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
;
47 struct cc_crypto_req creq
;
48 struct cc_hw_desc desc
[CC_MAX_DESC_SEQ_LEN
];
50 struct list_head list
;
54 static const u32 cc_cpp_int_masks
[CC_CPP_NUM_ALGS
][CC_CPP_NUM_SLOTS
] = {
55 { BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_0_INT_BIT_SHIFT
),
56 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_1_INT_BIT_SHIFT
),
57 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_2_INT_BIT_SHIFT
),
58 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_3_INT_BIT_SHIFT
),
59 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_4_INT_BIT_SHIFT
),
60 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_5_INT_BIT_SHIFT
),
61 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_6_INT_BIT_SHIFT
),
62 BIT(CC_HOST_IRR_REE_OP_ABORTED_AES_7_INT_BIT_SHIFT
) },
63 { BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_0_INT_BIT_SHIFT
),
64 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_1_INT_BIT_SHIFT
),
65 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_2_INT_BIT_SHIFT
),
66 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_3_INT_BIT_SHIFT
),
67 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_4_INT_BIT_SHIFT
),
68 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_5_INT_BIT_SHIFT
),
69 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_6_INT_BIT_SHIFT
),
70 BIT(CC_HOST_IRR_REE_OP_ABORTED_SM_7_INT_BIT_SHIFT
) }
73 static void comp_handler(unsigned long devarg
);
75 static void comp_work_handler(struct work_struct
*work
);
78 static inline u32
cc_cpp_int_mask(enum cc_cpp_alg alg
, int slot
)
80 alg
= array_index_nospec(alg
, CC_CPP_NUM_ALGS
);
81 slot
= array_index_nospec(slot
, CC_CPP_NUM_SLOTS
);
83 return cc_cpp_int_masks
[alg
][slot
];
86 void cc_req_mgr_fini(struct cc_drvdata
*drvdata
)
88 struct cc_req_mgr_handle
*req_mgr_h
= drvdata
->request_mgr_handle
;
89 struct device
*dev
= drvdata_to_dev(drvdata
);
92 return; /* Not allocated */
94 if (req_mgr_h
->dummy_comp_buff_dma
) {
95 dma_free_coherent(dev
, sizeof(u32
), req_mgr_h
->dummy_comp_buff
,
96 req_mgr_h
->dummy_comp_buff_dma
);
99 dev_dbg(dev
, "max_used_hw_slots=%d\n", (req_mgr_h
->hw_queue_size
-
100 req_mgr_h
->min_free_hw_slots
));
101 dev_dbg(dev
, "max_used_sw_slots=%d\n", req_mgr_h
->max_used_sw_slots
);
104 flush_workqueue(req_mgr_h
->workq
);
105 destroy_workqueue(req_mgr_h
->workq
);
108 tasklet_kill(&req_mgr_h
->comptask
);
110 kfree_sensitive(req_mgr_h
);
111 drvdata
->request_mgr_handle
= NULL
;
114 int cc_req_mgr_init(struct cc_drvdata
*drvdata
)
116 struct cc_req_mgr_handle
*req_mgr_h
;
117 struct device
*dev
= drvdata_to_dev(drvdata
);
120 req_mgr_h
= kzalloc(sizeof(*req_mgr_h
), GFP_KERNEL
);
123 goto req_mgr_init_err
;
126 drvdata
->request_mgr_handle
= req_mgr_h
;
128 spin_lock_init(&req_mgr_h
->hw_lock
);
