WIP FPC-III support
[linux/fpc-iii.git] / drivers / infiniband / hw / qedr / verbs.c
blob0eb6a7a618e0752d5ac5cec0f20130a4e964ac36
1 /* QLogic qedr NIC Driver
2 * Copyright (c) 2015-2016 QLogic Corporation
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and /or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
32 #include <linux/dma-mapping.h>
33 #include <linux/crc32.h>
34 #include <net/ip.h>
35 #include <net/ipv6.h>
36 #include <net/udp.h>
37 #include <linux/iommu.h>
39 #include <rdma/ib_verbs.h>
40 #include <rdma/ib_user_verbs.h>
41 #include <rdma/iw_cm.h>
42 #include <rdma/ib_umem.h>
43 #include <rdma/ib_addr.h>
44 #include <rdma/ib_cache.h>
45 #include <rdma/uverbs_ioctl.h>
47 #include <linux/qed/common_hsi.h>
48 #include "qedr_hsi_rdma.h"
49 #include <linux/qed/qed_if.h>
50 #include "qedr.h"
51 #include "verbs.h"
52 #include <rdma/qedr-abi.h>
53 #include "qedr_roce_cm.h"
54 #include "qedr_iw_cm.h"
56 #define QEDR_SRQ_WQE_ELEM_SIZE sizeof(union rdma_srq_elm)
57 #define RDMA_MAX_SGE_PER_SRQ (4)
58 #define RDMA_MAX_SRQ_WQE_SIZE (RDMA_MAX_SGE_PER_SRQ + 1)
60 #define DB_ADDR_SHIFT(addr) ((addr) << DB_PWM_ADDR_OFFSET_SHIFT)
62 enum {
63 QEDR_USER_MMAP_IO_WC = 0,
64 QEDR_USER_MMAP_PHYS_PAGE,
67 static inline int qedr_ib_copy_to_udata(struct ib_udata *udata, void *src,
68 size_t len)
70 size_t min_len = min_t(size_t, len, udata->outlen);
72 return ib_copy_to_udata(udata, src, min_len);
75 int qedr_query_pkey(struct ib_device *ibdev, u8 port, u16 index, u16 *pkey)
77 if (index >= QEDR_ROCE_PKEY_TABLE_LEN)
78 return -EINVAL;
80 *pkey = QEDR_ROCE_PKEY_DEFAULT;
81 return 0;
84 int qedr_iw_query_gid(struct ib_device *ibdev, u8 port,
85 int index, union ib_gid *sgid)
87 struct qedr_dev *dev = get_qedr_dev(ibdev);
89 memset(sgid->raw, 0, sizeof(sgid->raw));
90 ether_addr_copy(sgid->raw, dev->ndev->dev_addr);
92 DP_DEBUG(dev, QEDR_MSG_INIT, "QUERY sgid[%d]=%llx:%llx\n", index,
93 sgid->global.interface_id, sgid->global.subnet_prefix);
95 return 0;
98 int qedr_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *srq_attr)
100 struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
101 struct qedr_device_attr *qattr = &dev->attr;
102 struct qedr_srq *srq = get_qedr_srq(ibsrq);
104 srq_attr->srq_limit = srq->srq_limit;
105 srq_attr->max_wr = qattr->max_srq_wr;
106 srq_attr->max_sge = qattr->max_sge;
108 return 0;
111 int qedr_query_device(struct ib_device *ibdev,
112 struct ib_device_attr *attr, struct ib_udata *udata)
114 struct qedr_dev *dev = get_qedr_dev(ibdev);
115 struct qedr_device_attr *qattr = &dev->attr;
117 if (!dev->rdma_ctx) {
118 DP_ERR(dev,
119 "qedr_query_device called with invalid params rdma_ctx=%p\n",
120 dev->rdma_ctx);
121 return -EINVAL;
124 memset(attr, 0, sizeof(*attr));
126 attr->fw_ver = qattr->fw_ver;
127 attr->sys_image_guid = qattr->sys_image_guid;
128 attr->max_mr_size = qattr->max_mr_size;
129 attr->page_size_cap = qattr->page_size_caps;
130 attr->vendor_id = qattr->vendor_id;
131 attr->vendor_part_id = qattr->vendor_part_id;
132 attr->hw_ver = qattr->hw_ver;
133 attr->max_qp = qattr->max_qp;
134 attr->max_qp_wr = max_t(u32, qattr->max_sqe, qattr->max_rqe);
135 attr->device_cap_flags = IB_DEVICE_CURR_QP_STATE_MOD |
136 IB_DEVICE_RC_RNR_NAK_GEN |
137 IB_DEVICE_LOCAL_DMA_LKEY | IB_DEVICE_MEM_MGT_EXTENSIONS;
139 if (!rdma_protocol_iwarp(&dev->ibdev, 1))
140 attr->device_cap_flags |= IB_DEVICE_XRC;
141 attr->max_send_sge = qattr->max_sge;
142 attr->max_recv_sge = qattr->max_sge;
143 attr->max_sge_rd = qattr->max_sge;
144 attr->max_cq = qattr->max_cq;
145 attr->max_cqe = qattr->max_cqe;
146 attr->max_mr = qattr->max_mr;
147 attr->max_mw = qattr->max_mw;
148 attr->max_pd = qattr->max_pd;
149 attr->atomic_cap = dev->atomic_cap;
150 attr->max_qp_init_rd_atom =
151 1 << (fls(qattr->max_qp_req_rd_atomic_resc) - 1);
152 attr->max_qp_rd_atom =
153 min(1 << (fls(qattr->max_qp_resp_rd_atomic_resc) - 1),
154 attr->max_qp_init_rd_atom);
156 attr->max_srq = qattr->max_srq;
157 attr->max_srq_sge = qattr->max_srq_sge;
158 attr->max_srq_wr = qattr->max_srq_wr;
160 attr->local_ca_ack_delay = qattr->dev_ack_delay;
161 attr->max_fast_reg_page_list_len = qattr->max_mr / 8;
162 attr->max_pkeys = qattr->max_pkey;
163 attr->max_ah = qattr->max_ah;
165 return 0;
168 static inline void get_link_speed_and_width(int speed, u16 *ib_speed,
169 u8 *ib_width)
171 switch (speed) {
172 case 1000:
173 *ib_speed = IB_SPEED_SDR;
174 *ib_width = IB_WIDTH_1X;
175 break;
176 case 10000:
177 *ib_speed = IB_SPEED_QDR;
178 *ib_width = IB_WIDTH_1X;
179 break;
181 case 20000:
182 *ib_speed = IB_SPEED_DDR;
183 *ib_width = IB_WIDTH_4X;
184 break;
186 case 25000:
187 *ib_speed = IB_SPEED_EDR;
188 *ib_width = IB_WIDTH_1X;
189 break;
191 case 40000:
192 *ib_speed = IB_SPEED_QDR;
193 *ib_width = IB_WIDTH_4X;
194 break;
196 case 50000:
197 *ib_speed = IB_SPEED_HDR;
198 *ib_width = IB_WIDTH_1X;
199 break;
201 case 100000:
202 *ib_speed = IB_SPEED_EDR;
203 *ib_width = IB_WIDTH_4X;
204 break;
206 default:
207 /* Unsupported */
208 *ib_speed = IB_SPEED_SDR;
209 *ib_width = IB_WIDTH_1X;
213 int qedr_query_port(struct ib_device *ibdev, u8 port, struct ib_port_attr *attr)
215 struct qedr_dev *dev;
216 struct qed_rdma_port *rdma_port;
218 dev = get_qedr_dev(ibdev);
220 if (!dev->rdma_ctx) {
221 DP_ERR(dev, "rdma_ctx is NULL\n");
222 return -EINVAL;
225 rdma_port = dev->ops->rdma_query_port(dev->rdma_ctx);
227 /* *attr being zeroed by the caller, avoid zeroing it here */
228 if (rdma_port->port_state == QED_RDMA_PORT_UP) {
229 attr->state = IB_PORT_ACTIVE;
230 attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
231 } else {
232 attr->state = IB_PORT_DOWN;
233 attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
235 attr->max_mtu = IB_MTU_4096;
236 attr->lid = 0;
237 attr->lmc = 0;
238 attr->sm_lid = 0;
239 attr->sm_sl = 0;
240 attr->ip_gids = true;
241 if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
242 attr->active_mtu = iboe_get_mtu(dev->iwarp_max_mtu);
243 attr->gid_tbl_len = 1;
244 } else {
245 attr->active_mtu = iboe_get_mtu(dev->ndev->mtu);
246 attr->gid_tbl_len = QEDR_MAX_SGID;
247 attr->pkey_tbl_len = QEDR_ROCE_PKEY_TABLE_LEN;
249 attr->bad_pkey_cntr = rdma_port->pkey_bad_counter;
250 attr->qkey_viol_cntr = 0;
251 get_link_speed_and_width(rdma_port->link_speed,
252 &attr->active_speed, &attr->active_width);
253 attr->max_msg_sz = rdma_port->max_msg_size;
254 attr->max_vl_num = 4;
256 return 0;
259 int qedr_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata)
261 struct ib_device *ibdev = uctx->device;
262 int rc;
263 struct qedr_ucontext *ctx = get_qedr_ucontext(uctx);
264 struct qedr_alloc_ucontext_resp uresp = {};
265 struct qedr_alloc_ucontext_req ureq = {};
266 struct qedr_dev *dev = get_qedr_dev(ibdev);
267 struct qed_rdma_add_user_out_params oparams;
268 struct qedr_user_mmap_entry *entry;
270 if (!udata)
271 return -EFAULT;
273 if (udata->inlen) {
274 rc = ib_copy_from_udata(&ureq, udata,
275 min(sizeof(ureq), udata->inlen));
276 if (rc) {
277 DP_ERR(dev, "Problem copying data from user space\n");
278 return -EFAULT;
280 ctx->edpm_mode = !!(ureq.context_flags &
281 QEDR_ALLOC_UCTX_EDPM_MODE);
282 ctx->db_rec = !!(ureq.context_flags & QEDR_ALLOC_UCTX_DB_REC);
285 rc = dev->ops->rdma_add_user(dev->rdma_ctx, &oparams);
286 if (rc) {
287 DP_ERR(dev,
288 "failed to allocate a DPI for a new RoCE application, rc=%d. To overcome this consider to increase the number of DPIs, increase the doorbell BAR size or just close unnecessary RoCE applications. In order to increase the number of DPIs consult the qedr readme\n",
289 rc);
290 return rc;
293 ctx->dpi = oparams.dpi;
294 ctx->dpi_addr = oparams.dpi_addr;
295 ctx->dpi_phys_addr = oparams.dpi_phys_addr;
296 ctx->dpi_size = oparams.dpi_size;
297 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
298 if (!entry) {
299 rc = -ENOMEM;
300 goto err;
303 entry->io_address = ctx->dpi_phys_addr;
304 entry->length = ctx->dpi_size;
305 entry->mmap_flag = QEDR_USER_MMAP_IO_WC;
306 entry->dpi = ctx->dpi;
307 entry->dev = dev;
308 rc = rdma_user_mmap_entry_insert(uctx, &entry->rdma_entry,
309 ctx->dpi_size);
310 if (rc) {
311 kfree(entry);
312 goto err;
314 ctx->db_mmap_entry = &entry->rdma_entry;
316 if (!dev->user_dpm_enabled)
317 uresp.dpm_flags = 0;
318 else if (rdma_protocol_iwarp(&dev->ibdev, 1))
319 uresp.dpm_flags = QEDR_DPM_TYPE_IWARP_LEGACY;
320 else
321 uresp.dpm_flags = QEDR_DPM_TYPE_ROCE_ENHANCED |
322 QEDR_DPM_TYPE_ROCE_LEGACY |
323 QEDR_DPM_TYPE_ROCE_EDPM_MODE;
325 if (ureq.context_flags & QEDR_SUPPORT_DPM_SIZES) {
326 uresp.dpm_flags |= QEDR_DPM_SIZES_SET;
327 uresp.ldpm_limit_size = QEDR_LDPM_MAX_SIZE;
328 uresp.edpm_trans_size = QEDR_EDPM_TRANS_SIZE;
329 uresp.edpm_limit_size = QEDR_EDPM_MAX_SIZE;
332 uresp.wids_enabled = 1;
333 uresp.wid_count = oparams.wid_count;
334 uresp.db_pa = rdma_user_mmap_get_offset(ctx->db_mmap_entry);
335 uresp.db_size = ctx->dpi_size;
336 uresp.max_send_wr = dev->attr.max_sqe;
337 uresp.max_recv_wr = dev->attr.max_rqe;
338 uresp.max_srq_wr = dev->attr.max_srq_wr;
339 uresp.sges_per_send_wr = QEDR_MAX_SQE_ELEMENTS_PER_SQE;
340 uresp.sges_per_recv_wr = QEDR_MAX_RQE_ELEMENTS_PER_RQE;
341 uresp.sges_per_srq_wr = dev->attr.max_srq_sge;
342 uresp.max_cqes = QEDR_MAX_CQES;
344 rc = qedr_ib_copy_to_udata(udata, &uresp, sizeof(uresp));
345 if (rc)
346 goto err;
348 ctx->dev = dev;
350 DP_DEBUG(dev, QEDR_MSG_INIT, "Allocating user context %p\n",
351 &ctx->ibucontext);
352 return 0;
354 err:
355 if (!ctx->db_mmap_entry)
356 dev->ops->rdma_remove_user(dev->rdma_ctx, ctx->dpi);
357 else
358 rdma_user_mmap_entry_remove(ctx->db_mmap_entry);
360 return rc;
363 void qedr_dealloc_ucontext(struct ib_ucontext *ibctx)
365 struct qedr_ucontext *uctx = get_qedr_ucontext(ibctx);
367 DP_DEBUG(uctx->dev, QEDR_MSG_INIT, "Deallocating user context %p\n",
368 uctx);
370 rdma_user_mmap_entry_remove(uctx->db_mmap_entry);
373 void qedr_mmap_free(struct rdma_user_mmap_entry *rdma_entry)
375 struct qedr_user_mmap_entry *entry = get_qedr_mmap_entry(rdma_entry);
376 struct qedr_dev *dev = entry->dev;
378 if (entry->mmap_flag == QEDR_USER_MMAP_PHYS_PAGE)
379 free_page((unsigned long)entry->address);
380 else if (entry->mmap_flag == QEDR_USER_MMAP_IO_WC)
381 dev->ops->rdma_remove_user(dev->rdma_ctx, entry->dpi);
383 kfree(entry);
386 int qedr_mmap(struct ib_ucontext *ucontext, struct vm_area_struct *vma)
388 struct ib_device *dev = ucontext->device;
389 size_t length = vma->vm_end - vma->vm_start;
390 struct rdma_user_mmap_entry *rdma_entry;
391 struct qedr_user_mmap_entry *entry;
392 int rc = 0;
393 u64 pfn;
395 ibdev_dbg(dev,
396 "start %#lx, end %#lx, length = %#zx, pgoff = %#lx\n",
397 vma->vm_start, vma->vm_end, length, vma->vm_pgoff);
399 rdma_entry = rdma_user_mmap_entry_get(ucontext, vma);
400 if (!rdma_entry) {
401 ibdev_dbg(dev, "pgoff[%#lx] does not have valid entry\n",
402 vma->vm_pgoff);
403 return -EINVAL;
405 entry = get_qedr_mmap_entry(rdma_entry);
406 ibdev_dbg(dev,
407 "Mapping address[%#llx], length[%#zx], mmap_flag[%d]\n",
408 entry->io_address, length, entry->mmap_flag);
410 switch (entry->mmap_flag) {
411 case QEDR_USER_MMAP_IO_WC:
412 pfn = entry->io_address >> PAGE_SHIFT;
413 rc = rdma_user_mmap_io(ucontext, vma, pfn, length,
414 pgprot_writecombine(vma->vm_page_prot),
415 rdma_entry);
416 break;
417 case QEDR_USER_MMAP_PHYS_PAGE:
418 rc = vm_insert_page(vma, vma->vm_start,
419 virt_to_page(entry->address));
420 break;
421 default:
422 rc = -EINVAL;
425 if (rc)
426 ibdev_dbg(dev,
427 "Couldn't mmap address[%#llx] length[%#zx] mmap_flag[%d] err[%d]\n",
428 entry->io_address, length, entry->mmap_flag, rc);
430 rdma_user_mmap_entry_put(rdma_entry);
431 return rc;
434 int qedr_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
436 struct ib_device *ibdev = ibpd->device;
437 struct qedr_dev *dev = get_qedr_dev(ibdev);
438 struct qedr_pd *pd = get_qedr_pd(ibpd);
439 u16 pd_id;
440 int rc;
442 DP_DEBUG(dev, QEDR_MSG_INIT, "Function called from: %s\n",
443 udata ? "User Lib" : "Kernel");
445 if (!dev->rdma_ctx) {
446 DP_ERR(dev, "invalid RDMA context\n");
447 return -EINVAL;
450 rc = dev->ops->rdma_alloc_pd(dev->rdma_ctx, &pd_id);
451 if (rc)
452 return rc;
454 pd->pd_id = pd_id;
456 if (udata) {
457 struct qedr_alloc_pd_uresp uresp = {
458 .pd_id = pd_id,
460 struct qedr_ucontext *context = rdma_udata_to_drv_context(
461 udata, struct qedr_ucontext, ibucontext);
463 rc = qedr_ib_copy_to_udata(udata, &uresp, sizeof(uresp));
464 if (rc) {
465 DP_ERR(dev, "copy error pd_id=0x%x.\n", pd_id);
466 dev->ops->rdma_dealloc_pd(dev->rdma_ctx, pd_id);
467 return rc;
470 pd->uctx = context;
471 pd->uctx->pd = pd;
474 return 0;
477 int qedr_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
479 struct qedr_dev *dev = get_qedr_dev(ibpd->device);
480 struct qedr_pd *pd = get_qedr_pd(ibpd);
482 DP_DEBUG(dev, QEDR_MSG_INIT, "Deallocating PD %d\n", pd->pd_id);
483 dev->ops->rdma_dealloc_pd(dev->rdma_ctx, pd->pd_id);
484 return 0;
488 int qedr_alloc_xrcd(struct ib_xrcd *ibxrcd, struct ib_udata *udata)
490 struct qedr_dev *dev = get_qedr_dev(ibxrcd->device);
491 struct qedr_xrcd *xrcd = get_qedr_xrcd(ibxrcd);
493 return dev->ops->rdma_alloc_xrcd(dev->rdma_ctx, &xrcd->xrcd_id);
496 int qedr_dealloc_xrcd(struct ib_xrcd *ibxrcd, struct ib_udata *udata)
498 struct qedr_dev *dev = get_qedr_dev(ibxrcd->device);
499 u16 xrcd_id = get_qedr_xrcd(ibxrcd)->xrcd_id;
501 dev->ops->rdma_dealloc_xrcd(dev->rdma_ctx, xrcd_id);
502 return 0;
504 static void qedr_free_pbl(struct qedr_dev *dev,
505 struct qedr_pbl_info *pbl_info, struct qedr_pbl *pbl)
507 struct pci_dev *pdev = dev->pdev;
508 int i;
510 for (i = 0; i < pbl_info->num_pbls; i++) {
511 if (!pbl[i].va)
512 continue;
513 dma_free_coherent(&pdev->dev, pbl_info->pbl_size,
514 pbl[i].va, pbl[i].pa);
517 kfree(pbl);
520 #define MIN_FW_PBL_PAGE_SIZE (4 * 1024)
521 #define MAX_FW_PBL_PAGE_SIZE (64 * 1024)
523 #define NUM_PBES_ON_PAGE(_page_size) (_page_size / sizeof(u64))
524 #define MAX_PBES_ON_PAGE NUM_PBES_ON_PAGE(MAX_FW_PBL_PAGE_SIZE)
525 #define MAX_PBES_TWO_LAYER (MAX_PBES_ON_PAGE * MAX_PBES_ON_PAGE)
527 static struct qedr_pbl *qedr_alloc_pbl_tbl(struct qedr_dev *dev,
528 struct qedr_pbl_info *pbl_info,
529 gfp_t flags)
531 struct pci_dev *pdev = dev->pdev;
532 struct qedr_pbl *pbl_table;
533 dma_addr_t *pbl_main_tbl;
534 dma_addr_t pa;
535 void *va;
536 int i;
538 pbl_table = kcalloc(pbl_info->num_pbls, sizeof(*pbl_table), flags);
539 if (!pbl_table)
540 return ERR_PTR(-ENOMEM);
542 for (i = 0; i < pbl_info->num_pbls; i++) {
543 va = dma_alloc_coherent(&pdev->dev, pbl_info->pbl_size, &pa,
544 flags);
545 if (!va)
546 goto err;
548 pbl_table[i].va = va;
549 pbl_table[i].pa = pa;
552 /* Two-Layer PBLs, if we have more than one pbl we need to initialize
553 * the first one with physical pointers to all of the rest
555 pbl_main_tbl = (dma_addr_t *)pbl_table[0].va;
556 for (i = 0; i < pbl_info->num_pbls - 1; i++)
557 pbl_main_tbl[i] = pbl_table[i + 1].pa;
559 return pbl_table;
561 err:
562 for (i--; i >= 0; i--)
563 dma_free_coherent(&pdev->dev, pbl_info->pbl_size,
564 pbl_table[i].va, pbl_table[i].pa);
566 qedr_free_pbl(dev, pbl_info, pbl_table);
568 return ERR_PTR(-ENOMEM);
571 static int qedr_prepare_pbl_tbl(struct qedr_dev *dev,
572 struct qedr_pbl_info *pbl_info,
573 u32 num_pbes, int two_layer_capable)
575 u32 pbl_capacity;
576 u32 pbl_size;
577 u32 num_pbls;
579 if ((num_pbes > MAX_PBES_ON_PAGE) && two_layer_capable) {
580 if (num_pbes > MAX_PBES_TWO_LAYER) {
581 DP_ERR(dev, "prepare pbl table: too many pages %d\n",
582 num_pbes);
583 return -EINVAL;
586 /* calculate required pbl page size */
587 pbl_size = MIN_FW_PBL_PAGE_SIZE;
588 pbl_capacity = NUM_PBES_ON_PAGE(pbl_size) *
589 NUM_PBES_ON_PAGE(pbl_size);
591 while (pbl_capacity < num_pbes) {
592 pbl_size *= 2;
593 pbl_capacity = pbl_size / sizeof(u64);
594 pbl_capacity = pbl_capacity * pbl_capacity;
597 num_pbls = DIV_ROUND_UP(num_pbes, NUM_PBES_ON_PAGE(pbl_size));
598 num_pbls++; /* One for the layer0 ( points to the pbls) */
599 pbl_info->two_layered = true;
600 } else {
601 /* One layered PBL */
602 num_pbls = 1;
603 pbl_size = max_t(u32, MIN_FW_PBL_PAGE_SIZE,
604 roundup_pow_of_two((num_pbes * sizeof(u64))));
605 pbl_info->two_layered = false;
608 pbl_info->num_pbls = num_pbls;
609 pbl_info->pbl_size = pbl_size;
610 pbl_info->num_pbes = num_pbes;
612 DP_DEBUG(dev, QEDR_MSG_MR,
613 "prepare pbl table: num_pbes=%d, num_pbls=%d, pbl_size=%d\n",
614 pbl_info->num_pbes, pbl_info->num_pbls, pbl_info->pbl_size);
616 return 0;
619 static void qedr_populate_pbls(struct qedr_dev *dev, struct ib_umem *umem,
620 struct qedr_pbl *pbl,
621 struct qedr_pbl_info *pbl_info, u32 pg_shift)
623 int pbe_cnt, total_num_pbes = 0;
624 struct qedr_pbl *pbl_tbl;
625 struct ib_block_iter biter;
626 struct regpair *pbe;
628 if (!pbl_info->num_pbes)
629 return;
631 /* If we have a two layered pbl, the first pbl points to the rest
632 * of the pbls and the first entry lays on the second pbl in the table
634 if (pbl_info->two_layered)
635 pbl_tbl = &pbl[1];
636 else
637 pbl_tbl = pbl;
639 pbe = (struct regpair *)pbl_tbl->va;
640 if (!pbe) {
641 DP_ERR(dev, "cannot populate PBL due to a NULL PBE\n");
642 return;
645 pbe_cnt = 0;
647 rdma_umem_for_each_dma_block (umem, &biter, BIT(pg_shift)) {
648 u64 pg_addr = rdma_block_iter_dma_address(&biter);
650 pbe->lo = cpu_to_le32(pg_addr);
651 pbe->hi = cpu_to_le32(upper_32_bits(pg_addr));
653 pbe_cnt++;
654 total_num_pbes++;
655 pbe++;
657 if (total_num_pbes == pbl_info->num_pbes)
658 return;
660 /* If the given pbl is full storing the pbes, move to next pbl.
