test_data_pool", 16, 128, 0, 0);
629 	rtransport.data_wr_pool = spdk_mempool_create("test_wr_pool", 128,
630 				  sizeof(struct spdk_nvmf_rdma_request_data),
631 				  0, 0);
632 	MOCK_CLEAR(spdk_mempool_get);
633 
634 	device.attr.device_cap_flags = 0;
635 	device.map = (void *)0x0;
636 	g_rdma_mr.lkey = 0xABCD;
637 
638 	/* Test 1: single SGL READ request */
639 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_READ);
640 	rdma_req = create_req(&rqpair, rdma_recv);
641 	rqpair.current_recv_depth = 1;
642 	/* NEW -> EXECUTING */
643 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
644 	CU_ASSERT(progress == true);
645 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
646 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST);
647 	/* EXECUTED -> TRANSFERRING_C2H */
648 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
649 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
650 	CU_ASSERT(progress == true);
651 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST);
652 	CU_ASSERT(rdma_req->recv == NULL);
653 	CU_ASSERT(resources.recvs_to_post.first == &rdma_recv->wr);
654 	CU_ASSERT(resources.recvs_to_post.last == &rdma_recv->wr);
655 	/* COMPLETED -> FREE */
656 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
657 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
658 	CU_ASSERT(progress == true);
659 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
660 
661 	free_recv(rdma_recv);
662 	free_req(rdma_req);
663 	poller_reset(&poller, &group);
664 	qpair_reset(&rqpair, &poller, &device, &resources);
665 
666 	/* Test 2: single SGL WRITE request */
667 	rdma_recv = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
668 	rdma_req = create_req(&rqpair, rdma_recv);
669 	rqpair.current_recv_depth = 1;
670 	/* NEW -> TRANSFERRING_H2C */
671 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
672 	CU_ASSERT(progress == true);
673 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
674 	CU_ASSERT(rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
675 	STAILQ_INIT(&poller.qpairs_pending_send);
676 	/* READY_TO_EXECUTE -> EXECUTING */
677 	rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
678 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
679 	CU_ASSERT(progress == true);
680 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_EXECUTING);
681 	/* EXECUTED -> COMPLETING */
682 	rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
683 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
684 	CU_ASSERT(progress == true);
685 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_COMPLETING);
686 	CU_ASSERT(rdma_req->recv == NULL);
687 	CU_ASSERT(resources.recvs_to_post.first == &rdma_recv->wr);
688 	CU_ASSERT(resources.recvs_to_post.last == &rdma_recv->wr);
689 	/* COMPLETED -> FREE */
690 	rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
691 	progress = nvmf_rdma_request_process(&rtransport, rdma_req);
692 	CU_ASSERT(progress == true);
693 	CU_ASSERT(rdma_req->state == RDMA_REQUEST_STATE_FREE);
694 
695 	free_recv(rdma_recv);
696 	free_req(rdma_req);
697 	poller_reset(&poller, &group);
698 	qpair_reset(&rqpair, &poller, &device, &resources);
699 
700 	/* Test 3: WRITE+WRITE ibv_send batching */
701 	{
702 		struct spdk_nvmf_rdma_recv *recv1, *recv2;
703 		struct spdk_nvmf_rdma_request *req1, *req2;
704 		recv1 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
705 		req1 = create_req(&rqpair, recv1);
706 		recv2 = create_recv(&rqpair, SPDK_NVME_OPC_WRITE);
707 		req2 = create_req(&rqpair, recv2);
708 
709 		/* WRITE 1: NEW -> TRANSFERRING_H2C */
710 		rqpair.current_recv_depth = 1;
711 		nvmf_rdma_request_process(&rtransport, req1);
712 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
713 
714 		/* WRITE 2: NEW -> TRANSFERRING_H2C */
715 		rqpair.current_recv_depth = 2;
716 		nvmf_rdma_request_process(&rtransport, req2);
