xref: /linux/drivers/dma/mediatek/mtk-hsdma.c (revision 320fefa9e2edc67011e235ea1d50f0d00ddfe004)
1 // SPDX-License-Identifier: GPL-2.0
2 // Copyright (c) 2017-2018 MediaTek Inc.
3 
4 /*
5  * Driver for MediaTek High-Speed DMA Controller
6  *
7  * Author: Sean Wang <sean.wang@mediatek.com>
8  *
9  */
10 
11 #include <linux/bitops.h>
12 #include <linux/clk.h>
13 #include <linux/dmaengine.h>
14 #include <linux/dma-mapping.h>
15 #include <linux/err.h>
16 #include <linux/iopoll.h>
17 #include <linux/list.h>
18 #include <linux/module.h>
19 #include <linux/of.h>
20 #include <linux/of_device.h>
21 #include <linux/of_dma.h>
22 #include <linux/platform_device.h>
23 #include <linux/pm_runtime.h>
24 #include <linux/refcount.h>
25 #include <linux/slab.h>
26 
27 #include "../virt-dma.h"
28 
29 #define MTK_HSDMA_USEC_POLL		20
30 #define MTK_HSDMA_TIMEOUT_POLL		200000
31 #define MTK_HSDMA_DMA_BUSWIDTHS		BIT(DMA_SLAVE_BUSWIDTH_4_BYTES)
32 
33 /* The default number of virtual channel */
34 #define MTK_HSDMA_NR_VCHANS		3
35 
36 /* Only one physical channel supported */
37 #define MTK_HSDMA_NR_MAX_PCHANS		1
38 
39 /* Macro for physical descriptor (PD) manipulation */
40 /* The number of PD which must be 2 of power */
41 #define MTK_DMA_SIZE			64
42 #define MTK_HSDMA_NEXT_DESP_IDX(x, y)	(((x) + 1) & ((y) - 1))
43 #define MTK_HSDMA_LAST_DESP_IDX(x, y)	(((x) - 1) & ((y) - 1))
44 #define MTK_HSDMA_MAX_LEN		0x3f80
45 #define MTK_HSDMA_ALIGN_SIZE		4
46 #define MTK_HSDMA_PLEN_MASK		0x3fff
47 #define MTK_HSDMA_DESC_PLEN(x)		(((x) & MTK_HSDMA_PLEN_MASK) << 16)
48 #define MTK_HSDMA_DESC_PLEN_GET(x)	(((x) >> 16) & MTK_HSDMA_PLEN_MASK)
49 
50 /* Registers for underlying ring manipulation */
51 #define MTK_HSDMA_TX_BASE		0x0
52 #define MTK_HSDMA_TX_CNT		0x4
53 #define MTK_HSDMA_TX_CPU		0x8
54 #define MTK_HSDMA_TX_DMA		0xc
55 #define MTK_HSDMA_RX_BASE		0x100
56 #define MTK_HSDMA_RX_CNT		0x104
57 #define MTK_HSDMA_RX_CPU		0x108
58 #define MTK_HSDMA_RX_DMA		0x10c
59 
60 /* Registers for global setup */
61 #define MTK_HSDMA_GLO			0x204
62 #define MTK_HSDMA_GLO_MULTI_DMA		BIT(10)
63 #define MTK_HSDMA_TX_WB_DDONE		BIT(6)
64 #define MTK_HSDMA_BURST_64BYTES		(0x2 << 4)
65 #define MTK_HSDMA_GLO_RX_BUSY		BIT(3)
66 #define MTK_HSDMA_GLO_RX_DMA		BIT(2)
67 #define MTK_HSDMA_GLO_TX_BUSY		BIT(1)
68 #define MTK_HSDMA_GLO_TX_DMA		BIT(0)
69 #define MTK_HSDMA_GLO_DMA		(MTK_HSDMA_GLO_TX_DMA |	\
70 					 MTK_HSDMA_GLO_RX_DMA)
71 #define MTK_HSDMA_GLO_BUSY		(MTK_HSDMA_GLO_RX_BUSY | \
72 					 MTK_HSDMA_GLO_TX_BUSY)
73 #define MTK_HSDMA_GLO_DEFAULT		(MTK_HSDMA_GLO_TX_DMA | \
74 					 MTK_HSDMA_GLO_RX_DMA | \
75 					 MTK_HSDMA_TX_WB_DDONE | \
76 					 MTK_HSDMA_BURST_64BYTES | \
77 					 MTK_HSDMA_GLO_MULTI_DMA)
78 
79 /* Registers for reset */
80 #define MTK_HSDMA_RESET			0x208
81 #define MTK_HSDMA_RST_TX		BIT(0)
82 #define MTK_HSDMA_RST_RX		BIT(16)
83 
84 /* Registers for interrupt control */
85 #define MTK_HSDMA_DLYINT		0x20c
86 #define MTK_HSDMA_RXDLY_INT_EN		BIT(15)
87 
88 /* Interrupt fires when the pending number's more than the specified */
89 #define MTK_HSDMA_RXMAX_PINT(x)		(((x) & 0x7f) << 8)
90 
91 /* Interrupt fires when the pending time's more than the specified in 20 us */
92 #define MTK_HSDMA_RXMAX_PTIME(x)	((x) & 0x7f)
93 #define MTK_HSDMA_DLYINT_DEFAULT	(MTK_HSDMA_RXDLY_INT_EN | \
94 					 MTK_HSDMA_RXMAX_PINT(20) | \
95 					 MTK_HSDMA_RXMAX_PTIME(20))
96 #define MTK_HSDMA_INT_STATUS		0x220
97 #define MTK_HSDMA_INT_ENABLE		0x228
98 #define MTK_HSDMA_INT_RXDONE		BIT(16)
99 
100 enum mtk_hsdma_vdesc_flag {
101 	MTK_HSDMA_VDESC_FINISHED	= 0x01,
102 };
103 
104 #define IS_MTK_HSDMA_VDESC_FINISHED(x) ((x) == MTK_HSDMA_VDESC_FINISHED)
105 
106 /**
107  * struct mtk_hsdma_pdesc - This is the struct holding info describing physical
108  *			    descriptor (PD) and its placement must be kept at
109  *			    4-bytes alignment in little endian order.
