xref: /linux/drivers/net/ipa/gsi.c (revision 1e328ed559201c4a3733b148b0afe3cfb4ad8b45)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 /* Copyright (c) 2015-2018, The Linux Foundation. All rights reserved.
4  * Copyright (C) 2018-2020 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 
97 #define GSI_MHI_EVENT_ID_START		10	/* 1st reserved event id */
98 #define GSI_MHI_EVENT_ID_END		16	/* Last reserved event id */
99 
100 #define GSI_ISR_MAX_ITER		50	/* Detect interrupt storms */
101 
102 /* An entry in an event ring */
103 struct gsi_event {
104 	__le64 xfer_ptr;
105 	__le16 len;
106 	u8 reserved1;
107 	u8 code;
108 	__le16 reserved2;
109 	u8 type;
110 	u8 chid;
111 };
112 
113 /** gsi_channel_scratch_gpi - GPI protocol scratch register
114  * @max_outstanding_tre:
115  *	Defines the maximum number of TREs allowed in a single transaction
116  *	on a channel (in bytes).  This determines the amount of prefetch
117  *	performed by the hardware.  We configure this to equal the size of
118  *	the TLV FIFO for the channel.
119  * @outstanding_threshold:
120  *	Defines the threshold (in bytes) determining when the sequencer
121  *	should update the channel doorbell.  We configure this to equal
122  *	the size of two TREs.
123  */
124 struct gsi_channel_scratch_gpi {
125 	u64 reserved1;
126 	u16 reserved2;
127 	u16 max_outstanding_tre;
128 	u16 reserved3;
129 	u16 outstanding_threshold;
130 };
131 
132 /** gsi_channel_scratch - channel scratch configuration area
133  *
134  * The exact interpretation of this register is protocol-specific.
135  * We only use GPI channels; see struct gsi_channel_scratch_gpi, above.
136  */
137 union gsi_channel_scratch {
138 	struct gsi_channel_scratch_gpi gpi;
139 	struct {
140 		u32 word1;
141 		u32 word2;
142 		u32 word3;
143 		u32 word4;
144 	} data;
145 };
146 
147 /* Check things that can be validated at build time. */
148 static void gsi_validate_build(void)
149 {
150 	/* This is used as a divisor */
151 	BUILD_BUG_ON(!GSI_RING_ELEMENT_SIZE);
152 
153 	/* Code assumes the size of channel and event ring element are
154 	 * the same (and fixed).  Make sure the size of an event ring
155 	 * element is what's expected.
156 	 */
157 	BUILD_BUG_ON(sizeof(struct gsi_event) != GSI_RING_ELEMENT_SIZE);
158 
159 	/* Hardware requires a 2^n ring size.  We ensure the number of
160 	 * elements in an event ring is a power of 2 elsewhere; this
161 	 * ensure the elements themselves meet the requirement.
162 	 */
163 	BUILD_BUG_ON(!is_power_of_2(GSI_RING_ELEMENT_SIZE));
164 
165 	/* The channel element size must fit in this field */
166 	BUILD_BUG_ON(GSI_RING_ELEMENT_SIZE > field_max(ELEMENT_SIZE_FMASK));
167 
168 	/* The event ring element size must fit in this field */
169 	BUILD_BUG_ON(GSI_RING_ELEMENT_SIZE > field_max(EV_ELEMENT_SIZE_FMASK));
170 }
171 
172 /* Return the channel id associated with a given channel */
173 static u32 gsi_channel_id(struct gsi_channel *channel)
174 {
175 	return channel - &channel->gsi->channel[0];
176 }
177 
178 /* Update the GSI IRQ type register with the cached value */
179 static void gsi_irq_type_update(struct gsi *gsi, u32 val)
180 {
181 	gsi->type_enabled_bitmap = val;
182 	iowrite32(val, gsi->virt + GSI_CNTXT_TYPE_IRQ_MSK_OFFSET);
183 }
184 
185 static void gsi_irq_type_enable(struct gsi *gsi, enum gsi_irq_type_id type_id)
186 {
187 	gsi_irq_type_update(gsi, gsi->type_enabled_bitmap | BIT(type_id));
188 }
189 
190 static void gsi_irq_type_disable(struct gsi *gsi, enum gsi_irq_type_id type_id)
191 {
192 	gsi_irq_type_update(gsi, gsi->type_enabled_bitmap & ~BIT(type_id));
193 }
194 
195 /* Turn off all GSI interrupts initially */
196 static void gsi_irq_setup(struct gsi *gsi)
197 {
198 	u32 adjust;
199 
200 	/* Disable all interrupt types */
201 	gsi_irq_type_update(gsi, 0);
202 
203 	/* Clear all type-specific interrupt masks */
204 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_MSK_OFFSET);
205 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_MSK_OFFSET);
206 	iowrite32(0, gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
207 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_MSK_OFFSET);
208 
209 	/* Reverse the offset adjustment for inter-EE register offsets */
210 	adjust = gsi->version < IPA_VERSION_4_5 ? 0 : GSI_EE_REG_ADJUST;
211 	iowrite32(0, gsi->virt + adjust + GSI_INTER_EE_SRC_CH_IRQ_OFFSET);
212 	iowrite32(0, gsi->virt + adjust + GSI_INTER_EE_SRC_EV_CH_IRQ_OFFSET);
213 
214 	iowrite32(0, gsi->virt + GSI_CNTXT_GSI_IRQ_EN_OFFSET);
215 }
216 
217 /* Turn off all GSI interrupts when we're all done */
218 static void gsi_irq_teardown(struct gsi *gsi)
219 {
220 	/* Nothing to do */
221 }
222 
223 /* Event ring commands are performed one at a time.  Their completion
224  * is signaled by the event ring control GSI interrupt type, which is
225  * only enabled when we issue an event ring command.  Only the event
226  * ring being operated on has this interrupt enabled.
227  */
228 static void gsi_irq_ev_ctrl_enable(struct gsi *gsi, u32 evt_ring_id)
229 {
230 	u32 val = BIT(evt_ring_id);
231 
232 	/* There's a small chance that a previous command completed
233 	 * after the interrupt was disabled, so make sure we have no
234 	 * pending interrupts before we enable them.
235 	 */
236 	iowrite32(~0, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_CLR_OFFSET);
237 
238 	iowrite32(val, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_MSK_OFFSET);
239 	gsi_irq_type_enable(gsi, GSI_EV_CTRL);
240 }
241 
242 /* Disable event ring control interrupts */
243 static void gsi_irq_ev_ctrl_disable(struct gsi *gsi)
244 {
245 	gsi_irq_type_disable(gsi, GSI_EV_CTRL);
246 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_MSK_OFFSET);
247 }
248 
249 /* Channel commands are performed one at a time.  Their completion is
250  * signaled by the channel control GSI interrupt type, which is only
251  * enabled when we issue a channel command.  Only the channel being
252  * operated on has this interrupt enabled.
253  */
254 static void gsi_irq_ch_ctrl_enable(struct gsi *gsi, u32 channel_id)
255 {
256 	u32 val = BIT(channel_id);
257 
258 	/* There's a small chance that a previous command completed
259 	 * after the interrupt was disabled, so make sure we have no
260 	 * pending interrupts before we enable them.
261 	 */
262 	iowrite32(~0, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_CLR_OFFSET);
263 
264 	iowrite32(val, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_MSK_OFFSET);
265 	gsi_irq_type_enable(gsi, GSI_CH_CTRL);
266 }
267 
268 /* Disable channel control interrupts */
269 static void gsi_irq_ch_ctrl_disable(struct gsi *gsi)
270 {
271 	gsi_irq_type_disable(gsi, GSI_CH_CTRL);
272 	iowrite32(0, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_MSK_OFFSET);
273 }
274 
275 static void gsi_irq_ieob_enable_one(struct gsi *gsi, u32 evt_ring_id)
276 {
277 	bool enable_ieob = !gsi->ieob_enabled_bitmap;
278 	u32 val;
279 
280 	gsi->ieob_enabled_bitmap |= BIT(evt_ring_id);
281 	val = gsi->ieob_enabled_bitmap;
282 	iowrite32(val, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_MSK_OFFSET);
283 
284 	/* Enable the interrupt type if this is the first channel enabled */
285 	if (enable_ieob)
286 		gsi_irq_type_enable(gsi, GSI_IEOB);
287 }
288 
289 static void gsi_irq_ieob_disable(struct gsi *gsi, u32 event_mask)
290 {
291 	u32 val;
292 
293 	gsi->ieob_enabled_bitmap &= ~event_mask;
294 
295 	/* Disable the interrupt type if this was the last enabled channel */
296 	if (!gsi->ieob_enabled_bitmap)
297 		gsi_irq_type_disable(gsi, GSI_IEOB);
298 
299 	val = gsi->ieob_enabled_bitmap;
300 	iowrite32(val, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_MSK_OFFSET);
301 }
302 
303 static void gsi_irq_ieob_disable_one(struct gsi *gsi, u32 evt_ring_id)
304 {
305 	gsi_irq_ieob_disable(gsi, BIT(evt_ring_id));
306 }
307 
308 /* Enable all GSI_interrupt types */
309 static void gsi_irq_enable(struct gsi *gsi)
310 {
311 	u32 val;
312 
313 	/* Global interrupts include hardware error reports.  Enable
314 	 * that so we can at least report the error should it occur.
315 	 */
316 	iowrite32(BIT(ERROR_INT), gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
317 	gsi_irq_type_update(gsi, gsi->type_enabled_bitmap | BIT(GSI_GLOB_EE));
318 
319 	/* General GSI interrupts are reported to all EEs; if they occur
320 	 * they are unrecoverable (without reset).  A breakpoint interrupt
321 	 * also exists, but we don't support that.  We want to be notified
322 	 * of errors so we can report them, even if they can't be handled.
323 	 */
324 	val = BIT(BUS_ERROR);
325 	val |= BIT(CMD_FIFO_OVRFLOW);
326 	val |= BIT(MCS_STACK_OVRFLOW);
327 	iowrite32(val, gsi->virt + GSI_CNTXT_GSI_IRQ_EN_OFFSET);
328 	gsi_irq_type_update(gsi, gsi->type_enabled_bitmap | BIT(GSI_GENERAL));
329 }
330 
331 /* Disable all GSI interrupt types */
332 static void gsi_irq_disable(struct gsi *gsi)
333 {
334 	gsi_irq_type_update(gsi, 0);
335 
336 	/* Clear the type-specific interrupt masks set by gsi_irq_enable() */
337 	iowrite32(0, gsi->virt + GSI_CNTXT_GSI_IRQ_EN_OFFSET);
338 	iowrite32(0, gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
339 }
340 
341 /* Return the virtual address associated with a ring index */
342 void *gsi_ring_virt(struct gsi_ring *ring, u32 index)
343 {
344 	/* Note: index *must* be used modulo the ring count here */
345 	return ring->virt + (index % ring->count) * GSI_RING_ELEMENT_SIZE;
346 }
347 
348 /* Return the 32-bit DMA address associated with a ring index */
349 static u32 gsi_ring_addr(struct gsi_ring *ring, u32 index)
350 {
351 	return (ring->addr & GENMASK(31, 0)) + index * GSI_RING_ELEMENT_SIZE;
352 }
353 
354 /* Return the ring index of a 32-bit ring offset */
355 static u32 gsi_ring_index(struct gsi_ring *ring, u32 offset)
356 {
357 	return (offset - gsi_ring_addr(ring, 0)) / GSI_RING_ELEMENT_SIZE;
358 }
359 
360 /* Issue a GSI command by writing a value to a register, then wait for
361  * completion to be signaled.  Returns true if the command completes
362  * or false if it times out.
