1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16#include "qemu/osdep.h"
17#include "qapi/qapi-events-rdma.h"
18
19#include <infiniband/verbs.h>
20
21#include "contrib/rdmacm-mux/rdmacm-mux.h"
22#include "trace.h"
23#include "rdma_utils.h"
24#include "rdma_rm.h"
25#include "rdma_backend.h"
26
27#define THR_NAME_LEN 16
28#define THR_POLL_TO 5000
29
30#define MAD_HDR_SIZE sizeof(struct ibv_grh)
31
32typedef struct BackendCtx {
33 void *up_ctx;
34 struct ibv_sge sge;
35 RdmaBackendQP *backend_qp;
36 RdmaBackendSRQ *backend_srq;
37} BackendCtx;
38
39struct backend_umad {
40 struct ib_user_mad hdr;
41 char mad[RDMA_MAX_PRIVATE_DATA];
42};
43
44static void (*comp_handler)(void *ctx, struct ibv_wc *wc);
45
46static void dummy_comp_handler(void *ctx, struct ibv_wc *wc)
47{
48 rdma_error_report("No completion handler is registered");
49}
50
51static inline void complete_work(enum ibv_wc_status status, uint32_t vendor_err,
52 void *ctx)
53{
54 struct ibv_wc wc = {};
55
56 wc.status = status;
57 wc.vendor_err = vendor_err;
58
59 comp_handler(ctx, &wc);
60}
61
62static void free_cqe_ctx(gpointer data, gpointer user_data)
63{
64 BackendCtx *bctx;
65 RdmaDeviceResources *rdma_dev_res = user_data;
66 unsigned long cqe_ctx_id = GPOINTER_TO_INT(data);
67
68 bctx = rdma_rm_get_cqe_ctx(rdma_dev_res, cqe_ctx_id);
69 if (bctx) {
70 rdma_rm_dealloc_cqe_ctx(rdma_dev_res, cqe_ctx_id);
71 atomic_dec(&rdma_dev_res->stats.missing_cqe);
72 }
73 g_free(bctx);
74}
75
76static void clean_recv_mads(RdmaBackendDev *backend_dev)
77{
78 unsigned long cqe_ctx_id;
79
80 do {
81 cqe_ctx_id = rdma_protected_qlist_pop_int64(&backend_dev->
82 recv_mads_list);
83 if (cqe_ctx_id != -ENOENT) {
84 atomic_inc(&backend_dev->rdma_dev_res->stats.missing_cqe);
85 free_cqe_ctx(GINT_TO_POINTER(cqe_ctx_id),
86 backend_dev->rdma_dev_res);
87 }
88 } while (cqe_ctx_id != -ENOENT);
89}
90
91static int rdma_poll_cq(RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq)
92{
93 int i, ne, total_ne = 0;
94 BackendCtx *bctx;
95 struct ibv_wc wc[2];
96 RdmaProtectedGSList *cqe_ctx_list;
97
98 qemu_mutex_lock(&rdma_dev_res->lock);
99 do {
100 ne = ibv_poll_cq(ibcq, ARRAY_SIZE(wc), wc);
101
102 trace_rdma_poll_cq(ne, ibcq);
103
104 for (i = 0; i < ne; i++) {
105 bctx = rdma_rm_get_cqe_ctx(rdma_dev_res, wc[i].wr_id);
106 if (unlikely(!bctx)) {
107 rdma_error_report("No matching ctx for req %"PRId64,
108 wc[i].wr_id);
109 continue;
110 }
111
112 comp_handler(bctx->up_ctx, &wc[i]);
113
114 if (bctx->backend_qp) {
115 cqe_ctx_list = &bctx->backend_qp->cqe_ctx_list;
116 } else {
117 cqe_ctx_list = &bctx->backend_srq->cqe_ctx_list;
118 }
119
120 rdma_protected_gslist_remove_int32(cqe_ctx_list, wc[i].wr_id);
121 rdma_rm_dealloc_cqe_ctx(rdma_dev_res, wc[i].wr_id);
122 g_free(bctx);
123 }
124 total_ne += ne;
125 } while (ne > 0);
126 atomic_sub(&rdma_dev_res->stats.missing_cqe, total_ne);
127 qemu_mutex_unlock(&rdma_dev_res->lock);
128
129 if (ne < 0) {
130 rdma_error_report("ibv_poll_cq fail, rc=%d, errno=%d", ne, errno);
131 }
132
133 rdma_dev_res->stats.completions += total_ne;
134
135 return total_ne;
136}
137
138static void *comp_handler_thread(void *arg)
139{
140 RdmaBackendDev *backend_dev = (RdmaBackendDev *)arg;
141 int rc;
142 struct ibv_cq *ev_cq;
143 void *ev_ctx;
144 int flags;
145 GPollFD pfds[1];
146
147
148 flags = fcntl(backend_dev->channel->fd, F_GETFL);
149 rc = fcntl(backend_dev->channel->fd, F_SETFL, flags | O_NONBLOCK);
150 if (rc < 0) {
151 rdma_error_report("Failed to change backend channel FD to non-blocking");
152 return NULL;
153 }
154
155 pfds[0].fd = backend_dev->channel->fd;
156 pfds[0].events = G_IO_IN | G_IO_HUP | G_IO_ERR;
157
158 backend_dev->comp_thread.is_running = true;
159
160 while (backend_dev->comp_thread.run) {
161 do {
162 rc = qemu_poll_ns(pfds, 1, THR_POLL_TO * (int64_t)SCALE_MS);
163 if (!rc) {
164 backend_dev->rdma_dev_res->stats.poll_cq_ppoll_to++;
165 }
166 } while (!rc && backend_dev->comp_thread.run);
167
168 if (backend_dev->comp_thread.run) {
169 rc = ibv_get_cq_event(backend_dev->channel, &ev_cq, &ev_ctx);
170 if (unlikely(rc)) {
171 rdma_error_report("ibv_get_cq_event fail, rc=%d, errno=%d", rc,
172 errno);
173 continue;
174 }
175
176 rc = ibv_req_notify_cq(ev_cq, 0);
177 if (unlikely(rc)) {
178 rdma_error_report("ibv_req_notify_cq fail, rc=%d, errno=%d", rc,
179 errno);
180 }
181
182 backend_dev->rdma_dev_res->stats.poll_cq_from_bk++;
183 rdma_poll_cq(backend_dev->rdma_dev_res, ev_cq);
184
185 ibv_ack_cq_events(ev_cq, 1);
186 }
187 }
188
189 backend_dev->comp_thread.is_running = false;
190
191 qemu_thread_exit(0);
192
193 return NULL;
194}
195
196static inline void disable_rdmacm_mux_async(RdmaBackendDev *backend_dev)
197{
198 atomic_set(&backend_dev->rdmacm_mux.can_receive, 0);
199}
200
201static inline void enable_rdmacm_mux_async(RdmaBackendDev *backend_dev)
