cachepc-linux

Fork of AMDESE/linux with modifications for CachePC side-channel attack
git clone https://git.sinitax.com/sinitax/cachepc-linux
Log | Files | Refs | README | LICENSE | sfeed.txt

cc_request_mgr.c (18675B)


      1// SPDX-License-Identifier: GPL-2.0
      2/* Copyright (C) 2012-2019 ARM Limited (or its affiliates). */
      3
      4#include <linux/kernel.h>
      5#include <linux/nospec.h>
      6#include "cc_driver.h"
      7#include "cc_buffer_mgr.h"
      8#include "cc_request_mgr.h"
      9#include "cc_pm.h"
     10
     11#define CC_MAX_POLL_ITER	10
     12/* The highest descriptor count in used */
     13#define CC_MAX_DESC_SEQ_LEN	23
     14
     15struct 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];
     21	u32 req_queue_head;
     22	u32 req_queue_tail;
     23	u32 axi_completed;
     24	u32 q_free_slots;
     25	/* This lock protects access to HW register
     26	 * that must be single request at a time
     27	 */
     28	spinlock_t hw_lock;
     29	struct cc_hw_desc compl_desc;
     30	u8 *dummy_comp_buff;
     31	dma_addr_t dummy_comp_buff_dma;
     32
     33	/* backlog queue */
     34	struct list_head backlog;
     35	unsigned int bl_len;
     36	spinlock_t bl_lock; /* protect backlog queue */
     37
     38#ifdef COMP_IN_WQ
     39	struct workqueue_struct *workq;
     40	struct delayed_work compwork;
     41#else
     42	struct tasklet_struct comptask;
     43#endif
     44};
     45
     46struct cc_bl_item {
     47	struct cc_crypto_req creq;
     48	struct cc_hw_desc desc[CC_MAX_DESC_SEQ_LEN];
     49	unsigned int len;
     50	struct list_head list;
     51	bool notif;
     52};
     53
     54static 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) }
     71};
     72
     73static void comp_handler(unsigned long devarg);
     74#ifdef COMP_IN_WQ
     75static void comp_work_handler(struct work_struct *work);
     76#endif
     77
     78static inline u32 cc_cpp_int_mask(enum cc_cpp_alg alg, int slot)
     79{
     80	alg = array_index_nospec(alg, CC_CPP_NUM_ALGS);
     81	slot = array_index_nospec(slot, CC_CPP_NUM_SLOTS);
     82
     83	return cc_cpp_int_masks[alg][slot];
     84}
     85
     86void cc_req_mgr_fini(struct cc_drvdata *drvdata)
     87{
     88	struct cc_req_mgr_handle *req_mgr_h = drvdata->request_mgr_handle;
     89	struct device *dev = drvdata_to_dev(drvdata);
     90
     91	if (!req_mgr_h)
     92		return; /* Not allocated */
     93
     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);
     97	}
     98
     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);
    102
    103#ifdef COMP_IN_WQ
    104	destroy_workqueue(req_mgr_h->workq);
    105#else
    106	/* Kill tasklet */
    107	tasklet_kill(&req_mgr_h->comptask);
    108#endif
    109	kfree_sensitive(req_mgr_h);
    110	drvdata->request_mgr_handle = NULL;
    111}
    112
    113int cc_req_mgr_init(struct cc_drvdata *drvdata)
    114{
    115	struct cc_req_mgr_handle *req_mgr_h;
    116	struct device *dev = drvdata_to_dev(drvdata);
    117	int rc = 0;
    118
    119	req_mgr_h = kzalloc(sizeof(*req_mgr_h), GFP_KERNEL);
    120	if (!req_mgr_h) {
    121		rc = -ENOMEM;
    122		goto req_mgr_init_err;
    123	}
    124