717 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER);
718 
719 		STAILQ_INIT(&poller.qpairs_pending_send);
720 
721 		/* WRITE 1 completes before WRITE 2 has finished RDMA reading */
722 		/* WRITE 1: READY_TO_EXECUTE -> EXECUTING */
723 		req1->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
724 		nvmf_rdma_request_process(&rtransport, req1);
725 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_EXECUTING);
726 		/* WRITE 1: EXECUTED -> COMPLETING */
727 		req1->state = RDMA_REQUEST_STATE_EXECUTED;
728 		nvmf_rdma_request_process(&rtransport, req1);
729 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_COMPLETING);
730 		STAILQ_INIT(&poller.qpairs_pending_send);
731 		/* WRITE 1: COMPLETED -> FREE */
732 		req1->state = RDMA_REQUEST_STATE_COMPLETED;
733 		nvmf_rdma_request_process(&rtransport, req1);
734 		CU_ASSERT(req1->state == RDMA_REQUEST_STATE_FREE);
735 
736 		/* Now WRITE 2 has finished reading and completes */
737 		/* WRITE 2: COMPLETED -> FREE */
738 		/* WRITE 2: READY_TO_EXECUTE -> EXECUTING */
739 		req2->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
740 		nvmf_rdma_request_process(&rtransport, req2);
741 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_EXECUTING);
742 		/* WRITE 1: EXECUTED -> COMPLETING */
743 		req2->state = RDMA_REQUEST_STATE_EXECUTED;
744 		nvmf_rdma_request_process(&rtransport, req2);
745 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_COMPLETING);
746 		STAILQ_INIT(&poller.qpairs_pending_send);
747 		/* WRITE 1: COMPLETED -> FREE */
748 		req2->state = RDMA_REQUEST_STATE_COMPLETED;
749 		nvmf_rdma_request_process(&rtransport, req2);
750 		CU_ASSERT(req2->state == RDMA_REQUEST_STATE_FREE);
751 
752 		free_recv(recv1);
753 		free_req(req1);
754 		free_recv(recv2);
755 		free_req(req2);
756 		poller_reset(&poller, &group);
757 		qpair_reset(&rqpair, &poller, &device, &resources);
758 	}
759 
760 	/* Test 4, invalid command, check xfer type */
761 	{
762 		struct spdk_nvmf_rdma_recv *rdma_recv_inv;
763 		struct spdk_nvmf_rdma_request *rdma_req_inv;
764 		/* construct an opcode that specifies BIDIRECTIONAL transfer */
765 		uint8_t opc = 0x10 | SPDK_NVME_DATA_BIDIRECTIONAL;
766 
767 		rdma_recv_inv = create_recv(&rqpair, opc);
768 		rdma_req_inv = create_req(&rqpair, rdma_recv_inv);
769 
770 		/* NEW -> RDMA_REQUEST_STATE_COMPLETING */
771 		rqpair.current_recv_depth = 1;
772 		progress = nvmf_rdma_request_process(&rtransport, rdma_req_inv);
773 		CU_ASSERT(progress == true);
774 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_COMPLETING);
775 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sct == SPDK_NVME_SCT_GENERIC);
776 		CU_ASSERT(rdma_req_inv->req.rsp->nvme_cpl.status.sc == SPDK_NVME_SC_INVALID_OPCODE);
777 
778 		/* RDMA_REQUEST_STATE_COMPLETED -> FREE */
779 		rdma_req_inv->state = RDMA_REQUEST_STATE_COMPLETED;
780 		nvmf_rdma_request_process(&rtransport, rdma_req_inv);
781 		CU_ASSERT(rdma_req_inv->state == RDMA_REQUEST_STATE_FREE);
782 
783 		free_recv(rdma_recv_inv);
784 		free_req(rdma_req_inv);
785 		poller_reset(&poller, &group);
786 		qpair_reset(&rqpair, &poller, &device, &resources);
787 	}
788 
789 	spdk_mempool_free(rtransport.transport.data_buf_pool);
790 	spdk_mempool_free(rtransport.data_wr_pool);
791 }
792 
793 #define TEST_GROUPS_COUNT 5
794 static void
795 test_nvmf_rdma_get_optimal_poll_group(void)
796 {
797 	struct spdk_nvmf_rdma_transport rtransport = {};
798 	struct spdk_nvmf_transport *transport = &rtransport.transport;
799 	struct spdk_nvmf_rdma_qpair rqpair = {};
800 	struct spdk_nvmf_transport_poll_group *groups[TEST_GROUPS_COUNT];
801 	struct spdk_nvmf_rdma_poll_group *rgroups[TEST_GROUPS_COUNT];
802 	struct spdk_nvmf_transport_poll_group *result;
803 	uint32_t i;
804 
805 	rqpair.qpair.transport = transport;
