2 * Copyright (C) 2013, 2014 ARM Limited, All Rights Reserved.
3 * Author: Marc Zyngier <marc.zyngier@arm.com>
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.
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.
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/>.
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>
24 #include <linux/msi.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>
33 #include <linux/irqchip/arm-gic-v3.h>
35 #include <asm/cacheflush.h>
36 #include <asm/cputype.h>
37 #include <asm/exception.h>
41 #define ITS_FLAGS_CMDQ_NEEDS_FLUSHING (1 << 0)
43 #define RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING (1 << 0)
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
50 struct its_collection
{
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.
62 struct list_head entry
;
63 struct msi_controller msi_chip
;
64 struct irq_domain
*domain
;
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
;
76 #define ITS_ITT_ALIGN SZ_256
79 * The ITS view of a device - belongs to an ITS, a collection, owns an
80 * interrupt translation table, and a list of interrupts.
83 struct list_head entry
;
85 struct its_collection
*collection
;
87 unsigned long *lpi_map
;
88 irq_hw_number_t lpi_base
;
94 static LIST_HEAD(its_nodes
);
95 static DEFINE_SPINLOCK(its_lock
);
96 static struct device_node
*gic_root_node
;
97 static struct rdists
*gic_rdists
;
99 #define gic_data_rdist() (raw_cpu_ptr(gic_rdists->rdist))
100 #define gic_data_rdist_rd_base() (gic_data_rdist()->rd_base)
103 * ITS command descriptors - parameters to be encoded in a command
106 struct its_cmd_desc
{
109 struct its_device
*dev
;
114 struct its_device
*dev
;
119 struct its_device
*dev
;
124 struct its_collection
*col
;
129 struct its_device
*dev
;
135 struct its_device
*dev
;
136 struct its_collection
*col
;
141 struct its_device
*dev
;
146 struct its_collection
*col
;
152 * The ITS command block, which is what the ITS actually parses.
154 struct its_cmd_block
{
158 #define ITS_CMD_QUEUE_SZ SZ_64K
159 #define ITS_CMD_QUEUE_NR_ENTRIES (ITS_CMD_QUEUE_SZ / sizeof(struct its_cmd_block))
161 typedef struct its_collection
*(*its_cmd_builder_t
)(struct its_cmd_block
*,
162 struct its_cmd_desc
*);
164 static void its_encode_cmd(struct its_cmd_block
*cmd
, u8 cmd_nr
)
166 cmd
->raw_cmd
[0] &= ~0xffUL
;
167 cmd
->raw_cmd
[0] |= cmd_nr
;
170 static void its_encode_devid(struct its_cmd_block
*cmd
, u32 devid
)
172 cmd
->raw_cmd
[0] &= BIT_ULL(32) - 1;
173 cmd
->raw_cmd
[0] |= ((u64
)devid
) << 32;
176 static void its_encode_event_id(struct its_cmd_block
*cmd
, u32 id
)
178 cmd
->raw_cmd
[1] &= ~0xffffffffUL
;
179 cmd
->raw_cmd
[1] |= id
;
182 static void its_encode_phys_id(struct its_cmd_block
*cmd
, u32 phys_id
)
184 cmd
->raw_cmd
[1] &= 0xffffffffUL
;
185 cmd
->raw_cmd
[1] |= ((u64
)phys_id
) << 32;
188 static void its_encode_size(struct its_cmd_block
*cmd
, u8 size
)
190 cmd
->raw_cmd
[1] &= ~0x1fUL
;
191 cmd
->raw_cmd
[1] |= size
& 0x1f;
194 static void its_encode_itt(struct its_cmd_block
*cmd
, u64 itt_addr
)
196 cmd
->raw_cmd
[2] &= ~0xffffffffffffUL
;
197 cmd
->raw_cmd
[2] |= itt_addr
& 0xffffffffff00UL
;
200 static void its_encode_valid(struct its_cmd_block
*cmd
, int valid
)
202 cmd
->raw_cmd
[2] &= ~(1UL << 63);
203 cmd
->raw_cmd
[2] |= ((u64
)!!valid
) << 63;
206 static void its_encode_target(struct its_cmd_block
*cmd
, u64 target_addr
)
208 cmd
->raw_cmd
[2] &= ~(0xffffffffUL
<< 16);
209 cmd
->raw_cmd
[2] |= (target_addr
& (0xffffffffUL
<< 16));
212 static void its_encode_collection(struct its_cmd_block
*cmd
, u16 col
)
214 cmd
->raw_cmd
[2] &= ~0xffffUL
;
215 cmd
->raw_cmd
[2] |= col
;
218 static inline void its_fixup_cmd(struct its_cmd_block
*cmd
)
220 /* Let's fixup BE commands */
221 cmd
->raw_cmd
[0] = cpu_to_le64(cmd
->raw_cmd
[0]);
222 cmd
->raw_cmd
[1] = cpu_to_le64(cmd
->raw_cmd
[1]);
223 cmd
->raw_cmd
[2] = cpu_to_le64(cmd
->raw_cmd
[2]);
224 cmd
->raw_cmd
[3] = cpu_to_le64(cmd
->raw_cmd
[3]);
227 static struct its_collection
*its_build_mapd_cmd(struct its_cmd_block
*cmd
,
228 struct its_cmd_desc
*desc
)
230 unsigned long itt_addr
;
231 u8 size
= ilog2(desc
->its_mapd_cmd
.dev
->nr_ites
);
233 itt_addr
= virt_to_phys(desc
->its_mapd_cmd
.dev
->itt
);
234 itt_addr
= ALIGN(itt_addr
, ITS_ITT_ALIGN
);
236 its_encode_cmd(cmd
, GITS_CMD_MAPD
);
237 its_encode_devid(cmd
, desc
->its_mapd_cmd
.dev
->device_id
);
238 its_encode_size(cmd
, size
- 1);
239 its_encode_itt(cmd
, itt_addr
);
240 its_encode_valid(cmd
, desc
->its_mapd_cmd
.valid
);
244 return desc
->its_mapd_cmd
.dev
->collection
;
247 static struct its_collection
*its_build_mapc_cmd(struct its_cmd_block
*cmd
,
248 struct its_cmd_desc
*desc
)
250 its_encode_cmd(cmd
, GITS_CMD_MAPC
);