662 if (pbe_cnt == (pbl_info->pbl_size / sizeof(u64))) {
663 pbl_tbl++;
664 pbe = (struct regpair *)pbl_tbl->va;
665 pbe_cnt = 0;
670 static int qedr_db_recovery_add(struct qedr_dev *dev,
671 void __iomem *db_addr,
672 void *db_data,
673 enum qed_db_rec_width db_width,
674 enum qed_db_rec_space db_space)
676 if (!db_data) {
677 DP_DEBUG(dev, QEDR_MSG_INIT, "avoiding db rec since old lib\n");
678 return 0;
681 return dev->ops->common->db_recovery_add(dev->cdev, db_addr, db_data,
682 db_width, db_space);
685 static void qedr_db_recovery_del(struct qedr_dev *dev,
686 void __iomem *db_addr,
687 void *db_data)
689 if (!db_data) {
690 DP_DEBUG(dev, QEDR_MSG_INIT, "avoiding db rec since old lib\n");
691 return;
694 /* Ignore return code as there is not much we can do about it. Error
695 * log will be printed inside.
697 dev->ops->common->db_recovery_del(dev->cdev, db_addr, db_data);
700 static int qedr_copy_cq_uresp(struct qedr_dev *dev,
701 struct qedr_cq *cq, struct ib_udata *udata,
702 u32 db_offset)
704 struct qedr_create_cq_uresp uresp;
705 int rc;
707 memset(&uresp, 0, sizeof(uresp));
709 uresp.db_offset = db_offset;
710 uresp.icid = cq->icid;
711 if (cq->q.db_mmap_entry)
712 uresp.db_rec_addr =
713 rdma_user_mmap_get_offset(cq->q.db_mmap_entry);
715 rc = qedr_ib_copy_to_udata(udata, &uresp, sizeof(uresp));
716 if (rc)
717 DP_ERR(dev, "copy error cqid=0x%x.\n", cq->icid);
719 return rc;
722 static void consume_cqe(struct qedr_cq *cq)
724 if (cq->latest_cqe == cq->toggle_cqe)
725 cq->pbl_toggle ^= RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK;
727 cq->latest_cqe = qed_chain_consume(&cq->pbl);
730 static inline int qedr_align_cq_entries(int entries)
732 u64 size, aligned_size;
734 /* We allocate an extra entry that we don't report to the FW. */
735 size = (entries + 1) * QEDR_CQE_SIZE;
736 aligned_size = ALIGN(size, PAGE_SIZE);
738 return aligned_size / QEDR_CQE_SIZE;
741 static int qedr_init_user_db_rec(struct ib_udata *udata,
742 struct qedr_dev *dev, struct qedr_userq *q,
743 bool requires_db_rec)
745 struct qedr_ucontext *uctx =
746 rdma_udata_to_drv_context(udata, struct qedr_ucontext,
747 ibucontext);
748 struct qedr_user_mmap_entry *entry;
749 int rc;
751 /* Aborting for non doorbell userqueue (SRQ) or non-supporting lib */
752 if (requires_db_rec == 0 || !uctx->db_rec)
753 return 0;
755 /* Allocate a page for doorbell recovery, add to mmap */
756 q->db_rec_data = (void *)get_zeroed_page(GFP_USER);
757 if (!q->db_rec_data) {
758 DP_ERR(dev, "get_zeroed_page failed\n");
759 return -ENOMEM;
762 entry = kzalloc(sizeof(*entry), GFP_KERNEL);
763 if (!entry)
764 goto err_free_db_data;
766 entry->address = q->db_rec_data;
767 entry->length = PAGE_SIZE;
768 entry->mmap_flag = QEDR_USER_MMAP_PHYS_PAGE;
769 rc = rdma_user_mmap_entry_insert(&uctx->ibucontext,
770 &entry->rdma_entry,
771 PAGE_SIZE);
772 if (rc)
773 goto err_free_entry;
775 q->db_mmap_entry = &entry->rdma_entry;
777 return 0;
779 err_free_entry:
780 kfree(entry);
782 err_free_db_data:
783 free_page((unsigned long)q->db_rec_data);
784 q->db_rec_data = NULL;
785 return -ENOMEM;
788 static inline int qedr_init_user_queue(struct ib_udata *udata,
789 struct qedr_dev *dev,
790 struct qedr_userq *q, u64 buf_addr,
791 size_t buf_len, bool requires_db_rec,
792 int access,
793 int alloc_and_init)
795 u32 fw_pages;
796 int rc;
798 q->buf_addr = buf_addr;
799 q->buf_len = buf_len;
800 q->umem = ib_umem_get(&dev->ibdev, q->buf_addr, q->buf_len, access);
801 if (IS_ERR(q->umem)) {
802 DP_ERR(dev, "create user queue: failed ib_umem_get, got %ld\n",
803 PTR_ERR(q->umem));
804 return PTR_ERR(q->umem);
807 fw_pages = ib_umem_num_dma_blocks(q->umem, 1 << FW_PAGE_SHIFT);
808 rc = qedr_prepare_pbl_tbl(dev, &q->pbl_info, fw_pages, 0);
809 if (rc)
810 goto err0;
812 if (alloc_and_init) {
813 q->pbl_tbl = qedr_alloc_pbl_tbl(dev, &q->pbl_info, GFP_KERNEL);
814 if (IS_ERR(q->pbl_tbl)) {
815 rc = PTR_ERR(q->pbl_tbl);
816 goto err0;
818 qedr_populate_pbls(dev, q->umem, q->pbl_tbl, &q->pbl_info,
819 FW_PAGE_SHIFT);
820 } else {
821 q->pbl_tbl = kzalloc(sizeof(*q->pbl_tbl), GFP_KERNEL);
822 if (!q->pbl_tbl) {
823 rc = -ENOMEM;
824 goto err0;
828 /* mmap the user address used to store doorbell data for recovery */
829 return qedr_init_user_db_rec(udata, dev, q, requires_db_rec);
831 err0:
832 ib_umem_release(q->umem);
833 q->umem = NULL;
835 return rc;
838 static inline void qedr_init_cq_params(struct qedr_cq *cq,
839 struct qedr_ucontext *ctx,
840 struct qedr_dev *dev, int vector,
841 int chain_entries, int page_cnt,
842 u64 pbl_ptr,
843 struct qed_rdma_create_cq_in_params
844 *params)
846 memset(params, 0, sizeof(*params));
847 params->cq_handle_hi = upper_32_bits((uintptr_t)cq);
848 params->cq_handle_lo = lower_32_bits((uintptr_t)cq);
849 params->cnq_id = vector;
850 params->cq_size = chain_entries - 1;
851 params->dpi = (ctx) ? ctx->dpi : dev->dpi;
852 params->pbl_num_pages = page_cnt;
853 params->pbl_ptr = pbl_ptr;
854 params->pbl_two_level = 0;
857 static void doorbell_cq(struct qedr_cq *cq, u32 cons, u8 flags)
859 cq->db.data.agg_flags = flags;
860 cq->db.data.value = cpu_to_le32(cons);
861 writeq(cq->db.raw, cq->db_addr);
864 int qedr_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
866 struct qedr_cq *cq = get_qedr_cq(ibcq);
867 unsigned long sflags;
868 struct qedr_dev *dev;
870 dev = get_qedr_dev(ibcq->device);
872 if (cq->destroyed) {
873 DP_ERR(dev,
874 "warning: arm was invoked after destroy for cq %p (icid=%d)\n",
875 cq, cq->icid);
876 return -EINVAL;
880 if (cq->cq_type == QEDR_CQ_TYPE_GSI)
881 return 0;
883 spin_lock_irqsave(&cq->cq_lock, sflags);
885 cq->arm_flags = 0;
887 if (flags & IB_CQ_SOLICITED)
888 cq->arm_flags |= DQ_UCM_ROCE_CQ_ARM_SE_CF_CMD;
890 if (flags & IB_CQ_NEXT_COMP)
891 cq->arm_flags |= DQ_UCM_ROCE_CQ_ARM_CF_CMD;
893 doorbell_cq(cq, cq->cq_cons - 1, cq->arm_flags);
895 spin_unlock_irqrestore(&cq->cq_lock, sflags);
897 return 0;
900 int qedr_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
901 struct ib_udata *udata)
903 struct ib_device *ibdev = ibcq->device;
904 struct qedr_ucontext *ctx = rdma_udata_to_drv_context(
905 udata, struct qedr_ucontext, ibucontext);
906 struct qed_rdma_destroy_cq_out_params destroy_oparams;
907 struct qed_rdma_destroy_cq_in_params destroy_iparams;
908 struct qed_chain_init_params chain_params = {
909 .mode = QED_CHAIN_MODE_PBL,
910 .intended_use = QED_CHAIN_USE_TO_CONSUME,
911 .cnt_type = QED_CHAIN_CNT_TYPE_U32,
912 .elem_size = sizeof(union rdma_cqe),
914 struct qedr_dev *dev = get_qedr_dev(ibdev);
915 struct qed_rdma_create_cq_in_params params;
916 struct qedr_create_cq_ureq ureq = {};
917 int vector = attr->comp_vector;
918 int entries = attr->cqe;
919 struct qedr_cq *cq = get_qedr_cq(ibcq);
920 int chain_entries;
921 u32 db_offset;
922 int page_cnt;
923 u64 pbl_ptr;
924 u16 icid;
925 int rc;
927 DP_DEBUG(dev, QEDR_MSG_INIT,
928 "create_cq: called from %s. entries=%d, vector=%d\n",
929 udata ? "User Lib" : "Kernel", entries, vector);
931 if (attr->flags)
932 return -EOPNOTSUPP;
934 if (entries > QEDR_MAX_CQES) {
935 DP_ERR(dev,
936 "create cq: the number of entries %d is too high. Must be equal or below %d.\n",
937 entries, QEDR_MAX_CQES);
938 return -EINVAL;
941 chain_entries = qedr_align_cq_entries(entries);
942 chain_entries = min_t(int, chain_entries, QEDR_MAX_CQES);
943 chain_params.num_elems = chain_entries;
945 /* calc db offset. user will add DPI base, kernel will add db addr */
946 db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_UCM_RDMA_CQ_CONS_32BIT);
948 if (udata) {
949 if (ib_copy_from_udata(&ureq, udata, min(sizeof(ureq),
950 udata->inlen))) {
951 DP_ERR(dev,
952 "create cq: problem copying data from user space\n");
953 goto err0;
956 if (!ureq.len) {
957 DP_ERR(dev,
958 "create cq: cannot create a cq with 0 entries\n");
959 goto err0;
962 cq->cq_type = QEDR_CQ_TYPE_USER;
964 rc = qedr_init_user_queue(udata, dev, &cq->q, ureq.addr,
965 ureq.len, true, IB_ACCESS_LOCAL_WRITE,
967 if (rc)
968 goto err0;
970 pbl_ptr = cq->q.pbl_tbl->pa;
971 page_cnt = cq->q.pbl_info.num_pbes;
973 cq->ibcq.cqe = chain_entries;
974 cq->q.db_addr = ctx->dpi_addr + db_offset;
975 } else {
976 cq->cq_type = QEDR_CQ_TYPE_KERNEL;
978 rc = dev->ops->common->chain_alloc(dev->cdev, &cq->pbl,
979 &chain_params);
980 if (rc)
981 goto err0;
983 page_cnt = qed_chain_get_page_cnt(&cq->pbl);
984 pbl_ptr = qed_chain_get_pbl_phys(&cq->pbl);
985 cq->ibcq.cqe = cq->pbl.capacity;
988 qedr_init_cq_params(cq, ctx, dev, vector, chain_entries, page_cnt,
989 pbl_ptr, &params);
991 rc = dev->ops->rdma_create_cq(dev->rdma_ctx, &params, &icid);
992 if (rc)
993 goto err1;
995 cq->icid = icid;
996 cq->sig = QEDR_CQ_MAGIC_NUMBER;
997 spin_lock_init(&cq->cq_lock);
999 if (udata) {
1000 rc = qedr_copy_cq_uresp(dev, cq, udata, db_offset);
1001 if (rc)
1002 goto err2;
1004 rc = qedr_db_recovery_add(dev, cq->q.db_addr,
1005 &cq->q.db_rec_data->db_data,
1006 DB_REC_WIDTH_64B,
1007 DB_REC_USER);
1008 if (rc)
1009 goto err2;
1011 } else {
1012 /* Generate doorbell address. */
1013 cq->db.data.icid = cq->icid;
1014 cq->db_addr = dev->db_addr + db_offset;
1015 cq->db.data.params = DB_AGG_CMD_MAX <<
1016 RDMA_PWM_VAL32_DATA_AGG_CMD_SHIFT;
1018 /* point to the very last element, passing it we will toggle */
1019 cq->toggle_cqe = qed_chain_get_last_elem(&cq->pbl);
1020 cq->pbl_toggle = RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK;
1021 cq->latest_cqe = NULL;
1022 consume_cqe(cq);
1023 cq->cq_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
1025 rc = qedr_db_recovery_add(dev, cq->db_addr, &cq->db.data,
1026 DB_REC_WIDTH_64B, DB_REC_KERNEL);
1027 if (rc)
1028 goto err2;
1031 DP_DEBUG(dev, QEDR_MSG_CQ,
1032 "create cq: icid=0x%0x, addr=%p, size(entries)=0x%0x\n",
1033 cq->icid, cq, params.cq_size);
1035 return 0;
1037 err2:
1038 destroy_iparams.icid = cq->icid;
1039 dev->ops->rdma_destroy_cq(dev->rdma_ctx, &destroy_iparams,
1040 &destroy_oparams);
1041 err1:
1042 if (udata) {
1043 qedr_free_pbl(dev, &cq->q.pbl_info, cq->q.pbl_tbl);
1044 ib_umem_release(cq->q.umem);
1045 if (cq->q.db_mmap_entry)
1046 rdma_user_mmap_entry_remove(cq->q.db_mmap_entry);
1047 } else {
1048 dev->ops->common->chain_free(dev->cdev, &cq->pbl);
1050 err0:
1051 return -EINVAL;
1054 int qedr_resize_cq(struct ib_cq *ibcq, int new_cnt, struct ib_udata *udata)
1056 struct qedr_dev *dev = get_qedr_dev(ibcq->device);
1057 struct qedr_cq *cq = get_qedr_cq(ibcq);
1059 DP_ERR(dev, "cq %p RESIZE NOT SUPPORTED\n", cq);
1061 return 0;
1064 #define QEDR_DESTROY_CQ_MAX_ITERATIONS (10)
1065 #define QEDR_DESTROY_CQ_ITER_DURATION (10)
1067 int qedr_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
1069 struct qedr_dev *dev = get_qedr_dev(ibcq->device);
1070 struct qed_rdma_destroy_cq_out_params oparams;
1071 struct qed_rdma_destroy_cq_in_params iparams;
1072 struct qedr_cq *cq = get_qedr_cq(ibcq);
1073 int iter;
1075 DP_DEBUG(dev, QEDR_MSG_CQ, "destroy cq %p (icid=%d)\n", cq, cq->icid);
1077 cq->destroyed = 1;
1079 /* GSIs CQs are handled by driver, so they don't exist in the FW */
1080 if (cq->cq_type == QEDR_CQ_TYPE_GSI) {
1081 qedr_db_recovery_del(dev, cq->db_addr, &cq->db.data);
1082 return 0;
1085 iparams.icid = cq->icid;
1086 dev->ops->rdma_destroy_cq(dev->rdma_ctx, &iparams, &oparams);
1087 dev->ops->common->chain_free(dev->cdev, &cq->pbl);
1089 if (udata) {
1090 qedr_free_pbl(dev, &cq->q.pbl_info, cq->q.pbl_tbl);
1091 ib_umem_release(cq->q.umem);
1093 if (cq->q.db_rec_data) {
1094 qedr_db_recovery_del(dev, cq->q.db_addr,
1095 &cq->q.db_rec_data->db_data);
1096 rdma_user_mmap_entry_remove(cq->q.db_mmap_entry);
1098 } else {
1099 qedr_db_recovery_del(dev, cq->db_addr, &cq->db.data);
1102 /* We don't want the IRQ handler to handle a non-existing CQ so we
1103 * wait until all CNQ interrupts, if any, are received. This will always
1104 * happen and will always happen very fast. If not, then a serious error
1105 * has occured. That is why we can use a long delay.
1106 * We spin for a short time so we don’t lose time on context switching
1107 * in case all the completions are handled in that span. Otherwise
1108 * we sleep for a while and check again. Since the CNQ may be
1109 * associated with (only) the current CPU we use msleep to allow the
1110 * current CPU to be freed.
1111 * The CNQ notification is increased in qedr_irq_handler().
1113 iter = QEDR_DESTROY_CQ_MAX_ITERATIONS;
1114 while (oparams.num_cq_notif != READ_ONCE(cq->cnq_notif) && iter) {
1115 udelay(QEDR_DESTROY_CQ_ITER_DURATION);
1116 iter--;
1119 iter = QEDR_DESTROY_CQ_MAX_ITERATIONS;
1120 while (oparams.num_cq_notif != READ_ONCE(cq->cnq_notif) && iter) {
1121 msleep(QEDR_DESTROY_CQ_ITER_DURATION);
1122 iter--;
1125 /* Note that we don't need to have explicit code to wait for the
1126 * completion of the event handler because it is invoked from the EQ.
1127 * Since the destroy CQ ramrod has also been received on the EQ we can
1128 * be certain that there's no event handler in process.
