1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
2 /* Copyright (c) 2021, Microsoft Corporation. */
3
4 #include <linux/debugfs.h>
5 #include <linux/module.h>
6 #include <linux/pci.h>
7 #include <linux/utsname.h>
8 #include <linux/version.h>
9
10 #include <net/mana/mana.h>
11
12 struct dentry *mana_debugfs_root;
13
mana_gd_r32(struct gdma_context * g,u64 offset)14 static u32 mana_gd_r32(struct gdma_context *g, u64 offset)
15 {
16 return readl(g->bar0_va + offset);
17 }
18
mana_gd_r64(struct gdma_context * g,u64 offset)19 static u64 mana_gd_r64(struct gdma_context *g, u64 offset)
20 {
21 return readq(g->bar0_va + offset);
22 }
23
mana_gd_init_pf_regs(struct pci_dev * pdev)24 static void mana_gd_init_pf_regs(struct pci_dev *pdev)
25 {
26 struct gdma_context *gc = pci_get_drvdata(pdev);
27 void __iomem *sriov_base_va;
28 u64 sriov_base_off;
29
30 gc->db_page_size = mana_gd_r32(gc, GDMA_PF_REG_DB_PAGE_SIZE) & 0xFFFF;
31 gc->db_page_base = gc->bar0_va +
32 mana_gd_r64(gc, GDMA_PF_REG_DB_PAGE_OFF);
33
34 sriov_base_off = mana_gd_r64(gc, GDMA_SRIOV_REG_CFG_BASE_OFF);
35
36 sriov_base_va = gc->bar0_va + sriov_base_off;
37 gc->shm_base = sriov_base_va +
38 mana_gd_r64(gc, sriov_base_off + GDMA_PF_REG_SHM_OFF);
39 }
40
mana_gd_init_vf_regs(struct pci_dev * pdev)41 static void mana_gd_init_vf_regs(struct pci_dev *pdev)
42 {
43 struct gdma_context *gc = pci_get_drvdata(pdev);
44
45 gc->db_page_size = mana_gd_r32(gc, GDMA_REG_DB_PAGE_SIZE) & 0xFFFF;
46
47 gc->db_page_base = gc->bar0_va +
48 mana_gd_r64(gc, GDMA_REG_DB_PAGE_OFFSET);
49
50 gc->phys_db_page_base = gc->bar0_pa +
51 mana_gd_r64(gc, GDMA_REG_DB_PAGE_OFFSET);
52
53 gc->shm_base = gc->bar0_va + mana_gd_r64(gc, GDMA_REG_SHM_OFFSET);
54 }
55
mana_gd_init_registers(struct pci_dev * pdev)56 static void mana_gd_init_registers(struct pci_dev *pdev)
57 {
58 struct gdma_context *gc = pci_get_drvdata(pdev);
59
60 if (gc->is_pf)
61 mana_gd_init_pf_regs(pdev);
62 else
63 mana_gd_init_vf_regs(pdev);
64 }
65
mana_gd_query_max_resources(struct pci_dev * pdev)66 static int mana_gd_query_max_resources(struct pci_dev *pdev)
67 {
68 struct gdma_context *gc = pci_get_drvdata(pdev);
69 struct gdma_query_max_resources_resp resp = {};
70 struct gdma_general_req req = {};
71 int err;
72
73 mana_gd_init_req_hdr(&req.hdr, GDMA_QUERY_MAX_RESOURCES,
74 sizeof(req), sizeof(resp));
75
76 err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
77 if (err || resp.hdr.status) {
78 dev_err(gc->dev, "Failed to query resource info: %d, 0x%x\n",
79 err, resp.hdr.status);
80 return err ? err : -EPROTO;
81 }
82
83 if (gc->num_msix_usable > resp.max_msix)
84 gc->num_msix_usable = resp.max_msix;
85
86 if (gc->num_msix_usable <= 1)
87 return -ENOSPC;
88
89 gc->max_num_queues = num_online_cpus();
90 if (gc->max_num_queues > MANA_MAX_NUM_QUEUES)
91 gc->max_num_queues = MANA_MAX_NUM_QUEUES;
92
93 if (gc->max_num_queues > resp.max_eq)
94 gc->max_num_queues = resp.max_eq;
95
96 if (gc->max_num_queues > resp.max_cq)
97 gc->max_num_queues = resp.max_cq;
98
99 if (gc->max_num_queues > resp.max_sq)
100 gc->max_num_queues = resp.max_sq;
101
102 if (gc->max_num_queues > resp.max_rq)
103 gc->max_num_queues = resp.max_rq;
104
105 /* The Hardware Channel (HWC) used 1 MSI-X */
106 if (gc->max_num_queues > gc->num_msix_usable - 1)
107 gc->max_num_queues = gc->num_msix_usable - 1;
108
109 return 0;
110 }
111
mana_gd_query_hwc_timeout(struct pci_dev * pdev,u32 * timeout_val)112 static int mana_gd_query_hwc_timeout(struct pci_dev *pdev, u32 *timeout_val)
113 {
114 struct gdma_context *gc = pci_get_drvdata(pdev);
115 struct gdma_query_hwc_timeout_resp resp = {};
116 struct gdma_query_hwc_timeout_req req = {};
117 int err;
118
119 mana_gd_init_req_hdr(&req.hdr, GDMA_QUERY_HWC_TIMEOUT,
120 sizeof(req), sizeof(resp));
121 req.timeout_ms = *timeout_val;
122 err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
123 if (err || resp.hdr.status)
124 return err ? err : -EPROTO;
125
126 *timeout_val = resp.timeout_ms;
127
128 return 0;
129 }
130
mana_gd_detect_devices(struct pci_dev * pdev)131 static int mana_gd_detect_devices(struct pci_dev *pdev)
132 {
133 struct gdma_context *gc = pci_get_drvdata(pdev);
134 struct gdma_list_devices_resp resp = {};
135 struct gdma_general_req req = {};
136 struct gdma_dev_id dev;
137 u32 i, max_num_devs;
138 u16 dev_type;
139 int err;
140
141 mana_gd_init_req_hdr(&req.hdr, GDMA_LIST_DEVICES, sizeof(req),
142 sizeof(resp));
143
144 err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
145 if (err || resp.hdr.status) {
146 dev_err(gc->dev, "Failed to detect devices: %d, 0x%x\n", err,
147 resp.hdr.status);
148 return err ? err : -EPROTO;
149 }
150
151 max_num_devs = min_t(u32, MAX_NUM_GDMA_DEVICES, resp.num_of_devs);
152
153 for (i = 0; i < max_num_devs; i++) {
154 dev = resp.devs[i];
155 dev_type = dev.type;
156
157 /* HWC is already detected in mana_hwc_create_channel(). */
158 if (dev_type == GDMA_DEVICE_HWC)
159 continue;
160
161 if (dev_type == GDMA_DEVICE_MANA) {
162 gc->mana.gdma_context = gc;
163 gc->mana.dev_id = dev;
164 } else if (dev_type == GDMA_DEVICE_MANA_IB) {
165 gc->mana_ib.dev_id = dev;
166 gc->mana_ib.gdma_context = gc;
167 }
168 }
169
170 return gc->mana.dev_id.type == 0 ? -ENODEV : 0;
171 }
172
mana_gd_send_request(struct gdma_context * gc,u32 req_len,const void * req,u32 resp_len,void * resp)173 int mana_gd_send_request(struct gdma_context *gc, u32 req_len, const void *req,
174 u32 resp_len, void *resp)
175 {
176 struct hw_channel_context *hwc = gc->hwc.driver_data;
177
178 return mana_hwc_send_request(hwc, req_len, req, resp_len, resp);
179 }
180 EXPORT_SYMBOL_NS(mana_gd_send_request, "NET_MANA");
181
mana_gd_alloc_memory(struct gdma_context * gc,unsigned int length,struct gdma_mem_info * gmi)182 int mana_gd_alloc_memory(struct gdma_context *gc, unsigned int length,
183 struct gdma_mem_info *gmi)
184 {
185 dma_addr_t dma_handle;
186 void *buf;
187
188 if (length < MANA_PAGE_SIZE || !is_power_of_2(length))
189 return -EINVAL;
190
191 gmi->dev = gc->dev;
192 buf = dma_alloc_coherent(gmi->dev, length, &dma_handle, GFP_KERNEL);
193 if (!buf)
194 return -ENOMEM;
195
196 gmi->dma_handle = dma_handle;
197 gmi->virt_addr = buf;
198 gmi->length = length;
199
200 return 0;
201 }
202
mana_gd_free_memory(struct gdma_mem_info * gmi)203 void mana_gd_free_memory(struct gdma_mem_info *gmi)
204 {
205 dma_free_coherent(gmi->dev, gmi->length, gmi->virt_addr,
206 gmi->dma_handle);
207 }
208
mana_gd_create_hw_eq(struct gdma_context * gc,struct gdma_queue * queue)209 static int mana_gd_create_hw_eq(struct gdma_context *gc,
