xref: /openbmc/qemu/hw/rdma/vmw/pvrdma_cmd.c (revision 812b31d3)
1 /*
2  * QEMU paravirtual RDMA - Command channel
3  *
4  * Copyright (C) 2018 Oracle
5  * Copyright (C) 2018 Red Hat Inc
6  *
7  * Authors:
8  *     Yuval Shaia <yuval.shaia@oracle.com>
9  *     Marcel Apfelbaum <marcel@redhat.com>
10  *
11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
12  * See the COPYING file in the top-level directory.
13  *
14  */
15 
16 #include "qemu/osdep.h"
17 #include "cpu.h"
18 #include "hw/pci/pci.h"
19 #include "hw/pci/pci_ids.h"
20 
21 #include "../rdma_backend.h"
22 #include "../rdma_rm.h"
23 #include "../rdma_utils.h"
24 
25 #include "trace.h"
26 #include "pvrdma.h"
27 #include "standard-headers/rdma/vmw_pvrdma-abi.h"
28 
29 static void *pvrdma_map_to_pdir(PCIDevice *pdev, uint64_t pdir_dma,
30                                 uint32_t nchunks, size_t length)
31 {
32     uint64_t *dir, *tbl;
33     int tbl_idx, dir_idx, addr_idx;
34     void *host_virt = NULL, *curr_page;
35 
36     if (!nchunks) {
37         rdma_error_report("Got nchunks=0");
38         return NULL;
39     }
40 
41     length = ROUND_UP(length, TARGET_PAGE_SIZE);
42     if (nchunks * TARGET_PAGE_SIZE != length) {
43         rdma_error_report("Invalid nchunks/length (%u, %lu)", nchunks,
44                           (unsigned long)length);
45         return NULL;
46     }
47 
48     dir = rdma_pci_dma_map(pdev, pdir_dma, TARGET_PAGE_SIZE);
49     if (!dir) {
50         rdma_error_report("Failed to map to page directory");
51         return NULL;
52     }
53 
54     tbl = rdma_pci_dma_map(pdev, dir[0], TARGET_PAGE_SIZE);
55     if (!tbl) {
56         rdma_error_report("Failed to map to page table 0");
57         goto out_unmap_dir;
58     }
59 
60     curr_page = rdma_pci_dma_map(pdev, (dma_addr_t)tbl[0], TARGET_PAGE_SIZE);
61     if (!curr_page) {
62         rdma_error_report("Failed to map the page 0");
63         goto out_unmap_tbl;
64     }
65 
66     host_virt = mremap(curr_page, 0, length, MREMAP_MAYMOVE);
67     if (host_virt == MAP_FAILED) {
68         host_virt = NULL;
69         rdma_error_report("Failed to remap memory for host_virt");
70         goto out_unmap_tbl;
71     }
72     trace_pvrdma_map_to_pdir_host_virt(curr_page, host_virt);
73 
74     rdma_pci_dma_unmap(pdev, curr_page, TARGET_PAGE_SIZE);
75 
76     dir_idx = 0;
77     tbl_idx = 1;
78     addr_idx = 1;
79     while (addr_idx < nchunks) {
80         if (tbl_idx == TARGET_PAGE_SIZE / sizeof(uint64_t)) {
81             tbl_idx = 0;
82             dir_idx++;
83             rdma_pci_dma_unmap(pdev, tbl, TARGET_PAGE_SIZE);
84             tbl = rdma_pci_dma_map(pdev, dir[dir_idx], TARGET_PAGE_SIZE);
85             if (!tbl) {
86                 rdma_error_report("Failed to map to page table %d", dir_idx);
87                 goto out_unmap_host_virt;
88             }
89         }
90 
91         curr_page = rdma_pci_dma_map(pdev, (dma_addr_t)tbl[tbl_idx],
92                                      TARGET_PAGE_SIZE);
93         if (!curr_page) {
94             rdma_error_report("Failed to map to page %d, dir %d", tbl_idx,
95                               dir_idx);
96             goto out_unmap_host_virt;
97         }
98 
99         mremap(curr_page, 0, TARGET_PAGE_SIZE, MREMAP_MAYMOVE | MREMAP_FIXED,
100                host_virt + TARGET_PAGE_SIZE * addr_idx);
101 
102         trace_pvrdma_map_to_pdir_next_page(addr_idx, curr_page, host_virt +
103                                            TARGET_PAGE_SIZE * addr_idx);
