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.h>
34 #include <linux/irqchip/arm-gic-v3.h>
36 #include <asm/cacheflush.h>
37 #include <asm/cputype.h>
38 #include <asm/exception.h>
40 #include "irq-gic-common.h"
42 #define ITS_FLAGS_CMDQ_NEEDS_FLUSHING (1ULL << 0)
43 #define ITS_FLAGS_WORKAROUND_CAVIUM_22375 (1ULL << 1)
45 #define RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING (1 << 0)
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
52 struct its_collection {
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.
64 struct list_head entry;
66 unsigned long phys_base;
67 struct its_cmd_block *cmd_base;
68 struct its_cmd_block *cmd_write;
72 } tables[GITS_BASER_NR_REGS];
73 struct its_collection *collections;
74 struct list_head its_device_list;
79 #define ITS_ITT_ALIGN SZ_256
81 /* Convert page order to size in bytes */
82 #define PAGE_ORDER_TO_SIZE(o) (PAGE_SIZE << (o))
84 struct event_lpi_map {
85 unsigned long *lpi_map;
87 irq_hw_number_t lpi_base;
92 * The ITS view of a device - belongs to an ITS, a collection, owns an
93 * interrupt translation table, and a list of interrupts.
96 struct list_head entry;
98 struct event_lpi_map event_map;
104 static LIST_HEAD(its_nodes);
105 static DEFINE_SPINLOCK(its_lock);
106 static struct device_node *gic_root_node;
107 static struct rdists *gic_rdists;
109 #define gic_data_rdist() (raw_cpu_ptr(gic_rdists->rdist))
110 #define gic_data_rdist_rd_base() (gic_data_rdist()->rd_base)
112 static struct its_collection *dev_event_to_col(struct its_device *its_dev,
115 struct its_node *its = its_dev->its;
117 return its->collections + its_dev->event_map.col_map[event];
121 * ITS command descriptors - parameters to be encoded in a command
124 struct its_cmd_desc {
127 struct its_device *dev;
132 struct its_device *dev;
137 struct its_device *dev;
142 struct its_collection *col;
147 struct its_device *dev;
153 struct its_device *dev;
154 struct its_collection *col;
159 struct its_device *dev;
164 struct its_collection *col;
170 * The ITS command block, which is what the ITS actually parses.
172 struct its_cmd_block {
176 #define ITS_CMD_QUEUE_SZ SZ_64K
177 #define ITS_CMD_QUEUE_NR_ENTRIES (ITS_CMD_QUEUE_SZ / sizeof(struct its_cmd_block))
179 typedef struct its_collection *(*its_cmd_builder_t)(struct its_cmd_block *,
180 struct its_cmd_desc *);
182 static void its_encode_cmd(struct its_cmd_block *cmd, u8 cmd_nr)
184 cmd->raw_cmd[0] &= ~0xffUL;
185 cmd->raw_cmd[0] |= cmd_nr;
188 static void its_encode_devid(struct its_cmd_block *cmd, u32 devid)
190 cmd->raw_cmd[0] &= BIT_ULL(32) - 1;
191 cmd->raw_cmd[0] |= ((u64)devid) << 32;
194 static void its_encode_event_id(struct its_cmd_block *cmd, u32 id)
196 cmd->raw_cmd[1] &= ~0xffffffffUL;
197 cmd->raw_cmd[1] |= id;
200 static void its_encode_phys_id(struct its_cmd_block *cmd, u32 phys_id)
202 cmd->raw_cmd[1] &= 0xffffffffUL;
203 cmd->raw_cmd[1] |= ((u64)phys_id) << 32;
206 static void its_encode_size(struct its_cmd_block *cmd, u8 size)
208 cmd->raw_cmd[1] &= ~0x1fUL;
209 cmd->raw_cmd[1] |= size & 0x1f;
212 static void its_encode_itt(struct its_cmd_block *cmd, u64 itt_addr)
214 cmd->raw_cmd[2] &= ~0xffffffffffffUL;
215 cmd->raw_cmd[2] |= itt_addr & 0xffffffffff00UL;
218 static void its_encode_valid(struct its_cmd_block *cmd, int valid)
220 cmd->raw_cmd[2] &= ~(1UL << 63);
221 cmd->raw_cmd[2] |= ((u64)!!valid) << 63;