251 its_encode_collection(cmd
, desc
->its_mapc_cmd
.col
->col_id
);
252 its_encode_target(cmd
, desc
->its_mapc_cmd
.col
->target_address
);
253 its_encode_valid(cmd
, desc
->its_mapc_cmd
.valid
);
257 return desc
->its_mapc_cmd
.col
;
260 static struct its_collection
*its_build_mapvi_cmd(struct its_cmd_block
*cmd
,
261 struct its_cmd_desc
*desc
)
263 its_encode_cmd(cmd
, GITS_CMD_MAPVI
);
264 its_encode_devid(cmd
, desc
->its_mapvi_cmd
.dev
->device_id
);
265 its_encode_event_id(cmd
, desc
->its_mapvi_cmd
.event_id
);
266 its_encode_phys_id(cmd
, desc
->its_mapvi_cmd
.phys_id
);
267 its_encode_collection(cmd
, desc
->its_mapvi_cmd
.dev
->collection
->col_id
);
271 return desc
->its_mapvi_cmd
.dev
->collection
;
274 static struct its_collection
*its_build_movi_cmd(struct its_cmd_block
*cmd
,
275 struct its_cmd_desc
*desc
)
277 its_encode_cmd(cmd
, GITS_CMD_MOVI
);
278 its_encode_devid(cmd
, desc
->its_movi_cmd
.dev
->device_id
);
279 its_encode_event_id(cmd
, desc
->its_movi_cmd
.id
);
280 its_encode_collection(cmd
, desc
->its_movi_cmd
.col
->col_id
);
284 return desc
->its_movi_cmd
.dev
->collection
;
287 static struct its_collection
*its_build_discard_cmd(struct its_cmd_block
*cmd
,
288 struct its_cmd_desc
*desc
)
290 its_encode_cmd(cmd
, GITS_CMD_DISCARD
);
291 its_encode_devid(cmd
, desc
->its_discard_cmd
.dev
->device_id
);
292 its_encode_event_id(cmd
, desc
->its_discard_cmd
.event_id
);
296 return desc
->its_discard_cmd
.dev
->collection
;
299 static struct its_collection
*its_build_inv_cmd(struct its_cmd_block
*cmd
,
300 struct its_cmd_desc
*desc
)
302 its_encode_cmd(cmd
, GITS_CMD_INV
);
303 its_encode_devid(cmd
, desc
->its_inv_cmd
.dev
->device_id
);
304 its_encode_event_id(cmd
, desc
->its_inv_cmd
.event_id
);
308 return desc
->its_inv_cmd
.dev
->collection
;
311 static struct its_collection
*its_build_invall_cmd(struct its_cmd_block
*cmd
,
312 struct its_cmd_desc
*desc
)
314 its_encode_cmd(cmd
, GITS_CMD_INVALL
);
315 its_encode_collection(cmd
, desc
->its_mapc_cmd
.col
->col_id
);
322 static u64
its_cmd_ptr_to_offset(struct its_node
*its
,
323 struct its_cmd_block
*ptr
)
325 return (ptr
- its
->cmd_base
) * sizeof(*ptr
);
328 static int its_queue_full(struct its_node
*its
)
333 widx
= its
->cmd_write
- its
->cmd_base
;
334 ridx
= readl_relaxed(its
->base
+ GITS_CREADR
) / sizeof(struct its_cmd_block
);
336 /* This is incredibly unlikely to happen, unless the ITS locks up. */
337 if (((widx
+ 1) % ITS_CMD_QUEUE_NR_ENTRIES
) == ridx
)
343 static struct its_cmd_block
*its_allocate_entry(struct its_node
*its
)
345 struct its_cmd_block
*cmd
;
346 u32 count
= 1000000; /* 1s! */
348 while (its_queue_full(its
)) {
351 pr_err_ratelimited("ITS queue not draining\n");
358 cmd
= its
->cmd_write
++;
360 /* Handle queue wrapping */
361 if (its
->cmd_write
== (its
->cmd_base
+ ITS_CMD_QUEUE_NR_ENTRIES
))
362 its
->cmd_write
= its
->cmd_base
;
367 static struct its_cmd_block
*its_post_commands(struct its_node
*its
)
369 u64 wr
= its_cmd_ptr_to_offset(its
, its
->cmd_write
);
371 writel_relaxed(wr
, its
->base
+ GITS_CWRITER
);
373 return its
->cmd_write
;
376 static void its_flush_cmd(struct its_node
*its
, struct its_cmd_block
*cmd
)
379 * Make sure the commands written to memory are observable by
382 if (its
->flags
& ITS_FLAGS_CMDQ_NEEDS_FLUSHING
)
383 __flush_dcache_area(cmd
, sizeof(*cmd
));
388 static void its_wait_for_range_completion(struct its_node
*its
,
389 struct its_cmd_block
*from
,
390 struct its_cmd_block
*to
)
392 u64 rd_idx
, from_idx
, to_idx
;
393 u32 count
= 1000000; /* 1s! */
395 from_idx
= its_cmd_ptr_to_offset(its
, from
);
396 to_idx
= its_cmd_ptr_to_offset(its
, to
);
399 rd_idx
= readl_relaxed(its
->base
+ GITS_CREADR
);
400 if (rd_idx
>= to_idx
|| rd_idx
< from_idx
)
405 pr_err_ratelimited("ITS queue timeout\n");
413 static void its_send_single_command(struct its_node
*its
,
414 its_cmd_builder_t builder
,
415 struct its_cmd_desc
*desc
)
417 struct its_cmd_block
*cmd
, *sync_cmd
, *next_cmd
;
418 struct its_collection
*sync_col
;
421 raw_spin_lock_irqsave(&its
->lock
, flags
);
423 cmd
= its_allocate_entry(its
);
424 if (!cmd
) { /* We're soooooo screewed... */
425 pr_err_ratelimited("ITS can't allocate, dropping command\n");
426 raw_spin_unlock_irqrestore(&its
->lock
, flags
);
429 sync_col
= builder(cmd
, desc
);
430 its_flush_cmd(its
, cmd
);
433 sync_cmd
= its_allocate_entry(its
);
435 pr_err_ratelimited("ITS can't SYNC, skipping\n");
438 its_encode_cmd(sync_cmd
, GITS_CMD_SYNC
);
439 its_encode_target(sync_cmd
, sync_col
->target_address
);
440 its_fixup_cmd(sync_cmd
);
441 its_flush_cmd(its
, sync_cmd
);
445 next_cmd
= its_post_commands(its
);
446 raw_spin_unlock_irqrestore(&its
->lock
, flags
);
448 its_wait_for_range_completion(its