104 
105         rdma_pci_dma_unmap(pdev, curr_page, TARGET_PAGE_SIZE);
106 
107         addr_idx++;
108 
109         tbl_idx++;
110     }
111 
112     goto out_unmap_tbl;
113 
114 out_unmap_host_virt:
115     munmap(host_virt, length);
116     host_virt = NULL;
117 
118 out_unmap_tbl:
119     rdma_pci_dma_unmap(pdev, tbl, TARGET_PAGE_SIZE);
120 
121 out_unmap_dir:
122     rdma_pci_dma_unmap(pdev, dir, TARGET_PAGE_SIZE);
123 
124     return host_virt;
125 }
126 
127 static int query_port(PVRDMADev *dev, union pvrdma_cmd_req *req,
128                       union pvrdma_cmd_resp *rsp)
129 {
130     struct pvrdma_cmd_query_port *cmd = &req->query_port;
131     struct pvrdma_cmd_query_port_resp *resp = &rsp->query_port_resp;
132     struct pvrdma_port_attr attrs = {};
133 
134     if (cmd->port_num > MAX_PORTS) {
135         return -EINVAL;
136     }
137 
138     if (rdma_backend_query_port(&dev->backend_dev,
139                                 (struct ibv_port_attr *)&attrs)) {
140         return -ENOMEM;
141     }
142 
143     memset(resp, 0, sizeof(*resp));
144 
145     resp->attrs.state = dev->func0->device_active ? attrs.state :
146                                                     PVRDMA_PORT_DOWN;
147     resp->attrs.max_mtu = attrs.max_mtu;
148     resp->attrs.active_mtu = attrs.active_mtu;
149     resp->attrs.phys_state = attrs.phys_state;
150     resp->attrs.gid_tbl_len = MIN(MAX_PORT_GIDS, attrs.gid_tbl_len);
151     resp->attrs.max_msg_sz = 1024;
152     resp->attrs.pkey_tbl_len = MIN(MAX_PORT_PKEYS, attrs.pkey_tbl_len);
153     resp->attrs.active_width = 1;
154     resp->attrs.active_speed = 1;
155 
156     return 0;
157 }
158 
159 static int query_pkey(PVRDMADev *dev, union pvrdma_cmd_req *req,
160                       union pvrdma_cmd_resp *rsp)
161 {
162     struct pvrdma_cmd_query_pkey *cmd = &req->query_pkey;
163     struct pvrdma_cmd_query_pkey_resp *resp = &rsp->query_pkey_resp;
164 
165     if (cmd->port_num > MAX_PORTS) {
166         return -EINVAL;
167     }
168 
169     if (cmd->index > MAX_PKEYS) {
170         return -EINVAL;
171     }
172 
173     memset(resp, 0, sizeof(*resp));
174 
175     resp->pkey = PVRDMA_PKEY;
176 
177     return 0;
178 }
179 
180 static int create_pd(PVRDMADev *dev, union pvrdma_cmd_req *req,
181                      union pvrdma_cmd_resp *rsp)
182 {
183     struct pvrdma_cmd_create_pd *cmd = &req->create_pd;
184     struct pvrdma_cmd_create_pd_resp *resp = &rsp->create_pd_resp;
185     int rc;
186 
187     memset(resp, 0, sizeof(*resp));
188     rc = rdma_rm_alloc_pd(&dev->rdma_dev_res, &dev->backend_dev,
189                           &resp->pd_handle, cmd->ctx_handle);
190 
191     return rc;
192 }
193 
194 static int destroy_pd(PVRDMADev *dev, union pvrdma_cmd_req *req,
195                       union pvrdma_cmd_resp *rsp)
196 {
197     struct pvrdma_cmd_destroy_pd *cmd = &req->destroy_pd;
198 
199     rdma_rm_dealloc_pd(&dev->rdma_dev_res, cmd->pd_handle);
200 
201     return 0;
202 }
203 
204 static int create_mr(PVRDMADev *dev, union pvrdma_cmd_req *req,
205                      union pvrdma_cmd_resp *rsp)
206 {
207     struct pvrdma_cmd_create_mr *cmd = &req->create_mr;
208     struct pvrdma_cmd_create_mr_resp *resp = &rsp->create_mr_resp;
209     PCIDevice *pci_dev = PCI_DEVICE(dev);
210     void *host_virt = NULL;
211     int rc = 0;
212 
213     memset(resp, 0, sizeof(*resp));
214 
215     if (!(cmd->flags & PVRDMA_MR_FLAG_DMA)) {
216         host_virt = pvrdma_map_to_pdir(pci_dev, cmd->pdir_dma, cmd->nchunks,
