1 | /* |
2 | * Broadcom NetXtreme-E RoCE driver. |
3 | * |
4 | * Copyright (c) 2016 - 2017, Broadcom. All rights reserved. The term |
5 | * Broadcom refers to Broadcom Limited and/or its subsidiaries. |
6 | * |
7 | * This software is available to you under a choice of one of two |
8 | * licenses. You may choose to be licensed under the terms of the GNU |
9 | * General Public License (GPL) Version 2, available from the file |
10 | * COPYING in the main directory of this source tree, or the |
11 | * BSD license below: |
12 | * |
13 | * Redistribution and use in source and binary forms, with or without |
14 | * modification, are permitted provided that the following conditions |
15 | * are met: |
16 | * |
17 | * 1. Redistributions of source code must retain the above copyright |
18 | * notice, this list of conditions and the following disclaimer. |
19 | * 2. Redistributions in binary form must reproduce the above copyright |
20 | * notice, this list of conditions and the following disclaimer in |
21 | * the documentation and/or other materials provided with the |
22 | * distribution. |
23 | * |
24 | * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' |
25 | * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, |
26 | * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR |
27 | * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS |
28 | * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR |
29 | * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
30 | * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR |
31 | * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, |
32 | * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE |
33 | * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN |
34 | * IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
35 | * |
36 | * Description: Main component of the bnxt_re driver |
37 | */ |
38 | |
39 | #include <linux/module.h> |
40 | #include <linux/netdevice.h> |
41 | #include <linux/ethtool.h> |
42 | #include <linux/mutex.h> |
43 | #include <linux/list.h> |
44 | #include <linux/rculist.h> |
45 | #include <linux/spinlock.h> |
46 | #include <linux/pci.h> |
47 | #include <net/dcbnl.h> |
48 | #include <net/ipv6.h> |
49 | #include <net/addrconf.h> |
50 | #include <linux/if_ether.h> |
51 | #include <linux/auxiliary_bus.h> |
52 | |
53 | #include <rdma/ib_verbs.h> |
54 | #include <rdma/ib_user_verbs.h> |
55 | #include <rdma/ib_umem.h> |
56 | #include <rdma/ib_addr.h> |
57 | |
58 | #include "bnxt_ulp.h" |
59 | #include "roce_hsi.h" |
60 | #include "qplib_res.h" |
61 | #include "qplib_sp.h" |
62 | #include "qplib_fp.h" |
63 | #include "qplib_rcfw.h" |
64 | #include "bnxt_re.h" |
65 | #include "ib_verbs.h" |
66 | #include <rdma/bnxt_re-abi.h> |
67 | #include "bnxt.h" |
68 | #include "hw_counters.h" |
69 | |
70 | static char version[] = |
71 | BNXT_RE_DESC "\n" ; |
72 | |
73 | MODULE_AUTHOR("Eddie Wai <eddie.wai@broadcom.com>" ); |
74 | MODULE_DESCRIPTION(BNXT_RE_DESC " Driver" ); |
75 | MODULE_LICENSE("Dual BSD/GPL" ); |
76 | |
77 | /* globals */ |
78 | static DEFINE_MUTEX(bnxt_re_mutex); |
79 | |
80 | static void bnxt_re_stop_irq(void *handle); |
81 | static void bnxt_re_dev_stop(struct bnxt_re_dev *rdev); |
82 | static int bnxt_re_netdev_event(struct notifier_block *notifier, |
83 | unsigned long event, void *ptr); |
84 | static struct bnxt_re_dev *bnxt_re_from_netdev(struct net_device *netdev); |
85 | static void bnxt_re_dev_uninit(struct bnxt_re_dev *rdev); |
86 | static int bnxt_re_hwrm_qcaps(struct bnxt_re_dev *rdev); |
87 | |
88 | static int bnxt_re_hwrm_qcfg(struct bnxt_re_dev *rdev, u32 *db_len, |
89 | u32 *offset); |
90 | static void bnxt_re_set_db_offset(struct bnxt_re_dev *rdev) |
91 | { |
92 | struct bnxt_qplib_chip_ctx *cctx; |
93 | struct bnxt_en_dev *en_dev; |
94 | struct bnxt_qplib_res *res; |
95 | u32 l2db_len = 0; |
96 | u32 offset = 0; |
97 | u32 barlen; |
98 | int rc; |
99 | |
100 | res = &rdev->qplib_res; |
101 | en_dev = rdev->en_dev; |
102 | cctx = rdev->chip_ctx; |
103 | |
104 | /* Issue qcfg */ |
105 | rc = bnxt_re_hwrm_qcfg(rdev, db_len: &l2db_len, offset: &offset); |
106 | if (rc) |
107 | dev_info(rdev_to_dev(rdev), |
108 | "Couldn't get DB bar size, Low latency framework is disabled\n" ); |
109 | /* set register offsets for both UC and WC */ |
110 | res->dpi_tbl.ucreg.offset = res->is_vf ? BNXT_QPLIB_DBR_VF_DB_OFFSET : |
111 | BNXT_QPLIB_DBR_PF_DB_OFFSET; |
112 | res->dpi_tbl.wcreg.offset = res->dpi_tbl.ucreg.offset; |
113 | |
114 | /* If WC mapping is disabled by L2 driver then en_dev->l2_db_size |
115 | * is equal to the DB-Bar actual size. This indicates that L2 |
116 | * is mapping entire bar as UC-. RoCE driver can't enable WC mapping |
117 | * in such cases and DB-push will be disabled. |
118 | */ |
119 | barlen = pci_resource_len(res->pdev, RCFW_DBR_PCI_BAR_REGION); |
120 | if (cctx->modes.db_push && l2db_len && en_dev->l2_db_size != barlen) { |
121 | res->dpi_tbl.wcreg.offset = en_dev->l2_db_size; |
122 | dev_info(rdev_to_dev(rdev), "Low latency framework is enabled\n" ); |
123 | } |
124 | } |
125 | |
126 | static void bnxt_re_set_drv_mode(struct bnxt_re_dev *rdev, u8 mode) |
127 | { |
128 | struct bnxt_qplib_chip_ctx *cctx; |
129 | |
130 | cctx = rdev->chip_ctx; |
131 | cctx->modes.wqe_mode = bnxt_qplib_is_chip_gen_p5(cctx: rdev->chip_ctx) ? |
132 | mode : BNXT_QPLIB_WQE_MODE_STATIC; |
133 | if (bnxt_re_hwrm_qcaps(rdev)) |
134 | dev_err(rdev_to_dev(rdev), |
135 | "Failed to query hwrm qcaps\n" ); |
136 | } |
137 | |
138 | static void bnxt_re_destroy_chip_ctx(struct bnxt_re_dev *rdev) |
139 | { |
140 | struct bnxt_qplib_chip_ctx *chip_ctx; |
141 | |
142 | if (!rdev->chip_ctx) |
143 | return; |
144 | chip_ctx = rdev->chip_ctx; |
145 | rdev->chip_ctx = NULL; |
146 | rdev->rcfw.res = NULL; |
147 | rdev->qplib_res.cctx = NULL; |
148 | rdev->qplib_res.pdev = NULL; |
149 | rdev->qplib_res.netdev = NULL; |
150 | kfree(objp: chip_ctx); |
151 | } |
152 | |
153 | static int bnxt_re_setup_chip_ctx(struct bnxt_re_dev *rdev, u8 wqe_mode) |
154 | { |
155 | struct bnxt_qplib_chip_ctx *chip_ctx; |
156 | struct bnxt_en_dev *en_dev; |
157 | int rc; |
158 | |
159 | en_dev = rdev->en_dev; |
160 | |
161 | chip_ctx = kzalloc(size: sizeof(*chip_ctx), GFP_KERNEL); |
162 | if (!chip_ctx) |
163 | return -ENOMEM; |
164 | chip_ctx->chip_num = en_dev->chip_num; |
165 | chip_ctx->hw_stats_size = en_dev->hw_ring_stats_size; |
166 | |
167 | rdev->chip_ctx = chip_ctx; |
168 | /* rest members to follow eventually */ |
169 | |
170 | rdev->qplib_res.cctx = rdev->chip_ctx; |
171 | rdev->rcfw.res = &rdev->qplib_res; |
172 | rdev->qplib_res.dattr = &rdev->dev_attr; |
173 | rdev->qplib_res.is_vf = BNXT_EN_VF(en_dev); |
174 | |
175 | bnxt_re_set_drv_mode(rdev, mode: wqe_mode); |
176 | |
177 | bnxt_re_set_db_offset(rdev); |
178 | rc = bnxt_qplib_map_db_bar(res: &rdev->qplib_res); |
179 | if (rc) |
180 | return rc; |
181 | |
182 | if (bnxt_qplib_determine_atomics(dev: en_dev->pdev)) |
183 | ibdev_info(ibdev: &rdev->ibdev, |
184 | format: "platform doesn't support global atomics." ); |
185 | return 0; |
186 | } |
187 | |
188 | /* SR-IOV helper functions */ |
189 | |
190 | static void bnxt_re_get_sriov_func_type(struct bnxt_re_dev *rdev) |
191 | { |
192 | if (BNXT_EN_VF(rdev->en_dev)) |
193 | rdev->is_virtfn = 1; |
194 | } |
195 | |
196 | /* Set the maximum number of each resource that the driver actually wants |
197 | * to allocate. This may be up to the maximum number the firmware has |
198 | * reserved for the function. The driver may choose to allocate fewer |
199 | * resources than the firmware maximum. |
200 | */ |
201 | static void bnxt_re_limit_pf_res(struct bnxt_re_dev *rdev) |
202 | { |
203 | struct bnxt_qplib_dev_attr *attr; |
204 | struct bnxt_qplib_ctx *ctx; |
205 | int i; |
206 | |
207 | attr = &rdev->dev_attr; |
208 | ctx = &rdev->qplib_ctx; |
209 | |
210 | ctx->qpc_count = min_t(u32, BNXT_RE_MAX_QPC_COUNT, |
211 | attr->max_qp); |
212 | ctx->mrw_count = BNXT_RE_MAX_MRW_COUNT_256K; |
213 | /* Use max_mr from fw since max_mrw does not get set */ |
214 | ctx->mrw_count = min_t(u32, ctx->mrw_count, attr->max_mr); |
215 | ctx->srqc_count = min_t(u32, BNXT_RE_MAX_SRQC_COUNT, |
216 | attr->max_srq); |
217 | ctx->cq_count = min_t(u32, BNXT_RE_MAX_CQ_COUNT, attr->max_cq); |
218 | if (!bnxt_qplib_is_chip_gen_p5(cctx: rdev->chip_ctx)) |
219 | for (i = 0; i < MAX_TQM_ALLOC_REQ; i++) |
220 | rdev->qplib_ctx.tqm_ctx.qcount[i] = |
221 | rdev->dev_attr.tqm_alloc_reqs[i]; |
222 | } |
223 | |
224 | static void bnxt_re_limit_vf_res(struct bnxt_qplib_ctx *qplib_ctx, u32 num_vf) |
225 | { |
226 | struct bnxt_qplib_vf_res *vf_res; |
227 | u32 mrws = 0; |
228 | u32 vf_pct; |
229 | u32 nvfs; |
230 | |
231 | vf_res = &qplib_ctx->vf_res; |
232 | /* |
233 | * Reserve a set of resources for the PF. Divide the remaining |
234 | * resources among the VFs |
235 | */ |
236 | vf_pct = 100 - BNXT_RE_PCT_RSVD_FOR_PF; |
237 | nvfs = num_vf; |
238 | num_vf = 100 * num_vf; |
239 | vf_res->max_qp_per_vf = (qplib_ctx->qpc_count * vf_pct) / num_vf; |
240 | vf_res->max_srq_per_vf = (qplib_ctx->srqc_count * vf_pct) / num_vf; |
241 | vf_res->max_cq_per_vf = (qplib_ctx->cq_count * vf_pct) / num_vf; |
242 | /* |
243 | * The driver allows many more MRs than other resources. If the |
244 | * firmware does also, then reserve a fixed amount for the PF and |
245 | * divide the rest among VFs. VFs may use many MRs for NFS |
246 | * mounts, ISER, NVME applications, etc. If the firmware severely |
247 | * restricts the number of MRs, then let PF have half and divide |
248 | * the rest among VFs, as for the other resource types. |
