xref: /linux/drivers/crypto/caam/qi.c (revision 85ffc6e4ed3712f8b3fedb3fbe42afae644a699c)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * CAAM/SEC 4.x QI transport/backend driver
4  * Queue Interface backend functionality
5  *
6  * Copyright 2013-2016 Freescale Semiconductor, Inc.
7  * Copyright 2016-2017, 2019-2020 NXP
8  */
9 
10 #include <linux/cpumask.h>
11 #include <linux/device.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/kernel.h>
14 #include <linux/kthread.h>
15 #include <linux/netdevice.h>
16 #include <linux/platform_device.h>
17 #include <linux/slab.h>
18 #include <linux/string.h>
19 #include <soc/fsl/qman.h>
20 
21 #include "debugfs.h"
22 #include "regs.h"
23 #include "qi.h"
24 #include "desc.h"
25 #include "intern.h"
26 #include "desc_constr.h"
27 
28 #define PREHDR_RSLS_SHIFT	31
29 #define PREHDR_ABS		BIT(25)
30 
31 /*
32  * Use a reasonable backlog of frames (per CPU) as congestion threshold,
33  * so that resources used by the in-flight buffers do not become a memory hog.
34  */
35 #define MAX_RSP_FQ_BACKLOG_PER_CPU	256
36 
37 #define CAAM_QI_ENQUEUE_RETRIES	10000
38 
39 #define CAAM_NAPI_WEIGHT	63
40 
41 /*
42  * caam_napi - struct holding CAAM NAPI-related params
43  * @irqtask: IRQ task for QI backend
44  * @p: QMan portal
45  */
46 struct caam_napi {
47 	struct napi_struct irqtask;
48 	struct qman_portal *p;
49 };
50 
51 /*
52  * caam_qi_pcpu_priv - percpu private data structure to main list of pending
53  *                     responses expected on each cpu.
54  * @caam_napi: CAAM NAPI params
55  * @net_dev: netdev used by NAPI
56  * @rsp_fq: response FQ from CAAM
57  */
58 struct caam_qi_pcpu_priv {
59 	struct caam_napi caam_napi;
60 	struct net_device *net_dev;
61 	struct qman_fq *rsp_fq;
62 } ____cacheline_aligned;
63 
64 static DEFINE_PER_CPU(struct caam_qi_pcpu_priv, pcpu_qipriv);
65 static DEFINE_PER_CPU(int, last_cpu);
66 
67 /*
68  * caam_qi_priv - CAAM QI backend private params
69  * @cgr: QMan congestion group
70  */
71 struct caam_qi_priv {
72 	struct qman_cgr cgr;
73 };
74 
75 static struct caam_qi_priv qipriv ____cacheline_aligned;
76 
77 /*
78  * This is written by only one core - the one that initialized the CGR - and
79  * read by multiple cores (all the others).
80  */
81 bool caam_congested __read_mostly;
82 EXPORT_SYMBOL(caam_congested);
83 
84 /*
85  * This is a cache of buffers, from which the users of CAAM QI driver
86  * can allocate short (CAAM_QI_MEMCACHE_SIZE) buffers. It's faster than
87  * doing malloc on the hotpath.
88  * NOTE: A more elegant solution would be to have some headroom in the frames
89  *       being processed. This could be added by the dpaa-ethernet driver.
90  *       This would pose a problem for userspace application processing which
91  *       cannot know of this limitation. So for now, this will work.
