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, 2019-2020 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);
128 refcount_inc(&req->drv_ctx->refcnt);
135 } while (num_retries < CAAM_QI_ENQUEUE_RETRIES);
137 dev_err(qidev, "qman_enqueue failed: %d\n", ret);
141 EXPORT_SYMBOL(caam_qi_enqueue);
143 static void caam_fq_ern_cb(struct qman_portal *qm, struct qman_fq *fq,
144 const union qm_mr_entry *msg)
146 const struct qm_fd *fd;
147 struct caam_drv_req *drv_req;
148 struct device *qidev = &(raw_cpu_ptr(&pcpu_qipriv)->net_dev.dev);
149 struct caam_drv_private *priv = dev_get_drvdata(qidev);
153 drv_req = caam_iova_to_virt(priv->domain, qm_fd_addr_get64(fd));
156 "Can't find original request for CAAM response\n");
160 refcount_dec(&drv_req->drv_ctx->refcnt);
162 if (qm_fd_get_format(fd) != qm_fd_compound) {
163 dev_err(qidev, "Non-compound FD from CAAM\n");
167 dma_unmap_single(drv_req->drv_ctx->qidev, qm_fd_addr(fd),
168 sizeof(drv_req->fd_sgt), DMA_BIDIRECTIONAL);
171 drv_req->cbk(drv_req, be32_to_cpu(fd->status));
173 drv_req->cbk(drv_req, JRSTA_SSRC_QI);
176 static struct qman_fq *create_caam_req_fq(struct device *qidev,
177 struct qman_fq *rsp_fq,
182 struct qman_fq *req_fq;
183 struct qm_mcc_initfq opts;
185 req_fq = kzalloc(sizeof(*req_fq), GFP_ATOMIC);
187 return ERR_PTR(-ENOMEM);
189 req_fq->cb.ern = caam_fq_ern_cb;
190 req_fq->cb.fqs = NULL;
192 ret = qman_create_fq(0, QMAN_FQ_FLAG_DYNAMIC_FQID |
193 QMAN_FQ_FLAG_TO_DCPORTAL, req_fq);
195 dev_err(qidev, "Failed to create session req FQ\n");
196 goto create_req_fq_fail;
199 memset(&opts, 0, sizeof(opts));
200 opts.we_mask = cpu_to_be16(QM_INITFQ_WE_FQCTRL | QM_INITFQ_WE_DESTWQ |
201 QM_INITFQ_WE_CONTEXTB |
202 QM_INITFQ_WE_CONTEXTA | QM_INITFQ_WE_CGID);
203 opts.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_CPCSTASH | QM_FQCTRL_CGE);
204 qm_fqd_set_destwq(&opts.fqd, qm_channel_caam, 2);
205 opts.fqd.context_b = cpu_to_be32(qman_fq_fqid(rsp_fq));
206 qm_fqd_context_a_set64(&opts.fqd, hwdesc);
207 opts.fqd.cgid = qipriv.cgr.cgrid;
209 ret = qman_init_fq(req_fq, fq_sched_flag, &opts);
211 dev_err(qidev, "Failed to init session req FQ\n");
212 goto init_req_fq_fail;
215 dev_dbg(qidev, "Allocated request FQ %u for CPU %u\n", req_fq->fqid,
220 qman_destroy_fq(req_fq);
226 static int empty_retired_fq(struct device *qidev, struct qman_fq *fq)
230 ret = qman_volatile_dequeue(fq, QMAN_VOLATILE_FLAG_WAIT_INT |
231 QMAN_VOLATILE_FLAG_FINISH,
232 QM_VDQCR_PRECEDENCE_VDQCR |
233 QM_VDQCR_NUMFRAMES_TILLEMPTY);
235 dev_err(qidev, "Volatile dequeue fail for FQ: %u\n", fq->fqid);
240 struct qman_portal *p;
242 p = qman_get_affine_portal(smp_processor_id());
243 qman_p_poll_dqrr(p, 16);