249 | */ |
250 | if (qplib_ctx->mrw_count < BNXT_RE_MAX_MRW_COUNT_64K) { |
251 | mrws = qplib_ctx->mrw_count * vf_pct; |
252 | nvfs = num_vf; |
253 | } else { |
254 | mrws = qplib_ctx->mrw_count - BNXT_RE_RESVD_MR_FOR_PF; |
255 | } |
256 | vf_res->max_mrw_per_vf = (mrws / nvfs); |
257 | vf_res->max_gid_per_vf = BNXT_RE_MAX_GID_PER_VF; |
258 | } |
259 | |
260 | static void bnxt_re_set_resource_limits(struct bnxt_re_dev *rdev) |
261 | { |
262 | u32 num_vfs; |
263 | |
264 | memset(&rdev->qplib_ctx.vf_res, 0, sizeof(struct bnxt_qplib_vf_res)); |
265 | bnxt_re_limit_pf_res(rdev); |
266 | |
267 | num_vfs = bnxt_qplib_is_chip_gen_p5(cctx: rdev->chip_ctx) ? |
268 | BNXT_RE_GEN_P5_MAX_VF : rdev->num_vfs; |
269 | if (num_vfs) |
270 | bnxt_re_limit_vf_res(qplib_ctx: &rdev->qplib_ctx, num_vf: num_vfs); |
271 | } |
272 | |
273 | static void bnxt_re_vf_res_config(struct bnxt_re_dev *rdev) |
274 | { |
275 | |
276 | if (test_bit(BNXT_RE_FLAG_ERR_DEVICE_DETACHED, &rdev->flags)) |
277 | return; |
278 | rdev->num_vfs = pci_sriov_get_totalvfs(dev: rdev->en_dev->pdev); |
279 | if (!bnxt_qplib_is_chip_gen_p5(cctx: rdev->chip_ctx)) { |
280 | bnxt_re_set_resource_limits(rdev); |
281 | bnxt_qplib_set_func_resources(res: &rdev->qplib_res, rcfw: &rdev->rcfw, |
282 | ctx: &rdev->qplib_ctx); |
283 | } |
284 | } |
285 | |
286 | static void bnxt_re_shutdown(struct auxiliary_device *adev) |
287 | { |
288 | struct bnxt_re_dev *rdev = auxiliary_get_drvdata(auxdev: adev); |
289 | |
290 | if (!rdev) |
291 | return; |
292 | ib_unregister_device(device: &rdev->ibdev); |
293 | bnxt_re_dev_uninit(rdev); |
294 | } |
295 | |
296 | static void bnxt_re_stop_irq(void *handle) |
297 | { |
298 | struct bnxt_re_dev *rdev = (struct bnxt_re_dev *)handle; |
299 | struct bnxt_qplib_rcfw *rcfw = &rdev->rcfw; |
300 | struct bnxt_qplib_nq *nq; |
301 | int indx; |
302 | |
303 | for (indx = BNXT_RE_NQ_IDX; indx < rdev->num_msix; indx++) { |
304 | nq = &rdev->nq[indx - 1]; |
305 | bnxt_qplib_nq_stop_irq(nq, kill: false); |
306 | } |
307 | |
308 | bnxt_qplib_rcfw_stop_irq(rcfw, kill: false); |
309 | } |
310 | |
311 | static void bnxt_re_start_irq(void *handle, struct bnxt_msix_entry *ent) |
312 | { |
313 | struct bnxt_re_dev *rdev = (struct bnxt_re_dev *)handle; |
314 | struct bnxt_msix_entry *msix_ent = rdev->en_dev->msix_entries; |
315 | struct bnxt_qplib_rcfw *rcfw = &rdev->rcfw; |
316 | struct bnxt_qplib_nq *nq; |
317 | int indx, rc; |
318 | |
319 | if (!ent) { |
320 | /* Not setting the f/w timeout bit in rcfw. |
321 | * During the driver unload the first command |
322 | * to f/w will timeout and that will set the |
323 | * timeout bit. |
324 | */ |
325 | ibdev_err(ibdev: &rdev->ibdev, format: "Failed to re-start IRQs\n" ); |
326 | return; |
327 | } |
328 | |
329 | /* Vectors may change after restart, so update with new vectors |
330 | * in device sctructure. |
331 | */ |
332 | for (indx = 0; indx < rdev->num_msix; indx++) |
333 | rdev->en_dev->msix_entries[indx].vector = ent[indx].vector; |
334 | |
335 | rc = bnxt_qplib_rcfw_start_irq(rcfw, msix_ent[BNXT_RE_AEQ_IDX].vector, |
336 | false); |
337 | if (rc) { |
338 | ibdev_warn(ibdev: &rdev->ibdev, format: "Failed to reinit CREQ\n" ); |
339 | return; |
340 | } |
341 | for (indx = BNXT_RE_NQ_IDX ; indx < rdev->num_msix; indx++) { |
342 | nq = &rdev->nq[indx - 1]; |
343 | rc = bnxt_qplib_nq_start_irq(nq, indx - 1, |
344 | msix_ent[indx].vector, false); |
345 | if (rc) { |
346 | ibdev_warn(ibdev: &rdev->ibdev, format: "Failed to reinit NQ index %d\n" , |
347 | indx - 1); |
348 | return; |
349 | } |
350 | } |
351 | } |
352 | |
353 | static struct bnxt_ulp_ops bnxt_re_ulp_ops = { |
354 | .ulp_irq_stop = bnxt_re_stop_irq, |
355 | .ulp_irq_restart = bnxt_re_start_irq |
356 | }; |
357 | |
358 | /* RoCE -> Net driver */ |
359 | |
360 | static int bnxt_re_register_netdev(struct bnxt_re_dev *rdev) |
361 | { |
362 | struct bnxt_en_dev *en_dev; |
363 | int rc; |
364 | |
365 | en_dev = rdev->en_dev; |
366 | |
367 | rc = bnxt_register_dev(en_dev, &bnxt_re_ulp_ops, rdev); |
368 | if (!rc) |
369 | rdev->qplib_res.pdev = rdev->en_dev->pdev; |
370 | return rc; |
371 | } |
372 | |
373 | static void bnxt_re_init_hwrm_hdr(struct input *hdr, u16 opcd) |
374 | { |
375 | hdr->req_type = cpu_to_le16(opcd); |
376 | hdr->cmpl_ring = cpu_to_le16(-1); |
377 | hdr->target_id = cpu_to_le16(-1); |
378 | } |
379 | |
380 | static void bnxt_re_fill_fw_msg(struct bnxt_fw_msg *fw_msg, void *msg, |
381 | int msg_len, void *resp, int resp_max_len, |
382 | int timeout) |
383 | { |
384 | fw_msg->msg = msg; |
385 | fw_msg->msg_len = msg_len; |
386 | fw_msg->resp = resp; |
387 | fw_msg->resp_max_len = resp_max_len; |
388 | fw_msg->timeout = timeout; |
389 | } |
390 | |
391 | /* Query device config using common hwrm */ |
392 | static int bnxt_re_hwrm_qcfg(struct bnxt_re_dev *rdev, u32 *db_len, |
393 | u32 *offset) |
394 | { |
395 | struct bnxt_en_dev *en_dev = rdev->en_dev; |
396 | struct hwrm_func_qcfg_output resp = {0}; |
397 | struct hwrm_func_qcfg_input req = {0}; |
398 | struct bnxt_fw_msg fw_msg = {}; |
399 | int rc; |
400 | |
401 | bnxt_re_init_hwrm_hdr(hdr: (void *)&req, opcd: HWRM_FUNC_QCFG); |
402 | req.fid = cpu_to_le16(0xffff); |
403 | bnxt_re_fill_fw_msg(fw_msg: &fw_msg, msg: (void *)&req, msg_len: sizeof(req), resp: (void *)&resp, |
404 | resp_max_len: sizeof(resp), timeout: DFLT_HWRM_CMD_TIMEOUT); |
405 | rc = bnxt_send_msg(en_dev, &fw_msg); |
406 | if (!rc) { |
407 | *db_len = PAGE_ALIGN(le16_to_cpu(resp.l2_doorbell_bar_size_kb) * 1024); |
408 | *offset = PAGE_ALIGN(le16_to_cpu(resp.legacy_l2_db_size_kb) * 1024); |
409 | } |
410 | return rc; |
411 | } |
412 | |
413 | /* Query function capabilities using common hwrm */ |
414 | int bnxt_re_hwrm_qcaps(struct bnxt_re_dev *rdev) |
415 | { |
416 | struct bnxt_en_dev *en_dev = rdev->en_dev; |
417 | struct hwrm_func_qcaps_output resp = {}; |
418 | struct hwrm_func_qcaps_input req = {}; |
419 | struct bnxt_qplib_chip_ctx *cctx; |
420 | struct bnxt_fw_msg fw_msg = {}; |
421 | int rc; |
422 | |
423 | cctx = rdev->chip_ctx; |
424 | bnxt_re_init_hwrm_hdr(hdr: (void *)&req, opcd: HWRM_FUNC_QCAPS); |
425 | req.fid = cpu_to_le16(0xffff); |
426 | bnxt_re_fill_fw_msg(fw_msg: &fw_msg, msg: (void *)&req, msg_len: sizeof(req), resp: (void *)&resp, |
427 | resp_max_len: sizeof(resp), timeout: DFLT_HWRM_CMD_TIMEOUT); |
428 | |
429 | rc = bnxt_send_msg(en_dev, &fw_msg); |
430 | if (rc) |
431 | return rc; |
432 | cctx->modes.db_push = le32_to_cpu(resp.flags) & FUNC_QCAPS_RESP_FLAGS_WCB_PUSH_MODE; |
433 | |
434 | cctx->modes.dbr_pacing = |
435 | le32_to_cpu(resp.flags_ext2) & |
436 | FUNC_QCAPS_RESP_FLAGS_EXT2_DBR_PACING_EXT_SUPPORTED; |
437 | return 0; |
438 | } |
439 | |
440 | static int bnxt_re_hwrm_dbr_pacing_qcfg(struct bnxt_re_dev *rdev) |
441 | { |
442 | struct hwrm_func_dbr_pacing_qcfg_output resp = {}; |
443 | struct hwrm_func_dbr_pacing_qcfg_input req = {}; |
444 | struct bnxt_en_dev *en_dev = rdev->en_dev; |
445 | struct bnxt_qplib_chip_ctx *cctx; |
446 | struct bnxt_fw_msg fw_msg = {}; |
447 | int rc; |
448 | |
449 | cctx = rdev->chip_ctx; |
450 | bnxt_re_init_hwrm_hdr(hdr: (void *)&req, opcd: HWRM_FUNC_DBR_PACING_QCFG); |
451 | bnxt_re_fill_fw_msg(fw_msg: &fw_msg, msg: (void *)&req, msg_len: sizeof(req), resp: (void *)&resp, |
452 | resp_max_len: sizeof(resp), timeout: DFLT_HWRM_CMD_TIMEOUT); |
453 | rc = bnxt_send_msg(en_dev, &fw_msg); |
454 | if (rc) |
455 | return rc; |
456 | |
457 | if ((le32_to_cpu(resp.dbr_stat_db_fifo_reg) & |
458 | FUNC_DBR_PACING_QCFG_RESP_DBR_STAT_DB_FIFO_REG_ADDR_SPACE_MASK) == |
459 | FUNC_DBR_PACING_QCFG_RESP_DBR_STAT_DB_FIFO_REG_ADDR_SPACE_GRC) |
460 | cctx->dbr_stat_db_fifo = |
461 | le32_to_cpu(resp.dbr_stat_db_fifo_reg) & |
462 | ~FUNC_DBR_PACING_QCFG_RESP_DBR_STAT_DB_FIFO_REG_ADDR_SPACE_MASK; |
463 | return 0; |
464 | } |
465 | |
466 | /* Update the pacing tunable parameters to the default values */ |
467 | static void bnxt_re_set_default_pacing_data(struct bnxt_re_dev *rdev) |
468 | { |
469 | struct bnxt_qplib_db_pacing_data *pacing_data = rdev->qplib_res.pacing_data; |
470 | |
471 | pacing_data->do_pacing = rdev->pacing.dbr_def_do_pacing; |
472 | pacing_data->pacing_th = rdev->pacing.pacing_algo_th; |
473 | pacing_data->alarm_th = |
474 | pacing_data->pacing_th * BNXT_RE_PACING_ALARM_TH_MULTIPLE; |
475 | } |
476 | |
477 | static void __wait_for_fifo_occupancy_below_th(struct bnxt_re_dev *rdev) |
478 | { |
479 | u32 read_val, fifo_occup; |
480 | |
481 | /* loop shouldn't run infintely as the occupancy usually goes |
482 | * below pacing algo threshold as soon as pacing kicks in. |
483 | */ |
484 | while (1) { |
485 | read_val = readl(addr: rdev->en_dev->bar0 + rdev->pacing.dbr_db_fifo_reg_off); |
486 | fifo_occup = BNXT_RE_MAX_FIFO_DEPTH - |
487 | ((read_val & BNXT_RE_DB_FIFO_ROOM_MASK) >> |
488 | BNXT_RE_DB_FIFO_ROOM_SHIFT); |
489 | /* Fifo occupancy cannot be greater the MAX FIFO depth */ |
490 | if (fifo_occup > BNXT_RE_MAX_FIFO_DEPTH) |
491 | break; |
492 | |
493 | if (fifo_occup < rdev->qplib_res.pacing_data->pacing_th) |
494 | break; |
495 | } |
496 | } |
497 | |
498 | static void bnxt_re_db_fifo_check(struct work_struct *work) |
499 | { |
500 | struct bnxt_re_dev *rdev = container_of(work, struct bnxt_re_dev, |
501 | dbq_fifo_check_work); |
502 | struct bnxt_qplib_db_pacing_data *pacing_data; |
503 | u32 pacing_save; |
504 | |
505 | if (!mutex_trylock(lock: &rdev->pacing.dbq_lock)) |
506 | return; |
507 | pacing_data = rdev->qplib_res.pacing_data; |
508 | pacing_save = rdev->pacing.do_pacing_save; |
509 | __wait_for_fifo_occupancy_below_th(rdev); |