224 static void its_encode_target(struct its_cmd_block *cmd, u64 target_addr)
226 cmd->raw_cmd[2] &= ~(0xffffffffUL << 16);
227 cmd->raw_cmd[2] |= (target_addr & (0xffffffffUL << 16));
230 static void its_encode_collection(struct its_cmd_block *cmd, u16 col)
232 cmd->raw_cmd[2] &= ~0xffffUL;
233 cmd->raw_cmd[2] |= col;
236 static inline void its_fixup_cmd(struct its_cmd_block *cmd)
238 /* Let's fixup BE commands */
239 cmd->raw_cmd[0] = cpu_to_le64(cmd->raw_cmd[0]);
240 cmd->raw_cmd[1] = cpu_to_le64(cmd->raw_cmd[1]);
241 cmd->raw_cmd[2] = cpu_to_le64(cmd->raw_cmd[2]);
242 cmd->raw_cmd[3] = cpu_to_le64(cmd->raw_cmd[3]);
245 static struct its_collection *its_build_mapd_cmd(struct its_cmd_block *cmd,
246 struct its_cmd_desc *desc)
248 unsigned long itt_addr;
249 u8 size = ilog2(desc->its_mapd_cmd.dev->nr_ites);
251 itt_addr = virt_to_phys(desc->its_mapd_cmd.dev->itt);
252 itt_addr = ALIGN(itt_addr, ITS_ITT_ALIGN);
254 its_encode_cmd(cmd, GITS_CMD_MAPD);
255 its_encode_devid(cmd, desc->its_mapd_cmd.dev->device_id);
256 its_encode_size(cmd, size - 1);
257 its_encode_itt(cmd, itt_addr);
258 its_encode_valid(cmd, desc->its_mapd_cmd.valid);
265 static struct its_collection *its_build_mapc_cmd(struct its_cmd_block *cmd,
266 struct its_cmd_desc *desc)
268 its_encode_cmd(cmd, GITS_CMD_MAPC);
269 its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
270 its_encode_target(cmd, desc->its_mapc_cmd.col->target_address);
271 its_encode_valid(cmd, desc->its_mapc_cmd.valid);
275 return desc->its_mapc_cmd.col;
278 static struct its_collection *its_build_mapvi_cmd(struct its_cmd_block *cmd,
279 struct its_cmd_desc *desc)
281 struct its_collection *col;
283 col = dev_event_to_col(desc->its_mapvi_cmd.dev,
284 desc->its_mapvi_cmd.event_id);
286 its_encode_cmd(cmd, GITS_CMD_MAPVI);
287 its_encode_devid(cmd, desc->its_mapvi_cmd.dev->device_id);
288 its_encode_event_id(cmd, desc->its_mapvi_cmd.event_id);
289 its_encode_phys_id(cmd, desc->its_mapvi_cmd.phys_id);
290 its_encode_collection(cmd, col->col_id);
297 static struct its_collection *its_build_movi_cmd(struct its_cmd_block *cmd,
298 struct its_cmd_desc *desc)
300 struct its_collection *col;
302 col = dev_event_to_col(desc->its_movi_cmd.dev,
303 desc->its_movi_cmd.event_id);
305 its_encode_cmd(cmd, GITS_CMD_MOVI);
306 its_encode_devid(cmd, desc->its_movi_cmd.dev->device_id);
307 its_encode_event_id(cmd, desc->its_movi_cmd.event_id);
308 its_encode_collection(cmd, desc->its_movi_cmd.col->col_id);
315 static struct its_collection *its_build_discard_cmd(struct its_cmd_block *cmd,
316 struct its_cmd_desc *desc)
318 struct its_collection *col;
320 col = dev_event_to_col(desc->its_discard_cmd.dev,
321 desc->its_discard_cmd.event_id);
323 its_encode_cmd(cmd, GITS_CMD_DISCARD);
324 its_encode_devid(cmd, desc->its_discard_cmd.dev->device_id);
325 its_encode_event_id(cmd, desc->its_discard_cmd.event_id);
332 static struct its_collection *its_build_inv_cmd(struct its_cmd_block *cmd,
333 struct its_cmd_desc *desc)
335 struct its_collection *col;
337 col = dev_event_to_col(desc->its_inv_cmd.dev,
338 desc->its_inv_cmd.event_id);
340 its_encode_cmd(cmd, GITS_CMD_INV);
341 its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
342 its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
349 static struct its_collection *its_build_invall_cmd(struct its_cmd_block *cmd,
350 struct its_cmd_desc *desc)
352 its_encode_cmd(cmd, GITS_CMD_INVALL);
