xref: /linux/drivers/net/ipa/gsi.c (revision 5c8166419acf468b5bc3e48f928a040485d3e0c2)
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 	size_t size = evt_ring->ring.count * GSI_RING_ELEMENT_SIZE;
669 	u32 val;
670 
671 	/* We program all event rings as GPI type/protocol */
672 	val = u32_encode_bits(GSI_CHANNEL_TYPE_GPI, EV_CHTYPE_FMASK);
673 	val |= EV_INTYPE_FMASK;
674 	val |= u32_encode_bits(GSI_RING_ELEMENT_SIZE, EV_ELEMENT_SIZE_FMASK);
675 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_0_OFFSET(evt_ring_id));
676 
677 	val = ev_r_length_encoded(gsi->version, size);
678 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_1_OFFSET(evt_ring_id));
679 
680 	/* The context 2 and 3 registers store the low-order and
681 	 * high-order 32 bits of the address of the event ring,
682 	 * respectively.
683 	 */
684 	val = lower_32_bits(evt_ring->ring.addr);
685 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_2_OFFSET(evt_ring_id));
686 	val = upper_32_bits(evt_ring->ring.addr);
687 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_3_OFFSET(evt_ring_id));
688 
689 	/* Enable interrupt moderation by setting the moderation delay */
690 	val = u32_encode_bits(GSI_EVT_RING_INT_MODT, MODT_FMASK);
691 	val |= u32_encode_bits(1, MODC_FMASK);	/* comes from channel */
692 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_8_OFFSET(evt_ring_id));
693 
694 	/* No MSI write data, and MSI address high and low address is 0 */
695 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_9_OFFSET(evt_ring_id));
696 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_10_OFFSET(evt_ring_id));
697 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_11_OFFSET(evt_ring_id));
698 
699 	/* We don't need to get event read pointer updates */
700 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_12_OFFSET(evt_ring_id));
701 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_13_OFFSET(evt_ring_id));
702 
703 	/* Finally, tell the hardware we've completed event 0 (arbitrary) */
704 	gsi_evt_ring_doorbell(gsi, evt_ring_id, 0);
705 }
706 
707 /* Find the transaction whose completion indicates a channel is quiesced */
708 static struct gsi_trans *gsi_channel_trans_last(struct gsi_channel *channel)
709 {
710 	struct gsi_trans_info *trans_info = &channel->trans_info;
711 	const struct list_head *list;
712 	struct gsi_trans *trans;
713 
714 	spin_lock_bh(&trans_info->spinlock);
715 
716 	/* There is a small chance a TX transaction got allocated just
717 	 * before we disabled transmits, so check for that.
718 	 */
719 	if (channel->toward_ipa) {
720 		list = &trans_info->alloc;
721 		if (!list_empty(list))
722 			goto done;
723 		list = &trans_info->pending;
724 		if (!list_empty(list))
725 			goto done;
726 	}
727 
728 	/* Otherwise (TX or RX) we want to wait for anything that
729 	 * has completed, or has been polled but not released yet.
730 	 */
731 	list = &trans_info->complete;
732 	if (!list_empty(list))
733 		goto done;
734 	list = &trans_info->polled;
735 	if (list_empty(list))
736 		list = NULL;
737 done:
738 	trans = list ? list_last_entry(list, struct gsi_trans, links) : NULL;
739 
740 	/* Caller will wait for this, so take a reference */
741 	if (trans)
742 		refcount_inc(&trans->refcount);
743 
744 	spin_unlock_bh(&trans_info->spinlock);
745 
746 	return trans;
747 }
748 
749 /* Wait for transaction activity on a channel to complete */
750 static void gsi_channel_trans_quiesce(struct gsi_channel *channel)
751 {
752 	struct gsi_trans *trans;
753 
754 	/* Get the last transaction, and wait for it to complete */
755 	trans = gsi_channel_trans_last(channel);
756 	if (trans) {
757 		wait_for_completion(&trans->completion);
758 		gsi_trans_free(trans);
759 	}
760 }
761 
762 /* Program a channel for use; there is no gsi_channel_deprogram() */
763 static void gsi_channel_program(struct gsi_channel *channel, bool doorbell)
764 {
765 	size_t size = channel->tre_ring.count * GSI_RING_ELEMENT_SIZE;
766 	u32 channel_id = gsi_channel_id(channel);
767 	union gsi_channel_scratch scr = { };
768 	struct gsi_channel_scratch_gpi *gpi;
769 	struct gsi *gsi = channel->gsi;
770 	u32 wrr_weight = 0;
771 	u32 val;
772 
773 	/* Arbitrarily pick TRE 0 as the first channel element to use */
774 	channel->tre_ring.index = 0;
775 
776 	/* We program all channels as GPI type/protocol */
777 	val = chtype_protocol_encoded(gsi->version, GSI_CHANNEL_TYPE_GPI);
778 	if (channel->toward_ipa)
779 		val |= CHTYPE_DIR_FMASK;
780 	val |= u32_encode_bits(channel->evt_ring_id, ERINDEX_FMASK);
781 	val |= u32_encode_bits(GSI_RING_ELEMENT_SIZE, ELEMENT_SIZE_FMASK);
782 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_0_OFFSET(channel_id));
783 
784 	val = r_length_encoded(gsi->version, size);
785 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_1_OFFSET(channel_id));
786 
787 	/* The context 2 and 3 registers store the low-order and
788 	 * high-order 32 bits of the address of the channel ring,
789 	 * respectively.
790 	 */
791 	val = lower_32_bits(channel->tre_ring.addr);
792 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_2_OFFSET(channel_id));
793 	val = upper_32_bits(channel->tre_ring.addr);
794 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_3_OFFSET(channel_id));
795 
796 	/* Command channel gets low weighted round-robin priority */
797 	if (channel->command)
798 		wrr_weight = field_max(WRR_WEIGHT_FMASK);
799 	val = u32_encode_bits(wrr_weight, WRR_WEIGHT_FMASK);
800 
801 	/* Max prefetch is 1 segment (do not set MAX_PREFETCH_FMASK) */
802 
803 	/* No need to use the doorbell engine starting at IPA v4.0 */
804 	if (gsi->version < IPA_VERSION_4_0 && doorbell)
805 		val |= USE_DB_ENG_FMASK;
806 
807 	/* v4.0 introduces an escape buffer for prefetch.  We use it
808 	 * on all but the AP command channel.
809 	 */
810 	if (gsi->version >= IPA_VERSION_4_0 && !channel->command) {
811 		/* If not otherwise set, prefetch buffers are used */
812 		if (gsi->version < IPA_VERSION_4_5)
813 			val |= USE_ESCAPE_BUF_ONLY_FMASK;
814 		else
815 			val |= u32_encode_bits(GSI_ESCAPE_BUF_ONLY,
816 					       PREFETCH_MODE_FMASK);
817 	}
818 	/* All channels set DB_IN_BYTES */
819 	if (gsi->version >= IPA_VERSION_4_9)
820 		val |= DB_IN_BYTES;
821 
822 	iowrite32(val, gsi->virt + GSI_CH_C_QOS_OFFSET(channel_id));
823 
824 	/* Now update the scratch registers for GPI protocol */
825 	gpi = &scr.gpi;
826 	gpi->max_outstanding_tre = gsi_channel_trans_tre_max(gsi, channel_id) *
827 					GSI_RING_ELEMENT_SIZE;
828 	gpi->outstanding_threshold = 2 * GSI_RING_ELEMENT_SIZE;
829 
830 	val = scr.data.word1;
831 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_0_OFFSET(channel_id));
832 
833 	val = scr.data.word2;
834 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_1_OFFSET(channel_id));
835 
836 	val = scr.data.word3;
837 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_2_OFFSET(channel_id));
838 
839 	/* We must preserve the upper 16 bits of the last scratch register.
840 	 * The next sequence assumes those bits remain unchanged between the
841 	 * read and the write.