129 spin_lock_init(&req_mgr_h
->bl_lock
);
130 INIT_LIST_HEAD(&req_mgr_h
->backlog
);
133 dev_dbg(dev
, "Initializing completion workqueue\n");
134 req_mgr_h
->workq
= create_singlethread_workqueue("ccree");
135 if (!req_mgr_h
->workq
) {
136 dev_err(dev
, "Failed creating work queue\n");
138 goto req_mgr_init_err
;
140 INIT_DELAYED_WORK(&req_mgr_h
->compwork
, comp_work_handler
);
142 dev_dbg(dev
, "Initializing completion tasklet\n");
143 tasklet_init(&req_mgr_h
->comptask
, comp_handler
,
144 (unsigned long)drvdata
);
146 req_mgr_h
->hw_queue_size
= cc_ioread(drvdata
,
147 CC_REG(DSCRPTR_QUEUE_SRAM_SIZE
));
148 dev_dbg(dev
, "hw_queue_size=0x%08X\n", req_mgr_h
->hw_queue_size
);
149 if (req_mgr_h
->hw_queue_size
< MIN_HW_QUEUE_SIZE
) {
150 dev_err(dev
, "Invalid HW queue size = %u (Min. required is %u)\n",
151 req_mgr_h
->hw_queue_size
, MIN_HW_QUEUE_SIZE
);
153 goto req_mgr_init_err
;
155 req_mgr_h
->min_free_hw_slots
= req_mgr_h
->hw_queue_size
;
156 req_mgr_h
->max_used_sw_slots
= 0;
158 /* Allocate DMA word for "dummy" completion descriptor use */
159 req_mgr_h
->dummy_comp_buff
=
160 dma_alloc_coherent(dev
, sizeof(u32
),
161 &req_mgr_h
->dummy_comp_buff_dma
,
163 if (!req_mgr_h
->dummy_comp_buff
) {
164 dev_err(dev
, "Not enough memory to allocate DMA (%zu) dropped buffer\n",
167 goto req_mgr_init_err
;
170 /* Init. "dummy" completion descriptor */
171 hw_desc_init(&req_mgr_h
->compl_desc
);
172 set_din_const(&req_mgr_h
->compl_desc
, 0, sizeof(u32
));
173 set_dout_dlli(&req_mgr_h
->compl_desc
, req_mgr_h
->dummy_comp_buff_dma
,
174 sizeof(u32
), NS_BIT
, 1);
175 set_flow_mode(&req_mgr_h
->compl_desc
, BYPASS
);
176 set_queue_last_ind(drvdata
, &req_mgr_h
->compl_desc
);
181 cc_req_mgr_fini(drvdata
);
185 static void enqueue_seq(struct cc_drvdata
*drvdata
, struct cc_hw_desc seq
[],
186 unsigned int seq_len
)
189 void __iomem
*reg
= drvdata
->cc_base
+ CC_REG(DSCRPTR_QUEUE_WORD0
);
190 struct device
*dev
= drvdata_to_dev(drvdata
);
193 * We do indeed write all 6 command words to the same
194 * register. The HW supports this.
197 for (i
= 0; i
< seq_len
; i
++) {
198 for (w
= 0; w
<= 5; w
++)
199 writel_relaxed(seq
[i
].word
[w
], reg
);
202 dev_dbg(dev
, "desc[%02d]: 0x%08X 0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n",
203 i
, seq
[i
].word
[0], seq
[i
].word
[1],
204 seq
[i
].word
[2], seq
[i
].word
[3],
205 seq
[i
].word
[4], seq
[i
].word
[5]);
210 * request_mgr_complete() - Completion will take place if and only if user
211 * requested completion by cc_send_sync_request().
213 * @dev: Device pointer
214 * @dx_compl_h: The completion event to signal
215 * @dummy: unused error code
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 changed 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 * cc_do_send_request() - Enqueue caller request to crypto hardware.