510 | cancel_delayed_work_sync(dwork: &rdev->dbq_pacing_work); |
511 | if (pacing_save > rdev->pacing.dbr_def_do_pacing) { |
512 | /* Double the do_pacing value during the congestion */ |
513 | pacing_save = pacing_save << 1; |
514 | } else { |
515 | /* |
516 | * when a new congestion is detected increase the do_pacing |
517 | * by 8 times. And also increase the pacing_th by 4 times. The |
518 | * reason to increase pacing_th is to give more space for the |
519 | * queue to oscillate down without getting empty, but also more |
520 | * room for the queue to increase without causing another alarm. |
521 | */ |
522 | pacing_save = pacing_save << 3; |
523 | pacing_data->pacing_th = rdev->pacing.pacing_algo_th * 4; |
524 | } |
525 | |
526 | if (pacing_save > BNXT_RE_MAX_DBR_DO_PACING) |
527 | pacing_save = BNXT_RE_MAX_DBR_DO_PACING; |
528 | |
529 | pacing_data->do_pacing = pacing_save; |
530 | rdev->pacing.do_pacing_save = pacing_data->do_pacing; |
531 | pacing_data->alarm_th = |
532 | pacing_data->pacing_th * BNXT_RE_PACING_ALARM_TH_MULTIPLE; |
533 | schedule_delayed_work(dwork: &rdev->dbq_pacing_work, |
534 | delay: msecs_to_jiffies(m: rdev->pacing.dbq_pacing_time)); |
535 | rdev->stats.pacing.alerts++; |
536 | mutex_unlock(lock: &rdev->pacing.dbq_lock); |
537 | } |
538 | |
539 | static void bnxt_re_pacing_timer_exp(struct work_struct *work) |
540 | { |
541 | struct bnxt_re_dev *rdev = container_of(work, struct bnxt_re_dev, |
542 | dbq_pacing_work.work); |
543 | struct bnxt_qplib_db_pacing_data *pacing_data; |
544 | u32 read_val, fifo_occup; |
545 | |
546 | if (!mutex_trylock(lock: &rdev->pacing.dbq_lock)) |
547 | return; |
548 | |
549 | pacing_data = rdev->qplib_res.pacing_data; |
550 | read_val = readl(addr: rdev->en_dev->bar0 + rdev->pacing.dbr_db_fifo_reg_off); |
551 | fifo_occup = BNXT_RE_MAX_FIFO_DEPTH - |
552 | ((read_val & BNXT_RE_DB_FIFO_ROOM_MASK) >> |
553 | BNXT_RE_DB_FIFO_ROOM_SHIFT); |
554 | |
555 | if (fifo_occup > pacing_data->pacing_th) |
556 | goto restart_timer; |
557 | |
558 | /* |
559 | * Instead of immediately going back to the default do_pacing |
560 | * reduce it by 1/8 times and restart the timer. |
561 | */ |
562 | pacing_data->do_pacing = pacing_data->do_pacing - (pacing_data->do_pacing >> 3); |
563 | pacing_data->do_pacing = max_t(u32, rdev->pacing.dbr_def_do_pacing, pacing_data->do_pacing); |
564 | if (pacing_data->do_pacing <= rdev->pacing.dbr_def_do_pacing) { |
565 | bnxt_re_set_default_pacing_data(rdev); |
566 | rdev->stats.pacing.complete++; |
567 | goto dbq_unlock; |
568 | } |
569 | |
570 | restart_timer: |
571 | schedule_delayed_work(dwork: &rdev->dbq_pacing_work, |
572 | delay: msecs_to_jiffies(m: rdev->pacing.dbq_pacing_time)); |
573 | rdev->stats.pacing.resched++; |
574 | dbq_unlock: |
575 | rdev->pacing.do_pacing_save = pacing_data->do_pacing; |
576 | mutex_unlock(lock: &rdev->pacing.dbq_lock); |
577 | } |
578 | |
579 | void bnxt_re_pacing_alert(struct bnxt_re_dev *rdev) |
580 | { |
581 | struct bnxt_qplib_db_pacing_data *pacing_data; |
582 | |
583 | if (!rdev->pacing.dbr_pacing) |
584 | return; |
585 | mutex_lock(&rdev->pacing.dbq_lock); |
586 | pacing_data = rdev->qplib_res.pacing_data; |
587 | |
588 | /* |
589 | * Increase the alarm_th to max so that other user lib instances do not |
590 | * keep alerting the driver. |
591 | */ |
592 | pacing_data->alarm_th = BNXT_RE_MAX_FIFO_DEPTH; |
593 | pacing_data->do_pacing = BNXT_RE_MAX_DBR_DO_PACING; |
594 | cancel_work_sync(work: &rdev->dbq_fifo_check_work); |
595 | schedule_work(work: &rdev->dbq_fifo_check_work); |
596 | mutex_unlock(lock: &rdev->pacing.dbq_lock); |
597 | } |
598 | |
599 | static int bnxt_re_initialize_dbr_pacing(struct bnxt_re_dev *rdev) |
600 | { |
601 | if (bnxt_re_hwrm_dbr_pacing_qcfg(rdev)) |
602 | return -EIO; |
603 | |
604 | /* Allocate a page for app use */ |
605 | rdev->pacing.dbr_page = (void *)__get_free_page(GFP_KERNEL); |
606 | if (!rdev->pacing.dbr_page) |
607 | return -ENOMEM; |
608 | |
609 | memset((u8 *)rdev->pacing.dbr_page, 0, PAGE_SIZE); |
610 | rdev->qplib_res.pacing_data = (struct bnxt_qplib_db_pacing_data *)rdev->pacing.dbr_page; |
611 | |
612 | /* MAP HW window 2 for reading db fifo depth */ |
613 | writel(val: rdev->chip_ctx->dbr_stat_db_fifo & BNXT_GRC_BASE_MASK, |
614 | addr: rdev->en_dev->bar0 + BNXT_GRCPF_REG_WINDOW_BASE_OUT + 4); |
615 | rdev->pacing.dbr_db_fifo_reg_off = |
616 | (rdev->chip_ctx->dbr_stat_db_fifo & BNXT_GRC_OFFSET_MASK) + |
617 | BNXT_RE_GRC_FIFO_REG_BASE; |
618 | rdev->pacing.dbr_bar_addr = |
619 | pci_resource_start(rdev->qplib_res.pdev, 0) + rdev->pacing.dbr_db_fifo_reg_off; |
620 | |
621 | rdev->pacing.pacing_algo_th = BNXT_RE_PACING_ALGO_THRESHOLD; |
622 | rdev->pacing.dbq_pacing_time = BNXT_RE_DBR_PACING_TIME; |
623 | rdev->pacing.dbr_def_do_pacing = BNXT_RE_DBR_DO_PACING_NO_CONGESTION; |
624 | rdev->pacing.do_pacing_save = rdev->pacing.dbr_def_do_pacing; |
625 | rdev->qplib_res.pacing_data->fifo_max_depth = BNXT_RE_MAX_FIFO_DEPTH; |
626 | rdev->qplib_res.pacing_data->fifo_room_mask = BNXT_RE_DB_FIFO_ROOM_MASK; |
627 | rdev->qplib_res.pacing_data->fifo_room_shift = BNXT_RE_DB_FIFO_ROOM_SHIFT; |
628 | rdev->qplib_res.pacing_data->grc_reg_offset = rdev->pacing.dbr_db_fifo_reg_off; |
629 | bnxt_re_set_default_pacing_data(rdev); |
630 | /* Initialize worker for DBR Pacing */ |
631 | INIT_WORK(&rdev->dbq_fifo_check_work, bnxt_re_db_fifo_check); |
632 | INIT_DELAYED_WORK(&rdev->dbq_pacing_work, bnxt_re_pacing_timer_exp); |
633 | return 0; |
634 | } |
635 | |
636 | static void bnxt_re_deinitialize_dbr_pacing(struct bnxt_re_dev *rdev) |
637 | { |
638 | cancel_work_sync(work: &rdev->dbq_fifo_check_work); |
639 | cancel_delayed_work_sync(dwork: &rdev->dbq_pacing_work); |
640 | if (rdev->pacing.dbr_page) |
641 | free_page((u64)rdev->pacing.dbr_page); |
642 | |
643 | rdev->pacing.dbr_page = NULL; |
644 | rdev->pacing.dbr_pacing = false; |
645 | } |
646 | |
647 | static int bnxt_re_net_ring_free(struct bnxt_re_dev *rdev, |
648 | u16 fw_ring_id, int type) |
649 | { |
650 | struct bnxt_en_dev *en_dev; |
651 | struct hwrm_ring_free_input req = {}; |
652 | struct hwrm_ring_free_output resp; |
653 | struct bnxt_fw_msg fw_msg = {}; |
654 | int rc = -EINVAL; |
655 | |
656 | if (!rdev) |
657 | return rc; |
658 | |
659 | en_dev = rdev->en_dev; |
660 | |
661 | if (!en_dev) |
662 | return rc; |
663 | |
664 | if (test_bit(BNXT_RE_FLAG_ERR_DEVICE_DETACHED, &rdev->flags)) |
665 | return 0; |
666 | |
667 | bnxt_re_init_hwrm_hdr(hdr: (void *)&req, opcd: HWRM_RING_FREE); |
668 | req.ring_type = type; |
669 | req.ring_id = cpu_to_le16(fw_ring_id); |
670 | bnxt_re_fill_fw_msg(fw_msg: &fw_msg, msg: (void *)&req, msg_len: sizeof(req), resp: (void *)&resp, |
671 | resp_max_len: sizeof(resp), timeout: DFLT_HWRM_CMD_TIMEOUT); |
672 | rc = bnxt_send_msg(en_dev, &fw_msg); |
673 | if (rc) |
674 | ibdev_err(ibdev: &rdev->ibdev, format: "Failed to free HW ring:%d :%#x" , |
675 | req.ring_id, rc); |
676 | return rc; |
677 | } |
678 | |
679 | static int bnxt_re_net_ring_alloc(struct bnxt_re_dev *rdev, |
680 | struct bnxt_re_ring_attr *ring_attr, |
681 | u16 *fw_ring_id) |
682 | { |
683 | struct bnxt_en_dev *en_dev = rdev->en_dev; |
684 | struct hwrm_ring_alloc_input req = {}; |
685 | struct hwrm_ring_alloc_output resp; |
686 | struct bnxt_fw_msg fw_msg = {}; |
687 | int rc = -EINVAL; |
688 | |
689 | if (!en_dev) |
690 | return rc; |
691 | |
692 | bnxt_re_init_hwrm_hdr(hdr: (void *)&req, opcd: HWRM_RING_ALLOC); |
693 | req.enables = 0; |
694 | req.page_tbl_addr = cpu_to_le64(ring_attr->dma_arr[0]); |
695 | if (ring_attr->pages > 1) { |
696 | /* Page size is in log2 units */ |
697 | req.page_size = BNXT_PAGE_SHIFT; |
698 | req.page_tbl_depth = 1; |
699 | } |
700 | req.fbo = 0; |
701 | /* Association of ring index with doorbell index and MSIX number */ |
702 | req.logical_id = cpu_to_le16(ring_attr->lrid); |
703 | req.length = cpu_to_le32(ring_attr->depth + 1); |
704 | req.ring_type = ring_attr->type; |
705 | req.int_mode = ring_attr->mode; |
706 | bnxt_re_fill_fw_msg(fw_msg: &fw_msg, msg: (void *)&req, msg_len: sizeof(req), resp: (void *)&resp, |
707 | resp_max_len: sizeof(resp), timeout: DFLT_HWRM_CMD_TIMEOUT); |
708 | rc = bnxt_send_msg(en_dev, &fw_msg); |
709 | if (!rc) |
710 | *fw_ring_id = le16_to_cpu(resp.ring_id); |
711 | |
712 | return rc; |
713 | } |
714 | |
715 | static int bnxt_re_net_stats_ctx_free(struct bnxt_re_dev *rdev, |
716 | u32 fw_stats_ctx_id) |
717 | { |
718 | struct bnxt_en_dev *en_dev = rdev->en_dev; |
719 | struct hwrm_stat_ctx_free_input req = {}; |
720 | struct hwrm_stat_ctx_free_output resp = {}; |
721 | struct bnxt_fw_msg fw_msg = {}; |
722 | int rc = -EINVAL; |
723 | |
724 | if (!en_dev) |
725 | return rc; |
726 | |
727 | if (test_bit(BNXT_RE_FLAG_ERR_DEVICE_DETACHED, &rdev->flags)) |
728 | return 0; |
729 | |
730 | bnxt_re_init_hwrm_hdr(hdr: (void *)&req, opcd: HWRM_STAT_CTX_FREE); |
731 | req.stat_ctx_id = cpu_to_le32(fw_stats_ctx_id); |
732 | bnxt_re_fill_fw_msg(fw_msg: &fw_msg, msg: (void *)&req, msg_len: sizeof(req), resp: (void *)&resp, |
733 | resp_max_len: sizeof(resp), timeout: DFLT_HWRM_CMD_TIMEOUT); |
734 | rc = bnxt_send_msg(en_dev, &fw_msg); |
735 | if (rc) |
736 | ibdev_err(ibdev: &rdev->ibdev, format: "Failed to free HW stats context %#x" , |
737 | rc); |
738 | |
739 | return rc; |
740 | } |
741 | |
742 | static int bnxt_re_net_stats_ctx_alloc(struct bnxt_re_dev *rdev, |
743 | dma_addr_t dma_map, |
744 | u32 *fw_stats_ctx_id) |
745 | { |
746 | struct bnxt_qplib_chip_ctx *chip_ctx = rdev->chip_ctx; |
747 | struct hwrm_stat_ctx_alloc_output resp = {}; |
748 | struct hwrm_stat_ctx_alloc_input req = {}; |