    125	drvdata->request_mgr_handle = req_mgr_h;
    126
    127	spin_lock_init(&req_mgr_h->hw_lock);
    128	spin_lock_init(&req_mgr_h->bl_lock);
    129	INIT_LIST_HEAD(&req_mgr_h->backlog);
    130
    131#ifdef COMP_IN_WQ
    132	dev_dbg(dev, "Initializing completion workqueue\n");
    133	req_mgr_h->workq = create_singlethread_workqueue("ccree");
    134	if (!req_mgr_h->workq) {
    135		dev_err(dev, "Failed creating work queue\n");
    136		rc = -ENOMEM;
    137		goto req_mgr_init_err;
    138	}
    139	INIT_DELAYED_WORK(&req_mgr_h->compwork, comp_work_handler);
    140#else
    141	dev_dbg(dev, "Initializing completion tasklet\n");
    142	tasklet_init(&req_mgr_h->comptask, comp_handler,
    143		     (unsigned long)drvdata);
    144#endif
    145	req_mgr_h->hw_queue_size = cc_ioread(drvdata,
    146					     CC_REG(DSCRPTR_QUEUE_SRAM_SIZE));
    147	dev_dbg(dev, "hw_queue_size=0x%08X\n", req_mgr_h->hw_queue_size);
    148	if (req_mgr_h->hw_queue_size < MIN_HW_QUEUE_SIZE) {
    149		dev_err(dev, "Invalid HW queue size = %u (Min. required is %u)\n",
    150			req_mgr_h->hw_queue_size, MIN_HW_QUEUE_SIZE);
    151		rc = -ENOMEM;
    152		goto req_mgr_init_err;
    153	}
    154	req_mgr_h->min_free_hw_slots = req_mgr_h->hw_queue_size;
    155	req_mgr_h->max_used_sw_slots = 0;
    156
    157	/* Allocate DMA word for "dummy" completion descriptor use */
    158	req_mgr_h->dummy_comp_buff =
    159		dma_alloc_coherent(dev, sizeof(u32),
    160				   &req_mgr_h->dummy_comp_buff_dma,
    161				   GFP_KERNEL);
    162	if (!req_mgr_h->dummy_comp_buff) {
    163		dev_err(dev, "Not enough memory to allocate DMA (%zu) dropped buffer\n",
    164			sizeof(u32));
    165		rc = -ENOMEM;
    166		goto req_mgr_init_err;
    167	}
    168
    169	/* Init. "dummy" completion descriptor */
    170	hw_desc_init(&req_mgr_h->compl_desc);
    171	set_din_const(&req_mgr_h->compl_desc, 0, sizeof(u32));
    172	set_dout_dlli(&req_mgr_h->compl_desc, req_mgr_h->dummy_comp_buff_dma,
    173		      sizeof(u32), NS_BIT, 1);
    174	set_flow_mode(&req_mgr_h->compl_desc, BYPASS);
    175	set_queue_last_ind(drvdata, &req_mgr_h->compl_desc);
    176
    177	return 0;
    178
    179req_mgr_init_err:
    180	cc_req_mgr_fini(drvdata);
    181	return rc;
    182}
    183
    184static void enqueue_seq(struct cc_drvdata *drvdata, struct cc_hw_desc seq[],
    185			unsigned int seq_len)
    186{
    187	int i, w;
    188	void __iomem *reg = drvdata->cc_base + CC_REG(DSCRPTR_QUEUE_WORD0);
    189	struct device *dev = drvdata_to_dev(drvdata);
    190
    191	/*
    192	 * We do indeed write all 6 command words to the same
    193	 * register. The HW supports this.
    194	 */
    195
    196	for (i = 0; i < seq_len; i++) {
    197		for (w = 0; w <= 5; w++)
    198			writel_relaxed(seq[i].word[w], reg);
    199
    200		if (cc_dump_desc)
    201			dev_dbg(dev, "desc[%02d]: 0x%08X 0x%08X 0x%08X 0x%08X 0x%08X 0x%08X\n",
    202				i, seq[i].word[0], seq[i].word[1],
    203				seq[i].word[2], seq[i].word[3],
    204				seq[i].word[4], seq[i].word[5]);
    205	}
    206}
    207
    208/**
    209 * request_mgr_complete() - Completion will take place if and only if user
    210 * requested completion by cc_send_sync_request().
