xref: /linux/drivers/net/ipa/gsi.c (revision 4359a011e259a4608afc7fb3635370c9d4ba5943)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 /* Copyright (c) 2015-2018, The Linux Foundation. All rights reserved.
4  * Copyright (C) 2018-2021 Linaro Ltd.
5  */
6 
7 #include <linux/types.h>
8 #include <linux/bits.h>
9 #include <linux/bitfield.h>
10 #include <linux/mutex.h>
11 #include <linux/completion.h>
12 #include <linux/io.h>
13 #include <linux/bug.h>
14 #include <linux/interrupt.h>
15 #include <linux/platform_device.h>
16 #include <linux/netdevice.h>
17 
18 #include "gsi.h"
19 #include "gsi_reg.h"
20 #include "gsi_private.h"
21 #include "gsi_trans.h"
22 #include "ipa_gsi.h"
23 #include "ipa_data.h"
24 #include "ipa_version.h"
25 
26 /**
27  * DOC: The IPA Generic Software Interface
28  *
29  * The generic software interface (GSI) is an integral component of the IPA,
30  * providing a well-defined communication layer between the AP subsystem
31  * and the IPA core.  The modem uses the GSI layer as well.
32  *
33  *	--------	     ---------
34  *	|      |	     |	     |
35  *	|  AP  +<---.	.----+ Modem |
36  *	|      +--. |	| .->+	     |
37  *	|      |  | |	| |  |	     |
38  *	--------  | |	| |  ---------
39  *		  v |	v |
40  *		--+-+---+-+--
41  *		|    GSI    |
42  *		|-----------|
43  *		|	    |
44  *		|    IPA    |
45  *		|	    |
46  *		-------------
47  *
48  * In the above diagram, the AP and Modem represent "execution environments"
49  * (EEs), which are independent operating environments that use the IPA for
50  * data transfer.
51  *
52  * Each EE uses a set of unidirectional GSI "channels," which allow transfer
53  * of data to or from the IPA.  A channel is implemented as a ring buffer,
54  * with a DRAM-resident array of "transfer elements" (TREs) available to
55  * describe transfers to or from other EEs through the IPA.  A transfer
56  * element can also contain an immediate command, requesting the IPA perform
57  * actions other than data transfer.
58  *
59  * Each TRE refers to a block of data--also located DRAM.  After writing one
60  * or more TREs to a channel, the writer (either the IPA or an EE) writes a
61  * doorbell register to inform the receiving side how many elements have
62  * been written.
63  *
64  * Each channel has a GSI "event ring" associated with it.  An event ring
65  * is implemented very much like a channel ring, but is always directed from
66  * the IPA to an EE.  The IPA notifies an EE (such as the AP) about channel
67  * events by adding an entry to the event ring associated with the channel.
68  * The GSI then writes its doorbell for the event ring, causing the target
69  * EE to be interrupted.  Each entry in an event ring contains a pointer
70  * to the channel TRE whose completion the event represents.
71  *
72  * Each TRE in a channel ring has a set of flags.  One flag indicates whether
73  * the completion of the transfer operation generates an entry (and possibly
74  * an interrupt) in the channel's event ring.  Other flags allow transfer
75  * elements to be chained together, forming a single logical transaction.
76  * TRE flags are used to control whether and when interrupts are generated
77  * to signal completion of channel transfers.
78  *
79  * Elements in channel and event rings are completed (or consumed) strictly
80  * in order.  Completion of one entry implies the completion of all preceding
81  * entries.  A single completion interrupt can therefore communicate the
82  * completion of many transfers.
83  *
84  * Note that all GSI registers are little-endian, which is the assumed
85  * endianness of I/O space accesses.  The accessor functions perform byte
86  * swapping if needed (i.e., for a big endian CPU).
87  */
88 
89 /* Delay period for interrupt moderation (in 32KHz IPA internal timer ticks) */
90 #define GSI_EVT_RING_INT_MODT		(32 * 1) /* 1ms under 32KHz clock */
91 
92 #define GSI_CMD_TIMEOUT			50	/* milliseconds */
93 
94 #define GSI_CHANNEL_STOP_RETRIES	10
95 #define GSI_CHANNEL_MODEM_HALT_RETRIES	10
96 #define GSI_CHANNEL_MODEM_FLOW_RETRIES	5	/* disable flow control only */
97 
98 #define GSI_MHI_EVENT_ID_START		10	/* 1st reserved event id */
99 #define GSI_MHI_EVENT_ID_END		16	/* Last reserved event id */
100 
101 #define GSI_ISR_MAX_ITER		50	/* Detect interrupt storms */
102 
103 /* An entry in an event ring */
104 struct gsi_event {
105 	__le64 xfer_ptr;
106 	__le16 len;
107 	u8 reserved1;
108 	u8 code;
109 	__le16 reserved2;
110 	u8 type;
111 	u8 chid;
112 };
113 
114 /** gsi_channel_scratch_gpi - GPI protocol scratch register
115  * @max_outstanding_tre:
116  *	Defines the maximum number of TREs allowed in a single transaction
117  *	on a channel (in bytes).  This determines the amount of prefetch
118  *	performed by the hardware.  We configure this to equal the size of
119  *	the TLV FIFO for the channel.
120  * @outstanding_threshold:
121  *	Defines the threshold (in bytes) determining when the sequencer
122  *	should update the channel doorbell.  We configure this to equal
123  *	the size of two TREs.
124  */
125 struct gsi_channel_scratch_gpi {
126 	u64 reserved1;
127 	u16 reserved2;
128 	u16 max_outstanding_tre;
129 	u16 reserved3;
130 	u16 outstanding_threshold;
131 };
132 
133 /** gsi_channel_scratch - channel scratch configuration area
134  *
135  * The exact interpretation of this register is protocol-specific.
136  * We only use GPI channels; see struct gsi_channel_scratch_gpi, above.
137  */
138 union gsi_channel_scratch {
139 	struct gsi_channel_scratch_gpi gpi;
140 	struct {
141 		u32 word1;
142 		u32 word2;
143 		u32 word3;
144 		u32 word4;
145 	} data;
146 };
147 
148 /* Check things that can be validated at build time. */
149 static void gsi_validate_build(void)
150 {
151 	/* This is used as a divisor */
152 	BUILD_BUG_ON(!GSI_RING_ELEMENT_SIZE);
153 
154 	/* Code assumes the size of channel and event ring element are
155 	 * the same (and fixed).  Make sure the size of an event ring
156 	 * element is what's expected.
157 	 */
158 	BUILD_BUG_ON(sizeof(struct gsi_event) != GSI_RING_ELEMENT_SIZE);
159 
160 	/* Hardware requires a 2^n ring size.  We ensure the number of
161 	 * elements in an event ring is a power of 2 elsewhere; this
162 	 * ensure the elements themselves meet the requirement.
163 	 */
164 	BUILD_BUG_ON(!is_power_of_2(GSI_RING_ELEMENT_SIZE));
165 
166 	/* The channel element size must fit in this field */
167 	BUILD_BUG_ON(GSI_RING_ELEMENT_SIZE > field_max(ELEMENT_SIZE_FMASK));
168 
169 	/* The event ring element size must fit in this field */
170 	BUILD_BUG_ON(GSI_RING_ELEMENT_SIZE > field_max(EV_ELEMENT_SIZE_FMASK));
171 }
172 
173 /* Return the channel id associated with a given channel */
174 static u32 gsi_channel_id(struct gsi_channel *channel)
175 {
176 	return channel - &channel->gsi->channel[0];
177 }
178 
179 /* An initialized channel has a non-null GSI pointer */
180 static bool gsi_channel_initialized(struct gsi_channel *channel)
181 {
182 	return !!channel->gsi;
183 }
184 
185 /* Update the GSI IRQ type register with the cached value */
186 static void gsi_irq_type_update(struct gsi *gsi, u32 val)
187 {
188 	gsi->type_enabled_bitmap = val;
189 	iowrite32(val, gsi->virt + GSI_CNTXT_TYPE_IRQ_MSK_OFFSET);
190 }
191 
192 static void gsi_irq_type_enable(struct gsi *gsi, enum gsi_irq_type_id type_id)
193 {
194 	gsi_irq_type_update(gsi, gsi->type_enabled_bitmap | BIT(type_id));
195 }
196 
197 static void gsi_irq_type_disable(struct gsi *gsi, enum gsi_irq_type_id type_id)
198 {
199 	gsi_irq_type_update(gsi, gsi->type_enabled_bitmap & ~BIT(type_id));
200 }
201 
202 /* Event ring commands are performed one at a time.  Their completion
203  * is signaled by the event ring control GSI interrupt type, which is
204  * only enabled when we issue an event ring command.  Only the event
205  * ring being operated on has this interrupt enabled.
206  */
207 static void gsi_irq_ev_ctrl_enable(struct gsi *gsi, u32 evt_ring_id)
208 {
209 	u32 val = BIT(evt_ring_id);
210 
211 	/* There's a small chance that a previous command completed
212 	 * after the interrupt was disabled, so make sure we have no
213 	 * pending interrupts before we enable them.
214 	 */
215 	iowrite32(~0, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_CLR_OFFSET);
216 
217 	iowrite32(val, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_MSK_OFFSET);
218 	gsi_irq_type_enable(gsi, GSI_EV_CTRL);
219 }
220 
221 /* Disable event ring control interrupts */
222 static void gsi_irq_ev_ctrl_disable(struct gsi *gsi)
223 {
224 	gsi_irq_type_disable(gsi, GSI_EV_CTRL);
225 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_MSK_OFFSET);
226 }
227 
228 /* Channel commands are performed one at a time.  Their completion is
229  * signaled by the channel control GSI interrupt type, which is only
230  * enabled when we issue a channel command.  Only the channel being
231  * operated on has this interrupt enabled.
232  */
233 static void gsi_irq_ch_ctrl_enable(struct gsi *gsi, u32 channel_id)
234 {
235 	u32 val = BIT(channel_id);
236 
237 	/* There's a small chance that a previous command completed
238 	 * after the interrupt was disabled, so make sure we have no
239 	 * pending interrupts before we enable them.
240 	 */
241 	iowrite32(~0, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_CLR_OFFSET);
242 
243 	iowrite32(val, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_MSK_OFFSET);
244 	gsi_irq_type_enable(gsi, GSI_CH_CTRL);
245 }
246 
247 /* Disable channel control interrupts */
248 static void gsi_irq_ch_ctrl_disable(struct gsi *gsi)
249 {
250 	gsi_irq_type_disable(gsi, GSI_CH_CTRL);
251 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_MSK_OFFSET);
252 }
253 
254 static void gsi_irq_ieob_enable_one(struct gsi *gsi, u32 evt_ring_id)
255 {
256 	bool enable_ieob = !gsi->ieob_enabled_bitmap;
257 	u32 val;
258 
259 	gsi->ieob_enabled_bitmap |= BIT(evt_ring_id);
260 	val = gsi->ieob_enabled_bitmap;
261 	iowrite32(val, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_MSK_OFFSET);
262 
263 	/* Enable the interrupt type if this is the first channel enabled */
264 	if (enable_ieob)
265 		gsi_irq_type_enable(gsi, GSI_IEOB);
266 }
267 
268 static void gsi_irq_ieob_disable(struct gsi *gsi, u32 event_mask)
269 {
270 	u32 val;
271 
272 	gsi->ieob_enabled_bitmap &= ~event_mask;
273 
274 	/* Disable the interrupt type if this was the last enabled channel */
275 	if (!gsi->ieob_enabled_bitmap)
276 		gsi_irq_type_disable(gsi, GSI_IEOB);
277 
278 	val = gsi->ieob_enabled_bitmap;
279 	iowrite32(val, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_MSK_OFFSET);
280 }
281 
282 static void gsi_irq_ieob_disable_one(struct gsi *gsi, u32 evt_ring_id)
283 {
284 	gsi_irq_ieob_disable(gsi, BIT(evt_ring_id));
285 }
286 
287 /* Enable all GSI_interrupt types */
288 static void gsi_irq_enable(struct gsi *gsi)
289 {
290 	u32 val;
291 
292 	/* Global interrupts include hardware error reports.  Enable
293 	 * that so we can at least report the error should it occur.