210 struct gdma_queue *queue)
211 {
212 struct gdma_create_queue_resp resp = {};
213 struct gdma_create_queue_req req = {};
214 int err;
215
216 if (queue->type != GDMA_EQ)
217 return -EINVAL;
218
219 mana_gd_init_req_hdr(&req.hdr, GDMA_CREATE_QUEUE,
220 sizeof(req), sizeof(resp));
221
222 req.hdr.dev_id = queue->gdma_dev->dev_id;
223 req.type = queue->type;
224 req.pdid = queue->gdma_dev->pdid;
225 req.doolbell_id = queue->gdma_dev->doorbell;
226 req.gdma_region = queue->mem_info.dma_region_handle;
227 req.queue_size = queue->queue_size;
228 req.log2_throttle_limit = queue->eq.log2_throttle_limit;
229 req.eq_pci_msix_index = queue->eq.msix_index;
230
231 err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
232 if (err || resp.hdr.status) {
233 dev_err(gc->dev, "Failed to create queue: %d, 0x%x\n", err,
234 resp.hdr.status);
235 return err ? err : -EPROTO;
236 }
237
238 queue->id = resp.queue_index;
239 queue->eq.disable_needed = true;
240 queue->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION;
241 return 0;
242 }
243
mana_gd_disable_queue(struct gdma_queue * queue)244 static int mana_gd_disable_queue(struct gdma_queue *queue)
245 {
246 struct gdma_context *gc = queue->gdma_dev->gdma_context;
247 struct gdma_disable_queue_req req = {};
248 struct gdma_general_resp resp = {};
249 int err;
250
251 WARN_ON(queue->type != GDMA_EQ);
252
253 mana_gd_init_req_hdr(&req.hdr, GDMA_DISABLE_QUEUE,
254 sizeof(req), sizeof(resp));
255
256 req.hdr.dev_id = queue->gdma_dev->dev_id;
257 req.type = queue->type;
258 req.queue_index = queue->id;
259 req.alloc_res_id_on_creation = 1;
260
261 err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
262 if (err || resp.hdr.status) {
263 dev_err(gc->dev, "Failed to disable queue: %d, 0x%x\n", err,
264 resp.hdr.status);
265 return err ? err : -EPROTO;
266 }
267
268 return 0;
269 }
270
271 #define DOORBELL_OFFSET_SQ 0x0
272 #define DOORBELL_OFFSET_RQ 0x400
273 #define DOORBELL_OFFSET_CQ 0x800
274 #define DOORBELL_OFFSET_EQ 0xFF8
275
mana_gd_ring_doorbell(struct gdma_context * gc,u32 db_index,enum gdma_queue_type q_type,u32 qid,u32 tail_ptr,u8 num_req)276 static void mana_gd_ring_doorbell(struct gdma_context *gc, u32 db_index,
277 enum gdma_queue_type q_type, u32 qid,
278 u32 tail_ptr, u8 num_req)
279 {
280 void __iomem *addr = gc->db_page_base + gc->db_page_size * db_index;
281 union gdma_doorbell_entry e = {};
282
283 switch (q_type) {
284 case GDMA_EQ:
285 e.eq.id = qid;
286 e.eq.tail_ptr = tail_ptr;
287 e.eq.arm = num_req;
288
289 addr += DOORBELL_OFFSET_EQ;
290 break;
291
292 case GDMA_CQ:
293 e.cq.id = qid;
294 e.cq.tail_ptr = tail_ptr;
295 e.cq.arm = num_req;
296
297 addr += DOORBELL_OFFSET_CQ;
298 break;
299
300 case GDMA_RQ:
301 e.rq.id = qid;
302 e.rq.tail_ptr = tail_ptr;
303 e.rq.wqe_cnt = num_req;
304
305 addr += DOORBELL_OFFSET_RQ;
306 break;
307
308 case GDMA_SQ:
309 e.sq.id = qid;
310 e.sq.tail_ptr = tail_ptr;
311
312 addr += DOORBELL_OFFSET_SQ;
313 break;
314
315 default:
316 WARN_ON(1);
317 return;
318 }
319
320 /* Ensure all writes are done before ring doorbell */
321 wmb();
322
323 writeq(e.as_uint64, addr);
324 }
325
mana_gd_wq_ring_doorbell(struct gdma_context * gc,struct gdma_queue * queue)326 void mana_gd_wq_ring_doorbell(struct gdma_context *gc, struct gdma_queue *queue)
327 {
328 /* Hardware Spec specifies that software client should set 0 for
329 * wqe_cnt for Receive Queues. This value is not used in Send Queues.
330 */
331 mana_gd_ring_doorbell(gc, queue->gdma_dev->doorbell, queue->type,
332 queue->id, queue->head * GDMA_WQE_BU_SIZE, 0);
333 }
334
mana_gd_ring_cq(struct gdma_queue * cq,u8 arm_bit)335 void mana_gd_ring_cq(struct gdma_queue *cq, u8 arm_bit)
336 {
337 struct gdma_context *gc = cq->gdma_dev->gdma_context;
338
339 u32 num_cqe = cq->queue_size / GDMA_CQE_SIZE;
340
341 u32 head = cq->head % (num_cqe << GDMA_CQE_OWNER_BITS);
342
343 mana_gd_ring_doorbell(gc, cq->gdma_dev->doorbell, cq->type, cq->id,
344 head, arm_bit);
345 }
346
mana_gd_process_eqe(struct gdma_queue * eq)347 static void mana_gd_process_eqe(struct gdma_queue *eq)
348 {
349 u32 head = eq->head % (eq->queue_size / GDMA_EQE_SIZE);
350 struct gdma_context *gc = eq->gdma_dev->gdma_context;
351 struct gdma_eqe *eq_eqe_ptr = eq->queue_mem_ptr;
352 union gdma_eqe_info eqe_info;
353 enum gdma_eqe_type type;
354 struct gdma_event event;
355 struct gdma_queue *cq;
356 struct gdma_eqe *eqe;
357 u32 cq_id;
358
359 eqe = &eq_eqe_ptr[head];
360 eqe_info.as_uint32 = eqe->eqe_info;
361 type = eqe_info.type;
362
363 switch (type) {
364 case GDMA_EQE_COMPLETION:
365 cq_id = eqe->details[0] & 0xFFFFFF;
366 if (WARN_ON_ONCE(cq_id >= gc->max_num_cqs))
367 break;
368
369 cq = gc->cq_table[cq_id];
370 if (WARN_ON_ONCE(!cq || cq->type != GDMA_CQ || cq->id != cq_id))
371 break;
372
373 if (cq->cq.callback)
374 cq->cq.callback(cq->cq.context, cq);
375
376 break;
377
378 case GDMA_EQE_TEST_EVENT:
379 gc->test_event_eq_id = eq->id;
380 complete(&gc->eq_test_event);
381 break;
382
383 case GDMA_EQE_HWC_INIT_EQ_ID_DB:
384 case GDMA_EQE_HWC_INIT_DATA:
385 case GDMA_EQE_HWC_INIT_DONE:
386 case GDMA_EQE_RNIC_QP_FATAL:
387 if (!eq->eq.callback)
388 break;
389
390 event.type = type;
391 memcpy(&event.details, &eqe->details, GDMA_EVENT_DATA_SIZE);
392 eq->eq.callback(eq->eq.context, eq, &event);
393 break;
394
395 default:
396 break;
397 }
398 }
399
mana_gd_process_eq_events(void * arg)400 static void mana_gd_process_eq_events(void *arg)
401 {
402 u32 owner_bits, new_bits, old_bits;
403 union gdma_eqe_info eqe_info;
404 struct gdma_eqe *eq_eqe_ptr;
405 struct gdma_queue *eq = arg;
406 struct gdma_context *gc;
407 struct gdma_eqe *eqe;
408 u32 head, num_eqe;
409 int i;
410
411 gc = eq->gdma_dev->gdma_context;
412
413 num_eqe = eq->queue_size / GDMA_EQE_SIZE;
414 eq_eqe_ptr = eq->queue_mem_ptr;
415
416 /* Process up to 5 EQEs at a time, and update the HW head. */
417 for (i = 0; i < 5; i++) {
418 eqe = &eq_eqe_ptr[eq->head % num_eqe];
419 eqe_info.as_uint32 = eqe->eqe_info;
420 owner_bits = eqe_info.owner_bits;
421
422 old_bits = (eq->head / num_eqe - 1) & GDMA_EQE_OWNER_MASK;
423 /* No more entries */
424 if (owner_bits == old_bits) {
425 /* return here without ringing the doorbell */
426 if (i == 0)
427 return;
428 break;
429 }
430
431 new_bits = (eq->head / num_eqe) & GDMA_EQE_OWNER_MASK;
432 if (owner_bits != new_bits) {
433 dev_err(gc->dev, "EQ %d: overflow detected\n", eq->id);
434 break;
435 }
436
437 /* Per GDMA spec, rmb is necessary after checking owner_bits, before
438 * reading eqe.