    211 *
    212 * @dev: Device pointer
    213 * @dx_compl_h: The completion event to signal
    214 * @dummy: unused error code
    215 */
    216static void request_mgr_complete(struct device *dev, void *dx_compl_h,
    217				 int dummy)
    218{
    219	struct completion *this_compl = dx_compl_h;
    220
    221	complete(this_compl);
    222}
    223
    224static int cc_queues_status(struct cc_drvdata *drvdata,
    225			    struct cc_req_mgr_handle *req_mgr_h,
    226			    unsigned int total_seq_len)
    227{
    228	unsigned long poll_queue;
    229	struct device *dev = drvdata_to_dev(drvdata);
    230
    231	/* SW queue is checked only once as it will not
    232	 * be changed during the poll because the spinlock_bh
    233	 * is held by the thread
    234	 */
    235	if (((req_mgr_h->req_queue_head + 1) & (MAX_REQUEST_QUEUE_SIZE - 1)) ==
    236	    req_mgr_h->req_queue_tail) {
    237		dev_err(dev, "SW FIFO is full. req_queue_head=%d sw_fifo_len=%d\n",
    238			req_mgr_h->req_queue_head, MAX_REQUEST_QUEUE_SIZE);
    239		return -ENOSPC;
    240	}
    241
    242	if (req_mgr_h->q_free_slots >= total_seq_len)
    243		return 0;
    244
    245	/* Wait for space in HW queue. Poll constant num of iterations. */
    246	for (poll_queue = 0; poll_queue < CC_MAX_POLL_ITER ; poll_queue++) {
    247		req_mgr_h->q_free_slots =
    248			cc_ioread(drvdata, CC_REG(DSCRPTR_QUEUE_CONTENT));
    249		if (req_mgr_h->q_free_slots < req_mgr_h->min_free_hw_slots)
    250			req_mgr_h->min_free_hw_slots = req_mgr_h->q_free_slots;
    251
    252		if (req_mgr_h->q_free_slots >= total_seq_len) {
    253			/* If there is enough place return */
    254			return 0;
    255		}
    256
    257		dev_dbg(dev, "HW FIFO is full. q_free_slots=%d total_seq_len=%d\n",
    258			req_mgr_h->q_free_slots, total_seq_len);
    259	}
    260	/* No room in the HW queue try again later */
    261	dev_dbg(dev, "HW FIFO full, timeout. req_queue_head=%d sw_fifo_len=%d q_free_slots=%d total_seq_len=%d\n",
    262		req_mgr_h->req_queue_head, MAX_REQUEST_QUEUE_SIZE,
    263		req_mgr_h->q_free_slots, total_seq_len);
    264	return -ENOSPC;
    265}
    266
    267/**
    268 * cc_do_send_request() - Enqueue caller request to crypto hardware.
    269 * Need to be called with HW lock held and PM running
    270 *
    271 * @drvdata: Associated device driver context
    272 * @cc_req: The request to enqueue
    273 * @desc: The crypto sequence
    274 * @len: The crypto sequence length
    275 * @add_comp: If "true": add an artificial dout DMA to mark completion
    276 *
    277 */
    278static void cc_do_send_request(struct cc_drvdata *drvdata,
    279			       struct cc_crypto_req *cc_req,
    280			       struct cc_hw_desc *desc, unsigned int len,
    281			       bool add_comp)
    282{
    283	struct cc_req_mgr_handle *req_mgr_h = drvdata->request_mgr_handle;
    284	unsigned int used_sw_slots;
    285	unsigned int total_seq_len = len; /*initial sequence length*/
    286	struct device *dev = drvdata_to_dev(drvdata);
    287
    288	used_sw_slots = ((req_mgr_h->req_queue_head -
    289			  req_mgr_h->req_queue_tail) &
    290			 (MAX_REQUEST_QUEUE_SIZE - 1));
    291	if (used_sw_slots > req_mgr_h->max_used_sw_slots)
    292		req_mgr_h->max_used_sw_slots = used_sw_slots;
    293
    294	/* Enqueue request - must be locked with HW lock*/
    295	req_mgr_h->req_queue[req_mgr_h->req_queue_head] = *cc_req;
    296	req_mgr_h->req_queue_head = (req_mgr_h->req_queue_head + 1) &
    297				    (MAX_REQUEST_QUEUE_SIZE - 1);
    298
    299	dev_dbg(dev, "Enqueue request head=%u\n", req_mgr_h->req_queue_head);
    300
    301	/*
    302	 * We are about to push command to the HW via the command registers
    303	 * that may reference host memory. We need to issue a memory barrier
    304	 * to make sure there are no outstanding memory writes
    305	 */
    306	wmb();
    307
    308	/* STAT_PHASE_4: Push sequence */
    309
    310	enqueue_seq(drvdata, desc, len);
    311
    312	if (add_comp) {
    313		enqueue_seq(drvdata, &req_mgr_h->compl_desc, 1);
    314		total_seq_len++;
    315	}
    316
    317	if (req_mgr_h->q_free_slots < total_seq_len) {
    318		/* This situation should never occur. Maybe indicating problem
    319		 * with resuming power. Set the free slot count to 0 and hope
    320		 * for the best.