842 	 */
843 	val = ioread32(gsi->virt + GSI_CH_C_SCRATCH_3_OFFSET(channel_id));
844 	val = (scr.data.word4 & GENMASK(31, 16)) | (val & GENMASK(15, 0));
845 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_3_OFFSET(channel_id));
846 
847 	/* All done! */
848 }
849 
850 static int __gsi_channel_start(struct gsi_channel *channel, bool resume)
851 {
852 	struct gsi *gsi = channel->gsi;
853 	int ret;
854 
855 	/* Prior to IPA v4.0 suspend/resume is not implemented by GSI */
856 	if (resume && gsi->version < IPA_VERSION_4_0)
857 		return 0;
858 
859 	mutex_lock(&gsi->mutex);
860 
861 	ret = gsi_channel_start_command(channel);
862 
863 	mutex_unlock(&gsi->mutex);
864 
865 	return ret;
866 }
867 
868 /* Start an allocated GSI channel */
869 int gsi_channel_start(struct gsi *gsi, u32 channel_id)
870 {
871 	struct gsi_channel *channel = &gsi->channel[channel_id];
872 	int ret;
873 
874 	/* Enable NAPI and the completion interrupt */
875 	napi_enable(&channel->napi);
876 	gsi_irq_ieob_enable_one(gsi, channel->evt_ring_id);
877 
878 	ret = __gsi_channel_start(channel, false);
879 	if (ret) {
880 		gsi_irq_ieob_disable_one(gsi, channel->evt_ring_id);
881 		napi_disable(&channel->napi);
882 	}
883 
884 	return ret;
885 }
886 
887 static int gsi_channel_stop_retry(struct gsi_channel *channel)
888 {
889 	u32 retries = GSI_CHANNEL_STOP_RETRIES;
890 	int ret;
891 
892 	do {
893 		ret = gsi_channel_stop_command(channel);
894 		if (ret != -EAGAIN)
895 			break;
896 		usleep_range(3 * USEC_PER_MSEC, 5 * USEC_PER_MSEC);
897 	} while (retries--);
898 
899 	return ret;
900 }
901 
902 static int __gsi_channel_stop(struct gsi_channel *channel, bool suspend)
903 {
904 	struct gsi *gsi = channel->gsi;
905 	int ret;
906 
907 	/* Wait for any underway transactions to complete before stopping. */
908 	gsi_channel_trans_quiesce(channel);
909 
910 	/* Prior to IPA v4.0 suspend/resume is not implemented by GSI */
911 	if (suspend && gsi->version < IPA_VERSION_4_0)
912 		return 0;
913 
914 	mutex_lock(&gsi->mutex);
915 
916 	ret = gsi_channel_stop_retry(channel);
917 
918 	mutex_unlock(&gsi->mutex);
919 
920 	return ret;
921 }
922 
923 /* Stop a started channel */
924 int gsi_channel_stop(struct gsi *gsi, u32 channel_id)
925 {
926 	struct gsi_channel *channel = &gsi->channel[channel_id];
927 	int ret;
928 
929 	ret = __gsi_channel_stop(channel, false);
930 	if (ret)
931 		return ret;
932 
933 	/* Disable the completion interrupt and NAPI if successful */
934 	gsi_irq_ieob_disable_one(gsi, channel->evt_ring_id);
935 	napi_disable(&channel->napi);
936 
937 	return 0;
938 }
939 
940 /* Reset and reconfigure a channel, (possibly) enabling the doorbell engine */
941 void gsi_channel_reset(struct gsi *gsi, u32 channel_id, bool doorbell)
942 {
943 	struct gsi_channel *channel = &gsi->channel[channel_id];
944 
945 	mutex_lock(&gsi->mutex);
946 
947 	gsi_channel_reset_command(channel);
948 	/* Due to a hardware quirk we may need to reset RX channels twice. */
949 	if (gsi->version < IPA_VERSION_4_0 && !channel->toward_ipa)
950 		gsi_channel_reset_command(channel);
951 
952 	gsi_channel_program(channel, doorbell);
953 	gsi_channel_trans_cancel_pending(channel);
954 
955 	mutex_unlock(&gsi->mutex);
956 }
957 
958 /* Stop a started channel for suspend */
959 int gsi_channel_suspend(struct gsi *gsi, u32 channel_id)
960 {
961 	struct gsi_channel *channel = &gsi->channel[channel_id];
962 	int ret;
963 
964 	ret = __gsi_channel_stop(channel, true);
965 	if (ret)
966 		return ret;
967 
968 	/* Ensure NAPI polling has finished. */
969 	napi_synchronize(&channel->napi);
970 
971 	return 0;
972 }
973 
974 /* Resume a suspended channel (starting if stopped) */
975 int gsi_channel_resume(struct gsi *gsi, u32 channel_id)
976 {
977 	struct gsi_channel *channel = &gsi->channel[channel_id];
978 
979 	return __gsi_channel_start(channel, true);
980 }
981 
982 /* Prevent all GSI interrupts while suspended */
983 void gsi_suspend(struct gsi *gsi)
984 {
985 	disable_irq(gsi->irq);
986 }
987 
988 /* Allow all GSI interrupts again when resuming */
989 void gsi_resume(struct gsi *gsi)
990 {
991 	enable_irq(gsi->irq);
992 }
993 
994 /**
995  * gsi_channel_tx_queued() - Report queued TX transfers for a channel
996  * @channel:	Channel for which to report
997  *
998  * Report to the network stack the number of bytes and transactions that
999  * have been queued to hardware since last call.  This and the next function
1000  * supply information used by the network stack for throttling.
1001  *
1002  * For each channel we track the number of transactions used and bytes of
1003  * data those transactions represent.  We also track what those values are
1004  * each time this function is called.  Subtracting the two tells us
1005  * the number of bytes and transactions that have been added between
1006  * successive calls.
1007  *
1008  * Calling this each time we ring the channel doorbell allows us to
1009  * provide accurate information to the network stack about how much
1010  * work we've given the hardware at any point in time.
1011  */
1012 void gsi_channel_tx_queued(struct gsi_channel *channel)
1013 {
1014 	u32 trans_count;
1015 	u32 byte_count;
1016 
1017 	byte_count = channel->byte_count - channel->queued_byte_count;
1018 	trans_count = channel->trans_count - channel->queued_trans_count;
1019 	channel->queued_byte_count = channel->byte_count;
1020 	channel->queued_trans_count = channel->trans_count;
1021 
1022 	ipa_gsi_channel_tx_queued(channel->gsi, gsi_channel_id(channel),
1023 				  trans_count, byte_count);
1024 }
1025 
1026 /**
1027  * gsi_channel_tx_update() - Report completed TX transfers
1028  * @channel:	Channel that has completed transmitting packets
1029  * @trans:	Last transation known to be complete
1030  *
1031  * Compute the number of transactions and bytes that have been transferred
1032  * over a TX channel since the given transaction was committed.  Report this
1033  * information to the network stack.
1034  *
1035  * At the time a transaction is committed, we record its channel's
1036  * committed transaction and byte counts *in the transaction*.
1037  * Completions are signaled by the hardware with an interrupt, and
1038  * we can determine the latest completed transaction at that time.
1039  *
1040  * The difference between the byte/transaction count recorded in
1041  * the transaction and the count last time we recorded a completion
1042  * tells us exactly how much data has been transferred between
1043  * completions.
1044  *
1045  * Calling this each time we learn of a newly-completed transaction
1046  * allows us to provide accurate information to the network stack
1047  * about how much work has been completed by the hardware at a given
1048  * point in time.