92  * NOTE: The memcache is SMP-safe. No need to handle spinlocks in-here
93  */
94 static struct kmem_cache *qi_cache;
95 
caam_iova_to_virt(struct iommu_domain * domain,dma_addr_t iova_addr)96 static void *caam_iova_to_virt(struct iommu_domain *domain,
97 			       dma_addr_t iova_addr)
98 {
99 	phys_addr_t phys_addr;
100 
101 	phys_addr = domain ? iommu_iova_to_phys(domain, iova_addr) : iova_addr;
102 
103 	return phys_to_virt(phys_addr);
104 }
105 
caam_qi_enqueue(struct device * qidev,struct caam_drv_req * req)106 int caam_qi_enqueue(struct device *qidev, struct caam_drv_req *req)
107 {
108 	struct qm_fd fd;
109 	dma_addr_t addr;
110 	int ret;
111 	int num_retries = 0;
112 
113 	qm_fd_clear_fd(&fd);
114 	qm_fd_set_compound(&fd, qm_sg_entry_get_len(&req->fd_sgt[1]));
115 
116 	addr = dma_map_single(qidev, req->fd_sgt, sizeof(req->fd_sgt),
117 			      DMA_BIDIRECTIONAL);
118 	if (dma_mapping_error(qidev, addr)) {
119 		dev_err(qidev, "DMA mapping error for QI enqueue request\n");
120 		return -EIO;
121 	}
122 	qm_fd_addr_set64(&fd, addr);
123 
124 	do {
125 		ret = qman_enqueue(req->drv_ctx->req_fq, &fd);
126 		if (likely(!ret)) {
127 			refcount_inc(&req->drv_ctx->refcnt);
128 			return 0;
129 		}
130 
131 		if (ret != -EBUSY)
132 			break;
133 		num_retries++;
134 	} while (num_retries < CAAM_QI_ENQUEUE_RETRIES);
135 
136 	dev_err(qidev, "qman_enqueue failed: %d\n", ret);
137 
138 	return ret;
139 }
140 EXPORT_SYMBOL(caam_qi_enqueue);
141 
caam_fq_ern_cb(struct qman_portal * qm,struct qman_fq * fq,const union qm_mr_entry * msg)142 static void caam_fq_ern_cb(struct qman_portal *qm, struct qman_fq *fq,
143 			   const union qm_mr_entry *msg)
144 {
145 	const struct qm_fd *fd;
146 	struct caam_drv_req *drv_req;
147 	struct device *qidev = &(raw_cpu_ptr(&pcpu_qipriv)->net_dev->dev);
148 	struct caam_drv_private *priv = dev_get_drvdata(qidev);
149 
150 	fd = &msg->ern.fd;
151 
152 	drv_req = caam_iova_to_virt(priv->domain, qm_fd_addr_get64(fd));
153 	if (!drv_req) {
154 		dev_err(qidev,
155 			"Can't find original request for CAAM response\n");
156 		return;
157 	}
158 
159 	refcount_dec(&drv_req->drv_ctx->refcnt);
160 
161 	if (qm_fd_get_format(fd) != qm_fd_compound) {
162 		dev_err(qidev, "Non-compound FD from CAAM\n");
163 		return;
164 	}
165 
166 	dma_unmap_single(drv_req->drv_ctx->qidev, qm_fd_addr(fd),
167 			 sizeof(drv_req->fd_sgt), DMA_BIDIRECTIONAL);
168 
169 	if (fd->status)
170 		drv_req->cbk(drv_req, be32_to_cpu(fd->status));
171 	else
172 		drv_req->cbk(drv_req, JRSTA_SSRC_QI);
173 }
174 
create_caam_req_fq(struct device * qidev,struct qman_fq * rsp_fq,dma_addr_t hwdesc,int fq_sched_flag)175 static struct qman_fq *create_caam_req_fq(struct device *qidev,
176 					  struct qman_fq *rsp_fq,
177 					  dma_addr_t hwdesc,
178 					  int fq_sched_flag)
179 {
180 	int ret;
181 	struct qman_fq *req_fq;
182 	struct qm_mcc_initfq opts;
183 
184 	req_fq = kzalloc(sizeof(*req_fq), GFP_ATOMIC);
185 	if (!req_fq)
186 		return ERR_PTR(-ENOMEM);
187 
188 	req_fq->cb.ern = caam_fq_ern_cb;
189 	req_fq->cb.fqs = NULL;
190 
191 	ret = qman_create_fq(0, QMAN_FQ_FLAG_DYNAMIC_FQID |
192 				QMAN_FQ_FLAG_TO_DCPORTAL, req_fq);
193 	if (ret) {
194 		dev_err(qidev, "Failed to create session req FQ\n");
195 		goto create_req_fq_fail;
196 	}
197 