363  */
364 static bool
365 gsi_command(struct gsi *gsi, u32 reg, u32 val, struct completion *completion)
366 {
367 	unsigned long timeout = msecs_to_jiffies(GSI_CMD_TIMEOUT);
368 
369 	reinit_completion(completion);
370 
371 	iowrite32(val, gsi->virt + reg);
372 
373 	return !!wait_for_completion_timeout(completion, timeout);
374 }
375 
376 /* Return the hardware's notion of the current state of an event ring */
377 static enum gsi_evt_ring_state
378 gsi_evt_ring_state(struct gsi *gsi, u32 evt_ring_id)
379 {
380 	u32 val;
381 
382 	val = ioread32(gsi->virt + GSI_EV_CH_E_CNTXT_0_OFFSET(evt_ring_id));
383 
384 	return u32_get_bits(val, EV_CHSTATE_FMASK);
385 }
386 
387 /* Issue an event ring command and wait for it to complete */
388 static void gsi_evt_ring_command(struct gsi *gsi, u32 evt_ring_id,
389 				 enum gsi_evt_cmd_opcode opcode)
390 {
391 	struct gsi_evt_ring *evt_ring = &gsi->evt_ring[evt_ring_id];
392 	struct completion *completion = &evt_ring->completion;
393 	struct device *dev = gsi->dev;
394 	bool timeout;
395 	u32 val;
396 
397 	/* Enable the completion interrupt for the command */
398 	gsi_irq_ev_ctrl_enable(gsi, evt_ring_id);
399 
400 	val = u32_encode_bits(evt_ring_id, EV_CHID_FMASK);
401 	val |= u32_encode_bits(opcode, EV_OPCODE_FMASK);
402 
403 	timeout = !gsi_command(gsi, GSI_EV_CH_CMD_OFFSET, val, completion);
404 
405 	gsi_irq_ev_ctrl_disable(gsi);
406 
407 	if (!timeout)
408 		return;
409 
410 	dev_err(dev, "GSI command %u for event ring %u timed out, state %u\n",
411 		opcode, evt_ring_id, evt_ring->state);
412 }
413 
414 /* Allocate an event ring in NOT_ALLOCATED state */
415 static int gsi_evt_ring_alloc_command(struct gsi *gsi, u32 evt_ring_id)
416 {
417 	struct gsi_evt_ring *evt_ring = &gsi->evt_ring[evt_ring_id];
418 
419 	/* Get initial event ring state */
420 	evt_ring->state = gsi_evt_ring_state(gsi, evt_ring_id);
421 	if (evt_ring->state != GSI_EVT_RING_STATE_NOT_ALLOCATED) {
422 		dev_err(gsi->dev, "event ring %u bad state %u before alloc\n",
423 			evt_ring_id, evt_ring->state);
424 		return -EINVAL;
425 	}
426 
427 	gsi_evt_ring_command(gsi, evt_ring_id, GSI_EVT_ALLOCATE);
428 
429 	/* If successful the event ring state will have changed */
430 	if (evt_ring->state == GSI_EVT_RING_STATE_ALLOCATED)
431 		return 0;
432 
433 	dev_err(gsi->dev, "event ring %u bad state %u after alloc\n",
434 		evt_ring_id, evt_ring->state);
435 
436 	return -EIO;
437 }
438 
439 /* Reset a GSI event ring in ALLOCATED or ERROR state. */
440 static void gsi_evt_ring_reset_command(struct gsi *gsi, u32 evt_ring_id)
441 {
442 	struct gsi_evt_ring *evt_ring = &gsi->evt_ring[evt_ring_id];
443 	enum gsi_evt_ring_state state = evt_ring->state;
444 
445 	if (state != GSI_EVT_RING_STATE_ALLOCATED &&
446 	    state != GSI_EVT_RING_STATE_ERROR) {
447 		dev_err(gsi->dev, "event ring %u bad state %u before reset\n",
448 			evt_ring_id, evt_ring->state);
449 		return;
450 	}
451 
452 	gsi_evt_ring_command(gsi, evt_ring_id, GSI_EVT_RESET);
453 
454 	/* If successful the event ring state will have changed */
455 	if (evt_ring->state == GSI_EVT_RING_STATE_ALLOCATED)
456 		return;
457 
458 	dev_err(gsi->dev, "event ring %u bad state %u after reset\n",
459 		evt_ring_id, evt_ring->state);
460 }
461 
462 /* Issue a hardware de-allocation request for an allocated event ring */
463 static void gsi_evt_ring_de_alloc_command(struct gsi *gsi, u32 evt_ring_id)
464 {
465 	struct gsi_evt_ring *evt_ring = &gsi->evt_ring[evt_ring_id];
466 
467 	if (evt_ring->state != GSI_EVT_RING_STATE_ALLOCATED) {
468 		dev_err(gsi->dev, "event ring %u state %u before dealloc\n",
469 			evt_ring_id, evt_ring->state);
470 		return;
471 	}
472 
473 	gsi_evt_ring_command(gsi, evt_ring_id, GSI_EVT_DE_ALLOC);
474 
475 	/* If successful the event ring state will have changed */
476 	if (evt_ring->state == GSI_EVT_RING_STATE_NOT_ALLOCATED)
477 		return;
478 
479 	dev_err(gsi->dev, "event ring %u bad state %u after dealloc\n",
480 		evt_ring_id, evt_ring->state);
481 }
482 
483 /* Fetch the current state of a channel from hardware */
484 static enum gsi_channel_state gsi_channel_state(struct gsi_channel *channel)
485 {
486 	u32 channel_id = gsi_channel_id(channel);
487 	void *virt = channel->gsi->virt;
488 	u32 val;
489 
490 	val = ioread32(virt + GSI_CH_C_CNTXT_0_OFFSET(channel_id));
491 
492 	return u32_get_bits(val, CHSTATE_FMASK);
493 }
494 
495 /* Issue a channel command and wait for it to complete */
496 static void
497 gsi_channel_command(struct gsi_channel *channel, enum gsi_ch_cmd_opcode opcode)
498 {
499 	struct completion *completion = &channel->completion;
500 	u32 channel_id = gsi_channel_id(channel);
501 	struct gsi *gsi = channel->gsi;
502 	struct device *dev = gsi->dev;
503 	bool timeout;
504 	u32 val;
505 
506 	/* Enable the completion interrupt for the command */
507 	gsi_irq_ch_ctrl_enable(gsi, channel_id);
508 
509 	val = u32_encode_bits(channel_id, CH_CHID_FMASK);
510 	val |= u32_encode_bits(opcode, CH_OPCODE_FMASK);
511 	timeout = !gsi_command(gsi, GSI_CH_CMD_OFFSET, val, completion);
512 
513 	gsi_irq_ch_ctrl_disable(gsi);
514 
515 	if (!timeout)
516 		return;
517 
518 	dev_err(dev, "GSI command %u for channel %u timed out, state %u\n",
519 		opcode, channel_id, gsi_channel_state(channel));
520 }
521 
522 /* Allocate GSI channel in NOT_ALLOCATED state */
523 static int gsi_channel_alloc_command(struct gsi *gsi, u32 channel_id)
524 {
525 	struct gsi_channel *channel = &gsi->channel[channel_id];
526 	struct device *dev = gsi->dev;
527 	enum gsi_channel_state state;
528 
529 	/* Get initial channel state */
530 	state = gsi_channel_state(channel);
531 	if (state != GSI_CHANNEL_STATE_NOT_ALLOCATED) {
532 		dev_err(dev, "channel %u bad state %u before alloc\n",
533 			channel_id, state);
534 		return -EINVAL;
535 	}
536 
537 	gsi_channel_command(channel, GSI_CH_ALLOCATE);
538 
539 	/* If successful the channel state will have changed */
540 	state = gsi_channel_state(channel);
541 	if (state == GSI_CHANNEL_STATE_ALLOCATED)
542 		return 0;
543 
544 	dev_err(dev, "channel %u bad state %u after alloc\n",
545 		channel_id, state);
546 
547 	return -EIO;
548 }
549 
550 /* Start an ALLOCATED channel */
551 static int gsi_channel_start_command(struct gsi_channel *channel)
552 {
553 	struct device *dev = channel->gsi->dev;
554 	enum gsi_channel_state state;
555 
556 	state = gsi_channel_state(channel);
557 	if (state != GSI_CHANNEL_STATE_ALLOCATED &&
558 	    state != GSI_CHANNEL_STATE_STOPPED) {
559 		dev_err(dev, "channel %u bad state %u before start\n",
560 			gsi_channel_id(channel), state);
561 		return -EINVAL;
562 	}
563 
564 	gsi_channel_command(channel, GSI_CH_START);
565 
566 	/* If successful the channel state will have changed */
567 	state = gsi_channel_state(channel);
568 	if (state == GSI_CHANNEL_STATE_STARTED)
569 		return 0;
570 
571 	dev_err(dev, "channel %u bad state %u after start\n",
572 		gsi_channel_id(channel), state);
573 
574 	return -EIO;
575 }
576 
577 /* Stop a GSI channel in STARTED state */
578 static int gsi_channel_stop_command(struct gsi_channel *channel)
579 {
580 	struct device *dev = channel->gsi->dev;
581 	enum gsi_channel_state state;
582 
583 	state = gsi_channel_state(channel);
584 
585 	/* Channel could have entered STOPPED state since last call
586 	 * if it timed out.  If so, we're done.