, cmd
, next_cmd
);
451 static void its_send_inv(struct its_device
*dev
, u32 event_id
)
453 struct its_cmd_desc desc
;
455 desc
.its_inv_cmd
.dev
= dev
;
456 desc
.its_inv_cmd
.event_id
= event_id
;
458 its_send_single_command(dev
->its
, its_build_inv_cmd
, &desc
);
461 static void its_send_mapd(struct its_device
*dev
, int valid
)
463 struct its_cmd_desc desc
;
465 desc
.its_mapd_cmd
.dev
= dev
;
466 desc
.its_mapd_cmd
.valid
= !!valid
;
468 its_send_single_command(dev
->its
, its_build_mapd_cmd
, &desc
);
471 static void its_send_mapc(struct its_node
*its
, struct its_collection
*col
,
474 struct its_cmd_desc desc
;
476 desc
.its_mapc_cmd
.col
= col
;
477 desc
.its_mapc_cmd
.valid
= !!valid
;
479 its_send_single_command(its
, its_build_mapc_cmd
, &desc
);
482 static void its_send_mapvi(struct its_device
*dev
, u32 irq_id
, u32 id
)
484 struct its_cmd_desc desc
;
486 desc
.its_mapvi_cmd
.dev
= dev
;
487 desc
.its_mapvi_cmd
.phys_id
= irq_id
;
488 desc
.its_mapvi_cmd
.event_id
= id
;
490 its_send_single_command(dev
->its
, its_build_mapvi_cmd
, &desc
);
493 static void its_send_movi(struct its_device
*dev
,
494 struct its_collection
*col
, u32 id
)
496 struct its_cmd_desc desc
;
498 desc
.its_movi_cmd
.dev
= dev
;
499 desc
.its_movi_cmd
.col
= col
;
500 desc
.its_movi_cmd
.id
= id
;
502 its_send_single_command(dev
->its
, its_build_movi_cmd
, &desc
);
505 static void its_send_discard(struct its_device
*dev
, u32 id
)
507 struct its_cmd_desc desc
;
509 desc
.its_discard_cmd
.dev
= dev
;
510 desc
.its_discard_cmd
.event_id
= id
;
512 its_send_single_command(dev
->its
, its_build_discard_cmd
, &desc
);
515 static void its_send_invall(struct its_node
*its
, struct its_collection
*col
)
517 struct its_cmd_desc desc
;
519 desc
.its_invall_cmd
.col
= col
;
521 its_send_single_command(its
, its_build_invall_cmd
, &desc
);
525 * irqchip functions - assumes MSI, mostly.
528 static inline u32
its_get_event_id(struct irq_data
*d
)
530 struct its_device
*its_dev
= irq_data_get_irq_chip_data(d
);
531 return d
->hwirq
- its_dev
->lpi_base
;
534 static void lpi_set_config(struct irq_data
*d
, bool enable
)
536 struct its_device
*its_dev
= irq_data_get_irq_chip_data(d
);
537 irq_hw_number_t hwirq
= d
->hwirq
;
538 u32 id
= its_get_event_id(d
);
539 u8
*cfg
= page_address(gic_rdists
->prop_page
) + hwirq
- 8192;
542 *cfg
|= LPI_PROP_ENABLED
;
544 *cfg
&= ~LPI_PROP_ENABLED
;
547 * Make the above write visible to the redistributors.
548 * And yes, we're flushing exactly: One. Single. Byte.
551 if (gic_rdists
->flags
& RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING
)
552 __flush_dcache_area(cfg
, sizeof(*cfg
));
555 its_send_inv(its_dev
, id
);
558 static void its_mask_irq(struct irq_data
*d
)
560 lpi_set_config(d
, false);
563 static void its_unmask_irq(struct irq_data
*d
)
565 lpi_set_config(d
, true);
568 static void its_eoi_irq(struct irq_data
*d
)
570 gic_write_eoir(d
->hwirq
);
573 static int its_set_affinity(struct irq_data
*d
, const struct cpumask
*mask_val
,
576 unsigned int cpu
= cpumask_any_and(mask_val
, cpu_online_mask
);
577 struct its_device
*its_dev
= irq_data_get_irq_chip_data(d
);
578 struct its_collection
*target_col
;
579 u32 id
= its_get_event_id(d
);
581 if (cpu
>= nr_cpu_ids
)
584 target_col
= &its_dev
->its
->collections
[cpu
];
585 its_send_movi(its_dev
, target_col
, id
);
586 its_dev
->collection
= target_col
;
588 return IRQ_SET_MASK_OK_DONE
;
591 static void its_irq_compose_msi_msg(struct irq_data
*d
, struct msi_msg
*msg
)
593 struct its_device
*its_dev
= irq_data_get_irq_chip_data(d
);
594 struct its_node
*its
;
598 addr
= its
->phys_base
+ GITS_TRANSLATER
;
600 msg
->address_lo
= addr
& ((1UL << 32) - 1);
601 msg
->address_hi
= addr
>> 32;
602 msg
->data
= its_get_event_id(d
);
605 static struct irq_chip its_irq_chip
= {
607 .irq_mask
= its_mask_irq
,
608 .irq_unmask
= its_unmask_irq
,
609 .irq_eoi
= its_eoi_irq
,
610 .irq_set_affinity
= its_set_affinity
,
611 .irq_compose_msi_msg
= its_irq_compose_msi_msg
,
614 static void its_mask_msi_irq(struct irq_data
*d
)
617 irq_chip_mask_parent(d
);
620 static void its_unmask_msi_irq(struct irq_data
*d
)
622 pci_msi_unmask_irq(d
);
623 irq_chip_unmask_parent(d
);
626 static struct irq_chip its_msi_irq_chip
= {
628 .irq_unmask
= its_unmask_msi_irq
,
629 .irq_mask
= its_mask_msi_irq
,
630 .irq_eoi
= irq_chip_eoi_parent
,
631 .irq_write_msi_msg
= pci_msi_domain_write_msg
,
635 * How we allocate LPIs:
637 * The GIC has id_bits bits for interrupt identifiers. From there, we
638 * must subtract 8192 which are reserved for SGIs/PPIs/SPIs. Then, as
639 * we allocate LPIs by chunks of 32, we can shift the whole thing by 5
642 * This gives us (((1UL << id_bits) - 8192) >> 5) possible allocations.