110  * @desc1:		    | The control pad used to indicate hardware how to
111  * @desc2:		    | deal with the descriptor such as source and
112  * @desc3:		    | destination address and data length. The maximum
113  * @desc4:		    | data length each pdesc can handle is 0x3f80 bytes
114  */
115 struct mtk_hsdma_pdesc {
116 	__le32 desc1;
117 	__le32 desc2;
118 	__le32 desc3;
119 	__le32 desc4;
120 } __packed __aligned(4);
121 
122 /**
123  * struct mtk_hsdma_vdesc - This is the struct holding info describing virtual
124  *			    descriptor (VD)
125  * @vd:			    An instance for struct virt_dma_desc
126  * @len:		    The total data size device wants to move
127  * @residue:		    The remaining data size device will move
128  * @dest:		    The destination address device wants to move to
129  * @src:		    The source address device wants to move from
130  */
131 struct mtk_hsdma_vdesc {
132 	struct virt_dma_desc vd;
133 	size_t len;
134 	size_t residue;
135 	dma_addr_t dest;
136 	dma_addr_t src;
137 };
138 
139 /**
140  * struct mtk_hsdma_cb - This is the struct holding extra info required for RX
141  *			 ring to know what relevant VD the PD is being
142  *			 mapped to.
143  * @vd:			 Pointer to the relevant VD.
144  * @flag:		 Flag indicating what action should be taken when VD
145  *			 is completed.
146  */
147 struct mtk_hsdma_cb {
148 	struct virt_dma_desc *vd;
149 	enum mtk_hsdma_vdesc_flag flag;
150 };
151 
152 /**
153  * struct mtk_hsdma_ring - This struct holds info describing underlying ring
154  *			   space
155  * @txd:		   The descriptor TX ring which describes DMA source
156  *			   information
157  * @rxd:		   The descriptor RX ring which describes DMA
158  *			   destination information
159  * @cb:			   The extra information pointed at by RX ring
160  * @tphys:		   The physical addr of TX ring
161  * @rphys:		   The physical addr of RX ring
162  * @cur_tptr:		   Pointer to the next free descriptor used by the host
163  * @cur_rptr:		   Pointer to the last done descriptor by the device
164  */
165 struct mtk_hsdma_ring {
166 	struct mtk_hsdma_pdesc *txd;
167 	struct mtk_hsdma_pdesc *rxd;
168 	struct mtk_hsdma_cb *cb;
169 	dma_addr_t tphys;
170 	dma_addr_t rphys;
171 	u16 cur_tptr;
172 	u16 cur_rptr;
173 };
174 
175 /**
176  * struct mtk_hsdma_pchan - This is the struct holding info describing physical
177  *			   channel (PC)
178  * @ring:		   An instance for the underlying ring
179  * @sz_ring:		   Total size allocated for the ring
180  * @nr_free:		   Total number of free rooms in the ring. It would
181  *			   be accessed and updated frequently between IRQ
182  *			   context and user context to reflect whether ring
183  *			   can accept requests from VD.
184  */
185 struct mtk_hsdma_pchan {
186 	struct mtk_hsdma_ring ring;
187 	size_t sz_ring;
188 	atomic_t nr_free;
189 };
190 
191 /**
192  * struct mtk_hsdma_vchan - This is the struct holding info describing virtual
193  *			   channel (VC)
194  * @vc:			   An instance for struct virt_dma_chan
195  * @issue_completion:	   The wait for all issued descriptors completited
196  * @issue_synchronize:	   Bool indicating channel synchronization starts
197  * @desc_hw_processing:	   List those descriptors the hardware is processing,
198  *			   which is protected by vc.lock
199  */
200 struct mtk_hsdma_vchan {
201 	struct virt_dma_chan vc;
202 	struct completion issue_completion;
203 	bool issue_synchronize;
204 	struct list_head desc_hw_processing;
205 };
206 
207 /**
208  * struct mtk_hsdma_soc - This is the struct holding differences among SoCs
209  * @ddone:		  Bit mask for DDONE
210  * @ls0:		  Bit mask for LS0
211  */
212 struct mtk_hsdma_soc {
213 	__le32 ddone;
214 	__le32 ls0;
215 };
216 
217 /**
218  * struct mtk_hsdma_device - This is the struct holding info describing HSDMA
219  *			     device
220  * @ddev:		     An instance for struct dma_device
221  * @base:		     The mapped register I/O base
222  * @clk:		     The clock that device internal is using
223  * @irq:		     The IRQ that device are using
224  * @dma_requests:	     The number of VCs the device supports to
