1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (C) 2012-2019 ARM Limited (or its affiliates). */
4 #include <linux/kernel.h>
5 #include <linux/nospec.h>
7 #include "cc_buffer_mgr.h"
8 #include "cc_request_mgr.h"
11 #define CC_MAX_POLL_ITER 10
12 /* The highest descriptor count in used */
13 #define CC_MAX_DESC_SEQ_LEN 23
15 struct cc_req_mgr_handle {
16 /* Request manager resources */
17 unsigned int hw_queue_size; /* HW capability */
18 unsigned int min_free_hw_slots;
19 unsigned int max_used_sw_slots;
20 struct cc_crypto_req req_queue[MAX_REQUEST_QUEUE_SIZE];
25 /* This lock protects access to HW register
26 * that must be single request at a time
29 struct cc_hw_desc compl_desc;
31 dma_addr_t dummy_comp_buff_dma;
34 struct list_head backlog;
36 spinlock_t bl_lock; /* protect backlog queue */
39 struct workqueue_struct *workq;
40 struct delayed_work compwork;
42 struct tasklet_struct comptask;
44 bool is_runtime_suspended;
48 struct cc_crypto_req creq;
49 struct cc_hw_desc desc[CC_MAX_DESC_SEQ_LEN];
51 struct list_head list;
55 static const u32 cc_cpp_int_masks[CC_CPP_NUM_ALGS][CC_CPP_NUM_SLOTS] = {
56 { BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_0_INT_BIT_SHIFT),
57 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_1_INT_BIT_SHIFT),
58 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_2_INT_BIT_SHIFT),
59 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_3_INT_BIT_SHIFT),
60 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_4_INT_BIT_SHIFT),
61 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_5_INT_BIT_SHIFT),
62 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_6_INT_BIT_SHIFT),
63 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_7_INT_BIT_SHIFT) },
64 { BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_0_INT_BIT_SHIFT),
65 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_1_INT_BIT_SHIFT),
66 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_2_INT_BIT_SHIFT),
67 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_3_INT_BIT_SHIFT),
68 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_4_INT_BIT_SHIFT),
69 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_5_INT_BIT_SHIFT),
70 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_6_INT_BIT_SHIFT),
71 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_7_INT_BIT_SHIFT) }
74 static void comp_handler(unsigned long devarg);
76 static void comp_work_handler(struct work_struct *work);
79 static inline u32 cc_cpp_int_mask(enum cc_cpp_alg alg, int slot)
81 alg = array_index_nospec(alg, CC_CPP_NUM_ALGS);
82 slot = array_index_nospec(slot, CC_CPP_NUM_SLOTS);
84 return cc_cpp_int_masks[alg][slot];
87 void cc_req_mgr_fini(struct cc_drvdata *drvdata)
89 struct cc_req_mgr_handle *req_mgr_h = drvdata->request_mgr_handle;
90 struct device *dev = drvdata_to_dev(drvdata);
93 return; /* Not allocated */
95 if (req_mgr_h->dummy_comp_buff_dma) {
96 dma_free_coherent(dev, sizeof(u32), req_mgr_h->dummy_comp_buff,
97 req_mgr_h->dummy_comp_buff_dma);
100 dev_dbg(dev, "max_used_hw_slots=%d\n", (req_mgr_h->hw_queue_size -