217                                        cmd->length);
218         if (!host_virt) {
219             rdma_error_report("Failed to map to pdir");
220             return -EINVAL;
221         }
222     }
223 
224     rc = rdma_rm_alloc_mr(&dev->rdma_dev_res, cmd->pd_handle, cmd->start,
225                           cmd->length, host_virt, cmd->access_flags,
226                           &resp->mr_handle, &resp->lkey, &resp->rkey);
227     if (rc && host_virt) {
228         munmap(host_virt, cmd->length);
229     }
230 
231     return rc;
232 }
233 
234 static int destroy_mr(PVRDMADev *dev, union pvrdma_cmd_req *req,
235                       union pvrdma_cmd_resp *rsp)
236 {
237     struct pvrdma_cmd_destroy_mr *cmd = &req->destroy_mr;
238 
239     rdma_rm_dealloc_mr(&dev->rdma_dev_res, cmd->mr_handle);
240 
241     return 0;
242 }
243 
244 static int create_cq_ring(PCIDevice *pci_dev , PvrdmaRing **ring,
245                           uint64_t pdir_dma, uint32_t nchunks, uint32_t cqe)
246 {
247     uint64_t *dir = NULL, *tbl = NULL;
248     PvrdmaRing *r;
249     int rc = -EINVAL;
250     char ring_name[MAX_RING_NAME_SZ];
251 
252     if (!nchunks || nchunks > PVRDMA_MAX_FAST_REG_PAGES) {
253         rdma_error_report("Got invalid nchunks: %d", nchunks);
254         return rc;
255     }
256 
257     dir = rdma_pci_dma_map(pci_dev, pdir_dma, TARGET_PAGE_SIZE);
258     if (!dir) {
259         rdma_error_report("Failed to map to CQ page directory");
260         goto out;
261     }
262 
263     tbl = rdma_pci_dma_map(pci_dev, dir[0], TARGET_PAGE_SIZE);
264     if (!tbl) {
265         rdma_error_report("Failed to map to CQ page table");
266         goto out;
267     }
268 
269     r = g_malloc(sizeof(*r));
270     *ring = r;
271 
272     r->ring_state = (PvrdmaRingState *)
273         rdma_pci_dma_map(pci_dev, tbl[0], TARGET_PAGE_SIZE);
274 
275     if (!r->ring_state) {
276         rdma_error_report("Failed to map to CQ ring state");
277         goto out_free_ring;
278     }
279 
280     sprintf(ring_name, "cq_ring_%" PRIx64, pdir_dma);
281     rc = pvrdma_ring_init(r, ring_name, pci_dev, &r->ring_state[1],
282                           cqe, sizeof(struct pvrdma_cqe),
283                           /* first page is ring state */
284                           (dma_addr_t *)&tbl[1], nchunks - 1);
285     if (rc) {
286         goto out_unmap_ring_state;
287     }
288 
289     goto out;
290 
291 out_unmap_ring_state:
292     /* ring_state was in slot 1, not 0 so need to jump back */
293     rdma_pci_dma_unmap(pci_dev, --r->ring_state, TARGET_PAGE_SIZE);
294 
295 out_free_ring:
296     g_free(r);
297 
298 out:
299     rdma_pci_dma_unmap(pci_dev, tbl, TARGET_PAGE_SIZE);
300     rdma_pci_dma_unmap(pci_dev, dir, TARGET_PAGE_SIZE);
301 
302     return rc;
303 }
304 
305 static void destroy_cq_ring(PvrdmaRing *ring)
306 {
307     pvrdma_ring_free(ring);
308     /* ring_state was in slot 1, not 0 so need to jump back */
309     rdma_pci_dma_unmap(ring->dev, --ring->ring_state, TARGET_PAGE_SIZE);
310     g_free(ring);
311 }
312 
313 static int create_cq(PVRDMADev *dev, union pvrdma_cmd_req *req,
314                      union pvrdma_cmd_resp *rsp)
315 {
316     struct pvrdma_cmd_create_cq *cmd = &req->create_cq;
317     struct pvrdma_cmd_create_cq_resp *resp = &rsp->create_cq_resp;
318     PvrdmaRing *ring = NULL;
319     int rc;
320 
321     memset(resp, 0, sizeof(*resp));
322 
323     resp->cqe = cmd->cqe;
324 