587 	 */
588 	if (state == GSI_CHANNEL_STATE_STOPPED)
589 		return 0;
590 
591 	if (state != GSI_CHANNEL_STATE_STARTED &&
592 	    state != GSI_CHANNEL_STATE_STOP_IN_PROC) {
593 		dev_err(dev, "channel %u bad state %u before stop\n",
594 			gsi_channel_id(channel), state);
595 		return -EINVAL;
596 	}
597 
598 	gsi_channel_command(channel, GSI_CH_STOP);
599 
600 	/* If successful the channel state will have changed */
601 	state = gsi_channel_state(channel);
602 	if (state == GSI_CHANNEL_STATE_STOPPED)
603 		return 0;
604 
605 	/* We may have to try again if stop is in progress */
606 	if (state == GSI_CHANNEL_STATE_STOP_IN_PROC)
607 		return -EAGAIN;
608 
609 	dev_err(dev, "channel %u bad state %u after stop\n",
610 		gsi_channel_id(channel), state);
611 
612 	return -EIO;
613 }
614 
615 /* Reset a GSI channel in ALLOCATED or ERROR state. */
616 static void gsi_channel_reset_command(struct gsi_channel *channel)
617 {
618 	struct device *dev = channel->gsi->dev;
619 	enum gsi_channel_state state;
620 
621 	/* A short delay is required before a RESET command */
622 	usleep_range(USEC_PER_MSEC, 2 * USEC_PER_MSEC);
623 
624 	state = gsi_channel_state(channel);
625 	if (state != GSI_CHANNEL_STATE_STOPPED &&
626 	    state != GSI_CHANNEL_STATE_ERROR) {
627 		/* No need to reset a channel already in ALLOCATED state */
628 		if (state != GSI_CHANNEL_STATE_ALLOCATED)
629 			dev_err(dev, "channel %u bad state %u before reset\n",
630 				gsi_channel_id(channel), state);
631 		return;
632 	}
633 
634 	gsi_channel_command(channel, GSI_CH_RESET);
635 
636 	/* If successful the channel state will have changed */
637 	state = gsi_channel_state(channel);
638 	if (state != GSI_CHANNEL_STATE_ALLOCATED)
639 		dev_err(dev, "channel %u bad state %u after reset\n",
640 			gsi_channel_id(channel), state);
641 }
642 
643 /* Deallocate an ALLOCATED GSI channel */
644 static void gsi_channel_de_alloc_command(struct gsi *gsi, u32 channel_id)
645 {
646 	struct gsi_channel *channel = &gsi->channel[channel_id];
647 	struct device *dev = gsi->dev;
648 	enum gsi_channel_state state;
649 
650 	state = gsi_channel_state(channel);
651 	if (state != GSI_CHANNEL_STATE_ALLOCATED) {
652 		dev_err(dev, "channel %u bad state %u before dealloc\n",
653 			channel_id, state);
654 		return;
655 	}
656 
657 	gsi_channel_command(channel, GSI_CH_DE_ALLOC);
658 
659 	/* If successful the channel state will have changed */
660 	state = gsi_channel_state(channel);
661 
662 	if (state != GSI_CHANNEL_STATE_NOT_ALLOCATED)
663 		dev_err(dev, "channel %u bad state %u after dealloc\n",
664 			channel_id, state);
665 }
666 
667 /* Ring an event ring doorbell, reporting the last entry processed by the AP.
668  * The index argument (modulo the ring count) is the first unfilled entry, so
669  * we supply one less than that with the doorbell.  Update the event ring
670  * index field with the value provided.
671  */
672 static void gsi_evt_ring_doorbell(struct gsi *gsi, u32 evt_ring_id, u32 index)
673 {
674 	struct gsi_ring *ring = &gsi->evt_ring[evt_ring_id].ring;
675 	u32 val;
676 
677 	ring->index = index;	/* Next unused entry */
678 
679 	/* Note: index *must* be used modulo the ring count here */
680 	val = gsi_ring_addr(ring, (index - 1) % ring->count);
681 	iowrite32(val, gsi->virt + GSI_EV_CH_E_DOORBELL_0_OFFSET(evt_ring_id));
682 }
683 
684 /* Program an event ring for use */
685 static void gsi_evt_ring_program(struct gsi *gsi, u32 evt_ring_id)
686 {
687 	struct gsi_evt_ring *evt_ring = &gsi->evt_ring[evt_ring_id];
688 	size_t size = evt_ring->ring.count * GSI_RING_ELEMENT_SIZE;
689 	u32 val;
690 
691 	/* We program all event rings as GPI type/protocol */
692 	val = u32_encode_bits(GSI_CHANNEL_TYPE_GPI, EV_CHTYPE_FMASK);
693 	val |= EV_INTYPE_FMASK;
694 	val |= u32_encode_bits(GSI_RING_ELEMENT_SIZE, EV_ELEMENT_SIZE_FMASK);
695 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_0_OFFSET(evt_ring_id));
696 
697 	val = u32_encode_bits(size, EV_R_LENGTH_FMASK);
698 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_1_OFFSET(evt_ring_id));
699 
700 	/* The context 2 and 3 registers store the low-order and
701 	 * high-order 32 bits of the address of the event ring,
702 	 * respectively.
703 	 */
704 	val = evt_ring->ring.addr & GENMASK(31, 0);
705 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_2_OFFSET(evt_ring_id));
706 
707 	val = evt_ring->ring.addr >> 32;
708 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_3_OFFSET(evt_ring_id));
709 
710 	/* Enable interrupt moderation by setting the moderation delay */
711 	val = u32_encode_bits(GSI_EVT_RING_INT_MODT, MODT_FMASK);
712 	val |= u32_encode_bits(1, MODC_FMASK);	/* comes from channel */
713 	iowrite32(val, gsi->virt + GSI_EV_CH_E_CNTXT_8_OFFSET(evt_ring_id));
714 
715 	/* No MSI write data, and MSI address high and low address is 0 */
716 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_9_OFFSET(evt_ring_id));
717 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_10_OFFSET(evt_ring_id));
718 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_11_OFFSET(evt_ring_id));
719 
720 	/* We don't need to get event read pointer updates */
721 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_12_OFFSET(evt_ring_id));
722 	iowrite32(0, gsi->virt + GSI_EV_CH_E_CNTXT_13_OFFSET(evt_ring_id));
723 
724 	/* Finally, tell the hardware we've completed event 0 (arbitrary) */
725 	gsi_evt_ring_doorbell(gsi, evt_ring_id, 0);
726 }
727 
728 /* Return the last (most recent) transaction completed on a channel. */
729 static struct gsi_trans *gsi_channel_trans_last(struct gsi_channel *channel)
730 {
731 	struct gsi_trans_info *trans_info = &channel->trans_info;
732 	struct gsi_trans *trans;
733 
734 	spin_lock_bh(&trans_info->spinlock);
735 
736 	if (!list_empty(&trans_info->complete))
737 		trans = list_last_entry(&trans_info->complete,
738 					struct gsi_trans, links);
739 	else if (!list_empty(&trans_info->polled))
740 		trans = list_last_entry(&trans_info->polled,
741 					struct gsi_trans, links);
742 	else
743 		trans = 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 /* Stop channel activity.  Transactions may not be allocated until thawed. */
768 static void gsi_channel_freeze(struct gsi_channel *channel)
769 {
770 	gsi_channel_trans_quiesce(channel);
771 
772 	napi_disable(&channel->napi);
773 
774 	gsi_irq_ieob_disable_one(channel->gsi, channel->evt_ring_id);
775 }
776 
777 /* Allow transactions to be used on the channel again. */
778 static void gsi_channel_thaw(struct gsi_channel *channel)
779 {
780 	gsi_irq_ieob_enable_one(channel->gsi, channel->evt_ring_id);
781 
782 	napi_enable(&channel->napi);
783 }
784 
785 /* Program a channel for use */
786 static void gsi_channel_program(struct gsi_channel *channel, bool doorbell)
787 {
788 	size_t size = channel->tre_ring.count * GSI_RING_ELEMENT_SIZE;
789 	u32 channel_id = gsi_channel_id(channel);
790 	union gsi_channel_scratch scr = { };
791 	struct gsi_channel_scratch_gpi *gpi;
792 	struct gsi *gsi = channel->gsi;
793 	u32 wrr_weight = 0;
794 	u32 val;
795 
796 	/* Arbitrarily pick TRE 0 as the first channel element to use */
797 	channel->tre_ring.index = 0;
798 
799 	/* We program all channels as GPI type/protocol */
800 	val = u32_encode_bits(GSI_CHANNEL_TYPE_GPI, CHTYPE_PROTOCOL_FMASK);
801 	if (channel->toward_ipa)
802 		val |= CHTYPE_DIR_FMASK;
803 	val |= u32_encode_bits(channel->evt_ring_id, ERINDEX_FMASK);
804 	val |= u32_encode_bits(GSI_RING_ELEMENT_SIZE, ELEMENT_SIZE_FMASK);
805 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_0_OFFSET(channel_id));
806 
807 	val = u32_encode_bits(size, R_LENGTH_FMASK);
808 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_1_OFFSET(channel_id));
809 
810 	/* The context 2 and 3 registers store the low-order and
811 	 * high-order 32 bits of the address of the channel ring,
812 	 * respectively.
813 	 */
814 	val = channel->tre_ring.addr & GENMASK(31, 0);
815 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_2_OFFSET(channel_id));
816 
817 	val = channel->tre_ring.addr >> 32;
818 	iowrite32(val, gsi->virt + GSI_CH_C_CNTXT_3_OFFSET(channel_id));
819 
820 	/* Command channel gets low weighted round-robin priority */
821 	if (channel->command)
822 		wrr_weight = field_max(WRR_WEIGHT_FMASK);
823 	val = u32_encode_bits(wrr_weight, WRR_WEIGHT_FMASK);
824 
825 	/* Max prefetch is 1 segment (do not set MAX_PREFETCH_FMASK) */
826 
827 	/* We enable the doorbell engine for IPA v3.5.1 */
828 	if (gsi->version == IPA_VERSION_3_5_1 && doorbell)
829 		val |= USE_DB_ENG_FMASK;
830 
831 	/* v4.0 introduces an escape buffer for prefetch.  We use it
832 	 * on all but the AP command channel.
833 	 */
834 	if (gsi->version != IPA_VERSION_3_5_1 && !channel->command) {
835 		/* If not otherwise set, prefetch buffers are used */
836 		if (gsi->version < IPA_VERSION_4_5)
837 			val |= USE_ESCAPE_BUF_ONLY_FMASK;
838 		else
839 			val |= u32_encode_bits(GSI_ESCAPE_BUF_ONLY,
840 					       PREFETCH_MODE_FMASK);
841 	}
842 
843 	iowrite32(val, gsi->virt + GSI_CH_C_QOS_OFFSET(channel_id));
844 
845 	/* Now update the scratch registers for GPI protocol */
846 	gpi = &scr.gpi;
847 	gpi->max_outstanding_tre = gsi_channel_trans_tre_max(gsi, channel_id) *
848 					GSI_RING_ELEMENT_SIZE;
849 	gpi->outstanding_threshold = 2 * GSI_RING_ELEMENT_SIZE;
850 
851 	val = scr.data.word1;
852 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_0_OFFSET(channel_id));
853 
854 	val = scr.data.word2;
855 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_1_OFFSET(channel_id));
856 
857 	val = scr.data.word3;
858 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_2_OFFSET(channel_id));
859 
860 	/* We must preserve the upper 16 bits of the last scratch register.
861 	 * The next sequence assumes those bits remain unchanged between the
862 	 * read and the write.