1130 return 0;
1133 static inline int get_gid_info_from_table(struct ib_qp *ibqp,
1134 struct ib_qp_attr *attr,
1135 int attr_mask,
1136 struct qed_rdma_modify_qp_in_params
1137 *qp_params)
1139 const struct ib_gid_attr *gid_attr;
1140 enum rdma_network_type nw_type;
1141 const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
1142 u32 ipv4_addr;
1143 int ret;
1144 int i;
1146 gid_attr = grh->sgid_attr;
1147 ret = rdma_read_gid_l2_fields(gid_attr, &qp_params->vlan_id, NULL);
1148 if (ret)
1149 return ret;
1151 nw_type = rdma_gid_attr_network_type(gid_attr);
1152 switch (nw_type) {
1153 case RDMA_NETWORK_IPV6:
1154 memcpy(&qp_params->sgid.bytes[0], &gid_attr->gid.raw[0],
1155 sizeof(qp_params->sgid));
1156 memcpy(&qp_params->dgid.bytes[0],
1157 &grh->dgid,
1158 sizeof(qp_params->dgid));
1159 qp_params->roce_mode = ROCE_V2_IPV6;
1160 SET_FIELD(qp_params->modify_flags,
1161 QED_ROCE_MODIFY_QP_VALID_ROCE_MODE, 1);
1162 break;
1163 case RDMA_NETWORK_ROCE_V1:
1164 memcpy(&qp_params->sgid.bytes[0], &gid_attr->gid.raw[0],
1165 sizeof(qp_params->sgid));
1166 memcpy(&qp_params->dgid.bytes[0],
1167 &grh->dgid,
1168 sizeof(qp_params->dgid));
1169 qp_params->roce_mode = ROCE_V1;
1170 break;
1171 case RDMA_NETWORK_IPV4:
1172 memset(&qp_params->sgid, 0, sizeof(qp_params->sgid));
1173 memset(&qp_params->dgid, 0, sizeof(qp_params->dgid));
1174 ipv4_addr = qedr_get_ipv4_from_gid(gid_attr->gid.raw);
1175 qp_params->sgid.ipv4_addr = ipv4_addr;
1176 ipv4_addr =
1177 qedr_get_ipv4_from_gid(grh->dgid.raw);
1178 qp_params->dgid.ipv4_addr = ipv4_addr;
1179 SET_FIELD(qp_params->modify_flags,
1180 QED_ROCE_MODIFY_QP_VALID_ROCE_MODE, 1);
1181 qp_params->roce_mode = ROCE_V2_IPV4;
1182 break;
1183 default:
1184 return -EINVAL;
1187 for (i = 0; i < 4; i++) {
1188 qp_params->sgid.dwords[i] = ntohl(qp_params->sgid.dwords[i]);
1189 qp_params->dgid.dwords[i] = ntohl(qp_params->dgid.dwords[i]);
1192 if (qp_params->vlan_id >= VLAN_CFI_MASK)
1193 qp_params->vlan_id = 0;
1195 return 0;
1198 static int qedr_check_qp_attrs(struct ib_pd *ibpd, struct qedr_dev *dev,
1199 struct ib_qp_init_attr *attrs,
1200 struct ib_udata *udata)
1202 struct qedr_device_attr *qattr = &dev->attr;
1204 /* QP0... attrs->qp_type == IB_QPT_GSI */
1205 if (attrs->qp_type != IB_QPT_RC &&
1206 attrs->qp_type != IB_QPT_GSI &&
1207 attrs->qp_type != IB_QPT_XRC_INI &&
1208 attrs->qp_type != IB_QPT_XRC_TGT) {
1209 DP_DEBUG(dev, QEDR_MSG_QP,
1210 "create qp: unsupported qp type=0x%x requested\n",
1211 attrs->qp_type);
1212 return -EOPNOTSUPP;
1215 if (attrs->cap.max_send_wr > qattr->max_sqe) {
1216 DP_ERR(dev,
1217 "create qp: cannot create a SQ with %d elements (max_send_wr=0x%x)\n",
1218 attrs->cap.max_send_wr, qattr->max_sqe);
1219 return -EINVAL;
1222 if (attrs->cap.max_inline_data > qattr->max_inline) {
1223 DP_ERR(dev,
1224 "create qp: unsupported inline data size=0x%x requested (max_inline=0x%x)\n",
1225 attrs->cap.max_inline_data, qattr->max_inline);
1226 return -EINVAL;
1229 if (attrs->cap.max_send_sge > qattr->max_sge) {
1230 DP_ERR(dev,
1231 "create qp: unsupported send_sge=0x%x requested (max_send_sge=0x%x)\n",
1232 attrs->cap.max_send_sge, qattr->max_sge);
1233 return -EINVAL;
1236 if (attrs->cap.max_recv_sge > qattr->max_sge) {
1237 DP_ERR(dev,
1238 "create qp: unsupported recv_sge=0x%x requested (max_recv_sge=0x%x)\n",
1239 attrs->cap.max_recv_sge, qattr->max_sge);
1240 return -EINVAL;
1243 /* verify consumer QPs are not trying to use GSI QP's CQ.
1244 * TGT QP isn't associated with RQ/SQ
1246 if ((attrs->qp_type != IB_QPT_GSI) && (dev->gsi_qp_created) &&
1247 (attrs->qp_type != IB_QPT_XRC_TGT)) {
1248 struct qedr_cq *send_cq = get_qedr_cq(attrs->send_cq);
1249 struct qedr_cq *recv_cq = get_qedr_cq(attrs->recv_cq);
1251 if ((send_cq->cq_type == QEDR_CQ_TYPE_GSI) ||
1252 (recv_cq->cq_type == QEDR_CQ_TYPE_GSI)) {
1253 DP_ERR(dev,
1254 "create qp: consumer QP cannot use GSI CQs.\n");
1255 return -EINVAL;
1259 return 0;
1262 static int qedr_copy_srq_uresp(struct qedr_dev *dev,
1263 struct qedr_srq *srq, struct ib_udata *udata)
1265 struct qedr_create_srq_uresp uresp = {};
1266 int rc;
1268 uresp.srq_id = srq->srq_id;
1270 rc = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
1271 if (rc)
1272 DP_ERR(dev, "create srq: problem copying data to user space\n");
1274 return rc;
1277 static void qedr_copy_rq_uresp(struct qedr_dev *dev,
1278 struct qedr_create_qp_uresp *uresp,
1279 struct qedr_qp *qp)
1281 /* iWARP requires two doorbells per RQ. */
1282 if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
1283 uresp->rq_db_offset =
1284 DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_IWARP_RQ_PROD);
1285 uresp->rq_db2_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_FLAGS);
1286 } else {
1287 uresp->rq_db_offset =
1288 DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
1291 uresp->rq_icid = qp->icid;
1292 if (qp->urq.db_mmap_entry)
1293 uresp->rq_db_rec_addr =
1294 rdma_user_mmap_get_offset(qp->urq.db_mmap_entry);
1297 static void qedr_copy_sq_uresp(struct qedr_dev *dev,
1298 struct qedr_create_qp_uresp *uresp,
1299 struct qedr_qp *qp)
1301 uresp->sq_db_offset = DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
1303 /* iWARP uses the same cid for rq and sq */
1304 if (rdma_protocol_iwarp(&dev->ibdev, 1))
1305 uresp->sq_icid = qp->icid;
1306 else
1307 uresp->sq_icid = qp->icid + 1;
1309 if (qp->usq.db_mmap_entry)
1310 uresp->sq_db_rec_addr =
1311 rdma_user_mmap_get_offset(qp->usq.db_mmap_entry);
1314 static int qedr_copy_qp_uresp(struct qedr_dev *dev,
1315 struct qedr_qp *qp, struct ib_udata *udata,
1316 struct qedr_create_qp_uresp *uresp)
1318 int rc;
1320 memset(uresp, 0, sizeof(*uresp));
1322 if (qedr_qp_has_sq(qp))
1323 qedr_copy_sq_uresp(dev, uresp, qp);
1325 if (qedr_qp_has_rq(qp))
1326 qedr_copy_rq_uresp(dev, uresp, qp);
1328 uresp->atomic_supported = dev->atomic_cap != IB_ATOMIC_NONE;
1329 uresp->qp_id = qp->qp_id;
1331 rc = qedr_ib_copy_to_udata(udata, uresp, sizeof(*uresp));
1332 if (rc)
1333 DP_ERR(dev,
1334 "create qp: failed a copy to user space with qp icid=0x%x.\n",
1335 qp->icid);
1337 return rc;
1340 static void qedr_set_common_qp_params(struct qedr_dev *dev,
1341 struct qedr_qp *qp,
1342 struct qedr_pd *pd,
1343 struct ib_qp_init_attr *attrs)
1345 spin_lock_init(&qp->q_lock);
1346 if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
1347 kref_init(&qp->refcnt);
1348 init_completion(&qp->iwarp_cm_comp);
1351 qp->pd = pd;
1352 qp->qp_type = attrs->qp_type;
1353 qp->max_inline_data = attrs->cap.max_inline_data;
1354 qp->state = QED_ROCE_QP_STATE_RESET;
1355 qp->signaled = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR) ? true : false;
1356 qp->dev = dev;
1357 if (qedr_qp_has_sq(qp)) {
1358 qp->sq.max_sges = attrs->cap.max_send_sge;
1359 qp->sq_cq = get_qedr_cq(attrs->send_cq);
1360 DP_DEBUG(dev, QEDR_MSG_QP,
1361 "SQ params:\tsq_max_sges = %d, sq_cq_id = %d\n",
1362 qp->sq.max_sges, qp->sq_cq->icid);
1365 if (attrs->srq)
1366 qp->srq = get_qedr_srq(attrs->srq);
1368 if (qedr_qp_has_rq(qp)) {
1369 qp->rq_cq = get_qedr_cq(attrs->recv_cq);
1370 qp->rq.max_sges = attrs->cap.max_recv_sge;
1371 DP_DEBUG(dev, QEDR_MSG_QP,
1372 "RQ params:\trq_max_sges = %d, rq_cq_id = %d\n",
1373 qp->rq.max_sges, qp->rq_cq->icid);
1376 DP_DEBUG(dev, QEDR_MSG_QP,
1377 "QP params:\tpd = %d, qp_type = %d, max_inline_data = %d, state = %d, signaled = %d, use_srq=%d\n",
1378 pd->pd_id, qp->qp_type, qp->max_inline_data,
1379 qp->state, qp->signaled, (attrs->srq) ? 1 : 0);
1380 DP_DEBUG(dev, QEDR_MSG_QP,
1381 "SQ params:\tsq_max_sges = %d, sq_cq_id = %d\n",
1382 qp->sq.max_sges, qp->sq_cq->icid);
1385 static int qedr_set_roce_db_info(struct qedr_dev *dev, struct qedr_qp *qp)
1387 int rc = 0;
1389 if (qedr_qp_has_sq(qp)) {
1390 qp->sq.db = dev->db_addr +
1391 DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
1392 qp->sq.db_data.data.icid = qp->icid + 1;
1393 rc = qedr_db_recovery_add(dev, qp->sq.db, &qp->sq.db_data,
1394 DB_REC_WIDTH_32B, DB_REC_KERNEL);
1395 if (rc)
1396 return rc;
1399 if (qedr_qp_has_rq(qp)) {
1400 qp->rq.db = dev->db_addr +
1401 DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_ROCE_RQ_PROD);
1402 qp->rq.db_data.data.icid = qp->icid;
1403 rc = qedr_db_recovery_add(dev, qp->rq.db, &qp->rq.db_data,
1404 DB_REC_WIDTH_32B, DB_REC_KERNEL);
1405 if (rc && qedr_qp_has_sq(qp))
1406 qedr_db_recovery_del(dev, qp->sq.db, &qp->sq.db_data);
1409 return rc;
1412 static int qedr_check_srq_params(struct qedr_dev *dev,
1413 struct ib_srq_init_attr *attrs,
1414 struct ib_udata *udata)
1416 struct qedr_device_attr *qattr = &dev->attr;
1418 if (attrs->attr.max_wr > qattr->max_srq_wr) {
1419 DP_ERR(dev,
1420 "create srq: unsupported srq_wr=0x%x requested (max_srq_wr=0x%x)\n",
1421 attrs->attr.max_wr, qattr->max_srq_wr);
1422 return -EINVAL;
1425 if (attrs->attr.max_sge > qattr->max_sge) {
1426 DP_ERR(dev,
1427 "create srq: unsupported sge=0x%x requested (max_srq_sge=0x%x)\n",
1428 attrs->attr.max_sge, qattr->max_sge);
1431 if (!udata && attrs->srq_type == IB_SRQT_XRC) {
1432 DP_ERR(dev, "XRC SRQs are not supported in kernel-space\n");
1433 return -EINVAL;
1436 return 0;
1439 static void qedr_free_srq_user_params(struct qedr_srq *srq)
1441 qedr_free_pbl(srq->dev, &srq->usrq.pbl_info, srq->usrq.pbl_tbl);
1442 ib_umem_release(srq->usrq.umem);
1443 ib_umem_release(srq->prod_umem);
1446 static void qedr_free_srq_kernel_params(struct qedr_srq *srq)
1448 struct qedr_srq_hwq_info *hw_srq = &srq->hw_srq;
1449 struct qedr_dev *dev = srq->dev;
1451 dev->ops->common->chain_free(dev->cdev, &hw_srq->pbl);
1453 dma_free_coherent(&dev->pdev->dev, sizeof(struct rdma_srq_producers),
1454 hw_srq->virt_prod_pair_addr,
1455 hw_srq->phy_prod_pair_addr);
1458 static int qedr_init_srq_user_params(struct ib_udata *udata,
1459 struct qedr_srq *srq,
1460 struct qedr_create_srq_ureq *ureq,
1461 int access)
1463 struct scatterlist *sg;
1464 int rc;
1466 rc = qedr_init_user_queue(udata, srq->dev, &srq->usrq, ureq->srq_addr,
1467 ureq->srq_len, false, access, 1);
1468 if (rc)
1469 return rc;
1471 srq->prod_umem = ib_umem_get(srq->ibsrq.device, ureq->prod_pair_addr,
1472 sizeof(struct rdma_srq_producers), access);
1473 if (IS_ERR(srq->prod_umem)) {
1474 qedr_free_pbl(srq->dev, &srq->usrq.pbl_info, srq->usrq.pbl_tbl);
1475 ib_umem_release(srq->usrq.umem);
1476 DP_ERR(srq->dev,
1477 "create srq: failed ib_umem_get for producer, got %ld\n",
1478 PTR_ERR(srq->prod_umem));
1479 return PTR_ERR(srq->prod_umem);
1482 sg = srq->prod_umem->sg_head.sgl;
1483 srq->hw_srq.phy_prod_pair_addr = sg_dma_address(sg);
1485 return 0;
1488 static int qedr_alloc_srq_kernel_params(struct qedr_srq *srq,
1489 struct qedr_dev *dev,
1490 struct ib_srq_init_attr *init_attr)
1492 struct qedr_srq_hwq_info *hw_srq = &srq->hw_srq;
1493 struct qed_chain_init_params params = {
1494 .mode = QED_CHAIN_MODE_PBL,
1495 .intended_use = QED_CHAIN_USE_TO_CONSUME_PRODUCE,
1496 .cnt_type = QED_CHAIN_CNT_TYPE_U32,
1497 .elem_size = QEDR_SRQ_WQE_ELEM_SIZE,
1499 dma_addr_t phy_prod_pair_addr;
1500 u32 num_elems;
1501 void *va;
1502 int rc;
1504 va = dma_alloc_coherent(&dev->pdev->dev,
1505 sizeof(struct rdma_srq_producers),
1506 &phy_prod_pair_addr, GFP_KERNEL);
1507 if (!va) {
1508 DP_ERR(dev,
1509 "create srq: failed to allocate dma memory for producer\n");
1510 return -ENOMEM;
1513 hw_srq->phy_prod_pair_addr = phy_prod_pair_addr;
1514 hw_srq->virt_prod_pair_addr = va;
1516 num_elems = init_attr->attr.max_wr * RDMA_MAX_SRQ_WQE_SIZE;
1517 params.num_elems = num_elems;
1519 rc = dev->ops->common->chain_alloc(dev->cdev, &hw_srq->pbl, &params);
1520 if (rc)
1521 goto err0;
1523 hw_srq->num_elems = num_elems;
1525 return 0;
1527 err0:
1528 dma_free_coherent(&dev->pdev->dev, sizeof(struct rdma_srq_producers),
1529 va, phy_prod_pair_addr);
1530 return rc;
1533 int qedr_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init_attr,
1534 struct ib_udata *udata)
1536 struct qed_rdma_destroy_srq_in_params destroy_in_params;
1537 struct qed_rdma_create_srq_in_params in_params = {};
1538 struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
1539 struct qed_rdma_create_srq_out_params out_params;
1540 struct qedr_pd *pd = get_qedr_pd(ibsrq->pd);
1541 struct qedr_create_srq_ureq ureq = {};
1542 u64 pbl_base_addr, phy_prod_pair_addr;
1543 struct qedr_srq_hwq_info *hw_srq;
1544 u32 page_cnt, page_size;
1545 struct qedr_srq *srq = get_qedr_srq(ibsrq);
1546 int rc = 0;
1548 DP_DEBUG(dev, QEDR_MSG_QP,
1549 "create SRQ called from %s (pd %p)\n",
1550 (udata) ? "User lib" : "kernel", pd);
1552 if (init_attr->srq_type != IB_SRQT_BASIC &&
1553 init_attr->srq_type != IB_SRQT_XRC)
1554 return -EOPNOTSUPP;
1556 rc = qedr_check_srq_params(dev, init_attr, udata);
1557 if (rc)
1558 return -EINVAL;
1560 srq->dev = dev;
1561 srq->is_xrc = (init_attr->srq_type == IB_SRQT_XRC);
1562 hw_srq = &srq->hw_srq;
1563 spin_lock_init(&srq->lock);
1565 hw_srq->max_wr = init_attr->attr.max_wr;
1566 hw_srq->max_sges = init_attr->attr.max_sge;
1568 if (udata) {
1569 if (ib_copy_from_udata(&ureq, udata, min(sizeof(ureq),
1570 udata->inlen))) {
1571 DP_ERR(dev,
1572 "create srq: problem copying data from user space\n");
1573 goto err0;
1576 rc = qedr_init_srq_user_params(udata, srq, &ureq, 0);
1577 if (rc)
1578 goto err0;
1580 page_cnt = srq->usrq.pbl_info.num_pbes;
1581 pbl_base_addr = srq->usrq.pbl_tbl->pa;
1582 phy_prod_pair_addr = hw_srq->phy_prod_pair_addr;
1583 page_size = PAGE_SIZE;
1584 } else {
1585 struct qed_chain *pbl;
1587 rc = qedr_alloc_srq_kernel_params(srq, dev, init_attr);
1588 if (rc)
1589 goto err0;
1591 pbl = &hw_srq->pbl;
1592 page_cnt = qed_chain_get_page_cnt(pbl);
1593 pbl_base_addr = qed_chain_get_pbl_phys(pbl);
1594 phy_prod_pair_addr = hw_srq->phy_prod_pair_addr;
1595 page_size = QED_CHAIN_PAGE_SIZE;
1598 in_params.pd_id = pd->pd_id;
1599 in_params.pbl_base_addr = pbl_base_addr;
1600 in_params.prod_pair_addr = phy_prod_pair_addr;
1601 in_params.num_pages = page_cnt;
1602 in_params.page_size = page_size;
1603 if (srq->is_xrc) {
1604 struct qedr_xrcd *xrcd = get_qedr_xrcd(init_attr->ext.xrc.xrcd);
1605 struct qedr_cq *cq = get_qedr_cq(init_attr->ext.cq);
1607 in_params.is_xrc = 1;
1608 in_params.xrcd_id = xrcd->xrcd_id;
1609 in_params.cq_cid = cq->icid;
1612 rc = dev->ops->rdma_create_srq(dev->rdma_ctx, &in_params, &out_params);
1613 if (rc)
1614 goto err1;
1616 srq->srq_id = out_params.srq_id;
1618 if (udata) {
1619 rc = qedr_copy_srq_uresp(dev, srq, udata);
1620 if (rc)
1621 goto err2;
1624 rc = xa_insert_irq(&dev->srqs, srq->srq_id, srq, GFP_KERNEL);
1625 if (rc)
1626 goto err2;
1628 DP_DEBUG(dev, QEDR_MSG_SRQ,
1629 "create srq: created srq with srq_id=0x%0x\n", srq->srq_id);
1630 return 0;
1632 err2:
1633 destroy_in_params.srq_id = srq->srq_id;
1635 dev->ops->rdma_destroy_srq(dev->rdma_ctx, &destroy_in_params);
1636 err1:
1637 if (udata)
1638 qedr_free_srq_user_params(srq);
1639 else
1640 qedr_free_srq_kernel_params(srq);
1641 err0:
1642 return -EFAULT;
1645 int qedr_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
1647 struct qed_rdma_destroy_srq_in_params in_params = {};
1648 struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
1649 struct qedr_srq *srq = get_qedr_srq(ibsrq);
1651 xa_erase_irq(&dev->srqs, srq->srq_id);
