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/arm-gic-v3.h>
34
35#include <asm/cacheflush.h>
36#include <asm/cputype.h>
37#include <asm/exception.h>
38
39#include "irqchip.h"
40
41#define ITS_FLAGS_CMDQ_NEEDS_FLUSHING		(1 << 0)
42
43#define RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING	(1 << 0)
44
45/*
46 * Collection structure - just an ID, and a redistributor address to
47 * ping. We use one per CPU as a bag of interrupts assigned to this
48 * CPU.
49 */
50struct its_collection {
51	u64			target_address;
52	u16			col_id;
53};
54
55/*
56 * The ITS structure - contains most of the infrastructure, with the
57 * msi_controller, the command queue, the collections, and the list of
58 * devices writing to it.
59 */
60struct its_node {
61	raw_spinlock_t		lock;
62	struct list_head	entry;
63	struct msi_controller	msi_chip;
64	struct irq_domain	*domain;
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
78struct 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 */
89struct 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
98static LIST_HEAD(its_nodes);
99static DEFINE_SPINLOCK(its_lock);
100static struct device_node *gic_root_node;
101static 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
106static 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 */
118struct 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 */
166struct 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
173typedef struct its_collection *(*its_cmd_builder_t)(struct its_cmd_block *,
174						    struct its_cmd_desc *);
175
176static 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
182static 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
188static 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
194static 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
200static 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
206static 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
212static 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
218static 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
224static 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
230static 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
239static 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
259static 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
272static 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
291static 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
309static 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
326static 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
343static 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
354static 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
360static 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
375static 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
399static 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
408static 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
420static 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
445static 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
476post:
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
483static 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
493static 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
503static 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
514static 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
525static 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
537static 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
547static 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
560static 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
566static 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
590static void its_mask_irq(struct irq_data *d)
591{
592	lpi_set_config(d, false);
593}
594
595static void its_unmask_irq(struct irq_data *d)
596{
597	lpi_set_config(d, true);
598}
599
600static void its_eoi_irq(struct irq_data *d)
601{
602	gic_write_eoir(d->hwirq);
603}
604
605static 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
623static 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
637static 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
646static void its_mask_msi_irq(struct irq_data *d)
647{
648	pci_msi_mask_irq(d);
649	irq_chip_mask_parent(d);
650}
651
652static void its_unmask_msi_irq(struct irq_data *d)
653{
654	pci_msi_unmask_irq(d);
655	irq_chip_unmask_parent(d);
656}
657
658static struct irq_chip its_msi_irq_chip = {
659	.name			= "ITS-MSI",
660	.irq_unmask		= its_unmask_msi_irq,
661	.irq_mask		= its_mask_msi_irq,
662	.irq_eoi		= irq_chip_eoi_parent,
663	.irq_write_msi_msg	= pci_msi_domain_write_msg,
664};
665
666/*
667 * How we allocate LPIs:
668 *
669 * The GIC has id_bits bits for interrupt identifiers. From there, we
670 * must subtract 8192 which are reserved for SGIs/PPIs/SPIs. Then, as
671 * we allocate LPIs by chunks of 32, we can shift the whole thing by 5