225  * @vc:			     The pointer to all available VCs
226  * @pc:			     The pointer to the underlying PC
227  * @pc_refcnt:		     Track how many VCs are using the PC
228  * @lock:		     Lock protect agaisting multiple VCs access PC
229  * @soc:		     The pointer to area holding differences among
230  *			     vaious platform
231  */
232 struct mtk_hsdma_device {
233 	struct dma_device ddev;
234 	void __iomem *base;
235 	struct clk *clk;
236 	u32 irq;
237 
238 	u32 dma_requests;
239 	struct mtk_hsdma_vchan *vc;
240 	struct mtk_hsdma_pchan *pc;
241 	refcount_t pc_refcnt;
242 
243 	/* Lock used to protect against multiple VCs access PC */
244 	spinlock_t lock;
245 
246 	const struct mtk_hsdma_soc *soc;
247 };
248 
249 static struct mtk_hsdma_device *to_hsdma_dev(struct dma_chan *chan)
250 {
251 	return container_of(chan->device, struct mtk_hsdma_device, ddev);
252 }
253 
254 static inline struct mtk_hsdma_vchan *to_hsdma_vchan(struct dma_chan *chan)
255 {
256 	return container_of(chan, struct mtk_hsdma_vchan, vc.chan);
257 }
258 
259 static struct mtk_hsdma_vdesc *to_hsdma_vdesc(struct virt_dma_desc *vd)
260 {
261 	return container_of(vd, struct mtk_hsdma_vdesc, vd);
262 }
263 
264 static struct device *hsdma2dev(struct mtk_hsdma_device *hsdma)
265 {
266 	return hsdma->ddev.dev;
267 }
268 
269 static u32 mtk_dma_read(struct mtk_hsdma_device *hsdma, u32 reg)
270 {
271 	return readl(hsdma->base + reg);
272 }
273 
274 static void mtk_dma_write(struct mtk_hsdma_device *hsdma, u32 reg, u32 val)
275 {
276 	writel(val, hsdma->base + reg);
277 }
278 
279 static void mtk_dma_rmw(struct mtk_hsdma_device *hsdma, u32 reg,
280 			u32 mask, u32 set)
281 {
282 	u32 val;
283 
284 	val = mtk_dma_read(hsdma, reg);
285 	val &= ~mask;
286 	val |= set;
287 	mtk_dma_write(hsdma, reg, val);
288 }
289 
290 static void mtk_dma_set(struct mtk_hsdma_device *hsdma, u32 reg, u32 val)
291 {
292 	mtk_dma_rmw(hsdma, reg, 0, val);
293 }
294 
295 static void mtk_dma_clr(struct mtk_hsdma_device *hsdma, u32 reg, u32 val)
296 {
297 	mtk_dma_rmw(hsdma, reg, val, 0);
298 }
299 
300 static void mtk_hsdma_vdesc_free(struct virt_dma_desc *vd)
301 {
302 	kfree(container_of(vd, struct mtk_hsdma_vdesc, vd));
303 }
304 
305 static int mtk_hsdma_busy_wait(struct mtk_hsdma_device *hsdma)
306 {
307 	u32 status = 0;
308 
309 	return readl_poll_timeout(hsdma->base + MTK_HSDMA_GLO, status,
310 				  !(status & MTK_HSDMA_GLO_BUSY),
311 				  MTK_HSDMA_USEC_POLL,
312 				  MTK_HSDMA_TIMEOUT_POLL);
313 }
314 
315 static int mtk_hsdma_alloc_pchan(struct mtk_hsdma_device *hsdma,
316 				 struct mtk_hsdma_pchan *pc)
317 {
318 	struct mtk_hsdma_ring *ring = &pc->ring;
319 	int err;
320 
321 	memset(pc, 0, sizeof(*pc));
322 
323 	/*
324 	 * Allocate ring space where [0 ... MTK_DMA_SIZE - 1] is for TX ring
325 	 * and [MTK_DMA_SIZE ... 2 * MTK_DMA_SIZE - 1] is for RX ring.
326 	 */
327 	pc->sz_ring = 2 * MTK_DMA_SIZE * sizeof(*ring->txd);
328 	ring->txd = dma_alloc_coherent(hsdma2dev(hsdma), pc->sz_ring,
329 				       &ring->tphys, GFP_NOWAIT);
330 	if (!ring->txd)
331 		return -ENOMEM;
332 
333 	ring->rxd = &ring->txd[MTK_DMA_SIZE];
334 	ring->rphys = ring->tphys + MTK_DMA_SIZE * sizeof(*ring->txd);
335 	ring->cur_tptr = 0;
336 	ring->cur_rptr = MTK_DMA_SIZE - 1;
337 
338 	ring->cb = kcalloc(MTK_DMA_SIZE, sizeof(*ring->cb), GFP_NOWAIT);
339 	if (!ring->cb) {
340 		err = -ENOMEM;
341 		goto err_free_dma;
342 	}
343 
344 	atomic_set(&pc->nr_free, MTK_DMA_SIZE - 1);
345 
346 	/* Disable HSDMA and wait for the completion */
347 	mtk_dma_clr(hsdma, MTK_HSDMA_GLO, MTK_HSDMA_GLO_DMA);
348 	err = mtk_hsdma_busy_wait(hsdma);
349 	if (err)
350 		goto err_free_cb;
351 
352 	/* Reset */
353 	mtk_dma_set(hsdma, MTK_HSDMA_RESET,
354 		    MTK_HSDMA_RST_TX | MTK_HSDMA_RST_RX);
355 	mtk_dma_clr(hsdma, MTK_HSDMA_RESET,
356 		    MTK_HSDMA_RST_TX | MTK_HSDMA_RST_RX);
357 
358 	/* Setup HSDMA initial pointer in the ring */
359 	mtk_dma_write(hsdma, MTK_HSDMA_TX_BASE, ring->tphys);
360 	mtk_dma_write(hsdma, MTK_HSDMA_TX_CNT, MTK_DMA_SIZE);
361 	mtk_dma_write(hsdma, MTK_HSDMA_TX_CPU, ring->cur_tptr);
362 	mtk_dma_write(hsdma, MTK_HSDMA_TX_DMA, 0);