644 #define IRQS_PER_CHUNK_SHIFT 5
645 #define IRQS_PER_CHUNK (1 << IRQS_PER_CHUNK_SHIFT)
647 static unsigned long *lpi_bitmap
;
648 static u32 lpi_chunks
;
649 static DEFINE_SPINLOCK(lpi_lock
);
651 static int its_lpi_to_chunk(int lpi
)
653 return (lpi
- 8192) >> IRQS_PER_CHUNK_SHIFT
;
656 static int its_chunk_to_lpi(int chunk
)
658 return (chunk
<< IRQS_PER_CHUNK_SHIFT
) + 8192;
661 static int its_lpi_init(u32 id_bits
)
663 lpi_chunks
= its_lpi_to_chunk(1UL << id_bits
);
665 lpi_bitmap
= kzalloc(BITS_TO_LONGS(lpi_chunks
) * sizeof(long),
672 pr_info("ITS: Allocated %d chunks for LPIs\n", (int)lpi_chunks
);
676 static unsigned long *its_lpi_alloc_chunks(int nr_irqs
, int *base
, int *nr_ids
)
678 unsigned long *bitmap
= NULL
;
683 nr_chunks
= DIV_ROUND_UP(nr_irqs
, IRQS_PER_CHUNK
);
685 spin_lock(&lpi_lock
);
688 chunk_id
= bitmap_find_next_zero_area(lpi_bitmap
, lpi_chunks
,
690 if (chunk_id
< lpi_chunks
)
694 } while (nr_chunks
> 0);
699 bitmap
= kzalloc(BITS_TO_LONGS(nr_chunks
* IRQS_PER_CHUNK
) * sizeof (long),
704 for (i
= 0; i
< nr_chunks
; i
++)
705 set_bit(chunk_id
+ i
, lpi_bitmap
);
707 *base
= its_chunk_to_lpi(chunk_id
);
708 *nr_ids
= nr_chunks
* IRQS_PER_CHUNK
;
711 spin_unlock(&lpi_lock
);
716 static void its_lpi_free(unsigned long *bitmap
, int base
, int nr_ids
)
720 spin_lock(&lpi_lock
);
722 for (lpi
= base
; lpi
< (base
+ nr_ids
); lpi
+= IRQS_PER_CHUNK
) {
723 int chunk
= its_lpi_to_chunk(lpi
);
724 BUG_ON(chunk
> lpi_chunks
);
725 if (test_bit(chunk
, lpi_bitmap
)) {
726 clear_bit(chunk
, lpi_bitmap
);
728 pr_err("Bad LPI chunk %d\n", chunk
);
732 spin_unlock(&lpi_lock
);
738 * We allocate 64kB for PROPBASE. That gives us at most 64K LPIs to
739 * deal with (one configuration byte per interrupt). PENDBASE has to
740 * be 64kB aligned (one bit per LPI, plus 8192 bits for SPI/PPI/SGI).
742 #define LPI_PROPBASE_SZ SZ_64K
743 #define LPI_PENDBASE_SZ (LPI_PROPBASE_SZ / 8 + SZ_1K)
746 * This is how many bits of ID we need, including the useless ones.
748 #define LPI_NRBITS ilog2(LPI_PROPBASE_SZ + SZ_8K)
750 #define LPI_PROP_DEFAULT_PRIO 0xa0
752 static int __init
its_alloc_lpi_tables(void)
756 gic_rdists
->prop_page
= alloc_pages(GFP_NOWAIT
,
757 get_order(LPI_PROPBASE_SZ
));
758 if (!gic_rdists
->prop_page
) {
759 pr_err("Failed to allocate PROPBASE\n");
763 paddr
= page_to_phys(gic_rdists
->prop_page
);
764 pr_info("GIC: using LPI property table @%pa\n", &paddr
);
766 /* Priority 0xa0, Group-1, disabled */
767 memset(page_address(gic_rdists
->prop_page
),
768 LPI_PROP_DEFAULT_PRIO
| LPI_PROP_GROUP1
,
771 /* Make sure the GIC will observe the written configuration */
772 __flush_dcache_area(page_address(gic_rdists
->prop_page
), LPI_PROPBASE_SZ
);
777 static const char *its_base_type_string
[] = {
778 [GITS_BASER_TYPE_DEVICE
] = "Devices",
779 [GITS_BASER_TYPE_VCPU
] = "Virtual CPUs",
780 [GITS_BASER_TYPE_CPU
] = "Physical CPUs",
781 [GITS_BASER_TYPE_COLLECTION
] = "Interrupt Collections",
782 [GITS_BASER_TYPE_RESERVED5
] = "Reserved (5)",
783 [GITS_BASER_TYPE_RESERVED6
] = "Reserved (6)",
784 [GITS_BASER_TYPE_RESERVED7
] = "Reserved (7)",
787 static void its_free_tables(struct its_node
*its
)
791 for (i
= 0; i
< GITS_BASER_NR_REGS
; i
++) {
792 if (its
->tables
[i
]) {
793 free_page((unsigned long)its
->tables
[i
]);
794 its
->tables
[i
] = NULL
;
799 static int its_alloc_tables(struct its_node
*its
)
804 u64 shr
= GITS_BASER_InnerShareable
;
805 u64 cache
= GITS_BASER_WaWb
;
807 for (i
= 0; i
< GITS_BASER_NR_REGS
; i
++) {
808 u64 val
= readq_relaxed(its
->base
+ GITS_BASER
+ i
* 8);
809 u64 type
= GITS_BASER_TYPE(val
);
810 u64 entry_size
= GITS_BASER_ENTRY_SIZE(val
);
811 int order
= get_order(psz
);
816 if (type
== GITS_BASER_TYPE_NONE
)
820 * Allocate as many entries as required to fit the
821 * range of device IDs that the ITS can grok... The ID
822 * space being incredibly sparse, this results in a
823 * massive waste of memory.