439 */
440 rmb();
441
442 mana_gd_process_eqe(eq);
443
444 eq->head++;
445 }
446
447 head = eq->head % (num_eqe << GDMA_EQE_OWNER_BITS);
448
449 mana_gd_ring_doorbell(gc, eq->gdma_dev->doorbell, eq->type, eq->id,
450 head, SET_ARM_BIT);
451 }
452
mana_gd_register_irq(struct gdma_queue * queue,const struct gdma_queue_spec * spec)453 static int mana_gd_register_irq(struct gdma_queue *queue,
454 const struct gdma_queue_spec *spec)
455 {
456 struct gdma_dev *gd = queue->gdma_dev;
457 struct gdma_irq_context *gic;
458 struct gdma_context *gc;
459 unsigned int msi_index;
460 unsigned long flags;
461 struct device *dev;
462 int err = 0;
463
464 gc = gd->gdma_context;
465 dev = gc->dev;
466 msi_index = spec->eq.msix_index;
467
468 if (msi_index >= gc->num_msix_usable) {
469 err = -ENOSPC;
470 dev_err(dev, "Register IRQ err:%d, msi:%u nMSI:%u",
471 err, msi_index, gc->num_msix_usable);
472
473 return err;
474 }
475
476 queue->eq.msix_index = msi_index;
477 gic = &gc->irq_contexts[msi_index];
478
479 spin_lock_irqsave(&gic->lock, flags);
480 list_add_rcu(&queue->entry, &gic->eq_list);
481 spin_unlock_irqrestore(&gic->lock, flags);
482
483 return 0;
484 }
485
mana_gd_deregiser_irq(struct gdma_queue * queue)486 static void mana_gd_deregiser_irq(struct gdma_queue *queue)
487 {
488 struct gdma_dev *gd = queue->gdma_dev;
489 struct gdma_irq_context *gic;
490 struct gdma_context *gc;
491 unsigned int msix_index;
492 unsigned long flags;
493 struct gdma_queue *eq;
494
495 gc = gd->gdma_context;
496
497 /* At most num_online_cpus() + 1 interrupts are used. */
498 msix_index = queue->eq.msix_index;
499 if (WARN_ON(msix_index >= gc->num_msix_usable))
500 return;
501
502 gic = &gc->irq_contexts[msix_index];
503 spin_lock_irqsave(&gic->lock, flags);
504 list_for_each_entry_rcu(eq, &gic->eq_list, entry) {
505 if (queue == eq) {
506 list_del_rcu(&eq->entry);
507 break;
508 }
509 }
510 spin_unlock_irqrestore(&gic->lock, flags);
511
512 queue->eq.msix_index = INVALID_PCI_MSIX_INDEX;
513 synchronize_rcu();
514 }
515
mana_gd_test_eq(struct gdma_context * gc,struct gdma_queue * eq)516 int mana_gd_test_eq(struct gdma_context *gc, struct gdma_queue *eq)
517 {
518 struct gdma_generate_test_event_req req = {};
519 struct gdma_general_resp resp = {};
520 struct device *dev = gc->dev;
521 int err;
522
523 mutex_lock(&gc->eq_test_event_mutex);
524
525 init_completion(&gc->eq_test_event);
526 gc->test_event_eq_id = INVALID_QUEUE_ID;
527
528 mana_gd_init_req_hdr(&req.hdr, GDMA_GENERATE_TEST_EQE,
529 sizeof(req), sizeof(resp));
530
531 req.hdr.dev_id = eq->gdma_dev->dev_id;
532 req.queue_index = eq->id;
533
534 err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
535 if (err) {
536 dev_err(dev, "test_eq failed: %d\n", err);
537 goto out;
538 }
539
540 err = -EPROTO;
541
542 if (resp.hdr.status) {
543 dev_err(dev, "test_eq failed: 0x%x\n", resp.hdr.status);
544 goto out;
545 }
546
547 if (!wait_for_completion_timeout(&gc->eq_test_event, 30 * HZ)) {
548 dev_err(dev, "test_eq timed out on queue %d\n", eq->id);
549 goto out;
550 }
551
552 if (eq->id != gc->test_event_eq_id) {
553 dev_err(dev, "test_eq got an event on wrong queue %d (%d)\n",
554 gc->test_event_eq_id, eq->id);
555 goto out;
556 }
557
558 err = 0;
559 out:
560 mutex_unlock(&gc->eq_test_event_mutex);
561 return err;
562 }
563
mana_gd_destroy_eq(struct gdma_context * gc,bool flush_evenets,struct gdma_queue * queue)564 static void mana_gd_destroy_eq(struct gdma_context *gc, bool flush_evenets,
565 struct gdma_queue *queue)
566 {
567 int err;
568
569 if (flush_evenets) {
570 err = mana_gd_test_eq(gc, queue);
571 if (err)
572 dev_warn(gc->dev, "Failed to flush EQ: %d\n", err);
573 }
574
575 mana_gd_deregiser_irq(queue);
576
577 if (queue->eq.disable_needed)
578 mana_gd_disable_queue(queue);
579 }
580
mana_gd_create_eq(struct gdma_dev * gd,const struct gdma_queue_spec * spec,bool create_hwq,struct gdma_queue * queue)581 static int mana_gd_create_eq(struct gdma_dev *gd,
582 const struct gdma_queue_spec *spec,
583 bool create_hwq, struct gdma_queue *queue)
584 {
585 struct gdma_context *gc = gd->gdma_context;
586 struct device *dev = gc->dev;
587 u32 log2_num_entries;
588 int err;
589
590 queue->eq.msix_index = INVALID_PCI_MSIX_INDEX;
591 queue->id = INVALID_QUEUE_ID;
592
593 log2_num_entries = ilog2(queue->queue_size / GDMA_EQE_SIZE);
594
595 if (spec->eq.log2_throttle_limit > log2_num_entries) {
596 dev_err(dev, "EQ throttling limit (%lu) > maximum EQE (%u)\n",
597 spec->eq.log2_throttle_limit, log2_num_entries);
598 return -EINVAL;
599 }
600
601 err = mana_gd_register_irq(queue, spec);
602 if (err) {
603 dev_err(dev, "Failed to register irq: %d\n", err);
604 return err;
605 }
606
607 queue->eq.callback = spec->eq.callback;
608 queue->eq.context = spec->eq.context;
609 queue->head |= INITIALIZED_OWNER_BIT(log2_num_entries);
610 queue->eq.log2_throttle_limit = spec->eq.log2_throttle_limit ?: 1;
611
612 if (create_hwq) {
613 err = mana_gd_create_hw_eq(gc, queue);
614 if (err)
615 goto out;
616
617 err = mana_gd_test_eq(gc, queue);
618 if (err)
619 goto out;
620 }
621
622 return 0;
623 out:
624 dev_err(dev, "Failed to create EQ: %d\n", err);
625 mana_gd_destroy_eq(gc, false, queue);
626 return err;
627 }
628
mana_gd_create_cq(const struct gdma_queue_spec * spec,struct gdma_queue * queue)629 static void mana_gd_create_cq(const struct gdma_queue_spec *spec,
630 struct gdma_queue *queue)
631 {
632 u32 log2_num_entries = ilog2(spec->queue_size / GDMA_CQE_SIZE);
633
634 queue->head |= INITIALIZED_OWNER_BIT(log2_num_entries);
635 queue->cq.parent = spec->cq.parent_eq;
636 queue->cq.context = spec->cq.context;
637 queue->cq.callback = spec->cq.callback;
638 }
639
mana_gd_destroy_cq(struct gdma_context * gc,struct gdma_queue * queue)640 static void mana_gd_destroy_cq(struct gdma_context *gc,
641 struct gdma_queue *queue)
642 {
643 u32 id = queue->id;
644
645 if (id >= gc->max_num_cqs)
646 return;
647
648 if (!gc->cq_table[id])
649 return;
650
651 gc->cq_table[id] = NULL;
652 }
653