202{
203 atomic_set(&backend_dev->rdmacm_mux.can_receive, sizeof(RdmaCmMuxMsg));
204}
205
206static inline int rdmacm_mux_can_process_async(RdmaBackendDev *backend_dev)
207{
208 return atomic_read(&backend_dev->rdmacm_mux.can_receive);
209}
210
211static int rdmacm_mux_check_op_status(CharBackend *mad_chr_be)
212{
213 RdmaCmMuxMsg msg = {};
214 int ret;
215
216 ret = qemu_chr_fe_read_all(mad_chr_be, (uint8_t *)&msg, sizeof(msg));
217 if (ret != sizeof(msg)) {
218 rdma_error_report("Got invalid message from mux: size %d, expecting %d",
219 ret, (int)sizeof(msg));
220 return -EIO;
221 }
222
223 trace_rdmacm_mux_check_op_status(msg.hdr.msg_type, msg.hdr.op_code,
224 msg.hdr.err_code);
225
226 if (msg.hdr.msg_type != RDMACM_MUX_MSG_TYPE_RESP) {
227 rdma_error_report("Got invalid message type %d", msg.hdr.msg_type);
228 return -EIO;
229 }
230
231 if (msg.hdr.err_code != RDMACM_MUX_ERR_CODE_OK) {
232 rdma_error_report("Operation failed in mux, error code %d",
233 msg.hdr.err_code);
234 return -EIO;
235 }
236
237 return 0;
238}
239
240static int rdmacm_mux_send(RdmaBackendDev *backend_dev, RdmaCmMuxMsg *msg)
241{
242 int rc = 0;
243
244 msg->hdr.msg_type = RDMACM_MUX_MSG_TYPE_REQ;
245 trace_rdmacm_mux("send", msg->hdr.msg_type, msg->hdr.op_code);
246 disable_rdmacm_mux_async(backend_dev);
247 rc = qemu_chr_fe_write(backend_dev->rdmacm_mux.chr_be,
248 (const uint8_t *)msg, sizeof(*msg));
249 if (rc != sizeof(*msg)) {
250 enable_rdmacm_mux_async(backend_dev);
251 rdma_error_report("Failed to send request to rdmacm_mux (rc=%d)", rc);
252 return -EIO;
253 }
254
255 rc = rdmacm_mux_check_op_status(backend_dev->rdmacm_mux.chr_be);
256 if (rc) {
257 rdma_error_report("Failed to execute rdmacm_mux request %d (rc=%d)",
258 msg->hdr.op_code, rc);
259 }
260
261 enable_rdmacm_mux_async(backend_dev);
262
263 return 0;
264}
265
266static void stop_backend_thread(RdmaBackendThread *thread)
267{
268 thread->run = false;
269 while (thread->is_running) {
270 sleep(THR_POLL_TO / SCALE_US / 2);
271 }
272}
273
274static void start_comp_thread(RdmaBackendDev *backend_dev)
275{
276 char thread_name[THR_NAME_LEN] = {};
277
278 stop_backend_thread(&backend_dev->comp_thread);
279
280 snprintf(thread_name, sizeof(thread_name), "rdma_comp_%s",
281 ibv_get_device_name(backend_dev->ib_dev));
282 backend_dev->comp_thread.run = true;
283 qemu_thread_create(&backend_dev->comp_thread.thread, thread_name,
284 comp_handler_thread, backend_dev, QEMU_THREAD_DETACHED);
285}
286
287void rdma_backend_register_comp_handler(void (*handler)(void *ctx,
288 struct ibv_wc *wc))
289{
290 comp_handler = handler;
291}
292
293void rdma_backend_unregister_comp_handler(void)
294{
295 rdma_backend_register_comp_handler(dummy_comp_handler);
296}
297
298int rdma_backend_query_port(RdmaBackendDev *backend_dev,
299 struct ibv_port_attr *port_attr)
300{
301 int rc;
302
303 rc = ibv_query_port(backend_dev->context, backend_dev->port_num, port_attr);
304 if (rc) {
305 rdma_error_report("ibv_query_port fail, rc=%d, errno=%d", rc, errno);
306 return -EIO;
307 }
308
309 return 0;
310}
311
312void rdma_backend_poll_cq(RdmaDeviceResources *rdma_dev_res, RdmaBackendCQ *cq)
313{
314 int polled;
315
316 rdma_dev_res->stats.poll_cq_from_guest++;
317 polled = rdma_poll_cq(rdma_dev_res, cq->ibcq);
318 if (!polled) {
319 rdma_dev_res->stats.poll_cq_from_guest_empty++;
320 }
321}
322
323static GHashTable *ah_hash;
324
325static struct ibv_ah *create_ah(RdmaBackendDev *backend_dev, struct ibv_pd *pd,
326 uint8_t sgid_idx, union ibv_gid *dgid)
327{
328 GBytes *ah_key = g_bytes_new(dgid, sizeof(*dgid));
329 struct ibv_ah *ah = g_hash_table_lookup(ah_hash, ah_key);
330
331 if (ah) {
332 trace_rdma_create_ah_cache_hit(be64_to_cpu(dgid->global.subnet_prefix),
333 be64_to_cpu(dgid->global.interface_id));
334 g_bytes_unref(ah_key);
335 } else {
336 struct ibv_ah_attr ah_attr = {
337 .is_global = 1,
338 .port_num = backend_dev->port_num,
339 .grh.hop_limit = 1,
340 };
341
342 ah_attr.grh.dgid = *dgid;
343 ah_attr.grh.sgid_index = sgid_idx;
344
345 ah = ibv_create_ah(pd, &ah_attr);
346 if (ah) {
347 g_hash_table_insert(ah_hash, ah_key, ah);
348 } else {
349 g_bytes_unref(ah_key);
350 rdma_error_report("Failed to create AH for gid <0x%" PRIx64", 0x%"PRIx64">",
351 be64_to_cpu(dgid->global.subnet_prefix),
352 be64_to_cpu(dgid->global.interface_id));
353 }
354
355 trace_rdma_create_ah_cache_miss(be64_to_cpu(dgid->global.subnet_prefix),
356 be64_to_cpu(dgid->global.interface_id));
357 }
358
359 return ah;
360}
361
362static void destroy_ah_hash_key(gpointer data)
363{
364 g_bytes_unref(data);
365}
366
367static void destroy_ah_hast_data(gpointer data)
368{
369 struct ibv_ah *ah = data;
370
371 ibv_destroy_ah(ah);
372}
373
374static void ah_cache_init(void)
375{
376 ah_hash = g_hash_table_new_full(g_bytes_hash, g_bytes_equal,