325     rc = create_cq_ring(PCI_DEVICE(dev), &ring, cmd->pdir_dma, cmd->nchunks,
326                         cmd->cqe);
327     if (rc) {
328         return rc;
329     }
330 
331     rc = rdma_rm_alloc_cq(&dev->rdma_dev_res, &dev->backend_dev, cmd->cqe,
332                           &resp->cq_handle, ring);
333     if (rc) {
334         destroy_cq_ring(ring);
335     }
336 
337     resp->cqe = cmd->cqe;
338 
339     return rc;
340 }
341 
342 static int destroy_cq(PVRDMADev *dev, union pvrdma_cmd_req *req,
343                       union pvrdma_cmd_resp *rsp)
344 {
345     struct pvrdma_cmd_destroy_cq *cmd = &req->destroy_cq;
346     RdmaRmCQ *cq;
347     PvrdmaRing *ring;
348 
349     cq = rdma_rm_get_cq(&dev->rdma_dev_res, cmd->cq_handle);
350     if (!cq) {
351         rdma_error_report("Got invalid CQ handle");
352         return -EINVAL;
353     }
354 
355     ring = (PvrdmaRing *)cq->opaque;
356     destroy_cq_ring(ring);
357 
358     rdma_rm_dealloc_cq(&dev->rdma_dev_res, cmd->cq_handle);
359 
360     return 0;
361 }
362 
363 static int create_qp_rings(PCIDevice *pci_dev, uint64_t pdir_dma,
364                            PvrdmaRing **rings, uint32_t scqe, uint32_t smax_sge,
365                            uint32_t spages, uint32_t rcqe, uint32_t rmax_sge,
366                            uint32_t rpages, uint8_t is_srq)
367 {
368     uint64_t *dir = NULL, *tbl = NULL;
369     PvrdmaRing *sr, *rr;
370     int rc = -EINVAL;
371     char ring_name[MAX_RING_NAME_SZ];
372     uint32_t wqe_sz;
373 
374     if (!spages || spages > PVRDMA_MAX_FAST_REG_PAGES) {
375         rdma_error_report("Got invalid send page count for QP ring: %d",
376                           spages);
377         return rc;
378     }
379 
380     if (!is_srq && (!rpages || rpages > PVRDMA_MAX_FAST_REG_PAGES)) {
381         rdma_error_report("Got invalid recv page count for QP ring: %d",
382                           rpages);
383         return rc;
384     }
385 
386     dir = rdma_pci_dma_map(pci_dev, pdir_dma, TARGET_PAGE_SIZE);
387     if (!dir) {
388         rdma_error_report("Failed to map to QP page directory");
389         goto out;
390     }
391 
392     tbl = rdma_pci_dma_map(pci_dev, dir[0], TARGET_PAGE_SIZE);
393     if (!tbl) {
394         rdma_error_report("Failed to map to QP page table");
395         goto out;
396     }
397 
398     if (!is_srq) {
399         sr = g_malloc(2 * sizeof(*rr));
400         rr = &sr[1];
401     } else {
402         sr = g_malloc(sizeof(*sr));
403     }
404 
405     *rings = sr;
406 
407     /* Create send ring */
408     sr->ring_state = (PvrdmaRingState *)
409         rdma_pci_dma_map(pci_dev, tbl[0], TARGET_PAGE_SIZE);
410     if (!sr->ring_state) {
411         rdma_error_report("Failed to map to QP ring state");
412         goto out_free_sr_mem;
413     }
414 
415     wqe_sz = pow2ceil(sizeof(struct pvrdma_sq_wqe_hdr) +
416                       sizeof(struct pvrdma_sge) * smax_sge - 1);
417 
418     sprintf(ring_name, "qp_sring_%" PRIx64, pdir_dma);
419     rc = pvrdma_ring_init(sr, ring_name, pci_dev, sr->ring_state,
420                           scqe, wqe_sz, (dma_addr_t *)&tbl[1], spages);
421     if (rc) {
422         goto out_unmap_ring_state;
423     }
424 
425     if (!is_srq) {
426         /* Create recv ring */
427         rr->ring_state = &sr->ring_state[1];
428         wqe_sz = pow2ceil(sizeof(struct pvrdma_rq_wqe_hdr) +
429                           sizeof(struct pvrdma_sge) * rmax_sge - 1);
430         sprintf(ring_name, "qp_rring_%" PRIx64, pdir_dma);