825 * For other tables, only allocate a single page.
827 if (type
== GITS_BASER_TYPE_DEVICE
) {
828 u64 typer
= readq_relaxed(its
->base
+ GITS_TYPER
);
829 u32 ids
= GITS_TYPER_DEVBITS(typer
);
831 order
= get_order((1UL << ids
) * entry_size
);
832 if (order
>= MAX_ORDER
) {
833 order
= MAX_ORDER
- 1;
834 pr_warn("%s: Device Table too large, reduce its page order to %u\n",
835 its
->msi_chip
.of_node
->full_name
, order
);
839 alloc_size
= (1 << order
) * PAGE_SIZE
;
840 base
= (void *)__get_free_pages(GFP_KERNEL
| __GFP_ZERO
, order
);
846 its
->tables
[i
] = base
;
849 val
= (virt_to_phys(base
) |
850 (type
<< GITS_BASER_TYPE_SHIFT
) |
851 ((entry_size
- 1) << GITS_BASER_ENTRY_SIZE_SHIFT
) |
858 val
|= GITS_BASER_PAGE_SIZE_4K
;
861 val
|= GITS_BASER_PAGE_SIZE_16K
;
864 val
|= GITS_BASER_PAGE_SIZE_64K
;
868 val
|= (alloc_size
/ psz
) - 1;
870 writeq_relaxed(val
, its
->base
+ GITS_BASER
+ i
* 8);
871 tmp
= readq_relaxed(its
->base
+ GITS_BASER
+ i
* 8);
873 if ((val
^ tmp
) & GITS_BASER_SHAREABILITY_MASK
) {
875 * Shareability didn't stick. Just use
876 * whatever the read reported, which is likely
877 * to be the only thing this redistributor
878 * supports. If that's zero, make it
879 * non-cacheable as well.
881 shr
= tmp
& GITS_BASER_SHAREABILITY_MASK
;
883 cache
= GITS_BASER_nC
;
887 if ((val
^ tmp
) & GITS_BASER_PAGE_SIZE_MASK
) {
889 * Page size didn't stick. Let's try a smaller
890 * size and retry. If we reach 4K, then
891 * something is horribly wrong...
904 pr_err("ITS: %s: GITS_BASER%d doesn't stick: %lx %lx\n",
905 its
->msi_chip
.of_node
->full_name
, i
,
906 (unsigned long) val
, (unsigned long) tmp
);
911 pr_info("ITS: allocated %d %s @%lx (psz %dK, shr %d)\n",
912 (int)(alloc_size
/ entry_size
),
913 its_base_type_string
[type
],
914 (unsigned long)virt_to_phys(base
),
915 psz
/ SZ_1K
, (int)shr
>> GITS_BASER_SHAREABILITY_SHIFT
);
921 its_free_tables(its
);
926 static int its_alloc_collections(struct its_node
*its
)
928 its
->collections
= kzalloc(nr_cpu_ids
* sizeof(*its
->collections
),
930 if (!its
->collections
)
936 static void its_cpu_init_lpis(void)
938 void __iomem
*rbase
= gic_data_rdist_rd_base();
939 struct page
*pend_page
;
942 /* If we didn't allocate the pending table yet, do it now */
943 pend_page
= gic_data_rdist()->pend_page
;
947 * The pending pages have to be at least 64kB aligned,
948 * hence the 'max(LPI_PENDBASE_SZ, SZ_64K)' below.
950 pend_page
= alloc_pages(GFP_NOWAIT
| __GFP_ZERO
,
951 get_order(max(LPI_PENDBASE_SZ
, SZ_64K
)));
953 pr_err("Failed to allocate PENDBASE for CPU%d\n",
958 /* Make sure the GIC will observe the zero-ed page */
959 __flush_dcache_area(page_address(pend_page
), LPI_PENDBASE_SZ
);
961 paddr
= page_to_phys(pend_page
);
962 pr_info("CPU%d: using LPI pending table @%pa\n",
963 smp_processor_id(), &paddr
);
964 gic_data_rdist()->pend_page
= pend_page
;
968 val
= readl_relaxed(rbase
+ GICR_CTLR
);
969 val
&= ~GICR_CTLR_ENABLE_LPIS
;
970 writel_relaxed(val
, rbase
+ GICR_CTLR
);
973 * Make sure any change to the table is observable by the GIC.