863 	 */
864 	val = ioread32(gsi->virt + GSI_CH_C_SCRATCH_3_OFFSET(channel_id));
865 	val = (scr.data.word4 & GENMASK(31, 16)) | (val & GENMASK(15, 0));
866 	iowrite32(val, gsi->virt + GSI_CH_C_SCRATCH_3_OFFSET(channel_id));
867 
868 	/* All done! */
869 }
870 
871 static void gsi_channel_deprogram(struct gsi_channel *channel)
872 {
873 	/* Nothing to do */
874 }
875 
876 /* Start an allocated GSI channel */
877 int gsi_channel_start(struct gsi *gsi, u32 channel_id)
878 {
879 	struct gsi_channel *channel = &gsi->channel[channel_id];
880 	int ret;
881 
882 	mutex_lock(&gsi->mutex);
883 
884 	ret = gsi_channel_start_command(channel);
885 
886 	mutex_unlock(&gsi->mutex);
887 
888 	gsi_channel_thaw(channel);
889 
890 	return ret;
891 }
892 
893 /* Stop a started channel */
894 int gsi_channel_stop(struct gsi *gsi, u32 channel_id)
895 {
896 	struct gsi_channel *channel = &gsi->channel[channel_id];
897 	u32 retries = GSI_CHANNEL_STOP_RETRIES;
898 	int ret;
899 
900 	gsi_channel_freeze(channel);
901 
902 	mutex_lock(&gsi->mutex);
903 
904 	do {
905 		ret = gsi_channel_stop_command(channel);
906 		if (ret != -EAGAIN)
907 			break;
908 		usleep_range(3 * USEC_PER_MSEC, 5 * USEC_PER_MSEC);
909 	} while (retries--);
910 
911 	mutex_unlock(&gsi->mutex);
912 
913 	/* Thaw the channel if we need to retry (or on error) */
914 	if (ret)
915 		gsi_channel_thaw(channel);
916 
917 	return ret;
918 }
919 
920 /* Reset and reconfigure a channel, (possibly) enabling the doorbell engine */
921 void gsi_channel_reset(struct gsi *gsi, u32 channel_id, bool doorbell)
922 {
923 	struct gsi_channel *channel = &gsi->channel[channel_id];
924 
925 	mutex_lock(&gsi->mutex);
926 
927 	gsi_channel_reset_command(channel);
928 	/* Due to a hardware quirk we may need to reset RX channels twice. */
929 	if (gsi->version == IPA_VERSION_3_5_1 && !channel->toward_ipa)
930 		gsi_channel_reset_command(channel);
931 
932 	gsi_channel_program(channel, doorbell);
933 	gsi_channel_trans_cancel_pending(channel);
934 
935 	mutex_unlock(&gsi->mutex);
936 }
937 
938 /* Stop a STARTED channel for suspend (using stop if requested) */
939 int gsi_channel_suspend(struct gsi *gsi, u32 channel_id, bool stop)
940 {
941 	struct gsi_channel *channel = &gsi->channel[channel_id];
942 
943 	if (stop)
944 		return gsi_channel_stop(gsi, channel_id);
945 
946 	gsi_channel_freeze(channel);
947 
948 	return 0;
949 }
950 
951 /* Resume a suspended channel (starting will be requested if STOPPED) */
952 int gsi_channel_resume(struct gsi *gsi, u32 channel_id, bool start)
953 {
954 	struct gsi_channel *channel = &gsi->channel[channel_id];
955 
956 	if (start)
957 		return gsi_channel_start(gsi, channel_id);
958 
959 	gsi_channel_thaw(channel);
960 
961 	return 0;
962 }
963 
964 /**
965  * gsi_channel_tx_queued() - Report queued TX transfers for a channel
966  * @channel:	Channel for which to report
967  *
968  * Report to the network stack the number of bytes and transactions that
969  * have been queued to hardware since last call.  This and the next function
970  * supply information used by the network stack for throttling.
971  *
972  * For each channel we track the number of transactions used and bytes of
973  * data those transactions represent.  We also track what those values are
974  * each time this function is called.  Subtracting the two tells us
975  * the number of bytes and transactions that have been added between
976  * successive calls.
977  *
978  * Calling this each time we ring the channel doorbell allows us to
979  * provide accurate information to the network stack about how much
980  * work we've given the hardware at any point in time.
981  */
982 void gsi_channel_tx_queued(struct gsi_channel *channel)
983 {
984 	u32 trans_count;
985 	u32 byte_count;
986 
987 	byte_count = channel->byte_count - channel->queued_byte_count;
988 	trans_count = channel->trans_count - channel->queued_trans_count;
989 	channel->queued_byte_count = channel->byte_count;
990 	channel->queued_trans_count = channel->trans_count;
991 
992 	ipa_gsi_channel_tx_queued(channel->gsi, gsi_channel_id(channel),
993 				  trans_count, byte_count);
994 }
995 
996 /**
997  * gsi_channel_tx_update() - Report completed TX transfers
998  * @channel:	Channel that has completed transmitting packets
999  * @trans:	Last transation known to be complete
1000  *
1001  * Compute the number of transactions and bytes that have been transferred
1002  * over a TX channel since the given transaction was committed.  Report this
1003  * information to the network stack.
1004  *
1005  * At the time a transaction is committed, we record its channel's
1006  * committed transaction and byte counts *in the transaction*.
1007  * Completions are signaled by the hardware with an interrupt, and
1008  * we can determine the latest completed transaction at that time.
1009  *
1010  * The difference between the byte/transaction count recorded in
1011  * the transaction and the count last time we recorded a completion
1012  * tells us exactly how much data has been transferred between
1013  * completions.
1014  *
1015  * Calling this each time we learn of a newly-completed transaction
1016  * allows us to provide accurate information to the network stack
1017  * about how much work has been completed by the hardware at a given
1018  * point in time.
1019  */
1020 static void
1021 gsi_channel_tx_update(struct gsi_channel *channel, struct gsi_trans *trans)
1022 {
1023 	u64 byte_count = trans->byte_count + trans->len;
1024 	u64 trans_count = trans->trans_count + 1;
1025 
1026 	byte_count -= channel->compl_byte_count;
1027 	channel->compl_byte_count += byte_count;
1028 	trans_count -= channel->compl_trans_count;
1029 	channel->compl_trans_count += trans_count;
1030 
1031 	ipa_gsi_channel_tx_completed(channel->gsi, gsi_channel_id(channel),
1032 				     trans_count, byte_count);
1033 }
1034 
1035 /* Channel control interrupt handler */
1036 static void gsi_isr_chan_ctrl(struct gsi *gsi)
1037 {
1038 	u32 channel_mask;
1039 
1040 	channel_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_CH_IRQ_OFFSET);
1041 	iowrite32(channel_mask, gsi->virt + GSI_CNTXT_SRC_CH_IRQ_CLR_OFFSET);
1042 
1043 	while (channel_mask) {
1044 		u32 channel_id = __ffs(channel_mask);
1045 		struct gsi_channel *channel;
1046 
1047 		channel_mask ^= BIT(channel_id);
1048 
1049 		channel = &gsi->channel[channel_id];
1050 
1051 		complete(&channel->completion);
1052 	}
1053 }
1054 
1055 /* Event ring control interrupt handler */
1056 static void gsi_isr_evt_ctrl(struct gsi *gsi)
1057 {
1058 	u32 event_mask;
1059 
1060 	event_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_OFFSET);
1061 	iowrite32(event_mask, gsi->virt + GSI_CNTXT_SRC_EV_CH_IRQ_CLR_OFFSET);
1062 
1063 	while (event_mask) {
1064 		u32 evt_ring_id = __ffs(event_mask);
1065 		struct gsi_evt_ring *evt_ring;
1066 
1067 		event_mask ^= BIT(evt_ring_id);
1068 
1069 		evt_ring = &gsi->evt_ring[evt_ring_id];
1070 		evt_ring->state = gsi_evt_ring_state(gsi, evt_ring_id);
1071 
1072 		complete(&evt_ring->completion);
1073 	}
1074 }
1075 
1076 /* Global channel error interrupt handler */
1077 static void
1078 gsi_isr_glob_chan_err(struct gsi *gsi, u32 err_ee, u32 channel_id, u32 code)
1079 {
1080 	if (code == GSI_OUT_OF_RESOURCES) {
1081 		dev_err(gsi->dev, "channel %u out of resources\n", channel_id);
1082 		complete(&gsi->channel[channel_id].completion);
1083 		return;
1084 	}
1085 
1086 	/* Report, but otherwise ignore all other error codes */
1087 	dev_err(gsi->dev, "channel %u global error ee 0x%08x code 0x%08x\n",
1088 		channel_id, err_ee, code);
1089 }
1090 
1091 /* Global event error interrupt handler */
1092 static void
1093 gsi_isr_glob_evt_err(struct gsi *gsi, u32 err_ee, u32 evt_ring_id, u32 code)
1094 {
1095 	if (code == GSI_OUT_OF_RESOURCES) {
1096 		struct gsi_evt_ring *evt_ring = &gsi->evt_ring[evt_ring_id];
1097 		u32 channel_id = gsi_channel_id(evt_ring->channel);
1098 
1099 		complete(&evt_ring->completion);
1100 		dev_err(gsi->dev, "evt_ring for channel %u out of resources\n",
1101 			channel_id);
1102 		return;
1103 	}
1104 
1105 	/* Report, but otherwise ignore all other error codes */
1106 	dev_err(gsi->dev, "event ring %u global error ee %u code 0x%08x\n",
1107 		evt_ring_id, err_ee, code);
1108 }
1109 
1110 /* Global error interrupt handler */
1111 static void gsi_isr_glob_err(struct gsi *gsi)
1112 {
1113 	enum gsi_err_type type;
1114 	enum gsi_err_code code;
1115 	u32 which;
1116 	u32 val;
1117 	u32 ee;
1118 
1119 	/* Get the logged error, then reinitialize the log */
1120 	val = ioread32(gsi->virt + GSI_ERROR_LOG_OFFSET);
1121 	iowrite32(0, gsi->virt + GSI_ERROR_LOG_OFFSET);
1122 	iowrite32(~0, gsi->virt + GSI_ERROR_LOG_CLR_OFFSET);
1123 
1124 	ee = u32_get_bits(val, ERR_EE_FMASK);
1125 	type = u32_get_bits(val, ERR_TYPE_FMASK);
1126 	which = u32_get_bits(val, ERR_VIRT_IDX_FMASK);
1127 	code = u32_get_bits(val, ERR_CODE_FMASK);
1128 
1129 	if (type == GSI_ERR_TYPE_CHAN)
1130 		gsi_isr_glob_chan_err(gsi, ee, which, code);
1131 	else if (type == GSI_ERR_TYPE_EVT)
1132 		gsi_isr_glob_evt_err(gsi, ee, which, code);
1133 	else	/* type GSI_ERR_TYPE_GLOB should be fatal */
1134 		dev_err(gsi->dev, "unexpected global error 0x%08x\n", type);
1135 }
1136 
1137 /* Generic EE interrupt handler */
1138 static void gsi_isr_gp_int1(struct gsi *gsi)
1139 {
1140 	u32 result;
1141 	u32 val;
1142 
1143 	/* This interrupt is used to handle completions of the two GENERIC
1144 	 * GSI commands.  We use these to allocate and halt channels on
1145 	 * the modem's behalf due to a hardware quirk on IPA v4.2.  Once
1146 	 * allocated, the modem "owns" these channels, and as a result we
1147 	 * have no way of knowing the channel's state at any given time.