806 	pthread_mutex_init(&rtransport.lock, NULL);
807 	TAILQ_INIT(&rtransport.poll_groups);
808 
809 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
810 		groups[i] = nvmf_rdma_poll_group_create(transport);
811 		CU_ASSERT(groups[i] != NULL);
812 		rgroups[i] = SPDK_CONTAINEROF(groups[i], struct spdk_nvmf_rdma_poll_group, group);
813 		groups[i]->transport = transport;
814 	}
815 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[0]);
816 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[0]);
817 
818 	/* Emulate connection of %TEST_GROUPS_COUNT% initiators - each creates 1 admin and 1 io qp */
819 	for (i = 0; i < TEST_GROUPS_COUNT; i++) {
820 		rqpair.qpair.qid = 0;
821 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
822 		CU_ASSERT(result == groups[i]);
823 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
824 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i]);
825 
826 		rqpair.qpair.qid = 1;
827 		result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
828 		CU_ASSERT(result == groups[i]);
829 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
830 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[(i + 1) % TEST_GROUPS_COUNT]);
831 	}
832 	/* wrap around, admin/io pg point to the first pg
833 	   Destroy all poll groups except of the last one */
834 	for (i = 0; i < TEST_GROUPS_COUNT - 1; i++) {
835 		nvmf_rdma_poll_group_destroy(groups[i]);
836 		CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[i + 1]);
837 		CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[i + 1]);
838 	}
839 
840 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
841 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
842 
843 	/* Check that pointers to the next admin/io poll groups are not changed */
844 	rqpair.qpair.qid = 0;
845 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
846 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
847 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
848 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
849 
850 	rqpair.qpair.qid = 1;
851 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
852 	CU_ASSERT(result == groups[TEST_GROUPS_COUNT - 1]);
853 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == rgroups[TEST_GROUPS_COUNT - 1]);
854 	CU_ASSERT(rtransport.conn_sched.next_io_pg == rgroups[TEST_GROUPS_COUNT - 1]);
855 
856 	/* Remove the last poll group, check that pointers are NULL */
857 	nvmf_rdma_poll_group_destroy(groups[TEST_GROUPS_COUNT - 1]);
858 	CU_ASSERT(rtransport.conn_sched.next_admin_pg == NULL);
859 	CU_ASSERT(rtransport.conn_sched.next_io_pg == NULL);
860 
861 	/* Request optimal poll group, result must be NULL */
862 	rqpair.qpair.qid = 0;
863 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
864 	CU_ASSERT(result == NULL);
865 
866 	rqpair.qpair.qid = 1;
867 	result = nvmf_rdma_get_optimal_poll_group(&rqpair.qpair);
868 	CU_ASSERT(result == NULL);
869 
870 	pthread_mutex_destroy(&rtransport.lock);
871 }
872 #undef TEST_GROUPS_COUNT
873 
874 static void
875 test_spdk_nvmf_rdma_request_parse_sgl_with_md(void)
876 {
877 	struct spdk_nvmf_rdma_transport rtransport;
878 	struct spdk_nvmf_rdma_device device;
879 	struct spdk_nvmf_rdma_request rdma_req = {};
880 	struct spdk_nvmf_rdma_recv recv;
881 	struct spdk_nvmf_rdma_poll_group group;
882 	struct spdk_nvmf_rdma_qpair rqpair;
883 	struct spdk_nvmf_rdma_poller poller;
884 	union nvmf_c2h_msg cpl;
885 	union nvmf_h2c_msg cmd;
886 	struct spdk_nvme_sgl_descriptor *sgl;
887 	struct spdk_nvme_sgl_descriptor sgl_desc[SPDK_NVMF_MAX_SGL_ENTRIES] = {{0}};
888 	struct spdk_nvmf_rdma_request_data data;
889 	struct spdk_nvmf_transport_pg_cache_buf	buffer;
890 	struct spdk_nvmf_transport_pg_cache_buf	*buffer_ptr;
891 	const uint32_t data_bs = 512;
892 	const uint32_t md_size = 8;
893 	int rc, i;
894 	void *aligned_buffer;
895 