    321		 */
    322		dev_err(dev, "HW free slot count mismatch.");
    323		req_mgr_h->q_free_slots = 0;
    324	} else {
    325		/* Update the free slots in HW queue */
    326		req_mgr_h->q_free_slots -= total_seq_len;
    327	}
    328}
    329
    330static void cc_enqueue_backlog(struct cc_drvdata *drvdata,
    331			       struct cc_bl_item *bli)
    332{
    333	struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
    334	struct device *dev = drvdata_to_dev(drvdata);
    335
    336	spin_lock_bh(&mgr->bl_lock);
    337	list_add_tail(&bli->list, &mgr->backlog);
    338	++mgr->bl_len;
    339	dev_dbg(dev, "+++bl len: %d\n", mgr->bl_len);
    340	spin_unlock_bh(&mgr->bl_lock);
    341	tasklet_schedule(&mgr->comptask);
    342}
    343
    344static void cc_proc_backlog(struct cc_drvdata *drvdata)
    345{
    346	struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
    347	struct cc_bl_item *bli;
    348	struct cc_crypto_req *creq;
    349	void *req;
    350	struct device *dev = drvdata_to_dev(drvdata);
    351	int rc;
    352
    353	spin_lock(&mgr->bl_lock);
    354
    355	while (mgr->bl_len) {
    356		bli = list_first_entry(&mgr->backlog, struct cc_bl_item, list);
    357		dev_dbg(dev, "---bl len: %d\n", mgr->bl_len);
    358
    359		spin_unlock(&mgr->bl_lock);
    360
    361
    362		creq = &bli->creq;
    363		req = creq->user_arg;
    364
    365		/*
    366		 * Notify the request we're moving out of the backlog
    367		 * but only if we haven't done so already.
    368		 */
    369		if (!bli->notif) {
    370			creq->user_cb(dev, req, -EINPROGRESS);
    371			bli->notif = true;
    372		}
    373
    374		spin_lock(&mgr->hw_lock);
    375
    376		rc = cc_queues_status(drvdata, mgr, bli->len);
    377		if (rc) {
    378			/*
    379			 * There is still no room in the FIFO for
    380			 * this request. Bail out. We'll return here
    381			 * on the next completion irq.
    382			 */
    383			spin_unlock(&mgr->hw_lock);
    384			return;
    385		}
    386
    387		cc_do_send_request(drvdata, &bli->creq, bli->desc, bli->len,
    388				   false);
    389		spin_unlock(&mgr->hw_lock);
    390
    391		/* Remove ourselves from the backlog list */
    392		spin_lock(&mgr->bl_lock);
    393		list_del(&bli->list);
    394		--mgr->bl_len;
    395		kfree(bli);
    396	}
    397
    398	spin_unlock(&mgr->bl_lock);
    399}
    400
    401int cc_send_request(struct cc_drvdata *drvdata, struct cc_crypto_req *cc_req,
    402		    struct cc_hw_desc *desc, unsigned int len,
    403		    struct crypto_async_request *req)
    404{
    405	int rc;
    406	struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
    407	struct device *dev = drvdata_to_dev(drvdata);
    408	bool backlog_ok = req->flags & CRYPTO_TFM_REQ_MAY_BACKLOG;
    409	gfp_t flags = cc_gfp_flags(req);
    410	struct cc_bl_item *bli;
    411
    412	rc = cc_pm_get(dev);
    413	if (rc) {
    414		dev_err(dev, "cc_pm_get returned %x\n", rc);
    415		return rc;
    416	}
    417
    418	spin_lock_bh(&mgr->hw_lock);
    419	rc = cc_queues_status(drvdata, mgr, len);
    420
    421#ifdef CC_DEBUG_FORCE_BACKLOG
    422	if (backlog_ok)
    423		rc = -ENOSPC;
    424#endif /* CC_DEBUG_FORCE_BACKLOG */
    425
    426	if (rc == -ENOSPC && backlog_ok) {
    427		spin_unlock_bh(&mgr->hw_lock);
    428
    429		bli = kmalloc(sizeof(*bli), flags);
    430		if (!bli) {
    431			cc_pm_put_suspend(dev);
    432			return -ENOMEM;
    433		}
    434
    435		memcpy(&bli->creq, cc_req, sizeof(*cc_req));
    436		memcpy(&bli->desc, desc, len * sizeof(*desc));
    437		bli->len = len;
    438		bli->notif = false;
    439		cc_enqueue_backlog(drvdata, bli);
    440		return -EBUSY;
    441	}
    442
    443	if (!rc) {
    444		cc_do_send_request(drvdata, cc_req, desc, len, false);
    445		rc = -EINPROGRESS;
    446	}
    447
    448	spin_unlock_bh(&mgr->hw_lock);
    449	return rc;
    450}
    451
    452int cc_send_sync_request(struct cc_drvdata *drvdata,
    453			 struct cc_crypto_req *cc_req, struct cc_hw_desc *desc,