1148 	 *
1149 	 * It is recommended that we halt the modem channels we allocated
1150 	 * when shutting down, but it's possible the channel isn't running
1151 	 * at the time we issue the HALT command.  We'll get an error in
1152 	 * that case, but it's harmless (the channel is already halted).
1153 	 *
1154 	 * For this reason, we silently ignore a CHANNEL_NOT_RUNNING error
1155 	 * if we receive it.
1156 	 */
1157 	val = ioread32(gsi->virt + GSI_CNTXT_SCRATCH_0_OFFSET);
1158 	result = u32_get_bits(val, GENERIC_EE_RESULT_FMASK);
1159 
1160 	switch (result) {
1161 	case GENERIC_EE_SUCCESS:
1162 	case GENERIC_EE_CHANNEL_NOT_RUNNING:
1163 		gsi->result = 0;
1164 		break;
1165 
1166 	case GENERIC_EE_RETRY:
1167 		gsi->result = -EAGAIN;
1168 		break;
1169 
1170 	default:
1171 		dev_err(gsi->dev, "global INT1 generic result %u\n", result);
1172 		gsi->result = -EIO;
1173 		break;
1174 	}
1175 
1176 	complete(&gsi->completion);
1177 }
1178 
1179 /* Inter-EE interrupt handler */
1180 static void gsi_isr_glob_ee(struct gsi *gsi)
1181 {
1182 	u32 val;
1183 
1184 	val = ioread32(gsi->virt + GSI_CNTXT_GLOB_IRQ_STTS_OFFSET);
1185 
1186 	if (val & BIT(ERROR_INT))
1187 		gsi_isr_glob_err(gsi);
1188 
1189 	iowrite32(val, gsi->virt + GSI_CNTXT_GLOB_IRQ_CLR_OFFSET);
1190 
1191 	val &= ~BIT(ERROR_INT);
1192 
1193 	if (val & BIT(GP_INT1)) {
1194 		val ^= BIT(GP_INT1);
1195 		gsi_isr_gp_int1(gsi);
1196 	}
1197 
1198 	if (val)
1199 		dev_err(gsi->dev, "unexpected global interrupt 0x%08x\n", val);
1200 }
1201 
1202 /* I/O completion interrupt event */
1203 static void gsi_isr_ieob(struct gsi *gsi)
1204 {
1205 	u32 event_mask;
1206 
1207 	event_mask = ioread32(gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_OFFSET);
1208 	gsi_irq_ieob_disable(gsi, event_mask);
1209 	iowrite32(event_mask, gsi->virt + GSI_CNTXT_SRC_IEOB_IRQ_CLR_OFFSET);
1210 
1211 	while (event_mask) {
1212 		u32 evt_ring_id = __ffs(event_mask);
1213 
1214 		event_mask ^= BIT(evt_ring_id);
1215 
1216 		napi_schedule(&gsi->evt_ring[evt_ring_id].channel->napi);
1217 	}
1218 }
1219 
1220 /* General event interrupts represent serious problems, so report them */
1221 static void gsi_isr_general(struct gsi *gsi)
1222 {
1223 	struct device *dev = gsi->dev;
1224 	u32 val;
1225 
1226 	val = ioread32(gsi->virt + GSI_CNTXT_GSI_IRQ_STTS_OFFSET);
1227 	iowrite32(val, gsi->virt + GSI_CNTXT_GSI_IRQ_CLR_OFFSET);
1228 
1229 	dev_err(dev, "unexpected general interrupt 0x%08x\n", val);
1230 }
1231 
1232 /**
1233  * gsi_isr() - Top level GSI interrupt service routine
1234  * @irq:	Interrupt number (ignored)
1235  * @dev_id:	GSI pointer supplied to request_irq()
1236  *
1237  * This is the main handler function registered for the GSI IRQ. Each type
1238  * of interrupt has a separate handler function that is called from here.
1239  */
1240 static irqreturn_t gsi_isr(int irq, void *dev_id)
1241 {
1242 	struct gsi *gsi = dev_id;
1243 	u32 intr_mask;
1244 	u32 cnt = 0;
1245 
1246 	/* enum gsi_irq_type_id defines GSI interrupt types */
1247 	while ((intr_mask = ioread32(gsi->virt + GSI_CNTXT_TYPE_IRQ_OFFSET))) {
1248 		/* intr_mask contains bitmask of pending GSI interrupts */
1249 		do {
1250 			u32 gsi_intr = BIT(__ffs(intr_mask));
1251 
1252 			intr_mask ^= gsi_intr;
1253 
1254 			switch (gsi_intr) {
1255 			case BIT(GSI_CH_CTRL):
1256 				gsi_isr_chan_ctrl(gsi);
1257 				break;
1258 			case BIT(GSI_EV_CTRL):
1259 				gsi_isr_evt_ctrl(gsi);
1260 				break;
1261 			case BIT(GSI_GLOB_EE):
1262 				gsi_isr_glob_ee(gsi);
1263 				break;
1264 			case BIT(GSI_IEOB):
1265 				gsi_isr_ieob(gsi);
1266 				break;
1267 			case BIT(GSI_GENERAL):
1268 				gsi_isr_general(gsi);
1269 				break;
1270 			default:
1271 				dev_err(gsi->dev,
1272 					"unrecognized interrupt type 0x%08x\n",
1273 					gsi_intr);
1274 				break;
1275 			}
1276 		} while (intr_mask);
1277 
1278 		if (++cnt > GSI_ISR_MAX_ITER) {
1279 			dev_err(gsi->dev, "interrupt flood\n");
1280 			break;
1281 		}
1282 	}
1283 
1284 	return IRQ_HANDLED;
1285 }
1286 
1287 static int gsi_irq_init(struct gsi *gsi, struct platform_device *pdev)
1288 {
1289 	struct device *dev = &pdev->dev;
1290 	unsigned int irq;
1291 	int ret;
1292 
1293 	ret = platform_get_irq_byname(pdev, "gsi");
1294 	if (ret <= 0) {
1295 		dev_err(dev, "DT error %d getting \"gsi\" IRQ property\n", ret);
1296 		return ret ? : -EINVAL;
1297 	}
1298 	irq = ret;
1299 
1300 	ret = request_irq(irq, gsi_isr, 0, "gsi", gsi);
1301 	if (ret) {
1302 		dev_err(dev, "error %d requesting \"gsi\" IRQ\n", ret);
1303 		return ret;
1304 	}
1305 	gsi->irq = irq;
1306 
1307 	return 0;
1308 }
1309 
1310 static void gsi_irq_exit(struct gsi *gsi)
1311 {
1312 	free_irq(gsi->irq, gsi);
1313 }
1314 
1315 /* Return the transaction associated with a transfer completion event */
1316 static struct gsi_trans *gsi_event_trans(struct gsi_channel *channel,
1317 					 struct gsi_event *event)
1318 {
1319 	u32 tre_offset;
1320 	u32 tre_index;
1321 
1322 	/* Event xfer_ptr records the TRE it's associated with */
1323 	tre_offset = le64_to_cpu(event->xfer_ptr) & GENMASK(31, 0);
1324 	tre_index = gsi_ring_index(&channel->tre_ring, tre_offset);
1325 
1326 	return gsi_channel_trans_mapped(channel, tre_index);
1327 }
1328 
1329 /**
1330  * gsi_evt_ring_rx_update() - Record lengths of received data
1331  * @evt_ring:	Event ring associated with channel that received packets
1332  * @index:	Event index in ring reported by hardware
1333  *
1334  * Events for RX channels contain the actual number of bytes received into
1335  * the buffer.  Every event has a transaction associated with it, and here
1336  * we update transactions to record their actual received lengths.
1337  *
1338  * This function is called whenever we learn that the GSI hardware has filled
1339  * new events since the last time we checked.  The ring's index field tells
1340  * the first entry in need of processing.  The index provided is the
1341  * first *unfilled* event in the ring (following the last filled one).
1342  *
1343  * Events are sequential within the event ring, and transactions are
1344  * sequential within the transaction pool.
1345  *
1346  * Note that @index always refers to an element *within* the event ring.
1347  */
1348 static void gsi_evt_ring_rx_update(struct gsi_evt_ring *evt_ring, u32 index)
1349 {
1350 	struct gsi_channel *channel = evt_ring->channel;
1351 	struct gsi_ring *ring = &evt_ring->ring;
1352 	struct gsi_trans_info *trans_info;
1353 	struct gsi_event *event_done;
1354 	struct gsi_event *event;
1355 	struct gsi_trans *trans;
1356 	u32 byte_count = 0;
1357 	u32 old_index;
1358 	u32 event_avail;
1359 
1360 	trans_info = &channel->trans_info;
1361 
1362 	/* We'll start with the oldest un-processed event.  RX channels
1363 	 * replenish receive buffers in single-TRE transactions, so we
1364 	 * can just map that event to its transaction.  Transactions
1365 	 * associated with completion events are consecutive.
1366 	 */
1367 	old_index = ring->index;
1368 	event = gsi_ring_virt(ring, old_index);
1369 	trans = gsi_event_trans(channel, event);
1370 
1371 	/* Compute the number of events to process before we wrap,
1372 	 * and determine when we'll be done processing events.