431         rc = pvrdma_ring_init(rr, ring_name, pci_dev, rr->ring_state,
432                               rcqe, wqe_sz, (dma_addr_t *)&tbl[1 + spages],
433                               rpages);
434         if (rc) {
435             goto out_free_sr;
436         }
437     }
438 
439     goto out;
440 
441 out_free_sr:
442     pvrdma_ring_free(sr);
443 
444 out_unmap_ring_state:
445     rdma_pci_dma_unmap(pci_dev, sr->ring_state, TARGET_PAGE_SIZE);
446 
447 out_free_sr_mem:
448     g_free(sr);
449 
450 out:
451     rdma_pci_dma_unmap(pci_dev, tbl, TARGET_PAGE_SIZE);
452     rdma_pci_dma_unmap(pci_dev, dir, TARGET_PAGE_SIZE);
453 
454     return rc;
455 }
456 
457 static void destroy_qp_rings(PvrdmaRing *ring, uint8_t is_srq)
458 {
459     pvrdma_ring_free(&ring[0]);
460     if (!is_srq) {
461         pvrdma_ring_free(&ring[1]);
462     }
463 
464     rdma_pci_dma_unmap(ring->dev, ring->ring_state, TARGET_PAGE_SIZE);
465     g_free(ring);
466 }
467 
468 static int create_qp(PVRDMADev *dev, union pvrdma_cmd_req *req,
469                      union pvrdma_cmd_resp *rsp)
470 {
471     struct pvrdma_cmd_create_qp *cmd = &req->create_qp;
472     struct pvrdma_cmd_create_qp_resp *resp = &rsp->create_qp_resp;
473     PvrdmaRing *rings = NULL;
474     int rc;
475 
476     memset(resp, 0, sizeof(*resp));
477 
478     rc = create_qp_rings(PCI_DEVICE(dev), cmd->pdir_dma, &rings,
479                          cmd->max_send_wr, cmd->max_send_sge, cmd->send_chunks,
480                          cmd->max_recv_wr, cmd->max_recv_sge,
481                          cmd->total_chunks - cmd->send_chunks - 1, cmd->is_srq);
482     if (rc) {
483         return rc;
484     }
485 
486     rc = rdma_rm_alloc_qp(&dev->rdma_dev_res, cmd->pd_handle, cmd->qp_type,
487                           cmd->max_send_wr, cmd->max_send_sge,
488                           cmd->send_cq_handle, cmd->max_recv_wr,
489                           cmd->max_recv_sge, cmd->recv_cq_handle, rings,
490                           &resp->qpn, cmd->is_srq, cmd->srq_handle);
491     if (rc) {
492         destroy_qp_rings(rings, cmd->is_srq);
493         return rc;
494     }
495 
496     resp->max_send_wr = cmd->max_send_wr;
497     resp->max_recv_wr = cmd->max_recv_wr;
498     resp->max_send_sge = cmd->max_send_sge;
499     resp->max_recv_sge = cmd->max_recv_sge;
500     resp->max_inline_data = cmd->max_inline_data;
501 
502     return 0;
503 }
504 
505 static int modify_qp(PVRDMADev *dev, union pvrdma_cmd_req *req,
506                      union pvrdma_cmd_resp *rsp)
507 {
508     struct pvrdma_cmd_modify_qp *cmd = &req->modify_qp;
509     int rc;
510 
511     /* No need to verify sgid_index since it is u8 */
512 
513     rc = rdma_rm_modify_qp(&dev->rdma_dev_res, &dev->backend_dev,
514                            cmd->qp_handle, cmd->attr_mask,
515                            cmd->attrs.ah_attr.grh.sgid_index,
516                            (union ibv_gid *)&cmd->attrs.ah_attr.grh.dgid,
517                            cmd->attrs.dest_qp_num,
518                            (enum ibv_qp_state)cmd->attrs.qp_state,
519                            cmd->attrs.qkey, cmd->attrs.rq_psn,
520                            cmd->attrs.sq_psn);
521 
522     return rc;
523 }
524 
525 static int query_qp(PVRDMADev *dev, union pvrdma_cmd_req *req,
526                      union pvrdma_cmd_resp *rsp)
527 {
528     struct pvrdma_cmd_query_qp *cmd = &req->query_qp;
529     struct pvrdma_cmd_query_qp_resp *resp = &rsp->query_qp_resp;
