1 /*
2  * Copyright (C) 2013, 2014 ARM Limited, All Rights Reserved.
3  * Author: Marc Zyngier <marc.zyngier@arm.com>
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License version 2 as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
16  */
17 
18 #include <linux/bitmap.h>
19 #include <linux/cpu.h>
20 #include <linux/delay.h>
21 #include <linux/interrupt.h>
22 #include <linux/log2.h>
23 #include <linux/mm.h>
24 #include <linux/msi.h>
25 #include <linux/of.h>
26 #include <linux/of_address.h>
27 #include <linux/of_irq.h>
28 #include <linux/of_pci.h>
29 #include <linux/of_platform.h>
30 #include <linux/percpu.h>
31 #include <linux/slab.h>
32 
33 #include <linux/irqchip.h>
34 #include <linux/irqchip/arm-gic-v3.h>
35 
36 #include <asm/cacheflush.h>
37 #include <asm/cputype.h>
38 #include <asm/exception.h>
39 
40 #include "irq-gic-common.h"
41 
42 #define ITS_FLAGS_CMDQ_NEEDS_FLUSHING		(1ULL << 0)
43 #define ITS_FLAGS_WORKAROUND_CAVIUM_22375	(1ULL << 1)
44 
45 #define RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING	(1 << 0)
46 
47 /*
48  * Collection structure - just an ID, and a redistributor address to
49  * ping. We use one per CPU as a bag of interrupts assigned to this
50  * CPU.
51  */
52 struct its_collection {
53 	u64			target_address;
54 	u16			col_id;
55 };
56 
57 /*
58  * The ITS structure - contains most of the infrastructure, with the
59  * top-level MSI domain, the command queue, the collections, and the
60  * list of devices writing to it.
61  */
62 struct its_node {
63 	raw_spinlock_t		lock;
64 	struct list_head	entry;
65 	void __iomem		*base;
66 	unsigned long		phys_base;
67 	struct its_cmd_block	*cmd_base;
68 	struct its_cmd_block	*cmd_write;
69 	void			*tables[GITS_BASER_NR_REGS];
70 	struct its_collection	*collections;
71 	struct list_head	its_device_list;
72 	u64			flags;
73 	u32			ite_size;
74 };
75 
76 #define ITS_ITT_ALIGN		SZ_256
77 
78 struct event_lpi_map {
79 	unsigned long		*lpi_map;
80 	u16			*col_map;
81 	irq_hw_number_t		lpi_base;
82 	int			nr_lpis;
83 };
84 
85 /*
86  * The ITS view of a device - belongs to an ITS, a collection, owns an
87  * interrupt translation table, and a list of interrupts.
88  */
89 struct its_device {
90 	struct list_head	entry;
91 	struct its_node		*its;
92 	struct event_lpi_map	event_map;
93 	void			*itt;
94 	u32			nr_ites;
95 	u32			device_id;
96 };
97 
98 static LIST_HEAD(its_nodes);
99 static DEFINE_SPINLOCK(its_lock);
100 static struct device_node *gic_root_node;
101 static struct rdists *gic_rdists;
102 
103 #define gic_data_rdist()		(raw_cpu_ptr(gic_rdists->rdist))
104 #define gic_data_rdist_rd_base()	(gic_data_rdist()->rd_base)
105 
106 static struct its_collection *dev_event_to_col(struct its_device *its_dev,
107 					       u32 event)
108 {
109 	struct its_node *its = its_dev->its;
110 
111 	return its->collections + its_dev->event_map.col_map[event];
112 }
113 
114 /*
115  * ITS command descriptors - parameters to be encoded in a command
116  * block.
117  */
118 struct its_cmd_desc {
119 	union {
120 		struct {
121 			struct its_device *dev;
122 			u32 event_id;
123 		} its_inv_cmd;
124 
125 		struct {
126 			struct its_device *dev;
127 			u32 event_id;
128 		} its_int_cmd;
129 
130 		struct {
131 			struct its_device *dev;
132 			int valid;
133 		} its_mapd_cmd;
134 
135 		struct {
136 			struct its_collection *col;
137 			int valid;
138 		} its_mapc_cmd;
139 
140 		struct {
141 			struct its_device *dev;
142 			u32 phys_id;
143 			u32 event_id;
144 		} its_mapvi_cmd;
145 
146 		struct {
147 			struct its_device *dev;
148 			struct its_collection *col;
149 			u32 event_id;
150 		} its_movi_cmd;
151 
152 		struct {
153 			struct its_device *dev;
154 			u32 event_id;
155 		} its_discard_cmd;
156 
157 		struct {
158 			struct its_collection *col;
159 		} its_invall_cmd;
160 	};
161 };
162 
163 /*
164  * The ITS command block, which is what the ITS actually parses.
165  */
166 struct its_cmd_block {
167 	u64	raw_cmd[4];
168 };
169 
170 #define ITS_CMD_QUEUE_SZ		SZ_64K
171 #define ITS_CMD_QUEUE_NR_ENTRIES	(ITS_CMD_QUEUE_SZ / sizeof(struct its_cmd_block))
172 
173 typedef struct its_collection *(*its_cmd_builder_t)(struct its_cmd_block *,
174 						    struct its_cmd_desc *);
175 
176 static void its_encode_cmd(struct its_cmd_block *cmd, u8 cmd_nr)
177 {
178 	cmd->raw_cmd[0] &= ~0xffUL;
179 	cmd->raw_cmd[0] |= cmd_nr;
180 }
181 
182 static void its_encode_devid(struct its_cmd_block *cmd, u32 devid)
183 {
184 	cmd->raw_cmd[0] &= BIT_ULL(32) - 1;
185 	cmd->raw_cmd[0] |= ((u64)devid) << 32;
186 }
187 
188 static void its_encode_event_id(struct its_cmd_block *cmd, u32 id)
189 {
190 	cmd->raw_cmd[1] &= ~0xffffffffUL;
191 	cmd->raw_cmd[1] |= id;
192 }
193 
194 static void its_encode_phys_id(struct its_cmd_block *cmd, u32 phys_id)
195 {
196 	cmd->raw_cmd[1] &= 0xffffffffUL;
197 	cmd->raw_cmd[1] |= ((u64)phys_id) << 32;
198 }
199 
200 static void its_encode_size(struct its_cmd_block *cmd, u8 size)
201 {
202 	cmd->raw_cmd[1] &= ~0x1fUL;
203 	cmd->raw_cmd[1] |= size & 0x1f;
204 }
205 
206 static void its_encode_itt(struct its_cmd_block *cmd, u64 itt_addr)
207 {
208 	cmd->raw_cmd[2] &= ~0xffffffffffffUL;
209 	cmd->raw_cmd[2] |= itt_addr & 0xffffffffff00UL;
210 }
211 
212 static void its_encode_valid(struct its_cmd_block *cmd, int valid)
213 {
214 	cmd->raw_cmd[2] &= ~(1UL << 63);
215 	cmd->raw_cmd[2] |= ((u64)!!valid) << 63;
216 }
217 
218 static void its_encode_target(struct its_cmd_block *cmd, u64 target_addr)
219 {
220 	cmd->raw_cmd[2] &= ~(0xffffffffUL << 16);
221 	cmd->raw_cmd[2] |= (target_addr & (0xffffffffUL << 16));