1049  */
1050 static void
1051 gsi_channel_tx_update(struct gsi_channel *channel, struct gsi_trans *trans)
1052 {
1053 	u64 byte_count = trans->byte_count + trans->len;
1054 	u64 trans_count = trans->trans_count + 1;
1055 
1056 	byte_count -= channel->compl_byte_count;
1057 	channel->compl_byte_count += byte_count;
1058 	trans_count -= channel->compl_trans_count;
1059 	channel->compl_trans_count += trans_count;
1060 
1061 	ipa_gsi_channel_tx_completed(channel->gsi, gsi_channel_id(channel),
1062 				     trans_count, byte_count);
1063 }
1064 
1065 /* Channel control interrupt handler */
1066 static void gsi_isr_chan_ctrl(struct gsi *gsi)
1067 {
1068 	u32 channel_mask;
1069 
1070 	channel_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_CH_IRQ_OFFSET);
1071 	iowrite32(channel_mask, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_CLR_OFFSET);
1072 
1073 	while (channel_mask) {
1074 		u32 channel_id = __ffs(channel_mask);
1075 
1076 		channel_mask ^= BIT(channel_id);
1077 
1078 		complete(&gsi->completion);
1079 	}
1080 }
1081 
1082 /* Event ring control interrupt handler */
1083 static void gsi_isr_evt_ctrl(struct gsi *gsi)
1084 {
1085 	u32 event_mask;
1086 
1087 	event_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_OFFSET);
1088 	iowrite32(event_mask, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_CLR_OFFSET);
1089 
1090 	while (event_mask) {
1091 		u32 evt_ring_id = __ffs(event_mask);
1092 
1093 		event_mask ^= BIT(evt_ring_id);
1094 
1095 		complete(&gsi->completion);
1096 	}
1097 }
1098 
1099 /* Global channel error interrupt handler */
1100 static void
1101 gsi_isr_glob_chan_err(struct gsi *gsi, u32 err_ee, u32 channel_id, u32 code)
1102 {
1103 	if (code == GSI_OUT_OF_RESOURCES) {
1104 		dev_err(gsi->dev, "channel %u out of resources\n", channel_id);
1105 		complete(&gsi->completion);
1106 		return;
1107 	}
1108 
1109 	/* Report, but otherwise ignore all other error codes */
1110 	dev_err(gsi->dev, "channel %u global error ee 0x%08x code 0x%08x\n",
1111 		channel_id, err_ee, code);
1112 }
1113 
1114 /* Global event error interrupt handler */
1115 static void
1116 gsi_isr_glob_evt_err(struct gsi *gsi, u32 err_ee, u32 evt_ring_id, u32 code)
1117 {
1118 	if (code == GSI_OUT_OF_RESOURCES) {
1119 		struct gsi_evt_ring *evt_ring = &gsi->evt_ring[evt_ring_id];
1120 		u32 channel_id = gsi_channel_id(evt_ring->channel);
1121 
1122 		complete(&gsi->completion);
1123 		dev_err(gsi->dev, "evt_ring for channel %u out of resources\n",
1124 			channel_id);
1125 		return;
1126 	}
1127 
1128 	/* Report, but otherwise ignore all other error codes */
1129 	dev_err(gsi->dev, "event ring %u global error ee %u code 0x%08x\n",
1130 		evt_ring_id, err_ee, code);
1131 }
1132 
1133 /* Global error interrupt handler */
1134 static void gsi_isr_glob_err(struct gsi *gsi)
1135 {
1136 	enum gsi_err_type type;
1137 	enum gsi_err_code code;
1138 	u32 which;
1139 	u32 val;
1140 	u32 ee;
1141 
1142 	/* Get the logged error, then reinitialize the log */
1143 	val = ioread32(gsi->virt + GSI_ERROR_LOG_OFFSET);
1144 	iowrite32(0, gsi->virt + GSI_ERROR_LOG_OFFSET);
1145 	iowrite32(~0, gsi->virt + GSI_ERROR_LOG_CLR_OFFSET);
1146 
1147 	ee = u32_get_bits(val, ERR_EE_FMASK);
1148 	type = u32_get_bits(val, ERR_TYPE_FMASK);
1149 	which = u32_get_bits(val, ERR_VIRT_IDX_FMASK);
1150 	code = u32_get_bits(val, ERR_CODE_FMASK);
1151 
1152 	if (type == GSI_ERR_TYPE_CHAN)
1153 		gsi_isr_glob_chan_err(gsi, ee, which, code);
1154 	else if (type == GSI_ERR_TYPE_EVT)
1155 		gsi_isr_glob_evt_err(gsi, ee, which, code);
1156 	else	/* type GSI_ERR_TYPE_GLOB should be fatal */
1157 		dev_err(gsi->dev, "unexpected global error 0x%08x\n", type);
1158 }
1159 
1160 /* Generic EE interrupt handler */
1161 static void gsi_isr_gp_int1(struct gsi *gsi)
1162 {
1163 	u32 result;
1164 	u32 val;
1165 
1166 	/* This interrupt is used to handle completions of GENERIC GSI
1167 	 * commands.  We use these to allocate and halt channels on the
1168 	 * modem's behalf due to a hardware quirk on IPA v4.2.  The modem
1169 	 * "owns" channels even when the AP allocates them, and have no
1170 	 * way of knowing whether a modem channel's state has been changed.
1171 	 *
1172 	 * We also use GENERIC commands to enable/disable channel flow
1173 	 * control for IPA v4.2+.
1174 	 *
1175 	 * It is recommended that we halt the modem channels we allocated
1176 	 * when shutting down, but it's possible the channel isn't running
1177 	 * at the time we issue the HALT command.  We'll get an error in
1178 	 * that case, but it's harmless (the channel is already halted).
1179 	 * Similarly, we could get an error back when updating flow control
1180 	 * on a channel because it's not in the proper state.
1181 	 *
1182 	 * In either case, we silently ignore a CHANNEL_NOT_RUNNING error
1183 	 * if we receive it.
1184 	 */
1185 	val = ioread32(gsi->virt + GSI_CNTXT_SCRATCH_0_OFFSET);
1186 	result = u32_get_bits(val, GENERIC_EE_RESULT_FMASK);
1187 
1188 	switch (result) {
1189 	case GENERIC_EE_SUCCESS:
1190 	case GENERIC_EE_CHANNEL_NOT_RUNNING:
1191 		gsi->result = 0;
1192 		break;
1193 
1194 	case GENERIC_EE_RETRY:
1195 		gsi->result = -EAGAIN;
1196 		break;
1197 
1198 	default:
1199 		dev_err(gsi->dev, "global INT1 generic result %u\n", result);
1200 		gsi->result = -EIO;
1201 		break;
1202 	}
1203 
1204 	complete(&gsi->completion);
1205 }
1206 
1207 /* Inter-EE interrupt handler */
1208 static void gsi_isr_glob_ee(struct gsi *gsi)
1209 {
1210 	u32 val;
1211 
1212 	val = ioread32(gsi->virt + GSI_CNTXT_GLOB_IRQ_STTS_OFFSET);
1213 
1214 	if (val & BIT(ERROR_INT))
1215 		gsi_isr_glob_err(gsi);
1216 
1217 	iowrite32(val, gsi->virt + GSI_CNTXT_GLOB_IRQ_CLR_OFFSET);
1218 
1219 	val &= ~BIT(ERROR_INT);
1220 
1221 	if (val & BIT(GP_INT1)) {
1222 		val ^= BIT(GP_INT1);
1223 		gsi_isr_gp_int1(gsi);
1224 	}
1225 
1226 	if (val)
1227 		dev_err(gsi->dev, "unexpected global interrupt 0x%08x\n", val);
1228 }
1229 
1230 /* I/O completion interrupt event */
1231 static void gsi_isr_ieob(struct gsi *gsi)
1232 {
1233 	u32 event_mask;
1234 
1235 	event_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_OFFSET);
1236 	gsi_irq_ieob_disable(gsi, event_mask);
1237 	iowrite32(event_mask, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_CLR_OFFSET);
1238 
1239 	while (event_mask) {
1240 		u32 evt_ring_id = __ffs(event_mask);
1241 
1242 		event_mask ^= BIT(evt_ring_id);
1243 
1244 		napi_schedule(&gsi->evt_ring[evt_ring_id].channel->napi);
1245 	}
1246 }
1247 
1248 /* General event interrupts represent serious problems, so report them */
1249 static void gsi_isr_general(struct gsi *gsi)
1250 {
1251 	struct device *dev = gsi->dev;
1252 	u32 val;
1253 
1254 	val = ioread32(gsi->virt + GSI_CNTXT_GSI_IRQ_STTS_OFFSET);
1255 	iowrite32(val, gsi->virt + GSI_CNTXT_GSI_IRQ_CLR_OFFSET);
1256 
1257 	dev_err(dev, "unexpected general interrupt 0x%08x\n", val);
1258 }
1259 
1260 /**
1261  * gsi_isr() - Top level GSI interrupt service routine
1262  * @irq:	Interrupt number (ignored)
1263  * @dev_id:	GSI pointer supplied to request_irq()
1264  *
1265  * This is the main handler function registered for the GSI IRQ. Each type
1266  * of interrupt has a separate handler function that is called from here.