1373 	 */
1374 	event_avail = ring->count - old_index % ring->count;
1375 	event_done = gsi_ring_virt(ring, index);
1376 	do {
1377 		trans->len = __le16_to_cpu(event->len);
1378 		byte_count += trans->len;
1379 
1380 		/* Move on to the next event and transaction */
1381 		if (--event_avail)
1382 			event++;
1383 		else
1384 			event = gsi_ring_virt(ring, 0);
1385 		trans = gsi_trans_pool_next(&trans_info->pool, trans);
1386 	} while (event != event_done);
1387 
1388 	/* We record RX bytes when they are received */
1389 	channel->byte_count += byte_count;
1390 	channel->trans_count++;
1391 }
1392 
1393 /* Initialize a ring, including allocating DMA memory for its entries */
1394 static int gsi_ring_alloc(struct gsi *gsi, struct gsi_ring *ring, u32 count)
1395 {
1396 	size_t size = count * GSI_RING_ELEMENT_SIZE;
1397 	struct device *dev = gsi->dev;
1398 	dma_addr_t addr;
1399 
1400 	/* Hardware requires a 2^n ring size, with alignment equal to size */
1401 	ring->virt = dma_alloc_coherent(dev, size, &addr, GFP_KERNEL);
1402 	if (ring->virt && addr % size) {
1403 		dma_free_coherent(dev, size, ring->virt, ring->addr);
1404 		dev_err(dev, "unable to alloc 0x%zx-aligned ring buffer\n",
1405 			size);
1406 		return -EINVAL;	/* Not a good error value, but distinct */
1407 	} else if (!ring->virt) {
1408 		return -ENOMEM;
1409 	}
1410 	ring->addr = addr;
1411 	ring->count = count;
1412 
1413 	return 0;
1414 }
1415 
1416 /* Free a previously-allocated ring */
1417 static void gsi_ring_free(struct gsi *gsi, struct gsi_ring *ring)
1418 {
1419 	size_t size = ring->count * GSI_RING_ELEMENT_SIZE;
1420 
1421 	dma_free_coherent(gsi->dev, size, ring->virt, ring->addr);
1422 }
1423 
1424 /* Allocate an available event ring id */
1425 static int gsi_evt_ring_id_alloc(struct gsi *gsi)
1426 {
1427 	u32 evt_ring_id;
1428 
1429 	if (gsi->event_bitmap == ~0U) {
1430 		dev_err(gsi->dev, "event rings exhausted\n");
1431 		return -ENOSPC;
1432 	}
1433 
1434 	evt_ring_id = ffz(gsi->event_bitmap);
1435 	gsi->event_bitmap |= BIT(evt_ring_id);
1436 
1437 	return (int)evt_ring_id;
1438 }
1439 
1440 /* Free a previously-allocated event ring id */
1441 static void gsi_evt_ring_id_free(struct gsi *gsi, u32 evt_ring_id)
1442 {
1443 	gsi->event_bitmap &= ~BIT(evt_ring_id);
1444 }
1445 
1446 /* Ring a channel doorbell, reporting the first un-filled entry */
1447 void gsi_channel_doorbell(struct gsi_channel *channel)
1448 {
1449 	struct gsi_ring *tre_ring = &channel->tre_ring;
1450 	u32 channel_id = gsi_channel_id(channel);
1451 	struct gsi *gsi = channel->gsi;
1452 	u32 val;
1453 
1454 	/* Note: index *must* be used modulo the ring count here */
1455 	val = gsi_ring_addr(tre_ring, tre_ring->index % tre_ring->count);
1456 	iowrite32(val, gsi->virt + GSI_CH_C_DOORBELL_0_OFFSET(channel_id));
1457 }
1458 
1459 /* Consult hardware, move any newly completed transactions to completed list */
1460 static struct gsi_trans *gsi_channel_update(struct gsi_channel *channel)
1461 {
1462 	u32 evt_ring_id = channel->evt_ring_id;
1463 	struct gsi *gsi = channel->gsi;
1464 	struct gsi_evt_ring *evt_ring;
1465 	struct gsi_trans *trans;
1466 	struct gsi_ring *ring;
1467 	u32 offset;
1468 	u32 index;
1469 
1470 	evt_ring = &gsi->evt_ring[evt_ring_id];
1471 	ring = &evt_ring->ring;
1472 
1473 	/* See if there's anything new to process; if not, we're done.  Note
1474 	 * that index always refers to an entry *within* the event ring.
1475 	 */
1476 	offset = GSI_EV_CH_E_CNTXT_4_OFFSET(evt_ring_id);
1477 	index = gsi_ring_index(ring, ioread32(gsi->virt + offset));
1478 	if (index == ring->index % ring->count)
1479 		return NULL;
1480 
1481 	/* Get the transaction for the latest completed event.  Take a
1482 	 * reference to keep it from completing before we give the events
1483 	 * for this and previous transactions back to the hardware.
1484 	 */
1485 	trans = gsi_event_trans(channel, gsi_ring_virt(ring, index - 1));
1486 	refcount_inc(&trans->refcount);
1487 
1488 	/* For RX channels, update each completed transaction with the number
1489 	 * of bytes that were actually received.  For TX channels, report
1490 	 * the number of transactions and bytes this completion represents
1491 	 * up the network stack.
1492 	 */
1493 	if (channel->toward_ipa)
1494 		gsi_channel_tx_update(channel, trans);
1495 	else
1496 		gsi_evt_ring_rx_update(evt_ring, index);
1497 
1498 	gsi_trans_move_complete(trans);
1499 
1500 	/* Tell the hardware we've handled these events */
1501 	gsi_evt_ring_doorbell(channel->gsi, channel->evt_ring_id, index);
1502 
1503 	gsi_trans_free(trans);
1504 
1505 	return gsi_channel_trans_complete(channel);
1506 }
1507 
1508 /**
1509  * gsi_channel_poll_one() - Return a single completed transaction on a channel
1510  * @channel:	Channel to be polled
1511  *
1512  * Return:	Transaction pointer, or null if none are available
1513  *
1514  * This function returns the first entry on a channel's completed transaction
1515  * list.  If that list is empty, the hardware is consulted to determine
1516  * whether any new transactions have completed.  If so, they're moved to the
1517  * completed list and the new first entry is returned.  If there are no more
1518  * completed transactions, a null pointer is returned.
1519  */
1520 static struct gsi_trans *gsi_channel_poll_one(struct gsi_channel *channel)
1521 {
1522 	struct gsi_trans *trans;
1523 
1524 	/* Get the first transaction from the completed list */
1525 	trans = gsi_channel_trans_complete(channel);
1526 	if (!trans)	/* List is empty; see if there's more to do */
1527 		trans = gsi_channel_update(channel);
1528 
1529 	if (trans)
1530 		gsi_trans_move_polled(trans);
1531 
1532 	return trans;
1533 }
1534 
1535 /**
1536  * gsi_channel_poll() - NAPI poll function for a channel
1537  * @napi:	NAPI structure for the channel
1538  * @budget:	Budget supplied by NAPI core
1539  *
1540  * Return:	Number of items polled (<= budget)
1541  *
1542  * Single transactions completed by hardware are polled until either
1543  * the budget is exhausted, or there are no more.  Each transaction
1544  * polled is passed to gsi_trans_complete(), to perform remaining
1545  * completion processing and retire/free the transaction.
1546  */
1547 static int gsi_channel_poll(struct napi_struct *napi, int budget)
1548 {
1549 	struct gsi_channel *channel;
1550 	int count;
1551 
1552 	channel = container_of(napi, struct gsi_channel, napi);
1553 	for (count = 0; count < budget; count++) {
1554 		struct gsi_trans *trans;
1555 
1556 		trans = gsi_channel_poll_one(channel);
1557 		if (!trans)
1558 			break;
1559 		gsi_trans_complete(trans);
1560 	}
1561 
1562 	if (count < budget && napi_complete(napi))
1563 		gsi_irq_ieob_enable_one(channel->gsi, channel->evt_ring_id);
1564 
1565 	return count;
1566 }
1567 
1568 /* The event bitmap represents which event ids are available for allocation.
1569  * Set bits are not available, clear bits can be used.  This function
1570  * initializes the map so all events supported by the hardware are available,
1571  * then precludes any reserved events from being allocated.
1572  */
1573 static u32 gsi_event_bitmap_init(u32 evt_ring_max)
1574 {
1575 	u32 event_bitmap = GENMASK(BITS_PER_LONG - 1, evt_ring_max);
1576 
1577 	event_bitmap |= GENMASK(GSI_MHI_EVENT_ID_END, GSI_MHI_EVENT_ID_START);
1578 
1579 	return event_bitmap;
1580 }
1581 
1582 /* Setup function for event rings */
1583 static void gsi_evt_ring_setup(struct gsi *gsi)
1584 {
1585 	/* Nothing to do */
1586 }
1587 
1588 /* Inverse of gsi_evt_ring_setup() */
1589 static void gsi_evt_ring_teardown(struct gsi *gsi)
1590 {
1591 	/* Nothing to do */
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 (!channel->gsi)
1602 		return 0;	/* Ignore uninitialized channels */
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 (!channel->gsi)
1639 		return;		/* Ignore uninitialized channels */
1640 
1641 	netif_napi_del(&channel->napi);
1642 
1643 	gsi_channel_deprogram(channel);
1644 	gsi_channel_de_alloc_command(gsi, channel_id);
1645 	gsi_evt_ring_reset_command(gsi, evt_ring_id);
1646 	gsi_evt_ring_de_alloc_command(gsi, evt_ring_id);
1647 }
1648 
1649 static int gsi_generic_command(struct gsi *gsi, u32 channel_id,
1650 			       enum gsi_generic_cmd_opcode opcode)
1651 {
1652 	struct completion *completion = &gsi->completion;
1653 	bool timeout;
1654 	u32 val;
1655 
1656 	/* The error global interrupt type is always enabled (until we
1657 	 * teardown), so we won't change that.  A generic EE command
1658 	 * completes with a GSI global interrupt of type GP_INT1.  We
1659 	 * only perform one generic command at a time (to allocate or
1660 	 * halt a modem channel) and only from this function.  So we
1661 	 * enable the GP_INT1 IRQ type here while we're expecting it.