672 * bits to the right.
673 *
674 * This gives us (((1UL << id_bits) - 8192) >> 5) possible allocations.
675 */
676#define IRQS_PER_CHUNK_SHIFT	5
677#define IRQS_PER_CHUNK		(1 << IRQS_PER_CHUNK_SHIFT)
678
679static unsigned long *lpi_bitmap;
680static u32 lpi_chunks;
681static DEFINE_SPINLOCK(lpi_lock);
682
683static int its_lpi_to_chunk(int lpi)
684{
685	return (lpi - 8192) >> IRQS_PER_CHUNK_SHIFT;
686}
687
688static int its_chunk_to_lpi(int chunk)
689{
690	return (chunk << IRQS_PER_CHUNK_SHIFT) + 8192;
691}
692
693static int its_lpi_init(u32 id_bits)
694{
695	lpi_chunks = its_lpi_to_chunk(1UL << id_bits);
696
697	lpi_bitmap = kzalloc(BITS_TO_LONGS(lpi_chunks) * sizeof(long),
698			     GFP_KERNEL);
699	if (!lpi_bitmap) {
700		lpi_chunks = 0;
701		return -ENOMEM;
702	}
703
704	pr_info("ITS: Allocated %d chunks for LPIs\n", (int)lpi_chunks);
705	return 0;
706}
707
708static unsigned long *its_lpi_alloc_chunks(int nr_irqs, int *base, int *nr_ids)
709{
710	unsigned long *bitmap = NULL;
711	int chunk_id;
712	int nr_chunks;
713	int i;
714
715	nr_chunks = DIV_ROUND_UP(nr_irqs, IRQS_PER_CHUNK);
716
717	spin_lock(&lpi_lock);
718
719	do {
720		chunk_id = bitmap_find_next_zero_area(lpi_bitmap, lpi_chunks,
721						      0, nr_chunks, 0);
722		if (chunk_id < lpi_chunks)
723			break;
724
725		nr_chunks--;
726	} while (nr_chunks > 0);
727
728	if (!nr_chunks)
729		goto out;
730
731	bitmap = kzalloc(BITS_TO_LONGS(nr_chunks * IRQS_PER_CHUNK) * sizeof (long),
732			 GFP_ATOMIC);
733	if (!bitmap)
734		goto out;
735
736	for (i = 0; i < nr_chunks; i++)
737		set_bit(chunk_id + i, lpi_bitmap);
738
739	*base = its_chunk_to_lpi(chunk_id);
740	*nr_ids = nr_chunks * IRQS_PER_CHUNK;
741
742out:
743	spin_unlock(&lpi_lock);
744
745	return bitmap;
746}
747
748static void its_lpi_free(struct event_lpi_map *map)
749{
750	int base = map->lpi_base;
751	int nr_ids = map->nr_lpis;
752	int lpi;
753
754	spin_lock(&lpi_lock);
755
756	for (lpi = base; lpi < (base + nr_ids); lpi += IRQS_PER_CHUNK) {
757		int chunk = its_lpi_to_chunk(lpi);
758		BUG_ON(chunk > lpi_chunks);
759		if (test_bit(chunk, lpi_bitmap)) {
760			clear_bit(chunk, lpi_bitmap);
761		} else {
762			pr_err("Bad LPI chunk %d\n", chunk);
763		}
764	}
765
766	spin_unlock(&lpi_lock);
767
768	kfree(map->lpi_map);
769	kfree(map->col_map);
770}
771
772/*
773 * We allocate 64kB for PROPBASE. That gives us at most 64K LPIs to
774 * deal with (one configuration byte per interrupt). PENDBASE has to
775 * be 64kB aligned (one bit per LPI, plus 8192 bits for SPI/PPI/SGI).
776 */
777#define LPI_PROPBASE_SZ		SZ_64K
778#define LPI_PENDBASE_SZ		(LPI_PROPBASE_SZ / 8 + SZ_1K)
779
780/*
781 * This is how many bits of ID we need, including the useless ones.
782 */
783#define LPI_NRBITS		ilog2(LPI_PROPBASE_SZ + SZ_8K)
784
785#define LPI_PROP_DEFAULT_PRIO	0xa0
786
787static int __init its_alloc_lpi_tables(void)
788{
789	phys_addr_t paddr;
790
791	gic_rdists->prop_page = alloc_pages(GFP_NOWAIT,
792					   get_order(LPI_PROPBASE_SZ));
793	if (!gic_rdists->prop_page) {
794		pr_err("Failed to allocate PROPBASE\n");
795		return -ENOMEM;
796	}
797
798	paddr = page_to_phys(gic_rdists->prop_page);
799	pr_info("GIC: using LPI property table @%pa\n", &paddr);
800
801	/* Priority 0xa0, Group-1, disabled */
802	memset(page_address(gic_rdists->prop_page),
803	       LPI_PROP_DEFAULT_PRIO | LPI_PROP_GROUP1,
804	       LPI_PROPBASE_SZ);
805
806	/* Make sure the GIC will observe the written configuration */
807	__flush_dcache_area(page_address(gic_rdists->prop_page), LPI_PROPBASE_SZ);
808
809	return 0;
810}
811
812static const char *its_base_type_string[] = {
813	[GITS_BASER_TYPE_DEVICE]	= "Devices",
814	[GITS_BASER_TYPE_VCPU]		= "Virtual CPUs",
815	[GITS_BASER_TYPE_CPU]		= "Physical CPUs",
816	[GITS_BASER_TYPE_COLLECTION]	= "Interrupt Collections",
817	[GITS_BASER_TYPE_RESERVED5] 	= "Reserved (5)",
818	[GITS_BASER_TYPE_RESERVED6] 	= "Reserved (6)",
819	[GITS_BASER_TYPE_RESERVED7] 	= "Reserved (7)",
820};
821
822static void its_free_tables(struct its_node *its)
823{
824	int i;
825
826	for (i = 0; i < GITS_BASER_NR_REGS; i++) {
827		if (its->tables[i]) {
828			free_page((unsigned long)its->tables[i]);
829			its->tables[i] = NULL;
830		}
831	}
832}
833
834static int its_alloc_tables(struct its_node *its)
835{
836	int err;
837	int i;
838	int psz = SZ_64K;
839	u64 shr = GITS_BASER_InnerShareable;
840	u64 cache = GITS_BASER_WaWb;
841
842	for (i = 0; i < GITS_BASER_NR_REGS; i++) {
843		u64 val = readq_relaxed(its->base + GITS_BASER + i * 8);
844		u64 type = GITS_BASER_TYPE(val);
845		u64 entry_size = GITS_BASER_ENTRY_SIZE(val);
846		int order = get_order(psz);
847		int alloc_size;
848		u64 tmp;
849		void *base;
850
851		if (type == GITS_BASER_TYPE_NONE)
852			continue;
853
854		/*
855		 * Allocate as many entries as required to fit the
856		 * range of device IDs that the ITS can grok... The ID
857		 * space being incredibly sparse, this results in a
858		 * massive waste of memory.