mana_gd_create_hwc_queue(struct gdma_dev * gd,const struct gdma_queue_spec * spec,struct gdma_queue ** queue_ptr)654 int mana_gd_create_hwc_queue(struct gdma_dev *gd,
655 const struct gdma_queue_spec *spec,
656 struct gdma_queue **queue_ptr)
657 {
658 struct gdma_context *gc = gd->gdma_context;
659 struct gdma_mem_info *gmi;
660 struct gdma_queue *queue;
661 int err;
662
663 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
664 if (!queue)
665 return -ENOMEM;
666
667 gmi = &queue->mem_info;
668 err = mana_gd_alloc_memory(gc, spec->queue_size, gmi);
669 if (err)
670 goto free_q;
671
672 queue->head = 0;
673 queue->tail = 0;
674 queue->queue_mem_ptr = gmi->virt_addr;
675 queue->queue_size = spec->queue_size;
676 queue->monitor_avl_buf = spec->monitor_avl_buf;
677 queue->type = spec->type;
678 queue->gdma_dev = gd;
679
680 if (spec->type == GDMA_EQ)
681 err = mana_gd_create_eq(gd, spec, false, queue);
682 else if (spec->type == GDMA_CQ)
683 mana_gd_create_cq(spec, queue);
684
685 if (err)
686 goto out;
687
688 *queue_ptr = queue;
689 return 0;
690 out:
691 mana_gd_free_memory(gmi);
692 free_q:
693 kfree(queue);
694 return err;
695 }
696
mana_gd_destroy_dma_region(struct gdma_context * gc,u64 dma_region_handle)697 int mana_gd_destroy_dma_region(struct gdma_context *gc, u64 dma_region_handle)
698 {
699 struct gdma_destroy_dma_region_req req = {};
700 struct gdma_general_resp resp = {};
701 int err;
702
703 if (dma_region_handle == GDMA_INVALID_DMA_REGION)
704 return 0;
705
706 mana_gd_init_req_hdr(&req.hdr, GDMA_DESTROY_DMA_REGION, sizeof(req),
707 sizeof(resp));
708 req.dma_region_handle = dma_region_handle;
709
710 err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
711 if (err || resp.hdr.status) {
712 dev_err(gc->dev, "Failed to destroy DMA region: %d, 0x%x\n",
713 err, resp.hdr.status);
714 return -EPROTO;
715 }
716
717 return 0;
718 }
719 EXPORT_SYMBOL_NS(mana_gd_destroy_dma_region, "NET_MANA");
720
mana_gd_create_dma_region(struct gdma_dev * gd,struct gdma_mem_info * gmi)721 static int mana_gd_create_dma_region(struct gdma_dev *gd,
722 struct gdma_mem_info *gmi)
723 {
724 unsigned int num_page = gmi->length / MANA_PAGE_SIZE;
725 struct gdma_create_dma_region_req *req = NULL;
726 struct gdma_create_dma_region_resp resp = {};
727 struct gdma_context *gc = gd->gdma_context;
728 struct hw_channel_context *hwc;
729 u32 length = gmi->length;
730 size_t req_msg_size;
731 int err;
732 int i;
733
734 if (length < MANA_PAGE_SIZE || !is_power_of_2(length))
735 return -EINVAL;
736
737 if (!MANA_PAGE_ALIGNED(gmi->virt_addr))
738 return -EINVAL;
739
740 hwc = gc->hwc.driver_data;
741 req_msg_size = struct_size(req, page_addr_list, num_page);
742 if (req_msg_size > hwc->max_req_msg_size)
743 return -EINVAL;
744
745 req = kzalloc(req_msg_size, GFP_KERNEL);
746 if (!req)
747 return -ENOMEM;
748
749 mana_gd_init_req_hdr(&req->hdr, GDMA_CREATE_DMA_REGION,
750 req_msg_size, sizeof(resp));
751 req->length = length;
752 req->offset_in_page = 0;
753 req->gdma_page_type = GDMA_PAGE_TYPE_4K;
754 req->page_count = num_page;
755 req->page_addr_list_len = num_page;
756
757 for (i = 0; i < num_page; i++)
758 req->page_addr_list[i] = gmi->dma_handle + i * MANA_PAGE_SIZE;
759
760 err = mana_gd_send_request(gc, req_msg_size, req, sizeof(resp), &resp);
761 if (err)
762 goto out;
763
764 if (resp.hdr.status ||
765 resp.dma_region_handle == GDMA_INVALID_DMA_REGION) {
766 dev_err(gc->dev, "Failed to create DMA region: 0x%x\n",
767 resp.hdr.status);
768 err = -EPROTO;
769 goto out;
770 }
771
772 gmi->dma_region_handle = resp.dma_region_handle;
773 out:
774 kfree(req);
775 return err;
776 }
777
mana_gd_create_mana_eq(struct gdma_dev * gd,const struct gdma_queue_spec * spec,struct gdma_queue ** queue_ptr)778 int mana_gd_create_mana_eq(struct gdma_dev *gd,
779 const struct gdma_queue_spec *spec,
780 struct gdma_queue **queue_ptr)
781 {
782 struct gdma_context *gc = gd->gdma_context;
783 struct gdma_mem_info *gmi;
784 struct gdma_queue *queue;
785 int err;
786
787 if (spec->type != GDMA_EQ)
788 return -EINVAL;
789
790 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
791 if (!queue)
792 return -ENOMEM;
793
794 gmi = &queue->mem_info;
795 err = mana_gd_alloc_memory(gc, spec->queue_size, gmi);
796 if (err)
797 goto free_q;
798
799 err = mana_gd_create_dma_region(gd, gmi);
800 if (err)
801 goto out;
802
803 queue->head = 0;
804 queue->tail = 0;
805 queue->queue_mem_ptr = gmi->virt_addr;
806 queue->queue_size = spec->queue_size;
807 queue->monitor_avl_buf = spec->monitor_avl_buf;
808 queue->type = spec->type;
809 queue->gdma_dev = gd;
810
811 err = mana_gd_create_eq(gd, spec, true, queue);
812 if (err)
813 goto out;
814
815 *queue_ptr = queue;
816 return 0;
817 out:
818 mana_gd_free_memory(gmi);
819 free_q:
820 kfree(queue);
821 return err;
822 }
823 EXPORT_SYMBOL_NS(mana_gd_create_mana_eq, "NET_MANA");
824
mana_gd_create_mana_wq_cq(struct gdma_dev * gd,const struct gdma_queue_spec * spec,struct gdma_queue ** queue_ptr)825 int mana_gd_create_mana_wq_cq(struct gdma_dev *gd,
826 const struct gdma_queue_spec *spec,
827 struct gdma_queue **queue_ptr)
828 {
829 struct gdma_context *gc = gd->gdma_context;
830 struct gdma_mem_info *gmi;
831 struct gdma_queue *queue;
832 int err;
833
834 if (spec->type != GDMA_CQ && spec->type != GDMA_SQ &&
835 spec->type != GDMA_RQ)
836 return -EINVAL;
837
838 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
839 if (!queue)
840 return -ENOMEM;
841
842 gmi = &queue->mem_info;
843 err = mana_gd_alloc_memory(gc, spec->queue_size, gmi);
844 if (err)
845 goto free_q;
846
847 err = mana_gd_create_dma_region(gd, gmi);
848 if (err)
849 goto out;
850
851 queue->head = 0;
852 queue->tail = 0;
853 queue->queue_mem_ptr = gmi->virt_addr;
854 queue->queue_size = spec->queue_size;
855 queue->monitor_avl_buf = spec->monitor_avl_buf;
856 queue->type = spec->type;
857 queue->gdma_dev = gd;
858
859 if (spec->type == GDMA_CQ)
860 mana_gd_create_cq(spec, queue);
861
862 *queue_ptr = queue;
863 return 0;
864 out:
865 mana_gd_free_memory(gmi);
866 free_q:
867 kfree(queue);
868 return err;
869 }
870