1267  */
1268 static irqreturn_t gsi_isr(int irq, void *dev_id)
1269 {
1270 	struct gsi *gsi = dev_id;
1271 	u32 intr_mask;
1272 	u32 cnt = 0;
1273 
1274 	/* enum gsi_irq_type_id defines GSI interrupt types */
1275 	while ((intr_mask = ioread32(gsi->virt + GSI_CNTXT_TYPE_IRQ_OFFSET))) {
1276 		/* intr_mask contains bitmask of pending GSI interrupts */
1277 		do {
1278 			u32 gsi_intr = BIT(__ffs(intr_mask));
1279 
1280 			intr_mask ^= gsi_intr;
1281 
1282 			switch (gsi_intr) {
1283 			case BIT(GSI_CH_CTRL):
1284 				gsi_isr_chan_ctrl(gsi);
1285 				break;
1286 			case BIT(GSI_EV_CTRL):
1287 				gsi_isr_evt_ctrl(gsi);
1288 				break;
1289 			case BIT(GSI_GLOB_EE):
1290 				gsi_isr_glob_ee(gsi);
1291 				break;
1292 			case BIT(GSI_IEOB):
1293 				gsi_isr_ieob(gsi);
1294 				break;
1295 			case BIT(GSI_GENERAL):
1296 				gsi_isr_general(gsi);
1297 				break;
1298 			default:
1299 				dev_err(gsi->dev,
1300 					"unrecognized interrupt type 0x%08x\n",
1301 					gsi_intr);
1302 				break;
1303 			}
1304 		} while (intr_mask);
1305 
1306 		if (++cnt > GSI_ISR_MAX_ITER) {
1307 			dev_err(gsi->dev, "interrupt flood\n");
1308 			break;
1309 		}
1310 	}
1311 
1312 	return IRQ_HANDLED;
1313 }
1314 
1315 /* Init function for GSI IRQ lookup; there is no gsi_irq_exit() */
1316 static int gsi_irq_init(struct gsi *gsi, struct platform_device *pdev)
1317 {
1318 	int ret;
1319 
1320 	ret = platform_get_irq_byname(pdev, "gsi");
1321 	if (ret <= 0)
1322 		return ret ? : -EINVAL;
1323 
1324 	gsi->irq = ret;
1325 
1326 	return 0;
1327 }
1328 
1329 /* Return the transaction associated with a transfer completion event */
1330 static struct gsi_trans *gsi_event_trans(struct gsi_channel *channel,
1331 					 struct gsi_event *event)
1332 {
1333 	u32 tre_offset;
1334 	u32 tre_index;
1335 
1336 	/* Event xfer_ptr records the TRE it's associated with */
1337 	tre_offset = lower_32_bits(le64_to_cpu(event->xfer_ptr));
1338 	tre_index = gsi_ring_index(&channel->tre_ring, tre_offset);
1339 
1340 	return gsi_channel_trans_mapped(channel, tre_index);
1341 }
1342 
1343 /**
1344  * gsi_evt_ring_rx_update() - Record lengths of received data
1345  * @evt_ring:	Event ring associated with channel that received packets
1346  * @index:	Event index in ring reported by hardware
1347  *
1348  * Events for RX channels contain the actual number of bytes received into
1349  * the buffer.  Every event has a transaction associated with it, and here
1350  * we update transactions to record their actual received lengths.
1351  *
1352  * This function is called whenever we learn that the GSI hardware has filled
1353  * new events since the last time we checked.  The ring's index field tells
1354  * the first entry in need of processing.  The index provided is the
1355  * first *unfilled* event in the ring (following the last filled one).
1356  *
1357  * Events are sequential within the event ring, and transactions are
1358  * sequential within the transaction pool.
1359  *
1360  * Note that @index always refers to an element *within* the event ring.
1361  */
1362 static void gsi_evt_ring_rx_update(struct gsi_evt_ring *evt_ring, u32 index)
1363 {
1364 	struct gsi_channel *channel = evt_ring->channel;
1365 	struct gsi_ring *ring = &evt_ring->ring;
1366 	struct gsi_trans_info *trans_info;
1367 	struct gsi_event *event_done;
1368 	struct gsi_event *event;
1369 	struct gsi_trans *trans;
1370 	u32 byte_count = 0;
1371 	u32 old_index;
1372 	u32 event_avail;
1373 
1374 	trans_info = &channel->trans_info;
1375 
1376 	/* We'll start with the oldest un-processed event.  RX channels
1377 	 * replenish receive buffers in single-TRE transactions, so we
1378 	 * can just map that event to its transaction.  Transactions
1379 	 * associated with completion events are consecutive.
1380 	 */
1381 	old_index = ring->index;
1382 	event = gsi_ring_virt(ring, old_index);
1383 	trans = gsi_event_trans(channel, event);
1384 
1385 	/* Compute the number of events to process before we wrap,
1386 	 * and determine when we'll be done processing events.
1387 	 */
1388 	event_avail = ring->count - old_index % ring->count;
1389 	event_done = gsi_ring_virt(ring, index);
1390 	do {
1391 		trans->len = __le16_to_cpu(event->len);
1392 		byte_count += trans->len;
1393 
1394 		/* Move on to the next event and transaction */
1395 		if (--event_avail)
1396 			event++;
1397 		else
1398 			event = gsi_ring_virt(ring, 0);
1399 		trans = gsi_trans_pool_next(&trans_info->pool, trans);
1400 	} while (event != event_done);
1401 
1402 	/* We record RX bytes when they are received */
1403 	channel->byte_count += byte_count;
1404 	channel->trans_count++;
1405 }
1406 
1407 /* Initialize a ring, including allocating DMA memory for its entries */
1408 static int gsi_ring_alloc(struct gsi *gsi, struct gsi_ring *ring, u32 count)
1409 {
1410 	u32 size = count * GSI_RING_ELEMENT_SIZE;
1411 	struct device *dev = gsi->dev;
1412 	dma_addr_t addr;
1413 
1414 	/* Hardware requires a 2^n ring size, with alignment equal to size.
1415 	 * The DMA address returned by dma_alloc_coherent() is guaranteed to
1416 	 * be a power-of-2 number of pages, which satisfies the requirement.
1417 	 */
1418 	ring->virt = dma_alloc_coherent(dev, size, &addr, GFP_KERNEL);
1419 	if (!ring->virt)
1420 		return -ENOMEM;
1421 
1422 	ring->addr = addr;
1423 	ring->count = count;
1424 
1425 	return 0;
1426 }
1427 
1428 /* Free a previously-allocated ring */
1429 static void gsi_ring_free(struct gsi *gsi, struct gsi_ring *ring)
1430 {
1431 	size_t size = ring->count * GSI_RING_ELEMENT_SIZE;
1432 
1433 	dma_free_coherent(gsi->dev, size, ring->virt, ring->addr);
1434 }
1435 
1436 /* Allocate an available event ring id */
1437 static int gsi_evt_ring_id_alloc(struct gsi *gsi)
1438 {
1439 	u32 evt_ring_id;
1440 
1441 	if (gsi->event_bitmap == ~0U) {
1442 		dev_err(gsi->dev, "event rings exhausted\n");
1443 		return -ENOSPC;
1444 	}
1445 
1446 	evt_ring_id = ffz(gsi->event_bitmap);
1447 	gsi->event_bitmap |= BIT(evt_ring_id);
1448 
1449 	return (int)evt_ring_id;
1450 }
1451 
1452 /* Free a previously-allocated event ring id */
1453 static void gsi_evt_ring_id_free(struct gsi *gsi, u32 evt_ring_id)
1454 {
1455 	gsi->event_bitmap &= ~BIT(evt_ring_id);
1456 }
1457 
1458 /* Ring a channel doorbell, reporting the first un-filled entry */
1459 void gsi_channel_doorbell(struct gsi_channel *channel)
1460 {
1461 	struct gsi_ring *tre_ring = &channel->tre_ring;
1462 	u32 channel_id = gsi_channel_id(channel);
1463 	struct gsi *gsi = channel->gsi;
1464 	u32 val;
1465 
1466 	/* Note: index *must* be used modulo the ring count here */
1467 	val = gsi_ring_addr(tre_ring, tre_ring->index % tre_ring->count);
1468 	iowrite32(val, gsi->virt + GSI_CH_C_DOORBELL_0_OFFSET(channel_id));
1469 }
1470 
1471 /* Consult hardware, move any newly completed transactions to completed list */
1472 static struct gsi_trans *gsi_channel_update(struct gsi_channel *channel)
1473 {
1474 	u32 evt_ring_id = channel->evt_ring_id;
1475 	struct gsi *gsi = channel->gsi;
1476 	struct gsi_evt_ring *evt_ring;
1477 	struct gsi_trans *trans;
1478 	struct gsi_ring *ring;
1479 	u32 offset;
1480 	u32 index;
1481 
1482 	evt_ring = &gsi->evt_ring[evt_ring_id];
1483 	ring = &evt_ring->ring;
1484 
1485 	/* See if there's anything new to process; if not, we're done.  Note
1486 	 * that index always refers to an entry *within* the event ring.