530     struct ibv_qp_init_attr init_attr;
531     int rc;
532 
533     memset(resp, 0, sizeof(*resp));
534 
535     rc = rdma_rm_query_qp(&dev->rdma_dev_res, &dev->backend_dev, cmd->qp_handle,
536                           (struct ibv_qp_attr *)&resp->attrs, cmd->attr_mask,
537                           &init_attr);
538 
539     return rc;
540 }
541 
542 static int destroy_qp(PVRDMADev *dev, union pvrdma_cmd_req *req,
543                       union pvrdma_cmd_resp *rsp)
544 {
545     struct pvrdma_cmd_destroy_qp *cmd = &req->destroy_qp;
546     RdmaRmQP *qp;
547     PvrdmaRing *ring;
548 
549     qp = rdma_rm_get_qp(&dev->rdma_dev_res, cmd->qp_handle);
550     if (!qp) {
551         return -EINVAL;
552     }
553 
554     ring = (PvrdmaRing *)qp->opaque;
555     destroy_qp_rings(ring, qp->is_srq);
556     rdma_rm_dealloc_qp(&dev->rdma_dev_res, cmd->qp_handle);
557 
558     return 0;
559 }
560 
561 static int create_bind(PVRDMADev *dev, union pvrdma_cmd_req *req,
562                        union pvrdma_cmd_resp *rsp)
563 {
564     struct pvrdma_cmd_create_bind *cmd = &req->create_bind;
565     int rc;
566     union ibv_gid *gid = (union ibv_gid *)&cmd->new_gid;
567 
568     if (cmd->index >= MAX_PORT_GIDS) {
569         return -EINVAL;
570     }
571 
572     rc = rdma_rm_add_gid(&dev->rdma_dev_res, &dev->backend_dev,
573                          dev->backend_eth_device_name, gid, cmd->index);
574 
575     return rc;
576 }
577 
578 static int destroy_bind(PVRDMADev *dev, union pvrdma_cmd_req *req,
579                         union pvrdma_cmd_resp *rsp)
580 {
581     int rc;
582 
583     struct pvrdma_cmd_destroy_bind *cmd = &req->destroy_bind;
584 
585     if (cmd->index >= MAX_PORT_GIDS) {
586         return -EINVAL;
587     }
588 
589     rc = rdma_rm_del_gid(&dev->rdma_dev_res, &dev->backend_dev,
590                         dev->backend_eth_device_name, cmd->index);
591 
592     return rc;
593 }
594 
595 static int create_uc(PVRDMADev *dev, union pvrdma_cmd_req *req,
596                      union pvrdma_cmd_resp *rsp)
597 {
598     struct pvrdma_cmd_create_uc *cmd = &req->create_uc;
599     struct pvrdma_cmd_create_uc_resp *resp = &rsp->create_uc_resp;
600     int rc;
601 
602     memset(resp, 0, sizeof(*resp));
603     rc = rdma_rm_alloc_uc(&dev->rdma_dev_res, cmd->pfn, &resp->ctx_handle);
604 
605     return rc;
606 }
607 
608 static int destroy_uc(PVRDMADev *dev, union pvrdma_cmd_req *req,
609                       union pvrdma_cmd_resp *rsp)
610 {
611     struct pvrdma_cmd_destroy_uc *cmd = &req->destroy_uc;
612 
613     rdma_rm_dealloc_uc(&dev->rdma_dev_res, cmd->ctx_handle);
614 
615     return 0;
616 }
617 
618 static int create_srq_ring(PCIDevice *pci_dev, PvrdmaRing **ring,
619                            uint64_t pdir_dma, uint32_t max_wr,
620                            uint32_t max_sge, uint32_t nchunks)
621 {
622     uint64_t *dir = NULL, *tbl = NULL;
623     PvrdmaRing *r;
624     int rc = -EINVAL;
625     char ring_name[MAX_RING_NAME_SZ];
626     uint32_t wqe_sz;
627 
628     if (!nchunks || nchunks > PVRDMA_MAX_FAST_REG_PAGES) {
629         rdma_error_report("Got invalid page count for SRQ ring: %d",
630                           nchunks);
631         return rc;
632     }
633 
634     dir = rdma_pci_dma_map(pci_dev, pdir_dma, TARGET_PAGE_SIZE);
635     if (!dir) {
636         rdma_error_report("Failed to map to SRQ page directory");
637         goto out;
638     }
639 
640     tbl = rdma_pci_dma_map(pci_dev, dir[0], TARGET_PAGE_SIZE);