749 | struct bnxt_en_dev *en_dev = rdev->en_dev; |
750 | struct bnxt_fw_msg fw_msg = {}; |
751 | int rc = -EINVAL; |
752 | |
753 | *fw_stats_ctx_id = INVALID_STATS_CTX_ID; |
754 | |
755 | if (!en_dev) |
756 | return rc; |
757 | |
758 | bnxt_re_init_hwrm_hdr(hdr: (void *)&req, opcd: HWRM_STAT_CTX_ALLOC); |
759 | req.update_period_ms = cpu_to_le32(1000); |
760 | req.stats_dma_addr = cpu_to_le64(dma_map); |
761 | req.stats_dma_length = cpu_to_le16(chip_ctx->hw_stats_size); |
762 | req.stat_ctx_flags = STAT_CTX_ALLOC_REQ_STAT_CTX_FLAGS_ROCE; |
763 | bnxt_re_fill_fw_msg(fw_msg: &fw_msg, msg: (void *)&req, msg_len: sizeof(req), resp: (void *)&resp, |
764 | resp_max_len: sizeof(resp), timeout: DFLT_HWRM_CMD_TIMEOUT); |
765 | rc = bnxt_send_msg(en_dev, &fw_msg); |
766 | if (!rc) |
767 | *fw_stats_ctx_id = le32_to_cpu(resp.stat_ctx_id); |
768 | |
769 | return rc; |
770 | } |
771 | |
772 | static void bnxt_re_disassociate_ucontext(struct ib_ucontext *ibcontext) |
773 | { |
774 | } |
775 | |
776 | /* Device */ |
777 | |
778 | static struct bnxt_re_dev *bnxt_re_from_netdev(struct net_device *netdev) |
779 | { |
780 | struct ib_device *ibdev = |
781 | ib_device_get_by_netdev(ndev: netdev, driver_id: RDMA_DRIVER_BNXT_RE); |
782 | if (!ibdev) |
783 | return NULL; |
784 | |
785 | return container_of(ibdev, struct bnxt_re_dev, ibdev); |
786 | } |
787 | |
788 | static ssize_t hw_rev_show(struct device *device, struct device_attribute *attr, |
789 | char *buf) |
790 | { |
791 | struct bnxt_re_dev *rdev = |
792 | rdma_device_to_drv_device(device, struct bnxt_re_dev, ibdev); |
793 | |
794 | return sysfs_emit(buf, fmt: "0x%x\n" , rdev->en_dev->pdev->vendor); |
795 | } |
796 | static DEVICE_ATTR_RO(hw_rev); |
797 | |
798 | static ssize_t hca_type_show(struct device *device, |
799 | struct device_attribute *attr, char *buf) |
800 | { |
801 | struct bnxt_re_dev *rdev = |
802 | rdma_device_to_drv_device(device, struct bnxt_re_dev, ibdev); |
803 | |
804 | return sysfs_emit(buf, fmt: "%s\n" , rdev->ibdev.node_desc); |
805 | } |
806 | static DEVICE_ATTR_RO(hca_type); |
807 | |
808 | static struct attribute *bnxt_re_attributes[] = { |
809 | &dev_attr_hw_rev.attr, |
810 | &dev_attr_hca_type.attr, |
811 | NULL |
812 | }; |
813 | |
814 | static const struct attribute_group bnxt_re_dev_attr_group = { |
815 | .attrs = bnxt_re_attributes, |
816 | }; |
817 | |
818 | static const struct ib_device_ops bnxt_re_dev_ops = { |
819 | .owner = THIS_MODULE, |
820 | .driver_id = RDMA_DRIVER_BNXT_RE, |
821 | .uverbs_abi_ver = BNXT_RE_ABI_VERSION, |
822 | |
823 | .add_gid = bnxt_re_add_gid, |
824 | .alloc_hw_port_stats = bnxt_re_ib_alloc_hw_port_stats, |
825 | .alloc_mr = bnxt_re_alloc_mr, |
826 | .alloc_pd = bnxt_re_alloc_pd, |
827 | .alloc_ucontext = bnxt_re_alloc_ucontext, |
828 | .create_ah = bnxt_re_create_ah, |
829 | .create_cq = bnxt_re_create_cq, |
830 | .create_qp = bnxt_re_create_qp, |
831 | .create_srq = bnxt_re_create_srq, |
832 | .create_user_ah = bnxt_re_create_ah, |
833 | .dealloc_pd = bnxt_re_dealloc_pd, |
834 | .dealloc_ucontext = bnxt_re_dealloc_ucontext, |
835 | .del_gid = bnxt_re_del_gid, |
836 | .dereg_mr = bnxt_re_dereg_mr, |
837 | .destroy_ah = bnxt_re_destroy_ah, |
838 | .destroy_cq = bnxt_re_destroy_cq, |
839 | .destroy_qp = bnxt_re_destroy_qp, |
840 | .destroy_srq = bnxt_re_destroy_srq, |
841 | .device_group = &bnxt_re_dev_attr_group, |
842 | .disassociate_ucontext = bnxt_re_disassociate_ucontext, |
843 | .get_dev_fw_str = bnxt_re_query_fw_str, |
844 | .get_dma_mr = bnxt_re_get_dma_mr, |
845 | .get_hw_stats = bnxt_re_ib_get_hw_stats, |
846 | .get_link_layer = bnxt_re_get_link_layer, |
847 | .get_port_immutable = bnxt_re_get_port_immutable, |
848 | .map_mr_sg = bnxt_re_map_mr_sg, |
849 | .mmap = bnxt_re_mmap, |
850 | .mmap_free = bnxt_re_mmap_free, |
851 | .modify_qp = bnxt_re_modify_qp, |
852 | .modify_srq = bnxt_re_modify_srq, |
853 | .poll_cq = bnxt_re_poll_cq, |
854 | .post_recv = bnxt_re_post_recv, |
855 | .post_send = bnxt_re_post_send, |
856 | .post_srq_recv = bnxt_re_post_srq_recv, |
857 | .query_ah = bnxt_re_query_ah, |
858 | .query_device = bnxt_re_query_device, |
859 | .query_pkey = bnxt_re_query_pkey, |
860 | .query_port = bnxt_re_query_port, |
861 | .query_qp = bnxt_re_query_qp, |
862 | .query_srq = bnxt_re_query_srq, |
863 | .reg_user_mr = bnxt_re_reg_user_mr, |
864 | .reg_user_mr_dmabuf = bnxt_re_reg_user_mr_dmabuf, |
865 | .req_notify_cq = bnxt_re_req_notify_cq, |
866 | .resize_cq = bnxt_re_resize_cq, |
867 | INIT_RDMA_OBJ_SIZE(ib_ah, bnxt_re_ah, ib_ah), |
868 | INIT_RDMA_OBJ_SIZE(ib_cq, bnxt_re_cq, ib_cq), |
869 | INIT_RDMA_OBJ_SIZE(ib_pd, bnxt_re_pd, ib_pd), |
870 | INIT_RDMA_OBJ_SIZE(ib_qp, bnxt_re_qp, ib_qp), |
871 | INIT_RDMA_OBJ_SIZE(ib_srq, bnxt_re_srq, ib_srq), |
872 | INIT_RDMA_OBJ_SIZE(ib_ucontext, bnxt_re_ucontext, ib_uctx), |
873 | }; |
874 | |
875 | static int bnxt_re_register_ib(struct bnxt_re_dev *rdev) |
876 | { |
877 | struct ib_device *ibdev = &rdev->ibdev; |
878 | int ret; |
879 | |
880 | /* ib device init */ |
881 | ibdev->node_type = RDMA_NODE_IB_CA; |
882 | strscpy(p: ibdev->node_desc, BNXT_RE_DESC " HCA" , |
883 | strlen(BNXT_RE_DESC) + 5); |
884 | ibdev->phys_port_cnt = 1; |
885 | |
886 | addrconf_addr_eui48(eui: (u8 *)&ibdev->node_guid, addr: rdev->netdev->dev_addr); |
887 | |
888 | ibdev->num_comp_vectors = rdev->num_msix - 1; |
889 | ibdev->dev.parent = &rdev->en_dev->pdev->dev; |
890 | ibdev->local_dma_lkey = BNXT_QPLIB_RSVD_LKEY; |
891 | |
892 | if (IS_ENABLED(CONFIG_INFINIBAND_USER_ACCESS)) |
893 | ibdev->driver_def = bnxt_re_uapi_defs; |
894 | |
895 | ib_set_device_ops(device: ibdev, ops: &bnxt_re_dev_ops); |
896 | ret = ib_device_set_netdev(ib_dev: &rdev->ibdev, ndev: rdev->netdev, port: 1); |
897 | if (ret) |
898 | return ret; |
899 | |
900 | dma_set_max_seg_size(dev: &rdev->en_dev->pdev->dev, UINT_MAX); |
901 | ibdev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POLL_CQ); |
902 | return ib_register_device(device: ibdev, name: "bnxt_re%d" , dma_device: &rdev->en_dev->pdev->dev); |
903 | } |
904 | |
905 | static struct bnxt_re_dev *bnxt_re_dev_add(struct bnxt_aux_priv *aux_priv, |
906 | struct bnxt_en_dev *en_dev) |
907 | { |
908 | struct bnxt_re_dev *rdev; |
909 | |
910 | /* Allocate bnxt_re_dev instance here */ |
911 | rdev = ib_alloc_device(bnxt_re_dev, ibdev); |
912 | if (!rdev) { |
913 | ibdev_err(NULL, format: "%s: bnxt_re_dev allocation failure!" , |
914 | ROCE_DRV_MODULE_NAME); |
915 | return NULL; |
916 | } |
917 | /* Default values */ |
918 | rdev->nb.notifier_call = NULL; |
919 | rdev->netdev = en_dev->net; |
920 | rdev->en_dev = en_dev; |
921 | rdev->id = rdev->en_dev->pdev->devfn; |
922 | INIT_LIST_HEAD(list: &rdev->qp_list); |
923 | mutex_init(&rdev->qp_lock); |
924 | mutex_init(&rdev->pacing.dbq_lock); |
925 | atomic_set(v: &rdev->stats.res.qp_count, i: 0); |
926 | atomic_set(v: &rdev->stats.res.cq_count, i: 0); |
927 | atomic_set(v: &rdev->stats.res.srq_count, i: 0); |
928 | atomic_set(v: &rdev->stats.res.mr_count, i: 0); |
929 | atomic_set(v: &rdev->stats.res.mw_count, i: 0); |
930 | atomic_set(v: &rdev->stats.res.ah_count, i: 0); |
931 | atomic_set(v: &rdev->stats.res.pd_count, i: 0); |
932 | rdev->cosq[0] = 0xFFFF; |
933 | rdev->cosq[1] = 0xFFFF; |
934 | |
935 | return rdev; |
936 | } |
937 | |
938 | static int bnxt_re_handle_unaffi_async_event(struct creq_func_event |
939 | *unaffi_async) |
940 | { |
941 | switch (unaffi_async->event) { |
942 | case CREQ_FUNC_EVENT_EVENT_TX_WQE_ERROR: |
943 | break; |
944 | case CREQ_FUNC_EVENT_EVENT_TX_DATA_ERROR: |
945 | break; |
946 | case CREQ_FUNC_EVENT_EVENT_RX_WQE_ERROR: |
947 | break; |
948 | case CREQ_FUNC_EVENT_EVENT_RX_DATA_ERROR: |
949 | break; |
950 | case CREQ_FUNC_EVENT_EVENT_CQ_ERROR: |
951 | break; |
952 | case CREQ_FUNC_EVENT_EVENT_TQM_ERROR: |
953 | break; |
954 | case CREQ_FUNC_EVENT_EVENT_CFCQ_ERROR: |
955 | break; |
956 | case CREQ_FUNC_EVENT_EVENT_CFCS_ERROR: |
957 | break; |
958 | case CREQ_FUNC_EVENT_EVENT_CFCC_ERROR: |
959 | break; |
960 | case CREQ_FUNC_EVENT_EVENT_CFCM_ERROR: |
961 | break; |
962 | case CREQ_FUNC_EVENT_EVENT_TIM_ERROR: |
963 | break; |
964 | default: |
965 | return -EINVAL; |
966 | } |
967 | return 0; |
968 | } |
969 | |
970 | static int bnxt_re_handle_qp_async_event(struct creq_qp_event *qp_event, |
971 | struct bnxt_re_qp *qp) |
972 | { |
973 | struct bnxt_re_srq *srq = container_of(qp->qplib_qp.srq, struct bnxt_re_srq, |
974 | qplib_srq); |
975 | struct creq_qp_error_notification *err_event; |
976 | struct ib_event event = {}; |
977 | unsigned int flags; |
978 | |
979 | if (qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR && |
980 | rdma_is_kernel_res(res: &qp->ib_qp.res)) { |
981 | flags = bnxt_re_lock_cqs(qp); |
982 | bnxt_qplib_add_flush_qp(qp: &qp->qplib_qp); |
983 | bnxt_re_unlock_cqs(qp, flags); |
984 | } |
985 | |
986 | event.device = &qp->rdev->ibdev; |
987 | event.element.qp = &qp->ib_qp; |
988 | event.event = IB_EVENT_QP_FATAL; |
989 | |
990 | err_event = (struct creq_qp_error_notification *)qp_event; |
991 | |
992 | switch (err_event->req_err_state_reason) { |
993 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_OPCODE_ERROR: |
994 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_TIMEOUT_RETRY_LIMIT: |
995 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_RNR_TIMEOUT_RETRY_LIMIT: |
996 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_NAK_ARRIVAL_2: |
997 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_NAK_ARRIVAL_3: |
998 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_INVALID_READ_RESP: |
999 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_ILLEGAL_BIND: |