377 destroy_ah_hash_key, destroy_ah_hast_data);
378}
379
380static int build_host_sge_array(RdmaDeviceResources *rdma_dev_res,
381 struct ibv_sge *dsge, struct ibv_sge *ssge,
382 uint8_t num_sge, uint64_t *total_length)
383{
384 RdmaRmMR *mr;
385 int ssge_idx;
386
387 for (ssge_idx = 0; ssge_idx < num_sge; ssge_idx++) {
388 mr = rdma_rm_get_mr(rdma_dev_res, ssge[ssge_idx].lkey);
389 if (unlikely(!mr)) {
390 rdma_error_report("Invalid lkey 0x%x", ssge[ssge_idx].lkey);
391 return VENDOR_ERR_INVLKEY | ssge[ssge_idx].lkey;
392 }
393
394 dsge->addr = (uintptr_t)mr->virt + ssge[ssge_idx].addr - mr->start;
395 dsge->length = ssge[ssge_idx].length;
396 dsge->lkey = rdma_backend_mr_lkey(&mr->backend_mr);
397
398 *total_length += dsge->length;
399
400 dsge++;
401 }
402
403 return 0;
404}
405
406static void trace_mad_message(const char *title, char *buf, int len)
407{
408 int i;
409 char *b = g_malloc0(len * 3 + 1);
410 char b1[4];
411
412 for (i = 0; i < len; i++) {
413 sprintf(b1, "%.2X ", buf[i] & 0x000000FF);
414 strcat(b, b1);
415 }
416
417 trace_rdma_mad_message(title, len, b);
418
419 g_free(b);
420}
421
422static int mad_send(RdmaBackendDev *backend_dev, uint8_t sgid_idx,
423 union ibv_gid *sgid, struct ibv_sge *sge, uint32_t num_sge)
424{
425 RdmaCmMuxMsg msg = {};
426 char *hdr, *data;
427 int ret;
428
429 if (num_sge != 2) {
430 return -EINVAL;
431 }
432
433 msg.hdr.op_code = RDMACM_MUX_OP_CODE_MAD;
434 memcpy(msg.hdr.sgid.raw, sgid->raw, sizeof(msg.hdr.sgid));
435
436 msg.umad_len = sge[0].length + sge[1].length;
437
438 if (msg.umad_len > sizeof(msg.umad.mad)) {
439 return -ENOMEM;
440 }
441
442 msg.umad.hdr.addr.qpn = htobe32(1);
443 msg.umad.hdr.addr.grh_present = 1;
444 msg.umad.hdr.addr.gid_index = sgid_idx;
445 memcpy(msg.umad.hdr.addr.gid, sgid->raw, sizeof(msg.umad.hdr.addr.gid));
446 msg.umad.hdr.addr.hop_limit = 0xFF;
447
448 hdr = rdma_pci_dma_map(backend_dev->dev, sge[0].addr, sge[0].length);
449 if (!hdr) {
450 return -ENOMEM;
451 }
452 data = rdma_pci_dma_map(backend_dev->dev, sge[1].addr, sge[1].length);
453 if (!data) {
454 rdma_pci_dma_unmap(backend_dev->dev, hdr, sge[0].length);
455 return -ENOMEM;
456 }
457
458 memcpy(&msg.umad.mad[0], hdr, sge[0].length);
459 memcpy(&msg.umad.mad[sge[0].length], data, sge[1].length);
460
461 rdma_pci_dma_unmap(backend_dev->dev, data, sge[1].length);
462 rdma_pci_dma_unmap(backend_dev->dev, hdr, sge[0].length);
463
464 trace_mad_message("send", msg.umad.mad, msg.umad_len);
465
466 ret = rdmacm_mux_send(backend_dev, &msg);
467 if (ret) {
468 rdma_error_report("Failed to send MAD to rdma_umadmux (%d)", ret);
469 return -EIO;
470 }
471
472 return 0;
473}
474
475void rdma_backend_post_send(RdmaBackendDev *backend_dev,
476 RdmaBackendQP *qp, uint8_t qp_type,
477 struct ibv_sge *sge, uint32_t num_sge,
478 uint8_t sgid_idx, union ibv_gid *sgid,
479 union ibv_gid *dgid, uint32_t dqpn, uint32_t dqkey,
480 void *ctx)
481{
482 BackendCtx *bctx;
483 struct ibv_sge new_sge[MAX_SGE];
484 uint32_t bctx_id;
485 int rc;
486 struct ibv_send_wr wr = {}, *bad_wr;
487
488 if (!qp->ibqp) {
489 if (qp_type == IBV_QPT_SMI) {
490 rdma_error_report("Got QP0 request");
491 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_QP0, ctx);
492 } else if (qp_type == IBV_QPT_GSI) {
493 rc = mad_send(backend_dev, sgid_idx, sgid, sge, num_sge);
494 if (rc) {
495 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_MAD_SEND, ctx);
496 backend_dev->rdma_dev_res->stats.mad_tx_err++;
497 } else {
498 complete_work(IBV_WC_SUCCESS, 0, ctx);
499 backend_dev->rdma_dev_res->stats.mad_tx++;
500 }
501 }
502 return;
503 }
504
505 bctx = g_malloc0(sizeof(*bctx));
506 bctx->up_ctx = ctx;
507 bctx->backend_qp = qp;
508
509 rc = rdma_rm_alloc_cqe_ctx(backend_dev->rdma_dev_res, &bctx_id, bctx);
510 if (unlikely(rc)) {
511 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_NOMEM, ctx);
512 goto err_free_bctx;
513 }
514
515 rdma_protected_gslist_append_int32(&qp->cqe_ctx_list, bctx_id);
516
517 rc = build_host_sge_array(backend_dev->rdma_dev_res, new_sge, sge, num_sge,
518 &backend_dev->rdma_dev_res->stats.tx_len);
519 if (rc) {
520 complete_work(IBV_WC_GENERAL_ERR, rc, ctx);
521 goto err_dealloc_cqe_ctx;
522 }
523
524 if (qp_type == IBV_QPT_UD) {
525 wr.wr.ud.ah = create_ah(backend_dev, qp->ibpd, sgid_idx, dgid);
526 if (!wr.wr.ud.ah) {
527 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_FAIL_BACKEND, ctx);
528 goto err_dealloc_cqe_ctx;
529 }
530 wr.wr.ud.remote_qpn = dqpn;
531 wr.wr.ud.remote_qkey = dqkey;
532 }
533
534 wr.num_sge = num_sge;
535 wr.opcode = IBV_WR_SEND;
536 wr.send_flags = IBV_SEND_SIGNALED;
537 wr.sg_list = new_sge;
538 wr.wr_id = bctx_id;
539
540 rc = ibv_post_send(qp->ibqp, &wr, &bad_wr);
541 if (rc) {
542 rdma_error_report("ibv_post_send fail, qpn=0x%x, rc=%d, errno=%d",
543 qp->ibqp->qp_num, rc, errno);