1487 	 */
1488 	offset = GSI_EV_CH_E_CNTXT_4_OFFSET(evt_ring_id);
1489 	index = gsi_ring_index(ring, ioread32(gsi->virt + offset));
1490 	if (index == ring->index % ring->count)
1491 		return NULL;
1492 
1493 	/* Get the transaction for the latest completed event.  Take a
1494 	 * reference to keep it from completing before we give the events
1495 	 * for this and previous transactions back to the hardware.
1496 	 */
1497 	trans = gsi_event_trans(channel, gsi_ring_virt(ring, index - 1));
1498 	refcount_inc(&trans->refcount);
1499 
1500 	/* For RX channels, update each completed transaction with the number
1501 	 * of bytes that were actually received.  For TX channels, report
1502 	 * the number of transactions and bytes this completion represents
1503 	 * up the network stack.
1504 	 */
1505 	if (channel->toward_ipa)
1506 		gsi_channel_tx_update(channel, trans);
1507 	else
1508 		gsi_evt_ring_rx_update(evt_ring, index);
1509 
1510 	gsi_trans_move_complete(trans);
1511 
1512 	/* Tell the hardware we've handled these events */
1513 	gsi_evt_ring_doorbell(channel->gsi, channel->evt_ring_id, index);
1514 
1515 	gsi_trans_free(trans);
1516 
1517 	return gsi_channel_trans_complete(channel);
1518 }
1519 
1520 /**
1521  * gsi_channel_poll_one() - Return a single completed transaction on a channel
1522  * @channel:	Channel to be polled
1523  *
1524  * Return:	Transaction pointer, or null if none are available
1525  *
1526  * This function returns the first entry on a channel's completed transaction
1527  * list.  If that list is empty, the hardware is consulted to determine
1528  * whether any new transactions have completed.  If so, they're moved to the
1529  * completed list and the new first entry is returned.  If there are no more
1530  * completed transactions, a null pointer is returned.
1531  */
1532 static struct gsi_trans *gsi_channel_poll_one(struct gsi_channel *channel)
1533 {
1534 	struct gsi_trans *trans;
1535 
1536 	/* Get the first transaction from the completed list */
1537 	trans = gsi_channel_trans_complete(channel);
1538 	if (!trans)	/* List is empty; see if there's more to do */
1539 		trans = gsi_channel_update(channel);
1540 
1541 	if (trans)
1542 		gsi_trans_move_polled(trans);
1543 
1544 	return trans;
1545 }
1546 
1547 /**
1548  * gsi_channel_poll() - NAPI poll function for a channel
1549  * @napi:	NAPI structure for the channel
1550  * @budget:	Budget supplied by NAPI core
1551  *
1552  * Return:	Number of items polled (<= budget)
1553  *
1554  * Single transactions completed by hardware are polled until either
1555  * the budget is exhausted, or there are no more.  Each transaction
1556  * polled is passed to gsi_trans_complete(), to perform remaining
1557  * completion processing and retire/free the transaction.
1558  */
1559 static int gsi_channel_poll(struct napi_struct *napi, int budget)
1560 {
1561 	struct gsi_channel *channel;
1562 	int count;
1563 
1564 	channel = container_of(napi, struct gsi_channel, napi);
1565 	for (count = 0; count < budget; count++) {
1566 		struct gsi_trans *trans;
1567 
1568 		trans = gsi_channel_poll_one(channel);
1569 		if (!trans)
1570 			break;
1571 		gsi_trans_complete(trans);
1572 	}
1573 
1574 	if (count < budget && napi_complete(napi))
1575 		gsi_irq_ieob_enable_one(channel->gsi, channel->evt_ring_id);
1576 
1577 	return count;
1578 }
1579 
1580 /* The event bitmap represents which event ids are available for allocation.
1581  * Set bits are not available, clear bits can be used.  This function
1582  * initializes the map so all events supported by the hardware are available,
1583  * then precludes any reserved events from being allocated.
1584  */
1585 static u32 gsi_event_bitmap_init(u32 evt_ring_max)
1586 {
1587 	u32 event_bitmap = GENMASK(BITS_PER_LONG - 1, evt_ring_max);
1588 
1589 	event_bitmap |= GENMASK(GSI_MHI_EVENT_ID_END, GSI_MHI_EVENT_ID_START);
1590 
1591 	return event_bitmap;
1592 }
1593 
1594 /* Setup function for a single channel */
1595 static int gsi_channel_setup_one(struct gsi *gsi, u32 channel_id)
1596 {
1597 	struct gsi_channel *channel = &gsi->channel[channel_id];
1598 	u32 evt_ring_id = channel->evt_ring_id;
1599 	int ret;
1600 
1601 	if (!gsi_channel_initialized(channel))
1602 		return 0;
1603 
1604 	ret = gsi_evt_ring_alloc_command(gsi, evt_ring_id);
1605 	if (ret)
1606 		return ret;
1607 
1608 	gsi_evt_ring_program(gsi, evt_ring_id);
1609 
1610 	ret = gsi_channel_alloc_command(gsi, channel_id);
1611 	if (ret)
1612 		goto err_evt_ring_de_alloc;
1613 
1614 	gsi_channel_program(channel, true);
1615 
1616 	if (channel->toward_ipa)
1617 		netif_tx_napi_add(&gsi->dummy_dev, &channel->napi,
1618 				  gsi_channel_poll, NAPI_POLL_WEIGHT);
1619 	else
1620 		netif_napi_add(&gsi->dummy_dev, &channel->napi,
1621 			       gsi_channel_poll, NAPI_POLL_WEIGHT);
1622 
1623 	return 0;
1624 
1625 err_evt_ring_de_alloc:
1626 	/* We've done nothing with the event ring yet so don't reset */
1627 	gsi_evt_ring_de_alloc_command(gsi, evt_ring_id);
1628 
1629 	return ret;
1630 }
1631 
1632 /* Inverse of gsi_channel_setup_one() */
1633 static void gsi_channel_teardown_one(struct gsi *gsi, u32 channel_id)
1634 {
1635 	struct gsi_channel *channel = &gsi->channel[channel_id];
1636 	u32 evt_ring_id = channel->evt_ring_id;
1637 
1638 	if (!gsi_channel_initialized(channel))
1639 		return;
1640 
1641 	netif_napi_del(&channel->napi);
1642 
1643 	gsi_channel_de_alloc_command(gsi, channel_id);
1644 	gsi_evt_ring_reset_command(gsi, evt_ring_id);
1645 	gsi_evt_ring_de_alloc_command(gsi, evt_ring_id);
1646 }
1647 
1648 /* We use generic commands only to operate on modem channels.  We don't have
1649  * the ability to determine channel state for a modem channel, so we simply
1650  * issue the command and wait for it to complete.
1651  */
1652 static int gsi_generic_command(struct gsi *gsi, u32 channel_id,
1653 			       enum gsi_generic_cmd_opcode opcode,
1654 			       u8 params)
1655 {
1656 	bool timeout;
1657 	u32 val;
1658 
1659 	/* The error global interrupt type is always enabled (until we tear
1660 	 * down), so we will keep it enabled.
1661 	 *
1662 	 * A generic EE command completes with a GSI global interrupt of
1663 	 * type GP_INT1.  We only perform one generic command at a time
1664 	 * (to allocate, halt, or enable/disable flow control on a modem
1665 	 * channel), and only from this function.  So we enable the GP_INT1
1666 	 * IRQ type here, and disable it again after the command completes.