896 	data.wr.sg_list = data.sgl;
897 	STAILQ_INIT(&group.group.buf_cache);
898 	group.group.buf_cache_size = 0;
899 	group.group.buf_cache_count = 0;
900 	group.group.transport = &rtransport.transport;
901 	STAILQ_INIT(&group.retired_bufs);
902 	poller.group = &group;
903 	rqpair.poller = &poller;
904 	rqpair.max_send_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
905 
906 	sgl = &cmd.nvme_cmd.dptr.sgl1;
907 	rdma_req.recv = &recv;
908 	rdma_req.req.cmd = &cmd;
909 	rdma_req.req.rsp = &cpl;
910 	rdma_req.data.wr.sg_list = rdma_req.data.sgl;
911 	rdma_req.req.qpair = &rqpair.qpair;
912 	rdma_req.req.xfer = SPDK_NVME_DATA_CONTROLLER_TO_HOST;
913 
914 	rtransport.transport.opts = g_rdma_ut_transport_opts;
915 	rtransport.data_wr_pool = NULL;
916 	rtransport.transport.data_buf_pool = NULL;
917 
918 	device.attr.device_cap_flags = 0;
919 	device.map = NULL;
920 	g_rdma_mr.lkey = 0xABCD;
921 	sgl->keyed.key = 0xEEEE;
922 	sgl->address = 0xFFFF;
923 	rdma_req.recv->buf = (void *)0xDDDD;
924 
925 	/* Test 1: sgl type: keyed data block subtype: address */
926 	sgl->generic.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
927 	sgl->keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
928 
929 	/* Part 1: simple I/O, one SGL smaller than the transport io unit size, block size 512 */
930 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
931 	reset_nvmf_rdma_request(&rdma_req);
932 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
933 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
934 			  0, 0, 0, 0, 0);
935 	rdma_req.req.dif.dif_insert_or_strip = true;
936 	rtransport.transport.opts.io_unit_size = data_bs * 8;
937 	sgl->keyed.length = data_bs * 4;
938 
939 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
940 
941 	CU_ASSERT(rc == 0);
942 	CU_ASSERT(rdma_req.req.data_from_pool == true);
943 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
944 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
945 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
946 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
947 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
948 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
949 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
950 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
951 
952 	for (i = 0; i < 4; ++i) {
953 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
954 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
955 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
956 	}
957 
958 	/* Part 2: simple I/O, one SGL equal to io unit size, io_unit_size is not aligned with md_size,
959 		block size 512 */
960 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
961 	reset_nvmf_rdma_request(&rdma_req);
962 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
963 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
964 			  0, 0, 0, 0, 0);
965 	rdma_req.req.dif.dif_insert_or_strip = true;
966 	rtransport.transport.opts.io_unit_size = data_bs * 4;
967 	sgl->keyed.length = data_bs * 4;
968 
969 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
970 
971 	CU_ASSERT(rc == 0);
972 	CU_ASSERT(rdma_req.req.data_from_pool == true);
973 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
974 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
975 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
976 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
977 	CU_ASSERT(rdma_req.data.wr.num_sge == 5);
978 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
979 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
980 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
981 
982 	for (i = 0; i < 3; ++i) {
983 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
984 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
985 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