353 its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
360 static u64 its_cmd_ptr_to_offset(struct its_node *its,
361 struct its_cmd_block *ptr)
363 return (ptr - its->cmd_base) * sizeof(*ptr);
366 static int its_queue_full(struct its_node *its)
371 widx = its->cmd_write - its->cmd_base;
372 ridx = readl_relaxed(its->base + GITS_CREADR) / sizeof(struct its_cmd_block);
374 /* This is incredibly unlikely to happen, unless the ITS locks up. */
375 if (((widx + 1) % ITS_CMD_QUEUE_NR_ENTRIES) == ridx)
381 static struct its_cmd_block *its_allocate_entry(struct its_node *its)
383 struct its_cmd_block *cmd;
384 u32 count = 1000000; /* 1s! */
386 while (its_queue_full(its)) {
389 pr_err_ratelimited("ITS queue not draining\n");
396 cmd = its->cmd_write++;
398 /* Handle queue wrapping */
399 if (its->cmd_write == (its->cmd_base + ITS_CMD_QUEUE_NR_ENTRIES))
400 its->cmd_write = its->cmd_base;
405 static struct its_cmd_block *its_post_commands(struct its_node *its)
407 u64 wr = its_cmd_ptr_to_offset(its, its->cmd_write);
409 writel_relaxed(wr, its->base + GITS_CWRITER);
411 return its->cmd_write;
414 static void its_flush_cmd(struct its_node *its, struct its_cmd_block *cmd)
417 * Make sure the commands written to memory are observable by
420 if (its->flags & ITS_FLAGS_CMDQ_NEEDS_FLUSHING)
421 __flush_dcache_area(cmd, sizeof(*cmd));
426 static void its_wait_for_range_completion(struct its_node *its,
427 struct its_cmd_block *from,
428 struct its_cmd_block *to)
430 u64 rd_idx, from_idx, to_idx;
431 u32 count = 1000000; /* 1s! */
433 from_idx = its_cmd_ptr_to_offset(its, from);
434 to_idx = its_cmd_ptr_to_offset(its, to);
437 rd_idx = readl_relaxed(its->base + GITS_CREADR);
438 if (rd_idx >= to_idx || rd_idx < from_idx)
443 pr_err_ratelimited("ITS queue timeout\n");
451 static void its_send_single_command(struct its_node *its,
452 its_cmd_builder_t builder,
453 struct its_cmd_desc *desc)
455 struct its_cmd_block *cmd, *sync_cmd, *next_cmd;
456 struct its_collection *sync_col;
459 raw_spin_lock_irqsave(&its->lock, flags);
461 cmd = its_allocate_entry(its);
462 if (!cmd) { /* We're soooooo screewed... */
463 pr_err_ratelimited("ITS can't allocate, dropping command\n");
464 raw_spin_unlock_irqrestore(&its->lock, flags);
467 sync_col = builder(cmd, desc);
468 its_flush_cmd(its, cmd);
471 sync_cmd = its_allocate_entry(its);
473 pr_err_ratelimited("ITS can't SYNC, skipping\n");
476 its_encode_cmd(sync_cmd, GITS_CMD_SYNC);
477 its_encode_target(sync_cmd, sync_col->target_address);
478 its_fixup_cmd(sync_cmd);
479 its_flush_cmd(its, sync_cmd);
483 next_cmd = its_post_commands(its);
484 raw_spin_unlock_irqrestore(&its->lock, flags);
486 its_wait_for_range_completion(its, cmd, next_cmd);
489 static void its_send_inv(struct its_device *dev, u32 event_id)
491 struct its_cmd_desc desc;
493 desc.its_inv_cmd.dev = dev;
494 desc.its_inv_cmd.event_id = event_id;
496 its_send_single_command(dev->its, its_build_inv_cmd, &desc);
499 static void its_send_mapd(struct its_device *dev, int valid)
501 struct its_cmd_desc desc;
503 desc.its_mapd_cmd.dev = dev;
504 desc.its_mapd_cmd.valid = !!valid;
506 its_send_single_command(dev->its, its_build_mapd_cmd, &desc);
509 static void its_send_mapc(struct its_node *its, struct its_collection *col,
512 struct its_cmd_desc desc;
514 desc.its_mapc_cmd.col = col;
515 desc.its_mapc_cmd.valid = !!valid;