mana_gd_destroy_queue(struct gdma_context * gc,struct gdma_queue * queue)871 void mana_gd_destroy_queue(struct gdma_context *gc, struct gdma_queue *queue)
872 {
873 struct gdma_mem_info *gmi = &queue->mem_info;
874
875 switch (queue->type) {
876 case GDMA_EQ:
877 mana_gd_destroy_eq(gc, queue->eq.disable_needed, queue);
878 break;
879
880 case GDMA_CQ:
881 mana_gd_destroy_cq(gc, queue);
882 break;
883
884 case GDMA_RQ:
885 break;
886
887 case GDMA_SQ:
888 break;
889
890 default:
891 dev_err(gc->dev, "Can't destroy unknown queue: type=%d\n",
892 queue->type);
893 return;
894 }
895
896 mana_gd_destroy_dma_region(gc, gmi->dma_region_handle);
897 mana_gd_free_memory(gmi);
898 kfree(queue);
899 }
900 EXPORT_SYMBOL_NS(mana_gd_destroy_queue, "NET_MANA");
901
mana_gd_verify_vf_version(struct pci_dev * pdev)902 int mana_gd_verify_vf_version(struct pci_dev *pdev)
903 {
904 struct gdma_context *gc = pci_get_drvdata(pdev);
905 struct gdma_verify_ver_resp resp = {};
906 struct gdma_verify_ver_req req = {};
907 struct hw_channel_context *hwc;
908 int err;
909
910 hwc = gc->hwc.driver_data;
911 mana_gd_init_req_hdr(&req.hdr, GDMA_VERIFY_VF_DRIVER_VERSION,
912 sizeof(req), sizeof(resp));
913
914 req.protocol_ver_min = GDMA_PROTOCOL_FIRST;
915 req.protocol_ver_max = GDMA_PROTOCOL_LAST;
916
917 req.gd_drv_cap_flags1 = GDMA_DRV_CAP_FLAGS1;
918 req.gd_drv_cap_flags2 = GDMA_DRV_CAP_FLAGS2;
919 req.gd_drv_cap_flags3 = GDMA_DRV_CAP_FLAGS3;
920 req.gd_drv_cap_flags4 = GDMA_DRV_CAP_FLAGS4;
921
922 req.drv_ver = 0; /* Unused*/
923 req.os_type = 0x10; /* Linux */
924 req.os_ver_major = LINUX_VERSION_MAJOR;
925 req.os_ver_minor = LINUX_VERSION_PATCHLEVEL;
926 req.os_ver_build = LINUX_VERSION_SUBLEVEL;
927 strscpy(req.os_ver_str1, utsname()->sysname, sizeof(req.os_ver_str1));
928 strscpy(req.os_ver_str2, utsname()->release, sizeof(req.os_ver_str2));
929 strscpy(req.os_ver_str3, utsname()->version, sizeof(req.os_ver_str3));
930
931 err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
932 if (err || resp.hdr.status) {
933 dev_err(gc->dev, "VfVerifyVersionOutput: %d, status=0x%x\n",
934 err, resp.hdr.status);
935 return err ? err : -EPROTO;
936 }
937 if (resp.pf_cap_flags1 & GDMA_DRV_CAP_FLAG_1_HWC_TIMEOUT_RECONFIG) {
938 err = mana_gd_query_hwc_timeout(pdev, &hwc->hwc_timeout);
939 if (err) {
940 dev_err(gc->dev, "Failed to set the hwc timeout %d\n", err);
941 return err;
942 }
943 dev_dbg(gc->dev, "set the hwc timeout to %u\n", hwc->hwc_timeout);
944 }
945 return 0;
946 }
947
mana_gd_register_device(struct gdma_dev * gd)948 int mana_gd_register_device(struct gdma_dev *gd)
949 {
950 struct gdma_context *gc = gd->gdma_context;
951 struct gdma_register_device_resp resp = {};
952 struct gdma_general_req req = {};
953 int err;
954
955 gd->pdid = INVALID_PDID;
956 gd->doorbell = INVALID_DOORBELL;
957 gd->gpa_mkey = INVALID_MEM_KEY;
958
959 mana_gd_init_req_hdr(&req.hdr, GDMA_REGISTER_DEVICE, sizeof(req),
960 sizeof(resp));
961
962 req.hdr.dev_id = gd->dev_id;
963
964 err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
965 if (err || resp.hdr.status) {
966 dev_err(gc->dev, "gdma_register_device_resp failed: %d, 0x%x\n",
967 err, resp.hdr.status);
968 return err ? err : -EPROTO;
969 }
970
971 gd->pdid = resp.pdid;
972 gd->gpa_mkey = resp.gpa_mkey;
973 gd->doorbell = resp.db_id;
974
975 return 0;
976 }
977 EXPORT_SYMBOL_NS(mana_gd_register_device, "NET_MANA");
978
mana_gd_deregister_device(struct gdma_dev * gd)979 int mana_gd_deregister_device(struct gdma_dev *gd)
980 {
981 struct gdma_context *gc = gd->gdma_context;
982 struct gdma_general_resp resp = {};
983 struct gdma_general_req req = {};
984 int err;
985
986 if (gd->pdid == INVALID_PDID)
987 return -EINVAL;
988
989 mana_gd_init_req_hdr(&req.hdr, GDMA_DEREGISTER_DEVICE, sizeof(req),
990 sizeof(resp));
991
992 req.hdr.dev_id = gd->dev_id;
993
994 err = mana_gd_send_request(gc, sizeof(req), &req, sizeof(resp), &resp);
995 if (err || resp.hdr.status) {
996 dev_err(gc->dev, "Failed to deregister device: %d, 0x%x\n",
997 err, resp.hdr.status);
998 if (!err)
999 err = -EPROTO;
1000 }
1001
1002 gd->pdid = INVALID_PDID;
1003 gd->doorbell = INVALID_DOORBELL;
1004 gd->gpa_mkey = INVALID_MEM_KEY;
1005
1006 return err;
1007 }
1008 EXPORT_SYMBOL_NS(mana_gd_deregister_device, "NET_MANA");
1009
mana_gd_wq_avail_space(struct gdma_queue * wq)1010 u32 mana_gd_wq_avail_space(struct gdma_queue *wq)
1011 {
1012 u32 used_space = (wq->head - wq->tail) * GDMA_WQE_BU_SIZE;
1013 u32 wq_size = wq->queue_size;
1014
1015 WARN_ON_ONCE(used_space > wq_size);
1016
1017 return wq_size - used_space;
1018 }
1019
mana_gd_get_wqe_ptr(const struct gdma_queue * wq,u32 wqe_offset)1020 u8 *mana_gd_get_wqe_ptr(const struct gdma_queue *wq, u32 wqe_offset)
1021 {
1022 u32 offset = (wqe_offset * GDMA_WQE_BU_SIZE) & (wq->queue_size - 1);
1023
1024 WARN_ON_ONCE((offset + GDMA_WQE_BU_SIZE) > wq->queue_size);
1025
1026 return wq->queue_mem_ptr + offset;
1027 }
1028
mana_gd_write_client_oob(const struct gdma_wqe_request * wqe_req,enum gdma_queue_type q_type,u32 client_oob_size,u32 sgl_data_size,u8 * wqe_ptr)1029 static u32 mana_gd_write_client_oob(const struct gdma_wqe_request *wqe_req,
1030 enum gdma_queue_type q_type,
1031 u32 client_oob_size, u32 sgl_data_size,
1032 u8 *wqe_ptr)
1033 {
1034 bool oob_in_sgl = !!(wqe_req->flags & GDMA_WR_OOB_IN_SGL);
1035 bool pad_data = !!(wqe_req->flags & GDMA_WR_PAD_BY_SGE0);
1036 struct gdma_wqe *header = (struct gdma_wqe *)wqe_ptr;
1037 u8 *ptr;
1038
1039 memset(header, 0, sizeof(struct gdma_wqe));
1040 header->num_sge = wqe_req->num_sge;
1041 header->inline_oob_size_div4 = client_oob_size / sizeof(u32);
1042
1043 if (oob_in_sgl) {
1044 WARN_ON_ONCE(!pad_data || wqe_req->num_sge < 2);
1045
1046 header->client_oob_in_sgl = 1;
1047
1048 if (pad_data)
1049 header->last_vbytes = wqe_req->sgl[0].size;
1050 }
1051
1052 if (q_type == GDMA_SQ)
1053 header->client_data_unit = wqe_req->client_data_unit;
1054
1055 /* The size of gdma_wqe + client_oob_size must be less than or equal
1056 * to one Basic Unit (i.e. 32 bytes), so the pointer can't go beyond
1057 * the queue memory buffer boundary.