859		 *
860		 * For other tables, only allocate a single page.
861		 */
862		if (type == GITS_BASER_TYPE_DEVICE) {
863			u64 typer = readq_relaxed(its->base + GITS_TYPER);
864			u32 ids = GITS_TYPER_DEVBITS(typer);
865
866			/*
867			 * 'order' was initialized earlier to the default page
868			 * granule of the the ITS.  We can't have an allocation
869			 * smaller than that.  If the requested allocation
870			 * is smaller, round up to the default page granule.
871			 */
872			order = max(get_order((1UL << ids) * entry_size),
873				    order);
874			if (order >= MAX_ORDER) {
875				order = MAX_ORDER - 1;
876				pr_warn("%s: Device Table too large, reduce its page order to %u\n",
877					its->msi_chip.of_node->full_name, order);
878			}
879		}
880
881		alloc_size = (1 << order) * PAGE_SIZE;
882		base = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, order);
883		if (!base) {
884			err = -ENOMEM;
885			goto out_free;
886		}
887
888		its->tables[i] = base;
889
890retry_baser:
891		val = (virt_to_phys(base) 				 |
892		       (type << GITS_BASER_TYPE_SHIFT)			 |
893		       ((entry_size - 1) << GITS_BASER_ENTRY_SIZE_SHIFT) |
894		       cache						 |
895		       shr						 |
896		       GITS_BASER_VALID);
897
898		switch (psz) {
899		case SZ_4K:
900			val |= GITS_BASER_PAGE_SIZE_4K;
901			break;
902		case SZ_16K:
903			val |= GITS_BASER_PAGE_SIZE_16K;
904			break;
905		case SZ_64K:
906			val |= GITS_BASER_PAGE_SIZE_64K;
907			break;
908		}
909
910		val |= (alloc_size / psz) - 1;
911
912		writeq_relaxed(val, its->base + GITS_BASER + i * 8);
913		tmp = readq_relaxed(its->base + GITS_BASER + i * 8);
914
915		if ((val ^ tmp) & GITS_BASER_SHAREABILITY_MASK) {
916			/*
917			 * Shareability didn't stick. Just use
918			 * whatever the read reported, which is likely
919			 * to be the only thing this redistributor
920			 * supports. If that's zero, make it
921			 * non-cacheable as well.
922			 */
923			shr = tmp & GITS_BASER_SHAREABILITY_MASK;
924			if (!shr) {
925				cache = GITS_BASER_nC;
926				__flush_dcache_area(base, alloc_size);
927			}
928			goto retry_baser;
929		}
930
931		if ((val ^ tmp) & GITS_BASER_PAGE_SIZE_MASK) {
932			/*
933			 * Page size didn't stick. Let's try a smaller
934			 * size and retry. If we reach 4K, then
935			 * something is horribly wrong...
936			 */
937			switch (psz) {
938			case SZ_16K:
939				psz = SZ_4K;
940				goto retry_baser;
941			case SZ_64K:
942				psz = SZ_16K;
943				goto retry_baser;
944			}
945		}
946
947		if (val != tmp) {
948			pr_err("ITS: %s: GITS_BASER%d doesn't stick: %lx %lx\n",
949			       its->msi_chip.of_node->full_name, i,
950			       (unsigned long) val, (unsigned long) tmp);
951			err = -ENXIO;
952			goto out_free;
953		}
954
955		pr_info("ITS: allocated %d %s @%lx (psz %dK, shr %d)\n",
956			(int)(alloc_size / entry_size),
957			its_base_type_string[type],
958			(unsigned long)virt_to_phys(base),
959			psz / SZ_1K, (int)shr >> GITS_BASER_SHAREABILITY_SHIFT);
960	}
961
962	return 0;
963
964out_free:
965	its_free_tables(its);
966
967	return err;
968}
969
970static int its_alloc_collections(struct its_node *its)
971{
972	its->collections = kzalloc(nr_cpu_ids * sizeof(*its->collections),
973				   GFP_KERNEL);
974	if (!its->collections)
975		return -ENOMEM;
976
977	return 0;
978}
979
980static void its_cpu_init_lpis(void)
981{
982	void __iomem *rbase = gic_data_rdist_rd_base();
983	struct page *pend_page;
984	u64 val, tmp;
985
986	/* If we didn't allocate the pending table yet, do it now */
987	pend_page = gic_data_rdist()->pend_page;
988	if (!pend_page) {
989		phys_addr_t paddr;
990		/*
991		 * The pending pages have to be at least 64kB aligned,
992		 * hence the 'max(LPI_PENDBASE_SZ, SZ_64K)' below.