1667 	 */
1668 	val = BIT(ERROR_INT) | BIT(GP_INT1);
1669 	iowrite32(val, gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
1670 
1671 	/* First zero the result code field */
1672 	val = ioread32(gsi->virt + GSI_CNTXT_SCRATCH_0_OFFSET);
1673 	val &= ~GENERIC_EE_RESULT_FMASK;
1674 	iowrite32(val, gsi->virt + GSI_CNTXT_SCRATCH_0_OFFSET);
1675 
1676 	/* Now issue the command */
1677 	val = u32_encode_bits(opcode, GENERIC_OPCODE_FMASK);
1678 	val |= u32_encode_bits(channel_id, GENERIC_CHID_FMASK);
1679 	val |= u32_encode_bits(GSI_EE_MODEM, GENERIC_EE_FMASK);
1680 	val |= u32_encode_bits(params, GENERIC_PARAMS_FMASK);
1681 
1682 	timeout = !gsi_command(gsi, GSI_GENERIC_CMD_OFFSET, val);
1683 
1684 	/* Disable the GP_INT1 IRQ type again */
1685 	iowrite32(BIT(ERROR_INT), gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
1686 
1687 	if (!timeout)
1688 		return gsi->result;
1689 
1690 	dev_err(gsi->dev, "GSI generic command %u to channel %u timed out\n",
1691 		opcode, channel_id);
1692 
1693 	return -ETIMEDOUT;
1694 }
1695 
1696 static int gsi_modem_channel_alloc(struct gsi *gsi, u32 channel_id)
1697 {
1698 	return gsi_generic_command(gsi, channel_id,
1699 				   GSI_GENERIC_ALLOCATE_CHANNEL, 0);
1700 }
1701 
1702 static void gsi_modem_channel_halt(struct gsi *gsi, u32 channel_id)
1703 {
1704 	u32 retries = GSI_CHANNEL_MODEM_HALT_RETRIES;
1705 	int ret;
1706 
1707 	do
1708 		ret = gsi_generic_command(gsi, channel_id,
1709 					  GSI_GENERIC_HALT_CHANNEL, 0);
1710 	while (ret == -EAGAIN && retries--);
1711 
1712 	if (ret)
1713 		dev_err(gsi->dev, "error %d halting modem channel %u\n",
1714 			ret, channel_id);
1715 }
1716 
1717 /* Enable or disable flow control for a modem GSI TX channel (IPA v4.2+) */
1718 void
1719 gsi_modem_channel_flow_control(struct gsi *gsi, u32 channel_id, bool enable)
1720 {
1721 	u32 retries = 0;
1722 	u32 command;
1723 	int ret;
1724 
1725 	command = enable ? GSI_GENERIC_ENABLE_FLOW_CONTROL
1726 			 : GSI_GENERIC_DISABLE_FLOW_CONTROL;
1727 	/* Disabling flow control on IPA v4.11+ can return -EAGAIN if enable
1728 	 * is underway.  In this case we need to retry the command.
1729 	 */
1730 	if (!enable && gsi->version >= IPA_VERSION_4_11)
1731 		retries = GSI_CHANNEL_MODEM_FLOW_RETRIES;
1732 
1733 	do
1734 		ret = gsi_generic_command(gsi, channel_id, command, 0);
1735 	while (ret == -EAGAIN && retries--);
1736 
1737 	if (ret)
1738 		dev_err(gsi->dev,
1739 			"error %d %sabling mode channel %u flow control\n",
1740 			ret, enable ? "en" : "dis", channel_id);
1741 }
1742 
1743 /* Setup function for channels */
1744 static int gsi_channel_setup(struct gsi *gsi)
1745 {
1746 	u32 channel_id = 0;
1747 	u32 mask;
1748 	int ret;
1749 
1750 	gsi_irq_enable(gsi);
1751 
1752 	mutex_lock(&gsi->mutex);
1753 
1754 	do {
1755 		ret = gsi_channel_setup_one(gsi, channel_id);
1756 		if (ret)
1757 			goto err_unwind;
1758 	} while (++channel_id < gsi->channel_count);
1759 
1760 	/* Make sure no channels were defined that hardware does not support */
1761 	while (channel_id < GSI_CHANNEL_COUNT_MAX) {
1762 		struct gsi_channel *channel = &gsi->channel[channel_id++];
1763 
1764 		if (!gsi_channel_initialized(channel))
1765 			continue;
1766 
1767 		ret = -EINVAL;
1768 		dev_err(gsi->dev, "channel %u not supported by hardware\n",
1769 			channel_id - 1);
1770 		channel_id = gsi->channel_count;
1771 		goto err_unwind;
1772 	}
1773 
1774 	/* Allocate modem channels if necessary */
1775 	mask = gsi->modem_channel_bitmap;
1776 	while (mask) {
1777 		u32 modem_channel_id = __ffs(mask);
1778 
1779 		ret = gsi_modem_channel_alloc(gsi, modem_channel_id);
1780 		if (ret)
1781 			goto err_unwind_modem;
1782 
1783 		/* Clear bit from mask only after success (for unwind) */
1784 		mask ^= BIT(modem_channel_id);
1785 	}
1786 
1787 	mutex_unlock(&gsi->mutex);
1788 
1789 	return 0;
1790 
1791 err_unwind_modem:
1792 	/* Compute which modem channels need to be deallocated */
1793 	mask ^= gsi->modem_channel_bitmap;
1794 	while (mask) {
1795 		channel_id = __fls(mask);
1796 
1797 		mask ^= BIT(channel_id);
1798 
1799 		gsi_modem_channel_halt(gsi, channel_id);
1800 	}
1801 
1802 err_unwind:
1803 	while (channel_id--)
1804 		gsi_channel_teardown_one(gsi, channel_id);
1805 
1806 	mutex_unlock(&gsi->mutex);
1807 
1808 	gsi_irq_disable(gsi);
1809 
1810 	return ret;
1811 }
1812 
1813 /* Inverse of gsi_channel_setup() */
1814 static void gsi_channel_teardown(struct gsi *gsi)
1815 {
1816 	u32 mask = gsi->modem_channel_bitmap;
1817 	u32 channel_id;
1818 
1819 	mutex_lock(&gsi->mutex);
1820 
1821 	while (mask) {
1822 		channel_id = __fls(mask);
1823 
1824 		mask ^= BIT(channel_id);
1825 
1826 		gsi_modem_channel_halt(gsi, channel_id);
1827 	}
1828 
1829 	channel_id = gsi->channel_count - 1;
1830 	do
1831 		gsi_channel_teardown_one(gsi, channel_id);
1832 	while (channel_id--);
1833 
1834 	mutex_unlock(&gsi->mutex);
1835 
1836 	gsi_irq_disable(gsi);
1837 }
1838 
1839 /* Turn off all GSI interrupts initially */
1840 static int gsi_irq_setup(struct gsi *gsi)
1841 {
1842 	int ret;
1843 
1844 	/* Writing 1 indicates IRQ interrupts; 0 would be MSI */
1845 	iowrite32(1, gsi->virt + GSI_CNTXT_INTSET_OFFSET);
1846 
1847 	/* Disable all interrupt types */
1848 	gsi_irq_type_update(gsi, 0);
1849 
1850 	/* Clear all type-specific interrupt masks */
1851 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_MSK_OFFSET);
1852 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_MSK_OFFSET);
1853 	iowrite32(0, gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
1854 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_MSK_OFFSET);
1855 
1856 	/* The inter-EE interrupts are not supported for IPA v3.0-v3.1 */
1857 	if (gsi->version > IPA_VERSION_3_1) {
1858 		u32 offset;
1859 
1860 		/* These registers are in the non-adjusted address range */
1861 		offset = GSI_INTER_EE_SRC_CH_IRQ_MSK_OFFSET;
1862 		iowrite32(0, gsi->virt_raw + offset);
1863 		offset = GSI_INTER_EE_SRC_EV_CH_IRQ_MSK_OFFSET;
1864 		iowrite32(0, gsi->virt_raw + offset);
1865 	}
1866 
1867 	iowrite32(0, gsi->virt + GSI_CNTXT_GSI_IRQ_EN_OFFSET);
1868 
1869 	ret = request_irq(gsi->irq, gsi_isr, 0, "gsi", gsi);
1870 	if (ret)
1871 		dev_err(gsi->dev, "error %d requesting \"gsi\" IRQ\n", ret);
1872 
1873 	return ret;
1874 }
1875 
1876 static void gsi_irq_teardown(struct gsi *gsi)
1877 {
1878 	free_irq(gsi->irq, gsi);
1879 }
1880 
1881 /* Get # supported channel and event rings; there is no gsi_ring_teardown() */
1882 static int gsi_ring_setup(struct gsi *gsi)
1883 {
1884 	struct device *dev = gsi->dev;
1885 	u32 count;
1886 	u32 val;
1887 
1888 	if (gsi->version < IPA_VERSION_3_5_1) {
1889 		/* No HW_PARAM_2 register prior to IPA v3.5.1, assume the max */