1652 in_params.srq_id = srq->srq_id;
1653 in_params.is_xrc = srq->is_xrc;
1654 dev->ops->rdma_destroy_srq(dev->rdma_ctx, &in_params);
1656 if (ibsrq->uobject)
1657 qedr_free_srq_user_params(srq);
1658 else
1659 qedr_free_srq_kernel_params(srq);
1661 DP_DEBUG(dev, QEDR_MSG_SRQ,
1662 "destroy srq: destroyed srq with srq_id=0x%0x\n",
1663 srq->srq_id);
1664 return 0;
1667 int qedr_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
1668 enum ib_srq_attr_mask attr_mask, struct ib_udata *udata)
1670 struct qed_rdma_modify_srq_in_params in_params = {};
1671 struct qedr_dev *dev = get_qedr_dev(ibsrq->device);
1672 struct qedr_srq *srq = get_qedr_srq(ibsrq);
1673 int rc;
1675 if (attr_mask & IB_SRQ_MAX_WR) {
1676 DP_ERR(dev,
1677 "modify srq: invalid attribute mask=0x%x specified for %p\n",
1678 attr_mask, srq);
1679 return -EINVAL;
1682 if (attr_mask & IB_SRQ_LIMIT) {
1683 if (attr->srq_limit >= srq->hw_srq.max_wr) {
1684 DP_ERR(dev,
1685 "modify srq: invalid srq_limit=0x%x (max_srq_limit=0x%x)\n",
1686 attr->srq_limit, srq->hw_srq.max_wr);
1687 return -EINVAL;
1690 in_params.srq_id = srq->srq_id;
1691 in_params.wqe_limit = attr->srq_limit;
1692 rc = dev->ops->rdma_modify_srq(dev->rdma_ctx, &in_params);
1693 if (rc)
1694 return rc;
1697 srq->srq_limit = attr->srq_limit;
1699 DP_DEBUG(dev, QEDR_MSG_SRQ,
1700 "modify srq: modified srq with srq_id=0x%0x\n", srq->srq_id);
1702 return 0;
1705 static enum qed_rdma_qp_type qedr_ib_to_qed_qp_type(enum ib_qp_type ib_qp_type)
1707 switch (ib_qp_type) {
1708 case IB_QPT_RC:
1709 return QED_RDMA_QP_TYPE_RC;
1710 case IB_QPT_XRC_INI:
1711 return QED_RDMA_QP_TYPE_XRC_INI;
1712 case IB_QPT_XRC_TGT:
1713 return QED_RDMA_QP_TYPE_XRC_TGT;
1714 default:
1715 return QED_RDMA_QP_TYPE_INVAL;
1719 static inline void
1720 qedr_init_common_qp_in_params(struct qedr_dev *dev,
1721 struct qedr_pd *pd,
1722 struct qedr_qp *qp,
1723 struct ib_qp_init_attr *attrs,
1724 bool fmr_and_reserved_lkey,
1725 struct qed_rdma_create_qp_in_params *params)
1727 /* QP handle to be written in an async event */
1728 params->qp_handle_async_lo = lower_32_bits((uintptr_t) qp);
1729 params->qp_handle_async_hi = upper_32_bits((uintptr_t) qp);
1731 params->signal_all = (attrs->sq_sig_type == IB_SIGNAL_ALL_WR);
1732 params->fmr_and_reserved_lkey = fmr_and_reserved_lkey;
1733 params->qp_type = qedr_ib_to_qed_qp_type(attrs->qp_type);
1734 params->stats_queue = 0;
1736 if (pd) {
1737 params->pd = pd->pd_id;
1738 params->dpi = pd->uctx ? pd->uctx->dpi : dev->dpi;
1741 if (qedr_qp_has_sq(qp))
1742 params->sq_cq_id = get_qedr_cq(attrs->send_cq)->icid;
1744 if (qedr_qp_has_rq(qp))
1745 params->rq_cq_id = get_qedr_cq(attrs->recv_cq)->icid;
1747 if (qedr_qp_has_srq(qp)) {
1748 params->rq_cq_id = get_qedr_cq(attrs->recv_cq)->icid;
1749 params->srq_id = qp->srq->srq_id;
1750 params->use_srq = true;
1751 } else {
1752 params->srq_id = 0;
1753 params->use_srq = false;
1757 static inline void qedr_qp_user_print(struct qedr_dev *dev, struct qedr_qp *qp)
1759 DP_DEBUG(dev, QEDR_MSG_QP, "create qp: successfully created user QP. "
1760 "qp=%p. "
1761 "sq_addr=0x%llx, "
1762 "sq_len=%zd, "
1763 "rq_addr=0x%llx, "
1764 "rq_len=%zd"
1765 "\n",
1767 qedr_qp_has_sq(qp) ? qp->usq.buf_addr : 0x0,
1768 qedr_qp_has_sq(qp) ? qp->usq.buf_len : 0,
1769 qedr_qp_has_rq(qp) ? qp->urq.buf_addr : 0x0,
1770 qedr_qp_has_sq(qp) ? qp->urq.buf_len : 0);
1773 static inline void
1774 qedr_iwarp_populate_user_qp(struct qedr_dev *dev,
1775 struct qedr_qp *qp,
1776 struct qed_rdma_create_qp_out_params *out_params)
1778 qp->usq.pbl_tbl->va = out_params->sq_pbl_virt;
1779 qp->usq.pbl_tbl->pa = out_params->sq_pbl_phys;
1781 qedr_populate_pbls(dev, qp->usq.umem, qp->usq.pbl_tbl,
1782 &qp->usq.pbl_info, FW_PAGE_SHIFT);
1783 if (!qp->srq) {
1784 qp->urq.pbl_tbl->va = out_params->rq_pbl_virt;
1785 qp->urq.pbl_tbl->pa = out_params->rq_pbl_phys;
1788 qedr_populate_pbls(dev, qp->urq.umem, qp->urq.pbl_tbl,
1789 &qp->urq.pbl_info, FW_PAGE_SHIFT);
1792 static void qedr_cleanup_user(struct qedr_dev *dev,
1793 struct qedr_ucontext *ctx,
1794 struct qedr_qp *qp)
1796 if (qedr_qp_has_sq(qp)) {
1797 ib_umem_release(qp->usq.umem);
1798 qp->usq.umem = NULL;
1801 if (qedr_qp_has_rq(qp)) {
1802 ib_umem_release(qp->urq.umem);
1803 qp->urq.umem = NULL;
1806 if (rdma_protocol_roce(&dev->ibdev, 1)) {
1807 qedr_free_pbl(dev, &qp->usq.pbl_info, qp->usq.pbl_tbl);
1808 qedr_free_pbl(dev, &qp->urq.pbl_info, qp->urq.pbl_tbl);
1809 } else {
1810 kfree(qp->usq.pbl_tbl);
1811 kfree(qp->urq.pbl_tbl);
1814 if (qp->usq.db_rec_data) {
1815 qedr_db_recovery_del(dev, qp->usq.db_addr,
1816 &qp->usq.db_rec_data->db_data);
1817 rdma_user_mmap_entry_remove(qp->usq.db_mmap_entry);
1820 if (qp->urq.db_rec_data) {
1821 qedr_db_recovery_del(dev, qp->urq.db_addr,
1822 &qp->urq.db_rec_data->db_data);
1823 rdma_user_mmap_entry_remove(qp->urq.db_mmap_entry);
1826 if (rdma_protocol_iwarp(&dev->ibdev, 1))
1827 qedr_db_recovery_del(dev, qp->urq.db_rec_db2_addr,
1828 &qp->urq.db_rec_db2_data);
1831 static int qedr_create_user_qp(struct qedr_dev *dev,
1832 struct qedr_qp *qp,
1833 struct ib_pd *ibpd,
1834 struct ib_udata *udata,
1835 struct ib_qp_init_attr *attrs)
1837 struct qed_rdma_create_qp_in_params in_params;
1838 struct qed_rdma_create_qp_out_params out_params;
1839 struct qedr_create_qp_uresp uresp = {};
1840 struct qedr_create_qp_ureq ureq = {};
1841 int alloc_and_init = rdma_protocol_roce(&dev->ibdev, 1);
1842 struct qedr_ucontext *ctx = NULL;
1843 struct qedr_pd *pd = NULL;
1844 int rc = 0;
1846 qp->create_type = QEDR_QP_CREATE_USER;
1848 if (ibpd) {
1849 pd = get_qedr_pd(ibpd);
1850 ctx = pd->uctx;
1853 if (udata) {
1854 rc = ib_copy_from_udata(&ureq, udata, min(sizeof(ureq),
1855 udata->inlen));
1856 if (rc) {
1857 DP_ERR(dev, "Problem copying data from user space\n");
1858 return rc;
1862 if (qedr_qp_has_sq(qp)) {
1863 /* SQ - read access only (0) */
1864 rc = qedr_init_user_queue(udata, dev, &qp->usq, ureq.sq_addr,
1865 ureq.sq_len, true, 0, alloc_and_init);
1866 if (rc)
1867 return rc;
1870 if (qedr_qp_has_rq(qp)) {
1871 /* RQ - read access only (0) */
1872 rc = qedr_init_user_queue(udata, dev, &qp->urq, ureq.rq_addr,
1873 ureq.rq_len, true, 0, alloc_and_init);
1874 if (rc)
1875 return rc;
1878 memset(&in_params, 0, sizeof(in_params));
1879 qedr_init_common_qp_in_params(dev, pd, qp, attrs, false, &in_params);
1880 in_params.qp_handle_lo = ureq.qp_handle_lo;
1881 in_params.qp_handle_hi = ureq.qp_handle_hi;
1883 if (qp->qp_type == IB_QPT_XRC_TGT) {
1884 struct qedr_xrcd *xrcd = get_qedr_xrcd(attrs->xrcd);
1886 in_params.xrcd_id = xrcd->xrcd_id;
1887 in_params.qp_handle_lo = qp->qp_id;
1888 in_params.use_srq = 1;
1891 if (qedr_qp_has_sq(qp)) {
1892 in_params.sq_num_pages = qp->usq.pbl_info.num_pbes;
1893 in_params.sq_pbl_ptr = qp->usq.pbl_tbl->pa;
1896 if (qedr_qp_has_rq(qp)) {
1897 in_params.rq_num_pages = qp->urq.pbl_info.num_pbes;
1898 in_params.rq_pbl_ptr = qp->urq.pbl_tbl->pa;
1901 if (ctx)
1902 SET_FIELD(in_params.flags, QED_ROCE_EDPM_MODE, ctx->edpm_mode);
1904 qp->qed_qp = dev->ops->rdma_create_qp(dev->rdma_ctx,
1905 &in_params, &out_params);
1907 if (!qp->qed_qp) {
1908 rc = -ENOMEM;
1909 goto err1;
1912 if (rdma_protocol_iwarp(&dev->ibdev, 1))
1913 qedr_iwarp_populate_user_qp(dev, qp, &out_params);
1915 qp->qp_id = out_params.qp_id;
1916 qp->icid = out_params.icid;
1918 if (udata) {
1919 rc = qedr_copy_qp_uresp(dev, qp, udata, &uresp);
1920 if (rc)
1921 goto err;
1924 /* db offset was calculated in copy_qp_uresp, now set in the user q */
1925 if (qedr_qp_has_sq(qp)) {
1926 qp->usq.db_addr = ctx->dpi_addr + uresp.sq_db_offset;
1927 rc = qedr_db_recovery_add(dev, qp->usq.db_addr,
1928 &qp->usq.db_rec_data->db_data,
1929 DB_REC_WIDTH_32B,
1930 DB_REC_USER);
1931 if (rc)
1932 goto err;
1935 if (qedr_qp_has_rq(qp)) {
1936 qp->urq.db_addr = ctx->dpi_addr + uresp.rq_db_offset;
1937 rc = qedr_db_recovery_add(dev, qp->urq.db_addr,
1938 &qp->urq.db_rec_data->db_data,
1939 DB_REC_WIDTH_32B,
1940 DB_REC_USER);
1941 if (rc)
1942 goto err;
1945 if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
1946 qp->urq.db_rec_db2_addr = ctx->dpi_addr + uresp.rq_db2_offset;
1948 /* calculate the db_rec_db2 data since it is constant so no
1949 * need to reflect from user
1951 qp->urq.db_rec_db2_data.data.icid = cpu_to_le16(qp->icid);
1952 qp->urq.db_rec_db2_data.data.value =
1953 cpu_to_le16(DQ_TCM_IWARP_POST_RQ_CF_CMD);
1955 rc = qedr_db_recovery_add(dev, qp->urq.db_rec_db2_addr,
1956 &qp->urq.db_rec_db2_data,
1957 DB_REC_WIDTH_32B,
1958 DB_REC_USER);
1959 if (rc)
1960 goto err;
1962 qedr_qp_user_print(dev, qp);
1963 return rc;
1964 err:
1965 rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
1966 if (rc)
1967 DP_ERR(dev, "create qp: fatal fault. rc=%d", rc);
1969 err1:
1970 qedr_cleanup_user(dev, ctx, qp);
1971 return rc;
1974 static int qedr_set_iwarp_db_info(struct qedr_dev *dev, struct qedr_qp *qp)
1976 int rc;
1978 qp->sq.db = dev->db_addr +
1979 DB_ADDR_SHIFT(DQ_PWM_OFFSET_XCM_RDMA_SQ_PROD);
1980 qp->sq.db_data.data.icid = qp->icid;
1982 rc = qedr_db_recovery_add(dev, qp->sq.db,
1983 &qp->sq.db_data,
1984 DB_REC_WIDTH_32B,
1985 DB_REC_KERNEL);
1986 if (rc)
1987 return rc;
1989 qp->rq.db = dev->db_addr +
1990 DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_IWARP_RQ_PROD);
1991 qp->rq.db_data.data.icid = qp->icid;
1992 qp->rq.iwarp_db2 = dev->db_addr +
1993 DB_ADDR_SHIFT(DQ_PWM_OFFSET_TCM_FLAGS);
1994 qp->rq.iwarp_db2_data.data.icid = qp->icid;
1995 qp->rq.iwarp_db2_data.data.value = DQ_TCM_IWARP_POST_RQ_CF_CMD;
1997 rc = qedr_db_recovery_add(dev, qp->rq.db,
1998 &qp->rq.db_data,
1999 DB_REC_WIDTH_32B,
2000 DB_REC_KERNEL);
2001 if (rc)
2002 return rc;
2004 rc = qedr_db_recovery_add(dev, qp->rq.iwarp_db2,
2005 &qp->rq.iwarp_db2_data,
2006 DB_REC_WIDTH_32B,
2007 DB_REC_KERNEL);
2008 return rc;
2011 static int
2012 qedr_roce_create_kernel_qp(struct qedr_dev *dev,
2013 struct qedr_qp *qp,
2014 struct qed_rdma_create_qp_in_params *in_params,
2015 u32 n_sq_elems, u32 n_rq_elems)
2017 struct qed_rdma_create_qp_out_params out_params;
2018 struct qed_chain_init_params params = {
2019 .mode = QED_CHAIN_MODE_PBL,
2020 .cnt_type = QED_CHAIN_CNT_TYPE_U32,
2022 int rc;
2024 params.intended_use = QED_CHAIN_USE_TO_PRODUCE;
2025 params.num_elems = n_sq_elems;
2026 params.elem_size = QEDR_SQE_ELEMENT_SIZE;
2028 rc = dev->ops->common->chain_alloc(dev->cdev, &qp->sq.pbl, &params);
2029 if (rc)
2030 return rc;
2032 in_params->sq_num_pages = qed_chain_get_page_cnt(&qp->sq.pbl);
2033 in_params->sq_pbl_ptr = qed_chain_get_pbl_phys(&qp->sq.pbl);
2035 params.intended_use = QED_CHAIN_USE_TO_CONSUME_PRODUCE;
2036 params.num_elems = n_rq_elems;
2037 params.elem_size = QEDR_RQE_ELEMENT_SIZE;
2039 rc = dev->ops->common->chain_alloc(dev->cdev, &qp->rq.pbl, &params);
2040 if (rc)
2041 return rc;
2043 in_params->rq_num_pages = qed_chain_get_page_cnt(&qp->rq.pbl);
2044 in_params->rq_pbl_ptr = qed_chain_get_pbl_phys(&qp->rq.pbl);
2046 qp->qed_qp = dev->ops->rdma_create_qp(dev->rdma_ctx,
2047 in_params, &out_params);
2049 if (!qp->qed_qp)
2050 return -EINVAL;
2052 qp->qp_id = out_params.qp_id;
2053 qp->icid = out_params.icid;
2055 return qedr_set_roce_db_info(dev, qp);
2058 static int
2059 qedr_iwarp_create_kernel_qp(struct qedr_dev *dev,
2060 struct qedr_qp *qp,
2061 struct qed_rdma_create_qp_in_params *in_params,
2062 u32 n_sq_elems, u32 n_rq_elems)
2064 struct qed_rdma_create_qp_out_params out_params;
2065 struct qed_chain_init_params params = {
2066 .mode = QED_CHAIN_MODE_PBL,
2067 .cnt_type = QED_CHAIN_CNT_TYPE_U32,
2069 int rc;
2071 in_params->sq_num_pages = QED_CHAIN_PAGE_CNT(n_sq_elems,
2072 QEDR_SQE_ELEMENT_SIZE,
2073 QED_CHAIN_PAGE_SIZE,
2074 QED_CHAIN_MODE_PBL);
2075 in_params->rq_num_pages = QED_CHAIN_PAGE_CNT(n_rq_elems,
2076 QEDR_RQE_ELEMENT_SIZE,
2077 QED_CHAIN_PAGE_SIZE,
2078 QED_CHAIN_MODE_PBL);
2080 qp->qed_qp = dev->ops->rdma_create_qp(dev->rdma_ctx,
2081 in_params, &out_params);
2083 if (!qp->qed_qp)
2084 return -EINVAL;
2086 /* Now we allocate the chain */
2088 params.intended_use = QED_CHAIN_USE_TO_PRODUCE;
2089 params.num_elems = n_sq_elems;
2090 params.elem_size = QEDR_SQE_ELEMENT_SIZE;
2091 params.ext_pbl_virt = out_params.sq_pbl_virt;
2092 params.ext_pbl_phys = out_params.sq_pbl_phys;
2094 rc = dev->ops->common->chain_alloc(dev->cdev, &qp->sq.pbl, &params);
2095 if (rc)
2096 goto err;
2098 params.intended_use = QED_CHAIN_USE_TO_CONSUME_PRODUCE;
2099 params.num_elems = n_rq_elems;
2100 params.elem_size = QEDR_RQE_ELEMENT_SIZE;
2101 params.ext_pbl_virt = out_params.rq_pbl_virt;
2102 params.ext_pbl_phys = out_params.rq_pbl_phys;
2104 rc = dev->ops->common->chain_alloc(dev->cdev, &qp->rq.pbl, &params);
2105 if (rc)
2106 goto err;
2108 qp->qp_id = out_params.qp_id;
2109 qp->icid = out_params.icid;
2111 return qedr_set_iwarp_db_info(dev, qp);
2113 err:
2114 dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
2116 return rc;
2119 static void qedr_cleanup_kernel(struct qedr_dev *dev, struct qedr_qp *qp)
2121 dev->ops->common->chain_free(dev->cdev, &qp->sq.pbl);
2122 kfree(qp->wqe_wr_id);
2124 dev->ops->common->chain_free(dev->cdev, &qp->rq.pbl);
2125 kfree(qp->rqe_wr_id);
2127 /* GSI qp is not registered to db mechanism so no need to delete */
2128 if (qp->qp_type == IB_QPT_GSI)
2129 return;
2131 qedr_db_recovery_del(dev, qp->sq.db, &qp->sq.db_data);
2133 if (!qp->srq) {
2134 qedr_db_recovery_del(dev, qp->rq.db, &qp->rq.db_data);
2136 if (rdma_protocol_iwarp(&dev->ibdev, 1))
2137 qedr_db_recovery_del(dev, qp->rq.iwarp_db2,
2138 &qp->rq.iwarp_db2_data);
2142 static int qedr_create_kernel_qp(struct qedr_dev *dev,
2143 struct qedr_qp *qp,
2144 struct ib_pd *ibpd,
2145 struct ib_qp_init_attr *attrs)
2147 struct qed_rdma_create_qp_in_params in_params;
2148 struct qedr_pd *pd = get_qedr_pd(ibpd);
2149 int rc = -EINVAL;
2150 u32 n_rq_elems;
2151 u32 n_sq_elems;
2152 u32 n_sq_entries;
2154 memset(&in_params, 0, sizeof(in_params));
2155 qp->create_type = QEDR_QP_CREATE_KERNEL;
2157 /* A single work request may take up to QEDR_MAX_SQ_WQE_SIZE elements in
2158 * the ring. The ring should allow at least a single WR, even if the
2159 * user requested none, due to allocation issues.
2160 * We should add an extra WR since the prod and cons indices of
2161 * wqe_wr_id are managed in such a way that the WQ is considered full
2162 * when (prod+1)%max_wr==cons. We currently don't do that because we
2163 * double the number of entries due an iSER issue that pushes far more
2164 * WRs than indicated. If we decline its ib_post_send() then we get
2165 * error prints in the dmesg we'd like to avoid.
2167 qp->sq.max_wr = min_t(u32, attrs->cap.max_send_wr * dev->wq_multiplier,
2168 dev->attr.max_sqe);
2170 qp->wqe_wr_id = kcalloc(qp->sq.max_wr, sizeof(*qp->wqe_wr_id),
2171 GFP_KERNEL);
2172 if (!qp->wqe_wr_id) {
2173 DP_ERR(dev, "create qp: failed SQ shadow memory allocation\n");
2174 return -ENOMEM;
2177 /* QP handle to be written in CQE */
2178 in_params.qp_handle_lo = lower_32_bits((uintptr_t) qp);
2179 in_params.qp_handle_hi = upper_32_bits((uintptr_t) qp);
2181 /* A single work request may take up to QEDR_MAX_RQ_WQE_SIZE elements in
2182 * the ring. There ring should allow at least a single WR, even if the
2183 * user requested none, due to allocation issues.