993		 */
994		pend_page = alloc_pages(GFP_NOWAIT | __GFP_ZERO,
995					get_order(max(LPI_PENDBASE_SZ, SZ_64K)));
996		if (!pend_page) {
997			pr_err("Failed to allocate PENDBASE for CPU%d\n",
998			       smp_processor_id());
999			return;
1000		}
1001
1002		/* Make sure the GIC will observe the zero-ed page */
1003		__flush_dcache_area(page_address(pend_page), LPI_PENDBASE_SZ);
1004
1005		paddr = page_to_phys(pend_page);
1006		pr_info("CPU%d: using LPI pending table @%pa\n",
1007			smp_processor_id(), &paddr);
1008		gic_data_rdist()->pend_page = pend_page;
1009	}
1010
1011	/* Disable LPIs */
1012	val = readl_relaxed(rbase + GICR_CTLR);
1013	val &= ~GICR_CTLR_ENABLE_LPIS;
1014	writel_relaxed(val, rbase + GICR_CTLR);
1015
1016	/*
1017	 * Make sure any change to the table is observable by the GIC.
1018	 */
1019	dsb(sy);
1020
1021	/* set PROPBASE */
1022	val = (page_to_phys(gic_rdists->prop_page) |
1023	       GICR_PROPBASER_InnerShareable |
1024	       GICR_PROPBASER_WaWb |
1025	       ((LPI_NRBITS - 1) & GICR_PROPBASER_IDBITS_MASK));
1026
1027	writeq_relaxed(val, rbase + GICR_PROPBASER);
1028	tmp = readq_relaxed(rbase + GICR_PROPBASER);
1029
1030	if ((tmp ^ val) & GICR_PROPBASER_SHAREABILITY_MASK) {
1031		if (!(tmp & GICR_PROPBASER_SHAREABILITY_MASK)) {
1032			/*
1033			 * The HW reports non-shareable, we must
1034			 * remove the cacheability attributes as
1035			 * well.
1036			 */
1037			val &= ~(GICR_PROPBASER_SHAREABILITY_MASK |
1038				 GICR_PROPBASER_CACHEABILITY_MASK);
1039			val |= GICR_PROPBASER_nC;
1040			writeq_relaxed(val, rbase + GICR_PROPBASER);
1041		}
1042		pr_info_once("GIC: using cache flushing for LPI property table\n");
1043		gic_rdists->flags |= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING;
1044	}
1045
1046	/* set PENDBASE */
1047	val = (page_to_phys(pend_page) |
1048	       GICR_PENDBASER_InnerShareable |
1049	       GICR_PENDBASER_WaWb);
1050
1051	writeq_relaxed(val, rbase + GICR_PENDBASER);
1052	tmp = readq_relaxed(rbase + GICR_PENDBASER);
1053
1054	if (!(tmp & GICR_PENDBASER_SHAREABILITY_MASK)) {
1055		/*
1056		 * The HW reports non-shareable, we must remove the
1057		 * cacheability attributes as well.
1058		 */
1059		val &= ~(GICR_PENDBASER_SHAREABILITY_MASK |
1060			 GICR_PENDBASER_CACHEABILITY_MASK);
1061		val |= GICR_PENDBASER_nC;
1062		writeq_relaxed(val, rbase + GICR_PENDBASER);
1063	}
1064
1065	/* Enable LPIs */
1066	val = readl_relaxed(rbase + GICR_CTLR);
1067	val |= GICR_CTLR_ENABLE_LPIS;
1068	writel_relaxed(val, rbase + GICR_CTLR);
1069
1070	/* Make sure the GIC has seen the above */
1071	dsb(sy);
1072}
1073
1074static void its_cpu_init_collection(void)
1075{
1076	struct its_node *its;
1077	int cpu;
1078
1079	spin_lock(&its_lock);
1080	cpu = smp_processor_id();
1081
1082	list_for_each_entry(its, &its_nodes, entry) {
1083		u64 target;
1084
1085		/*
1086		 * We now have to bind each collection to its target
1087		 * redistributor.