986 	}
987 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
988 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
989 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == g_rdma_mr.lkey);
990 
991 	/* 2nd buffer consumed */
992 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
993 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
994 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == g_rdma_mr.lkey);
995 
996 	/* Part 3: simple I/O, one SGL equal io unit size, io_unit_size is equal to block size 512 bytes */
997 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
998 	reset_nvmf_rdma_request(&rdma_req);
999 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1000 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1001 			  0, 0, 0, 0, 0);
1002 	rdma_req.req.dif.dif_insert_or_strip = true;
1003 	rtransport.transport.opts.io_unit_size = data_bs;
1004 	sgl->keyed.length = data_bs;
1005 
1006 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1007 
1008 	CU_ASSERT(rc == 0);
1009 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1010 	CU_ASSERT(rdma_req.req.length == data_bs);
1011 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1012 	CU_ASSERT(rdma_req.req.dif.elba_length == data_bs + md_size);
1013 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1014 	CU_ASSERT(rdma_req.data.wr.num_sge == 1);
1015 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1016 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1017 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1018 
1019 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1020 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == data_bs);
1021 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
1022 
1023 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1024 	CU_ASSERT(rdma_req.req.iov[0].iov_base == (void *)((unsigned long)0x2000));
1025 	CU_ASSERT(rdma_req.req.iov[0].iov_len == data_bs);
1026 	/* 2nd buffer consumed for metadata */
1027 	CU_ASSERT(rdma_req.req.iov[1].iov_base == (void *)((unsigned long)0x2000));
1028 	CU_ASSERT(rdma_req.req.iov[1].iov_len == md_size);
1029 
1030 	/* Part 4: simple I/O, one SGL equal io unit size, io_unit_size is aligned with md_size,
1031 	   block size 512 */
1032 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1033 	reset_nvmf_rdma_request(&rdma_req);
1034 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1035 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1036 			  0, 0, 0, 0, 0);
1037 	rdma_req.req.dif.dif_insert_or_strip = true;
1038 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1039 	sgl->keyed.length = data_bs * 4;
1040 
1041 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1042 
1043 	CU_ASSERT(rc == 0);
1044 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1045 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1046 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1047 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1048 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1049 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1050 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1051 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1052 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1053 
1054 	for (i = 0; i < 4; ++i) {
1055 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1056 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1057 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
1058 	}
1059 
1060 	/* Part 5: simple I/O, one SGL equal to 2x io unit size, io_unit_size is aligned with md_size,
1061 	   block size 512 */
1062 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1063 	reset_nvmf_rdma_request(&rdma_req);
1064 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1065 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1066 			  0, 0, 0, 0, 0);
