1 // SPDX-License-Identifier: GPL-2.0
3 * CAAM/SEC 4.x QI transport/backend driver
4 * Queue Interface backend functionality
6 * Copyright 2013-2016 Freescale Semiconductor, Inc.
7 * Copyright 2016-2017, 2019 NXP
10 #include <linux/cpumask.h>
11 #include <linux/kthread.h>
12 #include <soc/fsl/qman.h>
18 #include "desc_constr.h"
20 #define PREHDR_RSLS_SHIFT 31
21 #define PREHDR_ABS BIT(25)
24 * Use a reasonable backlog of frames (per CPU) as congestion threshold,
25 * so that resources used by the in-flight buffers do not become a memory hog.
27 #define MAX_RSP_FQ_BACKLOG_PER_CPU 256
29 #define CAAM_QI_ENQUEUE_RETRIES 10000
31 #define CAAM_NAPI_WEIGHT 63
34 * caam_napi - struct holding CAAM NAPI-related params
35 * @irqtask: IRQ task for QI backend
39 struct napi_struct irqtask;
40 struct qman_portal *p;
44 * caam_qi_pcpu_priv - percpu private data structure to main list of pending
45 * responses expected on each cpu.
46 * @caam_napi: CAAM NAPI params
47 * @net_dev: netdev used by NAPI
48 * @rsp_fq: response FQ from CAAM
50 struct caam_qi_pcpu_priv {
51 struct caam_napi caam_napi;
52 struct net_device net_dev;
53 struct qman_fq *rsp_fq;
54 } ____cacheline_aligned;
56 static DEFINE_PER_CPU(struct caam_qi_pcpu_priv, pcpu_qipriv);
57 static DEFINE_PER_CPU(int, last_cpu);
60 * caam_qi_priv - CAAM QI backend private params
61 * @cgr: QMan congestion group
67 static struct caam_qi_priv qipriv ____cacheline_aligned;
70 * This is written by only one core - the one that initialized the CGR - and
71 * read by multiple cores (all the others).
73 bool caam_congested __read_mostly;
74 EXPORT_SYMBOL(caam_congested);
76 #ifdef CONFIG_DEBUG_FS
78 * This is a counter for the number of times the congestion group (where all
79 * the request and response queueus are) reached congestion. Incremented
80 * each time the congestion callback is called with congested == true.
82 static u64 times_congested;
86 * This is a a cache of buffers, from which the users of CAAM QI driver
87 * can allocate short (CAAM_QI_MEMCACHE_SIZE) buffers. It's faster than
88 * doing malloc on the hotpath.
89 * NOTE: A more elegant solution would be to have some headroom in the frames
90 * being processed. This could be added by the dpaa-ethernet driver.
91 * This would pose a problem for userspace application processing which
92 * cannot know of this limitation. So for now, this will work.
93 * NOTE: The memcache is SMP-safe. No need to handle spinlocks in-here
95 static struct kmem_cache *qi_cache;
97 static void *caam_iova_to_virt(struct iommu_domain *domain,
100 phys_addr_t phys_addr;
102 phys_addr = domain ? iommu_iova_to_phys(domain, iova_addr) : iova_addr;
104 return phys_to_virt(phys_addr);