978 val
= (page_to_phys(gic_rdists
->prop_page
) |
979 GICR_PROPBASER_InnerShareable
|
980 GICR_PROPBASER_WaWb
|
981 ((LPI_NRBITS
- 1) & GICR_PROPBASER_IDBITS_MASK
));
983 writeq_relaxed(val
, rbase
+ GICR_PROPBASER
);
984 tmp
= readq_relaxed(rbase
+ GICR_PROPBASER
);
986 if ((tmp
^ val
) & GICR_PROPBASER_SHAREABILITY_MASK
) {
987 if (!(tmp
& GICR_PROPBASER_SHAREABILITY_MASK
)) {
989 * The HW reports non-shareable, we must
990 * remove the cacheability attributes as
993 val
&= ~(GICR_PROPBASER_SHAREABILITY_MASK
|
994 GICR_PROPBASER_CACHEABILITY_MASK
);
995 val
|= GICR_PROPBASER_nC
;
996 writeq_relaxed(val
, rbase
+ GICR_PROPBASER
);
998 pr_info_once("GIC: using cache flushing for LPI property table\n");
999 gic_rdists
->flags
|= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING
;
1003 val
= (page_to_phys(pend_page
) |
1004 GICR_PENDBASER_InnerShareable
|
1005 GICR_PENDBASER_WaWb
);
1007 writeq_relaxed(val
, rbase
+ GICR_PENDBASER
);
1008 tmp
= readq_relaxed(rbase
+ GICR_PENDBASER
);
1010 if (!(tmp
& GICR_PENDBASER_SHAREABILITY_MASK
)) {
1012 * The HW reports non-shareable, we must remove the
1013 * cacheability attributes as well.
1015 val
&= ~(GICR_PENDBASER_SHAREABILITY_MASK
|
1016 GICR_PENDBASER_CACHEABILITY_MASK
);
1017 val
|= GICR_PENDBASER_nC
;
1018 writeq_relaxed(val
, rbase
+ GICR_PENDBASER
);
1022 val
= readl_relaxed(rbase
+ GICR_CTLR
);
1023 val
|= GICR_CTLR_ENABLE_LPIS
;
1024 writel_relaxed(val
, rbase
+ GICR_CTLR
);
1026 /* Make sure the GIC has seen the above */
1030 static void its_cpu_init_collection(void)
1032 struct its_node
*its
;
1035 spin_lock(&its_lock
);
1036 cpu
= smp_processor_id();
1038 list_for_each_entry(its
, &its_nodes
, entry
) {
1042 * We now have to bind each collection to its target
1045 if (readq_relaxed(its
->base
+ GITS_TYPER
) & GITS_TYPER_PTA
) {
1047 * This ITS wants the physical address of the
1050 target
= gic_data_rdist()->phys_base
;
1053 * This ITS wants a linear CPU number.
1055 target
= readq_relaxed(gic_data_rdist_rd_base() + GICR_TYPER
);
1056 target
= GICR_TYPER_CPU_NUMBER(target
) << 16;
1059 /* Perform collection mapping */
1060 its
->collections
[cpu
].target_address
= target
;
1061 its
->collections
[cpu
].col_id
= cpu
;
1063 its_send_mapc(its
, &its
->collections
[cpu
], 1);
1064 its_send_invall(its
, &its
->collections
[cpu
]);
1067 spin_unlock(&its_lock
);
1070 static struct its_device
*its_find_device(struct its_node
*its
, u32 dev_id
)
1072 struct its_device
*its_dev
= NULL
, *tmp
;
1073 unsigned long flags
;
1075 raw_spin_lock_irqsave(&its
->lock
, flags
);
1077 list_for_each_entry(tmp
, &its
->its_device_list
, entry
) {
1078 if (tmp
->device_id
== dev_id
) {
1084 raw_spin_unlock_irqrestore(&its
->lock
, flags
);
1089 static struct its_device
*its_create_device(struct its_node
*its
, u32 dev_id
,
1092 struct its_device
*dev
;
1093 unsigned long *lpi_map
;
1094 unsigned long flags
;
1102 dev
= kzalloc(sizeof(*dev
), GFP_KERNEL
);
1104 * At least one bit of EventID is being used, hence a minimum
1105 * of two entries. No, the architecture doesn't let you
1106 * express an ITT with a single entry.
1108 nr_ites
= max(2UL, roundup_pow_of_two(nvecs
));
1109 sz
= nr_ites
* its
->ite_size
;
1110 sz
= max(sz
, ITS_ITT_ALIGN
) + ITS_ITT_ALIGN
- 1;
1111 itt
= kzalloc(sz
, GFP_KERNEL
);
1112 lpi_map
= its_lpi_alloc_chunks(nvecs
, &lpi_base
, &nr_lpis
);
1114 if (!dev
|| !itt
|| !lpi_map
) {
1123 dev
->nr_ites
= nr_ites
;
1124 dev
->lpi_map
= lpi_map
;
1125 dev
->lpi_base
= lpi_base
;
1126 dev
->nr_lpis
= nr_lpis
;
1127 dev
->device_id
= dev_id
;
1128 INIT_LIST_HEAD(&dev
->entry
);
1130 raw_spin_lock_irqsave(&its
->lock
, flags
);
1131 list_add(&dev
->entry
, &its
->its_device_list
);
1132 raw_spin_unlock_irqrestore(&its
->lock
, flags
);
1134 /* Bind the device to the first possible CPU */
1135 cpu
= cpumask_first(cpu_online_mask
);
1136 dev
->collection
= &its
->collections
[cpu
];
1138 /* Map device to its ITT */
1139 its_send_mapd(dev
, 1);
1144 static void its_free_device(struct its_device
*its_dev
)
1146 unsigned long flags
;
1148 raw_spin_lock_irqsave(&its_dev
->its
->lock
, flags
);
1149 list_del(&its_dev
->entry
);
1150 raw_spin_unlock_irqrestore(&its_dev
->its
->lock
, flags
);
1151 kfree(its_dev
->itt
);
1155 static int its_alloc_device_irq(struct its_device
*dev
, irq_hw_number_t
*hwirq
)
1159 idx
= find_first_zero_bit(dev
->lpi_map
, dev
->nr_lpis
);
1160 if (idx
== dev
->nr_lpis
)
1163 *hwirq
= dev
->lpi_base
+ idx
;
1164 set_bit(idx
, dev
->lpi_map
);
1169 struct its_pci_alias
{
1170 struct pci_dev
*pdev
;
1175 static int its_pci_msi_vec_count(struct pci_dev
*pdev
)
1179 msi
= max(pci_msi_vec_count(pdev
), 0);
1180 msix
= max(pci_msix_vec_count(pdev
), 0);
1182 return max(msi
, msix
);
1185 static int its_get_pci_alias(struct pci_dev
*pdev
, u16 alias
, void *data
)
1187 struct its_pci_alias
*dev_alias
= data
;
1189 dev_alias
->dev_id
= alias
;
1190 if (pdev
!= dev_alias
->pdev
)
1191 dev_alias
->count
+= its_pci_msi_vec_count(dev_alias
->pdev
);
1196 static int its_msi_prepare(struct irq_domain
*domain
, struct device
*dev
,
1197 int nvec
, msi_alloc_info_t
*info
)
1199 struct pci_dev
*pdev
;
1200 struct its_node
*its
;
1201 struct its_device
*its_dev
;
1202 struct its_pci_alias dev_alias
;
1204 if (!dev_is_pci(dev
))
1207 pdev
= to_pci_dev(dev
);
1208 dev_alias
.pdev
= pdev
;
1209 dev_alias
.count
= nvec
;
1211 pci_for_each_dma_alias(pdev
, its_get_pci_alias
, &dev_alias
);
1212 its
= domain
->parent
->host_data
;
1214 its_dev
= its_find_device(its
, dev_alias
.dev_id
);
1217 * We already have seen this ID, probably through
1218 * another alias (PCI bridge of some sort). No need to
1219 * create the device.