1890 		gsi->channel_count = GSI_CHANNEL_COUNT_MAX;
1891 		gsi->evt_ring_count = GSI_EVT_RING_COUNT_MAX;
1892 
1893 		return 0;
1894 	}
1895 
1896 	val = ioread32(gsi->virt + GSI_GSI_HW_PARAM_2_OFFSET);
1897 
1898 	count = u32_get_bits(val, NUM_CH_PER_EE_FMASK);
1899 	if (!count) {
1900 		dev_err(dev, "GSI reports zero channels supported\n");
1901 		return -EINVAL;
1902 	}
1903 	if (count > GSI_CHANNEL_COUNT_MAX) {
1904 		dev_warn(dev, "limiting to %u channels; hardware supports %u\n",
1905 			 GSI_CHANNEL_COUNT_MAX, count);
1906 		count = GSI_CHANNEL_COUNT_MAX;
1907 	}
1908 	gsi->channel_count = count;
1909 
1910 	count = u32_get_bits(val, NUM_EV_PER_EE_FMASK);
1911 	if (!count) {
1912 		dev_err(dev, "GSI reports zero event rings supported\n");
1913 		return -EINVAL;
1914 	}
1915 	if (count > GSI_EVT_RING_COUNT_MAX) {
1916 		dev_warn(dev,
1917 			 "limiting to %u event rings; hardware supports %u\n",
1918 			 GSI_EVT_RING_COUNT_MAX, count);
1919 		count = GSI_EVT_RING_COUNT_MAX;
1920 	}
1921 	gsi->evt_ring_count = count;
1922 
1923 	return 0;
1924 }
1925 
1926 /* Setup function for GSI.  GSI firmware must be loaded and initialized */
1927 int gsi_setup(struct gsi *gsi)
1928 {
1929 	u32 val;
1930 	int ret;
1931 
1932 	/* Here is where we first touch the GSI hardware */
1933 	val = ioread32(gsi->virt + GSI_GSI_STATUS_OFFSET);
1934 	if (!(val & ENABLED_FMASK)) {
1935 		dev_err(gsi->dev, "GSI has not been enabled\n");
1936 		return -EIO;
1937 	}
1938 
1939 	ret = gsi_irq_setup(gsi);
1940 	if (ret)
1941 		return ret;
1942 
1943 	ret = gsi_ring_setup(gsi);	/* No matching teardown required */
1944 	if (ret)
1945 		goto err_irq_teardown;
1946 
1947 	/* Initialize the error log */
1948 	iowrite32(0, gsi->virt + GSI_ERROR_LOG_OFFSET);
1949 
1950 	ret = gsi_channel_setup(gsi);
1951 	if (ret)
1952 		goto err_irq_teardown;
1953 
1954 	return 0;
1955 
1956 err_irq_teardown:
1957 	gsi_irq_teardown(gsi);
1958 
1959 	return ret;
1960 }
1961 
1962 /* Inverse of gsi_setup() */
1963 void gsi_teardown(struct gsi *gsi)
1964 {
1965 	gsi_channel_teardown(gsi);
1966 	gsi_irq_teardown(gsi);
1967 }
1968 
1969 /* Initialize a channel's event ring */
1970 static int gsi_channel_evt_ring_init(struct gsi_channel *channel)
1971 {
1972 	struct gsi *gsi = channel->gsi;
1973 	struct gsi_evt_ring *evt_ring;
1974 	int ret;
1975 
1976 	ret = gsi_evt_ring_id_alloc(gsi);
1977 	if (ret < 0)
1978 		return ret;
1979 	channel->evt_ring_id = ret;
1980 
1981 	evt_ring = &gsi->evt_ring[channel->evt_ring_id];
1982 	evt_ring->channel = channel;
1983 
1984 	ret = gsi_ring_alloc(gsi, &evt_ring->ring, channel->event_count);
1985 	if (!ret)
1986 		return 0;	/* Success! */
1987 
1988 	dev_err(gsi->dev, "error %d allocating channel %u event ring\n",
1989 		ret, gsi_channel_id(channel));
1990 
1991 	gsi_evt_ring_id_free(gsi, channel->evt_ring_id);
1992 
1993 	return ret;
1994 }
1995 
1996 /* Inverse of gsi_channel_evt_ring_init() */
1997 static void gsi_channel_evt_ring_exit(struct gsi_channel *channel)
1998 {
1999 	u32 evt_ring_id = channel->evt_ring_id;
2000 	struct gsi *gsi = channel->gsi;
2001 	struct gsi_evt_ring *evt_ring;
2002 
2003 	evt_ring = &gsi->evt_ring[evt_ring_id];
2004 	gsi_ring_free(gsi, &evt_ring->ring);
2005 	gsi_evt_ring_id_free(gsi, evt_ring_id);
2006 }
2007 
2008 static bool gsi_channel_data_valid(struct gsi *gsi,
2009 				   const struct ipa_gsi_endpoint_data *data)
2010 {
2011 	u32 channel_id = data->channel_id;
2012 	struct device *dev = gsi->dev;
2013 
2014 	/* Make sure channel ids are in the range driver supports */
2015 	if (channel_id >= GSI_CHANNEL_COUNT_MAX) {
2016 		dev_err(dev, "bad channel id %u; must be less than %u\n",
2017 			channel_id, GSI_CHANNEL_COUNT_MAX);
2018 		return false;
2019 	}
2020 
2021 	if (data->ee_id != GSI_EE_AP && data->ee_id != GSI_EE_MODEM) {
2022 		dev_err(dev, "bad EE id %u; not AP or modem\n", data->ee_id);
2023 		return false;
2024 	}
2025 
2026 	if (!data->channel.tlv_count ||
2027 	    data->channel.tlv_count > GSI_TLV_MAX) {
2028 		dev_err(dev, "channel %u bad tlv_count %u; must be 1..%u\n",
2029 			channel_id, data->channel.tlv_count, GSI_TLV_MAX);
2030 		return false;
2031 	}
2032 
2033 	/* We have to allow at least one maximally-sized transaction to
2034 	 * be outstanding (which would use tlv_count TREs).  Given how
2035 	 * gsi_channel_tre_max() is computed, tre_count has to be almost
2036 	 * twice the TLV FIFO size to satisfy this requirement.
2037 	 */
2038 	if (data->channel.tre_count < 2 * data->channel.tlv_count - 1) {
2039 		dev_err(dev, "channel %u TLV count %u exceeds TRE count %u\n",
2040 			channel_id, data->channel.tlv_count,
2041 			data->channel.tre_count);
2042 		return false;
2043 	}
2044 
2045 	if (!is_power_of_2(data->channel.tre_count)) {
2046 		dev_err(dev, "channel %u bad tre_count %u; not power of 2\n",
2047 			channel_id, data->channel.tre_count);
2048 		return false;
2049 	}
2050 
2051 	if (!is_power_of_2(data->channel.event_count)) {
2052 		dev_err(dev, "channel %u bad event_count %u; not power of 2\n",
2053 			channel_id, data->channel.event_count);
2054 		return false;
2055 	}
2056 
2057 	return true;
2058 }
2059 
2060 /* Init function for a single channel */
2061 static int gsi_channel_init_one(struct gsi *gsi,
2062 				const struct ipa_gsi_endpoint_data *data,
2063 				bool command)
2064 {
2065 	struct gsi_channel *channel;
2066 	u32 tre_count;
2067 	int ret;
2068 
2069 	if (!gsi_channel_data_valid(gsi, data))
2070 		return -EINVAL;
2071 
2072 	/* Worst case we need an event for every outstanding TRE */
2073 	if (data->channel.tre_count > data->channel.event_count) {
2074 		tre_count = data->channel.event_count;
2075 		dev_warn(gsi->dev, "channel %u limited to %u TREs\n",
2076 			 data->channel_id, tre_count);
2077 	} else {
2078 		tre_count = data->channel.tre_count;
2079 	}
2080 
2081 	channel = &gsi->channel[data->channel_id];
2082 	memset(channel, 0, sizeof(*channel));
2083 
2084 	channel->gsi = gsi;
2085 	channel->toward_ipa = data->toward_ipa;
2086 	channel->command = command;
2087 	channel->tlv_count = data->channel.tlv_count;
2088 	channel->tre_count = tre_count;
2089 	channel->event_count = data->channel.event_count;
2090 
2091 	ret = gsi_channel_evt_ring_init(channel);
2092 	if (ret)
2093 		goto err_clear_gsi;
2094 
2095 	ret = gsi_ring_alloc(gsi, &channel->tre_ring, data->channel.tre_count);
2096 	if (ret) {
2097 		dev_err(gsi->dev, "error %d allocating channel %u ring\n",
2098 			ret, data->channel_id);
2099 		goto err_channel_evt_ring_exit;
2100 	}
2101 
2102 	ret = gsi_channel_trans_init(gsi, data->channel_id);
2103 	if (ret)
2104 		goto err_ring_free;
2105 