107 int caam_qi_enqueue(struct device *qidev, struct caam_drv_req *req)
115 qm_fd_set_compound(&fd, qm_sg_entry_get_len(&req->fd_sgt[1]));
117 addr = dma_map_single(qidev, req->fd_sgt, sizeof(req->fd_sgt),
119 if (dma_mapping_error(qidev, addr)) {
120 dev_err(qidev, "DMA mapping error for QI enqueue request\n");
123 qm_fd_addr_set64(&fd, addr);
126 ret = qman_enqueue(req->drv_ctx->req_fq, &fd);
133 } while (num_retries < CAAM_QI_ENQUEUE_RETRIES);
135 dev_err(qidev, "qman_enqueue failed: %d\n", ret);
139 EXPORT_SYMBOL(caam_qi_enqueue);
141 static void caam_fq_ern_cb(struct qman_portal *qm, struct qman_fq *fq,
142 const union qm_mr_entry *msg)
144 const struct qm_fd *fd;
145 struct caam_drv_req *drv_req;
146 struct device *qidev = &(raw_cpu_ptr(&pcpu_qipriv)->net_dev.dev);
147 struct caam_drv_private *priv = dev_get_drvdata(qidev);
151 if (qm_fd_get_format(fd) != qm_fd_compound) {
152 dev_err(qidev, "Non-compound FD from CAAM\n");
156 drv_req = caam_iova_to_virt(priv->domain, qm_fd_addr_get64(fd));
159 "Can't find original request for CAAM response\n");
163 dma_unmap_single(drv_req->drv_ctx->qidev, qm_fd_addr(fd),
164 sizeof(drv_req->fd_sgt), DMA_BIDIRECTIONAL);
167 drv_req->cbk(drv_req, be32_to_cpu(fd->status));
169 drv_req->cbk(drv_req, JRSTA_SSRC_QI);
172 static struct qman_fq *create_caam_req_fq(struct device *qidev,
173 struct qman_fq *rsp_fq,
178 struct qman_fq *req_fq;
179 struct qm_mcc_initfq opts;
181 req_fq = kzalloc(sizeof(*req_fq), GFP_ATOMIC);
183 return ERR_PTR(-ENOMEM);
185 req_fq->cb.ern = caam_fq_ern_cb;
186 req_fq->cb.fqs = NULL;
188 ret = qman_create_fq(0, QMAN_FQ_FLAG_DYNAMIC_FQID |
189 QMAN_FQ_FLAG_TO_DCPORTAL, req_fq);
191 dev_err(qidev, "Failed to create session req FQ\n");
192 goto create_req_fq_fail;
195 memset(&opts, 0, sizeof(opts));
196 opts.we_mask = cpu_to_be16(QM_INITFQ_WE_FQCTRL | QM_INITFQ_WE_DESTWQ |
197 QM_INITFQ_WE_CONTEXTB |
198 QM_INITFQ_WE_CONTEXTA | QM_INITFQ_WE_CGID);
199 opts.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_CPCSTASH | QM_FQCTRL_CGE);
200 qm_fqd_set_destwq(&opts.fqd, qm_channel_caam, 2);
201 opts.fqd.context_b = cpu_to_be32(qman_fq_fqid(rsp_fq));
202 qm_fqd_context_a_set64(&opts.fqd, hwdesc);
203 opts.fqd.cgid = qipriv.cgr.cgrid;
205 ret = qman_init_fq(req_fq, fq_sched_flag, &opts);
207 dev_err(qidev, "Failed to init session req FQ\n");
208 goto init_req_fq_fail;
211 dev_dbg(qidev, "Allocated request FQ %u for CPU %u\n", req_fq->fqid,
216 qman_destroy_fq(req_fq);
222 static int empty_retired_fq(struct device *qidev, struct qman_fq *fq)
226 ret = qman_volatile_dequeue(fq, QMAN_VOLATILE_FLAG_WAIT_INT |
227 QMAN_VOLATILE_FLAG_FINISH,
228 QM_VDQCR_PRECEDENCE_VDQCR |
229 QM_VDQCR_NUMFRAMES_TILLEMPTY);