294 	 */
295 	iowrite32(BIT(ERROR_INT), gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
296 	gsi_irq_type_update(gsi, gsi->type_enabled_bitmap | BIT(GSI_GLOB_EE));
297 
298 	/* General GSI interrupts are reported to all EEs; if they occur
299 	 * they are unrecoverable (without reset).  A breakpoint interrupt
300 	 * also exists, but we don't support that.  We want to be notified
301 	 * of errors so we can report them, even if they can't be handled.
302 	 */
303 	val = BIT(BUS_ERROR);
304 	val |= BIT(CMD_FIFO_OVRFLOW);
305 	val |= BIT(MCS_STACK_OVRFLOW);
306 	iowrite32(val, gsi->virt + GSI_CNTXT_GSI_IRQ_EN_OFFSET);
307 	gsi_irq_type_update(gsi, gsi->type_enabled_bitmap | BIT(GSI_GENERAL));
308 }
309 
310 /* Disable all GSI interrupt types */
311 static void gsi_irq_disable(struct gsi *gsi)
312 {
313 	gsi_irq_type_update(gsi, 0);
314 
315 	/* Clear the type-specific interrupt masks set by gsi_irq_enable() */
316 	iowrite32(0, gsi->virt + GSI_CNTXT_GSI_IRQ_EN_OFFSET);
317 	iowrite32(0, gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
318 }
319 
320 /* Return the virtual address associated with a ring index */
321 void *gsi_ring_virt(struct gsi_ring *ring, u32 index)
322 {
323 	/* Note: index *must* be used modulo the ring count here */
324 	return ring->virt + (index % ring->count) * GSI_RING_ELEMENT_SIZE;
325 }
326 
327 /* Return the 32-bit DMA address associated with a ring index */
328 static u32 gsi_ring_addr(struct gsi_ring *ring, u32 index)
329 {
330 	return lower_32_bits(ring->addr) + index * GSI_RING_ELEMENT_SIZE;
331 }
332 
333 /* Return the ring index of a 32-bit ring offset */
334 static u32 gsi_ring_index(struct gsi_ring *ring, u32 offset)
335 {
336 	return (offset - gsi_ring_addr(ring, 0)) / GSI_RING_ELEMENT_SIZE;
337 }
338 
339 /* Issue a GSI command by writing a value to a register, then wait for
340  * completion to be signaled.  Returns true if the command completes
341  * or false if it times out.
342  */
343 static bool gsi_command(struct gsi *gsi, u32 reg, u32 val)
344 {
345 	unsigned long timeout = msecs_to_jiffies(GSI_CMD_TIMEOUT);
346 	struct completion *completion = &gsi->completion;
347 
348 	reinit_completion(completion);
349 
350 	iowrite32(val, gsi->virt + reg);
351 
352 	return !!wait_for_completion_timeout(completion, timeout);
353 }
354 
355 /* Return the hardware's notion of the current state of an event ring */
356 static enum gsi_evt_ring_state
357 gsi_evt_ring_state(struct gsi *gsi, u32 evt_ring_id)
358 {
359 	u32 val;
360 
361 	val = ioread32(gsi->virt + GSI_EV_CH_E_CNTXT_0_OFFSET(evt_ring_id));
362 
363 	return u32_get_bits(val, EV_CHSTATE_FMASK);
364 }
365 
366 /* Issue an event ring command and wait for it to complete */
367 static void gsi_evt_ring_command(struct gsi *gsi, u32 evt_ring_id,
368 				 enum gsi_evt_cmd_opcode opcode)
369 {
370 	struct device *dev = gsi->dev;
371 	bool timeout;
372 	u32 val;
373 
374 	/* Enable the completion interrupt for the command */
375 	gsi_irq_ev_ctrl_enable(gsi, evt_ring_id);
376 
377 	val = u32_encode_bits(evt_ring_id, EV_CHID_FMASK);
378 	val |= u32_encode_bits(opcode, EV_OPCODE_FMASK);
379 
380 	timeout = !gsi_command(gsi, GSI_EV_CH_CMD_OFFSET, val);
381 
382 	gsi_irq_ev_ctrl_disable(gsi);
383 
384 	if (!timeout)
385 		return;
386 
387 	dev_err(dev, "GSI command %u for event ring %u timed out, state %u\n",
388 		opcode, evt_ring_id, gsi_evt_ring_state(gsi, evt_ring_id));
389 }
390 
391 /* Allocate an event ring in NOT_ALLOCATED state */
392 static int gsi_evt_ring_alloc_command(struct gsi *gsi, u32 evt_ring_id)
393 {
394 	enum gsi_evt_ring_state state;
395 
396 	/* Get initial event ring state */
397 	state = gsi_evt_ring_state(gsi, evt_ring_id);
398 	if (state != GSI_EVT_RING_STATE_NOT_ALLOCATED) {
399 		dev_err(gsi->dev, "event ring %u bad state %u before alloc\n",
400 			evt_ring_id, state);
401 		return -EINVAL;
402 	}
403 
404 	gsi_evt_ring_command(gsi, evt_ring_id, GSI_EVT_ALLOCATE);
405 
406 	/* If successful the event ring state will have changed */
407 	state = gsi_evt_ring_state(gsi, evt_ring_id);
408 	if (state == GSI_EVT_RING_STATE_ALLOCATED)
409 		return 0;
410 
411 	dev_err(gsi->dev, "event ring %u bad state %u after alloc\n",
412 		evt_ring_id, state);
413 
414 	return -EIO;
415 }
416 
417 /* Reset a GSI event ring in ALLOCATED or ERROR state. */
418 static void gsi_evt_ring_reset_command(struct gsi *gsi, u32 evt_ring_id)
419 {
420 	enum gsi_evt_ring_state state;
421 
422 	state = gsi_evt_ring_state(gsi, evt_ring_id);
423 	if (state != GSI_EVT_RING_STATE_ALLOCATED &&
424 	    state != GSI_EVT_RING_STATE_ERROR) {
425 		dev_err(gsi->dev, "event ring %u bad state %u before reset\n",
426 			evt_ring_id, state);
427 		return;
428 	}
429 
430 	gsi_evt_ring_command(gsi, evt_ring_id, GSI_EVT_RESET);
431 
432 	/* If successful the event ring state will have changed */
433 	state = gsi_evt_ring_state(gsi, evt_ring_id);
434 	if (state == GSI_EVT_RING_STATE_ALLOCATED)
435 		return;
436 
437 	dev_err(gsi->dev, "event ring %u bad state %u after reset\n",
438 		evt_ring_id, state);
439 }
440 
441 /* Issue a hardware de-allocation request for an allocated event ring */
442 static void gsi_evt_ring_de_alloc_command(struct gsi *gsi, u32 evt_ring_id)
443 {
444 	enum gsi_evt_ring_state state;
445 
446 	state = gsi_evt_ring_state(gsi, evt_ring_id);
447 	if (state != GSI_EVT_RING_STATE_ALLOCATED) {
448 		dev_err(gsi->dev, "event ring %u state %u before dealloc\n",
449 			evt_ring_id, state);
450 		return;
451 	}
452 
453 	gsi_evt_ring_command(gsi, evt_ring_id, GSI_EVT_DE_ALLOC);
454 
455 	/* If successful the event ring state will have changed */
456 	state = gsi_evt_ring_state(gsi, evt_ring_id);
457 	if (state == GSI_EVT_RING_STATE_NOT_ALLOCATED)
458 		return;
459 
460 	dev_err(gsi->dev, "event ring %u bad state %u after dealloc\n",
461 		evt_ring_id, state);
462 }
463 
464 /* Fetch the current state of a channel from hardware */
465 static enum gsi_channel_state gsi_channel_state(struct gsi_channel *channel)
466 {
467 	u32 channel_id = gsi_channel_id(channel);
468 	void __iomem *virt = channel->gsi->virt;
469 	u32 val;
470 
471 	val = ioread32(virt + GSI_CH_C_CNTXT_0_OFFSET(channel_id));
472 
473 	return u32_get_bits(val, CHSTATE_FMASK);
474 }
475 
476 /* Issue a channel command and wait for it to complete */
477 static void
478 gsi_channel_command(struct gsi_channel *channel, enum gsi_ch_cmd_opcode opcode)
479 {
480 	u32 channel_id = gsi_channel_id(channel);
481 	struct gsi *gsi = channel->gsi;
482 	struct device *dev = gsi->dev;
483 	bool timeout;
484 	u32 val;
485 
486 	/* Enable the completion interrupt for the command */
487 	gsi_irq_ch_ctrl_enable(gsi, channel_id);
488 
489 	val = u32_encode_bits(channel_id, CH_CHID_FMASK);
490 	val |= u32_encode_bits(opcode, CH_OPCODE_FMASK);
491 	timeout = !gsi_command(gsi, GSI_CH_CMD_OFFSET, val);
492 
493 	gsi_irq_ch_ctrl_disable(gsi);
494 
495 	if (!timeout)
496 		return;
497 
498 	dev_err(dev, "GSI command %u for channel %u timed out, state %u\n",
499 		opcode, channel_id, gsi_channel_state(channel));
500 }
501 
502 /* Allocate GSI channel in NOT_ALLOCATED state */
503 static int gsi_channel_alloc_command(struct gsi *gsi, u32 channel_id)
504 {
505 	struct gsi_channel *channel = &gsi->channel[channel_id];
506 	struct device *dev = gsi->dev;
507 	enum gsi_channel_state state;
508 
509 	/* Get initial channel state */
510 	state = gsi_channel_state(channel);
511 	if (state != GSI_CHANNEL_STATE_NOT_ALLOCATED) {
512 		dev_err(dev, "channel %u bad state %u before alloc\n",
513 			channel_id, state);
514 		return -EINVAL;
515 	}
516 
517 	gsi_channel_command(channel, GSI_CH_ALLOCATE);
518 
519 	/* If successful the channel state will have changed */
520 	state = gsi_channel_state(channel);
521 	if (state == GSI_CHANNEL_STATE_ALLOCATED)
522 		return 0;
523 
524 	dev_err(dev, "channel %u bad state %u after alloc\n",
525 		channel_id, state);
526 
527 	return -EIO;
528 }
529 
530 /* Start an ALLOCATED channel */
531 static int gsi_channel_start_command(struct gsi_channel *channel)
532 {
533 	struct device *dev = channel->gsi->dev;
534 	enum gsi_channel_state state;
535 
536 	state = gsi_channel_state(channel);
537 	if (state != GSI_CHANNEL_STATE_ALLOCATED &&
538 	    state != GSI_CHANNEL_STATE_STOPPED) {
539 		dev_err(dev, "channel %u bad state %u before start\n",
540 			gsi_channel_id(channel), state);
541 		return -EINVAL;
542 	}
543 
544 	gsi_channel_command(channel, GSI_CH_START);
545 
546 	/* If successful the channel state will have changed */
547 	state = gsi_channel_state(channel);
548 	if (state == GSI_CHANNEL_STATE_STARTED)
549 		return 0;
550 
551 	dev_err(dev, "channel %u bad state %u after start\n",
552 		gsi_channel_id(channel), state);
553 
554 	return -EIO;
555 }
556 
557 /* Stop a GSI channel in STARTED state */
558 static int gsi_channel_stop_command(struct gsi_channel *channel)
559 {
560 	struct device *dev = channel->gsi->dev;
561 	enum gsi_channel_state state;
562 
563 	state = gsi_channel_state(channel);
564 
565 	/* Channel could have entered STOPPED state since last call
566 	 * if it timed out.  If so, we're done.