363 	mtk_dma_write(hsdma, MTK_HSDMA_RX_BASE, ring->rphys);
364 	mtk_dma_write(hsdma, MTK_HSDMA_RX_CNT, MTK_DMA_SIZE);
365 	mtk_dma_write(hsdma, MTK_HSDMA_RX_CPU, ring->cur_rptr);
366 	mtk_dma_write(hsdma, MTK_HSDMA_RX_DMA, 0);
367 
368 	/* Enable HSDMA */
369 	mtk_dma_set(hsdma, MTK_HSDMA_GLO, MTK_HSDMA_GLO_DMA);
370 
371 	/* Setup delayed interrupt */
372 	mtk_dma_write(hsdma, MTK_HSDMA_DLYINT, MTK_HSDMA_DLYINT_DEFAULT);
373 
374 	/* Enable interrupt */
375 	mtk_dma_set(hsdma, MTK_HSDMA_INT_ENABLE, MTK_HSDMA_INT_RXDONE);
376 
377 	return 0;
378 
379 err_free_cb:
380 	kfree(ring->cb);
381 
382 err_free_dma:
383 	dma_free_coherent(hsdma2dev(hsdma),
384 			  pc->sz_ring, ring->txd, ring->tphys);
385 	return err;
386 }
387 
388 static void mtk_hsdma_free_pchan(struct mtk_hsdma_device *hsdma,
389 				 struct mtk_hsdma_pchan *pc)
390 {
391 	struct mtk_hsdma_ring *ring = &pc->ring;
392 
393 	/* Disable HSDMA and then wait for the completion */
394 	mtk_dma_clr(hsdma, MTK_HSDMA_GLO, MTK_HSDMA_GLO_DMA);
395 	mtk_hsdma_busy_wait(hsdma);
396 
397 	/* Reset pointer in the ring */
398 	mtk_dma_clr(hsdma, MTK_HSDMA_INT_ENABLE, MTK_HSDMA_INT_RXDONE);
399 	mtk_dma_write(hsdma, MTK_HSDMA_TX_BASE, 0);
400 	mtk_dma_write(hsdma, MTK_HSDMA_TX_CNT, 0);
401 	mtk_dma_write(hsdma, MTK_HSDMA_TX_CPU, 0);
402 	mtk_dma_write(hsdma, MTK_HSDMA_RX_BASE, 0);
403 	mtk_dma_write(hsdma, MTK_HSDMA_RX_CNT, 0);
404 	mtk_dma_write(hsdma, MTK_HSDMA_RX_CPU, MTK_DMA_SIZE - 1);
405 
406 	kfree(ring->cb);
407 
408 	dma_free_coherent(hsdma2dev(hsdma),
409 			  pc->sz_ring, ring->txd, ring->tphys);
410 }
411 
412 static int mtk_hsdma_issue_pending_vdesc(struct mtk_hsdma_device *hsdma,
413 					 struct mtk_hsdma_pchan *pc,
414 					 struct mtk_hsdma_vdesc *hvd)
415 {
416 	struct mtk_hsdma_ring *ring = &pc->ring;
417 	struct mtk_hsdma_pdesc *txd, *rxd;
418 	u16 reserved, prev, tlen, num_sgs;
419 	unsigned long flags;
420 
421 	/* Protect against PC is accessed by multiple VCs simultaneously */
422 	spin_lock_irqsave(&hsdma->lock, flags);
423 
424 	/*
425 	 * Reserve rooms, where pc->nr_free is used to track how many free
426 	 * rooms in the ring being updated in user and IRQ context.
427 	 */
428 	num_sgs = DIV_ROUND_UP(hvd->len, MTK_HSDMA_MAX_LEN);
429 	reserved = min_t(u16, num_sgs, atomic_read(&pc->nr_free));
430 
431 	if (!reserved) {
432 		spin_unlock_irqrestore(&hsdma->lock, flags);
433 		return -ENOSPC;
434 	}
435 
436 	atomic_sub(reserved, &pc->nr_free);
437 
438 	while (reserved--) {
439 		/* Limit size by PD capability for valid data moving */
440 		tlen = (hvd->len > MTK_HSDMA_MAX_LEN) ?
441 		       MTK_HSDMA_MAX_LEN : hvd->len;
442 
443 		/*
444 		 * Setup PDs using the remaining VD info mapped on those
445 		 * reserved rooms. And since RXD is shared memory between the
446 		 * host and the device allocated by dma_alloc_coherent call,
447 		 * the helper macro WRITE_ONCE can ensure the data written to
448 		 * RAM would really happens.
449 		 */
450 		txd = &ring->txd[ring->cur_tptr];
451 		WRITE_ONCE(txd->desc1, hvd->src);
452 		WRITE_ONCE(txd->desc2,
453 			   hsdma->soc->ls0 | MTK_HSDMA_DESC_PLEN(tlen));
454 
455 		rxd = &ring->rxd[ring->cur_tptr];
456 		WRITE_ONCE(rxd->desc1, hvd->dest);
457 		WRITE_ONCE(rxd->desc2, MTK_HSDMA_DESC_PLEN(tlen));
458 
459 		/* Associate VD, the PD belonged to */
460 		ring->cb[ring->cur_tptr].vd = &hvd->vd;
461 
462 		/* Move forward the pointer of TX ring */
463 		ring->cur_tptr = MTK_HSDMA_NEXT_DESP_IDX(ring->cur_tptr,
464 							 MTK_DMA_SIZE);
465 
466 		/* Update VD with remaining data */
467 		hvd->src  += tlen;
468 		hvd->dest += tlen;
469 		hvd->len  -= tlen;
470 	}
471 
472 	/*
473 	 * Tagging flag for the last PD for VD will be responsible for
474 	 * completing VD.
475 	 */
476 	if (!hvd->len) {
477 		prev = MTK_HSDMA_LAST_DESP_IDX(ring->cur_tptr, MTK_DMA_SIZE);
478 		ring->cb[prev].flag = MTK_HSDMA_VDESC_FINISHED;
479 	}
480 
481 	/* Ensure all changes indeed done before we're going on */
482 	wmb();
483 
484 	/*
485 	 * Updating into hardware the pointer of TX ring lets HSDMA to take
486 	 * action for those pending PDs.