231 dev_err(qidev, "Volatile dequeue fail for FQ: %u\n", fq->fqid);
236 struct qman_portal *p;
238 p = qman_get_affine_portal(smp_processor_id());
239 qman_p_poll_dqrr(p, 16);
240 } while (fq->flags & QMAN_FQ_STATE_NE);
245 static int kill_fq(struct device *qidev, struct qman_fq *fq)
250 ret = qman_retire_fq(fq, &flags);
252 dev_err(qidev, "qman_retire_fq failed: %d\n", ret);
259 /* Async FQ retirement condition */
261 /* Retry till FQ gets in retired state */
264 } while (fq->state != qman_fq_state_retired);
266 WARN_ON(fq->flags & QMAN_FQ_STATE_BLOCKOOS);
267 WARN_ON(fq->flags & QMAN_FQ_STATE_ORL);
271 if (fq->flags & QMAN_FQ_STATE_NE) {
272 ret = empty_retired_fq(qidev, fq);
274 dev_err(qidev, "empty_retired_fq fail for FQ: %u\n",
280 ret = qman_oos_fq(fq);
282 dev_err(qidev, "OOS of FQID: %u failed\n", fq->fqid);
290 static int empty_caam_fq(struct qman_fq *fq)
293 struct qm_mcr_queryfq_np np;
295 /* Wait till the older CAAM FQ get empty */
297 ret = qman_query_fq_np(fq, &np);
301 if (!qm_mcr_np_get(&np, frm_cnt))
308 * Give extra time for pending jobs from this FQ in holding tanks
315 int caam_drv_ctx_update(struct caam_drv_ctx *drv_ctx, u32 *sh_desc)
319 struct qman_fq *new_fq, *old_fq;
320 struct device *qidev = drv_ctx->qidev;
322 num_words = desc_len(sh_desc);
323 if (num_words > MAX_SDLEN) {
324 dev_err(qidev, "Invalid descriptor len: %d words\n", num_words);
328 /* Note down older req FQ */
329 old_fq = drv_ctx->req_fq;
331 /* Create a new req FQ in parked state */
332 new_fq = create_caam_req_fq(drv_ctx->qidev, drv_ctx->rsp_fq,
333 drv_ctx->context_a, 0);
334 if (IS_ERR(new_fq)) {
335 dev_err(qidev, "FQ allocation for shdesc update failed\n");
336 return PTR_ERR(new_fq);
339 /* Hook up new FQ to context so that new requests keep queuing */
340 drv_ctx->req_fq = new_fq;
342 /* Empty and remove the older FQ */
343 ret = empty_caam_fq(old_fq);
345 dev_err(qidev, "Old CAAM FQ empty failed: %d\n", ret);
347 /* We can revert to older FQ */
348 drv_ctx->req_fq = old_fq;
350 if (kill_fq(qidev, new_fq))
351 dev_warn(qidev, "New CAAM FQ kill failed\n");
357 * Re-initialise pre-header. Set RSLS and SDLEN.
358 * Update the shared descriptor for driver context.
360 drv_ctx->prehdr[0] = cpu_to_caam32((1 << PREHDR_RSLS_SHIFT) |
362 drv_ctx->prehdr[1] = cpu_to_caam32(PREHDR_ABS);
363 memcpy(drv_ctx->sh_desc, sh_desc, desc_bytes(sh_desc));
364 dma_sync_single_for_device(qidev, drv_ctx->context_a,
365 sizeof(drv_ctx->sh_desc) +
366 sizeof(drv_ctx->prehdr),
369 /* Put the new FQ in scheduled state */
370 ret = qman_schedule_fq(new_fq);
372 dev_err(qidev, "Fail to sched new CAAM FQ, ecode = %d\n", ret);
375 * We can kill new FQ and revert to old FQ.