567 	 */
568 	if (state == GSI_CHANNEL_STATE_STOPPED)
569 		return 0;
570 
571 	if (state != GSI_CHANNEL_STATE_STARTED &&
572 	    state != GSI_CHANNEL_STATE_STOP_IN_PROC) {
573 		dev_err(dev, "channel %u bad state %u before stop\n",
574 			gsi_channel_id(channel), state);
575 		return -EINVAL;
576 	}
577 
578 	gsi_channel_command(channel, GSI_CH_STOP);
579 
580 	/* If successful the channel state will have changed */
581 	state = gsi_channel_state(channel);
582 	if (state == GSI_CHANNEL_STATE_STOPPED)
583 		return 0;
584 
585 	/* We may have to try again if stop is in progress */
586 	if (state == GSI_CHANNEL_STATE_STOP_IN_PROC)
587 		return -EAGAIN;
588 
589 	dev_err(dev, "channel %u bad state %u after stop\n",
590 		gsi_channel_id(channel), state);
591 
592 	return -EIO;
593 }
594 
595 /* Reset a GSI channel in ALLOCATED or ERROR state. */
596 static void gsi_channel_reset_command(struct gsi_channel *channel)
597 {
598 	struct device *dev = channel->gsi->dev;
599 	enum gsi_channel_state state;
600 
601 	/* A short delay is required before a RESET command */
602 	usleep_range(USEC_PER_MSEC, 2 * USEC_PER_MSEC);
603 
604 	state = gsi_channel_state(channel);
605 	if (state != GSI_CHANNEL_STATE_STOPPED &&
606 	    state != GSI_CHANNEL_STATE_ERROR) {
607 		/* No need to reset a channel already in ALLOCATED state */
608 		if (state != GSI_CHANNEL_STATE_ALLOCATED)
609 			dev_err(dev, "channel %u bad state %u before reset\n",
610 				gsi_channel_id(channel), state);
611 		return;
612 	}
613 
614 	gsi_channel_command(channel, GSI_CH_RESET);
615 
616 	/* If successful the channel state will have changed */
617 	state = gsi_channel_state(channel);
618 	if (state != GSI_CHANNEL_STATE_ALLOCATED)
619 		dev_err(dev, "channel %u bad state %u after reset\n",
620 			gsi_channel_id(channel), state);
621 }
622 
623 /* Deallocate an ALLOCATED GSI channel */
624 static void gsi_channel_de_alloc_command(struct gsi *gsi, u32 channel_id)
625 {
626 	struct gsi_channel *channel = &gsi->channel[channel_id];
627 	struct device *dev = gsi->dev;
628 	enum gsi_channel_state state;
629 
630 	state = gsi_channel_state(channel);
631 	if (state != GSI_CHANNEL_STATE_ALLOCATED) {
632 		dev_err(dev, "channel %u bad state %u before dealloc\n",
633 			channel_id, state);
634 		return;
635 	}
636 
637 	gsi_channel_command(channel, GSI_CH_DE_ALLOC);
638 
639 	/* If successful the channel state will have changed */
640 	state = gsi_channel_state(channel);
641 
642 	if (state != GSI_CHANNEL_STATE_NOT_ALLOCATED)
643 		dev_err(dev, "channel %u bad state %u after dealloc\n",
644 			channel_id, state);
645 }
646 
647 /* Ring an event ring doorbell, reporting the last entry processed by the AP.
648  * The index argument (modulo the ring count) is the first unfilled entry, so
649  * we supply one less than that with the doorbell.  Update the event ring
650  * index field with the value provided.
651  */
652 static void gsi_evt_ring_doorbell(struct gsi *gsi, u32 evt_ring_id, u32 index)
653 {
654 	struct gsi_ring *ring = &gsi->evt_ring[evt_ring_id].ring;
655 	u32 val;
656 
657 	ring->index = index;	/* Next unused entry */
658 
659 	/* Note: index *must* be used modulo the ring count here */
660 	val = gsi_ring_addr(ring, (index - 1) % ring->count);
661 	iowrite32(val, gsi->virt + GSI_EV_CH_E_DOORBELL_0_OFFSET(evt_ring_id));
662 }
663 
664 /* Program an event ring for use */
665 static void gsi_evt_ring_program(struct gsi *gsi, u32 evt_ring_id)
666 {
667 	struct gsi_evt_ring *evt_ring = &gsi->evt_ring[evt_ring_id];
668 	struct gsi_ring *ring = &evt_ring->ring;
669 	size_t size;
670 	u32 val;
671 
672 	/* We program all event rings as GPI type/protocol */
673 	val = u32_encode_bits(GSI_CHANNEL_TYPE_GPI, EV_CHTYPE_FMASK);
674 	val |= EV_INTYPE_FMASK;
675 	val |= u32_encode_bits(GSI_RING_ELEMENT_SIZE, EV_ELEMENT_SIZE_FMASK);
676 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_0_OFFSET(evt_ring_id));
677 
678 	size = ring->count * GSI_RING_ELEMENT_SIZE;
679 	val = ev_r_length_encoded(gsi->version, size);
680 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_1_OFFSET(evt_ring_id));
681 
682 	/* The context 2 and 3 registers store the low-order and
683 	 * high-order 32 bits of the address of the event ring,
684 	 * respectively.
685 	 */
686 	val = lower_32_bits(ring->addr);
687 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_2_OFFSET(evt_ring_id));
688 	val = upper_32_bits(ring->addr);
689 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_3_OFFSET(evt_ring_id));
690 
691 	/* Enable interrupt moderation by setting the moderation delay */
692 	val = u32_encode_bits(GSI_EVT_RING_INT_MODT, MODT_FMASK);
693 	val |= u32_encode_bits(1, MODC_FMASK);	/* comes from channel */
694 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_8_OFFSET(evt_ring_id));
695 
696 	/* No MSI write data, and MSI address high and low address is 0 */
697 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_9_OFFSET(evt_ring_id));
698 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_10_OFFSET(evt_ring_id));
699 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_11_OFFSET(evt_ring_id));
700 
701 	/* We don't need to get event read pointer updates */
702 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_12_OFFSET(evt_ring_id));
703 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_13_OFFSET(evt_ring_id));
704 
705 	/* Finally, tell the hardware our "last processed" event (arbitrary) */
706 	gsi_evt_ring_doorbell(gsi, evt_ring_id, ring->index);
707 }
708 
709 /* Find the transaction whose completion indicates a channel is quiesced */
710 static struct gsi_trans *gsi_channel_trans_last(struct gsi_channel *channel)
711 {
712 	struct gsi_trans_info *trans_info = &channel->trans_info;
713 	const struct list_head *list;
714 	struct gsi_trans *trans;
715 
716 	spin_lock_bh(&trans_info->spinlock);
717 
718 	/* There is a small chance a TX transaction got allocated just
719 	 * before we disabled transmits, so check for that.
720 	 */
721 	if (channel->toward_ipa) {
722 		list = &trans_info->alloc;
723 		if (!list_empty(list))
724 			goto done;
725 		list = &trans_info->committed;
726 		if (!list_empty(list))
727 			goto done;
728 		list = &trans_info->pending;
729 		if (!list_empty(list))
730 			goto done;
731 	}
732 
733 	/* Otherwise (TX or RX) we want to wait for anything that
734 	 * has completed, or has been polled but not released yet.
735 	 */
736 	list = &trans_info->complete;
737 	if (!list_empty(list))
738 		goto done;
739 	list = &trans_info->polled;
740 	if (list_empty(list))
741 		list = NULL;
742 done:
743 	trans = list ? list_last_entry(list, struct gsi_trans, links) : NULL;
744 
745 	/* Caller will wait for this, so take a reference */
746 	if (trans)
747 		refcount_inc(&trans->refcount);
748 
749 	spin_unlock_bh(&trans_info->spinlock);
750 
751 	return trans;
752 }
753 
754 /* Wait for transaction activity on a channel to complete */
755 static void gsi_channel_trans_quiesce(struct gsi_channel *channel)
756 {
757 	struct gsi_trans *trans;
758 
759 	/* Get the last transaction, and wait for it to complete */
760 	trans = gsi_channel_trans_last(channel);
761 	if (trans) {
762 		wait_for_completion(&trans->completion);
763 		gsi_trans_free(trans);
764 	}
765 }
766 
767 /* Program a channel for use; there is no gsi_channel_deprogram() */
768 static void gsi_channel_program(struct gsi_channel *channel, bool doorbell)
769 {
770 	size_t size = channel->tre_ring.count * GSI_RING_ELEMENT_SIZE;
771 	u32 channel_id = gsi_channel_id(channel);
772 	union gsi_channel_scratch scr = { };
773 	struct gsi_channel_scratch_gpi *gpi;
774 	struct gsi *gsi = channel->gsi;
775 	u32 wrr_weight = 0;
776 	u32 val;
777 
778 	/* We program all channels as GPI type/protocol */
779 	val = chtype_protocol_encoded(gsi->version, GSI_CHANNEL_TYPE_GPI);
780 	if (channel->toward_ipa)
781 		val |= CHTYPE_DIR_FMASK;
782 	val |= u32_encode_bits(channel->evt_ring_id, ERINDEX_FMASK);
783 	val |= u32_encode_bits(GSI_RING_ELEMENT_SIZE, ELEMENT_SIZE_FMASK);
784 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_0_OFFSET(channel_id));
785 
786 	val = r_length_encoded(gsi->version, size);
787 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_1_OFFSET(channel_id));
788 
789 	/* The context 2 and 3 registers store the low-order and
790 	 * high-order 32 bits of the address of the channel ring,
791 	 * respectively.
792 	 */
793 	val = lower_32_bits(channel->tre_ring.addr);
794 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_2_OFFSET(channel_id));
795 	val = upper_32_bits(channel->tre_ring.addr);
796 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_3_OFFSET(channel_id));
797 
798 	/* Command channel gets low weighted round-robin priority */
799 	if (channel->command)
800 		wrr_weight = field_max(WRR_WEIGHT_FMASK);
801 	val = u32_encode_bits(wrr_weight, WRR_WEIGHT_FMASK);
802 
803 	/* Max prefetch is 1 segment (do not set MAX_PREFETCH_FMASK) */
804 
805 	/* No need to use the doorbell engine starting at IPA v4.0 */
806 	if (gsi->version < IPA_VERSION_4_0 && doorbell)
807 		val |= USE_DB_ENG_FMASK;
808 
809 	/* v4.0 introduces an escape buffer for prefetch.  We use it
810 	 * on all but the AP command channel.
811 	 */
812 	if (gsi->version >= IPA_VERSION_4_0 && !channel->command) {
813 		/* If not otherwise set, prefetch buffers are used */
814 		if (gsi->version < IPA_VERSION_4_5)
815 			val |= USE_ESCAPE_BUF_ONLY_FMASK;
816 		else
817 			val |= u32_encode_bits(GSI_ESCAPE_BUF_ONLY,
818 					       PREFETCH_MODE_FMASK);
819 	}
820 	/* All channels set DB_IN_BYTES */
821 	if (gsi->version >= IPA_VERSION_4_9)
822 		val |= DB_IN_BYTES;
823 
824 	iowrite32(val, gsi->virt + GSI_CH_C_QOS_OFFSET(channel_id));
825 
826 	/* Now update the scratch registers for GPI protocol */
827 	gpi = &scr.gpi;
828 	gpi->max_outstanding_tre = channel->trans_tre_max *
829 					GSI_RING_ELEMENT_SIZE;
830 	gpi->outstanding_threshold = 2 * GSI_RING_ELEMENT_SIZE;
831 
832 	val = scr.data.word1;
833 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_0_OFFSET(channel_id));
834 
835 	val = scr.data.word2;
836 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_1_OFFSET(channel_id));
837 
838 	val = scr.data.word3;
839 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_2_OFFSET(channel_id));
840 
841 	/* We must preserve the upper 16 bits of the last scratch register.
842 	 * The next sequence assumes those bits remain unchanged between the
843 	 * read and the write.