222 }
223 
224 static void its_encode_collection(struct its_cmd_block *cmd, u16 col)
225 {
226 	cmd->raw_cmd[2] &= ~0xffffUL;
227 	cmd->raw_cmd[2] |= col;
228 }
229 
230 static inline void its_fixup_cmd(struct its_cmd_block *cmd)
231 {
232 	/* Let's fixup BE commands */
233 	cmd->raw_cmd[0] = cpu_to_le64(cmd->raw_cmd[0]);
234 	cmd->raw_cmd[1] = cpu_to_le64(cmd->raw_cmd[1]);
235 	cmd->raw_cmd[2] = cpu_to_le64(cmd->raw_cmd[2]);
236 	cmd->raw_cmd[3] = cpu_to_le64(cmd->raw_cmd[3]);
237 }
238 
239 static struct its_collection *its_build_mapd_cmd(struct its_cmd_block *cmd,
240 						 struct its_cmd_desc *desc)
241 {
242 	unsigned long itt_addr;
243 	u8 size = ilog2(desc->its_mapd_cmd.dev->nr_ites);
244 
245 	itt_addr = virt_to_phys(desc->its_mapd_cmd.dev->itt);
246 	itt_addr = ALIGN(itt_addr, ITS_ITT_ALIGN);
247 
248 	its_encode_cmd(cmd, GITS_CMD_MAPD);
249 	its_encode_devid(cmd, desc->its_mapd_cmd.dev->device_id);
250 	its_encode_size(cmd, size - 1);
251 	its_encode_itt(cmd, itt_addr);
252 	its_encode_valid(cmd, desc->its_mapd_cmd.valid);
253 
254 	its_fixup_cmd(cmd);
255 
256 	return NULL;
257 }
258 
259 static struct its_collection *its_build_mapc_cmd(struct its_cmd_block *cmd,
260 						 struct its_cmd_desc *desc)
261 {
262 	its_encode_cmd(cmd, GITS_CMD_MAPC);
263 	its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
264 	its_encode_target(cmd, desc->its_mapc_cmd.col->target_address);
265 	its_encode_valid(cmd, desc->its_mapc_cmd.valid);
266 
267 	its_fixup_cmd(cmd);
268 
269 	return desc->its_mapc_cmd.col;
270 }
271 
272 static struct its_collection *its_build_mapvi_cmd(struct its_cmd_block *cmd,
273 						  struct its_cmd_desc *desc)
274 {
275 	struct its_collection *col;
276 
277 	col = dev_event_to_col(desc->its_mapvi_cmd.dev,
278 			       desc->its_mapvi_cmd.event_id);
279 
280 	its_encode_cmd(cmd, GITS_CMD_MAPVI);
281 	its_encode_devid(cmd, desc->its_mapvi_cmd.dev->device_id);
282 	its_encode_event_id(cmd, desc->its_mapvi_cmd.event_id);
283 	its_encode_phys_id(cmd, desc->its_mapvi_cmd.phys_id);
284 	its_encode_collection(cmd, col->col_id);
285 
286 	its_fixup_cmd(cmd);
287 
288 	return col;
289 }
290 
291 static struct its_collection *its_build_movi_cmd(struct its_cmd_block *cmd,
292 						 struct its_cmd_desc *desc)
293 {
294 	struct its_collection *col;
295 
296 	col = dev_event_to_col(desc->its_movi_cmd.dev,
297 			       desc->its_movi_cmd.event_id);
298 
299 	its_encode_cmd(cmd, GITS_CMD_MOVI);
300 	its_encode_devid(cmd, desc->its_movi_cmd.dev->device_id);
301 	its_encode_event_id(cmd, desc->its_movi_cmd.event_id);
302 	its_encode_collection(cmd, desc->its_movi_cmd.col->col_id);
303 
304 	its_fixup_cmd(cmd);
305 
306 	return col;
307 }
308 
309 static struct its_collection *its_build_discard_cmd(struct its_cmd_block *cmd,
310 						    struct its_cmd_desc *desc)
311 {
312 	struct its_collection *col;
313 
314 	col = dev_event_to_col(desc->its_discard_cmd.dev,
315 			       desc->its_discard_cmd.event_id);
316 
317 	its_encode_cmd(cmd, GITS_CMD_DISCARD);
318 	its_encode_devid(cmd, desc->its_discard_cmd.dev->device_id);
319 	its_encode_event_id(cmd, desc->its_discard_cmd.event_id);
320 
321 	its_fixup_cmd(cmd);
322 
323 	return col;
324 }
325 
326 static struct its_collection *its_build_inv_cmd(struct its_cmd_block *cmd,
327 						struct its_cmd_desc *desc)
328 {
329 	struct its_collection *col;
330 
331 	col = dev_event_to_col(desc->its_inv_cmd.dev,
332 			       desc->its_inv_cmd.event_id);
333 
334 	its_encode_cmd(cmd, GITS_CMD_INV);
335 	its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
336 	its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
337 
338 	its_fixup_cmd(cmd);
339 
340 	return col;
341 }
342 
343 static struct its_collection *its_build_invall_cmd(struct its_cmd_block *cmd,
344 						   struct its_cmd_desc *desc)
345 {
346 	its_encode_cmd(cmd, GITS_CMD_INVALL);
347 	its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
348 
349 	its_fixup_cmd(cmd);
350 
351 	return NULL;
352 }
353 
354 static u64 its_cmd_ptr_to_offset(struct its_node *its,
355 				 struct its_cmd_block *ptr)
356 {
357 	return (ptr - its->cmd_base) * sizeof(*ptr);
358 }
359 
360 static int its_queue_full(struct its_node *its)
361 {
362 	int widx;
363 	int ridx;
364 
365 	widx = its->cmd_write - its->cmd_base;
366 	ridx = readl_relaxed(its->base + GITS_CREADR) / sizeof(struct its_cmd_block);
367 
368 	/* This is incredibly unlikely to happen, unless the ITS locks up. */
369 	if (((widx + 1) % ITS_CMD_QUEUE_NR_ENTRIES) == ridx)
370 		return 1;
371 
372 	return 0;
373 }
374 
375 static struct its_cmd_block *its_allocate_entry(struct its_node *its)
376 {
377 	struct its_cmd_block *cmd;
378 	u32 count = 1000000;	/* 1s! */
379 
380 	while (its_queue_full(its)) {
381 		count--;
382 		if (!count) {
383 			pr_err_ratelimited("ITS queue not draining\n");
384 			return NULL;
385 		}
386 		cpu_relax();
387 		udelay(1);
388 	}
389 
390 	cmd = its->cmd_write++;
391 
392 	/* Handle queue wrapping */
393 	if (its->cmd_write == (its->cmd_base + ITS_CMD_QUEUE_NR_ENTRIES))
394 		its->cmd_write = its->cmd_base;
395 
396 	return cmd;
397 }
398 
399 static struct its_cmd_block *its_post_commands(struct its_node *its)
400 {
401 	u64 wr = its_cmd_ptr_to_offset(its, its->cmd_write);
402 
403 	writel_relaxed(wr, its->base + GITS_CWRITER);
404 
405 	return its->cmd_write;
406 }
407 
408 static void its_flush_cmd(struct its_node *its, struct its_cmd_block *cmd)
409 {
410 	/*
411 	 * Make sure the commands written to memory are observable by
412 	 * the ITS.