1000 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_ILLEGAL_FAST_REG: |
1001 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_ILLEGAL_INVALIDATE: |
1002 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_RETRAN_LOCAL_ERROR: |
1003 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_AV_DOMAIN_ERROR: |
1004 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_PROD_WQE_MSMTCH_ERROR: |
1005 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_PSN_RANGE_CHECK_ERROR: |
1006 | event.event = IB_EVENT_QP_ACCESS_ERR; |
1007 | break; |
1008 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_NAK_ARRIVAL_1: |
1009 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_NAK_ARRIVAL_4: |
1010 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_READ_RESP_LENGTH: |
1011 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_WQE_FORMAT_ERROR: |
1012 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_ORRQ_FORMAT_ERROR: |
1013 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_INVALID_AVID_ERROR: |
1014 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_SERV_TYPE_ERROR: |
1015 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_INVALID_OP_ERROR: |
1016 | event.event = IB_EVENT_QP_REQ_ERR; |
1017 | break; |
1018 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_RX_MEMORY_ERROR: |
1019 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_TX_MEMORY_ERROR: |
1020 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_CMP_ERROR: |
1021 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_CQ_LOAD_ERROR: |
1022 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_TX_PCI_ERROR: |
1023 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_RX_PCI_ERROR: |
1024 | case CREQ_QP_ERROR_NOTIFICATION_REQ_ERR_STATE_REASON_REQ_RETX_SETUP_ERROR: |
1025 | event.event = IB_EVENT_QP_FATAL; |
1026 | break; |
1027 | |
1028 | default: |
1029 | break; |
1030 | } |
1031 | |
1032 | switch (err_event->res_err_state_reason) { |
1033 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_EXCEED_MAX: |
1034 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_PAYLOAD_LENGTH_MISMATCH: |
1035 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_PSN_SEQ_ERROR_RETRY_LIMIT: |
1036 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_RX_INVALID_R_KEY: |
1037 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_RX_DOMAIN_ERROR: |
1038 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_RX_NO_PERMISSION: |
1039 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_RX_RANGE_ERROR: |
1040 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_TX_INVALID_R_KEY: |
1041 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_TX_DOMAIN_ERROR: |
1042 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_TX_NO_PERMISSION: |
1043 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_TX_RANGE_ERROR: |
1044 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_UNALIGN_ATOMIC: |
1045 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_PSN_NOT_FOUND: |
1046 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_INVALID_DUP_RKEY: |
1047 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_IRRQ_FORMAT_ERROR: |
1048 | event.event = IB_EVENT_QP_ACCESS_ERR; |
1049 | break; |
1050 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_EXCEEDS_WQE: |
1051 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_WQE_FORMAT_ERROR: |
1052 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_UNSUPPORTED_OPCODE: |
1053 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_REM_INVALIDATE: |
1054 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_OPCODE_ERROR: |
1055 | event.event = IB_EVENT_QP_REQ_ERR; |
1056 | break; |
1057 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_IRRQ_OFLOW: |
1058 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_CMP_ERROR: |
1059 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_CQ_LOAD_ERROR: |
1060 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_TX_PCI_ERROR: |
1061 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_RX_PCI_ERROR: |
1062 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_MEMORY_ERROR: |
1063 | event.event = IB_EVENT_QP_FATAL; |
1064 | break; |
1065 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_SRQ_LOAD_ERROR: |
1066 | case CREQ_QP_ERROR_NOTIFICATION_RES_ERR_STATE_REASON_RES_SRQ_ERROR: |
1067 | if (srq) |
1068 | event.event = IB_EVENT_SRQ_ERR; |
1069 | break; |
1070 | default: |
1071 | break; |
1072 | } |
1073 | |
1074 | if (err_event->res_err_state_reason || err_event->req_err_state_reason) { |
1075 | ibdev_dbg(&qp->rdev->ibdev, |
1076 | "%s %s qp_id: %d cons (%d %d) req (%d %d) res (%d %d)\n" , |
1077 | __func__, rdma_is_kernel_res(&qp->ib_qp.res) ? "kernel" : "user" , |
1078 | qp->qplib_qp.id, |
1079 | err_event->sq_cons_idx, |
1080 | err_event->rq_cons_idx, |
1081 | err_event->req_slow_path_state, |
1082 | err_event->req_err_state_reason, |
1083 | err_event->res_slow_path_state, |
1084 | err_event->res_err_state_reason); |
1085 | } else { |
1086 | if (srq) |
1087 | event.event = IB_EVENT_QP_LAST_WQE_REACHED; |
1088 | } |
1089 | |
1090 | if (event.event == IB_EVENT_SRQ_ERR && srq->ib_srq.event_handler) { |
1091 | (*srq->ib_srq.event_handler)(&event, |
1092 | srq->ib_srq.srq_context); |
1093 | } else if (event.device && qp->ib_qp.event_handler) { |
1094 | qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context); |
1095 | } |
1096 | |
1097 | return 0; |
1098 | } |
1099 | |
1100 | static int bnxt_re_handle_cq_async_error(void *event, struct bnxt_re_cq *cq) |
1101 | { |
1102 | struct creq_cq_error_notification *cqerr; |
1103 | struct ib_event ibevent = {}; |
1104 | |
1105 | cqerr = event; |
1106 | switch (cqerr->cq_err_reason) { |
1107 | case CREQ_CQ_ERROR_NOTIFICATION_CQ_ERR_REASON_REQ_CQ_INVALID_ERROR: |
1108 | case CREQ_CQ_ERROR_NOTIFICATION_CQ_ERR_REASON_REQ_CQ_OVERFLOW_ERROR: |
1109 | case CREQ_CQ_ERROR_NOTIFICATION_CQ_ERR_REASON_REQ_CQ_LOAD_ERROR: |
1110 | case CREQ_CQ_ERROR_NOTIFICATION_CQ_ERR_REASON_RES_CQ_INVALID_ERROR: |
1111 | case CREQ_CQ_ERROR_NOTIFICATION_CQ_ERR_REASON_RES_CQ_OVERFLOW_ERROR: |
1112 | case CREQ_CQ_ERROR_NOTIFICATION_CQ_ERR_REASON_RES_CQ_LOAD_ERROR: |
1113 | ibevent.event = IB_EVENT_CQ_ERR; |
1114 | break; |
1115 | default: |
1116 | break; |
1117 | } |
1118 | |
1119 | if (ibevent.event == IB_EVENT_CQ_ERR && cq->ib_cq.event_handler) { |
1120 | ibevent.element.cq = &cq->ib_cq; |
1121 | ibevent.device = &cq->rdev->ibdev; |
1122 | |
1123 | ibdev_dbg(&cq->rdev->ibdev, |
1124 | "%s err reason %d\n" , __func__, cqerr->cq_err_reason); |
1125 | cq->ib_cq.event_handler(&ibevent, cq->ib_cq.cq_context); |
1126 | } |
1127 | |
1128 | return 0; |
1129 | } |
1130 | |
1131 | static int bnxt_re_handle_affi_async_event(struct creq_qp_event *affi_async, |
1132 | void *obj) |
1133 | { |
1134 | struct bnxt_qplib_qp *lib_qp; |
1135 | struct bnxt_qplib_cq *lib_cq; |
1136 | struct bnxt_re_qp *qp; |
1137 | struct bnxt_re_cq *cq; |
1138 | int rc = 0; |
1139 | u8 event; |
1140 | |
1141 | if (!obj) |
1142 | return rc; /* QP was already dead, still return success */ |
1143 | |
1144 | event = affi_async->event; |
1145 | switch (event) { |
1146 | case CREQ_QP_EVENT_EVENT_QP_ERROR_NOTIFICATION: |
1147 | lib_qp = obj; |
1148 | qp = container_of(lib_qp, struct bnxt_re_qp, qplib_qp); |
1149 | rc = bnxt_re_handle_qp_async_event(qp_event: affi_async, qp); |
1150 | break; |
1151 | case CREQ_QP_EVENT_EVENT_CQ_ERROR_NOTIFICATION: |
1152 | lib_cq = obj; |
1153 | cq = container_of(lib_cq, struct bnxt_re_cq, qplib_cq); |
1154 | rc = bnxt_re_handle_cq_async_error(event: affi_async, cq); |
1155 | break; |
1156 | default: |
1157 | rc = -EINVAL; |
1158 | } |
1159 | return rc; |
1160 | } |
1161 | |
1162 | static int bnxt_re_aeq_handler(struct bnxt_qplib_rcfw *rcfw, |
1163 | void *aeqe, void *obj) |
1164 | { |
1165 | struct creq_qp_event *affi_async; |
1166 | struct creq_func_event *unaffi_async; |
1167 | u8 type; |
1168 | int rc; |
1169 | |
1170 | type = ((struct creq_base *)aeqe)->type; |
1171 | if (type == CREQ_BASE_TYPE_FUNC_EVENT) { |
1172 | unaffi_async = aeqe; |
1173 | rc = bnxt_re_handle_unaffi_async_event(unaffi_async); |
1174 | } else { |
1175 | affi_async = aeqe; |
1176 | rc = bnxt_re_handle_affi_async_event(affi_async, obj); |
1177 | } |
1178 | |
1179 | return rc; |
1180 | } |
1181 | |
1182 | static int bnxt_re_srqn_handler(struct bnxt_qplib_nq *nq, |
1183 | struct bnxt_qplib_srq *handle, u8 event) |
1184 | { |
1185 | struct bnxt_re_srq *srq = container_of(handle, struct bnxt_re_srq, |
1186 | qplib_srq); |
1187 | struct ib_event ib_event; |
1188 | |
1189 | ib_event.device = &srq->rdev->ibdev; |
1190 | ib_event.element.srq = &srq->ib_srq; |
1191 | |
1192 | if (srq->ib_srq.event_handler) { |
1193 | if (event == NQ_SRQ_EVENT_EVENT_SRQ_THRESHOLD_EVENT) |
1194 | ib_event.event = IB_EVENT_SRQ_LIMIT_REACHED; |
1195 | (*srq->ib_srq.event_handler)(&ib_event, |
1196 | srq->ib_srq.srq_context); |
1197 | } |
1198 | return 0; |
1199 | } |
1200 | |
1201 | static int bnxt_re_cqn_handler(struct bnxt_qplib_nq *nq, |
1202 | struct bnxt_qplib_cq *handle) |
1203 | { |
1204 | struct bnxt_re_cq *cq = container_of(handle, struct bnxt_re_cq, |
1205 | qplib_cq); |
1206 | |
1207 | if (cq->ib_cq.comp_handler) { |
1208 | /* Lock comp_handler? */ |
1209 | (*cq->ib_cq.comp_handler)(&cq->ib_cq, cq->ib_cq.cq_context); |
1210 | } |
1211 | |
1212 | return 0; |
1213 | } |
1214 | |
1215 | #define BNXT_RE_GEN_P5_PF_NQ_DB 0x10000 |
1216 | #define BNXT_RE_GEN_P5_VF_NQ_DB 0x4000 |
1217 | static u32 bnxt_re_get_nqdb_offset(struct bnxt_re_dev *rdev, u16 indx) |
1218 | { |