517 its_send_single_command(its, its_build_mapc_cmd, &desc);
520 static void its_send_mapvi(struct its_device *dev, u32 irq_id, u32 id)
522 struct its_cmd_desc desc;
524 desc.its_mapvi_cmd.dev = dev;
525 desc.its_mapvi_cmd.phys_id = irq_id;
526 desc.its_mapvi_cmd.event_id = id;
528 its_send_single_command(dev->its, its_build_mapvi_cmd, &desc);
531 static void its_send_movi(struct its_device *dev,
532 struct its_collection *col, u32 id)
534 struct its_cmd_desc desc;
536 desc.its_movi_cmd.dev = dev;
537 desc.its_movi_cmd.col = col;
538 desc.its_movi_cmd.event_id = id;
540 its_send_single_command(dev->its, its_build_movi_cmd, &desc);
543 static void its_send_discard(struct its_device *dev, u32 id)
545 struct its_cmd_desc desc;
547 desc.its_discard_cmd.dev = dev;
548 desc.its_discard_cmd.event_id = id;
550 its_send_single_command(dev->its, its_build_discard_cmd, &desc);
553 static void its_send_invall(struct its_node *its, struct its_collection *col)
555 struct its_cmd_desc desc;
557 desc.its_invall_cmd.col = col;
559 its_send_single_command(its, its_build_invall_cmd, &desc);
563 * irqchip functions - assumes MSI, mostly.
566 static inline u32 its_get_event_id(struct irq_data *d)
568 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
569 return d->hwirq - its_dev->event_map.lpi_base;
572 static void lpi_set_config(struct irq_data *d, bool enable)
574 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
575 irq_hw_number_t hwirq = d->hwirq;
576 u32 id = its_get_event_id(d);
577 u8 *cfg = page_address(gic_rdists->prop_page) + hwirq - 8192;
580 *cfg |= LPI_PROP_ENABLED;
582 *cfg &= ~LPI_PROP_ENABLED;
585 * Make the above write visible to the redistributors.
586 * And yes, we're flushing exactly: One. Single. Byte.
589 if (gic_rdists->flags & RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING)
590 __flush_dcache_area(cfg, sizeof(*cfg));
593 its_send_inv(its_dev, id);
596 static void its_mask_irq(struct irq_data *d)
598 lpi_set_config(d, false);
601 static void its_unmask_irq(struct irq_data *d)
603 lpi_set_config(d, true);
606 static int its_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
609 unsigned int cpu = cpumask_any_and(mask_val, cpu_online_mask);
610 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
611 struct its_collection *target_col;
612 u32 id = its_get_event_id(d);
614 if (cpu >= nr_cpu_ids)
617 target_col = &its_dev->its->collections[cpu];
618 its_send_movi(its_dev, target_col, id);
619 its_dev->event_map.col_map[id] = cpu;
621 return IRQ_SET_MASK_OK_DONE;
624 static void its_irq_compose_msi_msg(struct irq_data *d, struct msi_msg *msg)
626 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
627 struct its_node *its;
631 addr = its->phys_base + GITS_TRANSLATER;
633 msg->address_lo = addr & ((1UL << 32) - 1);
634 msg->address_hi = addr >> 32;
635 msg->data = its_get_event_id(d);
638 static struct irq_chip its_irq_chip = {
640 .irq_mask = its_mask_irq,
641 .irq_unmask = its_unmask_irq,
642 .irq_eoi = irq_chip_eoi_parent,
643 .irq_set_affinity = its_set_affinity,
644 .irq_compose_msi_msg = its_irq_compose_msi_msg,
648 * How we allocate LPIs:
650 * The GIC has id_bits bits for interrupt identifiers. From there, we
651 * must subtract 8192 which are reserved for SGIs/PPIs/SPIs. Then, as
652 * we allocate LPIs by chunks of 32, we can shift the whole thing by 5
655 * This gives us (((1UL << id_bits) - 8192) >> 5) possible allocations.