244 } while (fq->flags & QMAN_FQ_STATE_NE);
249 static int kill_fq(struct device *qidev, struct qman_fq *fq)
254 ret = qman_retire_fq(fq, &flags);
256 dev_err(qidev, "qman_retire_fq failed: %d\n", ret);
263 /* Async FQ retirement condition */
265 /* Retry till FQ gets in retired state */
268 } while (fq->state != qman_fq_state_retired);
270 WARN_ON(fq->flags & QMAN_FQ_STATE_BLOCKOOS);
271 WARN_ON(fq->flags & QMAN_FQ_STATE_ORL);
275 if (fq->flags & QMAN_FQ_STATE_NE) {
276 ret = empty_retired_fq(qidev, fq);
278 dev_err(qidev, "empty_retired_fq fail for FQ: %u\n",
284 ret = qman_oos_fq(fq);
286 dev_err(qidev, "OOS of FQID: %u failed\n", fq->fqid);
294 static int empty_caam_fq(struct qman_fq *fq, struct caam_drv_ctx *drv_ctx)
298 struct qm_mcr_queryfq_np np;
300 /* Wait till the older CAAM FQ get empty */
302 ret = qman_query_fq_np(fq, &np);
306 if (!qm_mcr_np_get(&np, frm_cnt))
312 /* Wait until pending jobs from this FQ are processed by CAAM */
314 if (refcount_read(&drv_ctx->refcnt) == 1)
321 dev_warn_once(drv_ctx->qidev, "%d frames from FQID %u still pending in CAAM\n",
322 refcount_read(&drv_ctx->refcnt), fq->fqid);
327 int caam_drv_ctx_update(struct caam_drv_ctx *drv_ctx, u32 *sh_desc)
331 struct qman_fq *new_fq, *old_fq;
332 struct device *qidev = drv_ctx->qidev;
334 num_words = desc_len(sh_desc);
335 if (num_words > MAX_SDLEN) {
336 dev_err(qidev, "Invalid descriptor len: %d words\n", num_words);
340 /* Note down older req FQ */
341 old_fq = drv_ctx->req_fq;
343 /* Create a new req FQ in parked state */
344 new_fq = create_caam_req_fq(drv_ctx->qidev, drv_ctx->rsp_fq,
345 drv_ctx->context_a, 0);
346 if (IS_ERR(new_fq)) {
347 dev_err(qidev, "FQ allocation for shdesc update failed\n");
348 return PTR_ERR(new_fq);
351 /* Hook up new FQ to context so that new requests keep queuing */
352 drv_ctx->req_fq = new_fq;
354 /* Empty and remove the older FQ */
355 ret = empty_caam_fq(old_fq, drv_ctx);
357 dev_err(qidev, "Old CAAM FQ empty failed: %d\n", ret);
359 /* We can revert to older FQ */
360 drv_ctx->req_fq = old_fq;
362 if (kill_fq(qidev, new_fq))
363 dev_warn(qidev, "New CAAM FQ kill failed\n");
369 * Re-initialise pre-header. Set RSLS and SDLEN.
370 * Update the shared descriptor for driver context.
372 drv_ctx->prehdr[0] = cpu_to_caam32((1 << PREHDR_RSLS_SHIFT) |
374 drv_ctx->prehdr[1] = cpu_to_caam32(PREHDR_ABS);
375 memcpy(drv_ctx->sh_desc, sh_desc, desc_bytes(sh_desc));
376 dma_sync_single_for_device(qidev, drv_ctx->context_a,
377 sizeof(drv_ctx->sh_desc) +
378 sizeof(drv_ctx->prehdr),
381 /* Put the new FQ in scheduled state */
382 ret = qman_schedule_fq(new_fq);
384 dev_err(qidev, "Fail to sched new CAAM FQ, ecode = %d\n", ret);
387 * We can kill new FQ and revert to old FQ.