544 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_FAIL_BACKEND, ctx);
545 goto err_dealloc_cqe_ctx;
546 }
547
548 atomic_inc(&backend_dev->rdma_dev_res->stats.missing_cqe);
549 backend_dev->rdma_dev_res->stats.tx++;
550
551 return;
552
553err_dealloc_cqe_ctx:
554 backend_dev->rdma_dev_res->stats.tx_err++;
555 rdma_rm_dealloc_cqe_ctx(backend_dev->rdma_dev_res, bctx_id);
556
557err_free_bctx:
558 g_free(bctx);
559}
560
561static unsigned int save_mad_recv_buffer(RdmaBackendDev *backend_dev,
562 struct ibv_sge *sge, uint32_t num_sge,
563 void *ctx)
564{
565 BackendCtx *bctx;
566 int rc;
567 uint32_t bctx_id;
568
569 if (num_sge != 1) {
570 rdma_error_report("Invalid num_sge (%d), expecting 1", num_sge);
571 return VENDOR_ERR_INV_NUM_SGE;
572 }
573
574 if (sge[0].length < RDMA_MAX_PRIVATE_DATA + sizeof(struct ibv_grh)) {
575 rdma_error_report("Too small buffer for MAD");
576 return VENDOR_ERR_INV_MAD_BUFF;
577 }
578
579 bctx = g_malloc0(sizeof(*bctx));
580
581 rc = rdma_rm_alloc_cqe_ctx(backend_dev->rdma_dev_res, &bctx_id, bctx);
582 if (unlikely(rc)) {
583 g_free(bctx);
584 return VENDOR_ERR_NOMEM;
585 }
586
587 bctx->up_ctx = ctx;
588 bctx->sge = *sge;
589
590 rdma_protected_qlist_append_int64(&backend_dev->recv_mads_list, bctx_id);
591
592 return 0;
593}
594
595void rdma_backend_post_recv(RdmaBackendDev *backend_dev,
596 RdmaBackendQP *qp, uint8_t qp_type,
597 struct ibv_sge *sge, uint32_t num_sge, void *ctx)
598{
599 BackendCtx *bctx;
600 struct ibv_sge new_sge[MAX_SGE];
601 uint32_t bctx_id;
602 int rc;
603 struct ibv_recv_wr wr = {}, *bad_wr;
604
605 if (!qp->ibqp) {
606 if (qp_type == IBV_QPT_SMI) {
607 rdma_error_report("Got QP0 request");
608 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_QP0, ctx);
609 }
610 if (qp_type == IBV_QPT_GSI) {
611 rc = save_mad_recv_buffer(backend_dev, sge, num_sge, ctx);
612 if (rc) {
613 complete_work(IBV_WC_GENERAL_ERR, rc, ctx);
614 backend_dev->rdma_dev_res->stats.mad_rx_bufs_err++;
615 } else {
616 backend_dev->rdma_dev_res->stats.mad_rx_bufs++;
617 }
618 }
619 return;
620 }
621
622 bctx = g_malloc0(sizeof(*bctx));
623 bctx->up_ctx = ctx;
624 bctx->backend_qp = qp;
625
626 rc = rdma_rm_alloc_cqe_ctx(backend_dev->rdma_dev_res, &bctx_id, bctx);
627 if (unlikely(rc)) {
628 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_NOMEM, ctx);
629 goto err_free_bctx;
630 }
631
632 rdma_protected_gslist_append_int32(&qp->cqe_ctx_list, bctx_id);
633
634 rc = build_host_sge_array(backend_dev->rdma_dev_res, new_sge, sge, num_sge,
635 &backend_dev->rdma_dev_res->stats.rx_bufs_len);
636 if (rc) {
637 complete_work(IBV_WC_GENERAL_ERR, rc, ctx);
638 goto err_dealloc_cqe_ctx;
639 }
640
641 wr.num_sge = num_sge;
642 wr.sg_list = new_sge;
643 wr.wr_id = bctx_id;
644 rc = ibv_post_recv(qp->ibqp, &wr, &bad_wr);
645 if (rc) {
646 rdma_error_report("ibv_post_recv fail, qpn=0x%x, rc=%d, errno=%d",
647 qp->ibqp->qp_num, rc, errno);
648 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_FAIL_BACKEND, ctx);
649 goto err_dealloc_cqe_ctx;
650 }
651
652 atomic_inc(&backend_dev->rdma_dev_res->stats.missing_cqe);
653 backend_dev->rdma_dev_res->stats.rx_bufs++;
654
655 return;
656
657err_dealloc_cqe_ctx:
658 backend_dev->rdma_dev_res->stats.rx_bufs_err++;
659 rdma_rm_dealloc_cqe_ctx(backend_dev->rdma_dev_res, bctx_id);
660
661err_free_bctx:
662 g_free(bctx);
663}
664
665void rdma_backend_post_srq_recv(RdmaBackendDev *backend_dev,
666 RdmaBackendSRQ *srq, struct ibv_sge *sge,
667 uint32_t num_sge, void *ctx)
668{
669 BackendCtx *bctx;
670 struct ibv_sge new_sge[MAX_SGE];
671 uint32_t bctx_id;
672 int rc;
673 struct ibv_recv_wr wr = {}, *bad_wr;
674
675 bctx = g_malloc0(sizeof(*bctx));
676 bctx->up_ctx = ctx;
677 bctx->backend_srq = srq;
678
679 rc = rdma_rm_alloc_cqe_ctx(backend_dev->rdma_dev_res, &bctx_id, bctx);
680 if (unlikely(rc)) {
681 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_NOMEM, ctx);
682 goto err_free_bctx;
683 }
684
685 rdma_protected_gslist_append_int32(&srq->cqe_ctx_list, bctx_id);
686
687 rc = build_host_sge_array(backend_dev->rdma_dev_res, new_sge, sge, num_sge,
688 &backend_dev->rdma_dev_res->stats.rx_bufs_len);
689 if (rc) {
690 complete_work(IBV_WC_GENERAL_ERR, rc, ctx);
691 goto err_dealloc_cqe_ctx;
692 }
693
694 wr.num_sge = num_sge;
695 wr.sg_list = new_sge;
696 wr.wr_id = bctx_id;
697 rc = ibv_post_srq_recv(srq->ibsrq, &wr, &bad_wr);
698 if (rc) {
699 rdma_error_report("ibv_post_srq_recv fail, srqn=0x%x, rc=%d, errno=%d",
700 srq->ibsrq->handle, rc, errno);
701 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_FAIL_BACKEND, ctx);
702 goto err_dealloc_cqe_ctx;
703 }
704
705 atomic_inc(&backend_dev->rdma_dev_res->stats.missing_cqe);
706 backend_dev->rdma_dev_res->stats.rx_bufs++;
707 backend_dev->rdma_dev_res->stats.rx_srq++;
708
709 return;
710
711err_dealloc_cqe_ctx:
712 backend_dev->rdma_dev_res->stats.rx_bufs_err++;