    454			 unsigned int len)
    455{
    456	int rc;
    457	struct device *dev = drvdata_to_dev(drvdata);
    458	struct cc_req_mgr_handle *mgr = drvdata->request_mgr_handle;
    459
    460	init_completion(&cc_req->seq_compl);
    461	cc_req->user_cb = request_mgr_complete;
    462	cc_req->user_arg = &cc_req->seq_compl;
    463
    464	rc = cc_pm_get(dev);
    465	if (rc) {
    466		dev_err(dev, "cc_pm_get returned %x\n", rc);
    467		return rc;
    468	}
    469
    470	while (true) {
    471		spin_lock_bh(&mgr->hw_lock);
    472		rc = cc_queues_status(drvdata, mgr, len + 1);
    473
    474		if (!rc)
    475			break;
    476
    477		spin_unlock_bh(&mgr->hw_lock);
    478		wait_for_completion_interruptible(&drvdata->hw_queue_avail);
    479		reinit_completion(&drvdata->hw_queue_avail);
    480	}
    481
    482	cc_do_send_request(drvdata, cc_req, desc, len, true);
    483	spin_unlock_bh(&mgr->hw_lock);
    484	wait_for_completion(&cc_req->seq_compl);
    485	return 0;
    486}
    487
    488/**
    489 * send_request_init() - Enqueue caller request to crypto hardware during init
    490 * process.
    491 * Assume this function is not called in the middle of a flow,
    492 * since we set QUEUE_LAST_IND flag in the last descriptor.
    493 *
    494 * @drvdata: Associated device driver context
    495 * @desc: The crypto sequence
    496 * @len: The crypto sequence length
    497 *
    498 * Return:
    499 * Returns "0" upon success
    500 */
    501int send_request_init(struct cc_drvdata *drvdata, struct cc_hw_desc *desc,
    502		      unsigned int len)
    503{
    504	struct cc_req_mgr_handle *req_mgr_h = drvdata->request_mgr_handle;
    505	unsigned int total_seq_len = len; /*initial sequence length*/
    506	int rc = 0;
    507
    508	/* Wait for space in HW and SW FIFO. Poll for as much as FIFO_TIMEOUT.
    509	 */
    510	rc = cc_queues_status(drvdata, req_mgr_h, total_seq_len);
    511	if (rc)
    512		return rc;
    513
    514	set_queue_last_ind(drvdata, &desc[(len - 1)]);
    515
    516	/*
    517	 * We are about to push command to the HW via the command registers
    518	 * that may reference host memory. We need to issue a memory barrier
    519	 * to make sure there are no outstanding memory writes
    520	 */
    521	wmb();
    522	enqueue_seq(drvdata, desc, len);
    523
    524	/* Update the free slots in HW queue */
    525	req_mgr_h->q_free_slots =
    526		cc_ioread(drvdata, CC_REG(DSCRPTR_QUEUE_CONTENT));
    527
    528	return 0;
    529}
    530
    531void complete_request(struct cc_drvdata *drvdata)
    532{
    533	struct cc_req_mgr_handle *request_mgr_handle =
    534						drvdata->request_mgr_handle;
    535
    536	complete(&drvdata->hw_queue_avail);
    537#ifdef COMP_IN_WQ
    538	queue_delayed_work(request_mgr_handle->workq,
    539			   &request_mgr_handle->compwork, 0);
    540#else
    541	tasklet_schedule(&request_mgr_handle->comptask);
    542#endif
    543}
    544
    545#ifdef COMP_IN_WQ
    546static void comp_work_handler(struct work_struct *work)
    547{
    548	struct cc_drvdata *drvdata =
    549		container_of(work, struct cc_drvdata, compwork.work);
    550
    551	comp_handler((unsigned long)drvdata);
    552}
    553#endif
    554
    555static void proc_completions(struct cc_drvdata *drvdata)
    556{
    557	struct cc_crypto_req *cc_req;
    558	struct device *dev = drvdata_to_dev(drvdata);
    559	struct cc_req_mgr_handle *request_mgr_handle =
    560						drvdata->request_mgr_handle;
    561	unsigned int *tail = &request_mgr_handle->req_queue_tail;
    562	unsigned int *head = &request_mgr_handle->req_queue_head;
    563	int rc;
    564	u32 mask;
    565
    566	while (request_mgr_handle->axi_completed) {
    567		request_mgr_handle->axi_completed--;
    568
    569		/* Dequeue request */
    570		if (*head == *tail) {
    571			/* We are supposed to handle a completion but our
    572			 * queue is empty. This is not normal. Return and
    573			 * hope for the best.