1219 | return bnxt_qplib_is_chip_gen_p5(cctx: rdev->chip_ctx) ? |
1220 | (rdev->is_virtfn ? BNXT_RE_GEN_P5_VF_NQ_DB : |
1221 | BNXT_RE_GEN_P5_PF_NQ_DB) : |
1222 | rdev->en_dev->msix_entries[indx].db_offset; |
1223 | } |
1224 | |
1225 | static void bnxt_re_cleanup_res(struct bnxt_re_dev *rdev) |
1226 | { |
1227 | int i; |
1228 | |
1229 | for (i = 1; i < rdev->num_msix; i++) |
1230 | bnxt_qplib_disable_nq(nq: &rdev->nq[i - 1]); |
1231 | |
1232 | if (rdev->qplib_res.rcfw) |
1233 | bnxt_qplib_cleanup_res(res: &rdev->qplib_res); |
1234 | } |
1235 | |
1236 | static int bnxt_re_init_res(struct bnxt_re_dev *rdev) |
1237 | { |
1238 | int num_vec_enabled = 0; |
1239 | int rc = 0, i; |
1240 | u32 db_offt; |
1241 | |
1242 | bnxt_qplib_init_res(res: &rdev->qplib_res); |
1243 | |
1244 | for (i = 1; i < rdev->num_msix ; i++) { |
1245 | db_offt = bnxt_re_get_nqdb_offset(rdev, indx: i); |
1246 | rc = bnxt_qplib_enable_nq(pdev: rdev->en_dev->pdev, nq: &rdev->nq[i - 1], |
1247 | nq_idx: i - 1, msix_vector: rdev->en_dev->msix_entries[i].vector, |
1248 | bar_reg_offset: db_offt, cqn_handler: &bnxt_re_cqn_handler, |
1249 | srq_handler: &bnxt_re_srqn_handler); |
1250 | if (rc) { |
1251 | ibdev_err(ibdev: &rdev->ibdev, |
1252 | format: "Failed to enable NQ with rc = 0x%x" , rc); |
1253 | goto fail; |
1254 | } |
1255 | num_vec_enabled++; |
1256 | } |
1257 | return 0; |
1258 | fail: |
1259 | for (i = num_vec_enabled; i >= 0; i--) |
1260 | bnxt_qplib_disable_nq(nq: &rdev->nq[i]); |
1261 | return rc; |
1262 | } |
1263 | |
1264 | static void bnxt_re_free_nq_res(struct bnxt_re_dev *rdev) |
1265 | { |
1266 | u8 type; |
1267 | int i; |
1268 | |
1269 | for (i = 0; i < rdev->num_msix - 1; i++) { |
1270 | type = bnxt_qplib_get_ring_type(cctx: rdev->chip_ctx); |
1271 | bnxt_re_net_ring_free(rdev, fw_ring_id: rdev->nq[i].ring_id, type); |
1272 | bnxt_qplib_free_nq(nq: &rdev->nq[i]); |
1273 | rdev->nq[i].res = NULL; |
1274 | } |
1275 | } |
1276 | |
1277 | static void bnxt_re_free_res(struct bnxt_re_dev *rdev) |
1278 | { |
1279 | bnxt_re_free_nq_res(rdev); |
1280 | |
1281 | if (rdev->qplib_res.dpi_tbl.max) { |
1282 | bnxt_qplib_dealloc_dpi(res: &rdev->qplib_res, |
1283 | dpi: &rdev->dpi_privileged); |
1284 | } |
1285 | if (rdev->qplib_res.rcfw) { |
1286 | bnxt_qplib_free_res(res: &rdev->qplib_res); |
1287 | rdev->qplib_res.rcfw = NULL; |
1288 | } |
1289 | } |
1290 | |
1291 | static int bnxt_re_alloc_res(struct bnxt_re_dev *rdev) |
1292 | { |
1293 | struct bnxt_re_ring_attr rattr = {}; |
1294 | int num_vec_created = 0; |
1295 | int rc, i; |
1296 | u8 type; |
1297 | |
1298 | /* Configure and allocate resources for qplib */ |
1299 | rdev->qplib_res.rcfw = &rdev->rcfw; |
1300 | rc = bnxt_qplib_get_dev_attr(rcfw: &rdev->rcfw, attr: &rdev->dev_attr); |
1301 | if (rc) |
1302 | goto fail; |
1303 | |
1304 | rc = bnxt_qplib_alloc_res(res: &rdev->qplib_res, pdev: rdev->en_dev->pdev, |
1305 | netdev: rdev->netdev, dev_attr: &rdev->dev_attr); |
1306 | if (rc) |
1307 | goto fail; |
1308 | |
1309 | rc = bnxt_qplib_alloc_dpi(res: &rdev->qplib_res, |
1310 | dpi: &rdev->dpi_privileged, |
1311 | app: rdev, type: BNXT_QPLIB_DPI_TYPE_KERNEL); |
1312 | if (rc) |
1313 | goto dealloc_res; |
1314 | |
1315 | for (i = 0; i < rdev->num_msix - 1; i++) { |
1316 | struct bnxt_qplib_nq *nq; |
1317 | |
1318 | nq = &rdev->nq[i]; |
1319 | nq->hwq.max_elements = BNXT_QPLIB_NQE_MAX_CNT; |
1320 | rc = bnxt_qplib_alloc_nq(res: &rdev->qplib_res, nq: &rdev->nq[i]); |
1321 | if (rc) { |
1322 | ibdev_err(ibdev: &rdev->ibdev, format: "Alloc Failed NQ%d rc:%#x" , |
1323 | i, rc); |
1324 | goto free_nq; |
1325 | } |
1326 | type = bnxt_qplib_get_ring_type(cctx: rdev->chip_ctx); |
1327 | rattr.dma_arr = nq->hwq.pbl[PBL_LVL_0].pg_map_arr; |
1328 | rattr.pages = nq->hwq.pbl[rdev->nq[i].hwq.level].pg_count; |
1329 | rattr.type = type; |
1330 | rattr.mode = RING_ALLOC_REQ_INT_MODE_MSIX; |
1331 | rattr.depth = BNXT_QPLIB_NQE_MAX_CNT - 1; |
1332 | rattr.lrid = rdev->en_dev->msix_entries[i + 1].ring_idx; |
1333 | rc = bnxt_re_net_ring_alloc(rdev, ring_attr: &rattr, fw_ring_id: &nq->ring_id); |
1334 | if (rc) { |
1335 | ibdev_err(ibdev: &rdev->ibdev, |
1336 | format: "Failed to allocate NQ fw id with rc = 0x%x" , |
1337 | rc); |
1338 | bnxt_qplib_free_nq(nq: &rdev->nq[i]); |
1339 | goto free_nq; |
1340 | } |
1341 | num_vec_created++; |
1342 | } |
1343 | return 0; |
1344 | free_nq: |
1345 | for (i = num_vec_created - 1; i >= 0; i--) { |
1346 | type = bnxt_qplib_get_ring_type(cctx: rdev->chip_ctx); |
1347 | bnxt_re_net_ring_free(rdev, fw_ring_id: rdev->nq[i].ring_id, type); |
1348 | bnxt_qplib_free_nq(nq: &rdev->nq[i]); |
1349 | } |
1350 | bnxt_qplib_dealloc_dpi(res: &rdev->qplib_res, |
1351 | dpi: &rdev->dpi_privileged); |
1352 | dealloc_res: |
1353 | bnxt_qplib_free_res(res: &rdev->qplib_res); |
1354 | |
1355 | fail: |
1356 | rdev->qplib_res.rcfw = NULL; |
1357 | return rc; |
1358 | } |
1359 | |
1360 | static void bnxt_re_dispatch_event(struct ib_device *ibdev, struct ib_qp *qp, |
1361 | u8 port_num, enum ib_event_type event) |
1362 | { |
1363 | struct ib_event ib_event; |
1364 | |
1365 | ib_event.device = ibdev; |
1366 | if (qp) { |
1367 | ib_event.element.qp = qp; |
1368 | ib_event.event = event; |
1369 | if (qp->event_handler) |
1370 | qp->event_handler(&ib_event, qp->qp_context); |
1371 | |
1372 | } else { |
1373 | ib_event.element.port_num = port_num; |
1374 | ib_event.event = event; |
1375 | ib_dispatch_event(event: &ib_event); |
1376 | } |
1377 | } |
1378 | |
1379 | static bool bnxt_re_is_qp1_or_shadow_qp(struct bnxt_re_dev *rdev, |
1380 | struct bnxt_re_qp *qp) |
1381 | { |
1382 | return (qp->ib_qp.qp_type == IB_QPT_GSI) || |
1383 | (qp == rdev->gsi_ctx.gsi_sqp); |
1384 | } |
1385 | |
1386 | static void bnxt_re_dev_stop(struct bnxt_re_dev *rdev) |
1387 | { |
1388 | int mask = IB_QP_STATE; |
1389 | struct ib_qp_attr qp_attr; |
1390 | struct bnxt_re_qp *qp; |
1391 | |
1392 | qp_attr.qp_state = IB_QPS_ERR; |
1393 | mutex_lock(&rdev->qp_lock); |
1394 | list_for_each_entry(qp, &rdev->qp_list, list) { |
1395 | /* Modify the state of all QPs except QP1/Shadow QP */ |
1396 | if (!bnxt_re_is_qp1_or_shadow_qp(rdev, qp)) { |
1397 | if (qp->qplib_qp.state != |
1398 | CMDQ_MODIFY_QP_NEW_STATE_RESET && |
1399 | qp->qplib_qp.state != |
1400 | CMDQ_MODIFY_QP_NEW_STATE_ERR) { |
1401 | bnxt_re_dispatch_event(ibdev: &rdev->ibdev, qp: &qp->ib_qp, |
1402 | port_num: 1, event: IB_EVENT_QP_FATAL); |
1403 | bnxt_re_modify_qp(qp: &qp->ib_qp, qp_attr: &qp_attr, qp_attr_mask: mask, |
1404 | NULL); |
1405 | } |
1406 | } |
1407 | } |
1408 | mutex_unlock(lock: &rdev->qp_lock); |
1409 | } |
1410 | |
1411 | static int bnxt_re_update_gid(struct bnxt_re_dev *rdev) |
1412 | { |
1413 | struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl; |
1414 | struct bnxt_qplib_gid gid; |
1415 | u16 gid_idx, index; |
1416 | int rc = 0; |
1417 | |
1418 | if (!ib_device_try_get(dev: &rdev->ibdev)) |
1419 | return 0; |
1420 | |
1421 | for (index = 0; index < sgid_tbl->active; index++) { |
1422 | gid_idx = sgid_tbl->hw_id[index]; |
1423 | |
1424 | if (!memcmp(p: &sgid_tbl->tbl[index], q: &bnxt_qplib_gid_zero, |
1425 | size: sizeof(bnxt_qplib_gid_zero))) |
1426 | continue; |
1427 | /* need to modify the VLAN enable setting of non VLAN GID only |
1428 | * as setting is done for VLAN GID while adding GID |
1429 | */ |
1430 | if (sgid_tbl->vlan[index]) |
1431 | continue; |
1432 | |
1433 | memcpy(&gid, &sgid_tbl->tbl[index], sizeof(gid)); |
1434 | |
1435 | rc = bnxt_qplib_update_sgid(sgid_tbl, gid: &gid, gid_idx, |
1436 | smac: rdev->qplib_res.netdev->dev_addr); |
1437 | } |
1438 | |
1439 | ib_device_put(device: &rdev->ibdev); |
1440 | return rc; |
1441 | } |
1442 | |
1443 | static u32 bnxt_re_get_priority_mask(struct bnxt_re_dev *rdev) |
1444 | { |
1445 | u32 prio_map = 0, tmp_map = 0; |
1446 | struct net_device *netdev; |
1447 | struct dcb_app app = {}; |
1448 | |
1449 | netdev = rdev->netdev; |
1450 | |
1451 | app.selector = IEEE_8021QAZ_APP_SEL_ETHERTYPE; |
1452 | app.protocol = ETH_P_IBOE; |
1453 | tmp_map = dcb_ieee_getapp_mask(netdev, &app); |
1454 | prio_map = tmp_map; |
1455 | |
1456 | app.selector = IEEE_8021QAZ_APP_SEL_DGRAM; |
1457 | app.protocol = ROCE_V2_UDP_DPORT; |
1458 | tmp_map = dcb_ieee_getapp_mask(netdev, &app); |
1459 | prio_map |= tmp_map; |
1460 | |
1461 | return prio_map; |
1462 | } |
1463 | |
1464 | static int bnxt_re_setup_qos(struct bnxt_re_dev *rdev) |
1465 | { |
1466 | u8 prio_map = 0; |
1467 | |
1468 | /* Get priority for roce */ |
1469 | prio_map = bnxt_re_get_priority_mask(rdev); |
1470 | |
1471 | if (prio_map == rdev->cur_prio_map) |
1472 | return 0; |
1473 | rdev->cur_prio_map = prio_map; |
1474 | /* Actual priorities are not programmed as they are already |
1475 | * done by L2 driver; just enable or disable priority vlan tagging |
1476 | */ |
1477 | if ((prio_map == 0 && rdev->qplib_res.prio) || |
1478 | (prio_map != 0 && !rdev->qplib_res.prio)) { |
1479 | rdev->qplib_res.prio = prio_map; |
1480 | bnxt_re_update_gid(rdev); |
1481 | } |
1482 | |
1483 | return 0; |
1484 | } |
1485 | |
1486 | static void bnxt_re_query_hwrm_intf_version(struct bnxt_re_dev *rdev) |
1487 | { |
1488 | struct bnxt_en_dev *en_dev = rdev->en_dev; |
1489 | struct hwrm_ver_get_output resp = {}; |
1490 | struct hwrm_ver_get_input req = {}; |
1491 | struct bnxt_qplib_chip_ctx *cctx; |
1492 | struct bnxt_fw_msg fw_msg = {}; |
1493 | int rc; |
1494 | |
1495 | bnxt_re_init_hwrm_hdr(hdr: (void *)&req, opcd: HWRM_VER_GET); |
1496 | req.hwrm_intf_maj = HWRM_VERSION_MAJOR; |
1497 | req.hwrm_intf_min = HWRM_VERSION_MINOR; |
1498 | req.hwrm_intf_upd = HWRM_VERSION_UPDATE; |