1662 	 */
1663 	val = BIT(ERROR_INT) | BIT(GP_INT1);
1664 	iowrite32(val, gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
1665 
1666 	/* First zero the result code field */
1667 	val = ioread32(gsi->virt + GSI_CNTXT_SCRATCH_0_OFFSET);
1668 	val &= ~GENERIC_EE_RESULT_FMASK;
1669 	iowrite32(val, gsi->virt + GSI_CNTXT_SCRATCH_0_OFFSET);
1670 
1671 	/* Now issue the command */
1672 	val = u32_encode_bits(opcode, GENERIC_OPCODE_FMASK);
1673 	val |= u32_encode_bits(channel_id, GENERIC_CHID_FMASK);
1674 	val |= u32_encode_bits(GSI_EE_MODEM, GENERIC_EE_FMASK);
1675 
1676 	timeout = !gsi_command(gsi, GSI_GENERIC_CMD_OFFSET, val, completion);
1677 
1678 	/* Disable the GP_INT1 IRQ type again */
1679 	iowrite32(BIT(ERROR_INT), gsi->virt + GSI_CNTXT_GLOB_IRQ_EN_OFFSET);
1680 
1681 	if (!timeout)
1682 		return gsi->result;
1683 
1684 	dev_err(gsi->dev, "GSI generic command %u to channel %u timed out\n",
1685 		opcode, channel_id);
1686 
1687 	return -ETIMEDOUT;
1688 }
1689 
1690 static int gsi_modem_channel_alloc(struct gsi *gsi, u32 channel_id)
1691 {
1692 	return gsi_generic_command(gsi, channel_id,
1693 				   GSI_GENERIC_ALLOCATE_CHANNEL);
1694 }
1695 
1696 static void gsi_modem_channel_halt(struct gsi *gsi, u32 channel_id)
1697 {
1698 	u32 retries = GSI_CHANNEL_MODEM_HALT_RETRIES;
1699 	int ret;
1700 
1701 	do
1702 		ret = gsi_generic_command(gsi, channel_id,
1703 					  GSI_GENERIC_HALT_CHANNEL);
1704 	while (ret == -EAGAIN && retries--);
1705 
1706 	if (ret)
1707 		dev_err(gsi->dev, "error %d halting modem channel %u\n",
1708 			ret, channel_id);
1709 }
1710 
1711 /* Setup function for channels */
1712 static int gsi_channel_setup(struct gsi *gsi)
1713 {
1714 	u32 channel_id = 0;
1715 	u32 mask;
1716 	int ret;
1717 
1718 	gsi_evt_ring_setup(gsi);
1719 	gsi_irq_enable(gsi);
1720 
1721 	mutex_lock(&gsi->mutex);
1722 
1723 	do {
1724 		ret = gsi_channel_setup_one(gsi, channel_id);
1725 		if (ret)
1726 			goto err_unwind;
1727 	} while (++channel_id < gsi->channel_count);
1728 
1729 	/* Make sure no channels were defined that hardware does not support */
1730 	while (channel_id < GSI_CHANNEL_COUNT_MAX) {
1731 		struct gsi_channel *channel = &gsi->channel[channel_id++];
1732 
1733 		if (!channel->gsi)
1734 			continue;	/* Ignore uninitialized channels */
1735 
1736 		dev_err(gsi->dev, "channel %u not supported by hardware\n",
1737 			channel_id - 1);
1738 		channel_id = gsi->channel_count;
1739 		goto err_unwind;
1740 	}
1741 
1742 	/* Allocate modem channels if necessary */
1743 	mask = gsi->modem_channel_bitmap;
1744 	while (mask) {
1745 		u32 modem_channel_id = __ffs(mask);
1746 
1747 		ret = gsi_modem_channel_alloc(gsi, modem_channel_id);
1748 		if (ret)
1749 			goto err_unwind_modem;
1750 
1751 		/* Clear bit from mask only after success (for unwind) */
1752 		mask ^= BIT(modem_channel_id);
1753 	}
1754 
1755 	mutex_unlock(&gsi->mutex);
1756 
1757 	return 0;
1758 
1759 err_unwind_modem:
1760 	/* Compute which modem channels need to be deallocated */
1761 	mask ^= gsi->modem_channel_bitmap;
1762 	while (mask) {
1763 		channel_id = __fls(mask);
1764 
1765 		mask ^= BIT(channel_id);
1766 
1767 		gsi_modem_channel_halt(gsi, channel_id);
1768 	}
1769 
1770 err_unwind:
1771 	while (channel_id--)
1772 		gsi_channel_teardown_one(gsi, channel_id);
1773 
1774 	mutex_unlock(&gsi->mutex);
1775 
1776 	gsi_irq_disable(gsi);
1777 	gsi_evt_ring_teardown(gsi);
1778 
1779 	return ret;
1780 }
1781 
1782 /* Inverse of gsi_channel_setup() */
1783 static void gsi_channel_teardown(struct gsi *gsi)
1784 {
1785 	u32 mask = gsi->modem_channel_bitmap;
1786 	u32 channel_id;
1787 
1788 	mutex_lock(&gsi->mutex);
1789 
1790 	while (mask) {
1791 		channel_id = __fls(mask);
1792 
1793 		mask ^= BIT(channel_id);
1794 
1795 		gsi_modem_channel_halt(gsi, channel_id);
1796 	}
1797 
1798 	channel_id = gsi->channel_count - 1;
1799 	do
1800 		gsi_channel_teardown_one(gsi, channel_id);
1801 	while (channel_id--);
1802 
1803 	mutex_unlock(&gsi->mutex);
1804 
1805 	gsi_irq_disable(gsi);
1806 	gsi_evt_ring_teardown(gsi);
1807 }
1808 
1809 /* Setup function for GSI.  GSI firmware must be loaded and initialized */
1810 int gsi_setup(struct gsi *gsi)
1811 {
1812 	struct device *dev = gsi->dev;
1813 	u32 val;
1814 	int ret;
1815 
1816 	/* Here is where we first touch the GSI hardware */
1817 	val = ioread32(gsi->virt + GSI_GSI_STATUS_OFFSET);
1818 	if (!(val & ENABLED_FMASK)) {
1819 		dev_err(dev, "GSI has not been enabled\n");
1820 		return -EIO;
1821 	}
1822 
1823 	gsi_irq_setup(gsi);
1824 
1825 	val = ioread32(gsi->virt + GSI_GSI_HW_PARAM_2_OFFSET);
1826 
1827 	gsi->channel_count = u32_get_bits(val, NUM_CH_PER_EE_FMASK);
1828 	if (!gsi->channel_count) {
1829 		dev_err(dev, "GSI reports zero channels supported\n");
1830 		return -EINVAL;
1831 	}
1832 	if (gsi->channel_count > GSI_CHANNEL_COUNT_MAX) {
1833 		dev_warn(dev,
1834 			 "limiting to %u channels; hardware supports %u\n",
1835 			 GSI_CHANNEL_COUNT_MAX, gsi->channel_count);
1836 		gsi->channel_count = GSI_CHANNEL_COUNT_MAX;
1837 	}
1838 
1839 	gsi->evt_ring_count = u32_get_bits(val, NUM_EV_PER_EE_FMASK);
1840 	if (!gsi->evt_ring_count) {
1841 		dev_err(dev, "GSI reports zero event rings supported\n");
1842 		return -EINVAL;
1843 	}
1844 	if (gsi->evt_ring_count > GSI_EVT_RING_COUNT_MAX) {
1845 		dev_warn(dev,
1846 			 "limiting to %u event rings; hardware supports %u\n",
1847 			 GSI_EVT_RING_COUNT_MAX, gsi->evt_ring_count);
1848 		gsi->evt_ring_count = GSI_EVT_RING_COUNT_MAX;
1849 	}
1850 
1851 	/* Initialize the error log */
1852 	iowrite32(0, gsi->virt + GSI_ERROR_LOG_OFFSET);
1853 
1854 	/* Writing 1 indicates IRQ interrupts; 0 would be MSI */
1855 	iowrite32(1, gsi->virt + GSI_CNTXT_INTSET_OFFSET);
1856 
1857 	ret = gsi_channel_setup(gsi);
1858 	if (ret)
1859 		gsi_irq_teardown(gsi);
1860 
1861 	return ret;
1862 }
1863 
1864 /* Inverse of gsi_setup() */
1865 void gsi_teardown(struct gsi *gsi)
1866 {
1867 	gsi_channel_teardown(gsi);
1868 	gsi_irq_teardown(gsi);
1869 }
1870 
1871 /* Initialize a channel's event ring */
1872 static int gsi_channel_evt_ring_init(struct gsi_channel *channel)
1873 {
1874 	struct gsi *gsi = channel->gsi;
1875 	struct gsi_evt_ring *evt_ring;
1876 	int ret;
1877 
1878 	ret = gsi_evt_ring_id_alloc(gsi);
1879 	if (ret < 0)
1880 		return ret;
1881 	channel->evt_ring_id = ret;
1882 
1883 	evt_ring = &gsi->evt_ring[channel->evt_ring_id];
1884 	evt_ring->channel = channel;
1885 
1886 	ret = gsi_ring_alloc(gsi, &evt_ring->ring, channel->event_count);
1887 	if (!ret)
1888 		return 0;	/* Success! */
1889 
1890 	dev_err(gsi->dev, "error %d allocating channel %u event ring\n",
1891 		ret, gsi_channel_id(channel));
1892 
1893 	gsi_evt_ring_id_free(gsi, channel->evt_ring_id);
1894 
1895 	return ret;
1896 }
1897 
1898 /* Inverse of gsi_channel_evt_ring_init() */
1899 static void gsi_channel_evt_ring_exit(struct gsi_channel *channel)
1900 {
1901 	u32 evt_ring_id = channel->evt_ring_id;
1902 	struct gsi *gsi = channel->gsi;
1903 	struct gsi_evt_ring *evt_ring;
1904 
1905 	evt_ring = &gsi->evt_ring[evt_ring_id];
1906 	gsi_ring_free(gsi, &evt_ring->ring);
1907 	gsi_evt_ring_id_free(gsi, evt_ring_id);
1908 }
1909 
1910 /* Init function for event rings */
1911 static void gsi_evt_ring_init(struct gsi *gsi)
1912 {
1913 	u32 evt_ring_id = 0;
1914 
1915 	gsi->event_bitmap = gsi_event_bitmap_init(GSI_EVT_RING_COUNT_MAX);
1916 	gsi->ieob_enabled_bitmap = 0;
1917 	do
1918 		init_completion(&gsi->evt_ring[evt_ring_id].completion);
1919 	while (++evt_ring_id < GSI_EVT_RING_COUNT_MAX);
1920 }
1921 
1922 /* Inverse of gsi_evt_ring_init() */
1923 static void gsi_evt_ring_exit(struct gsi *gsi)
1924 {
1925 	/* Nothing to do */
1926 }
1927 
1928 static bool gsi_channel_data_valid(struct gsi *gsi,
1929 				   const struct ipa_gsi_endpoint_data *data)
1930 {
1931 #ifdef IPA_VALIDATION
1932 	u32 channel_id = data->channel_id;
1933 	struct device *dev = gsi->dev;
1934 
1935 	/* Make sure channel ids are in the range driver supports */
1936 	if (channel_id >= GSI_CHANNEL_COUNT_MAX) {
1937 		dev_err(dev, "bad channel id %u; must be less than %u\n",
1938 			channel_id, GSI_CHANNEL_COUNT_MAX);
1939 		return false;
1940 	}
1941 
1942 	if (data->ee_id != GSI_EE_AP && data->ee_id != GSI_EE_MODEM) {
1943 		dev_err(dev, "bad EE id %u; not AP or modem\n", data->ee_id);
1944 		return false;
1945 	}
1946 
1947 	if (!data->channel.tlv_count ||
1948 	    data->channel.tlv_count > GSI_TLV_MAX) {
1949 		dev_err(dev, "channel %u bad tlv_count %u; must be 1..%u\n",
1950 			channel_id, data->channel.tlv_count, GSI_TLV_MAX);
1951 		return false;
1952 	}
1953 
1954 	/* We have to allow at least one maximally-sized transaction to
1955 	 * be outstanding (which would use tlv_count TREs).  Given how
1956 	 * gsi_channel_tre_max() is computed, tre_count has to be almost
1957 	 * twice the TLV FIFO size to satisfy this requirement.