657 #define IRQS_PER_CHUNK_SHIFT 5
658 #define IRQS_PER_CHUNK (1 << IRQS_PER_CHUNK_SHIFT)
660 static unsigned long *lpi_bitmap;
661 static u32 lpi_chunks;
662 static DEFINE_SPINLOCK(lpi_lock);
664 static int its_lpi_to_chunk(int lpi)
666 return (lpi - 8192) >> IRQS_PER_CHUNK_SHIFT;
669 static int its_chunk_to_lpi(int chunk)
671 return (chunk << IRQS_PER_CHUNK_SHIFT) + 8192;
674 static int its_lpi_init(u32 id_bits)
676 lpi_chunks = its_lpi_to_chunk(1UL << id_bits);
678 lpi_bitmap = kzalloc(BITS_TO_LONGS(lpi_chunks) * sizeof(long),
685 pr_info("ITS: Allocated %d chunks for LPIs\n", (int)lpi_chunks);
689 static unsigned long *its_lpi_alloc_chunks(int nr_irqs, int *base, int *nr_ids)
691 unsigned long *bitmap = NULL;
696 nr_chunks = DIV_ROUND_UP(nr_irqs, IRQS_PER_CHUNK);
698 spin_lock(&lpi_lock);
701 chunk_id = bitmap_find_next_zero_area(lpi_bitmap, lpi_chunks,
703 if (chunk_id < lpi_chunks)
707 } while (nr_chunks > 0);
712 bitmap = kzalloc(BITS_TO_LONGS(nr_chunks * IRQS_PER_CHUNK) * sizeof (long),
717 for (i = 0; i < nr_chunks; i++)
718 set_bit(chunk_id + i, lpi_bitmap);
720 *base = its_chunk_to_lpi(chunk_id);
721 *nr_ids = nr_chunks * IRQS_PER_CHUNK;
724 spin_unlock(&lpi_lock);
732 static void its_lpi_free(struct event_lpi_map *map)
734 int base = map->lpi_base;
735 int nr_ids = map->nr_lpis;
738 spin_lock(&lpi_lock);
740 for (lpi = base; lpi < (base + nr_ids); lpi += IRQS_PER_CHUNK) {
741 int chunk = its_lpi_to_chunk(lpi);
742 BUG_ON(chunk > lpi_chunks);
743 if (test_bit(chunk, lpi_bitmap)) {
744 clear_bit(chunk, lpi_bitmap);
746 pr_err("Bad LPI chunk %d\n", chunk);
750 spin_unlock(&lpi_lock);
757 * We allocate 64kB for PROPBASE. That gives us at most 64K LPIs to
758 * deal with (one configuration byte per interrupt). PENDBASE has to
759 * be 64kB aligned (one bit per LPI, plus 8192 bits for SPI/PPI/SGI).
761 #define LPI_PROPBASE_SZ SZ_64K
762 #define LPI_PENDBASE_SZ (LPI_PROPBASE_SZ / 8 + SZ_1K)
765 * This is how many bits of ID we need, including the useless ones.
767 #define LPI_NRBITS ilog2(LPI_PROPBASE_SZ + SZ_8K)
769 #define LPI_PROP_DEFAULT_PRIO 0xa0
771 static int __init its_alloc_lpi_tables(void)
775 gic_rdists->prop_page = alloc_pages(GFP_NOWAIT,
776 get_order(LPI_PROPBASE_SZ));
777 if (!gic_rdists->prop_page) {
778 pr_err("Failed to allocate PROPBASE\n");
782 paddr = page_to_phys(gic_rdists->prop_page);
783 pr_info("GIC: using LPI property table @%pa\n", &paddr);
785 /* Priority 0xa0, Group-1, disabled */
786 memset(page_address(gic_rdists->prop_page),
787 LPI_PROP_DEFAULT_PRIO | LPI_PROP_GROUP1,
790 /* Make sure the GIC will observe the written configuration */
791 __flush_dcache_area(page_address(gic_rdists->prop_page), LPI_PROPBASE_SZ);
796 static const char *its_base_type_string[] = {
797 [GITS_BASER_TYPE_DEVICE] = "Devices",
798 [GITS_BASER_TYPE_VCPU] = "Virtual CPUs",
799 [GITS_BASER_TYPE_CPU] = "Physical CPUs",
800 [GITS_BASER_TYPE_COLLECTION] = "Interrupt Collections",
801 [GITS_BASER_TYPE_RESERVED5] = "Reserved (5)",
802 [GITS_BASER_TYPE_RESERVED6] = "Reserved (6)",
803 [GITS_BASER_TYPE_RESERVED7] = "Reserved (7)",
806 static void its_free_tables(struct its_node *its)
810 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
811 if (its->tables[i].base) {
812 free_pages((unsigned long)its->tables[i].base,
813 its->tables[i].order);
814 its->tables[i].base = NULL;
819 static int its_alloc_tables(const char *node_name, struct its_node *its)
824 u64 shr = GITS_BASER_InnerShareable;
829 if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_22375) {
831 * erratum 22375: only alloc 8MB table size
832 * erratum 24313: ignore memory access type
835 ids = 0x14; /* 20 bits, 8MB */
837 cache = GITS_BASER_WaWb;
838 typer = readq_relaxed(its->base + GITS_TYPER);
839 ids = GITS_TYPER_DEVBITS(typer);
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);
851 if (type == GITS_BASER_TYPE_NONE)
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.