2185 qp->rq.max_wr = (u16) max_t(u32, attrs->cap.max_recv_wr, 1);
2187 /* Allocate driver internal RQ array */
2188 qp->rqe_wr_id = kcalloc(qp->rq.max_wr, sizeof(*qp->rqe_wr_id),
2189 GFP_KERNEL);
2190 if (!qp->rqe_wr_id) {
2191 DP_ERR(dev,
2192 "create qp: failed RQ shadow memory allocation\n");
2193 kfree(qp->wqe_wr_id);
2194 return -ENOMEM;
2197 qedr_init_common_qp_in_params(dev, pd, qp, attrs, true, &in_params);
2199 n_sq_entries = attrs->cap.max_send_wr;
2200 n_sq_entries = min_t(u32, n_sq_entries, dev->attr.max_sqe);
2201 n_sq_entries = max_t(u32, n_sq_entries, 1);
2202 n_sq_elems = n_sq_entries * QEDR_MAX_SQE_ELEMENTS_PER_SQE;
2204 n_rq_elems = qp->rq.max_wr * QEDR_MAX_RQE_ELEMENTS_PER_RQE;
2206 if (rdma_protocol_iwarp(&dev->ibdev, 1))
2207 rc = qedr_iwarp_create_kernel_qp(dev, qp, &in_params,
2208 n_sq_elems, n_rq_elems);
2209 else
2210 rc = qedr_roce_create_kernel_qp(dev, qp, &in_params,
2211 n_sq_elems, n_rq_elems);
2212 if (rc)
2213 qedr_cleanup_kernel(dev, qp);
2215 return rc;
2218 static int qedr_free_qp_resources(struct qedr_dev *dev, struct qedr_qp *qp,
2219 struct ib_udata *udata)
2221 struct qedr_ucontext *ctx =
2222 rdma_udata_to_drv_context(udata, struct qedr_ucontext,
2223 ibucontext);
2224 int rc;
2226 if (qp->qp_type != IB_QPT_GSI) {
2227 rc = dev->ops->rdma_destroy_qp(dev->rdma_ctx, qp->qed_qp);
2228 if (rc)
2229 return rc;
2232 if (qp->create_type == QEDR_QP_CREATE_USER)
2233 qedr_cleanup_user(dev, ctx, qp);
2234 else
2235 qedr_cleanup_kernel(dev, qp);
2237 return 0;
2240 struct ib_qp *qedr_create_qp(struct ib_pd *ibpd,
2241 struct ib_qp_init_attr *attrs,
2242 struct ib_udata *udata)
2244 struct qedr_xrcd *xrcd = NULL;
2245 struct qedr_pd *pd = NULL;
2246 struct qedr_dev *dev;
2247 struct qedr_qp *qp;
2248 struct ib_qp *ibqp;
2249 int rc = 0;
2251 if (attrs->create_flags)
2252 return ERR_PTR(-EOPNOTSUPP);
2254 if (attrs->qp_type == IB_QPT_XRC_TGT) {
2255 xrcd = get_qedr_xrcd(attrs->xrcd);
2256 dev = get_qedr_dev(xrcd->ibxrcd.device);
2257 } else {
2258 pd = get_qedr_pd(ibpd);
2259 dev = get_qedr_dev(ibpd->device);
2262 DP_DEBUG(dev, QEDR_MSG_QP, "create qp: called from %s, pd=%p\n",
2263 udata ? "user library" : "kernel", pd);
2265 rc = qedr_check_qp_attrs(ibpd, dev, attrs, udata);
2266 if (rc)
2267 return ERR_PTR(rc);
2269 DP_DEBUG(dev, QEDR_MSG_QP,
2270 "create qp: called from %s, event_handler=%p, eepd=%p sq_cq=%p, sq_icid=%d, rq_cq=%p, rq_icid=%d\n",
2271 udata ? "user library" : "kernel", attrs->event_handler, pd,
2272 get_qedr_cq(attrs->send_cq),
2273 get_qedr_cq(attrs->send_cq)->icid,
2274 get_qedr_cq(attrs->recv_cq),
2275 attrs->recv_cq ? get_qedr_cq(attrs->recv_cq)->icid : 0);
2277 qp = kzalloc(sizeof(*qp), GFP_KERNEL);
2278 if (!qp) {
2279 DP_ERR(dev, "create qp: failed allocating memory\n");
2280 return ERR_PTR(-ENOMEM);
2283 qedr_set_common_qp_params(dev, qp, pd, attrs);
2285 if (attrs->qp_type == IB_QPT_GSI) {
2286 ibqp = qedr_create_gsi_qp(dev, attrs, qp);
2287 if (IS_ERR(ibqp))
2288 kfree(qp);
2289 return ibqp;
2292 if (udata || xrcd)
2293 rc = qedr_create_user_qp(dev, qp, ibpd, udata, attrs);
2294 else
2295 rc = qedr_create_kernel_qp(dev, qp, ibpd, attrs);
2297 if (rc)
2298 goto out_free_qp;
2300 qp->ibqp.qp_num = qp->qp_id;
2302 if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
2303 rc = xa_insert(&dev->qps, qp->qp_id, qp, GFP_KERNEL);
2304 if (rc)
2305 goto out_free_qp_resources;
2308 return &qp->ibqp;
2310 out_free_qp_resources:
2311 qedr_free_qp_resources(dev, qp, udata);
2312 out_free_qp:
2313 kfree(qp);
2315 return ERR_PTR(-EFAULT);
2318 static enum ib_qp_state qedr_get_ibqp_state(enum qed_roce_qp_state qp_state)
2320 switch (qp_state) {
2321 case QED_ROCE_QP_STATE_RESET:
2322 return IB_QPS_RESET;
2323 case QED_ROCE_QP_STATE_INIT:
2324 return IB_QPS_INIT;
2325 case QED_ROCE_QP_STATE_RTR:
2326 return IB_QPS_RTR;
2327 case QED_ROCE_QP_STATE_RTS:
2328 return IB_QPS_RTS;
2329 case QED_ROCE_QP_STATE_SQD:
2330 return IB_QPS_SQD;
2331 case QED_ROCE_QP_STATE_ERR:
2332 return IB_QPS_ERR;
2333 case QED_ROCE_QP_STATE_SQE:
2334 return IB_QPS_SQE;
2336 return IB_QPS_ERR;
2339 static enum qed_roce_qp_state qedr_get_state_from_ibqp(
2340 enum ib_qp_state qp_state)
2342 switch (qp_state) {
2343 case IB_QPS_RESET:
2344 return QED_ROCE_QP_STATE_RESET;
2345 case IB_QPS_INIT:
2346 return QED_ROCE_QP_STATE_INIT;
2347 case IB_QPS_RTR:
2348 return QED_ROCE_QP_STATE_RTR;
2349 case IB_QPS_RTS:
2350 return QED_ROCE_QP_STATE_RTS;
2351 case IB_QPS_SQD:
2352 return QED_ROCE_QP_STATE_SQD;
2353 case IB_QPS_ERR:
2354 return QED_ROCE_QP_STATE_ERR;
2355 default:
2356 return QED_ROCE_QP_STATE_ERR;
2360 static void qedr_reset_qp_hwq_info(struct qedr_qp_hwq_info *qph)
2362 qed_chain_reset(&qph->pbl);
2363 qph->prod = 0;
2364 qph->cons = 0;
2365 qph->wqe_cons = 0;
2366 qph->db_data.data.value = cpu_to_le16(0);
2369 static int qedr_update_qp_state(struct qedr_dev *dev,
2370 struct qedr_qp *qp,
2371 enum qed_roce_qp_state cur_state,
2372 enum qed_roce_qp_state new_state)
2374 int status = 0;
2376 if (new_state == cur_state)
2377 return 0;
2379 switch (cur_state) {
2380 case QED_ROCE_QP_STATE_RESET:
2381 switch (new_state) {
2382 case QED_ROCE_QP_STATE_INIT:
2383 qp->prev_wqe_size = 0;
2384 qedr_reset_qp_hwq_info(&qp->sq);
2385 qedr_reset_qp_hwq_info(&qp->rq);
2386 break;
2387 default:
2388 status = -EINVAL;
2389 break;
2391 break;
2392 case QED_ROCE_QP_STATE_INIT:
2393 switch (new_state) {
2394 case QED_ROCE_QP_STATE_RTR:
2395 /* Update doorbell (in case post_recv was
2396 * done before move to RTR)
2399 if (rdma_protocol_roce(&dev->ibdev, 1)) {
2400 writel(qp->rq.db_data.raw, qp->rq.db);
2402 break;
2403 case QED_ROCE_QP_STATE_ERR:
2404 break;
2405 default:
2406 /* Invalid state change. */
2407 status = -EINVAL;
2408 break;
2410 break;
2411 case QED_ROCE_QP_STATE_RTR:
2412 /* RTR->XXX */
2413 switch (new_state) {
2414 case QED_ROCE_QP_STATE_RTS:
2415 break;
2416 case QED_ROCE_QP_STATE_ERR:
2417 break;
2418 default:
2419 /* Invalid state change. */
2420 status = -EINVAL;
2421 break;
2423 break;
2424 case QED_ROCE_QP_STATE_RTS:
2425 /* RTS->XXX */
2426 switch (new_state) {
2427 case QED_ROCE_QP_STATE_SQD:
2428 break;
2429 case QED_ROCE_QP_STATE_ERR:
2430 break;
2431 default:
2432 /* Invalid state change. */
2433 status = -EINVAL;
2434 break;
2436 break;
2437 case QED_ROCE_QP_STATE_SQD:
2438 /* SQD->XXX */
2439 switch (new_state) {
2440 case QED_ROCE_QP_STATE_RTS:
2441 case QED_ROCE_QP_STATE_ERR:
2442 break;
2443 default:
2444 /* Invalid state change. */
2445 status = -EINVAL;
2446 break;
2448 break;
2449 case QED_ROCE_QP_STATE_ERR:
2450 /* ERR->XXX */
2451 switch (new_state) {
2452 case QED_ROCE_QP_STATE_RESET:
2453 if ((qp->rq.prod != qp->rq.cons) ||
2454 (qp->sq.prod != qp->sq.cons)) {
2455 DP_NOTICE(dev,
2456 "Error->Reset with rq/sq not empty rq.prod=%x rq.cons=%x sq.prod=%x sq.cons=%x\n",
2457 qp->rq.prod, qp->rq.cons, qp->sq.prod,
2458 qp->sq.cons);
2459 status = -EINVAL;
2461 break;
2462 default:
2463 status = -EINVAL;
2464 break;
2466 break;
2467 default:
2468 status = -EINVAL;
2469 break;
2472 return status;
2475 int qedr_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
2476 int attr_mask, struct ib_udata *udata)
2478 struct qedr_qp *qp = get_qedr_qp(ibqp);
2479 struct qed_rdma_modify_qp_in_params qp_params = { 0 };
2480 struct qedr_dev *dev = get_qedr_dev(&qp->dev->ibdev);
2481 const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
2482 enum ib_qp_state old_qp_state, new_qp_state;
2483 enum qed_roce_qp_state cur_state;
2484 int rc = 0;
2486 DP_DEBUG(dev, QEDR_MSG_QP,
2487 "modify qp: qp %p attr_mask=0x%x, state=%d", qp, attr_mask,
2488 attr->qp_state);
2490 if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
2491 return -EOPNOTSUPP;
2493 old_qp_state = qedr_get_ibqp_state(qp->state);
2494 if (attr_mask & IB_QP_STATE)
2495 new_qp_state = attr->qp_state;
2496 else
2497 new_qp_state = old_qp_state;
2499 if (rdma_protocol_roce(&dev->ibdev, 1)) {
2500 if (!ib_modify_qp_is_ok(old_qp_state, new_qp_state,
2501 ibqp->qp_type, attr_mask)) {
2502 DP_ERR(dev,
2503 "modify qp: invalid attribute mask=0x%x specified for\n"
2504 "qpn=0x%x of type=0x%x old_qp_state=0x%x, new_qp_state=0x%x\n",
2505 attr_mask, qp->qp_id, ibqp->qp_type,
2506 old_qp_state, new_qp_state);
2507 rc = -EINVAL;
2508 goto err;
2512 /* Translate the masks... */
2513 if (attr_mask & IB_QP_STATE) {
2514 SET_FIELD(qp_params.modify_flags,
2515 QED_RDMA_MODIFY_QP_VALID_NEW_STATE, 1);
2516 qp_params.new_state = qedr_get_state_from_ibqp(attr->qp_state);
2519 if (attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY)
2520 qp_params.sqd_async = true;
2522 if (attr_mask & IB_QP_PKEY_INDEX) {
2523 SET_FIELD(qp_params.modify_flags,
2524 QED_ROCE_MODIFY_QP_VALID_PKEY, 1);
2525 if (attr->pkey_index >= QEDR_ROCE_PKEY_TABLE_LEN) {
2526 rc = -EINVAL;
2527 goto err;
2530 qp_params.pkey = QEDR_ROCE_PKEY_DEFAULT;
2533 if (attr_mask & IB_QP_QKEY)
2534 qp->qkey = attr->qkey;
2536 if (attr_mask & IB_QP_ACCESS_FLAGS) {
2537 SET_FIELD(qp_params.modify_flags,
2538 QED_RDMA_MODIFY_QP_VALID_RDMA_OPS_EN, 1);
2539 qp_params.incoming_rdma_read_en = attr->qp_access_flags &
2540 IB_ACCESS_REMOTE_READ;
2541 qp_params.incoming_rdma_write_en = attr->qp_access_flags &
2542 IB_ACCESS_REMOTE_WRITE;
2543 qp_params.incoming_atomic_en = attr->qp_access_flags &
2544 IB_ACCESS_REMOTE_ATOMIC;
2547 if (attr_mask & (IB_QP_AV | IB_QP_PATH_MTU)) {
2548 if (rdma_protocol_iwarp(&dev->ibdev, 1))
2549 return -EINVAL;
2551 if (attr_mask & IB_QP_PATH_MTU) {
2552 if (attr->path_mtu < IB_MTU_256 ||
2553 attr->path_mtu > IB_MTU_4096) {
2554 pr_err("error: Only MTU sizes of 256, 512, 1024, 2048 and 4096 are supported by RoCE\n");
2555 rc = -EINVAL;
2556 goto err;
2558 qp->mtu = min(ib_mtu_enum_to_int(attr->path_mtu),
2559 ib_mtu_enum_to_int(iboe_get_mtu
2560 (dev->ndev->mtu)));
2563 if (!qp->mtu) {
2564 qp->mtu =
2565 ib_mtu_enum_to_int(iboe_get_mtu(dev->ndev->mtu));
2566 pr_err("Fixing zeroed MTU to qp->mtu = %d\n", qp->mtu);
2569 SET_FIELD(qp_params.modify_flags,
2570 QED_ROCE_MODIFY_QP_VALID_ADDRESS_VECTOR, 1);
2572 qp_params.traffic_class_tos = grh->traffic_class;
2573 qp_params.flow_label = grh->flow_label;
2574 qp_params.hop_limit_ttl = grh->hop_limit;
2576 qp->sgid_idx = grh->sgid_index;
2578 rc = get_gid_info_from_table(ibqp, attr, attr_mask, &qp_params);
2579 if (rc) {
2580 DP_ERR(dev,
2581 "modify qp: problems with GID index %d (rc=%d)\n",
2582 grh->sgid_index, rc);
2583 return rc;
2586 rc = qedr_get_dmac(dev, &attr->ah_attr,
2587 qp_params.remote_mac_addr);
2588 if (rc)
2589 return rc;
2591 qp_params.use_local_mac = true;
2592 ether_addr_copy(qp_params.local_mac_addr, dev->ndev->dev_addr);
2594 DP_DEBUG(dev, QEDR_MSG_QP, "dgid=%x:%x:%x:%x\n",
2595 qp_params.dgid.dwords[0], qp_params.dgid.dwords[1],
2596 qp_params.dgid.dwords[2], qp_params.dgid.dwords[3]);
2597 DP_DEBUG(dev, QEDR_MSG_QP, "sgid=%x:%x:%x:%x\n",
2598 qp_params.sgid.dwords[0], qp_params.sgid.dwords[1],
2599 qp_params.sgid.dwords[2], qp_params.sgid.dwords[3]);
2600 DP_DEBUG(dev, QEDR_MSG_QP, "remote_mac=[%pM]\n",
2601 qp_params.remote_mac_addr);
2603 qp_params.mtu = qp->mtu;
2604 qp_params.lb_indication = false;
2607 if (!qp_params.mtu) {
2608 /* Stay with current MTU */
2609 if (qp->mtu)
2610 qp_params.mtu = qp->mtu;
2611 else
2612 qp_params.mtu =
2613 ib_mtu_enum_to_int(iboe_get_mtu(dev->ndev->mtu));
2616 if (attr_mask & IB_QP_TIMEOUT) {
2617 SET_FIELD(qp_params.modify_flags,
2618 QED_ROCE_MODIFY_QP_VALID_ACK_TIMEOUT, 1);
2620 /* The received timeout value is an exponent used like this:
2621 * "12.7.34 LOCAL ACK TIMEOUT
2622 * Value representing the transport (ACK) timeout for use by
2623 * the remote, expressed as: 4.096 * 2^timeout [usec]"
2624 * The FW expects timeout in msec so we need to divide the usec
2625 * result by 1000. We'll approximate 1000~2^10, and 4.096 ~ 2^2,
2626 * so we get: 2^2 * 2^timeout / 2^10 = 2^(timeout - 8).
2627 * The value of zero means infinite so we use a 'max_t' to make
2628 * sure that sub 1 msec values will be configured as 1 msec.
2630 if (attr->timeout)
2631 qp_params.ack_timeout =
2632 1 << max_t(int, attr->timeout - 8, 0);
2633 else
2634 qp_params.ack_timeout = 0;
2637 if (attr_mask & IB_QP_RETRY_CNT) {
2638 SET_FIELD(qp_params.modify_flags,
2639 QED_ROCE_MODIFY_QP_VALID_RETRY_CNT, 1);
2640 qp_params.retry_cnt = attr->retry_cnt;
2643 if (attr_mask & IB_QP_RNR_RETRY) {
2644 SET_FIELD(qp_params.modify_flags,
2645 QED_ROCE_MODIFY_QP_VALID_RNR_RETRY_CNT, 1);
2646 qp_params.rnr_retry_cnt = attr->rnr_retry;
2649 if (attr_mask & IB_QP_RQ_PSN) {
2650 SET_FIELD(qp_params.modify_flags,
2651 QED_ROCE_MODIFY_QP_VALID_RQ_PSN, 1);
2652 qp_params.rq_psn = attr->rq_psn;
2653 qp->rq_psn = attr->rq_psn;
2656 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
2657 if (attr->max_rd_atomic > dev->attr.max_qp_req_rd_atomic_resc) {
2658 rc = -EINVAL;
2659 DP_ERR(dev,
2660 "unsupported max_rd_atomic=%d, supported=%d\n",
2661 attr->max_rd_atomic,
2662 dev->attr.max_qp_req_rd_atomic_resc);
2663 goto err;
2666 SET_FIELD(qp_params.modify_flags,
2667 QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_REQ, 1);
2668 qp_params.max_rd_atomic_req = attr->max_rd_atomic;
2671 if (attr_mask & IB_QP_MIN_RNR_TIMER) {
2672 SET_FIELD(qp_params.modify_flags,
2673 QED_ROCE_MODIFY_QP_VALID_MIN_RNR_NAK_TIMER, 1);
2674 qp_params.min_rnr_nak_timer = attr->min_rnr_timer;
2677 if (attr_mask & IB_QP_SQ_PSN) {
2678 SET_FIELD(qp_params.modify_flags,
2679 QED_ROCE_MODIFY_QP_VALID_SQ_PSN, 1);
2680 qp_params.sq_psn = attr->sq_psn;
2681 qp->sq_psn = attr->sq_psn;
2684 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
2685 if (attr->max_dest_rd_atomic >
2686 dev->attr.max_qp_resp_rd_atomic_resc) {
2687 DP_ERR(dev,
2688 "unsupported max_dest_rd_atomic=%d, supported=%d\n",
2689 attr->max_dest_rd_atomic,
2690 dev->attr.max_qp_resp_rd_atomic_resc);
2692 rc = -EINVAL;
2693 goto err;
2696 SET_FIELD(qp_params.modify_flags,
2697 QED_RDMA_MODIFY_QP_VALID_MAX_RD_ATOMIC_RESP, 1);
2698 qp_params.max_rd_atomic_resp = attr->max_dest_rd_atomic;
2701 if (attr_mask & IB_QP_DEST_QPN) {
2702 SET_FIELD(qp_params.modify_flags,
2703 QED_ROCE_MODIFY_QP_VALID_DEST_QP, 1);
2705 qp_params.dest_qp = attr->dest_qp_num;
2706 qp->dest_qp_num = attr->dest_qp_num;
2709 cur_state = qp->state;
2711 /* Update the QP state before the actual ramrod to prevent a race with
2712 * fast path. Modifying the QP state to error will cause the device to
2713 * flush the CQEs and while polling the flushed CQEs will considered as
2714 * a potential issue if the QP isn't in error state.
2716 if ((attr_mask & IB_QP_STATE) && qp->qp_type != IB_QPT_GSI &&
2717 !udata && qp_params.new_state == QED_ROCE_QP_STATE_ERR)
2718 qp->state = QED_ROCE_QP_STATE_ERR;
2720 if (qp->qp_type != IB_QPT_GSI)
2721 rc = dev->ops->rdma_modify_qp(dev->rdma_ctx,
2722 qp->qed_qp, &qp_params);
2724 if (attr_mask & IB_QP_STATE) {
2725 if ((qp->qp_type != IB_QPT_GSI) && (!udata))
2726 rc = qedr_update_qp_state(dev, qp, cur_state,
2727 qp_params.new_state);
2728 qp->state = qp_params.new_state;
2731 err:
2732 return rc;
2735 static int qedr_to_ib_qp_acc_flags(struct qed_rdma_query_qp_out_params *params)
2737 int ib_qp_acc_flags = 0;
2739 if (params->incoming_rdma_write_en)
2740 ib_qp_acc_flags |= IB_ACCESS_REMOTE_WRITE;
2741 if (params->incoming_rdma_read_en)
2742 ib_qp_acc_flags |= IB_ACCESS_REMOTE_READ;
2743 if (params->incoming_atomic_en)
2744 ib_qp_acc_flags |= IB_ACCESS_REMOTE_ATOMIC;
2745 ib_qp_acc_flags |= IB_ACCESS_LOCAL_WRITE;
2746 return ib_qp_acc_flags;
2749 int qedr_query_qp(struct ib_qp *ibqp,
2750 struct ib_qp_attr *qp_attr,
2751 int attr_mask, struct ib_qp_init_attr *qp_init_attr)
2753 struct qed_rdma_query_qp_out_params params;
2754 struct qedr_qp *qp = get_qedr_qp(ibqp);
2755 struct qedr_dev *dev = qp->dev;
2756 int rc = 0;
2758 memset(&params, 0, sizeof(params));
2760 rc = dev->ops->rdma_query_qp(dev->rdma_ctx, qp->qed_qp, &params);
2761 if (rc)
2762 goto err;
2764 memset(qp_attr, 0, sizeof(*qp_attr));
2765 memset(qp_init_attr, 0, sizeof(*qp_init_attr));
2767 qp_attr->qp_state = qedr_get_ibqp_state(params.state);
2768 qp_attr->cur_qp_state = qedr_get_ibqp_state(params.state);
2769 qp_attr->path_mtu = ib_mtu_int_to_enum(params.mtu);
2770 qp_attr->path_mig_state = IB_MIG_MIGRATED;
2771 qp_attr->rq_psn = params.rq_psn;
2772 qp_attr->sq_psn = params.sq_psn;
2773 qp_attr->dest_qp_num = params.dest_qp;
2775 qp_attr->qp_access_flags = qedr_to_ib_qp_acc_flags(&params);
2777 qp_attr->cap.max_send_wr = qp->sq.max_wr;
2778 qp_attr->cap.max_recv_wr = qp->rq.max_wr;
2779 qp_attr->cap.max_send_sge = qp->sq.max_sges;
2780 qp_attr->cap.max_recv_sge = qp->rq.max_sges;
2781 qp_attr->cap.max_inline_data = dev->attr.max_inline;
2782 qp_init_attr->cap = qp_attr->cap;
2784 qp_attr->ah_attr.type = RDMA_AH_ATTR_TYPE_ROCE;
2785 rdma_ah_set_grh(&qp_attr->ah_attr, NULL,
2786 params.flow_label, qp->sgid_idx,
2787 params.hop_limit_ttl, params.traffic_class_tos);
2788 rdma_ah_set_dgid_raw(&qp_attr->ah_attr, &params.dgid.bytes[0]);
2789 rdma_ah_set_port_num(&qp_attr->ah_attr, 1);
2790 rdma_ah_set_sl(&qp_attr->ah_attr, 0);
2791 qp_attr->timeout = params.timeout;
2792 qp_attr->rnr_retry = params.rnr_retry;
2793 qp_attr->retry_cnt = params.retry_cnt;
2794 qp_attr->min_rnr_timer = params.min_rnr_nak_timer;
2795 qp_attr->pkey_index = params.pkey_index;
2796 qp_attr->port_num = 1;
2797 rdma_ah_set_path_bits(&qp_attr->ah_attr, 0);
2798 rdma_ah_set_static_rate(&qp_attr->ah_attr, 0);
2799 qp_attr->alt_pkey_index = 0;
2800 qp_attr->alt_port_num = 0;
2801 qp_attr->alt_timeout = 0;
2802 memset(&qp_attr->alt_ah_attr, 0, sizeof(qp_attr->alt_ah_attr));
2804 qp_attr->sq_draining = (params.state == QED_ROCE_QP_STATE_SQD) ? 1 : 0;
2805 qp_attr->max_dest_rd_atomic = params.max_dest_rd_atomic;
2806 qp_attr->max_rd_atomic = params.max_rd_atomic;
2807 qp_attr->en_sqd_async_notify = (params.sqd_async) ? 1 : 0;
2809 DP_DEBUG(dev, QEDR_MSG_QP, "QEDR_QUERY_QP: max_inline_data=%d\n",
2810 qp_attr->cap.max_inline_data);
2812 err:
2813 return rc;
2816 int qedr_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
2818 struct qedr_qp *qp = get_qedr_qp(ibqp);
2819 struct qedr_dev *dev = qp->dev;
2820 struct ib_qp_attr attr;
2821 int attr_mask = 0;
2823 DP_DEBUG(dev, QEDR_MSG_QP, "destroy qp: destroying %p, qp type=%d\n",
2824 qp, qp->qp_type);
2826 if (rdma_protocol_roce(&dev->ibdev, 1)) {
2827 if ((qp->state != QED_ROCE_QP_STATE_RESET) &&
2828 (qp->state != QED_ROCE_QP_STATE_ERR) &&
2829 (qp->state != QED_ROCE_QP_STATE_INIT)) {
2831 attr.qp_state = IB_QPS_ERR;
2832 attr_mask |= IB_QP_STATE;
2834 /* Change the QP state to ERROR */
2835 qedr_modify_qp(ibqp, &attr, attr_mask, NULL);
2837 } else {
2838 /* If connection establishment started the WAIT_FOR_CONNECT
2839 * bit will be on and we need to Wait for the establishment
2840 * to complete before destroying the qp.