388 * Since the desc is already modified, it is success case
391 drv_ctx->req_fq = old_fq;
393 if (kill_fq(qidev, new_fq))
394 dev_warn(qidev, "New CAAM FQ kill failed\n");
395 } else if (kill_fq(qidev, old_fq)) {
396 dev_warn(qidev, "Old CAAM FQ kill failed\n");
401 EXPORT_SYMBOL(caam_drv_ctx_update);
403 struct caam_drv_ctx *caam_drv_ctx_init(struct device *qidev,
410 struct caam_drv_ctx *drv_ctx;
411 const cpumask_t *cpus = qman_affine_cpus();
413 num_words = desc_len(sh_desc);
414 if (num_words > MAX_SDLEN) {
415 dev_err(qidev, "Invalid descriptor len: %d words\n",
417 return ERR_PTR(-EINVAL);
420 drv_ctx = kzalloc(sizeof(*drv_ctx), GFP_ATOMIC);
422 return ERR_PTR(-ENOMEM);
425 * Initialise pre-header - set RSLS and SDLEN - and shared descriptor
428 drv_ctx->prehdr[0] = cpu_to_caam32((1 << PREHDR_RSLS_SHIFT) |
430 drv_ctx->prehdr[1] = cpu_to_caam32(PREHDR_ABS);
431 memcpy(drv_ctx->sh_desc, sh_desc, desc_bytes(sh_desc));
432 size = sizeof(drv_ctx->prehdr) + sizeof(drv_ctx->sh_desc);
433 hwdesc = dma_map_single(qidev, drv_ctx->prehdr, size,
435 if (dma_mapping_error(qidev, hwdesc)) {
436 dev_err(qidev, "DMA map error for preheader + shdesc\n");
438 return ERR_PTR(-ENOMEM);
440 drv_ctx->context_a = hwdesc;
442 /* If given CPU does not own the portal, choose another one that does */
443 if (!cpumask_test_cpu(*cpu, cpus)) {
444 int *pcpu = &get_cpu_var(last_cpu);
446 *pcpu = cpumask_next(*pcpu, cpus);
447 if (*pcpu >= nr_cpu_ids)
448 *pcpu = cpumask_first(cpus);
451 put_cpu_var(last_cpu);
455 /* Find response FQ hooked with this CPU */
456 drv_ctx->rsp_fq = per_cpu(pcpu_qipriv.rsp_fq, drv_ctx->cpu);
458 /* Attach request FQ */
459 drv_ctx->req_fq = create_caam_req_fq(qidev, drv_ctx->rsp_fq, hwdesc,
460 QMAN_INITFQ_FLAG_SCHED);
461 if (IS_ERR(drv_ctx->req_fq)) {
462 dev_err(qidev, "create_caam_req_fq failed\n");
463 dma_unmap_single(qidev, hwdesc, size, DMA_BIDIRECTIONAL);
465 return ERR_PTR(-ENOMEM);
468 /* init reference counter used to track references to request FQ */
469 refcount_set(&drv_ctx->refcnt, 1);
471 drv_ctx->qidev = qidev;
474 EXPORT_SYMBOL(caam_drv_ctx_init);
476 void *qi_cache_alloc(gfp_t flags)
478 return kmem_cache_alloc(qi_cache, flags);
480 EXPORT_SYMBOL(qi_cache_alloc);
482 void qi_cache_free(void *obj)
484 kmem_cache_free(qi_cache, obj);
486 EXPORT_SYMBOL(qi_cache_free);
488 static int caam_qi_poll(struct napi_struct *napi, int budget)
490 struct caam_napi *np = container_of(napi, struct caam_napi, irqtask);
492 int cleaned = qman_p_poll_dqrr(np->p, budget);
494 if (cleaned < budget) {
496 qman_p_irqsource_add(np->p, QM_PIRQ_DQRI);
502 void caam_drv_ctx_rel(struct caam_drv_ctx *drv_ctx)
504 if (IS_ERR_OR_NULL(drv_ctx))
507 /* Remove request FQ */
508 if (kill_fq(drv_ctx->qidev, drv_ctx->req_fq))
509 dev_err(drv_ctx->qidev, "Crypto session req FQ kill failed\n");
511 dma_unmap_single(drv_ctx->qidev, drv_ctx->context_a,
512 sizeof(drv_ctx->sh_desc) + sizeof(drv_ctx->prehdr),
516 EXPORT_SYMBOL(caam_drv_ctx_rel);
518 static void caam_qi_shutdown(void *data)
521 struct device *qidev = data;
522 struct caam_qi_priv *priv = &qipriv;
523 const cpumask_t *cpus = qman_affine_cpus();
525 for_each_cpu(i, cpus) {
526 struct napi_struct *irqtask;
528 irqtask = &per_cpu_ptr(&pcpu_qipriv.caam_napi, i)->irqtask;
529 napi_disable(irqtask);
530 netif_napi_del(irqtask);
532 if (kill_fq(qidev, per_cpu(pcpu_qipriv.rsp_fq, i)))
533 dev_err(qidev, "Rsp FQ kill failed, cpu: %d\n", i);
536 qman_delete_cgr_safe(&priv->cgr);
537 qman_release_cgrid(priv->cgr.cgrid);
539 kmem_cache_destroy(qi_cache);
542 static void cgr_cb(struct qman_portal *qm, struct qman_cgr *cgr, int congested)
544 caam_congested = congested;
547 #ifdef CONFIG_DEBUG_FS
550 pr_debug_ratelimited("CAAM entered congestion\n");
553 pr_debug_ratelimited("CAAM exited congestion\n");
557 static int caam_qi_napi_schedule(struct qman_portal *p, struct caam_napi *np)
560 * In case of threaded ISR, for RT kernels in_irq() does not return
561 * appropriate value, so use in_serving_softirq to distinguish between
562 * softirq and irq contexts.