376 * Since the desc is already modified, it is success case
379 drv_ctx->req_fq = old_fq;
381 if (kill_fq(qidev, new_fq))
382 dev_warn(qidev, "New CAAM FQ kill failed\n");
383 } else if (kill_fq(qidev, old_fq)) {
384 dev_warn(qidev, "Old CAAM FQ kill failed\n");
389 EXPORT_SYMBOL(caam_drv_ctx_update);
391 struct caam_drv_ctx *caam_drv_ctx_init(struct device *qidev,
398 struct caam_drv_ctx *drv_ctx;
399 const cpumask_t *cpus = qman_affine_cpus();
401 num_words = desc_len(sh_desc);
402 if (num_words > MAX_SDLEN) {
403 dev_err(qidev, "Invalid descriptor len: %d words\n",
405 return ERR_PTR(-EINVAL);
408 drv_ctx = kzalloc(sizeof(*drv_ctx), GFP_ATOMIC);
410 return ERR_PTR(-ENOMEM);
413 * Initialise pre-header - set RSLS and SDLEN - and shared descriptor
416 drv_ctx->prehdr[0] = cpu_to_caam32((1 << PREHDR_RSLS_SHIFT) |
418 drv_ctx->prehdr[1] = cpu_to_caam32(PREHDR_ABS);
419 memcpy(drv_ctx->sh_desc, sh_desc, desc_bytes(sh_desc));
420 size = sizeof(drv_ctx->prehdr) + sizeof(drv_ctx->sh_desc);
421 hwdesc = dma_map_single(qidev, drv_ctx->prehdr, size,
423 if (dma_mapping_error(qidev, hwdesc)) {
424 dev_err(qidev, "DMA map error for preheader + shdesc\n");
426 return ERR_PTR(-ENOMEM);
428 drv_ctx->context_a = hwdesc;
430 /* If given CPU does not own the portal, choose another one that does */
431 if (!cpumask_test_cpu(*cpu, cpus)) {
432 int *pcpu = &get_cpu_var(last_cpu);
434 *pcpu = cpumask_next(*pcpu, cpus);
435 if (*pcpu >= nr_cpu_ids)
436 *pcpu = cpumask_first(cpus);
439 put_cpu_var(last_cpu);
443 /* Find response FQ hooked with this CPU */
444 drv_ctx->rsp_fq = per_cpu(pcpu_qipriv.rsp_fq, drv_ctx->cpu);
446 /* Attach request FQ */
447 drv_ctx->req_fq = create_caam_req_fq(qidev, drv_ctx->rsp_fq, hwdesc,
448 QMAN_INITFQ_FLAG_SCHED);
449 if (IS_ERR(drv_ctx->req_fq)) {
450 dev_err(qidev, "create_caam_req_fq failed\n");
451 dma_unmap_single(qidev, hwdesc, size, DMA_BIDIRECTIONAL);
453 return ERR_PTR(-ENOMEM);
456 drv_ctx->qidev = qidev;
459 EXPORT_SYMBOL(caam_drv_ctx_init);
461 void *qi_cache_alloc(gfp_t flags)
463 return kmem_cache_alloc(qi_cache, flags);
465 EXPORT_SYMBOL(qi_cache_alloc);
467 void qi_cache_free(void *obj)
469 kmem_cache_free(qi_cache, obj);
471 EXPORT_SYMBOL(qi_cache_free);
473 static int caam_qi_poll(struct napi_struct *napi, int budget)
475 struct caam_napi *np = container_of(napi, struct caam_napi, irqtask);
477 int cleaned = qman_p_poll_dqrr(np->p, budget);
479 if (cleaned < budget) {
481 qman_p_irqsource_add(np->p, QM_PIRQ_DQRI);
487 void caam_drv_ctx_rel(struct caam_drv_ctx *drv_ctx)
489 if (IS_ERR_OR_NULL(drv_ctx))
492 /* Remove request FQ */
493 if (kill_fq(drv_ctx->qidev, drv_ctx->req_fq))
494 dev_err(drv_ctx->qidev, "Crypto session req FQ kill failed\n");
496 dma_unmap_single(drv_ctx->qidev, drv_ctx->context_a,
497 sizeof(drv_ctx->sh_desc) + sizeof(drv_ctx->prehdr),
501 EXPORT_SYMBOL(caam_drv_ctx_rel);
503 void caam_qi_shutdown(struct device *qidev)
506 struct caam_qi_priv *priv = &qipriv;
507 const cpumask_t *cpus = qman_affine_cpus();
509 for_each_cpu(i, cpus) {
510 struct napi_struct *irqtask;
512 irqtask = &per_cpu_ptr(&pcpu_qipriv.caam_napi, i)->irqtask;
513 napi_disable(irqtask);
514 netif_napi_del(irqtask);
516 if (kill_fq(qidev, per_cpu(pcpu_qipriv.rsp_fq, i)))
517 dev_err(qidev, "Rsp FQ kill failed, cpu: %d\n", i);
520 qman_delete_cgr_safe(&priv->cgr);
521 qman_release_cgrid(priv->cgr.cgrid);
523 kmem_cache_destroy(qi_cache);
526 static void cgr_cb(struct qman_portal *qm, struct qman_cgr *cgr, int congested)
528 caam_congested = congested;
531 #ifdef CONFIG_DEBUG_FS
534 pr_debug_ratelimited("CAAM entered congestion\n");
537 pr_debug_ratelimited("CAAM exited congestion\n");
541 static int caam_qi_napi_schedule(struct qman_portal *p, struct caam_napi *np)
544 * In case of threaded ISR, for RT kernels in_irq() does not return
545 * appropriate value, so use in_serving_softirq to distinguish between
546 * softirq and irq contexts.