101 req_mgr_h->min_free_hw_slots));
102 dev_dbg(dev, "max_used_sw_slots=%d\n", req_mgr_h->max_used_sw_slots);
105 flush_workqueue(req_mgr_h->workq);
106 destroy_workqueue(req_mgr_h->workq);
109 tasklet_kill(&req_mgr_h->comptask);
112 drvdata->request_mgr_handle = NULL;
115 int cc_req_mgr_init(struct cc_drvdata *drvdata)
117 struct cc_req_mgr_handle *req_mgr_h;
118 struct device *dev = drvdata_to_dev(drvdata);
121 req_mgr_h = kzalloc(sizeof(*req_mgr_h), GFP_KERNEL);
124 goto req_mgr_init_err;
127 drvdata->request_mgr_handle = req_mgr_h;
129 spin_lock_init(&req_mgr_h->hw_lock);
130 spin_lock_init(&req_mgr_h->bl_lock);
131 INIT_LIST_HEAD(&req_mgr_h->backlog);
134 dev_dbg(dev, "Initializing completion workqueue\n");
135 req_mgr_h->workq = create_singlethread_workqueue("ccree");
136 if (!req_mgr_h->workq) {
137 dev_err(dev, "Failed creating work queue\n");
139 goto req_mgr_init_err;
141 INIT_DELAYED_WORK(&req_mgr_h->compwork, comp_work_handler);
143 dev_dbg(dev, "Initializing completion tasklet\n");
144 tasklet_init(&req_mgr_h->comptask, comp_handler,
145 (unsigned long)drvdata);
147 req_mgr_h->hw_queue_size = cc_ioread(drvdata,
148 CC_REG(DSCRPTR_QUEUE_SRAM_SIZE));
149 dev_dbg(dev, "hw_queue_size=0x%08X\n", req_mgr_h->hw_queue_size);
150 if (req_mgr_h->hw_queue_size < MIN_HW_QUEUE_SIZE) {
151 dev_err(dev, "Invalid HW queue size = %u (Min. required is %u)\n",
152 req_mgr_h->hw_queue_size, MIN_HW_QUEUE_SIZE);
154 goto req_mgr_init_err;
156 req_mgr_h->min_free_hw_slots = req_mgr_h->hw_queue_size;
157 req_mgr_h->max_used_sw_slots = 0;
159 /* Allocate DMA word for "dummy" completion descriptor use */
160 req_mgr_h->dummy_comp_buff =
161 dma_alloc_coherent(dev, sizeof(u32),
162 &req_mgr_h->dummy_comp_buff_dma,
164 if (!req_mgr_h->dummy_comp_buff) {
165 dev_err(dev, "Not enough memory to allocate DMA (%zu) dropped buffer\n",
168 goto req_mgr_init_err;
171 /* Init. "dummy" completion descriptor */
172 hw_desc_init(&req_mgr_h->compl_desc);
173 set_din_const(&req_mgr_h->compl_desc, 0, sizeof(u32));
174 set_dout_dlli(&req_mgr_h->compl_desc, req_mgr_h->dummy_comp_buff_dma,
175 sizeof(u32), NS_BIT, 1);
176 set_flow_mode(&req_mgr_h->compl_desc, BYPASS);
177 set_queue_last_ind(drvdata, &req_mgr_h->compl_desc);
182 cc_req_mgr_fini(drvdata);
186 static void enqueue_seq(struct cc_drvdata *drvdata, struct cc_hw_desc seq[],
187 unsigned int seq_len)
190 void __iomem *reg = drvdata->cc_base + CC_REG(DSCRPTR_QUEUE_WORD0);
191 struct device *dev = drvdata_to_dev(drvdata);
194 * We do indeed write all 6 command words to the same
195 * register. The HW supports this.
198 for (i = 0; i < seq_len; i++) {
199 for (w = 0; w <= 5; w++)
200 writel_relaxed(seq[i].word[w], reg);
203 dev_dbg(dev, "desc[%02d]: 0x%08X 0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n",
204 i, seq[i].word[0], seq[i].word[1],
205 seq[i].word[2], seq[i].word[3],
206 seq[i].word[4], seq[i].word[5]);
211 * Completion will take place if and only if user requested completion
212 * by cc_send_sync_request().