564 if (unlikely(in_irq() || !in_serving_softirq())) {
565 /* Disable QMan IRQ source and invoke NAPI */
566 qman_p_irqsource_remove(p, QM_PIRQ_DQRI);
568 napi_schedule(&np->irqtask);
574 static enum qman_cb_dqrr_result caam_rsp_fq_dqrr_cb(struct qman_portal *p,
575 struct qman_fq *rsp_fq,
576 const struct qm_dqrr_entry *dqrr)
578 struct caam_napi *caam_napi = raw_cpu_ptr(&pcpu_qipriv.caam_napi);
579 struct caam_drv_req *drv_req;
580 const struct qm_fd *fd;
581 struct device *qidev = &(raw_cpu_ptr(&pcpu_qipriv)->net_dev.dev);
582 struct caam_drv_private *priv = dev_get_drvdata(qidev);
585 if (caam_qi_napi_schedule(p, caam_napi))
586 return qman_cb_dqrr_stop;
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 refcount_dec(&drv_req->drv_ctx->refcnt);
599 status = be32_to_cpu(fd->status);
600 if (unlikely(status)) {
601 u32 ssrc = status & JRSTA_SSRC_MASK;
602 u8 err_id = status & JRSTA_CCBERR_ERRID_MASK;
604 if (ssrc != JRSTA_SSRC_CCB_ERROR ||
605 err_id != JRSTA_CCBERR_ERRID_ICVCHK)
606 dev_err_ratelimited(qidev,
607 "Error: %#x in CAAM response FD\n",
611 if (unlikely(qm_fd_get_format(fd) != qm_fd_compound)) {
612 dev_err(qidev, "Non-compound FD from CAAM\n");
613 return qman_cb_dqrr_consume;
616 dma_unmap_single(drv_req->drv_ctx->qidev, qm_fd_addr(fd),
617 sizeof(drv_req->fd_sgt), DMA_BIDIRECTIONAL);
619 drv_req->cbk(drv_req, status);
620 return qman_cb_dqrr_consume;
623 static int alloc_rsp_fq_cpu(struct device *qidev, unsigned int cpu)
625 struct qm_mcc_initfq opts;
629 fq = kzalloc(sizeof(*fq), GFP_KERNEL | GFP_DMA);
633 fq->cb.dqrr = caam_rsp_fq_dqrr_cb;
635 ret = qman_create_fq(0, QMAN_FQ_FLAG_NO_ENQUEUE |
636 QMAN_FQ_FLAG_DYNAMIC_FQID, fq);
638 dev_err(qidev, "Rsp FQ create failed\n");
643 memset(&opts, 0, sizeof(opts));
644 opts.we_mask = cpu_to_be16(QM_INITFQ_WE_FQCTRL | QM_INITFQ_WE_DESTWQ |
645 QM_INITFQ_WE_CONTEXTB |
646 QM_INITFQ_WE_CONTEXTA | QM_INITFQ_WE_CGID);
647 opts.fqd.fq_ctrl = cpu_to_be16(QM_FQCTRL_CTXASTASHING |
648 QM_FQCTRL_CPCSTASH | QM_FQCTRL_CGE);
649 qm_fqd_set_destwq(&opts.fqd, qman_affine_channel(cpu), 3);
650 opts.fqd.cgid = qipriv.cgr.cgrid;
651 opts.fqd.context_a.stashing.exclusive = QM_STASHING_EXCL_CTX |
652 QM_STASHING_EXCL_DATA;
653 qm_fqd_set_stashing(&opts.fqd, 0, 1, 1);
655 ret = qman_init_fq(fq, QMAN_INITFQ_FLAG_SCHED, &opts);
657 dev_err(qidev, "Rsp FQ init failed\n");
662 per_cpu(pcpu_qipriv.rsp_fq, cpu) = fq;
664 dev_dbg(qidev, "Allocated response FQ %u for CPU %u", fq->fqid, cpu);
668 static int init_cgr(struct device *qidev)