860 * For other tables, only allocate a single page.
862 if (type == GITS_BASER_TYPE_DEVICE) {
864 * 'order' was initialized earlier to the default page
865 * granule of the the ITS. We can't have an allocation
866 * smaller than that. If the requested allocation
867 * is smaller, round up to the default page granule.
869 order = max(get_order((1UL << ids) * entry_size),
871 if (order >= MAX_ORDER) {
872 order = MAX_ORDER - 1;
873 pr_warn("%s: Device Table too large, reduce its page order to %u\n",
879 alloc_pages = (PAGE_ORDER_TO_SIZE(order) / psz);
880 if (alloc_pages > GITS_BASER_PAGES_MAX) {
881 alloc_pages = GITS_BASER_PAGES_MAX;
882 order = get_order(GITS_BASER_PAGES_MAX * psz);
883 pr_warn("%s: Device Table too large, reduce its page order to %u (%u pages)\n",
884 node_name, order, alloc_pages);
887 base = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, order);
893 its->tables[i].base = base;
894 its->tables[i].order = order;
897 val = (virt_to_phys(base) |
898 (type << GITS_BASER_TYPE_SHIFT) |
899 ((entry_size - 1) << GITS_BASER_ENTRY_SIZE_SHIFT) |
906 val |= GITS_BASER_PAGE_SIZE_4K;
909 val |= GITS_BASER_PAGE_SIZE_16K;
912 val |= GITS_BASER_PAGE_SIZE_64K;
916 val |= alloc_pages - 1;
918 writeq_relaxed(val, its->base + GITS_BASER + i * 8);
919 tmp = readq_relaxed(its->base + GITS_BASER + i * 8);
921 if ((val ^ tmp) & GITS_BASER_SHAREABILITY_MASK) {
923 * Shareability didn't stick. Just use
924 * whatever the read reported, which is likely
925 * to be the only thing this redistributor
926 * supports. If that's zero, make it
927 * non-cacheable as well.
929 shr = tmp & GITS_BASER_SHAREABILITY_MASK;
931 cache = GITS_BASER_nC;
932 __flush_dcache_area(base, PAGE_ORDER_TO_SIZE(order));
937 if ((val ^ tmp) & GITS_BASER_PAGE_SIZE_MASK) {
939 * Page size didn't stick. Let's try a smaller
940 * size and retry. If we reach 4K, then
941 * something is horribly wrong...
943 free_pages((unsigned long)base, order);
944 its->tables[i].base = NULL;
949 goto retry_alloc_baser;
952 goto retry_alloc_baser;
957 pr_err("ITS: %s: GITS_BASER%d doesn't stick: %lx %lx\n",
959 (unsigned long) val, (unsigned long) tmp);
964 pr_info("ITS: allocated %d %s @%lx (psz %dK, shr %d)\n",
965 (int)(PAGE_ORDER_TO_SIZE(order) / entry_size),
966 its_base_type_string[type],
967 (unsigned long)virt_to_phys(base),
968 psz / SZ_1K, (int)shr >> GITS_BASER_SHAREABILITY_SHIFT);
974 its_free_tables(its);
979 static int its_alloc_collections(struct its_node *its)
981 its->collections = kzalloc(nr_cpu_ids * sizeof(*its->collections),
983 if (!its->collections)
989 static void its_cpu_init_lpis(void)
991 void __iomem *rbase = gic_data_rdist_rd_base();
992 struct page *pend_page;
995 /* If we didn't allocate the pending table yet, do it now */
996 pend_page = gic_data_rdist()->pend_page;
1000 * The pending pages have to be at least 64kB aligned,
1001 * hence the 'max(LPI_PENDBASE_SZ, SZ_64K)' below.