215 * \param dx_compl_h The completion event to signal
217 static void request_mgr_complete(struct device *dev, void *dx_compl_h,
220 struct completion *this_compl = dx_compl_h;
222 complete(this_compl);
225 static int cc_queues_status(struct cc_drvdata *drvdata,
226 struct cc_req_mgr_handle *req_mgr_h,
227 unsigned int total_seq_len)
229 unsigned long poll_queue;
230 struct device *dev = drvdata_to_dev(drvdata);
232 /* SW queue is checked only once as it will not
233 * be chaned during the poll because the spinlock_bh
234 * is held by the thread
236 if (((req_mgr_h->req_queue_head + 1) & (MAX_REQUEST_QUEUE_SIZE - 1)) ==
237 req_mgr_h->req_queue_tail) {
238 dev_err(dev, "SW FIFO is full. req_queue_head=%d sw_fifo_len=%d\n",
239 req_mgr_h->req_queue_head, MAX_REQUEST_QUEUE_SIZE);
243 if (req_mgr_h->q_free_slots >= total_seq_len)
246 /* Wait for space in HW queue. Poll constant num of iterations. */
247 for (poll_queue = 0; poll_queue < CC_MAX_POLL_ITER ; poll_queue++) {
248 req_mgr_h->q_free_slots =
249 cc_ioread(drvdata, CC_REG(DSCRPTR_QUEUE_CONTENT));
250 if (req_mgr_h->q_free_slots < req_mgr_h->min_free_hw_slots)
251 req_mgr_h->min_free_hw_slots = req_mgr_h->q_free_slots;
253 if (req_mgr_h->q_free_slots >= total_seq_len) {
254 /* If there is enough place return */
258 dev_dbg(dev, "HW FIFO is full. q_free_slots=%d total_seq_len=%d\n",
259 req_mgr_h->q_free_slots, total_seq_len);
261 /* No room in the HW queue try again later */
262 dev_dbg(dev, "HW FIFO full, timeout. req_queue_head=%d sw_fifo_len=%d q_free_slots=%d total_seq_len=%d\n",
263 req_mgr_h->req_queue_head, MAX_REQUEST_QUEUE_SIZE,
264 req_mgr_h->q_free_slots, total_seq_len);
269 * Enqueue caller request to crypto hardware.
270 * Need to be called with HW lock held and PM running
273 * \param cc_req The request to enqueue
274 * \param desc The crypto sequence
275 * \param len The crypto sequence length
276 * \param add_comp If "true": add an artificial dout DMA to mark completion
278 * \return int Returns -EINPROGRESS or error code
280 static int cc_do_send_request(struct cc_drvdata *drvdata,
281 struct cc_crypto_req *cc_req,
282 struct cc_hw_desc *desc, unsigned int len,
285 struct cc_req_mgr_handle *req_mgr_h = drvdata->request_mgr_handle;
286 unsigned int used_sw_slots;
287 unsigned int total_seq_len = len; /*initial sequence length*/
288 struct device *dev = drvdata_to_dev(drvdata);
290 used_sw_slots = ((req_mgr_h->req_queue_head -
291 req_mgr_h->req_queue_tail) &
292 (MAX_REQUEST_QUEUE_SIZE - 1));
293 if (used_sw_slots > req_mgr_h->max_used_sw_slots)
294 req_mgr_h->max_used_sw_slots = used_sw_slots;
296 /* Enqueue request - must be locked with HW lock*/
297 req_mgr_h->req_queue[req_mgr_h->req_queue_head] = *cc_req;
298 req_mgr_h->req_queue_head = (req_mgr_h->req_queue_head + 1) &
299 (MAX_REQUEST_QUEUE_SIZE - 1);
300 /* TODO: Use circ_buf.h ? */
302 dev_dbg(dev, "Enqueue request head=%u\n", req_mgr_h->req_queue_head);
305 * We are about to push command to the HW via the command registers
306 * that may refernece hsot memory. We need to issue a memory barrier
307 * to make sure there are no outstnading memory writes
311 /* STAT_PHASE_4: Push sequence */
313 enqueue_seq(drvdata, desc, len);
316 enqueue_seq(drvdata, &req_mgr_h->compl_desc, 1);
320 if (req_mgr_h->q_free_slots < total_seq_len) {
321 /* This situation should never occur. Maybe indicating problem
322 * with resuming power. Set the free slot count to 0 and hope
325 dev_err(dev, "HW free slot count mismatch.");
326 req_mgr_h->q_free_slots = 0;
328 /* Update the free slots in HW queue */