844 	 */
845 	val = ioread32(gsi->virt + GSI_CH_C_SCRATCH_3_OFFSET(channel_id));
846 	val = (scr.data.word4 & GENMASK(31, 16)) | (val & GENMASK(15, 0));
847 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_3_OFFSET(channel_id));
848 
849 	/* All done! */
850 }
851 
852 static int __gsi_channel_start(struct gsi_channel *channel, bool resume)
853 {
854 	struct gsi *gsi = channel->gsi;
855 	int ret;
856 
857 	/* Prior to IPA v4.0 suspend/resume is not implemented by GSI */
858 	if (resume && gsi->version < IPA_VERSION_4_0)
859 		return 0;
860 
861 	mutex_lock(&gsi->mutex);
862 
863 	ret = gsi_channel_start_command(channel);
864 
865 	mutex_unlock(&gsi->mutex);
866 
867 	return ret;
868 }
869 
870 /* Start an allocated GSI channel */
871 int gsi_channel_start(struct gsi *gsi, u32 channel_id)
872 {
873 	struct gsi_channel *channel = &gsi->channel[channel_id];
874 	int ret;
875 
876 	/* Enable NAPI and the completion interrupt */
877 	napi_enable(&channel->napi);
878 	gsi_irq_ieob_enable_one(gsi, channel->evt_ring_id);
879 
880 	ret = __gsi_channel_start(channel, false);
881 	if (ret) {
882 		gsi_irq_ieob_disable_one(gsi, channel->evt_ring_id);
883 		napi_disable(&channel->napi);
884 	}
885 
886 	return ret;
887 }
888 
889 static int gsi_channel_stop_retry(struct gsi_channel *channel)
890 {
891 	u32 retries = GSI_CHANNEL_STOP_RETRIES;
892 	int ret;
893 
894 	do {
895 		ret = gsi_channel_stop_command(channel);
896 		if (ret != -EAGAIN)
897 			break;
898 		usleep_range(3 * USEC_PER_MSEC, 5 * USEC_PER_MSEC);
899 	} while (retries--);
900 
901 	return ret;
902 }
903 
904 static int __gsi_channel_stop(struct gsi_channel *channel, bool suspend)
905 {
906 	struct gsi *gsi = channel->gsi;
907 	int ret;
908 
909 	/* Wait for any underway transactions to complete before stopping. */
910 	gsi_channel_trans_quiesce(channel);
911 
912 	/* Prior to IPA v4.0 suspend/resume is not implemented by GSI */
913 	if (suspend && gsi->version < IPA_VERSION_4_0)
914 		return 0;
915 
916 	mutex_lock(&gsi->mutex);
917 
918 	ret = gsi_channel_stop_retry(channel);
919 
920 	mutex_unlock(&gsi->mutex);
921 
922 	return ret;
923 }
924 
925 /* Stop a started channel */
926 int gsi_channel_stop(struct gsi *gsi, u32 channel_id)
927 {
928 	struct gsi_channel *channel = &gsi->channel[channel_id];
929 	int ret;
930 
931 	ret = __gsi_channel_stop(channel, false);
932 	if (ret)
933 		return ret;
934 
935 	/* Disable the completion interrupt and NAPI if successful */
936 	gsi_irq_ieob_disable_one(gsi, channel->evt_ring_id);
937 	napi_disable(&channel->napi);
938 
939 	return 0;
940 }
941 
942 /* Reset and reconfigure a channel, (possibly) enabling the doorbell engine */
943 void gsi_channel_reset(struct gsi *gsi, u32 channel_id, bool doorbell)
944 {
945 	struct gsi_channel *channel = &gsi->channel[channel_id];
946 
947 	mutex_lock(&gsi->mutex);
948 
949 	gsi_channel_reset_command(channel);
950 	/* Due to a hardware quirk we may need to reset RX channels twice. */
951 	if (gsi->version < IPA_VERSION_4_0 && !channel->toward_ipa)
952 		gsi_channel_reset_command(channel);
953 
954 	/* Hardware assumes this is 0 following reset */
955 	channel->tre_ring.index = 0;
956 	gsi_channel_program(channel, doorbell);
957 	gsi_channel_trans_cancel_pending(channel);
958 
959 	mutex_unlock(&gsi->mutex);
960 }
961 
962 /* Stop a started channel for suspend */
963 int gsi_channel_suspend(struct gsi *gsi, u32 channel_id)
964 {
965 	struct gsi_channel *channel = &gsi->channel[channel_id];
966 	int ret;
967 
968 	ret = __gsi_channel_stop(channel, true);
969 	if (ret)
970 		return ret;
971 
972 	/* Ensure NAPI polling has finished. */
973 	napi_synchronize(&channel->napi);
974 
975 	return 0;
976 }
977 
978 /* Resume a suspended channel (starting if stopped) */
979 int gsi_channel_resume(struct gsi *gsi, u32 channel_id)
980 {
981 	struct gsi_channel *channel = &gsi->channel[channel_id];
982 
983 	return __gsi_channel_start(channel, true);
984 }
985 
986 /* Prevent all GSI interrupts while suspended */
987 void gsi_suspend(struct gsi *gsi)
988 {
989 	disable_irq(gsi->irq);
990 }
991 
992 /* Allow all GSI interrupts again when resuming */
993 void gsi_resume(struct gsi *gsi)
994 {
995 	enable_irq(gsi->irq);
996 }
997 
998 void gsi_trans_tx_committed(struct gsi_trans *trans)
999 {
1000 	struct gsi_channel *channel = &trans->gsi->channel[trans->channel_id];
1001 
1002 	channel->trans_count++;
1003 	channel->byte_count += trans->len;
1004 
1005 	trans->trans_count = channel->trans_count;
1006 	trans->byte_count = channel->byte_count;
1007 }
1008 
1009 void gsi_trans_tx_queued(struct gsi_trans *trans)
1010 {
1011 	u32 channel_id = trans->channel_id;
1012 	struct gsi *gsi = trans->gsi;
1013 	struct gsi_channel *channel;
1014 	u32 trans_count;
1015 	u32 byte_count;
1016 
1017 	channel = &gsi->channel[channel_id];
1018 
1019 	byte_count = channel->byte_count - channel->queued_byte_count;
1020 	trans_count = channel->trans_count - channel->queued_trans_count;
1021 	channel->queued_byte_count = channel->byte_count;
1022 	channel->queued_trans_count = channel->trans_count;
1023 
1024 	ipa_gsi_channel_tx_queued(gsi, channel_id, trans_count, byte_count);
1025 }
1026 
1027 /**
1028  * gsi_trans_tx_completed() - Report completed TX transactions
1029  * @trans:	TX channel transaction that has completed
1030  *
1031  * Report that a transaction on a TX channel has completed.  At the time a
1032  * transaction is committed, we record *in the transaction* its channel's
1033  * committed transaction and byte counts.  Transactions are completed in
1034  * order, and the difference between the channel's byte/transaction count
1035  * when the transaction was committed and when it completes tells us
1036  * exactly how much data has been transferred while the transaction was
1037  * pending.
1038  *
1039  * We report this information to the network stack, which uses it to manage
1040  * the rate at which data is sent to hardware.
1041  */
1042 static void gsi_trans_tx_completed(struct gsi_trans *trans)
1043 {
1044 	u32 channel_id = trans->channel_id;
1045 	struct gsi *gsi = trans->gsi;
1046 	struct gsi_channel *channel;
1047 	u32 trans_count;
1048 	u32 byte_count;
1049 
1050 	channel = &gsi->channel[channel_id];
1051 	trans_count = trans->trans_count - channel->compl_trans_count;
1052 	byte_count = trans->byte_count - channel->compl_byte_count;
1053 
1054 	channel->compl_trans_count += trans_count;
1055 	channel->compl_byte_count += byte_count;
1056 
1057 	ipa_gsi_channel_tx_completed(gsi, channel_id, trans_count, byte_count);
1058 }
1059 
1060 /* Channel control interrupt handler */
1061 static void gsi_isr_chan_ctrl(struct gsi *gsi)
1062 {
1063 	u32 channel_mask;
1064 
1065 	channel_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_CH_IRQ_OFFSET);
1066 	iowrite32(channel_mask, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_CLR_OFFSET);
1067 
1068 	while (channel_mask) {
1069 		u32 channel_id = __ffs(channel_mask);
1070 
1071 		channel_mask ^= BIT(channel_id);
1072 
1073 		complete(&gsi->completion);
1074 	}
1075 }
1076 
1077 /* Event ring control interrupt handler */
1078 static void gsi_isr_evt_ctrl(struct gsi *gsi)
1079 {
1080 	u32 event_mask;
1081 
1082 	event_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_OFFSET);
1083 	iowrite32(event_mask, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_CLR_OFFSET);
1084 
1085 	while (event_mask) {
1086 		u32 evt_ring_id = __ffs(event_mask);
1087 
1088 		event_mask ^= BIT(evt_ring_id);
1089 
1090 		complete(&gsi->completion);
1091 	}
1092 }
1093 
1094 /* Global channel error interrupt handler */
1095 static void
1096 gsi_isr_glob_chan_err(struct gsi *gsi, u32 err_ee, u32 channel_id, u32 code)
1097 {
1098 	if (code == GSI_OUT_OF_RESOURCES) {
1099 		dev_err(gsi->dev, "channel %u out of resources\n", channel_id);
1100 		complete(&gsi->completion);
1101 		return;
1102 	}
1103 
1104 	/* Report, but otherwise ignore all other error codes */
1105 	dev_err(gsi->dev, "channel %u global error ee 0x%08x code 0x%08x\n",
1106 		channel_id, err_ee, code);
1107 }
1108 
1109 /* Global event error interrupt handler */
1110 static void
1111 gsi_isr_glob_evt_err(struct gsi *gsi, u32 err_ee, u32 evt_ring_id, u32 code)
1112 {
1113 	if (code == GSI_OUT_OF_RESOURCES) {
1114 		struct gsi_evt_ring *evt_ring = &gsi->evt_ring[evt_ring_id];
1115 		u32 channel_id = gsi_channel_id(evt_ring->channel);
1116 
1117 		complete(&gsi->completion);
1118 		dev_err(gsi->dev, "evt_ring for channel %u out of resources\n",
1119 			channel_id);
1120 		return;
1121 	}
1122 
1123 	/* Report, but otherwise ignore all other error codes */
1124 	dev_err(gsi->dev, "event ring %u global error ee %u code 0x%08x\n",
1125 		evt_ring_id, err_ee, code);
1126 }
1127 
1128 /* Global error interrupt handler */
1129 static void gsi_isr_glob_err(struct gsi *gsi)
1130 {
1131 	enum gsi_err_type type;
1132 	enum gsi_err_code code;
1133 	u32 which;
1134 	u32 val;
1135 	u32 ee;
1136 
1137 	/* Get the logged error, then reinitialize the log */
1138 	val = ioread32(gsi->virt + GSI_ERROR_LOG_OFFSET);
1139 	iowrite32(0, gsi->virt + GSI_ERROR_LOG_OFFSET);
1140 	iowrite32(~0, gsi->virt + GSI_ERROR_LOG_CLR_OFFSET);
1141 
1142 	ee = u32_get_bits(val, ERR_EE_FMASK);
1143 	type = u32_get_bits(val, ERR_TYPE_FMASK);
1144 	which = u32_get_bits(val, ERR_VIRT_IDX_FMASK);
1145 	code = u32_get_bits(val, ERR_CODE_FMASK);
1146 
1147 	if (type == GSI_ERR_TYPE_CHAN)
1148 		gsi_isr_glob_chan_err(gsi, ee, which, code);
1149 	else if (type == GSI_ERR_TYPE_EVT)
1150 		gsi_isr_glob_evt_err(gsi, ee, which, code);
1151 	else	/* type GSI_ERR_TYPE_GLOB should be fatal */
1152 		dev_err(gsi->dev, "unexpected global error 0x%08x\n", type);
1153 }
1154 
1155 /* Generic EE interrupt handler */
1156 static void gsi_isr_gp_int1(struct gsi *gsi)
1157 {
1158 	u32 result;
1159 	u32 val;
1160 
1161 	/* This interrupt is used to handle completions of GENERIC GSI
1162 	 * commands.  We use these to allocate and halt channels on the
1163 	 * modem's behalf due to a hardware quirk on IPA v4.2.  The modem
1164 	 * "owns" channels even when the AP allocates them, and have no
1165 	 * way of knowing whether a modem channel's state has been changed.