1067 	rdma_req.req.dif.dif_insert_or_strip = true;
1068 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 2;
1069 	sgl->keyed.length = data_bs * 4;
1070 
1071 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1072 
1073 	CU_ASSERT(rc == 0);
1074 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1075 	CU_ASSERT(rdma_req.req.length == data_bs * 4);
1076 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1077 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4);
1078 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1079 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1080 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1081 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1082 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1083 
1084 	for (i = 0; i < 2; ++i) {
1085 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1086 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1087 	}
1088 	for (i = 0; i < 2; ++i) {
1089 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].addr == 0x2000 + i * (data_bs + md_size));
1090 		CU_ASSERT(rdma_req.data.wr.sg_list[i + 2].length == data_bs);
1091 	}
1092 
1093 	/* Part 6: simple I/O, one SGL larger than the transport io unit size, io_unit_size is not aligned to md_size,
1094 	   block size 512 */
1095 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1096 	reset_nvmf_rdma_request(&rdma_req);
1097 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1098 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1099 			  0, 0, 0, 0, 0);
1100 	rdma_req.req.dif.dif_insert_or_strip = true;
1101 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1102 	sgl->keyed.length = data_bs * 6;
1103 
1104 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1105 
1106 	CU_ASSERT(rc == 0);
1107 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1108 	CU_ASSERT(rdma_req.req.length == data_bs * 6);
1109 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1110 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 6);
1111 	CU_ASSERT((uint64_t)rdma_req.req.data == 0x2000);
1112 	CU_ASSERT(rdma_req.data.wr.num_sge == 7);
1113 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1114 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1115 	CU_ASSERT((uint64_t)rdma_req.req.buffers[0] == 0x2000);
1116 
1117 	for (i = 0; i < 3; ++i) {
1118 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == 0x2000 + i * (data_bs + md_size));
1119 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1120 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
1121 	}
1122 	CU_ASSERT(rdma_req.data.wr.sg_list[3].addr == 0x2000 + 3 * (data_bs + md_size));
1123 	CU_ASSERT(rdma_req.data.wr.sg_list[3].length == 488);
1124 	CU_ASSERT(rdma_req.data.wr.sg_list[3].lkey == g_rdma_mr.lkey);
1125 
1126 	/* 2nd IO buffer consumed */
1127 	CU_ASSERT(rdma_req.data.wr.sg_list[4].addr == 0x2000);
1128 	CU_ASSERT(rdma_req.data.wr.sg_list[4].length == 24);
1129 	CU_ASSERT(rdma_req.data.wr.sg_list[4].lkey == g_rdma_mr.lkey);
1130 
1131 	CU_ASSERT(rdma_req.data.wr.sg_list[5].addr == 0x2000 + 24 + md_size);
1132 	CU_ASSERT(rdma_req.data.wr.sg_list[5].length == 512);
1133 	CU_ASSERT(rdma_req.data.wr.sg_list[5].lkey == g_rdma_mr.lkey);
1134 
1135 	CU_ASSERT(rdma_req.data.wr.sg_list[6].addr == 0x2000 + 24 + 512 + md_size * 2);
1136 	CU_ASSERT(rdma_req.data.wr.sg_list[6].length == 512);
1137 	CU_ASSERT(rdma_req.data.wr.sg_list[6].lkey == g_rdma_mr.lkey);
1138 
1139 	/* Part 7: simple I/O, number of SGL entries exceeds the number of entries
1140 	   one WR can hold. Additional WR is chained */
1141 	MOCK_SET(spdk_mempool_get, &data);
1142 	aligned_buffer = (void *)((uintptr_t)((char *)&data + NVMF_DATA_BUFFER_MASK) &
1143 				  ~NVMF_DATA_BUFFER_MASK);
1144 	reset_nvmf_rdma_request(&rdma_req);