270 * Need to be called with HW lock held and PM running
272 * @drvdata: Associated device driver context
273 * @cc_req: The request to enqueue
274 * @desc: The crypto sequence
275 * @len: The crypto sequence length
276 * @add_comp: If "true": add an artificial dout DMA to mark completion
279 static void cc_do_send_request(struct cc_drvdata
*drvdata
,
280 struct cc_crypto_req
*cc_req
,
281 struct cc_hw_desc
*desc
, unsigned int len
,
284 struct cc_req_mgr_handle
*req_mgr_h
= drvdata
->request_mgr_handle
;
285 unsigned int used_sw_slots
;
286 unsigned int total_seq_len
= len
; /*initial sequence length*/
287 struct device
*dev
= drvdata_to_dev(drvdata
);
289 used_sw_slots
= ((req_mgr_h
->req_queue_head
-
290 req_mgr_h
->req_queue_tail
) &
291 (MAX_REQUEST_QUEUE_SIZE
- 1));
292 if (used_sw_slots
> req_mgr_h
->max_used_sw_slots
)
293 req_mgr_h
->max_used_sw_slots
= used_sw_slots
;
295 /* Enqueue request - must be locked with HW lock*/
296 req_mgr_h
->req_queue
[req_mgr_h
->req_queue_head
] = *cc_req
;
297 req_mgr_h
->req_queue_head
= (req_mgr_h
->req_queue_head
+ 1) &
298 (MAX_REQUEST_QUEUE_SIZE
- 1);
300 dev_dbg(dev
, "Enqueue request head=%u\n", req_mgr_h
->req_queue_head
);
303 * We are about to push command to the HW via the command registers
304 * that may reference host memory. We need to issue a memory barrier
305 * to make sure there are no outstanding memory writes
309 /* STAT_PHASE_4: Push sequence */
311 enqueue_seq(drvdata
, desc
, len
);
314 enqueue_seq(drvdata
, &req_mgr_h
->compl_desc
, 1);
318 if (req_mgr_h
->q_free_slots
< total_seq_len
) {
319 /* This situation should never occur. Maybe indicating problem
320 * with resuming power. Set the free slot count to 0 and hope
323 dev_err(dev
, "HW free slot count mismatch.");
324 req_mgr_h
->q_free_slots
= 0;
326 /* Update the free slots in HW queue */
327 req_mgr_h
->q_free_slots
-= total_seq_len
;
331 static void cc_enqueue_backlog(struct cc_drvdata
*drvdata
,
332 struct cc_bl_item
*bli
)
334 struct cc_req_mgr_handle
*mgr
= drvdata
->request_mgr_handle
;
335 struct device
*dev
= drvdata_to_dev(drvdata
);
337 spin_lock_bh(&mgr
->bl_lock
);
338 list_add_tail(&bli
->list
, &mgr
->backlog
);
340 dev_dbg(dev
, "+++bl len: %d\n", mgr
->bl_len
);
341 spin_unlock_bh(&mgr
->bl_lock
);
342 tasklet_schedule(&mgr
->comptask
);
345 static void cc_proc_backlog(struct cc_drvdata
*drvdata
)
347 struct cc_req_mgr_handle
*mgr
= drvdata
->request_mgr_handle
;
348 struct cc_bl_item
*bli
;
349 struct cc_crypto_req
*creq
;
351 struct device
*dev
= drvdata_to_dev(drvdata
);
354 spin_lock(&mgr
->bl_lock
);
356 while (mgr
->bl_len
) {
357 bli
= list_first_entry(&mgr
->backlog
, struct cc_bl_item
, list
);
358 dev_dbg(dev
, "---bl len: %d\n", mgr
->bl_len
);
360 spin_unlock(&mgr
->bl_lock
);
364 req
= creq
->user_arg
;
367 * Notify the request we're moving out of the backlog
368 * but only if we haven't done so already.
371 creq
->user_cb(dev
, req
, -EINPROGRESS
);
375 spin_lock(&mgr
->hw_lock
);
377 rc
= cc_queues_status(drvdata
, mgr
, bli
->len
);
380 * There is still no room in the FIFO for
381 * this request. Bail out. We'll return here
382 * on the next completion irq.