2842 if (test_and_set_bit(QEDR_IWARP_CM_WAIT_FOR_CONNECT,
2843 &qp->iwarp_cm_flags))
2844 wait_for_completion(&qp->iwarp_cm_comp);
2846 /* If graceful disconnect started, the WAIT_FOR_DISCONNECT
2847 * bit will be on, and we need to wait for the disconnect to
2848 * complete before continuing. We can use the same completion,
2849 * iwarp_cm_comp, since this is the only place that waits for
2850 * this completion and it is sequential. In addition,
2851 * disconnect can't occur before the connection is fully
2852 * established, therefore if WAIT_FOR_DISCONNECT is on it
2853 * means WAIT_FOR_CONNECT is also on and the completion for
2854 * CONNECT already occurred.
2856 if (test_and_set_bit(QEDR_IWARP_CM_WAIT_FOR_DISCONNECT,
2857 &qp->iwarp_cm_flags))
2858 wait_for_completion(&qp->iwarp_cm_comp);
2861 if (qp->qp_type == IB_QPT_GSI)
2862 qedr_destroy_gsi_qp(dev);
2864 /* We need to remove the entry from the xarray before we release the
2865 * qp_id to avoid a race of the qp_id being reallocated and failing
2866 * on xa_insert
2868 if (rdma_protocol_iwarp(&dev->ibdev, 1))
2869 xa_erase(&dev->qps, qp->qp_id);
2871 qedr_free_qp_resources(dev, qp, udata);
2873 if (rdma_protocol_iwarp(&dev->ibdev, 1))
2874 qedr_iw_qp_rem_ref(&qp->ibqp);
2875 else
2876 kfree(qp);
2878 return 0;
2881 int qedr_create_ah(struct ib_ah *ibah, struct rdma_ah_init_attr *init_attr,
2882 struct ib_udata *udata)
2884 struct qedr_ah *ah = get_qedr_ah(ibah);
2886 rdma_copy_ah_attr(&ah->attr, init_attr->ah_attr);
2888 return 0;
2891 int qedr_destroy_ah(struct ib_ah *ibah, u32 flags)
2893 struct qedr_ah *ah = get_qedr_ah(ibah);
2895 rdma_destroy_ah_attr(&ah->attr);
2896 return 0;
2899 static void free_mr_info(struct qedr_dev *dev, struct mr_info *info)
2901 struct qedr_pbl *pbl, *tmp;
2903 if (info->pbl_table)
2904 list_add_tail(&info->pbl_table->list_entry,
2905 &info->free_pbl_list);
2907 if (!list_empty(&info->inuse_pbl_list))
2908 list_splice(&info->inuse_pbl_list, &info->free_pbl_list);
2910 list_for_each_entry_safe(pbl, tmp, &info->free_pbl_list, list_entry) {
2911 list_del(&pbl->list_entry);
2912 qedr_free_pbl(dev, &info->pbl_info, pbl);
2916 static int init_mr_info(struct qedr_dev *dev, struct mr_info *info,
2917 size_t page_list_len, bool two_layered)
2919 struct qedr_pbl *tmp;
2920 int rc;
2922 INIT_LIST_HEAD(&info->free_pbl_list);
2923 INIT_LIST_HEAD(&info->inuse_pbl_list);
2925 rc = qedr_prepare_pbl_tbl(dev, &info->pbl_info,
2926 page_list_len, two_layered);
2927 if (rc)
2928 goto done;
2930 info->pbl_table = qedr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL);
2931 if (IS_ERR(info->pbl_table)) {
2932 rc = PTR_ERR(info->pbl_table);
2933 goto done;
2936 DP_DEBUG(dev, QEDR_MSG_MR, "pbl_table_pa = %pa\n",
2937 &info->pbl_table->pa);
2939 /* in usual case we use 2 PBLs, so we add one to free
2940 * list and allocating another one
2942 tmp = qedr_alloc_pbl_tbl(dev, &info->pbl_info, GFP_KERNEL);
2943 if (IS_ERR(tmp)) {
2944 DP_DEBUG(dev, QEDR_MSG_MR, "Extra PBL is not allocated\n");
2945 goto done;
2948 list_add_tail(&tmp->list_entry, &info->free_pbl_list);
2950 DP_DEBUG(dev, QEDR_MSG_MR, "extra pbl_table_pa = %pa\n", &tmp->pa);
2952 done:
2953 if (rc)
2954 free_mr_info(dev, info);
2956 return rc;
2959 struct ib_mr *qedr_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
2960 u64 usr_addr, int acc, struct ib_udata *udata)
2962 struct qedr_dev *dev = get_qedr_dev(ibpd->device);
2963 struct qedr_mr *mr;
2964 struct qedr_pd *pd;
2965 int rc = -ENOMEM;
2967 pd = get_qedr_pd(ibpd);
2968 DP_DEBUG(dev, QEDR_MSG_MR,
2969 "qedr_register user mr pd = %d start = %lld, len = %lld, usr_addr = %lld, acc = %d\n",
2970 pd->pd_id, start, len, usr_addr, acc);
2972 if (acc & IB_ACCESS_REMOTE_WRITE && !(acc & IB_ACCESS_LOCAL_WRITE))
2973 return ERR_PTR(-EINVAL);
2975 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
2976 if (!mr)
2977 return ERR_PTR(rc);
2979 mr->type = QEDR_MR_USER;
2981 mr->umem = ib_umem_get(ibpd->device, start, len, acc);
2982 if (IS_ERR(mr->umem)) {
2983 rc = -EFAULT;
2984 goto err0;
2987 rc = init_mr_info(dev, &mr->info,
2988 ib_umem_num_dma_blocks(mr->umem, PAGE_SIZE), 1);
2989 if (rc)
2990 goto err1;
2992 qedr_populate_pbls(dev, mr->umem, mr->info.pbl_table,
2993 &mr->info.pbl_info, PAGE_SHIFT);
2995 rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
2996 if (rc) {
2997 DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
2998 goto err1;
3001 /* Index only, 18 bit long, lkey = itid << 8 | key */
3002 mr->hw_mr.tid_type = QED_RDMA_TID_REGISTERED_MR;
3003 mr->hw_mr.key = 0;
3004 mr->hw_mr.pd = pd->pd_id;
3005 mr->hw_mr.local_read = 1;
3006 mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
3007 mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
3008 mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
3009 mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
3010 mr->hw_mr.mw_bind = false;
3011 mr->hw_mr.pbl_ptr = mr->info.pbl_table[0].pa;
3012 mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered;
3013 mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size);
3014 mr->hw_mr.page_size_log = PAGE_SHIFT;
3015 mr->hw_mr.length = len;
3016 mr->hw_mr.vaddr = usr_addr;
3017 mr->hw_mr.phy_mr = false;
3018 mr->hw_mr.dma_mr = false;
3020 rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
3021 if (rc) {
3022 DP_ERR(dev, "roce register tid returned an error %d\n", rc);
3023 goto err2;
3026 mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
3027 if (mr->hw_mr.remote_write || mr->hw_mr.remote_read ||
3028 mr->hw_mr.remote_atomic)
3029 mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
3031 DP_DEBUG(dev, QEDR_MSG_MR, "register user mr lkey: %x\n",
3032 mr->ibmr.lkey);
3033 return &mr->ibmr;
3035 err2:
3036 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
3037 err1:
3038 qedr_free_pbl(dev, &mr->info.pbl_info, mr->info.pbl_table);
3039 err0:
3040 kfree(mr);
3041 return ERR_PTR(rc);
3044 int qedr_dereg_mr(struct ib_mr *ib_mr, struct ib_udata *udata)
3046 struct qedr_mr *mr = get_qedr_mr(ib_mr);
3047 struct qedr_dev *dev = get_qedr_dev(ib_mr->device);
3048 int rc = 0;
3050 rc = dev->ops->rdma_deregister_tid(dev->rdma_ctx, mr->hw_mr.itid);
3051 if (rc)
3052 return rc;
3054 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
3056 if (mr->type != QEDR_MR_DMA)
3057 free_mr_info(dev, &mr->info);
3059 /* it could be user registered memory. */
3060 ib_umem_release(mr->umem);
3062 kfree(mr);
3064 return rc;
3067 static struct qedr_mr *__qedr_alloc_mr(struct ib_pd *ibpd,
3068 int max_page_list_len)
3070 struct qedr_pd *pd = get_qedr_pd(ibpd);
3071 struct qedr_dev *dev = get_qedr_dev(ibpd->device);
3072 struct qedr_mr *mr;
3073 int rc = -ENOMEM;
3075 DP_DEBUG(dev, QEDR_MSG_MR,
3076 "qedr_alloc_frmr pd = %d max_page_list_len= %d\n", pd->pd_id,
3077 max_page_list_len);
3079 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3080 if (!mr)
3081 return ERR_PTR(rc);
3083 mr->dev = dev;
3084 mr->type = QEDR_MR_FRMR;
3086 rc = init_mr_info(dev, &mr->info, max_page_list_len, 1);
3087 if (rc)
3088 goto err0;
3090 rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
3091 if (rc) {
3092 DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
3093 goto err0;
3096 /* Index only, 18 bit long, lkey = itid << 8 | key */
3097 mr->hw_mr.tid_type = QED_RDMA_TID_FMR;
3098 mr->hw_mr.key = 0;
3099 mr->hw_mr.pd = pd->pd_id;
3100 mr->hw_mr.local_read = 1;
3101 mr->hw_mr.local_write = 0;
3102 mr->hw_mr.remote_read = 0;
3103 mr->hw_mr.remote_write = 0;
3104 mr->hw_mr.remote_atomic = 0;
3105 mr->hw_mr.mw_bind = false;
3106 mr->hw_mr.pbl_ptr = 0;
3107 mr->hw_mr.pbl_two_level = mr->info.pbl_info.two_layered;
3108 mr->hw_mr.pbl_page_size_log = ilog2(mr->info.pbl_info.pbl_size);
3109 mr->hw_mr.length = 0;
3110 mr->hw_mr.vaddr = 0;
3111 mr->hw_mr.phy_mr = true;
3112 mr->hw_mr.dma_mr = false;
3114 rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
3115 if (rc) {
3116 DP_ERR(dev, "roce register tid returned an error %d\n", rc);
3117 goto err1;
3120 mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
3121 mr->ibmr.rkey = mr->ibmr.lkey;
3123 DP_DEBUG(dev, QEDR_MSG_MR, "alloc frmr: %x\n", mr->ibmr.lkey);
3124 return mr;
3126 err1:
3127 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
3128 err0:
3129 kfree(mr);
3130 return ERR_PTR(rc);
3133 struct ib_mr *qedr_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
3134 u32 max_num_sg)
3136 struct qedr_mr *mr;
3138 if (mr_type != IB_MR_TYPE_MEM_REG)
3139 return ERR_PTR(-EINVAL);
3141 mr = __qedr_alloc_mr(ibpd, max_num_sg);
3143 if (IS_ERR(mr))
3144 return ERR_PTR(-EINVAL);
3146 return &mr->ibmr;
3149 static int qedr_set_page(struct ib_mr *ibmr, u64 addr)
3151 struct qedr_mr *mr = get_qedr_mr(ibmr);
3152 struct qedr_pbl *pbl_table;
3153 struct regpair *pbe;
3154 u32 pbes_in_page;
3156 if (unlikely(mr->npages == mr->info.pbl_info.num_pbes)) {
3157 DP_ERR(mr->dev, "qedr_set_page fails when %d\n", mr->npages);
3158 return -ENOMEM;
3161 DP_DEBUG(mr->dev, QEDR_MSG_MR, "qedr_set_page pages[%d] = 0x%llx\n",
3162 mr->npages, addr);
3164 pbes_in_page = mr->info.pbl_info.pbl_size / sizeof(u64);
3165 pbl_table = mr->info.pbl_table + (mr->npages / pbes_in_page);
3166 pbe = (struct regpair *)pbl_table->va;
3167 pbe += mr->npages % pbes_in_page;
3168 pbe->lo = cpu_to_le32((u32)addr);
3169 pbe->hi = cpu_to_le32((u32)upper_32_bits(addr));
3171 mr->npages++;
3173 return 0;
3176 static void handle_completed_mrs(struct qedr_dev *dev, struct mr_info *info)
3178 int work = info->completed - info->completed_handled - 1;
3180 DP_DEBUG(dev, QEDR_MSG_MR, "Special FMR work = %d\n", work);
3181 while (work-- > 0 && !list_empty(&info->inuse_pbl_list)) {
3182 struct qedr_pbl *pbl;
3184 /* Free all the page list that are possible to be freed
3185 * (all the ones that were invalidated), under the assumption
3186 * that if an FMR was completed successfully that means that
3187 * if there was an invalidate operation before it also ended
3189 pbl = list_first_entry(&info->inuse_pbl_list,
3190 struct qedr_pbl, list_entry);
3191 list_move_tail(&pbl->list_entry, &info->free_pbl_list);
3192 info->completed_handled++;
3196 int qedr_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
3197 int sg_nents, unsigned int *sg_offset)
3199 struct qedr_mr *mr = get_qedr_mr(ibmr);
3201 mr->npages = 0;
3203 handle_completed_mrs(mr->dev, &mr->info);
3204 return ib_sg_to_pages(ibmr, sg, sg_nents, NULL, qedr_set_page);
3207 struct ib_mr *qedr_get_dma_mr(struct ib_pd *ibpd, int acc)
3209 struct qedr_dev *dev = get_qedr_dev(ibpd->device);
3210 struct qedr_pd *pd = get_qedr_pd(ibpd);
3211 struct qedr_mr *mr;
3212 int rc;
3214 mr = kzalloc(sizeof(*mr), GFP_KERNEL);
3215 if (!mr)
3216 return ERR_PTR(-ENOMEM);
3218 mr->type = QEDR_MR_DMA;
3220 rc = dev->ops->rdma_alloc_tid(dev->rdma_ctx, &mr->hw_mr.itid);
3221 if (rc) {
3222 DP_ERR(dev, "roce alloc tid returned an error %d\n", rc);
3223 goto err1;
3226 /* index only, 18 bit long, lkey = itid << 8 | key */
3227 mr->hw_mr.tid_type = QED_RDMA_TID_REGISTERED_MR;
3228 mr->hw_mr.pd = pd->pd_id;
3229 mr->hw_mr.local_read = 1;
3230 mr->hw_mr.local_write = (acc & IB_ACCESS_LOCAL_WRITE) ? 1 : 0;
3231 mr->hw_mr.remote_read = (acc & IB_ACCESS_REMOTE_READ) ? 1 : 0;
3232 mr->hw_mr.remote_write = (acc & IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
3233 mr->hw_mr.remote_atomic = (acc & IB_ACCESS_REMOTE_ATOMIC) ? 1 : 0;
3234 mr->hw_mr.dma_mr = true;
3236 rc = dev->ops->rdma_register_tid(dev->rdma_ctx, &mr->hw_mr);
3237 if (rc) {
3238 DP_ERR(dev, "roce register tid returned an error %d\n", rc);
3239 goto err2;
3242 mr->ibmr.lkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
3243 if (mr->hw_mr.remote_write || mr->hw_mr.remote_read ||
3244 mr->hw_mr.remote_atomic)
3245 mr->ibmr.rkey = mr->hw_mr.itid << 8 | mr->hw_mr.key;
3247 DP_DEBUG(dev, QEDR_MSG_MR, "get dma mr: lkey = %x\n", mr->ibmr.lkey);
3248 return &mr->ibmr;
3250 err2:
3251 dev->ops->rdma_free_tid(dev->rdma_ctx, mr->hw_mr.itid);
3252 err1:
3253 kfree(mr);
3254 return ERR_PTR(rc);
3257 static inline int qedr_wq_is_full(struct qedr_qp_hwq_info *wq)
3259 return (((wq->prod + 1) % wq->max_wr) == wq->cons);
3262 static int sge_data_len(struct ib_sge *sg_list, int num_sge)
3264 int i, len = 0;
3266 for (i = 0; i < num_sge; i++)
3267 len += sg_list[i].length;
3269 return len;
3272 static void swap_wqe_data64(u64 *p)
3274 int i;
3276 for (i = 0; i < QEDR_SQE_ELEMENT_SIZE / sizeof(u64); i++, p++)
3277 *p = cpu_to_be64(cpu_to_le64(*p));
3280 static u32 qedr_prepare_sq_inline_data(struct qedr_dev *dev,
3281 struct qedr_qp *qp, u8 *wqe_size,
3282 const struct ib_send_wr *wr,
3283 const struct ib_send_wr **bad_wr,
3284 u8 *bits, u8 bit)
3286 u32 data_size = sge_data_len(wr->sg_list, wr->num_sge);
3287 char *seg_prt, *wqe;
3288 int i, seg_siz;
3290 if (data_size > ROCE_REQ_MAX_INLINE_DATA_SIZE) {
3291 DP_ERR(dev, "Too much inline data in WR: %d\n", data_size);
3292 *bad_wr = wr;
3293 return 0;
3296 if (!data_size)
3297 return data_size;
3299 *bits |= bit;
3301 seg_prt = NULL;
3302 wqe = NULL;
3303 seg_siz = 0;
3305 /* Copy data inline */
3306 for (i = 0; i < wr->num_sge; i++) {
3307 u32 len = wr->sg_list[i].length;
3308 void *src = (void *)(uintptr_t)wr->sg_list[i].addr;
3310 while (len > 0) {
3311 u32 cur;
3313 /* New segment required */
3314 if (!seg_siz) {
3315 wqe = (char *)qed_chain_produce(&qp->sq.pbl);
3316 seg_prt = wqe;
3317 seg_siz = sizeof(struct rdma_sq_common_wqe);
3318 (*wqe_size)++;
3321 /* Calculate currently allowed length */
3322 cur = min_t(u32, len, seg_siz);
3323 memcpy(seg_prt, src, cur);
3325 /* Update segment variables */
3326 seg_prt += cur;
3327 seg_siz -= cur;
3329 /* Update sge variables */
3330 src += cur;
3331 len -= cur;
3333 /* Swap fully-completed segments */
3334 if (!seg_siz)
3335 swap_wqe_data64((u64 *)wqe);
3339 /* swap last not completed segment */
3340 if (seg_siz)
3341 swap_wqe_data64((u64 *)wqe);
3343 return data_size;
3346 #define RQ_SGE_SET(sge, vaddr, vlength, vflags) \
3347 do { \
3348 DMA_REGPAIR_LE(sge->addr, vaddr); \
3349 (sge)->length = cpu_to_le32(vlength); \
3350 (sge)->flags = cpu_to_le32(vflags); \
3351 } while (0)
3353 #define SRQ_HDR_SET(hdr, vwr_id, num_sge) \
3354 do { \
3355 DMA_REGPAIR_LE(hdr->wr_id, vwr_id); \
3356 (hdr)->num_sges = num_sge; \
3357 } while (0)
3359 #define SRQ_SGE_SET(sge, vaddr, vlength, vlkey) \
3360 do { \
3361 DMA_REGPAIR_LE(sge->addr, vaddr); \
3362 (sge)->length = cpu_to_le32(vlength); \
3363 (sge)->l_key = cpu_to_le32(vlkey); \
3364 } while (0)
3366 static u32 qedr_prepare_sq_sges(struct qedr_qp *qp, u8 *wqe_size,
3367 const struct ib_send_wr *wr)
3369 u32 data_size = 0;
3370 int i;
3372 for (i = 0; i < wr->num_sge; i++) {
3373 struct rdma_sq_sge *sge = qed_chain_produce(&qp->sq.pbl);
3375 DMA_REGPAIR_LE(sge->addr, wr->sg_list[i].addr);
3376 sge->l_key = cpu_to_le32(wr->sg_list[i].lkey);
3377 sge->length = cpu_to_le32(wr->sg_list[i].length);
3378 data_size += wr->sg_list[i].length;
3381 if (wqe_size)
3382 *wqe_size += wr->num_sge;
3384 return data_size;
3387 static u32 qedr_prepare_sq_rdma_data(struct qedr_dev *dev,
3388 struct qedr_qp *qp,
3389 struct rdma_sq_rdma_wqe_1st *rwqe,
3390 struct rdma_sq_rdma_wqe_2nd *rwqe2,
3391 const struct ib_send_wr *wr,
3392 const struct ib_send_wr **bad_wr)
3394 rwqe2->r_key = cpu_to_le32(rdma_wr(wr)->rkey);
3395 DMA_REGPAIR_LE(rwqe2->remote_va, rdma_wr(wr)->remote_addr);