329 req_mgr_h->q_free_slots -= total_seq_len;
332 /* Operation still in process */
336 static void cc_enqueue_backlog(struct cc_drvdata *drvdata,
337 struct cc_bl_item *bli)
339 struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
340 struct device *dev = drvdata_to_dev(drvdata);
342 spin_lock_bh(&mgr->bl_lock);
343 list_add_tail(&bli->list, &mgr->backlog);
345 dev_dbg(dev, "+++bl len: %d\n", mgr->bl_len);
346 spin_unlock_bh(&mgr->bl_lock);
347 tasklet_schedule(&mgr->comptask);
350 static void cc_proc_backlog(struct cc_drvdata *drvdata)
352 struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
353 struct cc_bl_item *bli;
354 struct cc_crypto_req *creq;
356 struct device *dev = drvdata_to_dev(drvdata);
359 spin_lock(&mgr->bl_lock);
361 while (mgr->bl_len) {
362 bli = list_first_entry(&mgr->backlog, struct cc_bl_item, list);
363 dev_dbg(dev, "---bl len: %d\n", mgr->bl_len);
365 spin_unlock(&mgr->bl_lock);
369 req = creq->user_arg;
372 * Notify the request we're moving out of the backlog
373 * but only if we haven't done so already.
376 creq->user_cb(dev, req, -EINPROGRESS);
380 spin_lock(&mgr->hw_lock);
382 rc = cc_queues_status(drvdata, mgr, bli->len);
385 * There is still not room in the FIFO for
386 * this request. Bail out. We'll return here
387 * on the next completion irq.
389 spin_unlock(&mgr->hw_lock);
393 rc = cc_do_send_request(drvdata, &bli->creq, bli->desc,
396 spin_unlock(&mgr->hw_lock);
398 if (rc != -EINPROGRESS) {
399 cc_pm_put_suspend(dev);
400 creq->user_cb(dev, req, rc);
403 /* Remove ourselves from the backlog list */
404 spin_lock(&mgr->bl_lock);
405 list_del(&bli->list);
409 spin_unlock(&mgr->bl_lock);
412 int cc_send_request(struct cc_drvdata *drvdata, struct cc_crypto_req *cc_req,
413 struct cc_hw_desc *desc, unsigned int len,
414 struct crypto_async_request *req)
417 struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
418 struct device *dev = drvdata_to_dev(drvdata);
419 bool backlog_ok = req->flags & CRYPTO_TFM_REQ_MAY_BACKLOG;
420 gfp_t flags = cc_gfp_flags(req);
421 struct cc_bl_item *bli;
425 dev_err(dev, "ssi_power_mgr_runtime_get returned %x\n", rc);
429 spin_lock_bh(&mgr->hw_lock);
430 rc = cc_queues_status(drvdata, mgr, len);
432 #ifdef CC_DEBUG_FORCE_BACKLOG
435 #endif /* CC_DEBUG_FORCE_BACKLOG */
437 if (rc == -ENOSPC && backlog_ok) {
438 spin_unlock_bh(&mgr->hw_lock);
440 bli = kmalloc(sizeof(*bli), flags);
442 cc_pm_put_suspend(dev);
446 memcpy(&bli->creq, cc_req, sizeof(*cc_req));
447 memcpy(&bli->desc, desc, len * sizeof(*desc));
450 cc_enqueue_backlog(drvdata, bli);
455 rc = cc_do_send_request(drvdata, cc_req, desc, len, false);
457 spin_unlock_bh(&mgr->hw_lock);
461 int cc_send_sync_request(struct cc_drvdata *drvdata,
462 struct cc_crypto_req *cc_req, struct cc_hw_desc *desc,
466 struct device *dev = drvdata_to_dev(drvdata);
467 struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
469 init_completion(&cc_req->seq_compl);
470 cc_req->user_cb = request_mgr_complete;
471 cc_req->user_arg = &cc_req->seq_compl;
475 dev_err(dev, "ssi_power_mgr_runtime_get returned %x\n", rc);
480 spin_lock_bh(&mgr->hw_lock);
481 rc = cc_queues_status(drvdata, mgr, len + 1);
486 spin_unlock_bh(&mgr->hw_lock);
488 cc_pm_put_suspend(dev);
491 wait_for_completion_interruptible(&drvdata->hw_queue_avail);
492 reinit_completion(&drvdata->hw_queue_avail);
495 rc = cc_do_send_request(drvdata, cc_req, desc, len, true);
496 spin_unlock_bh(&mgr->hw_lock);
498 if (rc != -EINPROGRESS) {
499 cc_pm_put_suspend(dev);
503 wait_for_completion(&cc_req->seq_compl);
508 * Enqueue caller request to crypto hardware during init process.