198 	memset(&opts, 0, sizeof(opts));
199 	opts.we_mask = cpu_to_be16(QM_INITFQ_WE_FQCTRL | QM_INITFQ_WE_DESTWQ |
200 				   QM_INITFQ_WE_CONTEXTB |
201 				   QM_INITFQ_WE_CONTEXTA | QM_INITFQ_WE_CGID);
202 	opts.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_CPCSTASH | QM_FQCTRL_CGE);
203 	qm_fqd_set_destwq(&opts.fqd, qm_channel_caam, 2);
204 	opts.fqd.context_b = cpu_to_be32(qman_fq_fqid(rsp_fq));
205 	qm_fqd_context_a_set64(&opts.fqd, hwdesc);
206 	opts.fqd.cgid = qipriv.cgr.cgrid;
207 
208 	ret = qman_init_fq(req_fq, fq_sched_flag, &opts);
209 	if (ret) {
210 		dev_err(qidev, "Failed to init session req FQ\n");
211 		goto init_req_fq_fail;
212 	}
213 
214 	dev_dbg(qidev, "Allocated request FQ %u for CPU %u\n", req_fq->fqid,
215 		smp_processor_id());
216 	return req_fq;
217 
218 init_req_fq_fail:
219 	qman_destroy_fq(req_fq);
220 create_req_fq_fail:
221 	kfree(req_fq);
222 	return ERR_PTR(ret);
223 }
224 
empty_retired_fq(struct device * qidev,struct qman_fq * fq)225 static int empty_retired_fq(struct device *qidev, struct qman_fq *fq)
226 {
227 	int ret;
228 
229 	ret = qman_volatile_dequeue(fq, QMAN_VOLATILE_FLAG_WAIT_INT |
230 				    QMAN_VOLATILE_FLAG_FINISH,
231 				    QM_VDQCR_PRECEDENCE_VDQCR |
232 				    QM_VDQCR_NUMFRAMES_TILLEMPTY);
233 	if (ret) {
234 		dev_err(qidev, "Volatile dequeue fail for FQ: %u\n", fq->fqid);
235 		return ret;
236 	}
237 
238 	do {
239 		struct qman_portal *p;
240 
241 		p = qman_get_affine_portal(smp_processor_id());
242 		qman_p_poll_dqrr(p, 16);
243 	} while (fq->flags & QMAN_FQ_STATE_NE);
244 
245 	return 0;
246 }
247 
kill_fq(struct device * qidev,struct qman_fq * fq)248 static int kill_fq(struct device *qidev, struct qman_fq *fq)
249 {
250 	u32 flags;
251 	int ret;
252 
253 	ret = qman_retire_fq(fq, &flags);
254 	if (ret < 0) {
255 		dev_err(qidev, "qman_retire_fq failed: %d\n", ret);
256 		return ret;
257 	}
258 
259 	if (!ret)
260 		goto empty_fq;
261 
262 	/* Async FQ retirement condition */
263 	if (ret == 1) {
264 		/* Retry till FQ gets in retired state */
265 		do {
266 			msleep(20);
267 		} while (fq->state != qman_fq_state_retired);
268 
269 		WARN_ON(fq->flags & QMAN_FQ_STATE_BLOCKOOS);
270 		WARN_ON(fq->flags & QMAN_FQ_STATE_ORL);
271 	}
272 
273 empty_fq:
274 	if (fq->flags & QMAN_FQ_STATE_NE) {
275 		ret = empty_retired_fq(qidev, fq);
276 		if (ret) {
277 			dev_err(qidev, "empty_retired_fq fail for FQ: %u\n",
278 				fq->fqid);
279 			return ret;
280 		}
281 	}
282 
283 	ret = qman_oos_fq(fq);
284 	if (ret)
285 		dev_err(qidev, "OOS of FQID: %u failed\n", fq->fqid);
286 
287 	qman_destroy_fq(fq);
288 	kfree(fq);
289 
290 	return ret;
291 }
292 
empty_caam_fq(struct qman_fq * fq,struct caam_drv_ctx * drv_ctx)293 static int empty_caam_fq(struct qman_fq *fq, struct caam_drv_ctx *drv_ctx)
294 {
295 	int ret;
296 	int retries = 10;
297 	struct qm_mcr_queryfq_np np;
298 
299 	/* Wait till the older CAAM FQ get empty */
300 	do {
301 		ret = qman_query_fq_np(fq, &np);
302 		if (ret)
303 			return ret;
304 
305 		if (!qm_mcr_np_get(&np, frm_cnt))
306 			break;
307 
308 		msleep(20);
309 	} while (1);
310 
311 	/* Wait until pending jobs from this FQ are processed by CAAM */
312 	do {
313 		if (refcount_read(&drv_ctx->refcnt) == 1)
314 			break;
315 
316 		msleep(20);
317 	} while (--retries);
318 