487 	 */
488 	mtk_dma_write(hsdma, MTK_HSDMA_TX_CPU, ring->cur_tptr);
489 
490 	spin_unlock_irqrestore(&hsdma->lock, flags);
491 
492 	return 0;
493 }
494 
495 static void mtk_hsdma_issue_vchan_pending(struct mtk_hsdma_device *hsdma,
496 					  struct mtk_hsdma_vchan *hvc)
497 {
498 	struct virt_dma_desc *vd, *vd2;
499 	int err;
500 
501 	lockdep_assert_held(&hvc->vc.lock);
502 
503 	list_for_each_entry_safe(vd, vd2, &hvc->vc.desc_issued, node) {
504 		struct mtk_hsdma_vdesc *hvd;
505 
506 		hvd = to_hsdma_vdesc(vd);
507 
508 		/* Map VD into PC and all VCs shares a single PC */
509 		err = mtk_hsdma_issue_pending_vdesc(hsdma, hsdma->pc, hvd);
510 
511 		/*
512 		 * Move VD from desc_issued to desc_hw_processing when entire
513 		 * VD is fit into available PDs. Otherwise, the uncompleted
514 		 * VDs would stay in list desc_issued and then restart the
515 		 * processing as soon as possible once underlying ring space
516 		 * got freed.
517 		 */
518 		if (err == -ENOSPC || hvd->len > 0)
519 			break;
520 
521 		/*
522 		 * The extra list desc_hw_processing is used because
523 		 * hardware can't provide sufficient information allowing us
524 		 * to know what VDs are still working on the underlying ring.
525 		 * Through the additional list, it can help us to implement
526 		 * terminate_all, residue calculation and such thing needed
527 		 * to know detail descriptor status on the hardware.
528 		 */
529 		list_move_tail(&vd->node, &hvc->desc_hw_processing);
530 	}
531 }
532 
533 static void mtk_hsdma_free_rooms_in_ring(struct mtk_hsdma_device *hsdma)
534 {
535 	struct mtk_hsdma_vchan *hvc;
536 	struct mtk_hsdma_pdesc *rxd;
537 	struct mtk_hsdma_vdesc *hvd;
538 	struct mtk_hsdma_pchan *pc;
539 	struct mtk_hsdma_cb *cb;
540 	int i = MTK_DMA_SIZE;
541 	__le32 desc2;
542 	u32 status;
543 	u16 next;
544 
545 	/* Read IRQ status */
546 	status = mtk_dma_read(hsdma, MTK_HSDMA_INT_STATUS);
547 	if (unlikely(!(status & MTK_HSDMA_INT_RXDONE)))
548 		goto rx_done;
549 
550 	pc = hsdma->pc;
551 
552 	/*
553 	 * Using a fail-safe loop with iterations of up to MTK_DMA_SIZE to
554 	 * reclaim these finished descriptors: The most number of PDs the ISR
555 	 * can handle at one time shouldn't be more than MTK_DMA_SIZE so we
556 	 * take it as limited count instead of just using a dangerous infinite
557 	 * poll.
558 	 */
559 	while (i--) {
560 		next = MTK_HSDMA_NEXT_DESP_IDX(pc->ring.cur_rptr,
561 					       MTK_DMA_SIZE);
562 		rxd = &pc->ring.rxd[next];
563 
564 		/*
565 		 * If MTK_HSDMA_DESC_DDONE is no specified, that means data
566 		 * moving for the PD is still under going.
567 		 */
568 		desc2 = READ_ONCE(rxd->desc2);
569 		if (!(desc2 & hsdma->soc->ddone))
570 			break;
571 
572 		cb = &pc->ring.cb[next];
573 		if (unlikely(!cb->vd)) {
574 			dev_err(hsdma2dev(hsdma), "cb->vd cannot be null\n");
575 			break;
576 		}
577 
578 		/* Update residue of VD the associated PD belonged to */
579 		hvd = to_hsdma_vdesc(cb->vd);
580 		hvd->residue -= MTK_HSDMA_DESC_PLEN_GET(rxd->desc2);
581 
582 		/* Complete VD until the relevant last PD is finished */
583 		if (IS_MTK_HSDMA_VDESC_FINISHED(cb->flag)) {
584 			hvc = to_hsdma_vchan(cb->vd->tx.chan);
585 
586 			spin_lock(&hvc->vc.lock);
587 
588 			/* Remove VD from list desc_hw_processing */
589 			list_del(&cb->vd->node);
590 
591 			/* Add VD into list desc_completed */
592 			vchan_cookie_complete(cb->vd);
593 
594 			if (hvc->issue_synchronize &&
595 			    list_empty(&hvc->desc_hw_processing)) {
596 				complete(&hvc->issue_completion);
597 				hvc->issue_synchronize = false;
598 			}
599 			spin_unlock(&hvc->vc.lock);
600 
601 			cb->flag = 0;
602 		}
603 
604 		cb->vd = NULL;
605 
606 		/*
607 		 * Recycle the RXD with the helper WRITE_ONCE that can ensure
608 		 * data written into RAM would really happens.
609 		 */
610 		WRITE_ONCE(rxd->desc1, 0);
611 		WRITE_ONCE(rxd->desc2, 0);
612 		pc->ring.cur_rptr = next;
613 
614 		/* Release rooms */
615 		atomic_inc(&pc->nr_free);
616 	}
617 
618 	/* Ensure all changes indeed done before we're going on */
619 	wmb();
620 
621 	/* Update CPU pointer for those completed PDs */
622 	mtk_dma_write(hsdma, MTK_HSDMA_RX_CPU, pc->ring.cur_rptr);
623 
624 	/*
625 	 * Acking the pending IRQ allows hardware no longer to keep the used
626 	 * IRQ line in certain trigger state when software has completed all
627 	 * the finished physical descriptors.