1221 dev_dbg(dev
, "Reusing ITT for devID %x\n", dev_alias
.dev_id
);
1225 its_dev
= its_create_device(its
, dev_alias
.dev_id
, dev_alias
.count
);
1229 dev_dbg(&pdev
->dev
, "ITT %d entries, %d bits\n",
1230 dev_alias
.count
, ilog2(dev_alias
.count
));
1232 info
->scratchpad
[0].ptr
= its_dev
;
1233 info
->scratchpad
[1].ptr
= dev
;
1237 static struct msi_domain_ops its_pci_msi_ops
= {
1238 .msi_prepare
= its_msi_prepare
,
1241 static struct msi_domain_info its_pci_msi_domain_info
= {
1242 .flags
= (MSI_FLAG_USE_DEF_DOM_OPS
| MSI_FLAG_USE_DEF_CHIP_OPS
|
1243 MSI_FLAG_MULTI_PCI_MSI
| MSI_FLAG_PCI_MSIX
),
1244 .ops
= &its_pci_msi_ops
,
1245 .chip
= &its_msi_irq_chip
,
1248 static int its_irq_gic_domain_alloc(struct irq_domain
*domain
,
1250 irq_hw_number_t hwirq
)
1252 struct of_phandle_args args
;
1254 args
.np
= domain
->parent
->of_node
;
1255 args
.args_count
= 3;
1256 args
.args
[0] = GIC_IRQ_TYPE_LPI
;
1257 args
.args
[1] = hwirq
;
1258 args
.args
[2] = IRQ_TYPE_EDGE_RISING
;
1260 return irq_domain_alloc_irqs_parent(domain
, virq
, 1, &args
);
1263 static int its_irq_domain_alloc(struct irq_domain
*domain
, unsigned int virq
,
1264 unsigned int nr_irqs
, void *args
)
1266 msi_alloc_info_t
*info
= args
;
1267 struct its_device
*its_dev
= info
->scratchpad
[0].ptr
;
1268 irq_hw_number_t hwirq
;
1272 for (i
= 0; i
< nr_irqs
; i
++) {
1273 err
= its_alloc_device_irq(its_dev
, &hwirq
);
1277 err
= its_irq_gic_domain_alloc(domain
, virq
+ i
, hwirq
);
1281 irq_domain_set_hwirq_and_chip(domain
, virq
+ i
,
1282 hwirq
, &its_irq_chip
, its_dev
);
1283 dev_dbg(info
->scratchpad
[1].ptr
, "ID:%d pID:%d vID:%d\n",
1284 (int)(hwirq
- its_dev
->lpi_base
), (int)hwirq
, virq
+ i
);
1290 static void its_irq_domain_activate(struct irq_domain
*domain
,
1293 struct its_device
*its_dev
= irq_data_get_irq_chip_data(d
);
1294 u32 event
= its_get_event_id(d
);
1296 /* Map the GIC IRQ and event to the device */
1297 its_send_mapvi(its_dev
, d
->hwirq
, event
);
1300 static void its_irq_domain_deactivate(struct irq_domain
*domain
,
1303 struct its_device
*its_dev
= irq_data_get_irq_chip_data(d
);
1304 u32 event
= its_get_event_id(d
);
1306 /* Stop the delivery of interrupts */
1307 its_send_discard(its_dev
, event
);
1310 static void its_irq_domain_free(struct irq_domain
*domain
, unsigned int virq
,
1311 unsigned int nr_irqs
)
1313 struct irq_data
*d
= irq_domain_get_irq_data(domain
, virq
);
1314 struct its_device
*its_dev
= irq_data_get_irq_chip_data(d
);
1317 for (i
= 0; i
< nr_irqs
; i
++) {
1318 struct irq_data
*data
= irq_domain_get_irq_data(domain
,
1320 u32 event
= its_get_event_id(data
);
1322 /* Mark interrupt index as unused */
1323 clear_bit(event
, its_dev
->lpi_map
);
1325 /* Nuke the entry in the domain */
1326 irq_domain_reset_irq_data(data
);
1329 /* If all interrupts have been freed, start mopping the floor */
1330 if (bitmap_empty(its_dev
->lpi_map
, its_dev
->nr_lpis
)) {
1331 its_lpi_free(its_dev
->lpi_map
,
1335 /* Unmap device/itt */
1336 its_send_mapd(its_dev
, 0);
1337 its_free_device(its_dev
);
1340 irq_domain_free_irqs_parent(domain
, virq
, nr_irqs
);
1343 static const struct irq_domain_ops its_domain_ops
= {
1344 .alloc
= its_irq_domain_alloc
,
1345 .free
= its_irq_domain_free
,
1346 .activate
= its_irq_domain_activate
,
1347 .deactivate
= its_irq_domain_deactivate
,
1350 static int its_force_quiescent(void __iomem
*base
)
1352 u32 count
= 1000000; /* 1s */
1355 val
= readl_relaxed(base
+ GITS_CTLR
);
1356 if (val
& GITS_CTLR_QUIESCENT
)
1359 /* Disable the generation of all interrupts to this ITS */
1360 val
&= ~GITS_CTLR_ENABLE
;
1361 writel_relaxed(val
, base
+ GITS_CTLR
);
1363 /* Poll GITS_CTLR and wait until ITS becomes quiescent */
1365 val
= readl_relaxed(base
+ GITS_CTLR
);
1366 if (val
& GITS_CTLR_QUIESCENT
)
1378 static int its_probe(struct device_node
*node