413 	 */
414 	if (its->flags & ITS_FLAGS_CMDQ_NEEDS_FLUSHING)
415 		__flush_dcache_area(cmd, sizeof(*cmd));
416 	else
417 		dsb(ishst);
418 }
419 
420 static void its_wait_for_range_completion(struct its_node *its,
421 					  struct its_cmd_block *from,
422 					  struct its_cmd_block *to)
423 {
424 	u64 rd_idx, from_idx, to_idx;
425 	u32 count = 1000000;	/* 1s! */
426 
427 	from_idx = its_cmd_ptr_to_offset(its, from);
428 	to_idx = its_cmd_ptr_to_offset(its, to);
429 
430 	while (1) {
431 		rd_idx = readl_relaxed(its->base + GITS_CREADR);
432 		if (rd_idx >= to_idx || rd_idx < from_idx)
433 			break;
434 
435 		count--;
436 		if (!count) {
437 			pr_err_ratelimited("ITS queue timeout\n");
438 			return;
439 		}
440 		cpu_relax();
441 		udelay(1);
442 	}
443 }
444 
445 static void its_send_single_command(struct its_node *its,
446 				    its_cmd_builder_t builder,
447 				    struct its_cmd_desc *desc)
448 {
449 	struct its_cmd_block *cmd, *sync_cmd, *next_cmd;
450 	struct its_collection *sync_col;
451 	unsigned long flags;
452 
453 	raw_spin_lock_irqsave(&its->lock, flags);
454 
455 	cmd = its_allocate_entry(its);
456 	if (!cmd) {		/* We're soooooo screewed... */
457 		pr_err_ratelimited("ITS can't allocate, dropping command\n");
458 		raw_spin_unlock_irqrestore(&its->lock, flags);
459 		return;
460 	}
461 	sync_col = builder(cmd, desc);
462 	its_flush_cmd(its, cmd);
463 
464 	if (sync_col) {
465 		sync_cmd = its_allocate_entry(its);
466 		if (!sync_cmd) {
467 			pr_err_ratelimited("ITS can't SYNC, skipping\n");
468 			goto post;
469 		}
470 		its_encode_cmd(sync_cmd, GITS_CMD_SYNC);
471 		its_encode_target(sync_cmd, sync_col->target_address);
472 		its_fixup_cmd(sync_cmd);
473 		its_flush_cmd(its, sync_cmd);
474 	}
475 
476 post:
477 	next_cmd = its_post_commands(its);
478 	raw_spin_unlock_irqrestore(&its->lock, flags);
479 
480 	its_wait_for_range_completion(its, cmd, next_cmd);
481 }
482 
483 static void its_send_inv(struct its_device *dev, u32 event_id)
484 {
485 	struct its_cmd_desc desc;
486 
487 	desc.its_inv_cmd.dev = dev;
488 	desc.its_inv_cmd.event_id = event_id;
489 
490 	its_send_single_command(dev->its, its_build_inv_cmd, &desc);
491 }
492 
493 static void its_send_mapd(struct its_device *dev, int valid)
494 {
495 	struct its_cmd_desc desc;
496 
497 	desc.its_mapd_cmd.dev = dev;
498 	desc.its_mapd_cmd.valid = !!valid;
499 
500 	its_send_single_command(dev->its, its_build_mapd_cmd, &desc);
501 }
502 
503 static void its_send_mapc(struct its_node *its, struct its_collection *col,
504 			  int valid)
505 {
506 	struct its_cmd_desc desc;
507 
508 	desc.its_mapc_cmd.col = col;
509 	desc.its_mapc_cmd.valid = !!valid;
510 
511 	its_send_single_command(its, its_build_mapc_cmd, &desc);
512 }
513 
514 static void its_send_mapvi(struct its_device *dev, u32 irq_id, u32 id)
515 {
516 	struct its_cmd_desc desc;
517 
518 	desc.its_mapvi_cmd.dev = dev;
519 	desc.its_mapvi_cmd.phys_id = irq_id;
520 	desc.its_mapvi_cmd.event_id = id;
521 
522 	its_send_single_command(dev->its, its_build_mapvi_cmd, &desc);
523 }
524 
525 static void its_send_movi(struct its_device *dev,
526 			  struct its_collection *col, u32 id)
527 {
528 	struct its_cmd_desc desc;
529 
530 	desc.its_movi_cmd.dev = dev;
531 	desc.its_movi_cmd.col = col;
532 	desc.its_movi_cmd.event_id = id;
533 
534 	its_send_single_command(dev->its, its_build_movi_cmd, &desc);
535 }
536 
537 static void its_send_discard(struct its_device *dev, u32 id)
538 {
539 	struct its_cmd_desc desc;
540 
541 	desc.its_discard_cmd.dev = dev;
542 	desc.its_discard_cmd.event_id = id;
543 
544 	its_send_single_command(dev->its, its_build_discard_cmd, &desc);
545 }
546 
547 static void its_send_invall(struct its_node *its, struct its_collection *col)
548 {
549 	struct its_cmd_desc desc;
550 
551 	desc.its_invall_cmd.col = col;
552 
553 	its_send_single_command(its, its_build_invall_cmd, &desc);
554 }
555 
556 /*
557  * irqchip functions - assumes MSI, mostly.
558  */
559 
560 static inline u32 its_get_event_id(struct irq_data *d)
561 {
562 	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
563 	return d->hwirq - its_dev->event_map.lpi_base;
564 }
565 
566 static void lpi_set_config(struct irq_data *d, bool enable)
567 {
568 	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
569 	irq_hw_number_t hwirq = d->hwirq;
570 	u32 id = its_get_event_id(d);
571 	u8 *cfg = page_address(gic_rdists->prop_page) + hwirq - 8192;
572 
573 	if (enable)
574 		*cfg |= LPI_PROP_ENABLED;
575 	else
576 		*cfg &= ~LPI_PROP_ENABLED;
577 
578 	/*
579 	 * Make the above write visible to the redistributors.
580 	 * And yes, we're flushing exactly: One. Single. Byte.
581 	 * Humpf...
582 	 */
583 	if (gic_rdists->flags & RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING)
584 		__flush_dcache_area(cfg, sizeof(*cfg));
585 	else
586 		dsb(ishst);
587 	its_send_inv(its_dev, id);
588 }
589 
590 static void its_mask_irq(struct irq_data *d)
591 {
592 	lpi_set_config(d, false);
593 }
594 
595 static void its_unmask_irq(struct irq_data *d)
596 {
597 	lpi_set_config(d, true);
598 }
599 
600 static void its_eoi_irq(struct irq_data *d)
601 {
602 	gic_write_eoir(d->hwirq);
603 }
604 
605 static int its_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
606 			    bool force)
607 {
608 	unsigned int cpu = cpumask_any_and(mask_val, cpu_online_mask);
609 	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
610 	struct its_collection *target_col;
611 	u32 id = its_get_event_id(d);
612 
613 	if (cpu >= nr_cpu_ids)
614 		return -EINVAL;
615 
616 	target_col = &its_dev->its->collections[cpu];
617 	its_send_movi(its_dev, target_col, id);
618 	its_dev->event_map.col_map[id] = cpu;
619 
620 	return IRQ_SET_MASK_OK_DONE;
621 }
622 
623 static void its_irq_compose_msi_msg(struct irq_data *d, struct msi_msg *msg)
624 {
625 	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
626 	struct its_node *its;
627 	u64 addr;
628 
629 	its = its_dev->its;
630 	addr = its->phys_base + GITS_TRANSLATER;
631 
632 	msg->address_lo		= addr & ((1UL << 32) - 1);
633 	msg->address_hi		= addr >> 32;
634 	msg->data		= its_get_event_id(d);
635 }
636 
637 static struct irq_chip its_irq_chip = {
638 	.name			= "ITS",
639 	.irq_mask		= its_mask_irq,
640 	.irq_unmask		= its_unmask_irq,
641 	.irq_eoi		= its_eoi_irq,
642 	.irq_set_affinity	= its_set_affinity,
643 	.irq_compose_msi_msg	= its_irq_compose_msi_msg,
644 };
645 
646 /*
647  * How we allocate LPIs:
648  *
649  * The GIC has id_bits bits for interrupt identifiers. From there, we
650  * must subtract 8192 which are reserved for SGIs/PPIs/SPIs. Then, as
651  * we allocate LPIs by chunks of 32, we can shift the whole thing by 5
652  * bits to the right.