641     if (!tbl) {
642         rdma_error_report("Failed to map to SRQ page table");
643         goto out;
644     }
645 
646     r = g_malloc(sizeof(*r));
647     *ring = r;
648 
649     r->ring_state = (PvrdmaRingState *)
650             rdma_pci_dma_map(pci_dev, tbl[0], TARGET_PAGE_SIZE);
651     if (!r->ring_state) {
652         rdma_error_report("Failed to map tp SRQ ring state");
653         goto out_free_ring_mem;
654     }
655 
656     wqe_sz = pow2ceil(sizeof(struct pvrdma_rq_wqe_hdr) +
657                       sizeof(struct pvrdma_sge) * max_sge - 1);
658     sprintf(ring_name, "srq_ring_%" PRIx64, pdir_dma);
659     rc = pvrdma_ring_init(r, ring_name, pci_dev, &r->ring_state[1], max_wr,
660                           wqe_sz, (dma_addr_t *)&tbl[1], nchunks - 1);
661     if (rc) {
662         goto out_unmap_ring_state;
663     }
664 
665     goto out;
666 
667 out_unmap_ring_state:
668     rdma_pci_dma_unmap(pci_dev, r->ring_state, TARGET_PAGE_SIZE);
669 
670 out_free_ring_mem:
671     g_free(r);
672 
673 out:
674     rdma_pci_dma_unmap(pci_dev, tbl, TARGET_PAGE_SIZE);
675     rdma_pci_dma_unmap(pci_dev, dir, TARGET_PAGE_SIZE);
676 
677     return rc;
678 }
679 
680 static void destroy_srq_ring(PvrdmaRing *ring)
681 {
682     pvrdma_ring_free(ring);
683     rdma_pci_dma_unmap(ring->dev, ring->ring_state, TARGET_PAGE_SIZE);
684     g_free(ring);
685 }
686 
687 static int create_srq(PVRDMADev *dev, union pvrdma_cmd_req *req,
688                       union pvrdma_cmd_resp *rsp)
689 {
690     struct pvrdma_cmd_create_srq *cmd = &req->create_srq;
691     struct pvrdma_cmd_create_srq_resp *resp = &rsp->create_srq_resp;
692     PvrdmaRing *ring = NULL;
693     int rc;
694 
695     memset(resp, 0, sizeof(*resp));
696 
697     rc = create_srq_ring(PCI_DEVICE(dev), &ring, cmd->pdir_dma,
698                          cmd->attrs.max_wr, cmd->attrs.max_sge,
699                          cmd->nchunks);
700     if (rc) {
701         return rc;
702     }
703 
704     rc = rdma_rm_alloc_srq(&dev->rdma_dev_res, cmd->pd_handle,
705                            cmd->attrs.max_wr, cmd->attrs.max_sge,
706                            cmd->attrs.srq_limit, &resp->srqn, ring);
707     if (rc) {
708         destroy_srq_ring(ring);
709         return rc;
710     }
711 
712     return 0;
713 }
714 
715 static int query_srq(PVRDMADev *dev, union pvrdma_cmd_req *req,
716                      union pvrdma_cmd_resp *rsp)
717 {
718     struct pvrdma_cmd_query_srq *cmd = &req->query_srq;
719     struct pvrdma_cmd_query_srq_resp *resp = &rsp->query_srq_resp;
720 
721     memset(resp, 0, sizeof(*resp));
722 
723     return rdma_rm_query_srq(&dev->rdma_dev_res, cmd->srq_handle,
724                              (struct ibv_srq_attr *)&resp->attrs);
725 }
726 
727 static int modify_srq(PVRDMADev *dev, union pvrdma_cmd_req *req,
728                       union pvrdma_cmd_resp *rsp)
729 {
730     struct pvrdma_cmd_modify_srq *cmd = &req->modify_srq;
731 
732     /* Only support SRQ limit */
733     if (!(cmd->attr_mask & IBV_SRQ_LIMIT) ||
734         (cmd->attr_mask & IBV_SRQ_MAX_WR))
735             return -EINVAL;
736 
737     return rdma_rm_modify_srq(&dev->rdma_dev_res, cmd->srq_handle,
738                               (struct ibv_srq_attr *)&cmd->attrs,
739                               cmd->attr_mask);
740 }
741 
742 static int destroy_srq(PVRDMADev *dev, union pvrdma_cmd_req *req,
743                        union pvrdma_cmd_resp *rsp)
744 {