713 rdma_rm_dealloc_cqe_ctx(backend_dev->rdma_dev_res, bctx_id);
714
715err_free_bctx:
716 g_free(bctx);
717}
718
719int rdma_backend_create_pd(RdmaBackendDev *backend_dev, RdmaBackendPD *pd)
720{
721 pd->ibpd = ibv_alloc_pd(backend_dev->context);
722
723 if (!pd->ibpd) {
724 rdma_error_report("ibv_alloc_pd fail, errno=%d", errno);
725 return -EIO;
726 }
727
728 return 0;
729}
730
731void rdma_backend_destroy_pd(RdmaBackendPD *pd)
732{
733 if (pd->ibpd) {
734 ibv_dealloc_pd(pd->ibpd);
735 }
736}
737
738int rdma_backend_create_mr(RdmaBackendMR *mr, RdmaBackendPD *pd, void *addr,
739 size_t length, int access)
740{
741 mr->ibmr = ibv_reg_mr(pd->ibpd, addr, length, access);
742 if (!mr->ibmr) {
743 rdma_error_report("ibv_reg_mr fail, errno=%d", errno);
744 return -EIO;
745 }
746
747 mr->ibpd = pd->ibpd;
748
749 return 0;
750}
751
752void rdma_backend_destroy_mr(RdmaBackendMR *mr)
753{
754 if (mr->ibmr) {
755 ibv_dereg_mr(mr->ibmr);
756 }
757}
758
759int rdma_backend_create_cq(RdmaBackendDev *backend_dev, RdmaBackendCQ *cq,
760 int cqe)
761{
762 int rc;
763
764 cq->ibcq = ibv_create_cq(backend_dev->context, cqe + 1, NULL,
765 backend_dev->channel, 0);
766 if (!cq->ibcq) {
767 rdma_error_report("ibv_create_cq fail, errno=%d", errno);
768 return -EIO;
769 }
770
771 rc = ibv_req_notify_cq(cq->ibcq, 0);
772 if (rc) {
773 rdma_warn_report("ibv_req_notify_cq fail, rc=%d, errno=%d", rc, errno);
774 }
775
776 cq->backend_dev = backend_dev;
777
778 return 0;
779}
780
781void rdma_backend_destroy_cq(RdmaBackendCQ *cq)
782{
783 if (cq->ibcq) {
784 ibv_destroy_cq(cq->ibcq);
785 }
786}
787
788int rdma_backend_create_qp(RdmaBackendQP *qp, uint8_t qp_type,
789 RdmaBackendPD *pd, RdmaBackendCQ *scq,
790 RdmaBackendCQ *rcq, RdmaBackendSRQ *srq,
791 uint32_t max_send_wr, uint32_t max_recv_wr,
792 uint32_t max_send_sge, uint32_t max_recv_sge)
793{
794 struct ibv_qp_init_attr attr = {};
795
796 qp->ibqp = 0;
797
798 switch (qp_type) {
799 case IBV_QPT_GSI:
800 return 0;
801
802 case IBV_QPT_RC:
803
804 case IBV_QPT_UD:
805
806 break;
807
808 default:
809 rdma_error_report("Unsupported QP type %d", qp_type);
810 return -EIO;
811 }
812
813 attr.qp_type = qp_type;
814 attr.send_cq = scq->ibcq;
815 attr.recv_cq = rcq->ibcq;
816 attr.cap.max_send_wr = max_send_wr;
817 attr.cap.max_recv_wr = max_recv_wr;
818 attr.cap.max_send_sge = max_send_sge;
819 attr.cap.max_recv_sge = max_recv_sge;
820 if (srq) {
821 attr.srq = srq->ibsrq;
822 }
823
824 qp->ibqp = ibv_create_qp(pd->ibpd, &attr);
825 if (!qp->ibqp) {
826 rdma_error_report("ibv_create_qp fail, errno=%d", errno);
827 return -EIO;
828 }
829
830 rdma_protected_gslist_init(&qp->cqe_ctx_list);
831
832 qp->ibpd = pd->ibpd;
833
834
835
836 return 0;
837}
838
839int rdma_backend_qp_state_init(RdmaBackendDev *backend_dev, RdmaBackendQP *qp,
840 uint8_t qp_type, uint32_t qkey)
841{
842 struct ibv_qp_attr attr = {};
843 int rc, attr_mask;
844
845 attr_mask = IBV_QP_STATE | IBV_QP_PKEY_INDEX | IBV_QP_PORT;
846 attr.qp_state = IBV_QPS_INIT;
847 attr.pkey_index = 0;
848 attr.port_num = backend_dev->port_num;
849
850 switch (qp_type) {
851 case IBV_QPT_RC:
852 attr_mask |= IBV_QP_ACCESS_FLAGS;
853 trace_rdma_backend_rc_qp_state_init(qp->ibqp->qp_num);
854 break;
855
856 case IBV_QPT_UD:
857 attr.qkey = qkey;
858 attr_mask |= IBV_QP_QKEY;
859 trace_rdma_backend_ud_qp_state_init(qp->ibqp->qp_num, qkey);
860 break;
861
862 default:
863 rdma_error_report("Unsupported QP type %d", qp_type);
864 return -EIO;
865 }
866
867 rc = ibv_modify_qp(qp->ibqp, &attr, attr_mask);
868 if (rc) {
869 rdma_error_report("ibv_modify_qp fail, rc=%d, errno=%d", rc, errno);
870 return -EIO;
871 }
872
873 return 0;
874}
875
876int rdma_backend_qp_state_rtr(RdmaBackendDev *backend_dev, RdmaBackendQP *qp,
877 uint8_t qp_type, uint8_t sgid_idx,
878 union ibv_gid *dgid, uint32_t dqpn,
879 uint32_t rq_psn, uint32_t qkey, bool use_qkey)
880{
881 struct ibv_qp_attr attr = {};
882 union ibv_gid ibv_gid = {
883 .global.interface_id = dgid->global.interface_id,
884 .global.subnet_prefix = dgid->global.subnet_prefix
885 };
886 int rc, attr_mask;
887
888 attr.qp_state = IBV_QPS_RTR;
889 attr_mask = IBV_QP_STATE;
890
891 qp->sgid_idx = sgid_idx;
892
893 switch (qp_type) {
894 case IBV_QPT_RC:
895 attr.path_mtu = IBV_MTU_1024;
896 attr.dest_qp_num = dqpn;
897 attr.max_dest_rd_atomic = 1;
898 attr.min_rnr_timer = 12;
899 attr.ah_attr.port_num = backend_dev->port_num;
900 attr.ah_attr.is_global = 1;
901 attr.ah_attr.grh.hop_limit = 1;
902 attr.ah_attr.grh.dgid = ibv_gid;
903 attr.ah_attr.grh.sgid_index = qp->sgid_idx;
904 attr.rq_psn = rq_psn;
905
906 attr_mask |= IBV_QP_AV | IBV_QP_PATH_MTU | IBV_QP_DEST_QPN |
907 IBV_QP_RQ_PSN | IBV_QP_MAX_DEST_RD_ATOMIC |
908 IBV_QP_MIN_RNR_TIMER;
909
910 trace_rdma_backend_rc_qp_state_rtr(qp->ibqp->qp_num,
911 be64_to_cpu(ibv_gid.global.