319 	if (!retries)
320 		dev_warn_once(drv_ctx->qidev, "%d frames from FQID %u still pending in CAAM\n",
321 			      refcount_read(&drv_ctx->refcnt), fq->fqid);
322 
323 	return 0;
324 }
325 
caam_drv_ctx_update(struct caam_drv_ctx * drv_ctx,u32 * sh_desc)326 int caam_drv_ctx_update(struct caam_drv_ctx *drv_ctx, u32 *sh_desc)
327 {
328 	int ret;
329 	u32 num_words;
330 	struct qman_fq *new_fq, *old_fq;
331 	struct device *qidev = drv_ctx->qidev;
332 
333 	num_words = desc_len(sh_desc);
334 	if (num_words > MAX_SDLEN) {
335 		dev_err(qidev, "Invalid descriptor len: %d words\n", num_words);
336 		return -EINVAL;
337 	}
338 
339 	/* Note down older req FQ */
340 	old_fq = drv_ctx->req_fq;
341 
342 	/* Create a new req FQ in parked state */
343 	new_fq = create_caam_req_fq(drv_ctx->qidev, drv_ctx->rsp_fq,
344 				    drv_ctx->context_a, 0);
345 	if (IS_ERR(new_fq)) {
346 		dev_err(qidev, "FQ allocation for shdesc update failed\n");
347 		return PTR_ERR(new_fq);
348 	}
349 
350 	/* Hook up new FQ to context so that new requests keep queuing */
351 	drv_ctx->req_fq = new_fq;
352 
353 	/* Empty and remove the older FQ */
354 	ret = empty_caam_fq(old_fq, drv_ctx);
355 	if (ret) {
356 		dev_err(qidev, "Old CAAM FQ empty failed: %d\n", ret);
357 
358 		/* We can revert to older FQ */
359 		drv_ctx->req_fq = old_fq;
360 
361 		if (kill_fq(qidev, new_fq))
362 			dev_warn(qidev, "New CAAM FQ kill failed\n");
363 
364 		return ret;
365 	}
366 
367 	/*
368 	 * Re-initialise pre-header. Set RSLS and SDLEN.
369 	 * Update the shared descriptor for driver context.
370 	 */
371 	drv_ctx->prehdr[0] = cpu_to_caam32((1 << PREHDR_RSLS_SHIFT) |
372 					   num_words);
373 	drv_ctx->prehdr[1] = cpu_to_caam32(PREHDR_ABS);
374 	memcpy(drv_ctx->sh_desc, sh_desc, desc_bytes(sh_desc));
375 	dma_sync_single_for_device(qidev, drv_ctx->context_a,
376 				   sizeof(drv_ctx->sh_desc) +
377 				   sizeof(drv_ctx->prehdr),
378 				   DMA_BIDIRECTIONAL);
379 
380 	/* Put the new FQ in scheduled state */
381 	ret = qman_schedule_fq(new_fq);
382 	if (ret) {
383 		dev_err(qidev, "Fail to sched new CAAM FQ, ecode = %d\n", ret);
384 
385 		/*
386 		 * We can kill new FQ and revert to old FQ.
387 		 * Since the desc is already modified, it is success case
388 		 */
389 
390 		drv_ctx->req_fq = old_fq;
391 
392 		if (kill_fq(qidev, new_fq))
393 			dev_warn(qidev, "New CAAM FQ kill failed\n");
394 	} else if (kill_fq(qidev, old_fq)) {
395 		dev_warn(qidev, "Old CAAM FQ kill failed\n");
396 	}
397 
398 	return 0;
399 }
400 EXPORT_SYMBOL(caam_drv_ctx_update);
401 
caam_drv_ctx_init(struct device * qidev,int * cpu,u32 * sh_desc)402 struct caam_drv_ctx *caam_drv_ctx_init(struct device *qidev,
403 				       int *cpu,
404 				       u32 *sh_desc)
405 {
406 	size_t size;
407 	u32 num_words;
408 	dma_addr_t hwdesc;
409 	struct caam_drv_ctx *drv_ctx;
410 	const cpumask_t *cpus = qman_affine_cpus();
411 
412 	num_words = desc_len(sh_desc);
413 	if (num_words > MAX_SDLEN) {
414 		dev_err(qidev, "Invalid descriptor len: %d words\n",
415 			num_words);
416 		return ERR_PTR(-EINVAL);
417 	}
418 
419 	drv_ctx = kzalloc(sizeof(*drv_ctx), GFP_ATOMIC);
420 	if (!drv_ctx)
421 		return ERR_PTR(-ENOMEM);
422 
423 	/*
424 	 * Initialise pre-header - set RSLS and SDLEN - and shared descriptor
425 	 * and dma-map them.