1088		 */
1089		if (readq_relaxed(its->base + GITS_TYPER) & GITS_TYPER_PTA) {
1090			/*
1091			 * This ITS wants the physical address of the
1092			 * redistributor.
1093			 */
1094			target = gic_data_rdist()->phys_base;
1095		} else {
1096			/*
1097			 * This ITS wants a linear CPU number.
1098			 */
1099			target = readq_relaxed(gic_data_rdist_rd_base() + GICR_TYPER);
1100			target = GICR_TYPER_CPU_NUMBER(target) << 16;
1101		}
1102
1103		/* Perform collection mapping */
1104		its->collections[cpu].target_address = target;
1105		its->collections[cpu].col_id = cpu;
1106
1107		its_send_mapc(its, &its->collections[cpu], 1);
1108		its_send_invall(its, &its->collections[cpu]);
1109	}
1110
1111	spin_unlock(&its_lock);
1112}
1113
1114static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
1115{
1116	struct its_device *its_dev = NULL, *tmp;
1117	unsigned long flags;
1118
1119	raw_spin_lock_irqsave(&its->lock, flags);
1120
1121	list_for_each_entry(tmp, &its->its_device_list, entry) {
1122		if (tmp->device_id == dev_id) {
1123			its_dev = tmp;
1124			break;
1125		}
1126	}
1127
1128	raw_spin_unlock_irqrestore(&its->lock, flags);
1129
1130	return its_dev;
1131}
1132
1133static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
1134					    int nvecs)
1135{
1136	struct its_device *dev;
1137	unsigned long *lpi_map;
1138	unsigned long flags;
1139	u16 *col_map = NULL;
1140	void *itt;
1141	int lpi_base;
1142	int nr_lpis;
1143	int nr_ites;
1144	int sz;
1145
1146	dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1147	/*
1148	 * At least one bit of EventID is being used, hence a minimum
1149	 * of two entries. No, the architecture doesn't let you
1150	 * express an ITT with a single entry.
1151	 */
1152	nr_ites = max(2UL, roundup_pow_of_two(nvecs));
1153	sz = nr_ites * its->ite_size;
1154	sz = max(sz, ITS_ITT_ALIGN) + ITS_ITT_ALIGN - 1;
1155	itt = kzalloc(sz, GFP_KERNEL);
1156	lpi_map = its_lpi_alloc_chunks(nvecs, &lpi_base, &nr_lpis);
1157	if (lpi_map)
1158		col_map = kzalloc(sizeof(*col_map) * nr_lpis, GFP_KERNEL);
1159
1160	if (!dev || !itt || !lpi_map || !col_map) {
1161		kfree(dev);
1162		kfree(itt);
1163		kfree(lpi_map);
1164		kfree(col_map);
1165		return NULL;
1166	}
1167
1168	__flush_dcache_area(itt, sz);
1169
1170	dev->its = its;
1171	dev->itt = itt;
1172	dev->nr_ites = nr_ites;
1173	dev->event_map.lpi_map = lpi_map;
1174	dev->event_map.col_map = col_map;
1175	dev->event_map.lpi_base = lpi_base;
1176	dev->event_map.nr_lpis = nr_lpis;
1177	dev->device_id = dev_id;
1178	INIT_LIST_HEAD(&dev->entry);
1179
1180	raw_spin_lock_irqsave(&its->lock, flags);
1181	list_add(&dev->entry, &its->its_device_list);
1182	raw_spin_unlock_irqrestore(&its->lock, flags);
1183
1184	/* Map device to its ITT */
1185	its_send_mapd(dev, 1);
1186
1187	return dev;
1188}
1189
1190static void its_free_device(struct its_device *its_dev)
1191{
1192	unsigned long flags;
1193
1194	raw_spin_lock_irqsave(&its_dev->its->lock, flags);
1195	list_del(&its_dev->entry);
1196	raw_spin_unlock_irqrestore(&its_dev->its->lock, flags);
1197	kfree(its_dev->itt);
1198	kfree(its_dev);
1199}
1200
1201static int its_alloc_device_irq(struct its_device *dev, irq_hw_number_t *hwirq)
1202{
1203	int idx;
1204
1205	idx = find_first_zero_bit(dev->event_map.lpi_map,
1206				  dev->event_map.nr_lpis);
1207	if (idx == dev->event_map.nr_lpis)
1208		return -ENOSPC;
1209
1210	*hwirq = dev->event_map.lpi_base + idx;
1211	set_bit(idx, dev->event_map.lpi_map);
1212
1213	return 0;
1214}
1215
1216struct its_pci_alias {
1217	struct pci_dev	*pdev;
1218	u32		dev_id;
1219	u32		count;
1220};
1221
1222static int its_pci_msi_vec_count(struct pci_dev *pdev)
1223{
1224	int msi, msix;
1225
1226	msi = max(pci_msi_vec_count(pdev), 0);
1227	msix = max(pci_msix_vec_count(pdev), 0);
1228
1229	return max(msi, msix);
1230}
1231