1166 	 *
1167 	 * We also use GENERIC commands to enable/disable channel flow
1168 	 * control for IPA v4.2+.
1169 	 *
1170 	 * It is recommended that we halt the modem channels we allocated
1171 	 * when shutting down, but it's possible the channel isn't running
1172 	 * at the time we issue the HALT command.  We'll get an error in
1173 	 * that case, but it's harmless (the channel is already halted).
1174 	 * Similarly, we could get an error back when updating flow control
1175 	 * on a channel because it's not in the proper state.
1176 	 *
1177 	 * In either case, we silently ignore a INCORRECT_CHANNEL_STATE
1178 	 * error if we receive it.
1179 	 */
1180 	val = ioread32(gsi->virt + GSI_CNTXT_SCRATCH_0_OFFSET);
1181 	result = u32_get_bits(val, GENERIC_EE_RESULT_FMASK);
1182 
1183 	switch (result) {
1184 	case GENERIC_EE_SUCCESS:
1185 	case GENERIC_EE_INCORRECT_CHANNEL_STATE:
1186 		gsi->result = 0;
1187 		break;
1188 
1189 	case GENERIC_EE_RETRY:
1190 		gsi->result = -EAGAIN;
1191 		break;
1192 
1193 	default:
1194 		dev_err(gsi->dev, "global INT1 generic result %u\n", result);
1195 		gsi->result = -EIO;
1196 		break;
1197 	}
1198 
1199 	complete(&gsi->completion);
1200 }
1201 
1202 /* Inter-EE interrupt handler */
1203 static void gsi_isr_glob_ee(struct gsi *gsi)
1204 {
1205 	u32 val;
1206 
1207 	val = ioread32(gsi->virt + GSI_CNTXT_GLOB_IRQ_STTS_OFFSET);
1208 
1209 	if (val & BIT(ERROR_INT))
1210 		gsi_isr_glob_err(gsi);
1211 
1212 	iowrite32(val, gsi->virt + GSI_CNTXT_GLOB_IRQ_CLR_OFFSET);
1213 
1214 	val &= ~BIT(ERROR_INT);
1215 
1216 	if (val & BIT(GP_INT1)) {
1217 		val ^= BIT(GP_INT1);
1218 		gsi_isr_gp_int1(gsi);
1219 	}
1220 
1221 	if (val)
1222 		dev_err(gsi->dev, "unexpected global interrupt 0x%08x\n", val);
1223 }
1224 
1225 /* I/O completion interrupt event */
1226 static void gsi_isr_ieob(struct gsi *gsi)
1227 {
1228 	u32 event_mask;
1229 
1230 	event_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_OFFSET);
1231 	gsi_irq_ieob_disable(gsi, event_mask);
1232 	iowrite32(event_mask, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_CLR_OFFSET);
1233 
1234 	while (event_mask) {
1235 		u32 evt_ring_id = __ffs(event_mask);
1236 
1237 		event_mask ^= BIT(evt_ring_id);
1238 
1239 		napi_schedule(&gsi->evt_ring[evt_ring_id].channel->napi);
1240 	}
1241 }
1242 
1243 /* General event interrupts represent serious problems, so report them */
1244 static void gsi_isr_general(struct gsi *gsi)
1245 {
1246 	struct device *dev = gsi->dev;
1247 	u32 val;
1248 
1249 	val = ioread32(gsi->virt + GSI_CNTXT_GSI_IRQ_STTS_OFFSET);
1250 	iowrite32(val, gsi->virt + GSI_CNTXT_GSI_IRQ_CLR_OFFSET);
1251 
1252 	dev_err(dev, "unexpected general interrupt 0x%08x\n", val);
1253 }
1254 
1255 /**
1256  * gsi_isr() - Top level GSI interrupt service routine
1257  * @irq:	Interrupt number (ignored)
1258  * @dev_id:	GSI pointer supplied to request_irq()
1259  *
1260  * This is the main handler function registered for the GSI IRQ. Each type
1261  * of interrupt has a separate handler function that is called from here.
1262  */
1263 static irqreturn_t gsi_isr(int irq, void *dev_id)
1264 {
1265 	struct gsi *gsi = dev_id;
1266 	u32 intr_mask;
1267 	u32 cnt = 0;
1268 
1269 	/* enum gsi_irq_type_id defines GSI interrupt types */
1270 	while ((intr_mask = ioread32(gsi->virt + GSI_CNTXT_TYPE_IRQ_OFFSET))) {
1271 		/* intr_mask contains bitmask of pending GSI interrupts */
1272 		do {
1273 			u32 gsi_intr = BIT(__ffs(intr_mask));
1274 
1275 			intr_mask ^= gsi_intr;
1276 
1277 			switch (gsi_intr) {
1278 			case BIT(GSI_CH_CTRL):
1279 				gsi_isr_chan_ctrl(gsi);
1280 				break;
1281 			case BIT(GSI_EV_CTRL):
1282 				gsi_isr_evt_ctrl(gsi);
1283 				break;
1284 			case BIT(GSI_GLOB_EE):
1285 				gsi_isr_glob_ee(gsi);
1286 				break;
1287 			case BIT(GSI_IEOB):
1288 				gsi_isr_ieob(gsi);
1289 				break;
1290 			case BIT(GSI_GENERAL):
1291 				gsi_isr_general(gsi);
1292 				break;
1293 			default:
1294 				dev_err(gsi->dev,
1295 					"unrecognized interrupt type 0x%08x\n",
1296 					gsi_intr);
1297 				break;
1298 			}
1299 		} while (intr_mask);
1300 
1301 		if (++cnt > GSI_ISR_MAX_ITER) {
1302 			dev_err(gsi->dev, "interrupt flood\n");
1303 			break;
1304 		}
1305 	}
1306 
1307 	return IRQ_HANDLED;
1308 }
1309 
1310 /* Init function for GSI IRQ lookup; there is no gsi_irq_exit() */
1311 static int gsi_irq_init(struct gsi *gsi, struct platform_device *pdev)
1312 {
1313 	int ret;
1314 
1315 	ret = platform_get_irq_byname(pdev, "gsi");
1316 	if (ret <= 0)
1317 		return ret ? : -EINVAL;
1318 
1319 	gsi->irq = ret;
1320 
1321 	return 0;
1322 }
1323 
1324 /* Return the transaction associated with a transfer completion event */
1325 static struct gsi_trans *
1326 gsi_event_trans(struct gsi *gsi, struct gsi_event *event)
1327 {
1328 	u32 channel_id = event->chid;
1329 	struct gsi_channel *channel;
1330 	struct gsi_trans *trans;
1331 	u32 tre_offset;
1332 	u32 tre_index;
1333 
1334 	channel = &gsi->channel[channel_id];
1335 	if (WARN(!channel->gsi, "event has bad channel %u\n", channel_id))
1336 		return NULL;
1337 
1338 	/* Event xfer_ptr records the TRE it's associated with */
1339 	tre_offset = lower_32_bits(le64_to_cpu(event->xfer_ptr));
1340 	tre_index = gsi_ring_index(&channel->tre_ring, tre_offset);
1341 
1342 	trans = gsi_channel_trans_mapped(channel, tre_index);
1343 
1344 	if (WARN(!trans, "channel %u event with no transaction\n", channel_id))
1345 		return NULL;
1346 
1347 	return trans;
1348 }
1349 
1350 /**
1351  * gsi_evt_ring_update() - Update transaction state from hardware
1352  * @gsi:		GSI pointer
1353  * @evt_ring_id:	Event ring ID
1354  * @index:		Event index in ring reported by hardware
1355  *
1356  * Events for RX channels contain the actual number of bytes received into
1357  * the buffer.  Every event has a transaction associated with it, and here
1358  * we update transactions to record their actual received lengths.
1359  *
1360  * When an event for a TX channel arrives we use information in the
1361  * transaction to report the number of requests and bytes have been
1362  * transferred.
1363  *
1364  * This function is called whenever we learn that the GSI hardware has filled
1365  * new events since the last time we checked.  The ring's index field tells
1366  * the first entry in need of processing.  The index provided is the
1367  * first *unfilled* event in the ring (following the last filled one).
1368  *
1369  * Events are sequential within the event ring, and transactions are
1370  * sequential within the transaction array.
1371  *
1372  * Note that @index always refers to an element *within* the event ring.
1373  */
1374 static void gsi_evt_ring_update(struct gsi *gsi, u32 evt_ring_id, u32 index)
1375 {
1376 	struct gsi_evt_ring *evt_ring = &gsi->evt_ring[evt_ring_id];
1377 	struct gsi_ring *ring = &evt_ring->ring;
1378 	struct gsi_event *event_done;
1379 	struct gsi_event *event;
1380 	u32 event_avail;
1381 	u32 old_index;
1382 
1383 	/* Starting with the oldest un-processed event, determine which
1384 	 * transaction (and which channel) is associated with the event.
1385 	 * For RX channels, update each completed transaction with the
1386 	 * number of bytes that were actually received.  For TX channels
1387 	 * associated with a network device, report to the network stack
1388 	 * the number of transfers and bytes this completion represents.
1389 	 */
1390 	old_index = ring->index;
1391 	event = gsi_ring_virt(ring, old_index);
1392 
1393 	/* Compute the number of events to process before we wrap,
1394 	 * and determine when we'll be done processing events.
1395 	 */
1396 	event_avail = ring->count - old_index % ring->count;
1397 	event_done = gsi_ring_virt(ring, index);
1398 	do {
1399 		struct gsi_trans *trans;
1400 
1401 		trans = gsi_event_trans(gsi, event);
1402 		if (!trans)
1403 			return;
1404 
1405 		if (trans->direction == DMA_FROM_DEVICE)
1406 			trans->len = __le16_to_cpu(event->len);
1407 		else
1408 			gsi_trans_tx_completed(trans);
1409 
1410 		gsi_trans_move_complete(trans);
1411 
1412 		/* Move on to the next event and transaction */
1413 		if (--event_avail)
1414 			event++;
1415 		else
1416 			event = gsi_ring_virt(ring, 0);
1417 	} while (event != event_done);
1418 
1419 	/* Tell the hardware we've handled these events */
1420 	gsi_evt_ring_doorbell(gsi, evt_ring_id, index);
1421 }
1422 
1423 /* Initialize a ring, including allocating DMA memory for its entries */
1424 static int gsi_ring_alloc(struct gsi *gsi, struct gsi_ring *ring, u32 count)
1425 {
1426 	u32 size = count * GSI_RING_ELEMENT_SIZE;
1427 	struct device *dev = gsi->dev;
1428 	dma_addr_t addr;
1429 
1430 	/* Hardware requires a 2^n ring size, with alignment equal to size.
1431 	 * The DMA address returned by dma_alloc_coherent() is guaranteed to
1432 	 * be a power-of-2 number of pages, which satisfies the requirement.