653  *
654  * This gives us (((1UL << id_bits) - 8192) >> 5) possible allocations.
655  */
656 #define IRQS_PER_CHUNK_SHIFT	5
657 #define IRQS_PER_CHUNK		(1 << IRQS_PER_CHUNK_SHIFT)
658 
659 static unsigned long *lpi_bitmap;
660 static u32 lpi_chunks;
661 static DEFINE_SPINLOCK(lpi_lock);
662 
663 static int its_lpi_to_chunk(int lpi)
664 {
665 	return (lpi - 8192) >> IRQS_PER_CHUNK_SHIFT;
666 }
667 
668 static int its_chunk_to_lpi(int chunk)
669 {
670 	return (chunk << IRQS_PER_CHUNK_SHIFT) + 8192;
671 }
672 
673 static int its_lpi_init(u32 id_bits)
674 {
675 	lpi_chunks = its_lpi_to_chunk(1UL << id_bits);
676 
677 	lpi_bitmap = kzalloc(BITS_TO_LONGS(lpi_chunks) * sizeof(long),
678 			     GFP_KERNEL);
679 	if (!lpi_bitmap) {
680 		lpi_chunks = 0;
681 		return -ENOMEM;
682 	}
683 
684 	pr_info("ITS: Allocated %d chunks for LPIs\n", (int)lpi_chunks);
685 	return 0;
686 }
687 
688 static unsigned long *its_lpi_alloc_chunks(int nr_irqs, int *base, int *nr_ids)
689 {
690 	unsigned long *bitmap = NULL;
691 	int chunk_id;
692 	int nr_chunks;
693 	int i;
694 
695 	nr_chunks = DIV_ROUND_UP(nr_irqs, IRQS_PER_CHUNK);
696 
697 	spin_lock(&lpi_lock);
698 
699 	do {
700 		chunk_id = bitmap_find_next_zero_area(lpi_bitmap, lpi_chunks,
701 						      0, nr_chunks, 0);
702 		if (chunk_id < lpi_chunks)
703 			break;
704 
705 		nr_chunks--;
706 	} while (nr_chunks > 0);
707 
708 	if (!nr_chunks)
709 		goto out;
710 
711 	bitmap = kzalloc(BITS_TO_LONGS(nr_chunks * IRQS_PER_CHUNK) * sizeof (long),
712 			 GFP_ATOMIC);
713 	if (!bitmap)
714 		goto out;
715 
716 	for (i = 0; i < nr_chunks; i++)
717 		set_bit(chunk_id + i, lpi_bitmap);
718 
719 	*base = its_chunk_to_lpi(chunk_id);
720 	*nr_ids = nr_chunks * IRQS_PER_CHUNK;
721 
722 out:
723 	spin_unlock(&lpi_lock);
724 
725 	if (!bitmap)
726 		*base = *nr_ids = 0;
727 
728 	return bitmap;
729 }
730 
731 static void its_lpi_free(struct event_lpi_map *map)
732 {
733 	int base = map->lpi_base;
734 	int nr_ids = map->nr_lpis;
735 	int lpi;
736 
737 	spin_lock(&lpi_lock);
738 
739 	for (lpi = base; lpi < (base + nr_ids); lpi += IRQS_PER_CHUNK) {
740 		int chunk = its_lpi_to_chunk(lpi);
741 		BUG_ON(chunk > lpi_chunks);
742 		if (test_bit(chunk, lpi_bitmap)) {
743 			clear_bit(chunk, lpi_bitmap);
744 		} else {
745 			pr_err("Bad LPI chunk %d\n", chunk);
746 		}
747 	}
748 
749 	spin_unlock(&lpi_lock);
750 
751 	kfree(map->lpi_map);
752 	kfree(map->col_map);
753 }
754 
755 /*
756  * We allocate 64kB for PROPBASE. That gives us at most 64K LPIs to
757  * deal with (one configuration byte per interrupt). PENDBASE has to
758  * be 64kB aligned (one bit per LPI, plus 8192 bits for SPI/PPI/SGI).
759  */
760 #define LPI_PROPBASE_SZ		SZ_64K
761 #define LPI_PENDBASE_SZ		(LPI_PROPBASE_SZ / 8 + SZ_1K)
762 
763 /*
764  * This is how many bits of ID we need, including the useless ones.
765  */
766 #define LPI_NRBITS		ilog2(LPI_PROPBASE_SZ + SZ_8K)
767 
768 #define LPI_PROP_DEFAULT_PRIO	0xa0
769 
770 static int __init its_alloc_lpi_tables(void)
771 {
772 	phys_addr_t paddr;
773 
774 	gic_rdists->prop_page = alloc_pages(GFP_NOWAIT,
775 					   get_order(LPI_PROPBASE_SZ));
776 	if (!gic_rdists->prop_page) {
777 		pr_err("Failed to allocate PROPBASE\n");
778 		return -ENOMEM;
779 	}
780 
781 	paddr = page_to_phys(gic_rdists->prop_page);
782 	pr_info("GIC: using LPI property table @%pa\n", &paddr);
783 
784 	/* Priority 0xa0, Group-1, disabled */
785 	memset(page_address(gic_rdists->prop_page),
786 	       LPI_PROP_DEFAULT_PRIO | LPI_PROP_GROUP1,
787 	       LPI_PROPBASE_SZ);
788 
789 	/* Make sure the GIC will observe the written configuration */
790 	__flush_dcache_area(page_address(gic_rdists->prop_page), LPI_PROPBASE_SZ);
791 
792 	return 0;
793 }
794 
795 static const char *its_base_type_string[] = {
796 	[GITS_BASER_TYPE_DEVICE]	= "Devices",
797 	[GITS_BASER_TYPE_VCPU]		= "Virtual CPUs",
798 	[GITS_BASER_TYPE_CPU]		= "Physical CPUs",
799 	[GITS_BASER_TYPE_COLLECTION]	= "Interrupt Collections",
800 	[GITS_BASER_TYPE_RESERVED5] 	= "Reserved (5)",
801 	[GITS_BASER_TYPE_RESERVED6] 	= "Reserved (6)",
802 	[GITS_BASER_TYPE_RESERVED7] 	= "Reserved (7)",
803 };
804 
805 static void its_free_tables(struct its_node *its)
806 {
807 	int i;
808 
809 	for (i = 0; i < GITS_BASER_NR_REGS; i++) {
810 		if (its->tables[i]) {
811 			free_page((unsigned long)its->tables[i]);
812 			its->tables[i] = NULL;
813 		}
814 	}
815 }
816 
817 static int its_alloc_tables(const char *node_name, struct its_node *its)
818 {
819 	int err;
820 	int i;
821 	int psz = SZ_64K;
822 	u64 shr = GITS_BASER_InnerShareable;
823 	u64 cache;
824 	u64 typer;
825 	u32 ids;
826 
827 	if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_22375) {
828 		/*
829 		 * erratum 22375: only alloc 8MB table size
830 		 * erratum 24313: ignore memory access type
831 		 */
832 		cache	= 0;
833 		ids	= 0x14;			/* 20 bits, 8MB */
834 	} else {
835 		cache	= GITS_BASER_WaWb;
836 		typer	= readq_relaxed(its->base + GITS_TYPER);
837 		ids	= GITS_TYPER_DEVBITS(typer);
838 	}
839 
840 	for (i = 0; i < GITS_BASER_NR_REGS; i++) {
841 		u64 val = readq_relaxed(its->base + GITS_BASER + i * 8);
842 		u64 type = GITS_BASER_TYPE(val);
843 		u64 entry_size = GITS_BASER_ENTRY_SIZE(val);
844 		int order = get_order(psz);
845 		int alloc_size;
846 		int alloc_pages;
847 		u64 tmp;
848 		void *base;
849 
850 		if (type == GITS_BASER_TYPE_NONE)
851 			continue;
852 
853 		/*
854 		 * Allocate as many entries as required to fit the
855 		 * range of device IDs that the ITS can grok... The ID
856 		 * space being incredibly sparse, this results in a
857 		 * massive waste of memory.