384 spin_unlock(&mgr
->hw_lock
);
388 cc_do_send_request(drvdata
, &bli
->creq
, bli
->desc
, bli
->len
,
390 spin_unlock(&mgr
->hw_lock
);
392 /* Remove ourselves from the backlog list */
393 spin_lock(&mgr
->bl_lock
);
394 list_del(&bli
->list
);
399 spin_unlock(&mgr
->bl_lock
);
402 int cc_send_request(struct cc_drvdata
*drvdata
, struct cc_crypto_req
*cc_req
,
403 struct cc_hw_desc
*desc
, unsigned int len
,
404 struct crypto_async_request
*req
)
407 struct cc_req_mgr_handle
*mgr
= drvdata
->request_mgr_handle
;
408 struct device
*dev
= drvdata_to_dev(drvdata
);
409 bool backlog_ok
= req
->flags
& CRYPTO_TFM_REQ_MAY_BACKLOG
;
410 gfp_t flags
= cc_gfp_flags(req
);
411 struct cc_bl_item
*bli
;
415 dev_err(dev
, "cc_pm_get returned %x\n", rc
);
419 spin_lock_bh(&mgr
->hw_lock
);
420 rc
= cc_queues_status(drvdata
, mgr
, len
);
422 #ifdef CC_DEBUG_FORCE_BACKLOG
425 #endif /* CC_DEBUG_FORCE_BACKLOG */
427 if (rc
== -ENOSPC
&& backlog_ok
) {
428 spin_unlock_bh(&mgr
->hw_lock
);
430 bli
= kmalloc(sizeof(*bli
), flags
);
432 cc_pm_put_suspend(dev
);
436 memcpy(&bli
->creq
, cc_req
, sizeof(*cc_req
));
437 memcpy(&bli
->desc
, desc
, len
* sizeof(*desc
));
440 cc_enqueue_backlog(drvdata
, bli
);
445 cc_do_send_request(drvdata
, cc_req
, desc
, len
, false);
449 spin_unlock_bh(&mgr
->hw_lock
);
453 int cc_send_sync_request(struct cc_drvdata
*drvdata
,
454 struct cc_crypto_req
*cc_req
, struct cc_hw_desc
*desc
,
458 struct device
*dev
= drvdata_to_dev(drvdata
);
459 struct cc_req_mgr_handle
*mgr
= drvdata
->request_mgr_handle
;
461 init_completion(&cc_req
->seq_compl
);
462 cc_req
->user_cb
= request_mgr_complete
;
463 cc_req
->user_arg
= &cc_req
->seq_compl
;
467 dev_err(dev
, "cc_pm_get returned %x\n", rc
);
472 spin_lock_bh(&mgr
->hw_lock
);
473 rc
= cc_queues_status(drvdata
, mgr
, len
+ 1);
478 spin_unlock_bh(&mgr
->hw_lock
);
479 wait_for_completion_interruptible(&drvdata
->hw_queue_avail
);
480 reinit_completion(&drvdata
->hw_queue_avail
);
483 cc_do_send_request(drvdata
, cc_req
, desc
, len
, true);
484 spin_unlock_bh(&mgr
->hw_lock
);
485 wait_for_completion(&cc_req
->seq_compl
);
490 * send_request_init() - Enqueue caller request to crypto hardware during init
492 * Assume this function is not called in the middle of a flow,
493 * since we set QUEUE_LAST_IND flag in the last descriptor.
495 * @drvdata: Associated device driver context
496 * @desc: The crypto sequence
497 * @len: The crypto sequence length
500 * Returns "0" upon success
502 int send_request_init(struct cc_drvdata
*drvdata
, struct cc_hw_desc
*desc
,
505 struct cc_req_mgr_handle
*req_mgr_h
= drvdata
->request_mgr_handle
;
506 unsigned int total_seq_len
= len
; /*initial sequence length*/
509 /* Wait for space in HW and SW FIFO. Poll for as much as FIFO_TIMEOUT.
511 rc
= cc_queues_status(drvdata
, req_mgr_h
, total_seq_len
);
515 set_queue_last_ind(drvdata
, &desc
[(len
- 1)]);
518 * We are about to push command to the HW via the command registers
519 * that may reference host memory. We need to issue a memory barrier
520 * to make sure there are no outstanding memory writes
523 enqueue_seq(drvdata
, desc
, len
);
525 /* Update the free slots in HW queue */
526 req_mgr_h
->q_free_slots
=
527 cc_ioread(drvdata
, CC_REG(DSCRPTR_QUEUE_CONTENT
));
532 void complete_request(struct cc_drvdata
*drvdata
)
534 struct cc_req_mgr_handle
*request_mgr_handle
=
535 drvdata
->request_mgr_handle
;
537 complete(&drvdata
->hw_queue_avail
);