3397 if (wr->send_flags & IB_SEND_INLINE &&
3398 (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM ||
3399 wr->opcode == IB_WR_RDMA_WRITE)) {
3400 u8 flags = 0;
3402 SET_FIELD2(flags, RDMA_SQ_RDMA_WQE_1ST_INLINE_FLG, 1);
3403 return qedr_prepare_sq_inline_data(dev, qp, &rwqe->wqe_size, wr,
3404 bad_wr, &rwqe->flags, flags);
3407 return qedr_prepare_sq_sges(qp, &rwqe->wqe_size, wr);
3410 static u32 qedr_prepare_sq_send_data(struct qedr_dev *dev,
3411 struct qedr_qp *qp,
3412 struct rdma_sq_send_wqe_1st *swqe,
3413 struct rdma_sq_send_wqe_2st *swqe2,
3414 const struct ib_send_wr *wr,
3415 const struct ib_send_wr **bad_wr)
3417 memset(swqe2, 0, sizeof(*swqe2));
3418 if (wr->send_flags & IB_SEND_INLINE) {
3419 u8 flags = 0;
3421 SET_FIELD2(flags, RDMA_SQ_SEND_WQE_INLINE_FLG, 1);
3422 return qedr_prepare_sq_inline_data(dev, qp, &swqe->wqe_size, wr,
3423 bad_wr, &swqe->flags, flags);
3426 return qedr_prepare_sq_sges(qp, &swqe->wqe_size, wr);
3429 static int qedr_prepare_reg(struct qedr_qp *qp,
3430 struct rdma_sq_fmr_wqe_1st *fwqe1,
3431 const struct ib_reg_wr *wr)
3433 struct qedr_mr *mr = get_qedr_mr(wr->mr);
3434 struct rdma_sq_fmr_wqe_2nd *fwqe2;
3436 fwqe2 = (struct rdma_sq_fmr_wqe_2nd *)qed_chain_produce(&qp->sq.pbl);
3437 fwqe1->addr.hi = upper_32_bits(mr->ibmr.iova);
3438 fwqe1->addr.lo = lower_32_bits(mr->ibmr.iova);
3439 fwqe1->l_key = wr->key;
3441 fwqe2->access_ctrl = 0;
3443 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_READ,
3444 !!(wr->access & IB_ACCESS_REMOTE_READ));
3445 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_REMOTE_WRITE,
3446 !!(wr->access & IB_ACCESS_REMOTE_WRITE));
3447 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_ENABLE_ATOMIC,
3448 !!(wr->access & IB_ACCESS_REMOTE_ATOMIC));
3449 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_READ, 1);
3450 SET_FIELD2(fwqe2->access_ctrl, RDMA_SQ_FMR_WQE_2ND_LOCAL_WRITE,
3451 !!(wr->access & IB_ACCESS_LOCAL_WRITE));
3452 fwqe2->fmr_ctrl = 0;
3454 SET_FIELD2(fwqe2->fmr_ctrl, RDMA_SQ_FMR_WQE_2ND_PAGE_SIZE_LOG,
3455 ilog2(mr->ibmr.page_size) - 12);
3457 fwqe2->length_hi = 0;
3458 fwqe2->length_lo = mr->ibmr.length;
3459 fwqe2->pbl_addr.hi = upper_32_bits(mr->info.pbl_table->pa);
3460 fwqe2->pbl_addr.lo = lower_32_bits(mr->info.pbl_table->pa);
3462 qp->wqe_wr_id[qp->sq.prod].mr = mr;
3464 return 0;
3467 static enum ib_wc_opcode qedr_ib_to_wc_opcode(enum ib_wr_opcode opcode)
3469 switch (opcode) {
3470 case IB_WR_RDMA_WRITE:
3471 case IB_WR_RDMA_WRITE_WITH_IMM:
3472 return IB_WC_RDMA_WRITE;
3473 case IB_WR_SEND_WITH_IMM:
3474 case IB_WR_SEND:
3475 case IB_WR_SEND_WITH_INV:
3476 return IB_WC_SEND;
3477 case IB_WR_RDMA_READ:
3478 case IB_WR_RDMA_READ_WITH_INV:
3479 return IB_WC_RDMA_READ;
3480 case IB_WR_ATOMIC_CMP_AND_SWP:
3481 return IB_WC_COMP_SWAP;
3482 case IB_WR_ATOMIC_FETCH_AND_ADD:
3483 return IB_WC_FETCH_ADD;
3484 case IB_WR_REG_MR:
3485 return IB_WC_REG_MR;
3486 case IB_WR_LOCAL_INV:
3487 return IB_WC_LOCAL_INV;
3488 default:
3489 return IB_WC_SEND;
3493 static inline bool qedr_can_post_send(struct qedr_qp *qp,
3494 const struct ib_send_wr *wr)
3496 int wq_is_full, err_wr, pbl_is_full;
3497 struct qedr_dev *dev = qp->dev;
3499 /* prevent SQ overflow and/or processing of a bad WR */
3500 err_wr = wr->num_sge > qp->sq.max_sges;
3501 wq_is_full = qedr_wq_is_full(&qp->sq);
3502 pbl_is_full = qed_chain_get_elem_left_u32(&qp->sq.pbl) <
3503 QEDR_MAX_SQE_ELEMENTS_PER_SQE;
3504 if (wq_is_full || err_wr || pbl_is_full) {
3505 if (wq_is_full && !(qp->err_bitmap & QEDR_QP_ERR_SQ_FULL)) {
3506 DP_ERR(dev,
3507 "error: WQ is full. Post send on QP %p failed (this error appears only once)\n",
3508 qp);
3509 qp->err_bitmap |= QEDR_QP_ERR_SQ_FULL;
3512 if (err_wr && !(qp->err_bitmap & QEDR_QP_ERR_BAD_SR)) {
3513 DP_ERR(dev,
3514 "error: WR is bad. Post send on QP %p failed (this error appears only once)\n",
3515 qp);
3516 qp->err_bitmap |= QEDR_QP_ERR_BAD_SR;
3519 if (pbl_is_full &&
3520 !(qp->err_bitmap & QEDR_QP_ERR_SQ_PBL_FULL)) {
3521 DP_ERR(dev,
3522 "error: WQ PBL is full. Post send on QP %p failed (this error appears only once)\n",
3523 qp);
3524 qp->err_bitmap |= QEDR_QP_ERR_SQ_PBL_FULL;
3526 return false;
3528 return true;
3531 static int __qedr_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3532 const struct ib_send_wr **bad_wr)
3534 struct qedr_dev *dev = get_qedr_dev(ibqp->device);
3535 struct qedr_qp *qp = get_qedr_qp(ibqp);
3536 struct rdma_sq_atomic_wqe_1st *awqe1;
3537 struct rdma_sq_atomic_wqe_2nd *awqe2;
3538 struct rdma_sq_atomic_wqe_3rd *awqe3;
3539 struct rdma_sq_send_wqe_2st *swqe2;
3540 struct rdma_sq_local_inv_wqe *iwqe;
3541 struct rdma_sq_rdma_wqe_2nd *rwqe2;
3542 struct rdma_sq_send_wqe_1st *swqe;
3543 struct rdma_sq_rdma_wqe_1st *rwqe;
3544 struct rdma_sq_fmr_wqe_1st *fwqe1;
3545 struct rdma_sq_common_wqe *wqe;
3546 u32 length;
3547 int rc = 0;
3548 bool comp;
3550 if (!qedr_can_post_send(qp, wr)) {
3551 *bad_wr = wr;
3552 return -ENOMEM;
3555 wqe = qed_chain_produce(&qp->sq.pbl);
3556 qp->wqe_wr_id[qp->sq.prod].signaled =
3557 !!(wr->send_flags & IB_SEND_SIGNALED) || qp->signaled;
3559 wqe->flags = 0;
3560 SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_SE_FLG,
3561 !!(wr->send_flags & IB_SEND_SOLICITED));
3562 comp = (!!(wr->send_flags & IB_SEND_SIGNALED)) || qp->signaled;
3563 SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_COMP_FLG, comp);
3564 SET_FIELD2(wqe->flags, RDMA_SQ_SEND_WQE_RD_FENCE_FLG,
3565 !!(wr->send_flags & IB_SEND_FENCE));
3566 wqe->prev_wqe_size = qp->prev_wqe_size;
3568 qp->wqe_wr_id[qp->sq.prod].opcode = qedr_ib_to_wc_opcode(wr->opcode);
3570 switch (wr->opcode) {
3571 case IB_WR_SEND_WITH_IMM:
3572 if (unlikely(rdma_protocol_iwarp(&dev->ibdev, 1))) {
3573 rc = -EINVAL;
3574 *bad_wr = wr;
3575 break;
3577 wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_IMM;
3578 swqe = (struct rdma_sq_send_wqe_1st *)wqe;
3579 swqe->wqe_size = 2;
3580 swqe2 = qed_chain_produce(&qp->sq.pbl);
3582 swqe->inv_key_or_imm_data = cpu_to_le32(be32_to_cpu(wr->ex.imm_data));
3583 length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
3584 wr, bad_wr);
3585 swqe->length = cpu_to_le32(length);
3586 qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
3587 qp->prev_wqe_size = swqe->wqe_size;
3588 qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
3589 break;
3590 case IB_WR_SEND:
3591 wqe->req_type = RDMA_SQ_REQ_TYPE_SEND;
3592 swqe = (struct rdma_sq_send_wqe_1st *)wqe;
3594 swqe->wqe_size = 2;
3595 swqe2 = qed_chain_produce(&qp->sq.pbl);
3596 length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
3597 wr, bad_wr);
3598 swqe->length = cpu_to_le32(length);
3599 qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
3600 qp->prev_wqe_size = swqe->wqe_size;
3601 qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
3602 break;
3603 case IB_WR_SEND_WITH_INV:
3604 wqe->req_type = RDMA_SQ_REQ_TYPE_SEND_WITH_INVALIDATE;
3605 swqe = (struct rdma_sq_send_wqe_1st *)wqe;
3606 swqe2 = qed_chain_produce(&qp->sq.pbl);
3607 swqe->wqe_size = 2;
3608 swqe->inv_key_or_imm_data = cpu_to_le32(wr->ex.invalidate_rkey);
3609 length = qedr_prepare_sq_send_data(dev, qp, swqe, swqe2,
3610 wr, bad_wr);
3611 swqe->length = cpu_to_le32(length);
3612 qp->wqe_wr_id[qp->sq.prod].wqe_size = swqe->wqe_size;
3613 qp->prev_wqe_size = swqe->wqe_size;
3614 qp->wqe_wr_id[qp->sq.prod].bytes_len = swqe->length;
3615 break;
3617 case IB_WR_RDMA_WRITE_WITH_IMM:
3618 if (unlikely(rdma_protocol_iwarp(&dev->ibdev, 1))) {
3619 rc = -EINVAL;
3620 *bad_wr = wr;
3621 break;
3623 wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR_WITH_IMM;
3624 rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
3626 rwqe->wqe_size = 2;
3627 rwqe->imm_data = htonl(cpu_to_le32(wr->ex.imm_data));
3628 rwqe2 = qed_chain_produce(&qp->sq.pbl);
3629 length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
3630 wr, bad_wr);
3631 rwqe->length = cpu_to_le32(length);
3632 qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
3633 qp->prev_wqe_size = rwqe->wqe_size;
3634 qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
3635 break;
3636 case IB_WR_RDMA_WRITE:
3637 wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_WR;
3638 rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
3640 rwqe->wqe_size = 2;
3641 rwqe2 = qed_chain_produce(&qp->sq.pbl);
3642 length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
3643 wr, bad_wr);
3644 rwqe->length = cpu_to_le32(length);
3645 qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
3646 qp->prev_wqe_size = rwqe->wqe_size;
3647 qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
3648 break;
3649 case IB_WR_RDMA_READ_WITH_INV:
3650 SET_FIELD2(wqe->flags, RDMA_SQ_RDMA_WQE_1ST_READ_INV_FLG, 1);
3651 fallthrough; /* same is identical to RDMA READ */
3653 case IB_WR_RDMA_READ:
3654 wqe->req_type = RDMA_SQ_REQ_TYPE_RDMA_RD;
3655 rwqe = (struct rdma_sq_rdma_wqe_1st *)wqe;
3657 rwqe->wqe_size = 2;
3658 rwqe2 = qed_chain_produce(&qp->sq.pbl);
3659 length = qedr_prepare_sq_rdma_data(dev, qp, rwqe, rwqe2,
3660 wr, bad_wr);
3661 rwqe->length = cpu_to_le32(length);
3662 qp->wqe_wr_id[qp->sq.prod].wqe_size = rwqe->wqe_size;
3663 qp->prev_wqe_size = rwqe->wqe_size;
3664 qp->wqe_wr_id[qp->sq.prod].bytes_len = rwqe->length;
3665 break;
3667 case IB_WR_ATOMIC_CMP_AND_SWP:
3668 case IB_WR_ATOMIC_FETCH_AND_ADD:
3669 awqe1 = (struct rdma_sq_atomic_wqe_1st *)wqe;
3670 awqe1->wqe_size = 4;
3672 awqe2 = qed_chain_produce(&qp->sq.pbl);
3673 DMA_REGPAIR_LE(awqe2->remote_va, atomic_wr(wr)->remote_addr);
3674 awqe2->r_key = cpu_to_le32(atomic_wr(wr)->rkey);
3676 awqe3 = qed_chain_produce(&qp->sq.pbl);
3678 if (wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
3679 wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_ADD;
3680 DMA_REGPAIR_LE(awqe3->swap_data,
3681 atomic_wr(wr)->compare_add);
3682 } else {
3683 wqe->req_type = RDMA_SQ_REQ_TYPE_ATOMIC_CMP_AND_SWAP;
3684 DMA_REGPAIR_LE(awqe3->swap_data,
3685 atomic_wr(wr)->swap);
3686 DMA_REGPAIR_LE(awqe3->cmp_data,
3687 atomic_wr(wr)->compare_add);
3690 qedr_prepare_sq_sges(qp, NULL, wr);
3692 qp->wqe_wr_id[qp->sq.prod].wqe_size = awqe1->wqe_size;
3693 qp->prev_wqe_size = awqe1->wqe_size;
3694 break;
3696 case IB_WR_LOCAL_INV:
3697 iwqe = (struct rdma_sq_local_inv_wqe *)wqe;
3698 iwqe->wqe_size = 1;
3700 iwqe->req_type = RDMA_SQ_REQ_TYPE_LOCAL_INVALIDATE;
3701 iwqe->inv_l_key = wr->ex.invalidate_rkey;
3702 qp->wqe_wr_id[qp->sq.prod].wqe_size = iwqe->wqe_size;
3703 qp->prev_wqe_size = iwqe->wqe_size;
3704 break;
3705 case IB_WR_REG_MR:
3706 DP_DEBUG(dev, QEDR_MSG_CQ, "REG_MR\n");
3707 wqe->req_type = RDMA_SQ_REQ_TYPE_FAST_MR;
3708 fwqe1 = (struct rdma_sq_fmr_wqe_1st *)wqe;
3709 fwqe1->wqe_size = 2;
3711 rc = qedr_prepare_reg(qp, fwqe1, reg_wr(wr));
3712 if (rc) {
3713 DP_ERR(dev, "IB_REG_MR failed rc=%d\n", rc);
3714 *bad_wr = wr;
3715 break;
3718 qp->wqe_wr_id[qp->sq.prod].wqe_size = fwqe1->wqe_size;
3719 qp->prev_wqe_size = fwqe1->wqe_size;
3720 break;
3721 default:
3722 DP_ERR(dev, "invalid opcode 0x%x!\n", wr->opcode);
3723 rc = -EINVAL;
3724 *bad_wr = wr;
3725 break;
3728 if (*bad_wr) {
3729 u16 value;
3731 /* Restore prod to its position before
3732 * this WR was processed
3734 value = le16_to_cpu(qp->sq.db_data.data.value);
3735 qed_chain_set_prod(&qp->sq.pbl, value, wqe);
3737 /* Restore prev_wqe_size */
3738 qp->prev_wqe_size = wqe->prev_wqe_size;
3739 rc = -EINVAL;
3740 DP_ERR(dev, "POST SEND FAILED\n");
3743 return rc;
3746 int qedr_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
3747 const struct ib_send_wr **bad_wr)
3749 struct qedr_dev *dev = get_qedr_dev(ibqp->device);
3750 struct qedr_qp *qp = get_qedr_qp(ibqp);
3751 unsigned long flags;
3752 int rc = 0;
3754 *bad_wr = NULL;
3756 if (qp->qp_type == IB_QPT_GSI)
3757 return qedr_gsi_post_send(ibqp, wr, bad_wr);
3759 spin_lock_irqsave(&qp->q_lock, flags);
3761 if (rdma_protocol_roce(&dev->ibdev, 1)) {
3762 if ((qp->state != QED_ROCE_QP_STATE_RTS) &&
3763 (qp->state != QED_ROCE_QP_STATE_ERR) &&
3764 (qp->state != QED_ROCE_QP_STATE_SQD)) {
3765 spin_unlock_irqrestore(&qp->q_lock, flags);
3766 *bad_wr = wr;
3767 DP_DEBUG(dev, QEDR_MSG_CQ,
3768 "QP in wrong state! QP icid=0x%x state %d\n",
3769 qp->icid, qp->state);
3770 return -EINVAL;
3774 while (wr) {
3775 rc = __qedr_post_send(ibqp, wr, bad_wr);
3776 if (rc)
3777 break;
3779 qp->wqe_wr_id[qp->sq.prod].wr_id = wr->wr_id;
3781 qedr_inc_sw_prod(&qp->sq);
3783 qp->sq.db_data.data.value++;
3785 wr = wr->next;
3788 /* Trigger doorbell
3789 * If there was a failure in the first WR then it will be triggered in
3790 * vane. However this is not harmful (as long as the producer value is
3791 * unchanged). For performance reasons we avoid checking for this
3792 * redundant doorbell.
3794 * qp->wqe_wr_id is accessed during qedr_poll_cq, as
3795 * soon as we give the doorbell, we could get a completion
3796 * for this wr, therefore we need to make sure that the
3797 * memory is updated before giving the doorbell.
3798 * During qedr_poll_cq, rmb is called before accessing the
3799 * cqe. This covers for the smp_rmb as well.
3801 smp_wmb();
3802 writel(qp->sq.db_data.raw, qp->sq.db);
3804 spin_unlock_irqrestore(&qp->q_lock, flags);
3806 return rc;
3809 static u32 qedr_srq_elem_left(struct qedr_srq_hwq_info *hw_srq)
3811 u32 used;
3813 /* Calculate number of elements used based on producer
3814 * count and consumer count and subtract it from max
3815 * work request supported so that we get elements left.
3817 used = hw_srq->wr_prod_cnt - (u32)atomic_read(&hw_srq->wr_cons_cnt);
3819 return hw_srq->max_wr - used;
3822 int qedr_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
3823 const struct ib_recv_wr **bad_wr)
3825 struct qedr_srq *srq = get_qedr_srq(ibsrq);
3826 struct qedr_srq_hwq_info *hw_srq;
3827 struct qedr_dev *dev = srq->dev;
3828 struct qed_chain *pbl;
3829 unsigned long flags;
3830 int status = 0;
3831 u32 num_sge;
3833 spin_lock_irqsave(&srq->lock, flags);
3835 hw_srq = &srq->hw_srq;
3836 pbl = &srq->hw_srq.pbl;
3837 while (wr) {
3838 struct rdma_srq_wqe_header *hdr;
3839 int i;
3841 if (!qedr_srq_elem_left(hw_srq) ||
3842 wr->num_sge > srq->hw_srq.max_sges) {
3843 DP_ERR(dev, "Can't post WR (%d,%d) || (%d > %d)\n",
3844 hw_srq->wr_prod_cnt,
3845 atomic_read(&hw_srq->wr_cons_cnt),
3846 wr->num_sge, srq->hw_srq.max_sges);
3847 status = -ENOMEM;
3848 *bad_wr = wr;
3849 break;
3852 hdr = qed_chain_produce(pbl);
3853 num_sge = wr->num_sge;
3854 /* Set number of sge and work request id in header */
3855 SRQ_HDR_SET(hdr, wr->wr_id, num_sge);
3857 srq->hw_srq.wr_prod_cnt++;
3858 hw_srq->wqe_prod++;
3859 hw_srq->sge_prod++;
3861 DP_DEBUG(dev, QEDR_MSG_SRQ,
3862 "SRQ WR: SGEs: %d with wr_id[%d] = %llx\n",
3863 wr->num_sge, hw_srq->wqe_prod, wr->wr_id);
3865 for (i = 0; i < wr->num_sge; i++) {
3866 struct rdma_srq_sge *srq_sge = qed_chain_produce(pbl);
3868 /* Set SGE length, lkey and address */
3869 SRQ_SGE_SET(srq_sge, wr->sg_list[i].addr,
3870 wr->sg_list[i].length, wr->sg_list[i].lkey);
3872 DP_DEBUG(dev, QEDR_MSG_SRQ,
3873 "[%d]: len %d key %x addr %x:%x\n",
3874 i, srq_sge->length, srq_sge->l_key,
3875 srq_sge->addr.hi, srq_sge->addr.lo);
3876 hw_srq->sge_prod++;
3879 /* Update WQE and SGE information before
3880 * updating producer.
3882 dma_wmb();
3884 /* SRQ producer is 8 bytes. Need to update SGE producer index
3885 * in first 4 bytes and need to update WQE producer in
3886 * next 4 bytes.