858 		 *
859 		 * For other tables, only allocate a single page.
860 		 */
861 		if (type == GITS_BASER_TYPE_DEVICE) {
862 			/*
863 			 * 'order' was initialized earlier to the default page
864 			 * granule of the the ITS.  We can't have an allocation
865 			 * smaller than that.  If the requested allocation
866 			 * is smaller, round up to the default page granule.
867 			 */
868 			order = max(get_order((1UL << ids) * entry_size),
869 				    order);
870 			if (order >= MAX_ORDER) {
871 				order = MAX_ORDER - 1;
872 				pr_warn("%s: Device Table too large, reduce its page order to %u\n",
873 					node_name, order);
874 			}
875 		}
876 
877 		alloc_size = (1 << order) * PAGE_SIZE;
878 		alloc_pages = (alloc_size / psz);
879 		if (alloc_pages > GITS_BASER_PAGES_MAX) {
880 			alloc_pages = GITS_BASER_PAGES_MAX;
881 			order = get_order(GITS_BASER_PAGES_MAX * psz);
882 			pr_warn("%s: Device Table too large, reduce its page order to %u (%u pages)\n",
883 				node_name, order, alloc_pages);
884 		}
885 
886 		base = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, order);
887 		if (!base) {
888 			err = -ENOMEM;
889 			goto out_free;
890 		}
891 
892 		its->tables[i] = base;
893 
894 retry_baser:
895 		val = (virt_to_phys(base) 				 |
896 		       (type << GITS_BASER_TYPE_SHIFT)			 |
897 		       ((entry_size - 1) << GITS_BASER_ENTRY_SIZE_SHIFT) |
898 		       cache						 |
899 		       shr						 |
900 		       GITS_BASER_VALID);
901 
902 		switch (psz) {
903 		case SZ_4K:
904 			val |= GITS_BASER_PAGE_SIZE_4K;
905 			break;
906 		case SZ_16K:
907 			val |= GITS_BASER_PAGE_SIZE_16K;
908 			break;
909 		case SZ_64K:
910 			val |= GITS_BASER_PAGE_SIZE_64K;
911 			break;
912 		}
913 
914 		val |= alloc_pages - 1;
915 
916 		writeq_relaxed(val, its->base + GITS_BASER + i * 8);
917 		tmp = readq_relaxed(its->base + GITS_BASER + i * 8);
918 
919 		if ((val ^ tmp) & GITS_BASER_SHAREABILITY_MASK) {
920 			/*
921 			 * Shareability didn't stick. Just use
922 			 * whatever the read reported, which is likely
923 			 * to be the only thing this redistributor
924 			 * supports. If that's zero, make it
925 			 * non-cacheable as well.
926 			 */
927 			shr = tmp & GITS_BASER_SHAREABILITY_MASK;
928 			if (!shr) {
929 				cache = GITS_BASER_nC;
930 				__flush_dcache_area(base, alloc_size);
931 			}
932 			goto retry_baser;
933 		}
934 
935 		if ((val ^ tmp) & GITS_BASER_PAGE_SIZE_MASK) {
936 			/*
937 			 * Page size didn't stick. Let's try a smaller
938 			 * size and retry. If we reach 4K, then
939 			 * something is horribly wrong...
940 			 */
941 			switch (psz) {
942 			case SZ_16K:
943 				psz = SZ_4K;
944 				goto retry_baser;
945 			case SZ_64K:
946 				psz = SZ_16K;
947 				goto retry_baser;
948 			}
949 		}
950 
951 		if (val != tmp) {
952 			pr_err("ITS: %s: GITS_BASER%d doesn't stick: %lx %lx\n",
953 			       node_name, i,
954 			       (unsigned long) val, (unsigned long) tmp);
955 			err = -ENXIO;
956 			goto out_free;
957 		}
958 
959 		pr_info("ITS: allocated %d %s @%lx (psz %dK, shr %d)\n",
960 			(int)(alloc_size / entry_size),
961 			its_base_type_string[type],
962 			(unsigned long)virt_to_phys(base),
963 			psz / SZ_1K, (int)shr >> GITS_BASER_SHAREABILITY_SHIFT);
964 	}
965 
966 	return 0;
967 
968 out_free:
969 	its_free_tables(its);
970 
971 	return err;
972 }
973 
974 static int its_alloc_collections(struct its_node *its)
975 {
976 	its->collections = kzalloc(nr_cpu_ids * sizeof(*its->collections),
977 				   GFP_KERNEL);
978 	if (!its->collections)
979 		return -ENOMEM;
980 
981 	return 0;
982 }
983 
984 static void its_cpu_init_lpis(void)
985 {
986 	void __iomem *rbase = gic_data_rdist_rd_base();
987 	struct page *pend_page;
988 	u64 val, tmp;
989 
990 	/* If we didn't allocate the pending table yet, do it now */
991 	pend_page = gic_data_rdist()->pend_page;
992 	if (!pend_page) {
993 		phys_addr_t paddr;
994 		/*
995 		 * The pending pages have to be at least 64kB aligned,
996 		 * hence the 'max(LPI_PENDBASE_SZ, SZ_64K)' below.
997 		 */
998 		pend_page = alloc_pages(GFP_NOWAIT | __GFP_ZERO,
999 					get_order(max(LPI_PENDBASE_SZ, SZ_64K)));
1000 		if (!pend_page) {
1001 			pr_err("Failed to allocate PENDBASE for CPU%d\n",
1002 			       smp_processor_id());
1003 			return;
1004 		}
1005 
1006 		/* Make sure the GIC will observe the zero-ed page */
1007 		__flush_dcache_area(page_address(pend_page), LPI_PENDBASE_SZ);
1008 
1009 		paddr = page_to_phys(pend_page);
1010 		pr_info("CPU%d: using LPI pending table @%pa\n",
1011 			smp_processor_id(), &paddr);
1012 		gic_data_rdist()->pend_page = pend_page;
1013 	}
1014 
1015 	/* Disable LPIs */
1016 	val = readl_relaxed(rbase + GICR_CTLR);
1017 	val &= ~GICR_CTLR_ENABLE_LPIS;
1018 	writel_relaxed(val, rbase + GICR_CTLR);
1019 
1020 	/*
1021 	 * Make sure any change to the table is observable by the GIC.