509 * assume this function is not called in middle of a flow,
510 * since we set QUEUE_LAST_IND flag in the last descriptor.
513 * \param desc The crypto sequence
514 * \param len The crypto sequence length
516 * \return int Returns "0" upon success
518 int send_request_init(struct cc_drvdata *drvdata, struct cc_hw_desc *desc,
521 struct cc_req_mgr_handle *req_mgr_h = drvdata->request_mgr_handle;
522 unsigned int total_seq_len = len; /*initial sequence length*/
525 /* Wait for space in HW and SW FIFO. Poll for as much as FIFO_TIMEOUT.
527 rc = cc_queues_status(drvdata, req_mgr_h, total_seq_len);
531 set_queue_last_ind(drvdata, &desc[(len - 1)]);
534 * We are about to push command to the HW via the command registers
535 * that may refernece hsot memory. We need to issue a memory barrier
536 * to make sure there are no outstnading memory writes
539 enqueue_seq(drvdata, desc, len);
541 /* Update the free slots in HW queue */
542 req_mgr_h->q_free_slots =
543 cc_ioread(drvdata, CC_REG(DSCRPTR_QUEUE_CONTENT));
548 void complete_request(struct cc_drvdata *drvdata)
550 struct cc_req_mgr_handle *request_mgr_handle =
551 drvdata->request_mgr_handle;
553 complete(&drvdata->hw_queue_avail);
555 queue_delayed_work(request_mgr_handle->workq,
556 &request_mgr_handle->compwork, 0);
558 tasklet_schedule(&request_mgr_handle->comptask);
563 static void comp_work_handler(struct work_struct *work)
565 struct cc_drvdata *drvdata =
566 container_of(work, struct cc_drvdata, compwork.work);
568 comp_handler((unsigned long)drvdata);
572 static void proc_completions(struct cc_drvdata *drvdata)
574 struct cc_crypto_req *cc_req;
575 struct device *dev = drvdata_to_dev(drvdata);
576 struct cc_req_mgr_handle *request_mgr_handle =
577 drvdata->request_mgr_handle;
578 unsigned int *tail = &request_mgr_handle->req_queue_tail;
579 unsigned int *head = &request_mgr_handle->req_queue_head;
583 while (request_mgr_handle->axi_completed) {
584 request_mgr_handle->axi_completed--;
586 /* Dequeue request */
587 if (*head == *tail) {
588 /* We are supposed to handle a completion but our
589 * queue is empty. This is not normal. Return and
592 dev_err(dev, "Request queue is empty head == tail %u\n",
597 cc_req = &request_mgr_handle->req_queue[*tail];
599 if (cc_req->cpp.is_cpp) {
601 dev_dbg(dev, "CPP request completion slot: %d alg:%d\n",
602 cc_req->cpp.slot, cc_req->cpp.alg);
603 mask = cc_cpp_int_mask(cc_req->cpp.alg,
605 rc = (drvdata->irq & mask ? -EPERM : 0);
606 dev_dbg(dev, "Got mask: %x irq: %x rc: %d\n", mask,
609 dev_dbg(dev, "None CPP request completion\n");
614 cc_req->user_cb(dev, cc_req->user_arg, rc);
615 *tail = (*tail + 1) & (MAX_REQUEST_QUEUE_SIZE - 1);
616 dev_dbg(dev, "Dequeue request tail=%u\n", *tail);
617 dev_dbg(dev, "Request completed. axi_completed=%d\n",
618 request_mgr_handle->axi_completed);
619 cc_pm_put_suspend(dev);
623 static inline u32 cc_axi_comp_count(struct cc_drvdata *drvdata)