912 subnet_prefix),
913 be64_to_cpu(ibv_gid.global.
914 interface_id),
915 qp->sgid_idx, dqpn, rq_psn);
916 break;
917
918 case IBV_QPT_UD:
919 if (use_qkey) {
920 attr.qkey = qkey;
921 attr_mask |= IBV_QP_QKEY;
922 }
923 trace_rdma_backend_ud_qp_state_rtr(qp->ibqp->qp_num, use_qkey ? qkey :
924 0);
925 break;
926 }
927
928 rc = ibv_modify_qp(qp->ibqp, &attr, attr_mask);
929 if (rc) {
930 rdma_error_report("ibv_modify_qp fail, rc=%d, errno=%d", rc, errno);
931 return -EIO;
932 }
933
934 return 0;
935}
936
937int rdma_backend_qp_state_rts(RdmaBackendQP *qp, uint8_t qp_type,
938 uint32_t sq_psn, uint32_t qkey, bool use_qkey)
939{
940 struct ibv_qp_attr attr = {};
941 int rc, attr_mask;
942
943 attr.qp_state = IBV_QPS_RTS;
944 attr.sq_psn = sq_psn;
945 attr_mask = IBV_QP_STATE | IBV_QP_SQ_PSN;
946
947 switch (qp_type) {
948 case IBV_QPT_RC:
949 attr.timeout = 14;
950 attr.retry_cnt = 7;
951 attr.rnr_retry = 7;
952 attr.max_rd_atomic = 1;
953
954 attr_mask |= IBV_QP_TIMEOUT | IBV_QP_RETRY_CNT | IBV_QP_RNR_RETRY |
955 IBV_QP_MAX_QP_RD_ATOMIC;
956 trace_rdma_backend_rc_qp_state_rts(qp->ibqp->qp_num, sq_psn);
957 break;
958
959 case IBV_QPT_UD:
960 if (use_qkey) {
961 attr.qkey = qkey;
962 attr_mask |= IBV_QP_QKEY;
963 }
964 trace_rdma_backend_ud_qp_state_rts(qp->ibqp->qp_num, sq_psn,
965 use_qkey ? qkey : 0);
966 break;
967 }
968
969 rc = ibv_modify_qp(qp->ibqp, &attr, attr_mask);
970 if (rc) {
971 rdma_error_report("ibv_modify_qp fail, rc=%d, errno=%d", rc, errno);
972 return -EIO;
973 }
974
975 return 0;
976}
977
978int rdma_backend_query_qp(RdmaBackendQP *qp, struct ibv_qp_attr *attr,
979 int attr_mask, struct ibv_qp_init_attr *init_attr)
980{
981 if (!qp->ibqp) {
982 attr->qp_state = IBV_QPS_RTS;
983 return 0;
984 }
985
986 return ibv_query_qp(qp->ibqp, attr, attr_mask, init_attr);
987}
988
989void rdma_backend_destroy_qp(RdmaBackendQP *qp, RdmaDeviceResources *dev_res)
990{
991 if (qp->ibqp) {
992 ibv_destroy_qp(qp->ibqp);
993 }
994 g_slist_foreach(qp->cqe_ctx_list.list, free_cqe_ctx, dev_res);
995 rdma_protected_gslist_destroy(&qp->cqe_ctx_list);
996}
997
998int rdma_backend_create_srq(RdmaBackendSRQ *srq, RdmaBackendPD *pd,
999 uint32_t max_wr, uint32_t max_sge,
1000 uint32_t srq_limit)
1001{
1002 struct ibv_srq_init_attr srq_init_attr = {};
1003
1004 srq_init_attr.attr.max_wr = max_wr;
1005 srq_init_attr.attr.max_sge = max_sge;
1006 srq_init_attr.attr.srq_limit = srq_limit;
1007
1008 srq->ibsrq = ibv_create_srq(pd->ibpd, &srq_init_attr);
1009 if (!srq->ibsrq) {
1010 rdma_error_report("ibv_create_srq failed, errno=%d", errno);
1011 return -EIO;
1012 }
1013
1014 rdma_protected_gslist_init(&srq->cqe_ctx_list);
1015
1016 return 0;
1017}
1018
1019int rdma_backend_query_srq(RdmaBackendSRQ *srq, struct ibv_srq_attr *srq_attr)
1020{
1021 if (!srq->ibsrq) {
1022 return -EINVAL;
1023 }
1024
1025 return ibv_query_srq(srq->ibsrq, srq_attr);
1026}
1027
1028int rdma_backend_modify_srq(RdmaBackendSRQ *srq, struct ibv_srq_attr *srq_attr,
1029 int srq_attr_mask)
1030{
1031 if (!srq->ibsrq) {
1032 return -EINVAL;
1033 }
1034
1035 return ibv_modify_srq(srq->ibsrq, srq_attr, srq_attr_mask);
1036}
1037
1038void rdma_backend_destroy_srq(RdmaBackendSRQ *srq, RdmaDeviceResources *dev_res)
1039{
1040 if (srq->ibsrq) {
1041 ibv_destroy_srq(srq->ibsrq);
1042 }
1043 g_slist_foreach(srq->cqe_ctx_list.list, free_cqe_ctx, dev_res);
1044 rdma_protected_gslist_destroy(&srq->cqe_ctx_list);
1045}
1046
1047#define CHK_ATTR(req, dev, member, fmt) ({ \
1048 trace_rdma_check_dev_attr(#member, dev.member, req->member); \
1049 if (req->member > dev.member) { \
1050 rdma_warn_report("%s = "fmt" is higher than host device capability "fmt, \
1051 #member, req->member, dev.member); \
1052 req->member = dev.member; \
1053 } \
1054})
1055
1056static int init_device_caps(RdmaBackendDev *backend_dev,
1057 struct ibv_device_attr *dev_attr)
1058{
1059 struct ibv_device_attr bk_dev_attr;
1060 int rc;
1061
1062 rc = ibv_query_device(backend_dev->context, &bk_dev_attr);
1063 if (rc) {
1064 rdma_error_report("ibv_query_device fail, rc=%d, errno=%d", rc, errno);
1065 return -EIO;
1066 }
1067
1068 dev_attr->max_sge = MAX_SGE;
1069 dev_attr->max_srq_sge = MAX_SGE;
1070
1071 CHK_ATTR(dev_attr, bk_dev_attr, max_mr_size, "%" PRId64);
1072 CHK_ATTR(dev_attr, bk_dev_attr, max_qp, "%d");
1073 CHK_ATTR(dev_attr, bk_dev_attr, max_sge, "%d");
1074 CHK_ATTR(dev_attr, bk_dev_attr, max_cq, "%d");
1075 CHK_ATTR(dev_attr, bk_dev_attr, max_mr, "%d");
1076 CHK_ATTR(dev_attr, bk_dev_attr, max_pd, "%d");