1232static int its_get_pci_alias(struct pci_dev *pdev, u16 alias, void *data)
1233{
1234	struct its_pci_alias *dev_alias = data;
1235
1236	dev_alias->dev_id = alias;
1237	if (pdev != dev_alias->pdev)
1238		dev_alias->count += its_pci_msi_vec_count(dev_alias->pdev);
1239
1240	return 0;
1241}
1242
1243static int its_msi_prepare(struct irq_domain *domain, struct device *dev,
1244			   int nvec, msi_alloc_info_t *info)
1245{
1246	struct pci_dev *pdev;
1247	struct its_node *its;
1248	struct its_device *its_dev;
1249	struct its_pci_alias dev_alias;
1250
1251	if (!dev_is_pci(dev))
1252		return -EINVAL;
1253
1254	pdev = to_pci_dev(dev);
1255	dev_alias.pdev = pdev;
1256	dev_alias.count = nvec;
1257
1258	pci_for_each_dma_alias(pdev, its_get_pci_alias, &dev_alias);
1259	its = domain->parent->host_data;
1260
1261	its_dev = its_find_device(its, dev_alias.dev_id);
1262	if (its_dev) {
1263		/*
1264		 * We already have seen this ID, probably through
1265		 * another alias (PCI bridge of some sort). No need to
1266		 * create the device.
1267		 */
1268		dev_dbg(dev, "Reusing ITT for devID %x\n", dev_alias.dev_id);
1269		goto out;
1270	}
1271
1272	its_dev = its_create_device(its, dev_alias.dev_id, dev_alias.count);
1273	if (!its_dev)
1274		return -ENOMEM;
1275
1276	dev_dbg(&pdev->dev, "ITT %d entries, %d bits\n",
1277		dev_alias.count, ilog2(dev_alias.count));
1278out:
1279	info->scratchpad[0].ptr = its_dev;
1280	info->scratchpad[1].ptr = dev;
1281	return 0;
1282}
1283
1284static struct msi_domain_ops its_pci_msi_ops = {
1285	.msi_prepare	= its_msi_prepare,
1286};
1287
1288static struct msi_domain_info its_pci_msi_domain_info = {
1289	.flags	= (MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
1290		   MSI_FLAG_MULTI_PCI_MSI | MSI_FLAG_PCI_MSIX),
1291	.ops	= &its_pci_msi_ops,
1292	.chip	= &its_msi_irq_chip,
1293};
1294
1295static int its_irq_gic_domain_alloc(struct irq_domain *domain,
1296				    unsigned int virq,
1297				    irq_hw_number_t hwirq)
1298{
1299	struct of_phandle_args args;
1300
1301	args.np = domain->parent->of_node;
1302	args.args_count = 3;
1303	args.args[0] = GIC_IRQ_TYPE_LPI;
1304	args.args[1] = hwirq;
1305	args.args[2] = IRQ_TYPE_EDGE_RISING;
1306
1307	return irq_domain_alloc_irqs_parent(domain, virq, 1, &args);
1308}
1309
1310static int its_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
1311				unsigned int nr_irqs, void *args)
1312{
1313	msi_alloc_info_t *info = args;
1314	struct its_device *its_dev = info->scratchpad[0].ptr;
1315	irq_hw_number_t hwirq;
1316	int err;
1317	int i;
1318
1319	for (i = 0; i < nr_irqs; i++) {
1320		err = its_alloc_device_irq(its_dev, &hwirq);
1321		if (err)
1322			return err;
1323
1324		err = its_irq_gic_domain_alloc(domain, virq + i, hwirq);
1325		if (err)
1326			return err;
1327
1328		irq_domain_set_hwirq_and_chip(domain, virq + i,
1329					      hwirq, &its_irq_chip, its_dev);
1330		dev_dbg(info->scratchpad[1].ptr, "ID:%d pID:%d vID:%d\n",
1331			(int)(hwirq - its_dev->event_map.lpi_base),
1332			(int)hwirq, virq + i);
1333	}
1334
1335	return 0;
1336}
1337
1338static void its_irq_domain_activate(struct irq_domain *domain,
1339				    struct irq_data *d)
1340{
1341	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1342	u32 event = its_get_event_id(d);
1343
1344	/* Bind the LPI to the first possible CPU */
1345	its_dev->event_map.col_map[event] = cpumask_first(cpu_online_mask);
1346
1347	/* Map the GIC IRQ and event to the device */
1348	its_send_mapvi(its_dev, d->hwirq, event);
1349}
1350
1351static void its_irq_domain_deactivate(struct irq_domain *domain,
1352				      struct irq_data *d)
1353{
1354	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1355	u32 event = its_get_event_id(d);
1356
1357	/* Stop the delivery of interrupts */
1358	its_send_discard(its_dev, event);
1359}
1360