625 return FIELD_GET(AXIM_MON_COMP_VALUE,
626 cc_ioread(drvdata, drvdata->axim_mon_offset));
629 /* Deferred service handler, run as interrupt-fired tasklet */
630 static void comp_handler(unsigned long devarg)
632 struct cc_drvdata *drvdata = (struct cc_drvdata *)devarg;
633 struct cc_req_mgr_handle *request_mgr_handle =
634 drvdata->request_mgr_handle;
635 struct device *dev = drvdata_to_dev(drvdata);
638 dev_dbg(dev, "Completion handler called!\n");
639 irq = (drvdata->irq & drvdata->comp_mask);
641 /* To avoid the interrupt from firing as we unmask it,
644 cc_iowrite(drvdata, CC_REG(HOST_ICR), irq);
646 /* Avoid race with above clear: Test completion counter once more */
648 request_mgr_handle->axi_completed += cc_axi_comp_count(drvdata);
650 dev_dbg(dev, "AXI completion after updated: %d\n",
651 request_mgr_handle->axi_completed);
653 while (request_mgr_handle->axi_completed) {
655 drvdata->irq |= cc_ioread(drvdata, CC_REG(HOST_IRR));
656 irq = (drvdata->irq & drvdata->comp_mask);
657 proc_completions(drvdata);
659 /* At this point (after proc_completions()),
660 * request_mgr_handle->axi_completed is 0.
662 request_mgr_handle->axi_completed +=
663 cc_axi_comp_count(drvdata);
664 } while (request_mgr_handle->axi_completed > 0);
666 cc_iowrite(drvdata, CC_REG(HOST_ICR), irq);
668 request_mgr_handle->axi_completed += cc_axi_comp_count(drvdata);
671 /* after verifing that there is nothing to do,
672 * unmask AXI completion interrupt
674 cc_iowrite(drvdata, CC_REG(HOST_IMR),
675 cc_ioread(drvdata, CC_REG(HOST_IMR)) & ~drvdata->comp_mask);
677 cc_proc_backlog(drvdata);
678 dev_dbg(dev, "Comp. handler done.\n");
682 * resume the queue configuration - no need to take the lock as this happens
683 * inside the spin lock protection
685 #if defined(CONFIG_PM)
686 int cc_resume_req_queue(struct cc_drvdata *drvdata)
688 struct cc_req_mgr_handle *request_mgr_handle =
689 drvdata->request_mgr_handle;
691 spin_lock_bh(&request_mgr_handle->hw_lock);
692 request_mgr_handle->is_runtime_suspended = false;
693 spin_unlock_bh(&request_mgr_handle->hw_lock);
699 * suspend the queue configuration. Since it is used for the runtime suspend
700 * only verify that the queue can be suspended.
702 int cc_suspend_req_queue(struct cc_drvdata *drvdata)
704 struct cc_req_mgr_handle *request_mgr_handle =
705 drvdata->request_mgr_handle;
707 /* lock the send_request */
708 spin_lock_bh(&request_mgr_handle->hw_lock);
709 if (request_mgr_handle->req_queue_head !=
710 request_mgr_handle->req_queue_tail) {
711 spin_unlock_bh(&request_mgr_handle->hw_lock);
714 request_mgr_handle->is_runtime_suspended = true;
715 spin_unlock_bh(&request_mgr_handle->hw_lock);
720 bool cc_req_queue_suspended(struct cc_drvdata *drvdata)
722 struct cc_req_mgr_handle *request_mgr_handle =
723 drvdata->request_mgr_handle;
725 return request_mgr_handle->is_runtime_suspended;