548 if (unlikely(in_irq() || !in_serving_softirq())) {
549 /* Disable QMan IRQ source and invoke NAPI */
550 qman_p_irqsource_remove(p, QM_PIRQ_DQRI);
552 napi_schedule(&np->irqtask);
558 static enum qman_cb_dqrr_result caam_rsp_fq_dqrr_cb(struct qman_portal *p,
559 struct qman_fq *rsp_fq,
560 const struct qm_dqrr_entry *dqrr)
562 struct caam_napi *caam_napi = raw_cpu_ptr(&pcpu_qipriv.caam_napi);
563 struct caam_drv_req *drv_req;
564 const struct qm_fd *fd;
565 struct device *qidev = &(raw_cpu_ptr(&pcpu_qipriv)->net_dev.dev);
566 struct caam_drv_private *priv = dev_get_drvdata(qidev);
569 if (caam_qi_napi_schedule(p, caam_napi))
570 return qman_cb_dqrr_stop;
573 status = be32_to_cpu(fd->status);
574 if (unlikely(status)) {
575 u32 ssrc = status & JRSTA_SSRC_MASK;
576 u8 err_id = status & JRSTA_CCBERR_ERRID_MASK;
578 if (ssrc != JRSTA_SSRC_CCB_ERROR ||
579 err_id != JRSTA_CCBERR_ERRID_ICVCHK)
580 dev_err_ratelimited(qidev,
581 "Error: %#x in CAAM response FD\n",
585 if (unlikely(qm_fd_get_format(fd) != qm_fd_compound)) {
586 dev_err(qidev, "Non-compound FD from CAAM\n");
587 return qman_cb_dqrr_consume;
590 drv_req = caam_iova_to_virt(priv->domain, qm_fd_addr_get64(fd));
591 if (unlikely(!drv_req)) {
593 "Can't find original request for caam response\n");
594 return qman_cb_dqrr_consume;
597 dma_unmap_single(drv_req->drv_ctx->qidev, qm_fd_addr(fd),
598 sizeof(drv_req->fd_sgt), DMA_BIDIRECTIONAL);
600 drv_req->cbk(drv_req, status);
601 return qman_cb_dqrr_consume;
604 static int alloc_rsp_fq_cpu(struct device *qidev, unsigned int cpu)
606 struct qm_mcc_initfq opts;
610 fq = kzalloc(sizeof(*fq), GFP_KERNEL | GFP_DMA);
614 fq->cb.dqrr = caam_rsp_fq_dqrr_cb;
616 ret = qman_create_fq(0, QMAN_FQ_FLAG_NO_ENQUEUE |
617 QMAN_FQ_FLAG_DYNAMIC_FQID, fq);
619 dev_err(qidev, "Rsp FQ create failed\n");
624 memset(&opts, 0, sizeof(opts));
625 opts.we_mask = cpu_to_be16(QM_INITFQ_WE_FQCTRL | QM_INITFQ_WE_DESTWQ |
626 QM_INITFQ_WE_CONTEXTB |
627 QM_INITFQ_WE_CONTEXTA | QM_INITFQ_WE_CGID);
628 opts.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_CTXASTASHING |
629 QM_FQCTRL_CPCSTASH | QM_FQCTRL_CGE);
630 qm_fqd_set_destwq(&opts.fqd, qman_affine_channel(cpu), 3);
631 opts.fqd.cgid = qipriv.cgr.cgrid;
632 opts.fqd.context_a.stashing.exclusive = QM_STASHING_EXCL_CTX |
633 QM_STASHING_EXCL_DATA;
634 qm_fqd_set_stashing(&opts.fqd, 0, 1, 1);
636 ret = qman_init_fq(fq, QMAN_INITFQ_FLAG_SCHED, &opts);
638 dev_err(qidev, "Rsp FQ init failed\n");
643 per_cpu(pcpu_qipriv.rsp_fq, cpu) = fq;