1433 	 */
1434 	ring->virt = dma_alloc_coherent(dev, size, &addr, GFP_KERNEL);
1435 	if (!ring->virt)
1436 		return -ENOMEM;
1437 
1438 	ring->addr = addr;
1439 	ring->count = count;
1440 	ring->index = 0;
1441 
1442 	return 0;
1443 }
1444 
1445 /* Free a previously-allocated ring */
1446 static void gsi_ring_free(struct gsi *gsi, struct gsi_ring *ring)
1447 {
1448 	size_t size = ring->count * GSI_RING_ELEMENT_SIZE;
1449 
1450 	dma_free_coherent(gsi->dev, size, ring->virt, ring->addr);
1451 }
1452 
1453 /* Allocate an available event ring id */
1454 static int gsi_evt_ring_id_alloc(struct gsi *gsi)
1455 {
1456 	u32 evt_ring_id;
1457 
1458 	if (gsi->event_bitmap == ~0U) {
1459 		dev_err(gsi->dev, "event rings exhausted\n");
1460 		return -ENOSPC;
1461 	}
1462 
1463 	evt_ring_id = ffz(gsi->event_bitmap);
1464 	gsi->event_bitmap |= BIT(evt_ring_id);
1465 
1466 	return (int)evt_ring_id;
1467 }
1468 
1469 /* Free a previously-allocated event ring id */
1470 static void gsi_evt_ring_id_free(struct gsi *gsi, u32 evt_ring_id)
1471 {
1472 	gsi->event_bitmap &= ~BIT(evt_ring_id);
1473 }
1474 
1475 /* Ring a channel doorbell, reporting the first un-filled entry */
1476 void gsi_channel_doorbell(struct gsi_channel *channel)
1477 {
1478 	struct gsi_ring *tre_ring = &channel->tre_ring;
1479 	u32 channel_id = gsi_channel_id(channel);
1480 	struct gsi *gsi = channel->gsi;
1481 	u32 val;
1482 
1483 	/* Note: index *must* be used modulo the ring count here */
1484 	val = gsi_ring_addr(tre_ring, tre_ring->index % tre_ring->count);
1485 	iowrite32(val, gsi->virt + GSI_CH_C_DOORBELL_0_OFFSET(channel_id));
1486 }
1487 
1488 /* Consult hardware, move any newly completed transactions to completed list */
1489 static struct gsi_trans *gsi_channel_update(struct gsi_channel *channel)
1490 {
1491 	u32 evt_ring_id = channel->evt_ring_id;
1492 	struct gsi *gsi = channel->gsi;
1493 	struct gsi_evt_ring *evt_ring;
1494 	struct gsi_trans *trans;
1495 	struct gsi_ring *ring;
1496 	u32 offset;
1497 	u32 index;
1498 
1499 	evt_ring = &gsi->evt_ring[evt_ring_id];
1500 	ring = &evt_ring->ring;
1501 
1502 	/* See if there's anything new to process; if not, we're done.  Note
1503 	 * that index always refers to an entry *within* the event ring.
1504 	 */
1505 	offset = GSI_EV_CH_E_CNTXT_4_OFFSET(evt_ring_id);
1506 	index = gsi_ring_index(ring, ioread32(gsi->virt + offset));
1507 	if (index == ring->index % ring->count)
1508 		return NULL;
1509 
1510 	/* Get the transaction for the latest completed event. */
1511 	trans = gsi_event_trans(gsi, gsi_ring_virt(ring, index - 1));
1512 	if (!trans)
1513 		return NULL;
1514 
1515 	/* For RX channels, update each completed transaction with the number
1516 	 * of bytes that were actually received.  For TX channels, report
1517 	 * the number of transactions and bytes this completion represents
1518 	 * up the network stack.
1519 	 */
1520 	gsi_evt_ring_update(gsi, evt_ring_id, index);
1521 
1522 	return gsi_channel_trans_complete(channel);
1523 }
1524 
1525 /**
1526  * gsi_channel_poll_one() - Return a single completed transaction on a channel
1527  * @channel:	Channel to be polled
1528  *
1529  * Return:	Transaction pointer, or null if none are available
1530  *
1531  * This function returns the first entry on a channel's completed transaction
1532  * list.  If that list is empty, the hardware is consulted to determine
1533  * whether any new transactions have completed.  If so, they're moved to the
1534  * completed list and the new first entry is returned.  If there are no more
1535  * completed transactions, a null pointer is returned.
1536  */
1537 static struct gsi_trans *gsi_channel_poll_one(struct gsi_channel *channel)
1538 {
1539 	struct gsi_trans *trans;
1540 
1541 	/* Get the first transaction from the completed list */
1542 	trans = gsi_channel_trans_complete(channel);
1543 	if (!trans)	/* List is empty; see if there's more to do */
1544 		trans = gsi_channel_update(channel);
1545 
1546 	if (trans)
1547 		gsi_trans_move_polled(trans);
1548 
1549 	return trans;
1550 }
1551 
1552 /**
1553  * gsi_channel_poll() - NAPI poll function for a channel
1554  * @napi:	NAPI structure for the channel
1555  * @budget:	Budget supplied by NAPI core
1556  *
1557  * Return:	Number of items polled (<= budget)
1558  *
1559  * Single transactions completed by hardware are polled until either
1560  * the budget is exhausted, or there are no more.  Each transaction
1561  * polled is passed to gsi_trans_complete(), to perform remaining
1562  * completion processing and retire/free the transaction.
1563  */
1564 static int gsi_channel_poll(struct napi_struct *napi, int budget)
1565 {
1566 	struct gsi_channel *channel;
1567 	int count;
1568 
1569 	channel = container_of(napi, struct gsi_channel, napi);
1570 	for (count = 0; count < budget; count++) {
1571 		struct gsi_trans *trans;
1572 
1573 		trans = gsi_channel_poll_one(channel);
1574 		if (!trans)
1575 			break;
1576 		gsi_trans_complete(trans);
1577 	}
1578 
1579 	if (count < budget && napi_complete(napi))
1580 		gsi_irq_ieob_enable_one(channel->gsi, channel->evt_ring_id);
1581 
1582 	return count;
1583 }
1584 
1585 /* The event bitmap represents which event ids are available for allocation.
1586  * Set bits are not available, clear bits can be used.  This function
1587  * initializes the map so all events supported by the hardware are available,
1588  * then precludes any reserved events from being allocated.
1589  */
1590 static u32 gsi_event_bitmap_init(u32 evt_ring_max)
1591 {
1592 	u32 event_bitmap = GENMASK(BITS_PER_LONG - 1, evt_ring_max);
1593 
1594 	event_bitmap |= GENMASK(GSI_MHI_EVENT_ID_END, GSI_MHI_EVENT_ID_START);
1595 
1596 	return event_bitmap;
1597 }
1598 
1599 /* Setup function for a single channel */
1600 static int gsi_channel_setup_one(struct gsi *gsi, u32 channel_id)
1601 {
1602 	struct gsi_channel *channel = &gsi->channel[channel_id];
1603 	u32 evt_ring_id = channel->evt_ring_id;
1604 	int ret;
1605 
1606 	if (!gsi_channel_initialized(channel))
1607 		return 0;
1608 
1609 	ret = gsi_evt_ring_alloc_command(gsi, evt_ring_id);
1610 	if (ret)
1611 		return ret;
1612 
1613 	gsi_evt_ring_program(gsi, evt_ring_id);
1614 
1615 	ret = gsi_channel_alloc_command(gsi, channel_id);
1616 	if (ret)
1617 		goto err_evt_ring_de_alloc;
1618 
1619 	gsi_channel_program(channel, true);
1620 
1621 	if (channel->toward_ipa)
1622 		netif_napi_add_tx(&gsi->dummy_dev, &channel->napi,
1623 				  gsi_channel_poll);
1624 	else
1625 		netif_napi_add(&gsi->dummy_dev, &channel->napi,
1626 			       gsi_channel_poll, NAPI_POLL_WEIGHT);
1627 
1628 	return 0;
1629 
1630 err_evt_ring_de_alloc:
1631 	/* We've done nothing with the event ring yet so don't reset */
1632 	gsi_evt_ring_de_alloc_command(gsi, evt_ring_id);
1633 
1634 	return ret;
1635 }
1636 
1637 /* Inverse of gsi_channel_setup_one() */
1638 static void gsi_channel_teardown_one(struct gsi *gsi, u32 channel_id)
1639 {
1640 	struct gsi_channel *channel = &gsi->channel[channel_id];
1641 	u32 evt_ring_id = channel->evt_ring_id;
1642 
1643 	if (!gsi_channel_initialized(channel))
1644 		return;
1645 
1646 	netif_napi_del(&channel->napi);
1647 
1648 	gsi_channel_de_alloc_command(gsi, channel_id);
1649 	gsi_evt_ring_reset_command(gsi, evt_ring_id);
1650 	gsi_evt_ring_de_alloc_command(gsi, evt_ring_id);
1651 }
1652 
1653 /* We use generic commands only to operate on modem channels.  We don't have
1654  * the ability to determine channel state for a modem channel, so we simply
1655  * issue the command and wait for it to complete.
1656  */
1657 static int gsi_generic_command(struct gsi *gsi, u32 channel_id,
1658 			       enum gsi_generic_cmd_opcode opcode,
1659 			       u8 params)
1660 {
1661 	bool timeout;
1662 	u32 val;
1663 
1664 	/* The error global interrupt type is always enabled (until we tear
1665 	 * down), so we will keep it enabled.
1666 	 *
1667 	 * A generic EE command completes with a GSI global interrupt of
1668 	 * type GP_INT1.  We only perform one generic command at a time
1669 	 * (to allocate, halt, or enable/disable flow control on a modem
1670 	 * channel), and only from this function.  So we enable the GP_INT1
1671 	 * IRQ type here, and disable it again after the command completes.
1672 	 */
1673 	val = BIT(ERROR_INT) | BIT(GP_INT1);
1674 	iowrite32(val, gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
1675 
1676 	/* First zero the result code field */
1677 	val = ioread32(gsi->virt + GSI_CNTXT_SCRATCH_0_OFFSET);
1678 	val &= ~GENERIC_EE_RESULT_FMASK;
1679 	iowrite32(val, gsi->virt + GSI_CNTXT_SCRATCH_0_OFFSET);
1680 
1681 	/* Now issue the command */
1682 	val = u32_encode_bits(opcode, GENERIC_OPCODE_FMASK);
1683 	val |= u32_encode_bits(channel_id, GENERIC_CHID_FMASK);
1684 	val |= u32_encode_bits(GSI_EE_MODEM, GENERIC_EE_FMASK);
1685 	val |= u32_encode_bits(params, GENERIC_PARAMS_FMASK);
1686 
1687 	timeout = !gsi_command(gsi, GSI_GENERIC_CMD_OFFSET, val);
1688 
1689 	/* Disable the GP_INT1 IRQ type again */
1690 	iowrite32(BIT(ERROR_INT), gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
1691 
1692 	if (!timeout)
1693 		return gsi->result;
1694 
1695 	dev_err(gsi->dev, "GSI generic command %u to channel %u timed out\n",
1696 		opcode, channel_id);
1697 
1698 	return -ETIMEDOUT;
1699 }
1700 
1701 static int gsi_modem_channel_alloc(struct gsi *gsi, u32 channel_id)
1702 {
1703 	return gsi_generic_command(gsi, channel_id,
1704 				   GSI_GENERIC_ALLOCATE_CHANNEL, 0);
1705 }
1706 
1707 static void gsi_modem_channel_halt(struct gsi *gsi, u32 channel_id)
1708 {
1709 	u32 retries = GSI_CHANNEL_MODEM_HALT_RETRIES;
1710 	int ret;
1711 
1712 	do
1713 		ret = gsi_generic_command(gsi, channel_id,
1714 					  GSI_GENERIC_HALT_CHANNEL, 0);
1715 	while (ret == -EAGAIN && retries--);
1716 
1717 	if (ret)
1718 		dev_err(gsi->dev, "error %d halting modem channel %u\n",
1719 			ret, channel_id);
1720 }
1721 
1722 /* Enable or disable flow control for a modem GSI TX channel (IPA v4.2+) */
1723 void
1724 gsi_modem_channel_flow_control(struct gsi *gsi, u32 channel_id, bool enable)
1725 {
1726 	u32 retries = 0;
1727 	u32 command;
1728 	int ret;
1729 
1730 	command = enable ? GSI_GENERIC_ENABLE_FLOW_CONTROL
1731 			 : GSI_GENERIC_DISABLE_FLOW_CONTROL;
1732 	/* Disabling flow control on IPA v4.11+ can return -EAGAIN if enable
1733 	 * is underway.  In this case we need to retry the command.