628 	 */
629 	if (atomic_read(&pc->nr_free) >= MTK_DMA_SIZE - 1)
630 		mtk_dma_write(hsdma, MTK_HSDMA_INT_STATUS, status);
631 
632 	/* ASAP handles pending VDs in all VCs after freeing some rooms */
633 	for (i = 0; i < hsdma->dma_requests; i++) {
634 		hvc = &hsdma->vc[i];
635 		spin_lock(&hvc->vc.lock);
636 		mtk_hsdma_issue_vchan_pending(hsdma, hvc);
637 		spin_unlock(&hvc->vc.lock);
638 	}
639 
640 rx_done:
641 	/* All completed PDs are cleaned up, so enable interrupt again */
642 	mtk_dma_set(hsdma, MTK_HSDMA_INT_ENABLE, MTK_HSDMA_INT_RXDONE);
643 }
644 
645 static irqreturn_t mtk_hsdma_irq(int irq, void *devid)
646 {
647 	struct mtk_hsdma_device *hsdma = devid;
648 
649 	/*
650 	 * Disable interrupt until all completed PDs are cleaned up in
651 	 * mtk_hsdma_free_rooms call.
652 	 */
653 	mtk_dma_clr(hsdma, MTK_HSDMA_INT_ENABLE, MTK_HSDMA_INT_RXDONE);
654 
655 	mtk_hsdma_free_rooms_in_ring(hsdma);
656 
657 	return IRQ_HANDLED;
658 }
659 
660 static struct virt_dma_desc *mtk_hsdma_find_active_desc(struct dma_chan *c,
661 							dma_cookie_t cookie)
662 {
663 	struct mtk_hsdma_vchan *hvc = to_hsdma_vchan(c);
664 	struct virt_dma_desc *vd;
665 
666 	list_for_each_entry(vd, &hvc->desc_hw_processing, node)
667 		if (vd->tx.cookie == cookie)
668 			return vd;
669 
670 	list_for_each_entry(vd, &hvc->vc.desc_issued, node)
671 		if (vd->tx.cookie == cookie)
672 			return vd;
673 
674 	return NULL;
675 }
676 
677 static enum dma_status mtk_hsdma_tx_status(struct dma_chan *c,
678 					   dma_cookie_t cookie,
679 					   struct dma_tx_state *txstate)
680 {
681 	struct mtk_hsdma_vchan *hvc = to_hsdma_vchan(c);
682 	struct mtk_hsdma_vdesc *hvd;
683 	struct virt_dma_desc *vd;
684 	enum dma_status ret;
685 	unsigned long flags;
686 	size_t bytes = 0;
687 
688 	ret = dma_cookie_status(c, cookie, txstate);
689 	if (ret == DMA_COMPLETE || !txstate)
690 		return ret;
691 
692 	spin_lock_irqsave(&hvc->vc.lock, flags);
693 	vd = mtk_hsdma_find_active_desc(c, cookie);
694 	spin_unlock_irqrestore(&hvc->vc.lock, flags);
695 
696 	if (vd) {
697 		hvd = to_hsdma_vdesc(vd);
698 		bytes = hvd->residue;
699 	}
700 
701 	dma_set_residue(txstate, bytes);
702 
703 	return ret;
704 }
705 
706 static void mtk_hsdma_issue_pending(struct dma_chan *c)
707 {
708 	struct mtk_hsdma_device *hsdma = to_hsdma_dev(c);
709 	struct mtk_hsdma_vchan *hvc = to_hsdma_vchan(c);
710 	unsigned long flags;
711 
712 	spin_lock_irqsave(&hvc->vc.lock, flags);
713 
714 	if (vchan_issue_pending(&hvc->vc))
715 		mtk_hsdma_issue_vchan_pending(hsdma, hvc);
716 
717 	spin_unlock_irqrestore(&hvc->vc.lock, flags);
718 }
719 
720 static struct dma_async_tx_descriptor *
721 mtk_hsdma_prep_dma_memcpy(struct dma_chan *c, dma_addr_t dest,
722 			  dma_addr_t src, size_t len, unsigned long flags)
723 {
724 	struct mtk_hsdma_vdesc *hvd;
725 
726 	hvd = kzalloc(sizeof(*hvd), GFP_NOWAIT);
727 	if (!hvd)
728 		return NULL;
729 
730 	hvd->len = len;
731 	hvd->residue = len;
732 	hvd->src = src;
733 	hvd->dest = dest;
734 
735 	return vchan_tx_prep(to_virt_chan(c), &hvd->vd, flags);
736 }
737 
738 static int mtk_hsdma_free_inactive_desc(struct dma_chan *c)
739 {
740 	struct virt_dma_chan *vc = to_virt_chan(c);
741 	unsigned long flags;
742 	LIST_HEAD(head);
743 
744 	spin_lock_irqsave(&vc->lock, flags);
745 	list_splice_tail_init(&vc->desc_allocated, &head);
746 	list_splice_tail_init(&vc->desc_submitted, &head);
747 	list_splice_tail_init(&vc->desc_issued, &head);
748 	spin_unlock_irqrestore(&vc->lock, flags);
749 
750 	/* At the point, we don't expect users put descriptor into VC again */
751 	vchan_dma_desc_free_list(vc, &head);
752 
753 	return 0;
754 }
755 
756 static void mtk_hsdma_free_active_desc(struct dma_chan *c)
757 {
758 	struct mtk_hsdma_vchan *hvc = to_hsdma_vchan(c);
759 	bool sync_needed = false;
760 
761 	/*
762 	 * Once issue_synchronize is being set, which means once the hardware
763 	 * consumes all descriptors for the channel in the ring, the
764 	 * synchronization must be notified immediately it is completed.
765 	 */
766 	spin_lock(&hvc->vc.lock);
767 	if (!list_empty(&hvc->desc_hw_processing)) {
768 		hvc->issue_synchronize = true;
769 		sync_needed = true;
770 	}
771 	spin_unlock(&hvc->vc.lock);
772 
773 	if (sync_needed)
774 		wait_for_completion(&hvc->issue_completion);
775 	/*
776 	 * At the point, we expect that all remaining descriptors in the ring
777 	 * for the channel should be all processing done.