645 dev_dbg(qidev, "Allocated response FQ %u for CPU %u", fq->fqid, cpu);
649 static int init_cgr(struct device *qidev)
652 struct qm_mcc_initcgr opts;
653 const u64 val = (u64)cpumask_weight(qman_affine_cpus()) *
654 MAX_RSP_FQ_BACKLOG_PER_CPU;
656 ret = qman_alloc_cgrid(&qipriv.cgr.cgrid);
658 dev_err(qidev, "CGR alloc failed for rsp FQs: %d\n", ret);
662 qipriv.cgr.cb = cgr_cb;
663 memset(&opts, 0, sizeof(opts));
664 opts.we_mask = cpu_to_be16(QM_CGR_WE_CSCN_EN | QM_CGR_WE_CS_THRES |
666 opts.cgr.cscn_en = QM_CGR_EN;
667 opts.cgr.mode = QMAN_CGR_MODE_FRAME;
668 qm_cgr_cs_thres_set64(&opts.cgr.cs_thres, val, 1);
670 ret = qman_create_cgr(&qipriv.cgr, QMAN_CGR_FLAG_USE_INIT, &opts);
672 dev_err(qidev, "Error %d creating CAAM CGRID: %u\n", ret,
677 dev_dbg(qidev, "Congestion threshold set to %llu\n", val);
681 static int alloc_rsp_fqs(struct device *qidev)
684 const cpumask_t *cpus = qman_affine_cpus();
686 /*Now create response FQs*/
687 for_each_cpu(i, cpus) {
688 ret = alloc_rsp_fq_cpu(qidev, i);
690 dev_err(qidev, "CAAM rsp FQ alloc failed, cpu: %u", i);
698 static void free_rsp_fqs(void)
701 const cpumask_t *cpus = qman_affine_cpus();
703 for_each_cpu(i, cpus)
704 kfree(per_cpu(pcpu_qipriv.rsp_fq, i));
707 int caam_qi_init(struct platform_device *caam_pdev)
710 struct device *ctrldev = &caam_pdev->dev, *qidev;
711 struct caam_drv_private *ctrlpriv;
712 const cpumask_t *cpus = qman_affine_cpus();
714 ctrlpriv = dev_get_drvdata(ctrldev);
717 /* Initialize the congestion detection */
718 err = init_cgr(qidev);
720 dev_err(qidev, "CGR initialization failed: %d\n", err);
724 /* Initialise response FQs */
725 err = alloc_rsp_fqs(qidev);
727 dev_err(qidev, "Can't allocate CAAM response FQs: %d\n", err);
733 * Enable the NAPI contexts on each of the core which has an affine
736 for_each_cpu(i, cpus) {
737 struct caam_qi_pcpu_priv *priv = per_cpu_ptr(&pcpu_qipriv, i);
738 struct caam_napi *caam_napi = &priv->caam_napi;
739 struct napi_struct *irqtask = &caam_napi->irqtask;
740 struct net_device *net_dev = &priv->net_dev;
742 net_dev->dev = *qidev;
743 INIT_LIST_HEAD(&net_dev->napi_list);
745 netif_napi_add(net_dev, irqtask, caam_qi_poll,
748 napi_enable(irqtask);
751 qi_cache = kmem_cache_create("caamqicache", CAAM_QI_MEMCACHE_SIZE, 0,
752 SLAB_CACHE_DMA, NULL);
754 dev_err(qidev, "Can't allocate CAAM cache\n");
759 #ifdef CONFIG_DEBUG_FS
760 debugfs_create_file("qi_congested", 0444, ctrlpriv->ctl,
761 ×_congested, &caam_fops_u64_ro);
764 ctrlpriv->qi_init = 1;
765 dev_info(qidev, "Linux CAAM Queue I/F driver initialised\n");