671 struct qm_mcc_initcgr opts;
672 const u64 val = (u64)cpumask_weight(qman_affine_cpus()) *
673 MAX_RSP_FQ_BACKLOG_PER_CPU;
675 ret = qman_alloc_cgrid(&qipriv.cgr.cgrid);
677 dev_err(qidev, "CGR alloc failed for rsp FQs: %d\n", ret);
681 qipriv.cgr.cb = cgr_cb;
682 memset(&opts, 0, sizeof(opts));
683 opts.we_mask = cpu_to_be16(QM_CGR_WE_CSCN_EN | QM_CGR_WE_CS_THRES |
685 opts.cgr.cscn_en = QM_CGR_EN;
686 opts.cgr.mode = QMAN_CGR_MODE_FRAME;
687 qm_cgr_cs_thres_set64(&opts.cgr.cs_thres, val, 1);
689 ret = qman_create_cgr(&qipriv.cgr, QMAN_CGR_FLAG_USE_INIT, &opts);
691 dev_err(qidev, "Error %d creating CAAM CGRID: %u\n", ret,
696 dev_dbg(qidev, "Congestion threshold set to %llu\n", val);
700 static int alloc_rsp_fqs(struct device *qidev)
703 const cpumask_t *cpus = qman_affine_cpus();
705 /*Now create response FQs*/
706 for_each_cpu(i, cpus) {
707 ret = alloc_rsp_fq_cpu(qidev, i);
709 dev_err(qidev, "CAAM rsp FQ alloc failed, cpu: %u", i);
717 static void free_rsp_fqs(void)
720 const cpumask_t *cpus = qman_affine_cpus();
722 for_each_cpu(i, cpus)
723 kfree(per_cpu(pcpu_qipriv.rsp_fq, i));
726 int caam_qi_init(struct platform_device *caam_pdev)
729 struct device *ctrldev = &caam_pdev->dev, *qidev;
730 struct caam_drv_private *ctrlpriv;
731 const cpumask_t *cpus = qman_affine_cpus();
733 ctrlpriv = dev_get_drvdata(ctrldev);
736 /* Initialize the congestion detection */
737 err = init_cgr(qidev);
739 dev_err(qidev, "CGR initialization failed: %d\n", err);
743 /* Initialise response FQs */
744 err = alloc_rsp_fqs(qidev);
746 dev_err(qidev, "Can't allocate CAAM response FQs: %d\n", err);
752 * Enable the NAPI contexts on each of the core which has an affine
755 for_each_cpu(i, cpus) {
756 struct caam_qi_pcpu_priv *priv = per_cpu_ptr(&pcpu_qipriv, i);
757 struct caam_napi *caam_napi = &priv->caam_napi;
758 struct napi_struct *irqtask = &caam_napi->irqtask;
759 struct net_device *net_dev = &priv->net_dev;
761 net_dev->dev = *qidev;
762 INIT_LIST_HEAD(&net_dev->napi_list);
764 netif_napi_add(net_dev, irqtask, caam_qi_poll,
767 napi_enable(irqtask);
770 qi_cache = kmem_cache_create("caamqicache", CAAM_QI_MEMCACHE_SIZE, 0,
771 SLAB_CACHE_DMA, NULL);
773 dev_err(qidev, "Can't allocate CAAM cache\n");
778 #ifdef CONFIG_DEBUG_FS
779 debugfs_create_file("qi_congested", 0444, ctrlpriv->ctl,
780 ×_congested, &caam_fops_u64_ro);
783 err = devm_add_action_or_reset(qidev, caam_qi_shutdown, ctrlpriv);
787 dev_info(qidev, "Linux CAAM Queue I/F driver initialised\n");