2106 	if (command) {
2107 		u32 tre_max = gsi_channel_tre_max(gsi, data->channel_id);
2108 
2109 		ret = ipa_cmd_pool_init(channel, tre_max);
2110 	}
2111 	if (!ret)
2112 		return 0;	/* Success! */
2113 
2114 	gsi_channel_trans_exit(channel);
2115 err_ring_free:
2116 	gsi_ring_free(gsi, &channel->tre_ring);
2117 err_channel_evt_ring_exit:
2118 	gsi_channel_evt_ring_exit(channel);
2119 err_clear_gsi:
2120 	channel->gsi = NULL;	/* Mark it not (fully) initialized */
2121 
2122 	return ret;
2123 }
2124 
2125 /* Inverse of gsi_channel_init_one() */
2126 static void gsi_channel_exit_one(struct gsi_channel *channel)
2127 {
2128 	if (!gsi_channel_initialized(channel))
2129 		return;
2130 
2131 	if (channel->command)
2132 		ipa_cmd_pool_exit(channel);
2133 	gsi_channel_trans_exit(channel);
2134 	gsi_ring_free(channel->gsi, &channel->tre_ring);
2135 	gsi_channel_evt_ring_exit(channel);
2136 }
2137 
2138 /* Init function for channels */
2139 static int gsi_channel_init(struct gsi *gsi, u32 count,
2140 			    const struct ipa_gsi_endpoint_data *data)
2141 {
2142 	bool modem_alloc;
2143 	int ret = 0;
2144 	u32 i;
2145 
2146 	/* IPA v4.2 requires the AP to allocate channels for the modem */
2147 	modem_alloc = gsi->version == IPA_VERSION_4_2;
2148 
2149 	gsi->event_bitmap = gsi_event_bitmap_init(GSI_EVT_RING_COUNT_MAX);
2150 	gsi->ieob_enabled_bitmap = 0;
2151 
2152 	/* The endpoint data array is indexed by endpoint name */
2153 	for (i = 0; i < count; i++) {
2154 		bool command = i == IPA_ENDPOINT_AP_COMMAND_TX;
2155 
2156 		if (ipa_gsi_endpoint_data_empty(&data[i]))
2157 			continue;	/* Skip over empty slots */
2158 
2159 		/* Mark modem channels to be allocated (hardware workaround) */
2160 		if (data[i].ee_id == GSI_EE_MODEM) {
2161 			if (modem_alloc)
2162 				gsi->modem_channel_bitmap |=
2163 						BIT(data[i].channel_id);
2164 			continue;
2165 		}
2166 
2167 		ret = gsi_channel_init_one(gsi, &data[i], command);
2168 		if (ret)
2169 			goto err_unwind;
2170 	}
2171 
2172 	return ret;
2173 
2174 err_unwind:
2175 	while (i--) {
2176 		if (ipa_gsi_endpoint_data_empty(&data[i]))
2177 			continue;
2178 		if (modem_alloc && data[i].ee_id == GSI_EE_MODEM) {
2179 			gsi->modem_channel_bitmap &= ~BIT(data[i].channel_id);
2180 			continue;
2181 		}
2182 		gsi_channel_exit_one(&gsi->channel[data->channel_id]);
2183 	}
2184 
2185 	return ret;
2186 }
2187 
2188 /* Inverse of gsi_channel_init() */
2189 static void gsi_channel_exit(struct gsi *gsi)
2190 {
2191 	u32 channel_id = GSI_CHANNEL_COUNT_MAX - 1;
2192 
2193 	do
2194 		gsi_channel_exit_one(&gsi->channel[channel_id]);
2195 	while (channel_id--);
2196 	gsi->modem_channel_bitmap = 0;
2197 }
2198 
2199 /* Init function for GSI.  GSI hardware does not need to be "ready" */
2200 int gsi_init(struct gsi *gsi, struct platform_device *pdev,
2201 	     enum ipa_version version, u32 count,
2202 	     const struct ipa_gsi_endpoint_data *data)
2203 {
2204 	struct device *dev = &pdev->dev;
2205 	struct resource *res;
2206 	resource_size_t size;
2207 	u32 adjust;
2208 	int ret;
2209 
2210 	gsi_validate_build();
2211 
2212 	gsi->dev = dev;
2213 	gsi->version = version;
2214 
2215 	/* GSI uses NAPI on all channels.  Create a dummy network device
2216 	 * for the channel NAPI contexts to be associated with.
2217 	 */
2218 	init_dummy_netdev(&gsi->dummy_dev);
2219 
2220 	/* Get GSI memory range and map it */
2221 	res = platform_get_resource_byname(pdev, IORESOURCE_MEM, "gsi");
2222 	if (!res) {
2223 		dev_err(dev, "DT error getting \"gsi\" memory property\n");
2224 		return -ENODEV;
2225 	}
2226 
2227 	size = resource_size(res);
2228 	if (res->start > U32_MAX || size > U32_MAX - res->start) {
2229 		dev_err(dev, "DT memory resource \"gsi\" out of range\n");
2230 		return -EINVAL;
2231 	}
2232 
2233 	/* Make sure we can make our pointer adjustment if necessary */
2234 	adjust = gsi->version < IPA_VERSION_4_5 ? 0 : GSI_EE_REG_ADJUST;
2235 	if (res->start < adjust) {
2236 		dev_err(dev, "DT memory resource \"gsi\" too low (< %u)\n",
2237 			adjust);
2238 		return -EINVAL;
2239 	}
2240 
2241 	gsi->virt_raw = ioremap(res->start, size);
2242 	if (!gsi->virt_raw) {
2243 		dev_err(dev, "unable to remap \"gsi\" memory\n");
2244 		return -ENOMEM;
2245 	}
2246 	/* Most registers are accessed using an adjusted register range */
2247 	gsi->virt = gsi->virt_raw - adjust;
2248 
2249 	init_completion(&gsi->completion);
2250 
2251 	ret = gsi_irq_init(gsi, pdev);	/* No matching exit required */
2252 	if (ret)
2253 		goto err_iounmap;
2254 
2255 	ret = gsi_channel_init(gsi, count, data);
2256 	if (ret)
2257 		goto err_iounmap;
2258 
2259 	mutex_init(&gsi->mutex);
2260 
2261 	return 0;
2262 
2263 err_iounmap:
2264 	iounmap(gsi->virt_raw);
2265 
2266 	return ret;
2267 }
2268 
2269 /* Inverse of gsi_init() */
2270 void gsi_exit(struct gsi *gsi)
2271 {
2272 	mutex_destroy(&gsi->mutex);
2273 	gsi_channel_exit(gsi);
2274 	iounmap(gsi->virt_raw);
2275 }
2276 
2277 /* The maximum number of outstanding TREs on a channel.  This limits
2278  * a channel's maximum number of transactions outstanding (worst case
2279  * is one TRE per transaction).
2280  *
2281  * The absolute limit is the number of TREs in the channel's TRE ring,
2282  * and in theory we should be able use all of them.  But in practice,
2283  * doing that led to the hardware reporting exhaustion of event ring
2284  * slots for writing completion information.  So the hardware limit
2285  * would be (tre_count - 1).
2286  *
2287  * We reduce it a bit further though.  Transaction resource pools are
2288  * sized to be a little larger than this maximum, to allow resource
2289  * allocations to always be contiguous.  The number of entries in a
2290  * TRE ring buffer is a power of 2, and the extra resources in a pool
2291  * tends to nearly double the memory allocated for it.  Reducing the
2292  * maximum number of outstanding TREs allows the number of entries in
2293  * a pool to avoid crossing that power-of-2 boundary, and this can
2294  * substantially reduce pool memory requirements.  The number we
2295  * reduce it by matches the number added in gsi_trans_pool_init().
2296  */
2297 u32 gsi_channel_tre_max(struct gsi *gsi, u32 channel_id)
2298 {
2299 	struct gsi_channel *channel = &gsi->channel[channel_id];
2300 
2301 	/* Hardware limit is channel->tre_count - 1 */
2302 	return channel->tre_count - (channel->tlv_count - 1);
2303 }
2304 
2305 /* Returns the maximum number of TREs in a single transaction for a channel */
2306 u32 gsi_channel_trans_tre_max(struct gsi *gsi, u32 channel_id)
2307 {
2308 	struct gsi_channel *channel = &gsi->channel[channel_id];
2309 
2310 	return channel->tlv_count;
2311 }
2312