    574			 */
    575			dev_err(dev, "Request queue is empty head == tail %u\n",
    576				*head);
    577			break;
    578		}
    579
    580		cc_req = &request_mgr_handle->req_queue[*tail];
    581
    582		if (cc_req->cpp.is_cpp) {
    583
    584			dev_dbg(dev, "CPP request completion slot: %d alg:%d\n",
    585				cc_req->cpp.slot, cc_req->cpp.alg);
    586			mask = cc_cpp_int_mask(cc_req->cpp.alg,
    587					       cc_req->cpp.slot);
    588			rc = (drvdata->irq & mask ? -EPERM : 0);
    589			dev_dbg(dev, "Got mask: %x irq: %x rc: %d\n", mask,
    590				drvdata->irq, rc);
    591		} else {
    592			dev_dbg(dev, "None CPP request completion\n");
    593			rc = 0;
    594		}
    595
    596		if (cc_req->user_cb)
    597			cc_req->user_cb(dev, cc_req->user_arg, rc);
    598		*tail = (*tail + 1) & (MAX_REQUEST_QUEUE_SIZE - 1);
    599		dev_dbg(dev, "Dequeue request tail=%u\n", *tail);
    600		dev_dbg(dev, "Request completed. axi_completed=%d\n",
    601			request_mgr_handle->axi_completed);
    602		cc_pm_put_suspend(dev);
    603	}
    604}
    605
    606static inline u32 cc_axi_comp_count(struct cc_drvdata *drvdata)
    607{
    608	return FIELD_GET(AXIM_MON_COMP_VALUE,
    609			 cc_ioread(drvdata, drvdata->axim_mon_offset));
    610}
    611
    612/* Deferred service handler, run as interrupt-fired tasklet */
    613static void comp_handler(unsigned long devarg)
    614{
    615	struct cc_drvdata *drvdata = (struct cc_drvdata *)devarg;
    616	struct cc_req_mgr_handle *request_mgr_handle =
    617						drvdata->request_mgr_handle;
    618	struct device *dev = drvdata_to_dev(drvdata);
    619	u32 irq;
    620
    621	dev_dbg(dev, "Completion handler called!\n");
    622	irq = (drvdata->irq & drvdata->comp_mask);
    623
    624	/* To avoid the interrupt from firing as we unmask it,
    625	 * we clear it now
    626	 */
    627	cc_iowrite(drvdata, CC_REG(HOST_ICR), irq);
    628
    629	/* Avoid race with above clear: Test completion counter once more */
    630
    631	request_mgr_handle->axi_completed += cc_axi_comp_count(drvdata);
    632
    633	dev_dbg(dev, "AXI completion after updated: %d\n",
    634		request_mgr_handle->axi_completed);
    635
    636	while (request_mgr_handle->axi_completed) {
    637		do {
    638			drvdata->irq |= cc_ioread(drvdata, CC_REG(HOST_IRR));
    639			irq = (drvdata->irq & drvdata->comp_mask);
    640			proc_completions(drvdata);
    641
    642			/* At this point (after proc_completions()),
    643			 * request_mgr_handle->axi_completed is 0.
    644			 */
    645			request_mgr_handle->axi_completed +=
    646						cc_axi_comp_count(drvdata);
    647		} while (request_mgr_handle->axi_completed > 0);
    648
    649		cc_iowrite(drvdata, CC_REG(HOST_ICR), irq);
    650
    651		request_mgr_handle->axi_completed += cc_axi_comp_count(drvdata);
    652	}
    653
    654	/* after verifying that there is nothing to do,
    655	 * unmask AXI completion interrupt
    656	 */
    657	cc_iowrite(drvdata, CC_REG(HOST_IMR),
    658		   cc_ioread(drvdata, CC_REG(HOST_IMR)) & ~drvdata->comp_mask);
    659
    660	cc_proc_backlog(drvdata);
    661	dev_dbg(dev, "Comp. handler done.\n");
    662}