1734 	 */
1735 	if (!enable && gsi->version >= IPA_VERSION_4_11)
1736 		retries = GSI_CHANNEL_MODEM_FLOW_RETRIES;
1737 
1738 	do
1739 		ret = gsi_generic_command(gsi, channel_id, command, 0);
1740 	while (ret == -EAGAIN && retries--);
1741 
1742 	if (ret)
1743 		dev_err(gsi->dev,
1744 			"error %d %sabling mode channel %u flow control\n",
1745 			ret, enable ? "en" : "dis", channel_id);
1746 }
1747 
1748 /* Setup function for channels */
1749 static int gsi_channel_setup(struct gsi *gsi)
1750 {
1751 	u32 channel_id = 0;
1752 	u32 mask;
1753 	int ret;
1754 
1755 	gsi_irq_enable(gsi);
1756 
1757 	mutex_lock(&gsi->mutex);
1758 
1759 	do {
1760 		ret = gsi_channel_setup_one(gsi, channel_id);
1761 		if (ret)
1762 			goto err_unwind;
1763 	} while (++channel_id < gsi->channel_count);
1764 
1765 	/* Make sure no channels were defined that hardware does not support */
1766 	while (channel_id < GSI_CHANNEL_COUNT_MAX) {
1767 		struct gsi_channel *channel = &gsi->channel[channel_id++];
1768 
1769 		if (!gsi_channel_initialized(channel))
1770 			continue;
1771 
1772 		ret = -EINVAL;
1773 		dev_err(gsi->dev, "channel %u not supported by hardware\n",
1774 			channel_id - 1);
1775 		channel_id = gsi->channel_count;
1776 		goto err_unwind;
1777 	}
1778 
1779 	/* Allocate modem channels if necessary */
1780 	mask = gsi->modem_channel_bitmap;
1781 	while (mask) {
1782 		u32 modem_channel_id = __ffs(mask);
1783 
1784 		ret = gsi_modem_channel_alloc(gsi, modem_channel_id);
1785 		if (ret)
1786 			goto err_unwind_modem;
1787 
1788 		/* Clear bit from mask only after success (for unwind) */
1789 		mask ^= BIT(modem_channel_id);
1790 	}
1791 
1792 	mutex_unlock(&gsi->mutex);
1793 
1794 	return 0;
1795 
1796 err_unwind_modem:
1797 	/* Compute which modem channels need to be deallocated */
1798 	mask ^= gsi->modem_channel_bitmap;
1799 	while (mask) {
1800 		channel_id = __fls(mask);
1801 
1802 		mask ^= BIT(channel_id);
1803 
1804 		gsi_modem_channel_halt(gsi, channel_id);
1805 	}
1806 
1807 err_unwind:
1808 	while (channel_id--)
1809 		gsi_channel_teardown_one(gsi, channel_id);
1810 
1811 	mutex_unlock(&gsi->mutex);
1812 
1813 	gsi_irq_disable(gsi);
1814 
1815 	return ret;
1816 }
1817 
1818 /* Inverse of gsi_channel_setup() */
1819 static void gsi_channel_teardown(struct gsi *gsi)
1820 {
1821 	u32 mask = gsi->modem_channel_bitmap;
1822 	u32 channel_id;
1823 
1824 	mutex_lock(&gsi->mutex);
1825 
1826 	while (mask) {
1827 		channel_id = __fls(mask);
1828 
1829 		mask ^= BIT(channel_id);
1830 
1831 		gsi_modem_channel_halt(gsi, channel_id);
1832 	}
1833 
1834 	channel_id = gsi->channel_count - 1;
1835 	do
1836 		gsi_channel_teardown_one(gsi, channel_id);
1837 	while (channel_id--);
1838 
1839 	mutex_unlock(&gsi->mutex);
1840 
1841 	gsi_irq_disable(gsi);
1842 }
1843 
1844 /* Turn off all GSI interrupts initially */
1845 static int gsi_irq_setup(struct gsi *gsi)
1846 {
1847 	int ret;
1848 
1849 	/* Writing 1 indicates IRQ interrupts; 0 would be MSI */
1850 	iowrite32(1, gsi->virt + GSI_CNTXT_INTSET_OFFSET);
1851 
1852 	/* Disable all interrupt types */
1853 	gsi_irq_type_update(gsi, 0);
1854 
1855 	/* Clear all type-specific interrupt masks */
1856 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_MSK_OFFSET);
1857 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_MSK_OFFSET);
1858 	iowrite32(0, gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
1859 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_MSK_OFFSET);
1860 
1861 	/* The inter-EE interrupts are not supported for IPA v3.0-v3.1 */
1862 	if (gsi->version > IPA_VERSION_3_1) {
1863 		u32 offset;
1864 
1865 		/* These registers are in the non-adjusted address range */
1866 		offset = GSI_INTER_EE_SRC_CH_IRQ_MSK_OFFSET;
1867 		iowrite32(0, gsi->virt_raw + offset);
1868 		offset = GSI_INTER_EE_SRC_EV_CH_IRQ_MSK_OFFSET;
1869 		iowrite32(0, gsi->virt_raw + offset);
1870 	}
1871 
1872 	iowrite32(0, gsi->virt + GSI_CNTXT_GSI_IRQ_EN_OFFSET);
1873 
1874 	ret = request_irq(gsi->irq, gsi_isr, 0, "gsi", gsi);
1875 	if (ret)
1876 		dev_err(gsi->dev, "error %d requesting \"gsi\" IRQ\n", ret);
1877 
1878 	return ret;
1879 }
1880 
1881 static void gsi_irq_teardown(struct gsi *gsi)
1882 {
1883 	free_irq(gsi->irq, gsi);
1884 }
1885 
1886 /* Get # supported channel and event rings; there is no gsi_ring_teardown() */
1887 static int gsi_ring_setup(struct gsi *gsi)
1888 {
1889 	struct device *dev = gsi->dev;
1890 	u32 count;
1891 	u32 val;
1892 
1893 	if (gsi->version < IPA_VERSION_3_5_1) {
1894 		/* No HW_PARAM_2 register prior to IPA v3.5.1, assume the max */
1895 		gsi->channel_count = GSI_CHANNEL_COUNT_MAX;
1896 		gsi->evt_ring_count = GSI_EVT_RING_COUNT_MAX;
1897 
1898 		return 0;
1899 	}
1900 
1901 	val = ioread32(gsi->virt + GSI_GSI_HW_PARAM_2_OFFSET);
1902 
1903 	count = u32_get_bits(val, NUM_CH_PER_EE_FMASK);
1904 	if (!count) {
1905 		dev_err(dev, "GSI reports zero channels supported\n");
1906 		return -EINVAL;
1907 	}
1908 	if (count > GSI_CHANNEL_COUNT_MAX) {
1909 		dev_warn(dev, "limiting to %u channels; hardware supports %u\n",
1910 			 GSI_CHANNEL_COUNT_MAX, count);
1911 		count = GSI_CHANNEL_COUNT_MAX;
1912 	}
1913 	gsi->channel_count = count;
1914 
1915 	count = u32_get_bits(val, NUM_EV_PER_EE_FMASK);
1916 	if (!count) {
1917 		dev_err(dev, "GSI reports zero event rings supported\n");
1918 		return -EINVAL;
1919 	}
1920 	if (count > GSI_EVT_RING_COUNT_MAX) {
1921 		dev_warn(dev,
1922 			 "limiting to %u event rings; hardware supports %u\n",
1923 			 GSI_EVT_RING_COUNT_MAX, count);
1924 		count = GSI_EVT_RING_COUNT_MAX;
1925 	}
1926 	gsi->evt_ring_count = count;
1927 
1928 	return 0;
1929 }
1930 
1931 /* Setup function for GSI.  GSI firmware must be loaded and initialized */
1932 int gsi_setup(struct gsi *gsi)
1933 {
1934 	u32 val;
1935 	int ret;
1936 
1937 	/* Here is where we first touch the GSI hardware */
1938 	val = ioread32(gsi->virt + GSI_GSI_STATUS_OFFSET);
1939 	if (!(val & ENABLED_FMASK)) {
1940 		dev_err(gsi->dev, "GSI has not been enabled\n");
1941 		return -EIO;
1942 	}
1943 
1944 	ret = gsi_irq_setup(gsi);
1945 	if (ret)
1946 		return ret;
1947 
1948 	ret = gsi_ring_setup(gsi);	/* No matching teardown required */
1949 	if (ret)
1950 		goto err_irq_teardown;
1951 
1952 	/* Initialize the error log */
1953 	iowrite32(0, gsi->virt + GSI_ERROR_LOG_OFFSET);
1954 
1955 	ret = gsi_channel_setup(gsi);
1956 	if (ret)
1957 		goto err_irq_teardown;
1958 
1959 	return 0;
1960 
1961 err_irq_teardown:
1962 	gsi_irq_teardown(gsi);
1963 
1964 	return ret;
1965 }
1966 
1967 /* Inverse of gsi_setup() */
1968 void gsi_teardown(struct gsi *gsi)
1969 {
1970 	gsi_channel_teardown(gsi);
1971 	gsi_irq_teardown(gsi);
1972 }
1973 
1974 /* Initialize a channel's event ring */
1975 static int gsi_channel_evt_ring_init(struct gsi_channel *channel)
1976 {
1977 	struct gsi *gsi = channel->gsi;
1978 	struct gsi_evt_ring *evt_ring;
1979 	int ret;
1980 
1981 	ret = gsi_evt_ring_id_alloc(gsi);
1982 	if (ret < 0)
1983 		return ret;
1984 	channel->evt_ring_id = ret;
1985 
1986 	evt_ring = &gsi->evt_ring[channel->evt_ring_id];
1987 	evt_ring->channel = channel;
1988 
1989 	ret = gsi_ring_alloc(gsi, &evt_ring->ring, channel->event_count);
1990 	if (!ret)
1991 		return 0;	/* Success! */
1992 
1993 	dev_err(gsi->dev, "error %d allocating channel %u event ring\n",
1994 		ret, gsi_channel_id(channel));
1995 
1996 	gsi_evt_ring_id_free(gsi, channel->evt_ring_id);
1997 
1998 	return ret;
1999 }
2000 
2001 /* Inverse of gsi_channel_evt_ring_init() */
2002 static void gsi_channel_evt_ring_exit(struct gsi_channel *channel)
2003 {
2004 	u32 evt_ring_id = channel->evt_ring_id;
2005 	struct gsi *gsi = channel->gsi;
2006 	struct gsi_evt_ring *evt_ring;
2007 
2008 	evt_ring = &gsi->evt_ring[evt_ring_id];
2009 	gsi_ring_free(gsi, &evt_ring->ring);
2010 	gsi_evt_ring_id_free(gsi, evt_ring_id);
2011 }
2012 
2013 static bool gsi_channel_data_valid(struct gsi *gsi, bool command,
2014 				   const struct ipa_gsi_endpoint_data *data)
2015 {
2016 	const struct gsi_channel_data *channel_data;
2017 	u32 channel_id = data->channel_id;
2018 	struct device *dev = gsi->dev;
2019 
2020 	/* Make sure channel ids are in the range driver supports */
2021 	if (channel_id >= GSI_CHANNEL_COUNT_MAX) {
2022 		dev_err(dev, "bad channel id %u; must be less than %u\n",
2023 			channel_id, GSI_CHANNEL_COUNT_MAX);
2024 		return false;
2025 	}
2026 
2027 	if (data->ee_id != GSI_EE_AP && data->ee_id != GSI_EE_MODEM) {
2028 		dev_err(dev, "bad EE id %u; not AP or modem\n", data->ee_id);
2029 		return false;
2030 	}
2031 
2032 	if (command && !data->toward_ipa) {
2033 		dev_err(dev, "command channel %u is not TX\n", channel_id);
2034 		return false;
2035 	}
2036 
2037 	channel_data = &data->channel;
2038 
2039 	if (!channel_data->tlv_count ||
2040 	    channel_data->tlv_count > GSI_TLV_MAX) {
2041 		dev_err(dev, "channel %u bad tlv_count %u; must be 1..%u\n",
2042 			channel_id, channel_data->tlv_count, GSI_TLV_MAX);
2043 		return false;
2044 	}
2045 
2046 	if (command && IPA_COMMAND_TRANS_TRE_MAX > channel_data->tlv_count) {
2047 		dev_err(dev, "command TRE max too big for channel %u (%u > %u)\n",
2048 			channel_id, IPA_COMMAND_TRANS_TRE_MAX,
2049 			channel_data->tlv_count);
2050 		return false;
2051 	}
2052 
2053 	/* We have to allow at least one maximally-sized transaction to
2054 	 * be outstanding (which would use tlv_count TREs).  Given how
2055 	 * gsi_channel_tre_max() is computed, tre_count has to be almost
2056 	 * twice the TLV FIFO size to satisfy this requirement.