3888 srq->hw_srq.virt_prod_pair_addr->sge_prod = cpu_to_le32(hw_srq->sge_prod);
3889 /* Make sure sge producer is updated first */
3890 dma_wmb();
3891 srq->hw_srq.virt_prod_pair_addr->wqe_prod = cpu_to_le32(hw_srq->wqe_prod);
3893 wr = wr->next;
3896 DP_DEBUG(dev, QEDR_MSG_SRQ, "POST: Elements in S-RQ: %d\n",
3897 qed_chain_get_elem_left(pbl));
3898 spin_unlock_irqrestore(&srq->lock, flags);
3900 return status;
3903 int qedr_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
3904 const struct ib_recv_wr **bad_wr)
3906 struct qedr_qp *qp = get_qedr_qp(ibqp);
3907 struct qedr_dev *dev = qp->dev;
3908 unsigned long flags;
3909 int status = 0;
3911 if (qp->qp_type == IB_QPT_GSI)
3912 return qedr_gsi_post_recv(ibqp, wr, bad_wr);
3914 spin_lock_irqsave(&qp->q_lock, flags);
3916 if (qp->state == QED_ROCE_QP_STATE_RESET) {
3917 spin_unlock_irqrestore(&qp->q_lock, flags);
3918 *bad_wr = wr;
3919 return -EINVAL;
3922 while (wr) {
3923 int i;
3925 if (qed_chain_get_elem_left_u32(&qp->rq.pbl) <
3926 QEDR_MAX_RQE_ELEMENTS_PER_RQE ||
3927 wr->num_sge > qp->rq.max_sges) {
3928 DP_ERR(dev, "Can't post WR (%d < %d) || (%d > %d)\n",
3929 qed_chain_get_elem_left_u32(&qp->rq.pbl),
3930 QEDR_MAX_RQE_ELEMENTS_PER_RQE, wr->num_sge,
3931 qp->rq.max_sges);
3932 status = -ENOMEM;
3933 *bad_wr = wr;
3934 break;
3936 for (i = 0; i < wr->num_sge; i++) {
3937 u32 flags = 0;
3938 struct rdma_rq_sge *rqe =
3939 qed_chain_produce(&qp->rq.pbl);
3941 /* First one must include the number
3942 * of SGE in the list
3944 if (!i)
3945 SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES,
3946 wr->num_sge);
3948 SET_FIELD(flags, RDMA_RQ_SGE_L_KEY_LO,
3949 wr->sg_list[i].lkey);
3951 RQ_SGE_SET(rqe, wr->sg_list[i].addr,
3952 wr->sg_list[i].length, flags);
3955 /* Special case of no sges. FW requires between 1-4 sges...
3956 * in this case we need to post 1 sge with length zero. this is
3957 * because rdma write with immediate consumes an RQ.
3959 if (!wr->num_sge) {
3960 u32 flags = 0;
3961 struct rdma_rq_sge *rqe =
3962 qed_chain_produce(&qp->rq.pbl);
3964 /* First one must include the number
3965 * of SGE in the list
3967 SET_FIELD(flags, RDMA_RQ_SGE_L_KEY_LO, 0);
3968 SET_FIELD(flags, RDMA_RQ_SGE_NUM_SGES, 1);
3970 RQ_SGE_SET(rqe, 0, 0, flags);
3971 i = 1;
3974 qp->rqe_wr_id[qp->rq.prod].wr_id = wr->wr_id;
3975 qp->rqe_wr_id[qp->rq.prod].wqe_size = i;
3977 qedr_inc_sw_prod(&qp->rq);
3979 /* qp->rqe_wr_id is accessed during qedr_poll_cq, as
3980 * soon as we give the doorbell, we could get a completion
3981 * for this wr, therefore we need to make sure that the
3982 * memory is update before giving the doorbell.
3983 * During qedr_poll_cq, rmb is called before accessing the
3984 * cqe. This covers for the smp_rmb as well.
3986 smp_wmb();
3988 qp->rq.db_data.data.value++;
3990 writel(qp->rq.db_data.raw, qp->rq.db);
3992 if (rdma_protocol_iwarp(&dev->ibdev, 1)) {
3993 writel(qp->rq.iwarp_db2_data.raw, qp->rq.iwarp_db2);
3996 wr = wr->next;
3999 spin_unlock_irqrestore(&qp->q_lock, flags);
4001 return status;
4004 static int is_valid_cqe(struct qedr_cq *cq, union rdma_cqe *cqe)
4006 struct rdma_cqe_requester *resp_cqe = &cqe->req;
4008 return (resp_cqe->flags & RDMA_CQE_REQUESTER_TOGGLE_BIT_MASK) ==
4009 cq->pbl_toggle;
4012 static struct qedr_qp *cqe_get_qp(union rdma_cqe *cqe)
4014 struct rdma_cqe_requester *resp_cqe = &cqe->req;
4015 struct qedr_qp *qp;
4017 qp = (struct qedr_qp *)(uintptr_t)HILO_GEN(resp_cqe->qp_handle.hi,
4018 resp_cqe->qp_handle.lo,
4019 u64);
4020 return qp;
4023 static enum rdma_cqe_type cqe_get_type(union rdma_cqe *cqe)
4025 struct rdma_cqe_requester *resp_cqe = &cqe->req;
4027 return GET_FIELD(resp_cqe->flags, RDMA_CQE_REQUESTER_TYPE);
4030 /* Return latest CQE (needs processing) */
4031 static union rdma_cqe *get_cqe(struct qedr_cq *cq)
4033 return cq->latest_cqe;
4036 /* In fmr we need to increase the number of fmr completed counter for the fmr
4037 * algorithm determining whether we can free a pbl or not.
4038 * we need to perform this whether the work request was signaled or not. for
4039 * this purpose we call this function from the condition that checks if a wr
4040 * should be skipped, to make sure we don't miss it ( possibly this fmr
4041 * operation was not signalted)
4043 static inline void qedr_chk_if_fmr(struct qedr_qp *qp)
4045 if (qp->wqe_wr_id[qp->sq.cons].opcode == IB_WC_REG_MR)
4046 qp->wqe_wr_id[qp->sq.cons].mr->info.completed++;
4049 static int process_req(struct qedr_dev *dev, struct qedr_qp *qp,
4050 struct qedr_cq *cq, int num_entries,
4051 struct ib_wc *wc, u16 hw_cons, enum ib_wc_status status,
4052 int force)
4054 u16 cnt = 0;
4056 while (num_entries && qp->sq.wqe_cons != hw_cons) {
4057 if (!qp->wqe_wr_id[qp->sq.cons].signaled && !force) {
4058 qedr_chk_if_fmr(qp);
4059 /* skip WC */
4060 goto next_cqe;
4063 /* fill WC */
4064 wc->status = status;
4065 wc->vendor_err = 0;
4066 wc->wc_flags = 0;
4067 wc->src_qp = qp->id;
4068 wc->qp = &qp->ibqp;
4070 wc->wr_id = qp->wqe_wr_id[qp->sq.cons].wr_id;
4071 wc->opcode = qp->wqe_wr_id[qp->sq.cons].opcode;
4073 switch (wc->opcode) {
4074 case IB_WC_RDMA_WRITE:
4075 wc->byte_len = qp->wqe_wr_id[qp->sq.cons].bytes_len;
4076 break;
4077 case IB_WC_COMP_SWAP:
4078 case IB_WC_FETCH_ADD:
4079 wc->byte_len = 8;
4080 break;
4081 case IB_WC_REG_MR:
4082 qp->wqe_wr_id[qp->sq.cons].mr->info.completed++;
4083 break;
4084 case IB_WC_RDMA_READ:
4085 case IB_WC_SEND:
4086 wc->byte_len = qp->wqe_wr_id[qp->sq.cons].bytes_len;
4087 break;
4088 default:
4089 break;
4092 num_entries--;
4093 wc++;
4094 cnt++;
4095 next_cqe:
4096 while (qp->wqe_wr_id[qp->sq.cons].wqe_size--)
4097 qed_chain_consume(&qp->sq.pbl);
4098 qedr_inc_sw_cons(&qp->sq);
4101 return cnt;
4104 static int qedr_poll_cq_req(struct qedr_dev *dev,
4105 struct qedr_qp *qp, struct qedr_cq *cq,
4106 int num_entries, struct ib_wc *wc,
4107 struct rdma_cqe_requester *req)
4109 int cnt = 0;
4111 switch (req->status) {
4112 case RDMA_CQE_REQ_STS_OK:
4113 cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons,
4114 IB_WC_SUCCESS, 0);
4115 break;
4116 case RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR:
4117 if (qp->state != QED_ROCE_QP_STATE_ERR)
4118 DP_DEBUG(dev, QEDR_MSG_CQ,
4119 "Error: POLL CQ with RDMA_CQE_REQ_STS_WORK_REQUEST_FLUSHED_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4120 cq->icid, qp->icid);
4121 cnt = process_req(dev, qp, cq, num_entries, wc, req->sq_cons,
4122 IB_WC_WR_FLUSH_ERR, 1);
4123 break;
4124 default:
4125 /* process all WQE before the cosumer */
4126 qp->state = QED_ROCE_QP_STATE_ERR;
4127 cnt = process_req(dev, qp, cq, num_entries, wc,
4128 req->sq_cons - 1, IB_WC_SUCCESS, 0);
4129 wc += cnt;
4130 /* if we have extra WC fill it with actual error info */
4131 if (cnt < num_entries) {
4132 enum ib_wc_status wc_status;
4134 switch (req->status) {
4135 case RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR:
4136 DP_ERR(dev,
4137 "Error: POLL CQ with RDMA_CQE_REQ_STS_BAD_RESPONSE_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4138 cq->icid, qp->icid);
4139 wc_status = IB_WC_BAD_RESP_ERR;
4140 break;
4141 case RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR:
4142 DP_ERR(dev,
4143 "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_LENGTH_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4144 cq->icid, qp->icid);
4145 wc_status = IB_WC_LOC_LEN_ERR;
4146 break;
4147 case RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR:
4148 DP_ERR(dev,
4149 "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_QP_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4150 cq->icid, qp->icid);
4151 wc_status = IB_WC_LOC_QP_OP_ERR;
4152 break;
4153 case RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR:
4154 DP_ERR(dev,
4155 "Error: POLL CQ with RDMA_CQE_REQ_STS_LOCAL_PROTECTION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4156 cq->icid, qp->icid);
4157 wc_status = IB_WC_LOC_PROT_ERR;
4158 break;
4159 case RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR:
4160 DP_ERR(dev,
4161 "Error: POLL CQ with RDMA_CQE_REQ_STS_MEMORY_MGT_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4162 cq->icid, qp->icid);
4163 wc_status = IB_WC_MW_BIND_ERR;
4164 break;
4165 case RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR:
4166 DP_ERR(dev,
4167 "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_INVALID_REQUEST_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4168 cq->icid, qp->icid);
4169 wc_status = IB_WC_REM_INV_REQ_ERR;
4170 break;
4171 case RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR:
4172 DP_ERR(dev,
4173 "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_ACCESS_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4174 cq->icid, qp->icid);
4175 wc_status = IB_WC_REM_ACCESS_ERR;
4176 break;
4177 case RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR:
4178 DP_ERR(dev,
4179 "Error: POLL CQ with RDMA_CQE_REQ_STS_REMOTE_OPERATION_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4180 cq->icid, qp->icid);
4181 wc_status = IB_WC_REM_OP_ERR;
4182 break;
4183 case RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR:
4184 DP_ERR(dev,
4185 "Error: POLL CQ with RDMA_CQE_REQ_STS_RNR_NAK_RETRY_CNT_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4186 cq->icid, qp->icid);
4187 wc_status = IB_WC_RNR_RETRY_EXC_ERR;
4188 break;
4189 case RDMA_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR:
4190 DP_ERR(dev,
4191 "Error: POLL CQ with ROCE_CQE_REQ_STS_TRANSPORT_RETRY_CNT_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4192 cq->icid, qp->icid);
4193 wc_status = IB_WC_RETRY_EXC_ERR;
4194 break;
4195 default:
4196 DP_ERR(dev,
4197 "Error: POLL CQ with IB_WC_GENERAL_ERR. CQ icid=0x%x, QP icid=0x%x\n",
4198 cq->icid, qp->icid);
4199 wc_status = IB_WC_GENERAL_ERR;
4201 cnt += process_req(dev, qp, cq, 1, wc, req->sq_cons,
4202 wc_status, 1);
4206 return cnt;
4209 static inline int qedr_cqe_resp_status_to_ib(u8 status)
4211 switch (status) {
4212 case RDMA_CQE_RESP_STS_LOCAL_ACCESS_ERR:
4213 return IB_WC_LOC_ACCESS_ERR;
4214 case RDMA_CQE_RESP_STS_LOCAL_LENGTH_ERR:
4215 return IB_WC_LOC_LEN_ERR;
4216 case RDMA_CQE_RESP_STS_LOCAL_QP_OPERATION_ERR:
4217 return IB_WC_LOC_QP_OP_ERR;
4218 case RDMA_CQE_RESP_STS_LOCAL_PROTECTION_ERR:
4219 return IB_WC_LOC_PROT_ERR;
4220 case RDMA_CQE_RESP_STS_MEMORY_MGT_OPERATION_ERR:
4221 return IB_WC_MW_BIND_ERR;
4222 case RDMA_CQE_RESP_STS_REMOTE_INVALID_REQUEST_ERR:
4223 return IB_WC_REM_INV_RD_REQ_ERR;
4224 case RDMA_CQE_RESP_STS_OK:
4225 return IB_WC_SUCCESS;
4226 default:
4227 return IB_WC_GENERAL_ERR;
4231 static inline int qedr_set_ok_cqe_resp_wc(struct rdma_cqe_responder *resp,
4232 struct ib_wc *wc)
4234 wc->status = IB_WC_SUCCESS;
4235 wc->byte_len = le32_to_cpu(resp->length);
4237 if (resp->flags & QEDR_RESP_IMM) {
4238 wc->ex.imm_data = cpu_to_be32(le32_to_cpu(resp->imm_data_or_inv_r_Key));
4239 wc->wc_flags |= IB_WC_WITH_IMM;
4241 if (resp->flags & QEDR_RESP_RDMA)
4242 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
4244 if (resp->flags & QEDR_RESP_INV)
4245 return -EINVAL;
4247 } else if (resp->flags & QEDR_RESP_INV) {
4248 wc->ex.imm_data = le32_to_cpu(resp->imm_data_or_inv_r_Key);
4249 wc->wc_flags |= IB_WC_WITH_INVALIDATE;
4251 if (resp->flags & QEDR_RESP_RDMA)
4252 return -EINVAL;
4254 } else if (resp->flags & QEDR_RESP_RDMA) {
4255 return -EINVAL;
4258 return 0;
4261 static void __process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp,
4262 struct qedr_cq *cq, struct ib_wc *wc,
4263 struct rdma_cqe_responder *resp, u64 wr_id)
4265 /* Must fill fields before qedr_set_ok_cqe_resp_wc() */
4266 wc->opcode = IB_WC_RECV;
4267 wc->wc_flags = 0;
4269 if (likely(resp->status == RDMA_CQE_RESP_STS_OK)) {
4270 if (qedr_set_ok_cqe_resp_wc(resp, wc))
4271 DP_ERR(dev,
4272 "CQ %p (icid=%d) has invalid CQE responder flags=0x%x\n",
4273 cq, cq->icid, resp->flags);
4275 } else {
4276 wc->status = qedr_cqe_resp_status_to_ib(resp->status);
4277 if (wc->status == IB_WC_GENERAL_ERR)
4278 DP_ERR(dev,
4279 "CQ %p (icid=%d) contains an invalid CQE status %d\n",
4280 cq, cq->icid, resp->status);
4283 /* Fill the rest of the WC */
4284 wc->vendor_err = 0;
4285 wc->src_qp = qp->id;
4286 wc->qp = &qp->ibqp;
4287 wc->wr_id = wr_id;
4290 static int process_resp_one_srq(struct qedr_dev *dev, struct qedr_qp *qp,
4291 struct qedr_cq *cq, struct ib_wc *wc,
4292 struct rdma_cqe_responder *resp)
4294 struct qedr_srq *srq = qp->srq;
4295 u64 wr_id;
4297 wr_id = HILO_GEN(le32_to_cpu(resp->srq_wr_id.hi),
4298 le32_to_cpu(resp->srq_wr_id.lo), u64);
4300 if (resp->status == RDMA_CQE_RESP_STS_WORK_REQUEST_FLUSHED_ERR) {
4301 wc->status = IB_WC_WR_FLUSH_ERR;
4302 wc->vendor_err = 0;
4303 wc->wr_id = wr_id;
4304 wc->byte_len = 0;
4305 wc->src_qp = qp->id;
4306 wc->qp = &qp->ibqp;
4307 wc->wr_id = wr_id;
4308 } else {
4309 __process_resp_one(dev, qp, cq, wc, resp, wr_id);
4311 atomic_inc(&srq->hw_srq.wr_cons_cnt);
4313 return 1;
4315 static int process_resp_one(struct qedr_dev *dev, struct qedr_qp *qp,
4316 struct qedr_cq *cq, struct ib_wc *wc,
4317 struct rdma_cqe_responder *resp)
4319 u64 wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
4321 __process_resp_one(dev, qp, cq, wc, resp, wr_id);
4323 while (qp->rqe_wr_id[qp->rq.cons].wqe_size--)
4324 qed_chain_consume(&qp->rq.pbl);
4325 qedr_inc_sw_cons(&qp->rq);
4327 return 1;
4330 static int process_resp_flush(struct qedr_qp *qp, struct qedr_cq *cq,
4331 int num_entries, struct ib_wc *wc, u16 hw_cons)
4333 u16 cnt = 0;
4335 while (num_entries && qp->rq.wqe_cons != hw_cons) {
4336 /* fill WC */
4337 wc->status = IB_WC_WR_FLUSH_ERR;
4338 wc->vendor_err = 0;
4339 wc->wc_flags = 0;
4340 wc->src_qp = qp->id;
4341 wc->byte_len = 0;
4342 wc->wr_id = qp->rqe_wr_id[qp->rq.cons].wr_id;
4343 wc->qp = &qp->ibqp;
4344 num_entries--;
4345 wc++;
4346 cnt++;
4347 while (qp->rqe_wr_id[qp->rq.cons].wqe_size--)
4348 qed_chain_consume(&qp->rq.pbl);
4349 qedr_inc_sw_cons(&qp->rq);
4352 return cnt;
4355 static void try_consume_resp_cqe(struct qedr_cq *cq, struct qedr_qp *qp,
4356 struct rdma_cqe_responder *resp, int *update)
4358 if (le16_to_cpu(resp->rq_cons_or_srq_id) == qp->rq.wqe_cons) {
4359 consume_cqe(cq);
4360 *update |= 1;
4364 static int qedr_poll_cq_resp_srq(struct qedr_dev *dev, struct qedr_qp *qp,
4365 struct qedr_cq *cq, int num_entries,
4366 struct ib_wc *wc,
4367 struct rdma_cqe_responder *resp)
4369 int cnt;
4371 cnt = process_resp_one_srq(dev, qp, cq, wc, resp);
4372 consume_cqe(cq);
4374 return cnt;
4377 static int qedr_poll_cq_resp(struct qedr_dev *dev, struct qedr_qp *qp,
4378 struct qedr_cq *cq, int num_entries,
4379 struct ib_wc *wc, struct rdma_cqe_responder *resp,
4380 int *update)
4382 int cnt;
4384 if (resp->status == RDMA_CQE_RESP_STS_WORK_REQUEST_FLUSHED_ERR) {
4385 cnt = process_resp_flush(qp, cq, num_entries, wc,
4386 resp->rq_cons_or_srq_id);
4387 try_consume_resp_cqe(cq, qp, resp, update);
4388 } else {
4389 cnt = process_resp_one(dev, qp, cq, wc, resp);
4390 consume_cqe(cq);
4391 *update |= 1;
4394 return cnt;
4397 static void try_consume_req_cqe(struct qedr_cq *cq, struct qedr_qp *qp,
4398 struct rdma_cqe_requester *req, int *update)
4400 if (le16_to_cpu(req->sq_cons) == qp->sq.wqe_cons) {
4401 consume_cqe(cq);
4402 *update |= 1;
4406 int qedr_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
4408 struct qedr_dev *dev = get_qedr_dev(ibcq->device);
4409 struct qedr_cq *cq = get_qedr_cq(ibcq);
4410 union rdma_cqe *cqe;
4411 u32 old_cons, new_cons;
4412 unsigned long flags;
4413 int update = 0;
4414 int done = 0;
4416 if (cq->destroyed) {
4417 DP_ERR(dev,
4418 "warning: poll was invoked after destroy for cq %p (icid=%d)\n",
4419 cq, cq->icid);
4420 return 0;
4423 if (cq->cq_type == QEDR_CQ_TYPE_GSI)
4424 return qedr_gsi_poll_cq(ibcq, num_entries, wc);
4426 spin_lock_irqsave(&cq->cq_lock, flags);
4427 cqe = cq->latest_cqe;
4428 old_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
4429 while (num_entries && is_valid_cqe(cq, cqe)) {
4430 struct qedr_qp *qp;
4431 int cnt = 0;
4433 /* prevent speculative reads of any field of CQE */
4434 rmb();
4436 qp = cqe_get_qp(cqe);
4437 if (!qp) {
4438 WARN(1, "Error: CQE QP pointer is NULL. CQE=%p\n", cqe);
4439 break;
4442 wc->qp = &qp->ibqp;
4444 switch (cqe_get_type(cqe)) {
4445 case RDMA_CQE_TYPE_REQUESTER:
4446 cnt = qedr_poll_cq_req(dev, qp, cq, num_entries, wc,
4447 &cqe->req);
4448 try_consume_req_cqe(cq, qp, &cqe->req, &update);
4449 break;
4450 case RDMA_CQE_TYPE_RESPONDER_RQ:
4451 cnt = qedr_poll_cq_resp(dev, qp, cq, num_entries, wc,
4452 &cqe->resp, &update);
4453 break;
4454 case RDMA_CQE_TYPE_RESPONDER_SRQ:
4455 cnt = qedr_poll_cq_resp_srq(dev, qp, cq, num_entries,
4456 wc, &cqe->resp);
4457 update = 1;
4458 break;
4459 case RDMA_CQE_TYPE_INVALID:
4460 default:
4461 DP_ERR(dev, "Error: invalid CQE type = %d\n",
4462 cqe_get_type(cqe));
4464 num_entries -= cnt;
4465 wc += cnt;
4466 done += cnt;
4468 cqe = get_cqe(cq);
4470 new_cons = qed_chain_get_cons_idx_u32(&cq->pbl);
4472 cq->cq_cons += new_cons - old_cons;
4474 if (update)
4475 /* doorbell notifies abount latest VALID entry,
4476 * but chain already point to the next INVALID one
4478 doorbell_cq(cq, cq->cq_cons - 1, cq->arm_flags);
4480 spin_unlock_irqrestore(&cq->cq_lock, flags);
4481 return done;
4484 int qedr_process_mad(struct ib_device *ibdev, int process_mad_flags,
4485 u8 port_num, const struct ib_wc *in_wc,
4486 const struct ib_grh *in_grh, const struct ib_mad *in,
4487 struct ib_mad *out_mad, size_t *out_mad_size,
4488 u16 *out_mad_pkey_index)
4490 return IB_MAD_RESULT_SUCCESS;