426 	 */
427 	drv_ctx->prehdr[0] = cpu_to_caam32((1 << PREHDR_RSLS_SHIFT) |
428 					   num_words);
429 	drv_ctx->prehdr[1] = cpu_to_caam32(PREHDR_ABS);
430 	memcpy(drv_ctx->sh_desc, sh_desc, desc_bytes(sh_desc));
431 	size = sizeof(drv_ctx->prehdr) + sizeof(drv_ctx->sh_desc);
432 	hwdesc = dma_map_single(qidev, drv_ctx->prehdr, size,
433 				DMA_BIDIRECTIONAL);
434 	if (dma_mapping_error(qidev, hwdesc)) {
435 		dev_err(qidev, "DMA map error for preheader + shdesc\n");
436 		kfree(drv_ctx);
437 		return ERR_PTR(-ENOMEM);
438 	}
439 	drv_ctx->context_a = hwdesc;
440 
441 	/* If given CPU does not own the portal, choose another one that does */
442 	if (!cpumask_test_cpu(*cpu, cpus)) {
443 		int *pcpu = &get_cpu_var(last_cpu);
444 
445 		*pcpu = cpumask_next(*pcpu, cpus);
446 		if (*pcpu >= nr_cpu_ids)
447 			*pcpu = cpumask_first(cpus);
448 		*cpu = *pcpu;
449 
450 		put_cpu_var(last_cpu);
451 	}
452 	drv_ctx->cpu = *cpu;
453 
454 	/* Find response FQ hooked with this CPU */
455 	drv_ctx->rsp_fq = per_cpu(pcpu_qipriv.rsp_fq, drv_ctx->cpu);
456 
457 	/* Attach request FQ */
458 	drv_ctx->req_fq = create_caam_req_fq(qidev, drv_ctx->rsp_fq, hwdesc,
459 					     QMAN_INITFQ_FLAG_SCHED);
460 	if (IS_ERR(drv_ctx->req_fq)) {
461 		dev_err(qidev, "create_caam_req_fq failed\n");
462 		dma_unmap_single(qidev, hwdesc, size, DMA_BIDIRECTIONAL);
463 		kfree(drv_ctx);
464 		return ERR_PTR(-ENOMEM);
465 	}
466 
467 	/* init reference counter used to track references to request FQ */
468 	refcount_set(&drv_ctx->refcnt, 1);
469 
470 	drv_ctx->qidev = qidev;
471 	return drv_ctx;
472 }
473 EXPORT_SYMBOL(caam_drv_ctx_init);
474 
qi_cache_alloc(gfp_t flags)475 void *qi_cache_alloc(gfp_t flags)
476 {
477 	return kmem_cache_alloc(qi_cache, flags);
478 }
479 EXPORT_SYMBOL(qi_cache_alloc);
480 
qi_cache_free(void * obj)481 void qi_cache_free(void *obj)
482 {
483 	kmem_cache_free(qi_cache, obj);
484 }
485 EXPORT_SYMBOL(qi_cache_free);
486 
caam_qi_poll(struct napi_struct * napi,int budget)487 static int caam_qi_poll(struct napi_struct *napi, int budget)
488 {
489 	struct caam_napi *np = container_of(napi, struct caam_napi, irqtask);
490 
491 	int cleaned = qman_p_poll_dqrr(np->p, budget);
492 
493 	if (cleaned < budget) {
494 		napi_complete(napi);
495 		qman_p_irqsource_add(np->p, QM_PIRQ_DQRI);
496 	}
497 
498 	return cleaned;
499 }
500 
caam_drv_ctx_rel(struct caam_drv_ctx * drv_ctx)501 void caam_drv_ctx_rel(struct caam_drv_ctx *drv_ctx)
502 {
503 	if (IS_ERR_OR_NULL(drv_ctx))
504 		return;
505 
506 	/* Remove request FQ */
507 	if (kill_fq(drv_ctx->qidev, drv_ctx->req_fq))
508 		dev_err(drv_ctx->qidev, "Crypto session req FQ kill failed\n");
509 
510 	dma_unmap_single(drv_ctx->qidev, drv_ctx->context_a,
511 			 sizeof(drv_ctx->sh_desc) + sizeof(drv_ctx->prehdr),
512 			 DMA_BIDIRECTIONAL);
513 	kfree(drv_ctx);
514 }
515 EXPORT_SYMBOL(caam_drv_ctx_rel);
516 
caam_qi_shutdown(void * data)517 static void caam_qi_shutdown(void *data)
518 {
519 	int i;