1022 	 */
1023 	dsb(sy);
1024 
1025 	/* set PROPBASE */
1026 	val = (page_to_phys(gic_rdists->prop_page) |
1027 	       GICR_PROPBASER_InnerShareable |
1028 	       GICR_PROPBASER_WaWb |
1029 	       ((LPI_NRBITS - 1) & GICR_PROPBASER_IDBITS_MASK));
1030 
1031 	writeq_relaxed(val, rbase + GICR_PROPBASER);
1032 	tmp = readq_relaxed(rbase + GICR_PROPBASER);
1033 
1034 	if ((tmp ^ val) & GICR_PROPBASER_SHAREABILITY_MASK) {
1035 		if (!(tmp & GICR_PROPBASER_SHAREABILITY_MASK)) {
1036 			/*
1037 			 * The HW reports non-shareable, we must
1038 			 * remove the cacheability attributes as
1039 			 * well.
1040 			 */
1041 			val &= ~(GICR_PROPBASER_SHAREABILITY_MASK |
1042 				 GICR_PROPBASER_CACHEABILITY_MASK);
1043 			val |= GICR_PROPBASER_nC;
1044 			writeq_relaxed(val, rbase + GICR_PROPBASER);
1045 		}
1046 		pr_info_once("GIC: using cache flushing for LPI property table\n");
1047 		gic_rdists->flags |= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING;
1048 	}
1049 
1050 	/* set PENDBASE */
1051 	val = (page_to_phys(pend_page) |
1052 	       GICR_PENDBASER_InnerShareable |
1053 	       GICR_PENDBASER_WaWb);
1054 
1055 	writeq_relaxed(val, rbase + GICR_PENDBASER);
1056 	tmp = readq_relaxed(rbase + GICR_PENDBASER);
1057 
1058 	if (!(tmp & GICR_PENDBASER_SHAREABILITY_MASK)) {
1059 		/*
1060 		 * The HW reports non-shareable, we must remove the
1061 		 * cacheability attributes as well.
1062 		 */
1063 		val &= ~(GICR_PENDBASER_SHAREABILITY_MASK |
1064 			 GICR_PENDBASER_CACHEABILITY_MASK);
1065 		val |= GICR_PENDBASER_nC;
1066 		writeq_relaxed(val, rbase + GICR_PENDBASER);
1067 	}
1068 
1069 	/* Enable LPIs */
1070 	val = readl_relaxed(rbase + GICR_CTLR);
1071 	val |= GICR_CTLR_ENABLE_LPIS;
1072 	writel_relaxed(val, rbase + GICR_CTLR);
1073 
1074 	/* Make sure the GIC has seen the above */
1075 	dsb(sy);
1076 }
1077 
1078 static void its_cpu_init_collection(void)
1079 {
1080 	struct its_node *its;
1081 	int cpu;
1082 
1083 	spin_lock(&its_lock);
1084 	cpu = smp_processor_id();
1085 
1086 	list_for_each_entry(its, &its_nodes, entry) {
1087 		u64 target;
1088 
1089 		/*
1090 		 * We now have to bind each collection to its target
1091 		 * redistributor.
1092 		 */
1093 		if (readq_relaxed(its->base + GITS_TYPER) & GITS_TYPER_PTA) {
1094 			/*
1095 			 * This ITS wants the physical address of the
1096 			 * redistributor.
1097 			 */
1098 			target = gic_data_rdist()->phys_base;
1099 		} else {
1100 			/*
1101 			 * This ITS wants a linear CPU number.
1102 			 */
1103 			target = readq_relaxed(gic_data_rdist_rd_base() + GICR_TYPER);
1104 			target = GICR_TYPER_CPU_NUMBER(target) << 16;
1105 		}
1106 
1107 		/* Perform collection mapping */
1108 		its->collections[cpu].target_address = target;
1109 		its->collections[cpu].col_id = cpu;
1110 
1111 		its_send_mapc(its, &its->collections[cpu], 1);
1112 		its_send_invall(its, &its->collections[cpu]);
1113 	}
1114 
1115 	spin_unlock(&its_lock);
1116 }
1117 
1118 static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
1119 {
1120 	struct its_device *its_dev = NULL, *tmp;
1121 	unsigned long flags;
1122 
1123 	raw_spin_lock_irqsave(&its->lock, flags);
1124 
1125 	list_for_each_entry(tmp, &its->its_device_list, entry) {
1126 		if (tmp->device_id == dev_id) {
1127 			its_dev = tmp;
1128 			break;
1129 		}
1130 	}
1131 
1132 	raw_spin_unlock_irqrestore(&its->lock, flags);
1133 
1134 	return its_dev;
1135 }
1136 
1137 static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
1138 					    int nvecs)
1139 {
1140 	struct its_device *dev;
1141 	unsigned long *lpi_map;
1142 	unsigned long flags;
1143 	u16 *col_map = NULL;
1144 	void *itt;
1145 	int lpi_base;
1146 	int nr_lpis;
1147 	int nr_ites;
1148 	int sz;
1149 
1150 	dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1151 	/*
1152 	 * At least one bit of EventID is being used, hence a minimum
1153 	 * of two entries. No, the architecture doesn't let you
1154 	 * express an ITT with a single entry.
1155 	 */
1156 	nr_ites = max(2UL, roundup_pow_of_two(nvecs));
1157 	sz = nr_ites * its->ite_size;
1158 	sz = max(sz, ITS_ITT_ALIGN) + ITS_ITT_ALIGN - 1;
1159 	itt = kzalloc(sz, GFP_KERNEL);
1160 	lpi_map = its_lpi_alloc_chunks(nvecs, &lpi_base, &nr_lpis);
1161 	if (lpi_map)
1162 		col_map = kzalloc(sizeof(*col_map) * nr_lpis, GFP_KERNEL);
1163 
1164 	if (!dev || !itt || !lpi_map || !col_map) {
1165 		kfree(dev);
1166 		kfree(itt);
1167 		kfree(lpi_map);
1168 		kfree(col_map);
1169 		return NULL;
1170 	}
1171 
1172 	__flush_dcache_area(itt, sz);
1173 
1174 	dev->its = its;
1175 	dev->itt = itt;
1176 	dev->nr_ites = nr_ites;
1177 	dev->event_map.lpi_map = lpi_map;
1178 	dev->event_map.col_map = col_map;
1179 	dev->event_map.lpi_base = lpi_base;
1180 	dev->event_map.nr_lpis = nr_lpis;
1181 	dev->device_id = dev_id;
1182 	INIT_LIST_HEAD(&dev->entry);
1183 
1184 	raw_spin_lock_irqsave(&its->lock, flags);
1185 	list_add(&dev->entry, &its->its_device_list);
1186 	raw_spin_unlock_irqrestore(&its->lock, flags);
1187 
1188 	/* Map device to its ITT */
1189 	its_send_mapd(dev, 1);
1190 
1191 	return dev;
1192 }
1193 
1194 static void its_free_device(struct its_device *its_dev)
1195 {
1196 	unsigned long flags;
1197 
1198 	raw_spin_lock_irqsave(&its_dev->its->lock, flags);
1199 	list_del(&its_dev->entry);
1200 	raw_spin_unlock_irqrestore(&its_dev->its->lock, flags);
1201 	kfree(its_dev->itt);
1202 	kfree(its_dev);
1203 }
1204 
1205 static int its_alloc_device_irq(struct its_device *dev, irq_hw_number_t *hwirq)
1206 {
1207 	int idx;
1208 
1209 	idx = find_first_zero_bit(dev->event_map.lpi_map,
1210 				  dev->event_map.nr_lpis);
1211 	if (idx == dev->event_map.nr_lpis)
1212 		return -ENOSPC;
1213 
1214 	*hwirq = dev->event_map.lpi_base + idx;
1215 	set_bit(idx, dev->event_map.lpi_map);
1216 
1217 	return 0;
1218 }
1219 
1220 static int its_msi_prepare(struct irq_domain *domain, struct device *dev,
1221 			   int nvec, msi_alloc_info_t *info)
1222 {
1223 	struct its_node *its;
1224 	struct its_device *its_dev;
1225 	struct msi_domain_info *msi_info;
1226 	u32 dev_id;
1227 
1228 	/*
1229 	 * We ignore "dev" entierely, and rely on the dev_id that has
1230 	 * been passed via the scratchpad. This limits this domain's
1231 	 * usefulness to upper layers that definitely know that they
1232 	 * are built on top of the ITS.