1499 | bnxt_re_fill_fw_msg(fw_msg: &fw_msg, msg: (void *)&req, msg_len: sizeof(req), resp: (void *)&resp, |
1500 | resp_max_len: sizeof(resp), timeout: DFLT_HWRM_CMD_TIMEOUT); |
1501 | rc = bnxt_send_msg(en_dev, &fw_msg); |
1502 | if (rc) { |
1503 | ibdev_err(ibdev: &rdev->ibdev, format: "Failed to query HW version, rc = 0x%x" , |
1504 | rc); |
1505 | return; |
1506 | } |
1507 | |
1508 | cctx = rdev->chip_ctx; |
1509 | cctx->hwrm_intf_ver = |
1510 | (u64)le16_to_cpu(resp.hwrm_intf_major) << 48 | |
1511 | (u64)le16_to_cpu(resp.hwrm_intf_minor) << 32 | |
1512 | (u64)le16_to_cpu(resp.hwrm_intf_build) << 16 | |
1513 | le16_to_cpu(resp.hwrm_intf_patch); |
1514 | |
1515 | cctx->hwrm_cmd_max_timeout = le16_to_cpu(resp.max_req_timeout); |
1516 | |
1517 | if (!cctx->hwrm_cmd_max_timeout) |
1518 | cctx->hwrm_cmd_max_timeout = RCFW_FW_STALL_MAX_TIMEOUT; |
1519 | } |
1520 | |
1521 | static int bnxt_re_ib_init(struct bnxt_re_dev *rdev) |
1522 | { |
1523 | int rc; |
1524 | u32 event; |
1525 | |
1526 | /* Register ib dev */ |
1527 | rc = bnxt_re_register_ib(rdev); |
1528 | if (rc) { |
1529 | pr_err("Failed to register with IB: %#x\n" , rc); |
1530 | return rc; |
1531 | } |
1532 | dev_info(rdev_to_dev(rdev), "Device registered with IB successfully" ); |
1533 | set_bit(BNXT_RE_FLAG_ISSUE_ROCE_STATS, addr: &rdev->flags); |
1534 | |
1535 | event = netif_running(dev: rdev->netdev) && netif_carrier_ok(dev: rdev->netdev) ? |
1536 | IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR; |
1537 | |
1538 | bnxt_re_dispatch_event(ibdev: &rdev->ibdev, NULL, port_num: 1, event); |
1539 | |
1540 | return rc; |
1541 | } |
1542 | |
1543 | static void bnxt_re_dev_uninit(struct bnxt_re_dev *rdev) |
1544 | { |
1545 | u8 type; |
1546 | int rc; |
1547 | |
1548 | if (test_and_clear_bit(BNXT_RE_FLAG_QOS_WORK_REG, addr: &rdev->flags)) |
1549 | cancel_delayed_work_sync(dwork: &rdev->worker); |
1550 | |
1551 | if (test_and_clear_bit(BNXT_RE_FLAG_RESOURCES_INITIALIZED, |
1552 | addr: &rdev->flags)) |
1553 | bnxt_re_cleanup_res(rdev); |
1554 | if (test_and_clear_bit(BNXT_RE_FLAG_RESOURCES_ALLOCATED, addr: &rdev->flags)) |
1555 | bnxt_re_free_res(rdev); |
1556 | |
1557 | if (test_and_clear_bit(BNXT_RE_FLAG_RCFW_CHANNEL_EN, addr: &rdev->flags)) { |
1558 | rc = bnxt_qplib_deinit_rcfw(rcfw: &rdev->rcfw); |
1559 | if (rc) |
1560 | ibdev_warn(ibdev: &rdev->ibdev, |
1561 | format: "Failed to deinitialize RCFW: %#x" , rc); |
1562 | bnxt_re_net_stats_ctx_free(rdev, fw_stats_ctx_id: rdev->qplib_ctx.stats.fw_id); |
1563 | bnxt_qplib_free_ctx(res: &rdev->qplib_res, ctx: &rdev->qplib_ctx); |
1564 | bnxt_qplib_disable_rcfw_channel(rcfw: &rdev->rcfw); |
1565 | type = bnxt_qplib_get_ring_type(cctx: rdev->chip_ctx); |
1566 | bnxt_re_net_ring_free(rdev, fw_ring_id: rdev->rcfw.creq.ring_id, type); |
1567 | bnxt_qplib_free_rcfw_channel(rcfw: &rdev->rcfw); |
1568 | } |
1569 | |
1570 | rdev->num_msix = 0; |
1571 | |
1572 | if (rdev->pacing.dbr_pacing) |
1573 | bnxt_re_deinitialize_dbr_pacing(rdev); |
1574 | |
1575 | bnxt_re_destroy_chip_ctx(rdev); |
1576 | if (test_and_clear_bit(BNXT_RE_FLAG_NETDEV_REGISTERED, addr: &rdev->flags)) |
1577 | bnxt_unregister_dev(rdev->en_dev); |
1578 | } |
1579 | |
1580 | /* worker thread for polling periodic events. Now used for QoS programming*/ |
1581 | static void bnxt_re_worker(struct work_struct *work) |
1582 | { |
1583 | struct bnxt_re_dev *rdev = container_of(work, struct bnxt_re_dev, |
1584 | worker.work); |
1585 | |
1586 | bnxt_re_setup_qos(rdev); |
1587 | schedule_delayed_work(dwork: &rdev->worker, delay: msecs_to_jiffies(m: 30000)); |
1588 | } |
1589 | |
1590 | static int bnxt_re_dev_init(struct bnxt_re_dev *rdev, u8 wqe_mode) |
1591 | { |
1592 | struct bnxt_re_ring_attr rattr = {}; |
1593 | struct bnxt_qplib_creq_ctx *creq; |
1594 | u32 db_offt; |
1595 | int vid; |
1596 | u8 type; |
1597 | int rc; |
1598 | |
1599 | /* Registered a new RoCE device instance to netdev */ |
1600 | rc = bnxt_re_register_netdev(rdev); |
1601 | if (rc) { |
1602 | ibdev_err(ibdev: &rdev->ibdev, |
1603 | format: "Failed to register with netedev: %#x\n" , rc); |
1604 | return -EINVAL; |
1605 | } |
1606 | set_bit(BNXT_RE_FLAG_NETDEV_REGISTERED, addr: &rdev->flags); |
1607 | |
1608 | rc = bnxt_re_setup_chip_ctx(rdev, wqe_mode); |
1609 | if (rc) { |
1610 | bnxt_unregister_dev(rdev->en_dev); |
1611 | clear_bit(BNXT_RE_FLAG_NETDEV_REGISTERED, addr: &rdev->flags); |
1612 | ibdev_err(ibdev: &rdev->ibdev, format: "Failed to get chip context\n" ); |
1613 | return -EINVAL; |
1614 | } |
1615 | |
1616 | /* Check whether VF or PF */ |
1617 | bnxt_re_get_sriov_func_type(rdev); |
1618 | |
1619 | if (!rdev->en_dev->ulp_tbl->msix_requested) { |
1620 | ibdev_err(ibdev: &rdev->ibdev, |
1621 | format: "Failed to get MSI-X vectors: %#x\n" , rc); |
1622 | rc = -EINVAL; |
1623 | goto fail; |
1624 | } |
1625 | ibdev_dbg(&rdev->ibdev, "Got %d MSI-X vectors\n" , |
1626 | rdev->en_dev->ulp_tbl->msix_requested); |
1627 | rdev->num_msix = rdev->en_dev->ulp_tbl->msix_requested; |
1628 | |
1629 | bnxt_re_query_hwrm_intf_version(rdev); |
1630 | |
1631 | /* Establish RCFW Communication Channel to initialize the context |
1632 | * memory for the function and all child VFs |
1633 | */ |
1634 | rc = bnxt_qplib_alloc_rcfw_channel(res: &rdev->qplib_res, rcfw: &rdev->rcfw, |
1635 | ctx: &rdev->qplib_ctx, |
1636 | BNXT_RE_MAX_QPC_COUNT); |
1637 | if (rc) { |
1638 | ibdev_err(ibdev: &rdev->ibdev, |
1639 | format: "Failed to allocate RCFW Channel: %#x\n" , rc); |
1640 | goto fail; |
1641 | } |
1642 | |
1643 | type = bnxt_qplib_get_ring_type(cctx: rdev->chip_ctx); |
1644 | creq = &rdev->rcfw.creq; |
1645 | rattr.dma_arr = creq->hwq.pbl[PBL_LVL_0].pg_map_arr; |
1646 | rattr.pages = creq->hwq.pbl[creq->hwq.level].pg_count; |
1647 | rattr.type = type; |
1648 | rattr.mode = RING_ALLOC_REQ_INT_MODE_MSIX; |
1649 | rattr.depth = BNXT_QPLIB_CREQE_MAX_CNT - 1; |
1650 | rattr.lrid = rdev->en_dev->msix_entries[BNXT_RE_AEQ_IDX].ring_idx; |
1651 | rc = bnxt_re_net_ring_alloc(rdev, ring_attr: &rattr, fw_ring_id: &creq->ring_id); |
1652 | if (rc) { |
1653 | ibdev_err(ibdev: &rdev->ibdev, format: "Failed to allocate CREQ: %#x\n" , rc); |
1654 | goto free_rcfw; |
1655 | } |
1656 | db_offt = bnxt_re_get_nqdb_offset(rdev, BNXT_RE_AEQ_IDX); |
1657 | vid = rdev->en_dev->msix_entries[BNXT_RE_AEQ_IDX].vector; |
1658 | rc = bnxt_qplib_enable_rcfw_channel(rcfw: &rdev->rcfw, |
1659 | msix_vector: vid, cp_bar_reg_off: db_offt, |
1660 | aeq_handler: &bnxt_re_aeq_handler); |
1661 | if (rc) { |
1662 | ibdev_err(ibdev: &rdev->ibdev, format: "Failed to enable RCFW channel: %#x\n" , |
1663 | rc); |
1664 | goto free_ring; |
1665 | } |
1666 | |
1667 | if (bnxt_qplib_dbr_pacing_en(cctx: rdev->chip_ctx)) { |
1668 | rc = bnxt_re_initialize_dbr_pacing(rdev); |
1669 | if (!rc) { |
1670 | rdev->pacing.dbr_pacing = true; |
1671 | } else { |
1672 | ibdev_err(ibdev: &rdev->ibdev, |
1673 | format: "DBR pacing disabled with error : %d\n" , rc); |
1674 | rdev->pacing.dbr_pacing = false; |
1675 | } |
1676 | } |
1677 | rc = bnxt_qplib_get_dev_attr(rcfw: &rdev->rcfw, attr: &rdev->dev_attr); |
1678 | if (rc) |
1679 | goto disable_rcfw; |
1680 | |
1681 | bnxt_re_set_resource_limits(rdev); |
1682 | |
1683 | rc = bnxt_qplib_alloc_ctx(res: &rdev->qplib_res, ctx: &rdev->qplib_ctx, virt_fn: 0, |
1684 | is_p5: bnxt_qplib_is_chip_gen_p5(cctx: rdev->chip_ctx)); |
1685 | if (rc) { |
1686 | ibdev_err(ibdev: &rdev->ibdev, |
1687 | format: "Failed to allocate QPLIB context: %#x\n" , rc); |
1688 | goto disable_rcfw; |
1689 | } |
1690 | rc = bnxt_re_net_stats_ctx_alloc(rdev, |
1691 | dma_map: rdev->qplib_ctx.stats.dma_map, |
1692 | fw_stats_ctx_id: &rdev->qplib_ctx.stats.fw_id); |
1693 | if (rc) { |
1694 | ibdev_err(ibdev: &rdev->ibdev, |
1695 | format: "Failed to allocate stats context: %#x\n" , rc); |
1696 | goto free_ctx; |
1697 | } |
1698 | |
1699 | rc = bnxt_qplib_init_rcfw(rcfw: &rdev->rcfw, ctx: &rdev->qplib_ctx, |
1700 | is_virtfn: rdev->is_virtfn); |
1701 | if (rc) { |
1702 | ibdev_err(ibdev: &rdev->ibdev, |
1703 | format: "Failed to initialize RCFW: %#x\n" , rc); |
1704 | goto free_sctx; |
1705 | } |
1706 | set_bit(BNXT_RE_FLAG_RCFW_CHANNEL_EN, addr: &rdev->flags); |
1707 | |
1708 | /* Resources based on the 'new' device caps */ |
1709 | rc = bnxt_re_alloc_res(rdev); |
1710 | if (rc) { |
1711 | ibdev_err(ibdev: &rdev->ibdev, |
1712 | format: "Failed to allocate resources: %#x\n" , rc); |
1713 | goto fail; |
1714 | } |
1715 | set_bit(BNXT_RE_FLAG_RESOURCES_ALLOCATED, addr: &rdev->flags); |
1716 | rc = bnxt_re_init_res(rdev); |
1717 | if (rc) { |
1718 | ibdev_err(ibdev: &rdev->ibdev, |
1719 | format: "Failed to initialize resources: %#x\n" , rc); |
1720 | goto fail; |
1721 | } |
1722 | |
1723 | set_bit(BNXT_RE_FLAG_RESOURCES_INITIALIZED, addr: &rdev->flags); |
1724 | |
1725 | if (!rdev->is_virtfn) { |
1726 | rc = bnxt_re_setup_qos(rdev); |
1727 | if (rc) |
1728 | ibdev_info(ibdev: &rdev->ibdev, |
1729 | format: "RoCE priority not yet configured\n" ); |
1730 | |
1731 | INIT_DELAYED_WORK(&rdev->worker, bnxt_re_worker); |
1732 | set_bit(BNXT_RE_FLAG_QOS_WORK_REG, addr: &rdev->flags); |
1733 | schedule_delayed_work(dwork: &rdev->worker, delay: msecs_to_jiffies(m: 30000)); |
1734 | /* |
1735 | * Use the total VF count since the actual VF count may not be |
1736 | * available at this point. |
1737 | */ |
1738 | bnxt_re_vf_res_config(rdev); |
1739 | } |
1740 | |
1741 | return 0; |
1742 | free_sctx: |
1743 | bnxt_re_net_stats_ctx_free(rdev, fw_stats_ctx_id: rdev->qplib_ctx.stats.fw_id); |
1744 | free_ctx: |
1745 | bnxt_qplib_free_ctx(res: &rdev->qplib_res, ctx: &rdev->qplib_ctx); |