1958 	 */
1959 	if (data->channel.tre_count < 2 * data->channel.tlv_count - 1) {
1960 		dev_err(dev, "channel %u TLV count %u exceeds TRE count %u\n",
1961 			channel_id, data->channel.tlv_count,
1962 			data->channel.tre_count);
1963 		return false;
1964 	}
1965 
1966 	if (!is_power_of_2(data->channel.tre_count)) {
1967 		dev_err(dev, "channel %u bad tre_count %u; not power of 2\n",
1968 			channel_id, data->channel.tre_count);
1969 		return false;
1970 	}
1971 
1972 	if (!is_power_of_2(data->channel.event_count)) {
1973 		dev_err(dev, "channel %u bad event_count %u; not power of 2\n",
1974 			channel_id, data->channel.event_count);
1975 		return false;
1976 	}
1977 #endif /* IPA_VALIDATION */
1978 
1979 	return true;
1980 }
1981 
1982 /* Init function for a single channel */
1983 static int gsi_channel_init_one(struct gsi *gsi,
1984 				const struct ipa_gsi_endpoint_data *data,
1985 				bool command)
1986 {
1987 	struct gsi_channel *channel;
1988 	u32 tre_count;
1989 	int ret;
1990 
1991 	if (!gsi_channel_data_valid(gsi, data))
1992 		return -EINVAL;
1993 
1994 	/* Worst case we need an event for every outstanding TRE */
1995 	if (data->channel.tre_count > data->channel.event_count) {
1996 		tre_count = data->channel.event_count;
1997 		dev_warn(gsi->dev, "channel %u limited to %u TREs\n",
1998 			 data->channel_id, tre_count);
1999 	} else {
2000 		tre_count = data->channel.tre_count;
2001 	}
2002 
2003 	channel = &gsi->channel[data->channel_id];
2004 	memset(channel, 0, sizeof(*channel));
2005 
2006 	channel->gsi = gsi;
2007 	channel->toward_ipa = data->toward_ipa;
2008 	channel->command = command;
2009 	channel->tlv_count = data->channel.tlv_count;
2010 	channel->tre_count = tre_count;
2011 	channel->event_count = data->channel.event_count;
2012 	init_completion(&channel->completion);
2013 
2014 	ret = gsi_channel_evt_ring_init(channel);
2015 	if (ret)
2016 		goto err_clear_gsi;
2017 
2018 	ret = gsi_ring_alloc(gsi, &channel->tre_ring, data->channel.tre_count);
2019 	if (ret) {
2020 		dev_err(gsi->dev, "error %d allocating channel %u ring\n",
2021 			ret, data->channel_id);
2022 		goto err_channel_evt_ring_exit;
2023 	}
2024 
2025 	ret = gsi_channel_trans_init(gsi, data->channel_id);
2026 	if (ret)
2027 		goto err_ring_free;
2028 
2029 	if (command) {
2030 		u32 tre_max = gsi_channel_tre_max(gsi, data->channel_id);
2031 
2032 		ret = ipa_cmd_pool_init(channel, tre_max);
2033 	}
2034 	if (!ret)
2035 		return 0;	/* Success! */
2036 
2037 	gsi_channel_trans_exit(channel);
2038 err_ring_free:
2039 	gsi_ring_free(gsi, &channel->tre_ring);
2040 err_channel_evt_ring_exit:
2041 	gsi_channel_evt_ring_exit(channel);
2042 err_clear_gsi:
2043 	channel->gsi = NULL;	/* Mark it not (fully) initialized */
2044 
2045 	return ret;
2046 }
2047 
2048 /* Inverse of gsi_channel_init_one() */
2049 static void gsi_channel_exit_one(struct gsi_channel *channel)
2050 {
2051 	if (!channel->gsi)
2052 		return;		/* Ignore uninitialized channels */
2053 
2054 	if (channel->command)
2055 		ipa_cmd_pool_exit(channel);
2056 	gsi_channel_trans_exit(channel);
2057 	gsi_ring_free(channel->gsi, &channel->tre_ring);
2058 	gsi_channel_evt_ring_exit(channel);
2059 }
2060 
2061 /* Init function for channels */
2062 static int gsi_channel_init(struct gsi *gsi, u32 count,
2063 			    const struct ipa_gsi_endpoint_data *data)
2064 {
2065 	bool modem_alloc;
2066 	int ret = 0;
2067 	u32 i;
2068 
2069 	/* IPA v4.2 requires the AP to allocate channels for the modem */
2070 	modem_alloc = gsi->version == IPA_VERSION_4_2;
2071 
2072 	gsi_evt_ring_init(gsi);
2073 
2074 	/* The endpoint data array is indexed by endpoint name */
2075 	for (i = 0; i < count; i++) {
2076 		bool command = i == IPA_ENDPOINT_AP_COMMAND_TX;
2077 
2078 		if (ipa_gsi_endpoint_data_empty(&data[i]))
2079 			continue;	/* Skip over empty slots */
2080 
2081 		/* Mark modem channels to be allocated (hardware workaround) */
2082 		if (data[i].ee_id == GSI_EE_MODEM) {
2083 			if (modem_alloc)
2084 				gsi->modem_channel_bitmap |=
2085 						BIT(data[i].channel_id);
2086 			continue;
2087 		}
2088 
2089 		ret = gsi_channel_init_one(gsi, &data[i], command);
2090 		if (ret)
2091 			goto err_unwind;
2092 	}
2093 
2094 	return ret;
2095 
2096 err_unwind:
2097 	while (i--) {
2098 		if (ipa_gsi_endpoint_data_empty(&data[i]))
2099 			continue;
2100 		if (modem_alloc && data[i].ee_id == GSI_EE_MODEM) {
2101 			gsi->modem_channel_bitmap &= ~BIT(data[i].channel_id);
2102 			continue;
2103 		}
2104 		gsi_channel_exit_one(&gsi->channel[data->channel_id]);
2105 	}
2106 	gsi_evt_ring_exit(gsi);
2107 
2108 	return ret;
2109 }
2110 
2111 /* Inverse of gsi_channel_init() */
2112 static void gsi_channel_exit(struct gsi *gsi)
2113 {
2114 	u32 channel_id = GSI_CHANNEL_COUNT_MAX - 1;
2115 
2116 	do
2117 		gsi_channel_exit_one(&gsi->channel[channel_id]);
2118 	while (channel_id--);
2119 	gsi->modem_channel_bitmap = 0;
2120 
2121 	gsi_evt_ring_exit(gsi);
2122 }
2123 
2124 /* Init function for GSI.  GSI hardware does not need to be "ready" */
2125 int gsi_init(struct gsi *gsi, struct platform_device *pdev,
2126 	     enum ipa_version version, u32 count,
2127 	     const struct ipa_gsi_endpoint_data *data)
2128 {
2129 	struct device *dev = &pdev->dev;
2130 	struct resource *res;
2131 	resource_size_t size;
2132 	u32 adjust;
2133 	int ret;
2134 
2135 	gsi_validate_build();
2136 
2137 	gsi->dev = dev;
2138 	gsi->version = version;
2139 
2140 	/* The GSI layer performs NAPI on all endpoints.  NAPI requires a
2141 	 * network device structure, but the GSI layer does not have one,
2142 	 * so we must create a dummy network device for this purpose.
2143 	 */
2144 	init_dummy_netdev(&gsi->dummy_dev);
2145 
2146 	/* Get GSI memory range and map it */
2147 	res = platform_get_resource_byname(pdev, IORESOURCE_MEM, "gsi");
2148 	if (!res) {
2149 		dev_err(dev, "DT error getting \"gsi\" memory property\n");
2150 		return -ENODEV;
2151 	}
2152 
2153 	size = resource_size(res);
2154 	if (res->start > U32_MAX || size > U32_MAX - res->start) {
2155 		dev_err(dev, "DT memory resource \"gsi\" out of range\n");
2156 		return -EINVAL;
2157 	}
2158 
2159 	/* Make sure we can make our pointer adjustment if necessary */
2160 	adjust = gsi->version < IPA_VERSION_4_5 ? 0 : GSI_EE_REG_ADJUST;
2161 	if (res->start < adjust) {
2162 		dev_err(dev, "DT memory resource \"gsi\" too low (< %u)\n",
2163 			adjust);
2164 		return -EINVAL;
2165 	}
2166 
2167 	gsi->virt = ioremap(res->start, size);
2168 	if (!gsi->virt) {
2169 		dev_err(dev, "unable to remap \"gsi\" memory\n");
2170 		return -ENOMEM;
2171 	}
2172 	/* Adjust register range pointer downward for newer IPA versions */
2173 	gsi->virt -= adjust;
2174 
2175 	init_completion(&gsi->completion);
2176 
2177 	ret = gsi_irq_init(gsi, pdev);
2178 	if (ret)
2179 		goto err_iounmap;
2180 
2181 	ret = gsi_channel_init(gsi, count, data);
2182 	if (ret)
2183 		goto err_irq_exit;
2184 
2185 	mutex_init(&gsi->mutex);
2186 
2187 	return 0;
2188 
2189 err_irq_exit:
2190 	gsi_irq_exit(gsi);
2191 err_iounmap:
2192 	iounmap(gsi->virt);
2193 
2194 	return ret;
2195 }
2196 
2197 /* Inverse of gsi_init() */
2198 void gsi_exit(struct gsi *gsi)
2199 {
2200 	mutex_destroy(&gsi->mutex);
2201 	gsi_channel_exit(gsi);
2202 	gsi_irq_exit(gsi);
2203 	iounmap(gsi->virt);
2204 }
2205 
2206 /* The maximum number of outstanding TREs on a channel.  This limits
2207  * a channel's maximum number of transactions outstanding (worst case
2208  * is one TRE per transaction).
2209  *
2210  * The absolute limit is the number of TREs in the channel's TRE ring,
2211  * and in theory we should be able use all of them.  But in practice,
2212  * doing that led to the hardware reporting exhaustion of event ring
2213  * slots for writing completion information.  So the hardware limit
2214  * would be (tre_count - 1).
2215  *
2216  * We reduce it a bit further though.  Transaction resource pools are
2217  * sized to be a little larger than this maximum, to allow resource
2218  * allocations to always be contiguous.  The number of entries in a
2219  * TRE ring buffer is a power of 2, and the extra resources in a pool
2220  * tends to nearly double the memory allocated for it.  Reducing the
2221  * maximum number of outstanding TREs allows the number of entries in
2222  * a pool to avoid crossing that power-of-2 boundary, and this can
2223  * substantially reduce pool memory requirements.  The number we
2224  * reduce it by matches the number added in gsi_trans_pool_init().
2225  */
2226 u32 gsi_channel_tre_max(struct gsi *gsi, u32 channel_id)
2227 {
2228 	struct gsi_channel *channel = &gsi->channel[channel_id];
2229 
2230 	/* Hardware limit is channel->tre_count - 1 */
2231 	return channel->tre_count - (channel->tlv_count - 1);
2232 }
2233 
2234 /* Returns the maximum number of TREs in a single transaction for a channel */
2235 u32 gsi_channel_trans_tre_max(struct gsi *gsi, u32 channel_id)
2236 {
2237 	struct gsi_channel *channel = &gsi->channel[channel_id];
2238 
2239 	return channel->tlv_count;
2240 }
2241