1003 pend_page = alloc_pages(GFP_NOWAIT | __GFP_ZERO,
1004 get_order(max(LPI_PENDBASE_SZ, SZ_64K)));
1006 pr_err("Failed to allocate PENDBASE for CPU%d\n",
1007 smp_processor_id());
1011 /* Make sure the GIC will observe the zero-ed page */
1012 __flush_dcache_area(page_address(pend_page), LPI_PENDBASE_SZ);
1014 paddr = page_to_phys(pend_page);
1015 pr_info("CPU%d: using LPI pending table @%pa\n",
1016 smp_processor_id(), &paddr);
1017 gic_data_rdist()->pend_page = pend_page;
1021 val = readl_relaxed(rbase + GICR_CTLR);
1022 val &= ~GICR_CTLR_ENABLE_LPIS;
1023 writel_relaxed(val, rbase + GICR_CTLR);
1026 * Make sure any change to the table is observable by the GIC.
1031 val = (page_to_phys(gic_rdists->prop_page) |
1032 GICR_PROPBASER_InnerShareable |
1033 GICR_PROPBASER_WaWb |
1034 ((LPI_NRBITS - 1) & GICR_PROPBASER_IDBITS_MASK));
1036 writeq_relaxed(val, rbase + GICR_PROPBASER);
1037 tmp = readq_relaxed(rbase + GICR_PROPBASER);
1039 if ((tmp ^ val) & GICR_PROPBASER_SHAREABILITY_MASK) {
1040 if (!(tmp & GICR_PROPBASER_SHAREABILITY_MASK)) {
1042 * The HW reports non-shareable, we must
1043 * remove the cacheability attributes as
1046 val &= ~(GICR_PROPBASER_SHAREABILITY_MASK |
1047 GICR_PROPBASER_CACHEABILITY_MASK);
1048 val |= GICR_PROPBASER_nC;
1049 writeq_relaxed(val, rbase + GICR_PROPBASER);
1051 pr_info_once("GIC: using cache flushing for LPI property table\n");
1052 gic_rdists->flags |= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING;
1056 val = (page_to_phys(pend_page) |
1057 GICR_PENDBASER_InnerShareable |
1058 GICR_PENDBASER_WaWb);
1060 writeq_relaxed(val, rbase + GICR_PENDBASER);
1061 tmp = readq_relaxed(rbase + GICR_PENDBASER);
1063 if (!(tmp & GICR_PENDBASER_SHAREABILITY_MASK)) {
1065 * The HW reports non-shareable, we must remove the
1066 * cacheability attributes as well.
1068 val &= ~(GICR_PENDBASER_SHAREABILITY_MASK |
1069 GICR_PENDBASER_CACHEABILITY_MASK);
1070 val |= GICR_PENDBASER_nC;
1071 writeq_relaxed(val, rbase + GICR_PENDBASER);
1075 val = readl_relaxed(rbase + GICR_CTLR);
1076 val |= GICR_CTLR_ENABLE_LPIS;
1077 writel_relaxed(val, rbase + GICR_CTLR);
1079 /* Make sure the GIC has seen the above */
1083 static void its_cpu_init_collection(void)
1085 struct its_node *its;
1088 spin_lock(&its_lock);
1089 cpu = smp_processor_id();
1091 list_for_each_entry(its, &its_nodes, entry) {
1095 * We now have to bind each collection to its target
1098 if (readq_relaxed(its->base + GITS_TYPER) & GITS_TYPER_PTA) {
1100 * This ITS wants the physical address of the
1103 target = gic_data_rdist()->phys_base;
1106 * This ITS wants a linear CPU number.
1108 target = readq_relaxed(gic_data_rdist_rd_base() + GICR_TYPER);
1109 target = GICR_TYPER_CPU_NUMBER(target) << 16;
1112 /* Perform collection mapping */
1113 its->collections[cpu].target_address = target;
1114 its->collections[cpu].col_id = cpu;
1116 its_send_mapc(its, &its->collections[cpu], 1);
1117 its_send_invall(its, &its->collections[cpu]);
1120 spin_unlock(&its_lock);
1123 static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
1125 struct its_device *its_dev = NULL, *tmp;
1126 unsigned long flags;
1128 raw_spin_lock_irqsave(&its->lock, flags);
1130 list_for_each_entry(tmp, &its->its_device_list, entry) {
1131 if (tmp->device_id == dev_id) {
1137 raw_spin_unlock_irqrestore(&its->lock, flags);
1142 static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
1145 struct its_device *dev;
1146 unsigned long *lpi_map;
1147 unsigned long flags;
1148 u16 *col_map = NULL;
1155 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1157 * At least one bit of EventID is being used, hence a minimum
1158 * of two entries. No, the architecture doesn't let you
1159 * express an ITT with a single entry.