520 	struct device *qidev = data;
521 	struct caam_qi_priv *priv = &qipriv;
522 	const cpumask_t *cpus = qman_affine_cpus();
523 
524 	for_each_cpu(i, cpus) {
525 		struct napi_struct *irqtask;
526 
527 		irqtask = &per_cpu_ptr(&pcpu_qipriv.caam_napi, i)->irqtask;
528 		napi_disable(irqtask);
529 		netif_napi_del(irqtask);
530 
531 		if (kill_fq(qidev, per_cpu(pcpu_qipriv.rsp_fq, i)))
532 			dev_err(qidev, "Rsp FQ kill failed, cpu: %d\n", i);
533 		free_netdev(per_cpu(pcpu_qipriv.net_dev, i));
534 	}
535 
536 	qman_delete_cgr_safe(&priv->cgr);
537 	qman_release_cgrid(priv->cgr.cgrid);
538 
539 	kmem_cache_destroy(qi_cache);
540 }
541 
cgr_cb(struct qman_portal * qm,struct qman_cgr * cgr,int congested)542 static void cgr_cb(struct qman_portal *qm, struct qman_cgr *cgr, int congested)
543 {
544 	caam_congested = congested;
545 
546 	if (congested) {
547 		caam_debugfs_qi_congested();
548 
549 		pr_debug_ratelimited("CAAM entered congestion\n");
550 
551 	} else {
552 		pr_debug_ratelimited("CAAM exited congestion\n");
553 	}
554 }
555 
caam_qi_napi_schedule(struct qman_portal * p,struct caam_napi * np,bool sched_napi)556 static int caam_qi_napi_schedule(struct qman_portal *p, struct caam_napi *np,
557 				 bool sched_napi)
558 {
559 	if (sched_napi) {
560 		/* Disable QMan IRQ source and invoke NAPI */
561 		qman_p_irqsource_remove(p, QM_PIRQ_DQRI);
562 		np->p = p;
563 		napi_schedule(&np->irqtask);
564 		return 1;
565 	}
566 	return 0;
567 }
568 
caam_rsp_fq_dqrr_cb(struct qman_portal * p,struct qman_fq * rsp_fq,const struct qm_dqrr_entry * dqrr,bool sched_napi)569 static enum qman_cb_dqrr_result caam_rsp_fq_dqrr_cb(struct qman_portal *p,
570 						    struct qman_fq *rsp_fq,
571 						    const struct qm_dqrr_entry *dqrr,
572 						    bool sched_napi)
573 {
574 	struct caam_napi *caam_napi = raw_cpu_ptr(&pcpu_qipriv.caam_napi);
575 	struct caam_drv_req *drv_req;
576 	const struct qm_fd *fd;
577 	struct device *qidev = &(raw_cpu_ptr(&pcpu_qipriv)->net_dev->dev);
578 	struct caam_drv_private *priv = dev_get_drvdata(qidev);
579 	u32 status;
580 
581 	if (caam_qi_napi_schedule(p, caam_napi, sched_napi))
582 		return qman_cb_dqrr_stop;
583 
584 	fd = &dqrr->fd;
585 
586 	drv_req = caam_iova_to_virt(priv->domain, qm_fd_addr_get64(fd));
587 	if (unlikely(!drv_req)) {
588 		dev_err(qidev,
589 			"Can't find original request for caam response\n");
590 		return qman_cb_dqrr_consume;
591 	}
592 
593 	refcount_dec(&drv_req->drv_ctx->refcnt);
594 
595 	status = be32_to_cpu(fd->status);
596 	if (unlikely(status)) {
597 		u32 ssrc = status & JRSTA_SSRC_MASK;
598 		u8 err_id = status & JRSTA_CCBERR_ERRID_MASK;
599 
600 		if (ssrc != JRSTA_SSRC_CCB_ERROR ||
601 		    err_id != JRSTA_CCBERR_ERRID_ICVCHK)
602 			dev_err_ratelimited(qidev,
603 					    "Error: %#x in CAAM response FD\n",
604 					    status);
605 	}
606 
607 	if (unlikely(qm_fd_get_format(fd) != qm_fd_compound)) {
608 		dev_err(qidev, "Non-compound FD from CAAM\n");
609 		return qman_cb_dqrr_consume;