1077 CHK_ATTR(dev_attr, bk_dev_attr, max_qp_rd_atom, "%d");
1078 CHK_ATTR(dev_attr, bk_dev_attr, max_qp_init_rd_atom, "%d");
1079 CHK_ATTR(dev_attr, bk_dev_attr, max_ah, "%d");
1080 CHK_ATTR(dev_attr, bk_dev_attr, max_srq, "%d");
1081
1082 return 0;
1083}
1084
1085static inline void build_mad_hdr(struct ibv_grh *grh, union ibv_gid *sgid,
1086 union ibv_gid *my_gid, int paylen)
1087{
1088 grh->paylen = htons(paylen);
1089 grh->sgid = *sgid;
1090 grh->dgid = *my_gid;
1091}
1092
1093static void process_incoming_mad_req(RdmaBackendDev *backend_dev,
1094 RdmaCmMuxMsg *msg)
1095{
1096 unsigned long cqe_ctx_id;
1097 BackendCtx *bctx;
1098 char *mad;
1099
1100 trace_mad_message("recv", msg->umad.mad, msg->umad_len);
1101
1102 cqe_ctx_id = rdma_protected_qlist_pop_int64(&backend_dev->recv_mads_list);
1103 if (cqe_ctx_id == -ENOENT) {
1104 rdma_warn_report("No more free MADs buffers, waiting for a while");
1105 sleep(THR_POLL_TO);
1106 return;
1107 }
1108
1109 bctx = rdma_rm_get_cqe_ctx(backend_dev->rdma_dev_res, cqe_ctx_id);
1110 if (unlikely(!bctx)) {
1111 rdma_error_report("No matching ctx for req %ld", cqe_ctx_id);
1112 backend_dev->rdma_dev_res->stats.mad_rx_err++;
1113 return;
1114 }
1115
1116 mad = rdma_pci_dma_map(backend_dev->dev, bctx->sge.addr,
1117 bctx->sge.length);
1118 if (!mad || bctx->sge.length < msg->umad_len + MAD_HDR_SIZE) {
1119 backend_dev->rdma_dev_res->stats.mad_rx_err++;
1120 complete_work(IBV_WC_GENERAL_ERR, VENDOR_ERR_INV_MAD_BUFF,
1121 bctx->up_ctx);
1122 } else {
1123 struct ibv_wc wc = {};
1124 memset(mad, 0, bctx->sge.length);
1125 build_mad_hdr((struct ibv_grh *)mad,
1126 (union ibv_gid *)&msg->umad.hdr.addr.gid, &msg->hdr.sgid,
1127 msg->umad_len);
1128 memcpy(&mad[MAD_HDR_SIZE], msg->umad.mad, msg->umad_len);
1129 rdma_pci_dma_unmap(backend_dev->dev, mad, bctx->sge.length);
1130
1131 wc.byte_len = msg->umad_len;
1132 wc.status = IBV_WC_SUCCESS;
1133 wc.wc_flags = IBV_WC_GRH;
1134 backend_dev->rdma_dev_res->stats.mad_rx++;
1135 comp_handler(bctx->up_ctx, &wc);
1136 }
1137
1138 g_free(bctx);
1139 rdma_rm_dealloc_cqe_ctx(backend_dev->rdma_dev_res, cqe_ctx_id);
1140}
1141
1142static inline int rdmacm_mux_can_receive(void *opaque)
1143{
1144 RdmaBackendDev *backend_dev = (RdmaBackendDev *)opaque;
1145
1146 return rdmacm_mux_can_process_async(backend_dev);
1147}
1148
1149static void rdmacm_mux_read(void *opaque, const uint8_t *buf, int size)
1150{
1151 RdmaBackendDev *backend_dev = (RdmaBackendDev *)opaque;
1152 RdmaCmMuxMsg *msg = (RdmaCmMuxMsg *)buf;
1153
1154 trace_rdmacm_mux("read", msg->hdr.msg_type, msg->hdr.op_code);
1155
1156 if (msg->hdr.msg_type != RDMACM_MUX_MSG_TYPE_REQ &&
1157 msg->hdr.op_code != RDMACM_MUX_OP_CODE_MAD) {
1158 rdma_error_report("Error: Not a MAD request, skipping");
1159 return;
1160 }
1161 process_incoming_mad_req(backend_dev, msg);
1162}
1163
1164static int mad_init(RdmaBackendDev *backend_dev, CharBackend *mad_chr_be)
1165{
1166 int ret;
1167
1168 backend_dev->rdmacm_mux.chr_be = mad_chr_be;
1169
1170 ret = qemu_chr_fe_backend_connected(backend_dev->rdmacm_mux.chr_be);
1171 if (!ret) {
1172 rdma_error_report("Missing chardev for MAD multiplexer");
1173 return -EIO;
1174 }
1175
1176 rdma_protected_qlist_init(&backend_dev->recv_mads_list);
1177
1178 enable_rdmacm_mux_async(backend_dev);
1179
1180 qemu_chr_fe_set_handlers(backend_dev->rdmacm_mux.chr_be,
1181 rdmacm_mux_can_receive, rdmacm_mux_read, NULL,
1182 NULL, backend_dev, NULL, true);
1183
1184 return 0;
1185}
1186
1187static void mad_stop(RdmaBackendDev *backend_dev)
1188{
1189 clean_recv_mads(backend_dev);
1190}
1191
1192static void mad_fini(RdmaBackendDev *backend_dev)
1193{
1194 disable_rdmacm_mux_async(backend_dev);
1195 qemu_chr_fe_disconnect(backend_dev->rdmacm_mux.chr_be);
1196 rdma_protected_qlist_destroy(&backend_dev->recv_mads_list);
1197}
1198
1199int rdma_backend_get_gid_index(RdmaBackendDev *backend_dev,
1200 union ibv_gid *gid)
1201{
1202 union ibv_gid sgid;
1203 int ret;
1204 int i = 0;
1205
1206 do {
1207 ret = ibv_query_gid(backend_dev->context, backend_dev->port_num, i,
1208 &sgid);
1209 i++;
1210 } while (!ret && (memcmp(&sgid, gid, sizeof(*gid))));
1211
1212 trace_rdma_backend_get_gid_index(be64_to_cpu(gid->global.subnet_prefix),
1213 be64_to_cpu(gid->global.interface_id),
1214 i - 1);
1215
1216 return ret ? ret : i - 1;
1217}
1218
1219int rdma_backend_add_gid(RdmaBackendDev *backend_dev, const char *ifname,
1220 union ibv_gid *gid)
1221{
1222 RdmaCmMuxMsg msg = {};
1223 int ret;
1224