778 	 */
779 	WARN_ONCE(!list_empty(&hvc->desc_hw_processing),
780 		  "Desc pending still in list desc_hw_processing\n");
781 
782 	/* Free all descriptors in list desc_completed */
783 	vchan_synchronize(&hvc->vc);
784 
785 	WARN_ONCE(!list_empty(&hvc->vc.desc_completed),
786 		  "Desc pending still in list desc_completed\n");
787 }
788 
789 static int mtk_hsdma_terminate_all(struct dma_chan *c)
790 {
791 	/*
792 	 * Free pending descriptors not processed yet by hardware that have
793 	 * previously been submitted to the channel.
794 	 */
795 	mtk_hsdma_free_inactive_desc(c);
796 
797 	/*
798 	 * However, the DMA engine doesn't provide any way to stop these
799 	 * descriptors being processed currently by hardware. The only way is
800 	 * to just waiting until these descriptors are all processed completely
801 	 * through mtk_hsdma_free_active_desc call.
802 	 */
803 	mtk_hsdma_free_active_desc(c);
804 
805 	return 0;
806 }
807 
808 static int mtk_hsdma_alloc_chan_resources(struct dma_chan *c)
809 {
810 	struct mtk_hsdma_device *hsdma = to_hsdma_dev(c);
811 	int err;
812 
813 	/*
814 	 * Since HSDMA has only one PC, the resource for PC is being allocated
815 	 * when the first VC is being created and the other VCs would run on
816 	 * the same PC.
817 	 */
818 	if (!refcount_read(&hsdma->pc_refcnt)) {
819 		err = mtk_hsdma_alloc_pchan(hsdma, hsdma->pc);
820 		if (err)
821 			return err;
822 		/*
823 		 * refcount_inc would complain increment on 0; use-after-free.
824 		 * Thus, we need to explicitly set it as 1 initially.
825 		 */
826 		refcount_set(&hsdma->pc_refcnt, 1);
827 	} else {
828 		refcount_inc(&hsdma->pc_refcnt);
829 	}
830 
831 	return 0;
832 }
833 
834 static void mtk_hsdma_free_chan_resources(struct dma_chan *c)
835 {
836 	struct mtk_hsdma_device *hsdma = to_hsdma_dev(c);
837 
838 	/* Free all descriptors in all lists on the VC */
839 	mtk_hsdma_terminate_all(c);
840 
841 	/* The resource for PC is not freed until all the VCs are destroyed */
842 	if (!refcount_dec_and_test(&hsdma->pc_refcnt))
843 		return;
844 
845 	mtk_hsdma_free_pchan(hsdma, hsdma->pc);
846 }
847 
848 static int mtk_hsdma_hw_init(struct mtk_hsdma_device *hsdma)
849 {
850 	int err;
851 
852 	pm_runtime_enable(hsdma2dev(hsdma));
853 	pm_runtime_get_sync(hsdma2dev(hsdma));
854 
855 	err = clk_prepare_enable(hsdma->clk);
856 	if (err)
857 		return err;
858 
859 	mtk_dma_write(hsdma, MTK_HSDMA_INT_ENABLE, 0);
860 	mtk_dma_write(hsdma, MTK_HSDMA_GLO, MTK_HSDMA_GLO_DEFAULT);
861 
862 	return 0;
863 }
864 
865 static int mtk_hsdma_hw_deinit(struct mtk_hsdma_device *hsdma)
866 {
867 	mtk_dma_write(hsdma, MTK_HSDMA_GLO, 0);
868 
869 	clk_disable_unprepare(hsdma->clk);
870 
871 	pm_runtime_put_sync(hsdma2dev(hsdma));
872 	pm_runtime_disable(hsdma2dev(hsdma));
873 
874 	return 0;
875 }
876 
877 static const struct mtk_hsdma_soc mt7623_soc = {
878 	.ddone = BIT(31),
879 	.ls0 = BIT(30),
880 };
881 
882 static const struct mtk_hsdma_soc mt7622_soc = {
883 	.ddone = BIT(15),
884 	.ls0 = BIT(14),
885 };
886 
887 static const struct of_device_id mtk_hsdma_match[] = {
888 	{ .compatible = "mediatek,mt7623-hsdma", .data = &mt7623_soc},
889 	{ .compatible = "mediatek,mt7622-hsdma", .data = &mt7622_soc},
890 	{ /* sentinel */ }
891 };
892 MODULE_DEVICE_TABLE(of, mtk_hsdma_match);
893 
894 static int mtk_hsdma_probe(struct platform_device *pdev)
895 {
896 	struct mtk_hsdma_device *hsdma;
897 	struct mtk_hsdma_vchan *vc;
898 	struct dma_device *dd;
899 	struct resource *res;
900 	int i, err;
901 
902 	hsdma = devm_kzalloc(&pdev->dev, sizeof(*hsdma), GFP_KERNEL);
903 	if (!hsdma)
904 		return -ENOMEM;
905 
906 	dd = &hsdma->ddev;
907 
908 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
909 	hsdma->base = devm_ioremap_resource(&pdev->dev, res);
910 	if (IS_ERR(hsdma->base))
911 		return PTR_ERR(hsdma->base);
912 
913 	hsdma->soc = of_device_get_match_data(&pdev->dev);
914 	if (!hsdma->soc) {
915 		dev_err(&pdev->dev, "No device match found\n");
916 		return -ENODEV;
917 	}
918 
919 	hsdma->clk = devm_clk_get(&pdev->dev, "hsdma");
920 	if (IS_ERR(hsdma->clk)) {