1361static void its_irq_domain_free(struct irq_domain *domain, unsigned int virq,
1362				unsigned int nr_irqs)
1363{
1364	struct irq_data *d = irq_domain_get_irq_data(domain, virq);
1365	struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1366	int i;
1367
1368	for (i = 0; i < nr_irqs; i++) {
1369		struct irq_data *data = irq_domain_get_irq_data(domain,
1370								virq + i);
1371		u32 event = its_get_event_id(data);
1372
1373		/* Mark interrupt index as unused */
1374		clear_bit(event, its_dev->event_map.lpi_map);
1375
1376		/* Nuke the entry in the domain */
1377		irq_domain_reset_irq_data(data);
1378	}
1379
1380	/* If all interrupts have been freed, start mopping the floor */
1381	if (bitmap_empty(its_dev->event_map.lpi_map,
1382			 its_dev->event_map.nr_lpis)) {
1383		its_lpi_free(&its_dev->event_map);
1384
1385		/* Unmap device/itt */
1386		its_send_mapd(its_dev, 0);
1387		its_free_device(its_dev);
1388	}
1389
1390	irq_domain_free_irqs_parent(domain, virq, nr_irqs);
1391}
1392
1393static const struct irq_domain_ops its_domain_ops = {
1394	.alloc			= its_irq_domain_alloc,
1395	.free			= its_irq_domain_free,
1396	.activate		= its_irq_domain_activate,
1397	.deactivate		= its_irq_domain_deactivate,
1398};
1399
1400static int its_force_quiescent(void __iomem *base)
1401{
1402	u32 count = 1000000;	/* 1s */
1403	u32 val;
1404
1405	val = readl_relaxed(base + GITS_CTLR);
1406	if (val & GITS_CTLR_QUIESCENT)
1407		return 0;
1408
1409	/* Disable the generation of all interrupts to this ITS */
1410	val &= ~GITS_CTLR_ENABLE;
1411	writel_relaxed(val, base + GITS_CTLR);
1412
1413	/* Poll GITS_CTLR and wait until ITS becomes quiescent */
1414	while (1) {
1415		val = readl_relaxed(base + GITS_CTLR);
1416		if (val & GITS_CTLR_QUIESCENT)
1417			return 0;
1418
1419		count--;
1420		if (!count)
1421			return -EBUSY;
1422
1423		cpu_relax();
1424		udelay(1);
1425	}
1426}
1427
1428static 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	u32 val;
1434	u64 baser, tmp;
1435	int err;
1436
1437	err = of_address_to_resource(node, 0, &res);
1438	if (err) {
1439		pr_warn("%s: no regs?\n", node->full_name);
1440		return -ENXIO;
1441	}
1442
1443	its_base = ioremap(res.start, resource_size(&res));
1444	if (!its_base) {
1445		pr_warn("%s: unable to map registers\n", node->full_name);
1446		return -ENOMEM;
1447	}
1448
1449	val = readl_relaxed(its_base + GITS_PIDR2) & GIC_PIDR2_ARCH_MASK;
1450	if (val != 0x30 && val != 0x40) {
1451		pr_warn("%s: no ITS detected, giving up\n", node->full_name);
1452		err = -ENODEV;
1453		goto out_unmap;
1454	}
1455
1456	err = its_force_quiescent(its_base);
1457	if (err) {
1458		pr_warn("%s: failed to quiesce, giving up\n",
1459			node->full_name);
1460		goto out_unmap;
1461	}
1462
1463	pr_info("ITS: %s\n", node->full_name);
1464
1465	its = kzalloc(sizeof(*its), GFP_KERNEL);
1466	if (!its) {
1467		err = -ENOMEM;
1468		goto out_unmap;
1469	}
1470
1471	raw_spin_lock_init(&its->lock);
1472	INIT_LIST_HEAD(&its->entry);
1473	INIT_LIST_HEAD(&its->its_device_list);
1474	its->base = its_base;
1475	its->phys_base = res.start;
1476	its->msi_chip.of_node = node;
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	err = its_alloc_tables(its);
1487	if (err)
1488		goto out_free_cmd;
1489
1490	err = its_alloc_collections(its);
1491	if (err)
1492		goto out_free_tables;
1493
1494	baser = (virt_to_phys(its->cmd_base)	|
1495		 GITS_CBASER_WaWb		|
1496		 GITS_CBASER_InnerShareable	|
1497		 (ITS_CMD_QUEUE_SZ / SZ_4K - 1)	|
1498		 GITS_CBASER_VALID);
1499
1500	writeq_relaxed(baser, its->base + GITS_CBASER);
1501	tmp = readq_relaxed(its->base + GITS_CBASER);
1502
1503	if ((tmp ^ baser) & GITS_CBASER_SHAREABILITY_MASK) {
1504		if (!(tmp & GITS_CBASER_SHAREABILITY_MASK)) {
1505			/*
1506			 * The HW reports non-shareable, we must
1507			 * remove the cacheability attributes as
1508			 * well.