, struct irq_domain
*parent
)
1380 struct resource res
;
1381 struct its_node
*its
;
1382 void __iomem
*its_base
;
1387 err
= of_address_to_resource(node
, 0, &res
);
1389 pr_warn("%s: no regs?\n", node
->full_name
);
1393 its_base
= ioremap(res
.start
, resource_size(&res
));
1395 pr_warn("%s: unable to map registers\n", node
->full_name
);
1399 val
= readl_relaxed(its_base
+ GITS_PIDR2
) & GIC_PIDR2_ARCH_MASK
;
1400 if (val
!= 0x30 && val
!= 0x40) {
1401 pr_warn("%s: no ITS detected, giving up\n", node
->full_name
);
1406 err
= its_force_quiescent(its_base
);
1408 pr_warn("%s: failed to quiesce, giving up\n",
1413 pr_info("ITS: %s\n", node
->full_name
);
1415 its
= kzalloc(sizeof(*its
), GFP_KERNEL
);
1421 raw_spin_lock_init(&its
->lock
);
1422 INIT_LIST_HEAD(&its
->entry
);
1423 INIT_LIST_HEAD(&its
->its_device_list
);
1424 its
->base
= its_base
;
1425 its
->phys_base
= res
.start
;
1426 its
->msi_chip
.of_node
= node
;
1427 its
->ite_size
= ((readl_relaxed(its_base
+ GITS_TYPER
) >> 4) & 0xf) + 1;
1429 its
->cmd_base
= kzalloc(ITS_CMD_QUEUE_SZ
, GFP_KERNEL
);
1430 if (!its
->cmd_base
) {
1434 its
->cmd_write
= its
->cmd_base
;
1436 err
= its_alloc_tables(its
);
1440 err
= its_alloc_collections(its
);
1442 goto out_free_tables
;
1444 baser
= (virt_to_phys(its
->cmd_base
) |
1446 GITS_CBASER_InnerShareable
|
1447 (ITS_CMD_QUEUE_SZ
/ SZ_4K
- 1) |
1450 writeq_relaxed(baser
, its
->base
+ GITS_CBASER
);
1451 tmp
= readq_relaxed(its
->base
+ GITS_CBASER
);
1453 if ((tmp
^ baser
) & GITS_CBASER_SHAREABILITY_MASK
) {
1454 if (!(tmp
& GITS_CBASER_SHAREABILITY_MASK
)) {
1456 * The HW reports non-shareable, we must
1457 * remove the cacheability attributes as
1460 baser
&= ~(GITS_CBASER_SHAREABILITY_MASK
|
1461 GITS_CBASER_CACHEABILITY_MASK
);
1462 baser
|= GITS_CBASER_nC
;
1463 writeq_relaxed(baser
, its
->base
+ GITS_CBASER
);
1465 pr_info("ITS: using cache flushing for cmd queue\n");
1466 its
->flags
|= ITS_FLAGS_CMDQ_NEEDS_FLUSHING
;
1469 writeq_relaxed(0, its
->base
+ GITS_CWRITER
);
1470 writel_relaxed(GITS_CTLR_ENABLE
, its
->base
+ GITS_CTLR
);
1472 if (of_property_read_bool(its
->msi_chip
.of_node
, "msi-controller")) {
1473 its
->domain
= irq_domain_add_tree(NULL
, &its_domain_ops
, its
);
1476 goto out_free_tables
;
1479 its
->domain
->parent
= parent
;
1481 its
->msi_chip
.domain
= pci_msi_create_irq_domain(node
,
1482 &its_pci_msi_domain_info
,
1484 if (!its
->msi_chip
.domain
) {
1486 goto out_free_domains
;
1489 err
= of_pci_msi_chip_add(&its
->msi_chip
);
1491 goto out_free_domains
;
1494 spin_lock(&its_lock
);
1495 list_add(&its
->entry
, &its_nodes
);
1496 spin_unlock(&its_lock
);
1501 if (its
->msi_chip
.domain
)
1502 irq_domain_remove(its
->msi_chip
.domain
);
1504 irq_domain_remove(its
->domain
);
1506 its_free_tables(its
);
1508 kfree(its
->cmd_base
);
1513 pr_err("ITS: failed probing %s (%d)\n", node
->full_name
, err
);
1517 static bool gic_rdists_supports_plpis(void)
1519 return !!(readl_relaxed(gic_data_rdist_rd_base() + GICR_TYPER
) & GICR_TYPER_PLPIS
);
1522 int its_cpu_init(void)
1524 if (!list_empty(&its_nodes
)) {
1525 if (!gic_rdists_supports_plpis()) {
1526 pr_info("CPU%d: LPIs not supported\n", smp_processor_id());
1529 its_cpu_init_lpis();
1530 its_cpu_init_collection();
1536 static struct of_device_id its_device_id
[] = {
1537 { .compatible
= "arm,gic-v3-its", },
1541 int its_init(struct device_node
*node
, struct rdists
*rdists
,
1542 struct irq_domain
*parent_domain
)
1544 struct device_node
*np
;
1546 for (np
= of_find_matching_node(node
, its_device_id
); np
;
1547 np
= of_find_matching_node(np
, its_device_id
)) {
1548 its_probe(np
, parent_domain
);
1551 if (list_empty(&its_nodes
)) {
1552 pr_warn("ITS: No ITS available, not enabling LPIs\n");
1556 gic_rdists
= rdists
;
1557 gic_root_node
= node
;
1559 its_alloc_lpi_tables();
1560 its_lpi_init(rdists
->id_bits
);