1225 trace_rdma_backend_gid_change("add", be64_to_cpu(gid->global.subnet_prefix),
1226 be64_to_cpu(gid->global.interface_id));
1227
1228 msg.hdr.op_code = RDMACM_MUX_OP_CODE_REG;
1229 memcpy(msg.hdr.sgid.raw, gid->raw, sizeof(msg.hdr.sgid));
1230
1231 ret = rdmacm_mux_send(backend_dev, &msg);
1232 if (ret) {
1233 rdma_error_report("Failed to register GID to rdma_umadmux (%d)", ret);
1234 return -EIO;
1235 }
1236
1237 qapi_event_send_rdma_gid_status_changed(ifname, true,
1238 gid->global.subnet_prefix,
1239 gid->global.interface_id);
1240
1241 return ret;
1242}
1243
1244int rdma_backend_del_gid(RdmaBackendDev *backend_dev, const char *ifname,
1245 union ibv_gid *gid)
1246{
1247 RdmaCmMuxMsg msg = {};
1248 int ret;
1249
1250 trace_rdma_backend_gid_change("del", be64_to_cpu(gid->global.subnet_prefix),
1251 be64_to_cpu(gid->global.interface_id));
1252
1253 msg.hdr.op_code = RDMACM_MUX_OP_CODE_UNREG;
1254 memcpy(msg.hdr.sgid.raw, gid->raw, sizeof(msg.hdr.sgid));
1255
1256 ret = rdmacm_mux_send(backend_dev, &msg);
1257 if (ret) {
1258 rdma_error_report("Failed to unregister GID from rdma_umadmux (%d)",
1259 ret);
1260 return -EIO;
1261 }
1262
1263 qapi_event_send_rdma_gid_status_changed(ifname, false,
1264 gid->global.subnet_prefix,
1265 gid->global.interface_id);
1266
1267 return 0;
1268}
1269
1270int rdma_backend_init(RdmaBackendDev *backend_dev, PCIDevice *pdev,
1271 RdmaDeviceResources *rdma_dev_res,
1272 const char *backend_device_name, uint8_t port_num,
1273 struct ibv_device_attr *dev_attr, CharBackend *mad_chr_be)
1274{
1275 int i;
1276 int ret = 0;
1277 int num_ibv_devices;
1278 struct ibv_device **dev_list;
1279
1280 memset(backend_dev, 0, sizeof(*backend_dev));
1281
1282 backend_dev->dev = pdev;
1283 backend_dev->port_num = port_num;
1284 backend_dev->rdma_dev_res = rdma_dev_res;
1285
1286 rdma_backend_register_comp_handler(dummy_comp_handler);
1287
1288 dev_list = ibv_get_device_list(&num_ibv_devices);
1289 if (!dev_list) {
1290 rdma_error_report("Failed to get IB devices list");
1291 return -EIO;
1292 }
1293
1294 if (num_ibv_devices == 0) {
1295 rdma_error_report("No IB devices were found");
1296 ret = -ENXIO;
1297 goto out_free_dev_list;
1298 }
1299
1300 if (backend_device_name) {
1301 for (i = 0; dev_list[i]; ++i) {
1302 if (!strcmp(ibv_get_device_name(dev_list[i]),
1303 backend_device_name)) {
1304 break;
1305 }
1306 }
1307
1308 backend_dev->ib_dev = dev_list[i];
1309 if (!backend_dev->ib_dev) {
1310 rdma_error_report("Failed to find IB device %s",
1311 backend_device_name);
1312 ret = -EIO;
1313 goto out_free_dev_list;
1314 }
1315 } else {
1316 backend_dev->ib_dev = *dev_list;
1317 }
1318
1319 rdma_info_report("uverb device %s", backend_dev->ib_dev->dev_name);
1320
1321 backend_dev->context = ibv_open_device(backend_dev->ib_dev);
1322 if (!backend_dev->context) {
1323 rdma_error_report("Failed to open IB device %s",
1324 ibv_get_device_name(backend_dev->ib_dev));
1325 ret = -EIO;
1326 goto out;
1327 }
1328
1329 backend_dev->channel = ibv_create_comp_channel(backend_dev->context);
1330 if (!backend_dev->channel) {
1331 rdma_error_report("Failed to create IB communication channel");
1332 ret = -EIO;
1333 goto out_close_device;
1334 }
1335
1336 ret = init_device_caps(backend_dev, dev_attr);
1337 if (ret) {
1338 rdma_error_report("Failed to initialize device capabilities");
1339 ret = -EIO;
1340 goto out_destroy_comm_channel;
1341 }
1342
1343
1344 ret = mad_init(backend_dev, mad_chr_be);
1345 if (ret) {
1346 rdma_error_report("Failed to initialize mad");
1347 ret = -EIO;
1348 goto out_destroy_comm_channel;
1349 }
1350
1351 backend_dev->comp_thread.run = false;
1352 backend_dev->comp_thread.is_running = false;
1353
1354 ah_cache_init();
1355
1356 goto out_free_dev_list;
1357
1358out_destroy_comm_channel:
1359 ibv_destroy_comp_channel(backend_dev->channel);
1360
1361out_close_device:
1362 ibv_close_device(backend_dev->context);
1363
1364out_free_dev_list:
1365 ibv_free_device_list(dev_list);
1366
1367out:
1368 return ret;
1369}
1370
1371
1372void rdma_backend_start(RdmaBackendDev *backend_dev)
1373{
1374 start_comp_thread(backend_dev);
1375}
1376
1377void rdma_backend_stop(RdmaBackendDev *backend_dev)
1378{
1379 mad_stop(backend_dev);
1380 stop_backend_thread(&backend_dev->comp_thread);
1381}
1382
1383void rdma_backend_fini(RdmaBackendDev *backend_dev)
1384{
1385 mad_fini(backend_dev);
1386 g_hash_table_destroy(ah_hash);
1387 ibv_destroy_comp_channel(backend_dev->channel);
1388 ibv_close_device(backend_dev->context);
1389}
1390