1509			 */
1510			baser &= ~(GITS_CBASER_SHAREABILITY_MASK |
1511				   GITS_CBASER_CACHEABILITY_MASK);
1512			baser |= GITS_CBASER_nC;
1513			writeq_relaxed(baser, its->base + GITS_CBASER);
1514		}
1515		pr_info("ITS: using cache flushing for cmd queue\n");
1516		its->flags |= ITS_FLAGS_CMDQ_NEEDS_FLUSHING;
1517	}
1518
1519	writeq_relaxed(0, its->base + GITS_CWRITER);
1520	writel_relaxed(GITS_CTLR_ENABLE, its->base + GITS_CTLR);
1521
1522	if (of_property_read_bool(its->msi_chip.of_node, "msi-controller")) {
1523		its->domain = irq_domain_add_tree(NULL, &its_domain_ops, its);
1524		if (!its->domain) {
1525			err = -ENOMEM;
1526			goto out_free_tables;
1527		}
1528
1529		its->domain->parent = parent;
1530
1531		its->msi_chip.domain = pci_msi_create_irq_domain(node,
1532								 &its_pci_msi_domain_info,
1533								 its->domain);
1534		if (!its->msi_chip.domain) {
1535			err = -ENOMEM;
1536			goto out_free_domains;
1537		}
1538
1539		err = of_pci_msi_chip_add(&its->msi_chip);
1540		if (err)
1541			goto out_free_domains;
1542	}
1543
1544	spin_lock(&its_lock);
1545	list_add(&its->entry, &its_nodes);
1546	spin_unlock(&its_lock);
1547
1548	return 0;
1549
1550out_free_domains:
1551	if (its->msi_chip.domain)
1552		irq_domain_remove(its->msi_chip.domain);
1553	if (its->domain)
1554		irq_domain_remove(its->domain);
1555out_free_tables:
1556	its_free_tables(its);
1557out_free_cmd:
1558	kfree(its->cmd_base);
1559out_free_its:
1560	kfree(its);
1561out_unmap:
1562	iounmap(its_base);
1563	pr_err("ITS: failed probing %s (%d)\n", node->full_name, err);
1564	return err;
1565}
1566
1567static bool gic_rdists_supports_plpis(void)
1568{
1569	return !!(readl_relaxed(gic_data_rdist_rd_base() + GICR_TYPER) & GICR_TYPER_PLPIS);
1570}
1571
1572int its_cpu_init(void)
1573{
1574	if (!list_empty(&its_nodes)) {
1575		if (!gic_rdists_supports_plpis()) {
1576			pr_info("CPU%d: LPIs not supported\n", smp_processor_id());
1577			return -ENXIO;
1578		}
1579		its_cpu_init_lpis();
1580		its_cpu_init_collection();
1581	}
1582
1583	return 0;
1584}
1585
1586static struct of_device_id its_device_id[] = {
1587	{	.compatible	= "arm,gic-v3-its",	},
1588	{},
1589};
1590
1591int its_init(struct device_node *node, struct rdists *rdists,
1592	     struct irq_domain *parent_domain)
1593{
1594	struct device_node *np;
1595
1596	for (np = of_find_matching_node(node, its_device_id); np;
1597	     np = of_find_matching_node(np, its_device_id)) {
1598		its_probe(np, parent_domain);
1599	}
1600
1601	if (list_empty(&its_nodes)) {
1602		pr_warn("ITS: No ITS available, not enabling LPIs\n");
1603		return -ENXIO;
1604	}
1605
1606	gic_rdists = rdists;
1607	gic_root_node = node;
1608
1609	its_alloc_lpi_tables();
1610	its_lpi_init(rdists->id_bits);
1611
1612	return 0;
1613}
1614