745     struct pvrdma_cmd_destroy_srq *cmd = &req->destroy_srq;
746     RdmaRmSRQ *srq;
747     PvrdmaRing *ring;
748 
749     srq = rdma_rm_get_srq(&dev->rdma_dev_res, cmd->srq_handle);
750     if (!srq) {
751         return -EINVAL;
752     }
753 
754     ring = (PvrdmaRing *)srq->opaque;
755     destroy_srq_ring(ring);
756     rdma_rm_dealloc_srq(&dev->rdma_dev_res, cmd->srq_handle);
757 
758     return 0;
759 }
760 
761 struct cmd_handler {
762     uint32_t cmd;
763     uint32_t ack;
764     int (*exec)(PVRDMADev *dev, union pvrdma_cmd_req *req,
765             union pvrdma_cmd_resp *rsp);
766 };
767 
768 static struct cmd_handler cmd_handlers[] = {
769     {PVRDMA_CMD_QUERY_PORT,   PVRDMA_CMD_QUERY_PORT_RESP,        query_port},
770     {PVRDMA_CMD_QUERY_PKEY,   PVRDMA_CMD_QUERY_PKEY_RESP,        query_pkey},
771     {PVRDMA_CMD_CREATE_PD,    PVRDMA_CMD_CREATE_PD_RESP,         create_pd},
772     {PVRDMA_CMD_DESTROY_PD,   PVRDMA_CMD_DESTROY_PD_RESP_NOOP,   destroy_pd},
773     {PVRDMA_CMD_CREATE_MR,    PVRDMA_CMD_CREATE_MR_RESP,         create_mr},
774     {PVRDMA_CMD_DESTROY_MR,   PVRDMA_CMD_DESTROY_MR_RESP_NOOP,   destroy_mr},
775     {PVRDMA_CMD_CREATE_CQ,    PVRDMA_CMD_CREATE_CQ_RESP,         create_cq},
776     {PVRDMA_CMD_RESIZE_CQ,    PVRDMA_CMD_RESIZE_CQ_RESP,         NULL},
777     {PVRDMA_CMD_DESTROY_CQ,   PVRDMA_CMD_DESTROY_CQ_RESP_NOOP,   destroy_cq},
778     {PVRDMA_CMD_CREATE_QP,    PVRDMA_CMD_CREATE_QP_RESP,         create_qp},
779     {PVRDMA_CMD_MODIFY_QP,    PVRDMA_CMD_MODIFY_QP_RESP,         modify_qp},
780     {PVRDMA_CMD_QUERY_QP,     PVRDMA_CMD_QUERY_QP_RESP,          query_qp},
781     {PVRDMA_CMD_DESTROY_QP,   PVRDMA_CMD_DESTROY_QP_RESP,        destroy_qp},
782     {PVRDMA_CMD_CREATE_UC,    PVRDMA_CMD_CREATE_UC_RESP,         create_uc},
783     {PVRDMA_CMD_DESTROY_UC,   PVRDMA_CMD_DESTROY_UC_RESP_NOOP,   destroy_uc},
784     {PVRDMA_CMD_CREATE_BIND,  PVRDMA_CMD_CREATE_BIND_RESP_NOOP,  create_bind},
785     {PVRDMA_CMD_DESTROY_BIND, PVRDMA_CMD_DESTROY_BIND_RESP_NOOP, destroy_bind},
786     {PVRDMA_CMD_CREATE_SRQ,   PVRDMA_CMD_CREATE_SRQ_RESP,        create_srq},
787     {PVRDMA_CMD_QUERY_SRQ,    PVRDMA_CMD_QUERY_SRQ_RESP,         query_srq},
788     {PVRDMA_CMD_MODIFY_SRQ,   PVRDMA_CMD_MODIFY_SRQ_RESP,        modify_srq},
789     {PVRDMA_CMD_DESTROY_SRQ,  PVRDMA_CMD_DESTROY_SRQ_RESP,       destroy_srq},
790 };
791 
792 int pvrdma_exec_cmd(PVRDMADev *dev)
793 {
794     int err = 0xFFFF;
795     DSRInfo *dsr_info;
796 
797     dsr_info = &dev->dsr_info;
798 
799     if (dsr_info->req->hdr.cmd >= sizeof(cmd_handlers) /
800                       sizeof(struct cmd_handler)) {
801         rdma_error_report("Unsupported command");
802         goto out;
803     }
804 
805     if (!cmd_handlers[dsr_info->req->hdr.cmd].exec) {
806         rdma_error_report("Unsupported command (not implemented yet)");
807         goto out;
808     }
809 
810     err = cmd_handlers[dsr_info->req->hdr.cmd].exec(dev, dsr_info->req,
811                                                     dsr_info->rsp);
812     dsr_info->rsp->hdr.response = dsr_info->req->hdr.response;
813     dsr_info->rsp->hdr.ack = cmd_handlers[dsr_info->req->hdr.cmd].ack;
814     dsr_info->rsp->hdr.err = err < 0 ? -err : 0;
815 
816     trace_pvrdma_exec_cmd(dsr_info->req->hdr.cmd, dsr_info->rsp->hdr.err);
817 
818     dev->stats.commands++;
819 
820 out:
821     set_reg_val(dev, PVRDMA_REG_ERR, err);
822     post_interrupt(dev, INTR_VEC_CMD_RING);
823 
824     return (err == 0) ? 0 : -EINVAL;
825 }
826