1145 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1146 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1147 			  0, 0, 0, 0, 0);
1148 	rdma_req.req.dif.dif_insert_or_strip = true;
1149 	rtransport.transport.opts.io_unit_size = data_bs * 16;
1150 	sgl->keyed.length = data_bs * 16;
1151 
1152 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1153 
1154 	CU_ASSERT(rc == 0);
1155 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1156 	CU_ASSERT(rdma_req.req.length == data_bs * 16);
1157 	CU_ASSERT(rdma_req.req.iovcnt == 2);
1158 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1159 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 16);
1160 	CU_ASSERT(rdma_req.req.data == aligned_buffer);
1161 	CU_ASSERT(rdma_req.data.wr.num_sge == 16);
1162 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1163 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1164 	/* additional wr from pool */
1165 	CU_ASSERT(rdma_req.data.wr.next == (void *)&data.wr);
1166 	CU_ASSERT(rdma_req.data.wr.next->num_sge == 1);
1167 	CU_ASSERT(rdma_req.data.wr.next->next == &rdma_req.rsp.wr);
1168 
1169 	/* Part 8: simple I/O, data with metadata do not fit to 1 io_buffer */
1170 	MOCK_SET(spdk_mempool_get, (void *)0x2000);
1171 	reset_nvmf_rdma_request(&rdma_req);
1172 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1173 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1174 			  0, 0, 0, 0, 0);
1175 	rdma_req.req.dif.dif_insert_or_strip = true;
1176 	rtransport.transport.opts.io_unit_size = 516;
1177 	sgl->keyed.length = data_bs * 2;
1178 
1179 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1180 
1181 	CU_ASSERT(rc == 0);
1182 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1183 	CU_ASSERT(rdma_req.req.length == data_bs * 2);
1184 	CU_ASSERT(rdma_req.req.iovcnt == 3);
1185 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1186 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 2);
1187 	CU_ASSERT(rdma_req.req.data == (void *)0x2000);
1188 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1189 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1190 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1191 
1192 	CU_ASSERT(rdma_req.data.wr.sg_list[0].addr == 0x2000);
1193 	CU_ASSERT(rdma_req.data.wr.sg_list[0].length == 512);
1194 	CU_ASSERT(rdma_req.data.wr.sg_list[0].lkey == g_rdma_mr.lkey);
1195 
1196 	/* 2nd IO buffer consumed, offset 4 bytes due to part of the metadata
1197 	  is located at the beginning of that buffer */
1198 	CU_ASSERT(rdma_req.data.wr.sg_list[1].addr == 0x2000 + 4);
1199 	CU_ASSERT(rdma_req.data.wr.sg_list[1].length == 512);
1200 	CU_ASSERT(rdma_req.data.wr.sg_list[1].lkey == g_rdma_mr.lkey);
1201 
1202 	/* Test 9 dealing with a buffer split over two Memory Regions */
1203 	MOCK_SET(spdk_mempool_get, (void *)&buffer);
1204 	reset_nvmf_rdma_request(&rdma_req);
1205 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1206 			  SPDK_DIF_TYPE1, SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK,
1207 			  0, 0, 0, 0, 0);
1208 	rdma_req.req.dif.dif_insert_or_strip = true;
1209 	rtransport.transport.opts.io_unit_size = data_bs * 4;
1210 	sgl->keyed.length = data_bs * 2;
1211 	g_mr_size = data_bs;
1212 	g_mr_next_size = rtransport.transport.opts.io_unit_size;
1213 
1214 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1215 	SPDK_CU_ASSERT_FATAL(rc == 0);
1216 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1217 	CU_ASSERT(rdma_req.req.length == rtransport.transport.opts.io_unit_size / 2);
1218 	CU_ASSERT((uint64_t)rdma_req.req.data == (((uint64_t)&buffer + NVMF_DATA_BUFFER_MASK) &
1219 			~NVMF_DATA_BUFFER_MASK));
1220 	CU_ASSERT(rdma_req.data.wr.num_sge == 2);
1221 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0xEEEE);
1222 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0xFFFF);
1223 	CU_ASSERT(rdma_req.req.buffers[0] == &buffer);