921 		dev_err(&pdev->dev, "No clock for %s\n",
922 			dev_name(&pdev->dev));
923 		return PTR_ERR(hsdma->clk);
924 	}
925 
926 	err = platform_get_irq(pdev, 0);
927 	if (err < 0)
928 		return err;
929 	hsdma->irq = err;
930 
931 	refcount_set(&hsdma->pc_refcnt, 0);
932 	spin_lock_init(&hsdma->lock);
933 
934 	dma_cap_set(DMA_MEMCPY, dd->cap_mask);
935 
936 	dd->copy_align = MTK_HSDMA_ALIGN_SIZE;
937 	dd->device_alloc_chan_resources = mtk_hsdma_alloc_chan_resources;
938 	dd->device_free_chan_resources = mtk_hsdma_free_chan_resources;
939 	dd->device_tx_status = mtk_hsdma_tx_status;
940 	dd->device_issue_pending = mtk_hsdma_issue_pending;
941 	dd->device_prep_dma_memcpy = mtk_hsdma_prep_dma_memcpy;
942 	dd->device_terminate_all = mtk_hsdma_terminate_all;
943 	dd->src_addr_widths = MTK_HSDMA_DMA_BUSWIDTHS;
944 	dd->dst_addr_widths = MTK_HSDMA_DMA_BUSWIDTHS;
945 	dd->directions = BIT(DMA_MEM_TO_MEM);
946 	dd->residue_granularity = DMA_RESIDUE_GRANULARITY_SEGMENT;
947 	dd->dev = &pdev->dev;
948 	INIT_LIST_HEAD(&dd->channels);
949 
950 	hsdma->dma_requests = MTK_HSDMA_NR_VCHANS;
951 	if (pdev->dev.of_node && of_property_read_u32(pdev->dev.of_node,
952 						      "dma-requests",
953 						      &hsdma->dma_requests)) {
954 		dev_info(&pdev->dev,
955 			 "Using %u as missing dma-requests property\n",
956 			 MTK_HSDMA_NR_VCHANS);
957 	}
958 
959 	hsdma->pc = devm_kcalloc(&pdev->dev, MTK_HSDMA_NR_MAX_PCHANS,
960 				 sizeof(*hsdma->pc), GFP_KERNEL);
961 	if (!hsdma->pc)
962 		return -ENOMEM;
963 
964 	hsdma->vc = devm_kcalloc(&pdev->dev, hsdma->dma_requests,
965 				 sizeof(*hsdma->vc), GFP_KERNEL);
966 	if (!hsdma->vc)
967 		return -ENOMEM;
968 
969 	for (i = 0; i < hsdma->dma_requests; i++) {
970 		vc = &hsdma->vc[i];
971 		vc->vc.desc_free = mtk_hsdma_vdesc_free;
972 		vchan_init(&vc->vc, dd);
973 		init_completion(&vc->issue_completion);
974 		INIT_LIST_HEAD(&vc->desc_hw_processing);
975 	}
976 
977 	err = dma_async_device_register(dd);
978 	if (err)
979 		return err;
980 
981 	err = of_dma_controller_register(pdev->dev.of_node,
982 					 of_dma_xlate_by_chan_id, hsdma);
983 	if (err) {
984 		dev_err(&pdev->dev,
985 			"MediaTek HSDMA OF registration failed %d\n", err);
986 		goto err_unregister;
987 	}
988 
989 	mtk_hsdma_hw_init(hsdma);
990 
991 	err = devm_request_irq(&pdev->dev, hsdma->irq,
992 			       mtk_hsdma_irq, 0,
993 			       dev_name(&pdev->dev), hsdma);
994 	if (err) {
995 		dev_err(&pdev->dev,
996 			"request_irq failed with err %d\n", err);
997 		goto err_free;
998 	}
999 
1000 	platform_set_drvdata(pdev, hsdma);
1001 
1002 	dev_info(&pdev->dev, "MediaTek HSDMA driver registered\n");
1003 
1004 	return 0;
1005 
1006 err_free:
1007 	mtk_hsdma_hw_deinit(hsdma);
1008 	of_dma_controller_free(pdev->dev.of_node);
1009 err_unregister:
1010 	dma_async_device_unregister(dd);
1011 
1012 	return err;
1013 }
1014 
1015 static int mtk_hsdma_remove(struct platform_device *pdev)
1016 {
1017 	struct mtk_hsdma_device *hsdma = platform_get_drvdata(pdev);
1018 	struct mtk_hsdma_vchan *vc;
1019 	int i;
1020 
1021 	/* Kill VC task */
1022 	for (i = 0; i < hsdma->dma_requests; i++) {
1023 		vc = &hsdma->vc[i];
1024 
1025 		list_del(&vc->vc.chan.device_node);
1026 		tasklet_kill(&vc->vc.task);
1027 	}
1028 
1029 	/* Disable DMA interrupt */
1030 	mtk_dma_write(hsdma, MTK_HSDMA_INT_ENABLE, 0);
1031 
1032 	/* Waits for any pending IRQ handlers to complete */
1033 	synchronize_irq(hsdma->irq);
1034 
1035 	/* Disable hardware */
1036 	mtk_hsdma_hw_deinit(hsdma);
1037 
1038 	dma_async_device_unregister(&hsdma->ddev);
1039 	of_dma_controller_free(pdev->dev.of_node);
1040 
1041 	return 0;
1042 }
1043 
1044 static struct platform_driver mtk_hsdma_driver = {
1045 	.probe		= mtk_hsdma_probe,
1046 	.remove		= mtk_hsdma_remove,
1047 	.driver = {
1048 		.name		= KBUILD_MODNAME,
1049 		.of_match_table	= mtk_hsdma_match,
1050 	},
1051 };
1052 module_platform_driver(mtk_hsdma_driver);
1053 
1054 MODULE_DESCRIPTION("MediaTek High-Speed DMA Controller Driver");
1055 MODULE_AUTHOR("Sean Wang <sean.wang@mediatek.com>");
1056 MODULE_LICENSE("GPL v2");
1057