1233 	 */
1234 	dev_id = info->scratchpad[0].ul;
1235 
1236 	msi_info = msi_get_domain_info(domain);
1237 	its = msi_info->data;
1238 
1239 	its_dev = its_find_device(its, dev_id);
1240 	if (its_dev) {
1241 		/*
1242 		 * We already have seen this ID, probably through
1243 		 * another alias (PCI bridge of some sort). No need to
1244 		 * create the device.
1245 		 */
1246 		pr_debug("Reusing ITT for devID %x\n", dev_id);
1247 		goto out;
1248 	}
1249 
1250 	its_dev = its_create_device(its, dev_id, nvec);
1251 	if (!its_dev)
1252 		return -ENOMEM;
1253 
1254 	pr_debug("ITT %d entries, %d bits\n", nvec, ilog2(nvec));
1255 out:
1256 	info->scratchpad[0].ptr = its_dev;
1257 	return 0;
1258 }
1259 
1260 static struct msi_domain_ops its_msi_domain_ops = {
1261 	.msi_prepare	= its_msi_prepare,
1262 };
1263 
1264 static int its_irq_gic_domain_alloc(struct irq_domain *domain,
1265 				    unsigned int virq,
1266 				    irq_hw_number_t hwirq)
1267 {
1268 	struct irq_fwspec fwspec;
1269 
1270 	if (irq_domain_get_of_node(domain->parent)) {
1271 		fwspec.fwnode = domain->parent->fwnode;
1272 		fwspec.param_count = 3;
1273 		fwspec.param[0] = GIC_IRQ_TYPE_LPI;
1274 		fwspec.param[1] = hwirq;
1275 		fwspec.param[2] = IRQ_TYPE_EDGE_RISING;
1276 	} else {
1277 		return -EINVAL;
1278 	}
1279 
1280 	return irq_domain_alloc_irqs_parent(domain, virq, 1, &fwspec);
1281 }
1282 
1283 static int its_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
1284 				unsigned int nr_irqs, void *args)
1285 {
1286 	msi_alloc_info_t *info = args;
1287 	struct its_device *its_dev = info->scratchpad[0].ptr;
1288 	irq_hw_number_t hwirq;
1289 	int err;
1290 	int i;
1291 
1292 	for (i = 0; i < nr_irqs; i++) {
1293 		err = its_alloc_device_irq(its_dev, &hwirq);
1294 		if (err)
1295 			return err;
1296 
1297 		err = its_irq_gic_domain_alloc(domain, virq + i, hwirq);
1298 		if (err)
1299 			return err;
1300 
1301 		irq_domain_set_hwirq_and_chip(domain, virq + i,
1302 					      hwirq, &its_irq_chip, its_dev);
1303 		pr_debug("ID:%d pID:%d vID:%d\n",
1304 			 (int)(hwirq - its_dev->event_map.lpi_base),
1305 			 (int) hwirq, virq + i);
1306 	}
1307 
1308 	return 0;
1309 }
1310 
1311 static void its_irq_domain_activate(struct irq_domain *domain,
1312 				    struct irq_data *d)
1313 {
1314 	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1315 	u32 event = its_get_event_id(d);
1316 
1317 	/* Bind the LPI to the first possible CPU */
1318 	its_dev->event_map.col_map[event] = cpumask_first(cpu_online_mask);
1319 
1320 	/* Map the GIC IRQ and event to the device */
1321 	its_send_mapvi(its_dev, d->hwirq, event);
1322 }
1323 
1324 static void its_irq_domain_deactivate(struct irq_domain *domain,
1325 				      struct irq_data *d)
1326 {
1327 	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1328 	u32 event = its_get_event_id(d);
1329 
1330 	/* Stop the delivery of interrupts */
1331 	its_send_discard(its_dev, event);
1332 }
1333 
1334 static void its_irq_domain_free(struct irq_domain *domain, unsigned int virq,
1335 				unsigned int nr_irqs)
1336 {
1337 	struct irq_data *d = irq_domain_get_irq_data(domain, virq);
1338 	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1339 	int i;
1340 
1341 	for (i = 0; i < nr_irqs; i++) {
1342 		struct irq_data *data = irq_domain_get_irq_data(domain,
1343 								virq + i);
1344 		u32 event = its_get_event_id(data);
1345 
1346 		/* Mark interrupt index as unused */
1347 		clear_bit(event, its_dev->event_map.lpi_map);
1348 
1349 		/* Nuke the entry in the domain */
1350 		irq_domain_reset_irq_data(data);
1351 	}
1352 
1353 	/* If all interrupts have been freed, start mopping the floor */
1354 	if (bitmap_empty(its_dev->event_map.lpi_map,
1355 			 its_dev->event_map.nr_lpis)) {
1356 		its_lpi_free(&its_dev->event_map);
1357 
1358 		/* Unmap device/itt */
1359 		its_send_mapd(its_dev, 0);
1360 		its_free_device(its_dev);
1361 	}
1362 
1363 	irq_domain_free_irqs_parent(domain, virq, nr_irqs);
1364 }
1365 
1366 static const struct irq_domain_ops its_domain_ops = {
1367 	.alloc			= its_irq_domain_alloc,
1368 	.free			= its_irq_domain_free,
1369 	.activate		= its_irq_domain_activate,
1370 	.deactivate		= its_irq_domain_deactivate,
1371 };
1372 
1373 static int its_force_quiescent(void __iomem *base)
1374 {
1375 	u32 count = 1000000;	/* 1s */
1376 	u32 val;
1377 
1378 	val = readl_relaxed(base + GITS_CTLR);
1379 	if (val & GITS_CTLR_QUIESCENT)
1380 		return 0;
1381 
1382 	/* Disable the generation of all interrupts to this ITS */
1383 	val &= ~GITS_CTLR_ENABLE;
1384 	writel_relaxed(val, base + GITS_CTLR);
1385 
1386 	/* Poll GITS_CTLR and wait until ITS becomes quiescent */
1387 	while (1) {
1388 		val = readl_relaxed(base + GITS_CTLR);
1389 		if (val & GITS_CTLR_QUIESCENT)
1390 			return 0;
1391 
1392 		count--;
1393 		if (!count)
1394 			return -EBUSY;
1395 
1396 		cpu_relax();
1397 		udelay(1);
1398 	}
1399 }
1400 
1401 static void __maybe_unused its_enable_quirk_cavium_22375(void *data)
1402 {
1403 	struct its_node *its = data;
1404 
1405 	its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_22375;
1406 }
1407 
1408 static const struct gic_quirk its_quirks[] = {
1409 #ifdef CONFIG_CAVIUM_ERRATUM_22375
1410 	{
1411 		.desc	= "ITS: Cavium errata 22375, 24313",
1412 		.iidr	= 0xa100034c,	/* ThunderX pass 1.x */
1413 		.mask	= 0xffff0fff,
1414 		.init	= its_enable_quirk_cavium_22375,