1224 	for (i = 0; i < 2; i++) {
1225 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uint64_t)rdma_req.req.data + i *
1226 			  (data_bs + md_size));
1227 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1228 		CU_ASSERT(rdma_req.data.wr.sg_list[i].lkey == g_rdma_mr.lkey);
1229 	}
1230 	buffer_ptr = STAILQ_FIRST(&group.retired_bufs);
1231 	CU_ASSERT(buffer_ptr == &buffer);
1232 	STAILQ_REMOVE(&group.retired_bufs, buffer_ptr, spdk_nvmf_transport_pg_cache_buf, link);
1233 	CU_ASSERT(STAILQ_EMPTY(&group.retired_bufs));
1234 	g_mr_size = 0;
1235 	g_mr_next_size = 0;
1236 
1237 	/* Test 2: Multi SGL */
1238 	sgl->generic.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1239 	sgl->unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1240 	sgl->address = 0;
1241 	rdma_req.recv->buf = (void *)&sgl_desc;
1242 	MOCK_SET(spdk_mempool_get, &data);
1243 	aligned_buffer = (void *)((uintptr_t)((char *)&data + NVMF_DATA_BUFFER_MASK) &
1244 				  ~NVMF_DATA_BUFFER_MASK);
1245 
1246 	/* part 1: 2 segments each with 1 wr. io_unit_size is aligned with data_bs + md_size */
1247 	reset_nvmf_rdma_request(&rdma_req);
1248 	spdk_dif_ctx_init(&rdma_req.req.dif.dif_ctx, data_bs + md_size, md_size, true, false,
1249 			  SPDK_DIF_TYPE1,
1250 			  SPDK_DIF_FLAGS_GUARD_CHECK | SPDK_DIF_FLAGS_REFTAG_CHECK, 0, 0, 0, 0, 0);
1251 	rdma_req.req.dif.dif_insert_or_strip = true;
1252 	rtransport.transport.opts.io_unit_size = (data_bs + md_size) * 4;
1253 	sgl->unkeyed.length = 2 * sizeof(struct spdk_nvme_sgl_descriptor);
1254 
1255 	for (i = 0; i < 2; i++) {
1256 		sgl_desc[i].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1257 		sgl_desc[i].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1258 		sgl_desc[i].keyed.length = data_bs * 4;
1259 		sgl_desc[i].address = 0x4000 + i * data_bs * 4;
1260 		sgl_desc[i].keyed.key = 0x44;
1261 	}
1262 
1263 	rc = nvmf_rdma_request_parse_sgl(&rtransport, &device, &rdma_req);
1264 
1265 	CU_ASSERT(rc == 0);
1266 	CU_ASSERT(rdma_req.req.data_from_pool == true);
1267 	CU_ASSERT(rdma_req.req.length == data_bs * 4 * 2);
1268 	CU_ASSERT(rdma_req.req.dif.orig_length == rdma_req.req.length);
1269 	CU_ASSERT(rdma_req.req.dif.elba_length == (data_bs + md_size) * 4 * 2);
1270 	CU_ASSERT(rdma_req.data.wr.num_sge == 4);
1271 	for (i = 0; i < 4; ++i) {
1272 		CU_ASSERT(rdma_req.data.wr.sg_list[i].addr == (uintptr_t)((unsigned char *)aligned_buffer) + i *
1273 			  (data_bs + md_size));
1274 		CU_ASSERT(rdma_req.data.wr.sg_list[i].length == data_bs);
1275 	}
1276 
1277 	CU_ASSERT(rdma_req.data.wr.wr.rdma.rkey == 0x44);
1278 	CU_ASSERT(rdma_req.data.wr.wr.rdma.remote_addr == 0x4000);
1279 	CU_ASSERT(rdma_req.data.wr.next == &data.wr);
1280 	CU_ASSERT(data.wr.wr.rdma.rkey == 0x44);
1281 	CU_ASSERT(data.wr.wr.rdma.remote_addr == 0x4000 + data_bs * 4);
1282 	CU_ASSERT(data.wr.num_sge == 4);
1283 	for (i = 0; i < 4; ++i) {
1284 		CU_ASSERT(data.wr.sg_list[i].addr == (uintptr_t)((unsigned char *)aligned_buffer) + i *
1285 			  (data_bs + md_size));
1286 		CU_ASSERT(data.wr.sg_list[i].length == data_bs);
1287 	}
1288 
1289 	CU_ASSERT(data.wr.next == &rdma_req.rsp.wr);
1290 }
1291 
1292 int main(int argc, char **argv)
1293 {
1294 	CU_pSuite	suite = NULL;
1295 	unsigned int	num_failures;
1296 
1297 	CU_set_error_action(CUEA_ABORT);
1298 	CU_initialize_registry();
1299 
1300 	suite = CU_add_suite("nvmf", NULL, NULL);
1301 
1302 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl);
1303 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_process);
1304 	CU_ADD_TEST(suite, test_nvmf_rdma_get_optimal_poll_group);
1305 	CU_ADD_TEST(suite, test_spdk_nvmf_rdma_request_parse_sgl_with_md);
1306 
1307 	CU_basic_set_mode(CU_BRM_VERBOSE);
1308 	CU_basic_run_tests();
1309 	num_failures = CU_get_number_of_failures();
1310 	CU_cleanup_registry();
1311 	return num_failures;
1312 }
1313