539 queue_delayed_work(request_mgr_handle
->workq
,
540 &request_mgr_handle
->compwork
, 0);
542 tasklet_schedule(&request_mgr_handle
->comptask
);
547 static void comp_work_handler(struct work_struct
*work
)
549 struct cc_drvdata
*drvdata
=
550 container_of(work
, struct cc_drvdata
, compwork
.work
);
552 comp_handler((unsigned long)drvdata
);
556 static void proc_completions(struct cc_drvdata
*drvdata
)
558 struct cc_crypto_req
*cc_req
;
559 struct device
*dev
= drvdata_to_dev(drvdata
);
560 struct cc_req_mgr_handle
*request_mgr_handle
=
561 drvdata
->request_mgr_handle
;
562 unsigned int *tail
= &request_mgr_handle
->req_queue_tail
;
563 unsigned int *head
= &request_mgr_handle
->req_queue_head
;
567 while (request_mgr_handle
->axi_completed
) {
568 request_mgr_handle
->axi_completed
--;
570 /* Dequeue request */
571 if (*head
== *tail
) {
572 /* We are supposed to handle a completion but our
573 * queue is empty. This is not normal. Return and
576 dev_err(dev
, "Request queue is empty head == tail %u\n",
581 cc_req
= &request_mgr_handle
->req_queue
[*tail
];
583 if (cc_req
->cpp
.is_cpp
) {
585 dev_dbg(dev
, "CPP request completion slot: %d alg:%d\n",
586 cc_req
->cpp
.slot
, cc_req
->cpp
.alg
);
587 mask
= cc_cpp_int_mask(cc_req
->cpp
.alg
,
589 rc
= (drvdata
->irq
& mask
? -EPERM
: 0);
590 dev_dbg(dev
, "Got mask: %x irq: %x rc: %d\n", mask
,
593 dev_dbg(dev
, "None CPP request completion\n");
598 cc_req
->user_cb(dev
, cc_req
->user_arg
, rc
);
599 *tail
= (*tail
+ 1) & (MAX_REQUEST_QUEUE_SIZE
- 1);
600 dev_dbg(dev
, "Dequeue request tail=%u\n", *tail
);
601 dev_dbg(dev
, "Request completed. axi_completed=%d\n",
602 request_mgr_handle
->axi_completed
);
603 cc_pm_put_suspend(dev
);
607 static inline u32
cc_axi_comp_count(struct cc_drvdata
*drvdata
)
609 return FIELD_GET(AXIM_MON_COMP_VALUE
,
610 cc_ioread(drvdata
, drvdata
->axim_mon_offset
));
613 /* Deferred service handler, run as interrupt-fired tasklet */
614 static void comp_handler(unsigned long devarg
)
616 struct cc_drvdata
*drvdata
= (struct cc_drvdata
*)devarg
;
617 struct cc_req_mgr_handle
*request_mgr_handle
=
618 drvdata
->request_mgr_handle
;
619 struct device
*dev
= drvdata_to_dev(drvdata
);
622 dev_dbg(dev
, "Completion handler called!\n");
623 irq
= (drvdata
->irq
& drvdata
->comp_mask
);
625 /* To avoid the interrupt from firing as we unmask it,
628 cc_iowrite(drvdata
, CC_REG(HOST_ICR
), irq
);
630 /* Avoid race with above clear: Test completion counter once more */
632 request_mgr_handle
->axi_completed
+= cc_axi_comp_count(drvdata
);
634 dev_dbg(dev
, "AXI completion after updated: %d\n",
635 request_mgr_handle
->axi_completed
);
637 while (request_mgr_handle
->axi_completed
) {
639 drvdata
->irq
|= cc_ioread(drvdata
, CC_REG(HOST_IRR
));
640 irq
= (drvdata
->irq
& drvdata
->comp_mask
);
641 proc_completions(drvdata
);
643 /* At this point (after proc_completions()),
644 * request_mgr_handle->axi_completed is 0.
646 request_mgr_handle
->axi_completed
+=
647 cc_axi_comp_count(drvdata
);
648 } while (request_mgr_handle
->axi_completed
> 0);
650 cc_iowrite(drvdata
, CC_REG(HOST_ICR
), irq
);
652 request_mgr_handle
->axi_completed
+= cc_axi_comp_count(drvdata
);
655 /* after verifying that there is nothing to do,
656 * unmask AXI completion interrupt
658 cc_iowrite(drvdata
, CC_REG(HOST_IMR
),
659 cc_ioread(drvdata
, CC_REG(HOST_IMR
)) & ~drvdata
->comp_mask
);
661 cc_proc_backlog(drvdata
);
662 dev_dbg(dev
, "Comp. handler done.\n");