610 	}
611 
612 	dma_unmap_single(drv_req->drv_ctx->qidev, qm_fd_addr(fd),
613 			 sizeof(drv_req->fd_sgt), DMA_BIDIRECTIONAL);
614 
615 	drv_req->cbk(drv_req, status);
616 	return qman_cb_dqrr_consume;
617 }
618 
alloc_rsp_fq_cpu(struct device * qidev,unsigned int cpu)619 static int alloc_rsp_fq_cpu(struct device *qidev, unsigned int cpu)
620 {
621 	struct qm_mcc_initfq opts;
622 	struct qman_fq *fq;
623 	int ret;
624 
625 	fq = kzalloc(sizeof(*fq), GFP_KERNEL);
626 	if (!fq)
627 		return -ENOMEM;
628 
629 	fq->cb.dqrr = caam_rsp_fq_dqrr_cb;
630 
631 	ret = qman_create_fq(0, QMAN_FQ_FLAG_NO_ENQUEUE |
632 			     QMAN_FQ_FLAG_DYNAMIC_FQID, fq);
633 	if (ret) {
634 		dev_err(qidev, "Rsp FQ create failed\n");
635 		kfree(fq);
636 		return -ENODEV;
637 	}
638 
639 	memset(&opts, 0, sizeof(opts));
640 	opts.we_mask = cpu_to_be16(QM_INITFQ_WE_FQCTRL | QM_INITFQ_WE_DESTWQ |
641 				   QM_INITFQ_WE_CONTEXTB |
642 				   QM_INITFQ_WE_CONTEXTA | QM_INITFQ_WE_CGID);
643 	opts.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_CTXASTASHING |
644 				       QM_FQCTRL_CPCSTASH | QM_FQCTRL_CGE);
645 	qm_fqd_set_destwq(&opts.fqd, qman_affine_channel(cpu), 3);
646 	opts.fqd.cgid = qipriv.cgr.cgrid;
647 	opts.fqd.context_a.stashing.exclusive =	QM_STASHING_EXCL_CTX |
648 						QM_STASHING_EXCL_DATA;
649 	qm_fqd_set_stashing(&opts.fqd, 0, 1, 1);
650 
651 	ret = qman_init_fq(fq, QMAN_INITFQ_FLAG_SCHED, &opts);
652 	if (ret) {
653 		dev_err(qidev, "Rsp FQ init failed\n");
654 		kfree(fq);
655 		return -ENODEV;
656 	}
657 
658 	per_cpu(pcpu_qipriv.rsp_fq, cpu) = fq;
659 
660 	dev_dbg(qidev, "Allocated response FQ %u for CPU %u", fq->fqid, cpu);
661 	return 0;
662 }
663 
init_cgr(struct device * qidev)664 static int init_cgr(struct device *qidev)
665 {
666 	int ret;
667 	struct qm_mcc_initcgr opts;
668 	const u64 val = (u64)cpumask_weight(qman_affine_cpus()) *
669 			MAX_RSP_FQ_BACKLOG_PER_CPU;
670 
671 	ret = qman_alloc_cgrid(&qipriv.cgr.cgrid);
672 	if (ret) {
673 		dev_err(qidev, "CGR alloc failed for rsp FQs: %d\n", ret);
674 		return ret;
675 	}
676 
677 	qipriv.cgr.cb = cgr_cb;
678 	memset(&opts, 0, sizeof(opts));
679 	opts.we_mask = cpu_to_be16(QM_CGR_WE_CSCN_EN | QM_CGR_WE_CS_THRES |
680 				   QM_CGR_WE_MODE);
681 	opts.cgr.cscn_en = QM_CGR_EN;
682 	opts.cgr.mode = QMAN_CGR_MODE_FRAME;
683 	qm_cgr_cs_thres_set64(&opts.cgr.cs_thres, val, 1);
684 
685 	ret = qman_create_cgr(&qipriv.cgr, QMAN_CGR_FLAG_USE_INIT, &opts);
686 	if (ret) {
687 		dev_err(qidev, "Error %d creating CAAM CGRID: %u\n", ret,
688 			qipriv.cgr.cgrid);
689 		return ret;
690 	}
691 
692 	dev_dbg(qidev, "Congestion threshold set to %llu\n", val);
693 	return 0;
694 }
695 
alloc_rsp_fqs(struct device * qidev)696 static int alloc_rsp_fqs(struct device *qidev)
697 {
698 	int ret, i;
699 	const cpumask_t *cpus = qman_affine_cpus();
700 
701 	/*Now create response FQs*/
702 	for_each_cpu(i, cpus) {
703 		ret = alloc_rsp_fq_cpu(qidev, i);
704 		if (ret) {