1415 	},
1416 #endif
1417 	{
1418 	}
1419 };
1420 
1421 static void its_enable_quirks(struct its_node *its)
1422 {
1423 	u32 iidr = readl_relaxed(its->base + GITS_IIDR);
1424 
1425 	gic_enable_quirks(iidr, its_quirks, its);
1426 }
1427 
1428 static int its_probe(struct device_node *node, struct irq_domain *parent)
1429 {
1430 	struct resource res;
1431 	struct its_node *its;
1432 	void __iomem *its_base;
1433 	struct irq_domain *inner_domain;
1434 	u32 val;
1435 	u64 baser, tmp;
1436 	int err;
1437 
1438 	err = of_address_to_resource(node, 0, &res);
1439 	if (err) {
1440 		pr_warn("%s: no regs?\n", node->full_name);
1441 		return -ENXIO;
1442 	}
1443 
1444 	its_base = ioremap(res.start, resource_size(&res));
1445 	if (!its_base) {
1446 		pr_warn("%s: unable to map registers\n", node->full_name);
1447 		return -ENOMEM;
1448 	}
1449 
1450 	val = readl_relaxed(its_base + GITS_PIDR2) & GIC_PIDR2_ARCH_MASK;
1451 	if (val != 0x30 && val != 0x40) {
1452 		pr_warn("%s: no ITS detected, giving up\n", node->full_name);
1453 		err = -ENODEV;
1454 		goto out_unmap;
1455 	}
1456 
1457 	err = its_force_quiescent(its_base);
1458 	if (err) {
1459 		pr_warn("%s: failed to quiesce, giving up\n",
1460 			node->full_name);
1461 		goto out_unmap;
1462 	}
1463 
1464 	pr_info("ITS: %s\n", node->full_name);
1465 
1466 	its = kzalloc(sizeof(*its), GFP_KERNEL);
1467 	if (!its) {
1468 		err = -ENOMEM;
1469 		goto out_unmap;
1470 	}
1471 
1472 	raw_spin_lock_init(&its->lock);
1473 	INIT_LIST_HEAD(&its->entry);
1474 	INIT_LIST_HEAD(&its->its_device_list);
1475 	its->base = its_base;
1476 	its->phys_base = res.start;
1477 	its->ite_size = ((readl_relaxed(its_base + GITS_TYPER) >> 4) & 0xf) + 1;
1478 
1479 	its->cmd_base = kzalloc(ITS_CMD_QUEUE_SZ, GFP_KERNEL);
1480 	if (!its->cmd_base) {
1481 		err = -ENOMEM;
1482 		goto out_free_its;
1483 	}
1484 	its->cmd_write = its->cmd_base;
1485 
1486 	its_enable_quirks(its);
1487 
1488 	err = its_alloc_tables(node->full_name, its);
1489 	if (err)
1490 		goto out_free_cmd;
1491 
1492 	err = its_alloc_collections(its);
1493 	if (err)
1494 		goto out_free_tables;
1495 
1496 	baser = (virt_to_phys(its->cmd_base)	|
1497 		 GITS_CBASER_WaWb		|
1498 		 GITS_CBASER_InnerShareable	|
1499 		 (ITS_CMD_QUEUE_SZ / SZ_4K - 1)	|
1500 		 GITS_CBASER_VALID);
1501 
1502 	writeq_relaxed(baser, its->base + GITS_CBASER);
1503 	tmp = readq_relaxed(its->base + GITS_CBASER);
1504 
1505 	if ((tmp ^ baser) & GITS_CBASER_SHAREABILITY_MASK) {
1506 		if (!(tmp & GITS_CBASER_SHAREABILITY_MASK)) {
1507 			/*
1508 			 * The HW reports non-shareable, we must
1509 			 * remove the cacheability attributes as
1510 			 * well.
1511 			 */
1512 			baser &= ~(GITS_CBASER_SHAREABILITY_MASK |
1513 				   GITS_CBASER_CACHEABILITY_MASK);
1514 			baser |= GITS_CBASER_nC;
1515 			writeq_relaxed(baser, its->base + GITS_CBASER);
1516 		}
1517 		pr_info("ITS: using cache flushing for cmd queue\n");
1518 		its->flags |= ITS_FLAGS_CMDQ_NEEDS_FLUSHING;
1519 	}
1520 
1521 	writeq_relaxed(0, its->base + GITS_CWRITER);
1522 	writel_relaxed(GITS_CTLR_ENABLE, its->base + GITS_CTLR);
1523 
1524 	if (of_property_read_bool(node, "msi-controller")) {
1525 		struct msi_domain_info *info;
1526 
1527 		info = kzalloc(sizeof(*info), GFP_KERNEL);
1528 		if (!info) {
1529 			err = -ENOMEM;
1530 			goto out_free_tables;
1531 		}
1532 
1533 		inner_domain = irq_domain_add_tree(node, &its_domain_ops, its);
1534 		if (!inner_domain) {
1535 			err = -ENOMEM;
1536 			kfree(info);
1537 			goto out_free_tables;
1538 		}
1539 
1540 		inner_domain->parent = parent;
1541 		inner_domain->bus_token = DOMAIN_BUS_NEXUS;
1542 		info->ops = &its_msi_domain_ops;
1543 		info->data = its;
1544 		inner_domain->host_data = info;
1545 	}
1546 
1547 	spin_lock(&its_lock);
1548 	list_add(&its->entry, &its_nodes);
1549 	spin_unlock(&its_lock);
1550 
1551 	return 0;
1552 
1553 out_free_tables:
1554 	its_free_tables(its);
1555 out_free_cmd:
1556 	kfree(its->cmd_base);
1557 out_free_its:
1558 	kfree(its);
1559 out_unmap:
1560 	iounmap(its_base);
1561 	pr_err("ITS: failed probing %s (%d)\n", node->full_name, err);
1562 	return err;
1563 }
1564 
1565 static bool gic_rdists_supports_plpis(void)
1566 {
1567 	return !!(readl_relaxed(gic_data_rdist_rd_base() + GICR_TYPER) & GICR_TYPER_PLPIS);
1568 }
1569 
1570 int its_cpu_init(void)
1571 {
1572 	if (!list_empty(&its_nodes)) {
1573 		if (!gic_rdists_supports_plpis()) {
1574 			pr_info("CPU%d: LPIs not supported\n", smp_processor_id());
1575 			return -ENXIO;
1576 		}
1577 		its_cpu_init_lpis();
1578 		its_cpu_init_collection();
1579 	}
1580 
1581 	return 0;
1582 }
1583 
1584 static struct of_device_id its_device_id[] = {
1585 	{	.compatible	= "arm,gic-v3-its",	},
1586 	{},
1587 };
1588 
1589 int its_init(struct device_node *node, struct rdists *rdists,
1590 	     struct irq_domain *parent_domain)
1591 {
1592 	struct device_node *np;
1593 
1594 	for (np = of_find_matching_node(node, its_device_id); np;
1595 	     np = of_find_matching_node(np, its_device_id)) {
1596 		its_probe(np, parent_domain);
1597 	}
1598 
1599 	if (list_empty(&its_nodes)) {
1600 		pr_warn("ITS: No ITS available, not enabling LPIs\n");
1601 		return -ENXIO;
1602 	}
1603 
1604 	gic_rdists = rdists;
1605 	gic_root_node = node;
1606 
1607 	its_alloc_lpi_tables();
1608 	its_lpi_init(rdists->id_bits);
1609 
1610 	return 0;
1611 }
1612