1058 */
1059 ptr = wqe_ptr + sizeof(header);
1060
1061 if (wqe_req->inline_oob_data && wqe_req->inline_oob_size > 0) {
1062 memcpy(ptr, wqe_req->inline_oob_data, wqe_req->inline_oob_size);
1063
1064 if (client_oob_size > wqe_req->inline_oob_size)
1065 memset(ptr + wqe_req->inline_oob_size, 0,
1066 client_oob_size - wqe_req->inline_oob_size);
1067 }
1068
1069 return sizeof(header) + client_oob_size;
1070 }
1071
mana_gd_write_sgl(struct gdma_queue * wq,u8 * wqe_ptr,const struct gdma_wqe_request * wqe_req)1072 static void mana_gd_write_sgl(struct gdma_queue *wq, u8 *wqe_ptr,
1073 const struct gdma_wqe_request *wqe_req)
1074 {
1075 u32 sgl_size = sizeof(struct gdma_sge) * wqe_req->num_sge;
1076 const u8 *address = (u8 *)wqe_req->sgl;
1077 u8 *base_ptr, *end_ptr;
1078 u32 size_to_end;
1079
1080 base_ptr = wq->queue_mem_ptr;
1081 end_ptr = base_ptr + wq->queue_size;
1082 size_to_end = (u32)(end_ptr - wqe_ptr);
1083
1084 if (size_to_end < sgl_size) {
1085 memcpy(wqe_ptr, address, size_to_end);
1086
1087 wqe_ptr = base_ptr;
1088 address += size_to_end;
1089 sgl_size -= size_to_end;
1090 }
1091
1092 memcpy(wqe_ptr, address, sgl_size);
1093 }
1094
mana_gd_post_work_request(struct gdma_queue * wq,const struct gdma_wqe_request * wqe_req,struct gdma_posted_wqe_info * wqe_info)1095 int mana_gd_post_work_request(struct gdma_queue *wq,
1096 const struct gdma_wqe_request *wqe_req,
1097 struct gdma_posted_wqe_info *wqe_info)
1098 {
1099 u32 client_oob_size = wqe_req->inline_oob_size;
1100 struct gdma_context *gc;
1101 u32 sgl_data_size;
1102 u32 max_wqe_size;
1103 u32 wqe_size;
1104 u8 *wqe_ptr;
1105
1106 if (wqe_req->num_sge == 0)
1107 return -EINVAL;
1108
1109 if (wq->type == GDMA_RQ) {
1110 if (client_oob_size != 0)
1111 return -EINVAL;
1112
1113 client_oob_size = INLINE_OOB_SMALL_SIZE;
1114
1115 max_wqe_size = GDMA_MAX_RQE_SIZE;
1116 } else {
1117 if (client_oob_size != INLINE_OOB_SMALL_SIZE &&
1118 client_oob_size != INLINE_OOB_LARGE_SIZE)
1119 return -EINVAL;
1120
1121 max_wqe_size = GDMA_MAX_SQE_SIZE;
1122 }
1123
1124 sgl_data_size = sizeof(struct gdma_sge) * wqe_req->num_sge;
1125 wqe_size = ALIGN(sizeof(struct gdma_wqe) + client_oob_size +
1126 sgl_data_size, GDMA_WQE_BU_SIZE);
1127 if (wqe_size > max_wqe_size)
1128 return -EINVAL;
1129
1130 if (wq->monitor_avl_buf && wqe_size > mana_gd_wq_avail_space(wq)) {
1131 gc = wq->gdma_dev->gdma_context;
1132 dev_err(gc->dev, "unsuccessful flow control!\n");
1133 return -ENOSPC;
1134 }
1135
1136 if (wqe_info)
1137 wqe_info->wqe_size_in_bu = wqe_size / GDMA_WQE_BU_SIZE;
1138
1139 wqe_ptr = mana_gd_get_wqe_ptr(wq, wq->head);
1140 wqe_ptr += mana_gd_write_client_oob(wqe_req, wq->type, client_oob_size,
1141 sgl_data_size, wqe_ptr);
1142 if (wqe_ptr >= (u8 *)wq->queue_mem_ptr + wq->queue_size)
1143 wqe_ptr -= wq->queue_size;
1144
1145 mana_gd_write_sgl(wq, wqe_ptr, wqe_req);
1146
1147 wq->head += wqe_size / GDMA_WQE_BU_SIZE;
1148
1149 return 0;
1150 }
1151
mana_gd_post_and_ring(struct gdma_queue * queue,const struct gdma_wqe_request * wqe_req,struct gdma_posted_wqe_info * wqe_info)1152 int mana_gd_post_and_ring(struct gdma_queue *queue,
1153 const struct gdma_wqe_request *wqe_req,
1154 struct gdma_posted_wqe_info *wqe_info)
1155 {
1156 struct gdma_context *gc = queue->gdma_dev->gdma_context;
1157 int err;
1158
1159 err = mana_gd_post_work_request(queue, wqe_req, wqe_info);
1160 if (err)
1161 return err;
1162
1163 mana_gd_wq_ring_doorbell(gc, queue);
1164
1165 return 0;
1166 }
1167
mana_gd_read_cqe(struct gdma_queue * cq,struct gdma_comp * comp)1168 static int mana_gd_read_cqe(struct gdma_queue *cq, struct gdma_comp *comp)
1169 {
1170 unsigned int num_cqe = cq->queue_size / sizeof(struct gdma_cqe);
1171 struct gdma_cqe *cq_cqe = cq->queue_mem_ptr;
1172 u32 owner_bits, new_bits, old_bits;
1173 struct gdma_cqe *cqe;
1174
1175 cqe = &cq_cqe[cq->head % num_cqe];
1176 owner_bits = cqe->cqe_info.owner_bits;
1177
1178 old_bits = (cq->head / num_cqe - 1) & GDMA_CQE_OWNER_MASK;
1179 /* Return 0 if no more entries. */
1180 if (owner_bits == old_bits)
1181 return 0;
1182
1183 new_bits = (cq->head / num_cqe) & GDMA_CQE_OWNER_MASK;
1184 /* Return -1 if overflow detected. */
1185 if (WARN_ON_ONCE(owner_bits != new_bits))
1186 return -1;
1187
1188 /* Per GDMA spec, rmb is necessary after checking owner_bits, before
1189 * reading completion info
1190 */
1191 rmb();
1192
1193 comp->wq_num = cqe->cqe_info.wq_num;
1194 comp->is_sq = cqe->cqe_info.is_sq;
1195 memcpy(comp->cqe_data, cqe->cqe_data, GDMA_COMP_DATA_SIZE);
1196
1197 return 1;
1198 }
1199
mana_gd_poll_cq(struct gdma_queue * cq,struct gdma_comp * comp,int num_cqe)1200 int mana_gd_poll_cq(struct gdma_queue *cq, struct gdma_comp *comp, int num_cqe)
1201 {
1202 int cqe_idx;
1203 int ret;
1204
1205 for (cqe_idx = 0; cqe_idx < num_cqe; cqe_idx++) {
1206 ret = mana_gd_read_cqe(cq, &comp[cqe_idx]);
1207
1208 if (ret < 0) {
1209 cq->head -= cqe_idx;
1210 return ret;
1211 }
1212
1213 if (ret == 0)
1214 break;
1215
1216 cq->head++;
1217 }
1218
1219 return cqe_idx;
1220 }
1221
mana_gd_intr(int irq,void * arg)1222 static irqreturn_t mana_gd_intr(int irq, void *arg)
1223 {
1224 struct gdma_irq_context *gic = arg;
1225 struct list_head *eq_list = &gic->eq_list;
1226 struct gdma_queue *eq;
1227
1228 rcu_read_lock();
1229 list_for_each_entry_rcu(eq, eq_list, entry) {
1230 gic->handler(eq);
1231 }
1232 rcu_read_unlock();
1233
1234 return IRQ_HANDLED;
1235 }
1236
mana_gd_alloc_res_map(u32 res_avail,struct gdma_resource * r)1237 int mana_gd_alloc_res_map(u32 res_avail, struct gdma_resource *r)
1238 {
1239 r->map = bitmap_zalloc(res_avail, GFP_KERNEL);
1240 if (!r->map)
1241 return -ENOMEM;
1242
1243 r->size = res_avail;
1244 spin_lock_init(&r->lock);