705 			dev_err(qidev, "CAAM rsp FQ alloc failed, cpu: %u", i);
706 			return ret;
707 		}
708 	}
709 
710 	return 0;
711 }
712 
free_rsp_fqs(void)713 static void free_rsp_fqs(void)
714 {
715 	int i;
716 	const cpumask_t *cpus = qman_affine_cpus();
717 
718 	for_each_cpu(i, cpus)
719 		kfree(per_cpu(pcpu_qipriv.rsp_fq, i));
720 }
721 
free_caam_qi_pcpu_netdev(const cpumask_t * cpus)722 static void free_caam_qi_pcpu_netdev(const cpumask_t *cpus)
723 {
724 	struct caam_qi_pcpu_priv *priv;
725 	int i;
726 
727 	for_each_cpu(i, cpus) {
728 		priv = per_cpu_ptr(&pcpu_qipriv, i);
729 		free_netdev(priv->net_dev);
730 	}
731 }
732 
caam_qi_init(struct platform_device * caam_pdev)733 int caam_qi_init(struct platform_device *caam_pdev)
734 {
735 	int err, i;
736 	struct device *ctrldev = &caam_pdev->dev, *qidev;
737 	struct caam_drv_private *ctrlpriv;
738 	const cpumask_t *cpus = qman_affine_cpus();
739 	cpumask_var_t clean_mask;
740 
741 	err = -ENOMEM;
742 	if (!zalloc_cpumask_var(&clean_mask, GFP_KERNEL))
743 		goto fail_cpumask;
744 
745 	ctrlpriv = dev_get_drvdata(ctrldev);
746 	qidev = ctrldev;
747 
748 	/* Initialize the congestion detection */
749 	err = init_cgr(qidev);
750 	if (err) {
751 		dev_err(qidev, "CGR initialization failed: %d\n", err);
752 		goto fail_cgr;
753 	}
754 
755 	/* Initialise response FQs */
756 	err = alloc_rsp_fqs(qidev);
757 	if (err) {
758 		dev_err(qidev, "Can't allocate CAAM response FQs: %d\n", err);
759 		goto fail_fqs;
760 	}
761 
762 	/*
763 	 * Enable the NAPI contexts on each of the core which has an affine
764 	 * portal.
765 	 */
766 	for_each_cpu(i, cpus) {
767 		struct caam_qi_pcpu_priv *priv = per_cpu_ptr(&pcpu_qipriv, i);
768 		struct caam_napi *caam_napi = &priv->caam_napi;
769 		struct napi_struct *irqtask = &caam_napi->irqtask;
770 		struct net_device *net_dev;
771 
772 		net_dev = alloc_netdev_dummy(0);
773 		if (!net_dev) {
774 			err = -ENOMEM;
775 			goto fail;
776 		}
777 		cpumask_set_cpu(i, clean_mask);
778 		priv->net_dev = net_dev;
779 		net_dev->dev = *qidev;
780 
781 		netif_napi_add_tx_weight(net_dev, irqtask, caam_qi_poll,
782 					 CAAM_NAPI_WEIGHT);
783 
784 		napi_enable(irqtask);
785 	}
786 
787 	qi_cache = kmem_cache_create("caamqicache", CAAM_QI_MEMCACHE_SIZE,
788 				     dma_get_cache_alignment(), 0, NULL);
789 	if (!qi_cache) {
790 		dev_err(qidev, "Can't allocate CAAM cache\n");
791 		err = -ENOMEM;
792 		goto fail;
793 	}
794 
795 	caam_debugfs_qi_init(ctrlpriv);
796 
797 	err = devm_add_action_or_reset(qidev, caam_qi_shutdown, ctrlpriv);
798 	if (err)
799 		goto fail2;
800 
801 	dev_info(qidev, "Linux CAAM Queue I/F driver initialised\n");
802 	goto free_cpumask;
803 
804 fail2:
805 	kmem_cache_destroy(qi_cache);
806 fail:
807 	free_caam_qi_pcpu_netdev(clean_mask);
808 fail_fqs:
809 	free_rsp_fqs();
810 	qman_delete_cgr_safe(&qipriv.cgr);
811 	qman_release_cgrid(qipriv.cgr.cgrid);
812 fail_cgr:
813 free_cpumask:
814 	free_cpumask_var(clean_mask);
815 fail_cpumask:
816 	return err;
817 }
818