2057 	 */
2058 	if (channel_data->tre_count < 2 * channel_data->tlv_count - 1) {
2059 		dev_err(dev, "channel %u TLV count %u exceeds TRE count %u\n",
2060 			channel_id, channel_data->tlv_count,
2061 			channel_data->tre_count);
2062 		return false;
2063 	}
2064 
2065 	if (!is_power_of_2(channel_data->tre_count)) {
2066 		dev_err(dev, "channel %u bad tre_count %u; not power of 2\n",
2067 			channel_id, channel_data->tre_count);
2068 		return false;
2069 	}
2070 
2071 	if (!is_power_of_2(channel_data->event_count)) {
2072 		dev_err(dev, "channel %u bad event_count %u; not power of 2\n",
2073 			channel_id, channel_data->event_count);
2074 		return false;
2075 	}
2076 
2077 	return true;
2078 }
2079 
2080 /* Init function for a single channel */
2081 static int gsi_channel_init_one(struct gsi *gsi,
2082 				const struct ipa_gsi_endpoint_data *data,
2083 				bool command)
2084 {
2085 	struct gsi_channel *channel;
2086 	u32 tre_count;
2087 	int ret;
2088 
2089 	if (!gsi_channel_data_valid(gsi, command, data))
2090 		return -EINVAL;
2091 
2092 	/* Worst case we need an event for every outstanding TRE */
2093 	if (data->channel.tre_count > data->channel.event_count) {
2094 		tre_count = data->channel.event_count;
2095 		dev_warn(gsi->dev, "channel %u limited to %u TREs\n",
2096 			 data->channel_id, tre_count);
2097 	} else {
2098 		tre_count = data->channel.tre_count;
2099 	}
2100 
2101 	channel = &gsi->channel[data->channel_id];
2102 	memset(channel, 0, sizeof(*channel));
2103 
2104 	channel->gsi = gsi;
2105 	channel->toward_ipa = data->toward_ipa;
2106 	channel->command = command;
2107 	channel->trans_tre_max = data->channel.tlv_count;
2108 	channel->tre_count = tre_count;
2109 	channel->event_count = data->channel.event_count;
2110 
2111 	ret = gsi_channel_evt_ring_init(channel);
2112 	if (ret)
2113 		goto err_clear_gsi;
2114 
2115 	ret = gsi_ring_alloc(gsi, &channel->tre_ring, data->channel.tre_count);
2116 	if (ret) {
2117 		dev_err(gsi->dev, "error %d allocating channel %u ring\n",
2118 			ret, data->channel_id);
2119 		goto err_channel_evt_ring_exit;
2120 	}
2121 
2122 	ret = gsi_channel_trans_init(gsi, data->channel_id);
2123 	if (ret)
2124 		goto err_ring_free;
2125 
2126 	if (command) {
2127 		u32 tre_max = gsi_channel_tre_max(gsi, data->channel_id);
2128 
2129 		ret = ipa_cmd_pool_init(channel, tre_max);
2130 	}
2131 	if (!ret)
2132 		return 0;	/* Success! */
2133 
2134 	gsi_channel_trans_exit(channel);
2135 err_ring_free:
2136 	gsi_ring_free(gsi, &channel->tre_ring);
2137 err_channel_evt_ring_exit:
2138 	gsi_channel_evt_ring_exit(channel);
2139 err_clear_gsi:
2140 	channel->gsi = NULL;	/* Mark it not (fully) initialized */
2141 
2142 	return ret;
2143 }
2144 
2145 /* Inverse of gsi_channel_init_one() */
2146 static void gsi_channel_exit_one(struct gsi_channel *channel)
2147 {
2148 	if (!gsi_channel_initialized(channel))
2149 		return;
2150 
2151 	if (channel->command)
2152 		ipa_cmd_pool_exit(channel);
2153 	gsi_channel_trans_exit(channel);
2154 	gsi_ring_free(channel->gsi, &channel->tre_ring);
2155 	gsi_channel_evt_ring_exit(channel);
2156 }
2157 
2158 /* Init function for channels */
2159 static int gsi_channel_init(struct gsi *gsi, u32 count,
2160 			    const struct ipa_gsi_endpoint_data *data)
2161 {
2162 	bool modem_alloc;
2163 	int ret = 0;
2164 	u32 i;
2165 
2166 	/* IPA v4.2 requires the AP to allocate channels for the modem */
2167 	modem_alloc = gsi->version == IPA_VERSION_4_2;
2168 
2169 	gsi->event_bitmap = gsi_event_bitmap_init(GSI_EVT_RING_COUNT_MAX);
2170 	gsi->ieob_enabled_bitmap = 0;
2171 
2172 	/* The endpoint data array is indexed by endpoint name */
2173 	for (i = 0; i < count; i++) {
2174 		bool command = i == IPA_ENDPOINT_AP_COMMAND_TX;
2175 
2176 		if (ipa_gsi_endpoint_data_empty(&data[i]))
2177 			continue;	/* Skip over empty slots */
2178 
2179 		/* Mark modem channels to be allocated (hardware workaround) */
2180 		if (data[i].ee_id == GSI_EE_MODEM) {
2181 			if (modem_alloc)
2182 				gsi->modem_channel_bitmap |=
2183 						BIT(data[i].channel_id);
2184 			continue;
2185 		}
2186 
2187 		ret = gsi_channel_init_one(gsi, &data[i], command);
2188 		if (ret)
2189 			goto err_unwind;
2190 	}
2191 
2192 	return ret;
2193 
2194 err_unwind:
2195 	while (i--) {
2196 		if (ipa_gsi_endpoint_data_empty(&data[i]))
2197 			continue;
2198 		if (modem_alloc && data[i].ee_id == GSI_EE_MODEM) {
2199 			gsi->modem_channel_bitmap &= ~BIT(data[i].channel_id);
2200 			continue;
2201 		}
2202 		gsi_channel_exit_one(&gsi->channel[data->channel_id]);
2203 	}
2204 
2205 	return ret;
2206 }
2207 
2208 /* Inverse of gsi_channel_init() */
2209 static void gsi_channel_exit(struct gsi *gsi)
2210 {
2211 	u32 channel_id = GSI_CHANNEL_COUNT_MAX - 1;
2212 
2213 	do
2214 		gsi_channel_exit_one(&gsi->channel[channel_id]);
2215 	while (channel_id--);
2216 	gsi->modem_channel_bitmap = 0;
2217 }
2218 
2219 /* Init function for GSI.  GSI hardware does not need to be "ready" */
2220 int gsi_init(struct gsi *gsi, struct platform_device *pdev,
2221 	     enum ipa_version version, u32 count,
2222 	     const struct ipa_gsi_endpoint_data *data)
2223 {
2224 	struct device *dev = &pdev->dev;
2225 	struct resource *res;
2226 	resource_size_t size;
2227 	u32 adjust;
2228 	int ret;
2229 
2230 	gsi_validate_build();
2231 
2232 	gsi->dev = dev;
2233 	gsi->version = version;
2234 
2235 	/* GSI uses NAPI on all channels.  Create a dummy network device
2236 	 * for the channel NAPI contexts to be associated with.
2237 	 */
2238 	init_dummy_netdev(&gsi->dummy_dev);
2239 
2240 	/* Get GSI memory range and map it */
2241 	res = platform_get_resource_byname(pdev, IORESOURCE_MEM, "gsi");
2242 	if (!res) {
2243 		dev_err(dev, "DT error getting \"gsi\" memory property\n");
2244 		return -ENODEV;
2245 	}
2246 
2247 	size = resource_size(res);
2248 	if (res->start > U32_MAX || size > U32_MAX - res->start) {
2249 		dev_err(dev, "DT memory resource \"gsi\" out of range\n");
2250 		return -EINVAL;
2251 	}
2252 
2253 	/* Make sure we can make our pointer adjustment if necessary */
2254 	adjust = gsi->version < IPA_VERSION_4_5 ? 0 : GSI_EE_REG_ADJUST;
2255 	if (res->start < adjust) {
2256 		dev_err(dev, "DT memory resource \"gsi\" too low (< %u)\n",
2257 			adjust);
2258 		return -EINVAL;
2259 	}
2260 
2261 	gsi->virt_raw = ioremap(res->start, size);
2262 	if (!gsi->virt_raw) {
2263 		dev_err(dev, "unable to remap \"gsi\" memory\n");
2264 		return -ENOMEM;
2265 	}
2266 	/* Most registers are accessed using an adjusted register range */
2267 	gsi->virt = gsi->virt_raw - adjust;
2268 
2269 	init_completion(&gsi->completion);
2270 
2271 	ret = gsi_irq_init(gsi, pdev);	/* No matching exit required */
2272 	if (ret)
2273 		goto err_iounmap;
2274 
2275 	ret = gsi_channel_init(gsi, count, data);
2276 	if (ret)
2277 		goto err_iounmap;
2278 
2279 	mutex_init(&gsi->mutex);
2280 
2281 	return 0;
2282 
2283 err_iounmap:
2284 	iounmap(gsi->virt_raw);
2285 
2286 	return ret;
2287 }
2288 
2289 /* Inverse of gsi_init() */
2290 void gsi_exit(struct gsi *gsi)
2291 {
2292 	mutex_destroy(&gsi->mutex);
2293 	gsi_channel_exit(gsi);
2294 	iounmap(gsi->virt_raw);
2295 }
2296 
2297 /* The maximum number of outstanding TREs on a channel.  This limits
2298  * a channel's maximum number of transactions outstanding (worst case
2299  * is one TRE per transaction).
2300  *
2301  * The absolute limit is the number of TREs in the channel's TRE ring,
2302  * and in theory we should be able use all of them.  But in practice,
2303  * doing that led to the hardware reporting exhaustion of event ring
2304  * slots for writing completion information.  So the hardware limit
2305  * would be (tre_count - 1).
2306  *
2307  * We reduce it a bit further though.  Transaction resource pools are
2308  * sized to be a little larger than this maximum, to allow resource
2309  * allocations to always be contiguous.  The number of entries in a
2310  * TRE ring buffer is a power of 2, and the extra resources in a pool
2311  * tends to nearly double the memory allocated for it.  Reducing the
2312  * maximum number of outstanding TREs allows the number of entries in
2313  * a pool to avoid crossing that power-of-2 boundary, and this can
2314  * substantially reduce pool memory requirements.  The number we
2315  * reduce it by matches the number added in gsi_trans_pool_init().
2316  */
2317 u32 gsi_channel_tre_max(struct gsi *gsi, u32 channel_id)
2318 {
2319 	struct gsi_channel *channel = &gsi->channel[channel_id];
2320 
2321 	/* Hardware limit is channel->tre_count - 1 */
2322 	return channel->tre_count - (channel->trans_tre_max - 1);
2323 }
2324