1245
1246 return 0;
1247 }
1248
mana_gd_free_res_map(struct gdma_resource * r)1249 void mana_gd_free_res_map(struct gdma_resource *r)
1250 {
1251 bitmap_free(r->map);
1252 r->map = NULL;
1253 r->size = 0;
1254 }
1255
irq_setup(unsigned int * irqs,unsigned int len,int node)1256 static int irq_setup(unsigned int *irqs, unsigned int len, int node)
1257 {
1258 const struct cpumask *next, *prev = cpu_none_mask;
1259 cpumask_var_t cpus __free(free_cpumask_var);
1260 int cpu, weight;
1261
1262 if (!alloc_cpumask_var(&cpus, GFP_KERNEL))
1263 return -ENOMEM;
1264
1265 rcu_read_lock();
1266 for_each_numa_hop_mask(next, node) {
1267 weight = cpumask_weight_andnot(next, prev);
1268 while (weight > 0) {
1269 cpumask_andnot(cpus, next, prev);
1270 for_each_cpu(cpu, cpus) {
1271 if (len-- == 0)
1272 goto done;
1273 irq_set_affinity_and_hint(*irqs++, topology_sibling_cpumask(cpu));
1274 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu));
1275 --weight;
1276 }
1277 }
1278 prev = next;
1279 }
1280 done:
1281 rcu_read_unlock();
1282 return 0;
1283 }
1284
mana_gd_setup_irqs(struct pci_dev * pdev)1285 static int mana_gd_setup_irqs(struct pci_dev *pdev)
1286 {
1287 struct gdma_context *gc = pci_get_drvdata(pdev);
1288 unsigned int max_queues_per_port;
1289 struct gdma_irq_context *gic;
1290 unsigned int max_irqs, cpu;
1291 int start_irq_index = 1;
1292 int nvec, *irqs, irq;
1293 int err, i = 0, j;
1294
1295 cpus_read_lock();
1296 max_queues_per_port = num_online_cpus();
1297 if (max_queues_per_port > MANA_MAX_NUM_QUEUES)
1298 max_queues_per_port = MANA_MAX_NUM_QUEUES;
1299
1300 /* Need 1 interrupt for the Hardware communication Channel (HWC) */
1301 max_irqs = max_queues_per_port + 1;
1302
1303 nvec = pci_alloc_irq_vectors(pdev, 2, max_irqs, PCI_IRQ_MSIX);
1304 if (nvec < 0) {
1305 cpus_read_unlock();
1306 return nvec;
1307 }
1308 if (nvec <= num_online_cpus())
1309 start_irq_index = 0;
1310
1311 irqs = kmalloc_array((nvec - start_irq_index), sizeof(int), GFP_KERNEL);
1312 if (!irqs) {
1313 err = -ENOMEM;
1314 goto free_irq_vector;
1315 }
1316
1317 gc->irq_contexts = kcalloc(nvec, sizeof(struct gdma_irq_context),
1318 GFP_KERNEL);
1319 if (!gc->irq_contexts) {
1320 err = -ENOMEM;
1321 goto free_irq_array;
1322 }
1323
1324 for (i = 0; i < nvec; i++) {
1325 gic = &gc->irq_contexts[i];
1326 gic->handler = mana_gd_process_eq_events;
1327 INIT_LIST_HEAD(&gic->eq_list);
1328 spin_lock_init(&gic->lock);
1329
1330 if (!i)
1331 snprintf(gic->name, MANA_IRQ_NAME_SZ, "mana_hwc@pci:%s",
1332 pci_name(pdev));
1333 else
1334 snprintf(gic->name, MANA_IRQ_NAME_SZ, "mana_q%d@pci:%s",
1335 i - 1, pci_name(pdev));
1336
1337 irq = pci_irq_vector(pdev, i);
1338 if (irq < 0) {
1339 err = irq;
1340 goto free_irq;
1341 }
1342
1343 if (!i) {
1344 err = request_irq(irq, mana_gd_intr, 0, gic->name, gic);
1345 if (err)
1346 goto free_irq;
1347
1348 /* If number of IRQ is one extra than number of online CPUs,
1349 * then we need to assign IRQ0 (hwc irq) and IRQ1 to
1350 * same CPU.
1351 * Else we will use different CPUs for IRQ0 and IRQ1.
1352 * Also we are using cpumask_local_spread instead of
1353 * cpumask_first for the node, because the node can be
1354 * mem only.
1355 */
1356 if (start_irq_index) {
1357 cpu = cpumask_local_spread(i, gc->numa_node);
1358 irq_set_affinity_and_hint(irq, cpumask_of(cpu));
1359 } else {
1360 irqs[start_irq_index] = irq;
1361 }
1362 } else {
1363 irqs[i - start_irq_index] = irq;
1364 err = request_irq(irqs[i - start_irq_index], mana_gd_intr, 0,
1365 gic->name, gic);
1366 if (err)
1367 goto free_irq;
1368 }
1369 }
1370
1371 err = irq_setup(irqs, (nvec - start_irq_index), gc->numa_node);
1372 if (err)
1373 goto free_irq;
1374
1375 gc->max_num_msix = nvec;
1376 gc->num_msix_usable = nvec;
1377 cpus_read_unlock();
1378 kfree(irqs);
1379 return 0;
1380
1381 free_irq:
1382 for (j = i - 1; j >= 0; j--) {
1383 irq = pci_irq_vector(pdev, j);
1384 gic = &gc->irq_contexts[j];
1385
1386 irq_update_affinity_hint(irq, NULL);
1387 free_irq(irq, gic);
1388 }
1389
1390 kfree(gc->irq_contexts);
1391 gc->irq_contexts = NULL;
1392 free_irq_array:
1393 kfree(irqs);
1394 free_irq_vector:
1395 cpus_read_unlock();
1396 pci_free_irq_vectors(pdev);
1397 return err;
1398 }
1399
mana_gd_remove_irqs(struct pci_dev * pdev)1400 static void mana_gd_remove_irqs(struct pci_dev *pdev)
1401 {
1402 struct gdma_context *gc = pci_get_drvdata(pdev);
1403 struct gdma_irq_context *gic;
1404 int irq, i;
1405
1406 if (gc->max_num_msix < 1)
1407 return;
1408
1409 for (i = 0; i < gc->max_num_msix; i++) {
1410 irq = pci_irq_vector(pdev, i);
1411 if (irq < 0)
1412 continue;
1413
1414 gic = &gc->irq_contexts[i];
1415
1416 /* Need to clear the hint before free_irq */
1417 irq_update_affinity_hint(irq, NULL);
1418 free_irq(irq, gic);
1419 }
1420
1421 pci_free_irq_vectors(pdev);
1422
1423 gc->max_num_msix = 0;
1424 gc->num_msix_usable = 0;
1425 kfree(gc->irq_contexts);
1426 gc->irq_contexts = NULL;
1427 }
1428
mana_gd_setup(struct pci_dev * pdev)1429 static int mana_gd_setup(struct pci_dev *pdev)
1430 {
1431 struct gdma_context *gc = pci_get_drvdata(pdev);
1432 int err;
1433
1434 mana_gd_init_registers(pdev);
1435 mana_smc_init(&gc->shm_channel, gc->dev, gc->shm_base);
1436
1437 err = mana_gd_setup_irqs(pdev);
1438 if (err)
1439 return err;
1440
1441 err = mana_hwc_create_channel(gc);
1442 if (err)
1443 goto remove_irq;
1444
1445 err = mana_gd_verify_vf_version(pdev);
1446 if (err)
1447 goto destroy_hwc;
1448
1449 err = mana_gd_query_max_resources(pdev);
1450 if (err)
1451 goto destroy_hwc;
1452
1453 err = mana_gd_detect_devices(pdev);
1454 if (err)
1455 goto destroy_hwc;
1456
1457 return 0;
1458
1459 destroy_hwc:
1460 mana_hwc_destroy_channel(gc);
1461 remove_irq:
1462 mana_gd_remove_irqs(pdev);
1463 return err;
1464 }
1465
mana_gd_cleanup(struct pci_dev * pdev)1466 static void mana_gd_cleanup(struct pci_dev *pdev)
1467 {
1468 struct gdma_context *gc = pci_get_drvdata(pdev);
1469
1470 mana_hwc_destroy_channel(gc);
1471
1472 mana_gd_remove_irqs(pdev);
1473 }
1474
mana_is_pf(unsigned short dev_id)1475 static bool mana_is_pf(unsigned short dev_id)
1476 {
1477 return dev_id == MANA_PF_DEVICE_ID;
1478 }
1479
mana_gd_probe(struct pci_dev * pdev,const struct pci_device_id * ent)1480 static int mana_gd_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
1481 {
1482 struct gdma_context *gc;
1483 void __iomem *bar0_va;
1484 int bar = 0;
1485 int err;
1486
1487 /* Each port has 2 CQs, each CQ has at most 1 EQE at a time */
1488 BUILD_BUG_ON(2 * MAX_PORTS_IN_MANA_DEV * GDMA_EQE_SIZE > EQ_SIZE);
1489
1490 err = pci_enable_device(pdev);
1491 if (err)
1492 return -ENXIO;
1493
1494 pci_set_master(pdev);
1495
1496 err = pci_request_regions(pdev, "mana");
1497 if (err)
1498 goto disable_dev;
1499
1500 err = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(64));
1501 if (err)
1502 goto release_region;
1503
1504 dma_set_max_seg_size(&pdev->dev, UINT_MAX);
1505
1506 err = -ENOMEM;
1507 gc = vzalloc(sizeof(*gc));
1508 if (!gc)
1509 goto release_region;
1510
1511 mutex_init(&gc->eq_test_event_mutex);
1512 pci_set_drvdata(pdev, gc);
1513 gc->bar0_pa = pci_resource_start(pdev, 0);
1514
1515 bar0_va = pci_iomap(pdev, bar, 0);
1516 if (!bar0_va)
1517 goto free_gc;
1518
1519 gc->numa_node = dev_to_node(&pdev->dev);
1520 gc->is_pf = mana_is_pf(pdev->device);
1521 gc->bar0_va = bar0_va;
1522 gc->dev = &pdev->dev;
1523
1524 if (gc->is_pf)
1525 gc->mana_pci_debugfs = debugfs_create_dir("0", mana_debugfs_root);
1526 else
1527 gc->mana_pci_debugfs = debugfs_create_dir(pci_slot_name(pdev->slot),
1528 mana_debugfs_root);
1529
1530 err = mana_gd_setup(pdev);
1531 if (err)
1532 goto unmap_bar;
1533
1534 err = mana_probe(&gc->mana, false);
1535 if (err)
1536 goto cleanup_gd;
1537
1538 return 0;
1539
1540 cleanup_gd:
1541 mana_gd_cleanup(pdev);
1542 unmap_bar:
1543 /*
1544 * at this point we know that the other debugfs child dir/files
1545 * are either not yet created or are already cleaned up.
1546 * The pci debugfs folder clean-up now, will only be cleaning up
1547 * adapter-MTU file and apc->mana_pci_debugfs folder.
1548 */
1549 debugfs_remove_recursive(gc->mana_pci_debugfs);
1550 pci_iounmap(pdev, bar0_va);
1551 free_gc:
1552 pci_set_drvdata(pdev, NULL);
1553 vfree(gc);
1554 release_region:
1555 pci_release_regions(pdev);
1556 disable_dev:
1557 pci_disable_device(pdev);
1558 dev_err(&pdev->dev, "gdma probe failed: err = %d\n", err);
1559 return err;
1560 }
1561
mana_gd_remove(struct pci_dev * pdev)1562 static void mana_gd_remove(struct pci_dev *pdev)
1563 {
1564 struct gdma_context *gc = pci_get_drvdata(pdev);
1565
1566 mana_remove(&gc->mana, false);
1567
1568 mana_gd_cleanup(pdev);
1569
1570 debugfs_remove_recursive(gc->mana_pci_debugfs);
1571
1572 pci_iounmap(pdev, gc->bar0_va);
1573
1574 vfree(gc);
1575
1576 pci_release_regions(pdev);
1577 pci_disable_device(pdev);
1578 }
1579
1580 /* The 'state' parameter is not used. */
mana_gd_suspend(struct pci_dev * pdev,pm_message_t state)1581 static int mana_gd_suspend(struct pci_dev *pdev, pm_message_t state)
1582 {
1583 struct gdma_context *gc = pci_get_drvdata(pdev);
1584
1585 mana_remove(&gc->mana, true);
1586
1587 mana_gd_cleanup(pdev);
1588
1589 return 0;
1590 }
1591
1592 /* In case the NIC hardware stops working, the suspend and resume callbacks will
1593 * fail -- if this happens, it's safer to just report an error than try to undo
1594 * what has been done.
1595 */
mana_gd_resume(struct pci_dev * pdev)1596 static int mana_gd_resume(struct pci_dev *pdev)
1597 {
1598 struct gdma_context *gc = pci_get_drvdata(pdev);
1599 int err;
1600
1601 err = mana_gd_setup(pdev);
1602 if (err)
1603 return err;
1604
1605 err = mana_probe(&gc->mana, true);
1606 if (err)
1607 return err;
1608
1609 return 0;
1610 }
1611
1612 /* Quiesce the device for kexec. This is also called upon reboot/shutdown. */
mana_gd_shutdown(struct pci_dev * pdev)1613 static void mana_gd_shutdown(struct pci_dev *pdev)
1614 {
1615 struct gdma_context *gc = pci_get_drvdata(pdev);
1616
1617 dev_info(&pdev->dev, "Shutdown was called\n");
1618
1619 mana_remove(&gc->mana, true);
1620
1621 mana_gd_cleanup(pdev);
1622
1623 debugfs_remove_recursive(gc->mana_pci_debugfs);
1624
1625 pci_disable_device(pdev);
1626 }
1627
1628 static const struct pci_device_id mana_id_table[] = {
1629 { PCI_DEVICE(PCI_VENDOR_ID_MICROSOFT, MANA_PF_DEVICE_ID) },
1630 { PCI_DEVICE(PCI_VENDOR_ID_MICROSOFT, MANA_VF_DEVICE_ID) },
1631 { }
1632 };
1633
1634 static struct pci_driver mana_driver = {
1635 .name = "mana",
1636 .id_table = mana_id_table,
1637 .probe = mana_gd_probe,
1638 .remove = mana_gd_remove,
1639 .suspend = mana_gd_suspend,
1640 .resume = mana_gd_resume,
1641 .shutdown = mana_gd_shutdown,
1642 };
1643
mana_driver_init(void)1644 static int __init mana_driver_init(void)
1645 {
1646 int err;
1647
1648 mana_debugfs_root = debugfs_create_dir("mana", NULL);
1649
1650 err = pci_register_driver(&mana_driver);
1651 if (err)
1652 debugfs_remove(mana_debugfs_root);
1653
1654 return err;
1655 }
1656
mana_driver_exit(void)1657 static void __exit mana_driver_exit(void)
1658 {
1659 debugfs_remove(mana_debugfs_root);
1660
1661 pci_unregister_driver(&mana_driver);
1662 }
1663
1664 module_init(mana_driver_init);
1665 module_exit(mana_driver_exit);
1666
1667 MODULE_DEVICE_TABLE(pci, mana_id_table);
1668
1669 MODULE_LICENSE("Dual BSD/GPL");
1670 MODULE_DESCRIPTION("Microsoft Azure Network Adapter driver");
1671