1746 | disable_rcfw: |
1747 | bnxt_qplib_disable_rcfw_channel(rcfw: &rdev->rcfw); |
1748 | free_ring: |
1749 | type = bnxt_qplib_get_ring_type(cctx: rdev->chip_ctx); |
1750 | bnxt_re_net_ring_free(rdev, fw_ring_id: rdev->rcfw.creq.ring_id, type); |
1751 | free_rcfw: |
1752 | bnxt_qplib_free_rcfw_channel(rcfw: &rdev->rcfw); |
1753 | fail: |
1754 | bnxt_re_dev_uninit(rdev); |
1755 | |
1756 | return rc; |
1757 | } |
1758 | |
1759 | static int bnxt_re_add_device(struct auxiliary_device *adev, u8 wqe_mode) |
1760 | { |
1761 | struct bnxt_aux_priv *aux_priv = |
1762 | container_of(adev, struct bnxt_aux_priv, aux_dev); |
1763 | struct bnxt_en_dev *en_dev; |
1764 | struct bnxt_re_dev *rdev; |
1765 | int rc; |
1766 | |
1767 | /* en_dev should never be NULL as long as adev and aux_dev are valid. */ |
1768 | en_dev = aux_priv->edev; |
1769 | |
1770 | rdev = bnxt_re_dev_add(aux_priv, en_dev); |
1771 | if (!rdev || !rdev_to_dev(rdev)) { |
1772 | rc = -ENOMEM; |
1773 | goto exit; |
1774 | } |
1775 | |
1776 | rc = bnxt_re_dev_init(rdev, wqe_mode); |
1777 | if (rc) |
1778 | goto re_dev_dealloc; |
1779 | |
1780 | rc = bnxt_re_ib_init(rdev); |
1781 | if (rc) { |
1782 | pr_err("Failed to register with IB: %s" , |
1783 | aux_priv->aux_dev.name); |
1784 | goto re_dev_uninit; |
1785 | } |
1786 | auxiliary_set_drvdata(auxdev: adev, data: rdev); |
1787 | |
1788 | return 0; |
1789 | |
1790 | re_dev_uninit: |
1791 | bnxt_re_dev_uninit(rdev); |
1792 | re_dev_dealloc: |
1793 | ib_dealloc_device(device: &rdev->ibdev); |
1794 | exit: |
1795 | return rc; |
1796 | } |
1797 | |
1798 | static void bnxt_re_setup_cc(struct bnxt_re_dev *rdev, bool enable) |
1799 | { |
1800 | struct bnxt_qplib_cc_param cc_param = {}; |
1801 | |
1802 | /* Do not enable congestion control on VFs */ |
1803 | if (rdev->is_virtfn) |
1804 | return; |
1805 | |
1806 | /* Currently enabling only for GenP5 adapters */ |
1807 | if (!bnxt_qplib_is_chip_gen_p5(cctx: rdev->chip_ctx)) |
1808 | return; |
1809 | |
1810 | if (enable) { |
1811 | cc_param.enable = 1; |
1812 | cc_param.cc_mode = CMDQ_MODIFY_ROCE_CC_CC_MODE_PROBABILISTIC_CC_MODE; |
1813 | } |
1814 | |
1815 | cc_param.mask = (CMDQ_MODIFY_ROCE_CC_MODIFY_MASK_CC_MODE | |
1816 | CMDQ_MODIFY_ROCE_CC_MODIFY_MASK_ENABLE_CC | |
1817 | CMDQ_MODIFY_ROCE_CC_MODIFY_MASK_TOS_ECN); |
1818 | |
1819 | if (bnxt_qplib_modify_cc(res: &rdev->qplib_res, cc_param: &cc_param)) |
1820 | ibdev_err(ibdev: &rdev->ibdev, format: "Failed to setup CC enable = %d\n" , enable); |
1821 | } |
1822 | |
1823 | /* |
1824 | * "Notifier chain callback can be invoked for the same chain from |
1825 | * different CPUs at the same time". |
1826 | * |
1827 | * For cases when the netdev is already present, our call to the |
1828 | * register_netdevice_notifier() will actually get the rtnl_lock() |
1829 | * before sending NETDEV_REGISTER and (if up) NETDEV_UP |
1830 | * events. |
1831 | * |
1832 | * But for cases when the netdev is not already present, the notifier |
1833 | * chain is subjected to be invoked from different CPUs simultaneously. |
1834 | * |
1835 | * This is protected by the netdev_mutex. |
1836 | */ |
1837 | static int bnxt_re_netdev_event(struct notifier_block *notifier, |
1838 | unsigned long event, void *ptr) |
1839 | { |
1840 | struct net_device *real_dev, *netdev = netdev_notifier_info_to_dev(info: ptr); |
1841 | struct bnxt_re_dev *rdev; |
1842 | |
1843 | real_dev = rdma_vlan_dev_real_dev(dev: netdev); |
1844 | if (!real_dev) |
1845 | real_dev = netdev; |
1846 | |
1847 | if (real_dev != netdev) |
1848 | goto exit; |
1849 | |
1850 | rdev = bnxt_re_from_netdev(netdev: real_dev); |
1851 | if (!rdev) |
1852 | return NOTIFY_DONE; |
1853 | |
1854 | |
1855 | switch (event) { |
1856 | case NETDEV_UP: |
1857 | case NETDEV_DOWN: |
1858 | case NETDEV_CHANGE: |
1859 | bnxt_re_dispatch_event(ibdev: &rdev->ibdev, NULL, port_num: 1, |
1860 | event: netif_carrier_ok(dev: real_dev) ? |
1861 | IB_EVENT_PORT_ACTIVE : |
1862 | IB_EVENT_PORT_ERR); |
1863 | break; |
1864 | default: |
1865 | break; |
1866 | } |
1867 | ib_device_put(device: &rdev->ibdev); |
1868 | exit: |
1869 | return NOTIFY_DONE; |
1870 | } |
1871 | |
1872 | #define BNXT_ADEV_NAME "bnxt_en" |
1873 | |
1874 | static void bnxt_re_remove(struct auxiliary_device *adev) |
1875 | { |
1876 | struct bnxt_re_dev *rdev = auxiliary_get_drvdata(auxdev: adev); |
1877 | |
1878 | if (!rdev) |
1879 | return; |
1880 | |
1881 | mutex_lock(&bnxt_re_mutex); |
1882 | if (rdev->nb.notifier_call) { |
1883 | unregister_netdevice_notifier(nb: &rdev->nb); |
1884 | rdev->nb.notifier_call = NULL; |
1885 | } else { |
1886 | /* If notifier is null, we should have already done a |
1887 | * clean up before coming here. |
1888 | */ |
1889 | goto skip_remove; |
1890 | } |
1891 | bnxt_re_setup_cc(rdev, enable: false); |
1892 | ib_unregister_device(device: &rdev->ibdev); |
1893 | bnxt_re_dev_uninit(rdev); |
1894 | ib_dealloc_device(device: &rdev->ibdev); |
1895 | skip_remove: |
1896 | mutex_unlock(lock: &bnxt_re_mutex); |
1897 | } |
1898 | |
1899 | static int bnxt_re_probe(struct auxiliary_device *adev, |
1900 | const struct auxiliary_device_id *id) |
1901 | { |
1902 | struct bnxt_re_dev *rdev; |
1903 | int rc; |
1904 | |
1905 | mutex_lock(&bnxt_re_mutex); |
1906 | rc = bnxt_re_add_device(adev, wqe_mode: BNXT_QPLIB_WQE_MODE_STATIC); |
1907 | if (rc) { |
1908 | mutex_unlock(lock: &bnxt_re_mutex); |
1909 | return rc; |
1910 | } |
1911 | |
1912 | rdev = auxiliary_get_drvdata(auxdev: adev); |
1913 | |
1914 | rdev->nb.notifier_call = bnxt_re_netdev_event; |
1915 | rc = register_netdevice_notifier(nb: &rdev->nb); |
1916 | if (rc) { |
1917 | rdev->nb.notifier_call = NULL; |
1918 | pr_err("%s: Cannot register to netdevice_notifier" , |
1919 | ROCE_DRV_MODULE_NAME); |
1920 | goto err; |
1921 | } |
1922 | |
1923 | bnxt_re_setup_cc(rdev, enable: true); |
1924 | mutex_unlock(lock: &bnxt_re_mutex); |
1925 | return 0; |
1926 | |
1927 | err: |
1928 | mutex_unlock(lock: &bnxt_re_mutex); |
1929 | bnxt_re_remove(adev); |
1930 | |
1931 | return rc; |
1932 | } |
1933 | |
1934 | static int bnxt_re_suspend(struct auxiliary_device *adev, pm_message_t state) |
1935 | { |
1936 | struct bnxt_re_dev *rdev = auxiliary_get_drvdata(auxdev: adev); |
1937 | |
1938 | if (!rdev) |
1939 | return 0; |
1940 | |
1941 | mutex_lock(&bnxt_re_mutex); |
1942 | /* L2 driver may invoke this callback during device error/crash or device |
1943 | * reset. Current RoCE driver doesn't recover the device in case of |
1944 | * error. Handle the error by dispatching fatal events to all qps |
1945 | * ie. by calling bnxt_re_dev_stop and release the MSIx vectors as |
1946 | * L2 driver want to modify the MSIx table. |
1947 | */ |
1948 | |
1949 | ibdev_info(ibdev: &rdev->ibdev, format: "Handle device suspend call" ); |
1950 | /* Check the current device state from bnxt_en_dev and move the |
1951 | * device to detached state if FW_FATAL_COND is set. |
1952 | * This prevents more commands to HW during clean-up, |
1953 | * in case the device is already in error. |
1954 | */ |
1955 | if (test_bit(BNXT_STATE_FW_FATAL_COND, &rdev->en_dev->en_state)) |
1956 | set_bit(ERR_DEVICE_DETACHED, addr: &rdev->rcfw.cmdq.flags); |
1957 | |
1958 | bnxt_re_dev_stop(rdev); |
1959 | bnxt_re_stop_irq(handle: rdev); |
1960 | /* Move the device states to detached and avoid sending any more |
1961 | * commands to HW |
1962 | */ |
1963 | set_bit(BNXT_RE_FLAG_ERR_DEVICE_DETACHED, addr: &rdev->flags); |
1964 | set_bit(ERR_DEVICE_DETACHED, addr: &rdev->rcfw.cmdq.flags); |
1965 | wake_up_all(&rdev->rcfw.cmdq.waitq); |
1966 | mutex_unlock(lock: &bnxt_re_mutex); |
1967 | |
1968 | return 0; |
1969 | } |
1970 | |
1971 | static int bnxt_re_resume(struct auxiliary_device *adev) |
1972 | { |
1973 | struct bnxt_re_dev *rdev = auxiliary_get_drvdata(auxdev: adev); |
1974 | |
1975 | if (!rdev) |
1976 | return 0; |
1977 | |
1978 | mutex_lock(&bnxt_re_mutex); |
1979 | /* L2 driver may invoke this callback during device recovery, resume. |
1980 | * reset. Current RoCE driver doesn't recover the device in case of |
1981 | * error. Handle the error by dispatching fatal events to all qps |
1982 | * ie. by calling bnxt_re_dev_stop and release the MSIx vectors as |
1983 | * L2 driver want to modify the MSIx table. |
1984 | */ |
1985 | |
1986 | ibdev_info(ibdev: &rdev->ibdev, format: "Handle device resume call" ); |
1987 | mutex_unlock(lock: &bnxt_re_mutex); |
1988 | |
1989 | return 0; |
1990 | } |
1991 | |
1992 | static const struct auxiliary_device_id bnxt_re_id_table[] = { |
1993 | { .name = BNXT_ADEV_NAME ".rdma" , }, |
1994 | {}, |
1995 | }; |
1996 | |
1997 | MODULE_DEVICE_TABLE(auxiliary, bnxt_re_id_table); |
1998 | |
1999 | static struct auxiliary_driver bnxt_re_driver = { |
2000 | .name = "rdma" , |
2001 | .probe = bnxt_re_probe, |
2002 | .remove = bnxt_re_remove, |
2003 | .shutdown = bnxt_re_shutdown, |
2004 | .suspend = bnxt_re_suspend, |
2005 | .resume = bnxt_re_resume, |
2006 | .id_table = bnxt_re_id_table, |
2007 | }; |
2008 | |
2009 | static int __init bnxt_re_mod_init(void) |
2010 | { |
2011 | int rc; |
2012 | |
2013 | pr_info("%s: %s" , ROCE_DRV_MODULE_NAME, version); |
2014 | rc = auxiliary_driver_register(&bnxt_re_driver); |
2015 | if (rc) { |
2016 | pr_err("%s: Failed to register auxiliary driver\n" , |
2017 | ROCE_DRV_MODULE_NAME); |
2018 | return rc; |
2019 | } |
2020 | return 0; |
2021 | } |
2022 | |
2023 | static void __exit bnxt_re_mod_exit(void) |
2024 | { |
2025 | auxiliary_driver_unregister(auxdrv: &bnxt_re_driver); |
2026 | } |
2027 | |
2028 | module_init(bnxt_re_mod_init); |
2029 | module_exit(bnxt_re_mod_exit); |
2030 | |