1161 nr_ites = max(2UL, roundup_pow_of_two(nvecs));
1162 sz = nr_ites * its->ite_size;
1163 sz = max(sz, ITS_ITT_ALIGN) + ITS_ITT_ALIGN - 1;
1164 itt = kzalloc(sz, GFP_KERNEL);
1165 lpi_map = its_lpi_alloc_chunks(nvecs, &lpi_base, &nr_lpis);
1167 col_map = kzalloc(sizeof(*col_map) * nr_lpis, GFP_KERNEL);
1169 if (!dev || !itt || !lpi_map || !col_map) {
1177 __flush_dcache_area(itt, sz);
1181 dev->nr_ites = nr_ites;
1182 dev->event_map.lpi_map = lpi_map;
1183 dev->event_map.col_map = col_map;
1184 dev->event_map.lpi_base = lpi_base;
1185 dev->event_map.nr_lpis = nr_lpis;
1186 dev->device_id = dev_id;
1187 INIT_LIST_HEAD(&dev->entry);
1189 raw_spin_lock_irqsave(&its->lock, flags);
1190 list_add(&dev->entry, &its->its_device_list);
1191 raw_spin_unlock_irqrestore(&its->lock, flags);
1193 /* Map device to its ITT */
1194 its_send_mapd(dev, 1);
1199 static void its_free_device(struct its_device *its_dev)
1201 unsigned long flags;
1203 raw_spin_lock_irqsave(&its_dev->its->lock, flags);
1204 list_del(&its_dev->entry);
1205 raw_spin_unlock_irqrestore(&its_dev->its->lock, flags);
1206 kfree(its_dev->itt);
1210 static int its_alloc_device_irq(struct its_device *dev, irq_hw_number_t *hwirq)
1214 idx = find_first_zero_bit(dev->event_map.lpi_map,
1215 dev->event_map.nr_lpis);
1216 if (idx == dev->event_map.nr_lpis)
1219 *hwirq = dev->event_map.lpi_base + idx;
1220 set_bit(idx, dev->event_map.lpi_map);
1225 static int its_msi_prepare(struct irq_domain *domain, struct device *dev,
1226 int nvec, msi_alloc_info_t *info)
1228 struct its_node *its;
1229 struct its_device *its_dev;
1230 struct msi_domain_info *msi_info;
1234 * We ignore "dev" entierely, and rely on the dev_id that has
1235 * been passed via the scratchpad. This limits this domain's
1236 * usefulness to upper layers that definitely know that they
1237 * are built on top of the ITS.
1239 dev_id = info->scratchpad[0].ul;
1241 msi_info = msi_get_domain_info(domain);
1242 its = msi_info->data;
1244 its_dev = its_find_device(its, dev_id);
1247 * We already have seen this ID, probably through
1248 * another alias (PCI bridge of some sort). No need to
1249 * create the device.
1251 pr_debug("Reusing ITT for devID %x\n", dev_id);
1255 its_dev = its_create_device(its, dev_id, nvec);
1259 pr_debug("ITT %d entries, %d bits\n", nvec, ilog2(nvec));
1261 info->scratchpad[0].ptr = its_dev;
1265 static struct msi_domain_ops its_msi_domain_ops = {
1266 .msi_prepare = its_msi_prepare,
1269 static int its_irq_gic_domain_alloc(struct irq_domain *domain,
1271 irq_hw_number_t hwirq)
1273 struct irq_fwspec fwspec;
1275 if (irq_domain_get_of_node(domain->parent)) {
1276 fwspec.fwnode = domain->parent->fwnode;
1277 fwspec.param_count = 3;
1278 fwspec.param[0] = GIC_IRQ_TYPE_LPI;
1279 fwspec.param[1] = hwirq;
1280 fwspec.param[2] = IRQ_TYPE_EDGE_RISING;
1285 return irq_domain_alloc_irqs_parent(domain, virq, 1, &fwspec);
1288 static int its_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
1289 unsigned int nr_irqs, void *args)
1291 msi_alloc_info_t *info = args;
1292 struct its_device *its_dev = info->scratchpad[0].ptr;
1293 irq_hw_number_t hwirq;
1297 for (i = 0; i < nr_irqs; i++) {
1298 err = its_alloc_device_irq(its_dev, &hwirq);
1302 err = its_irq_gic_domain_alloc(domain, virq + i, hwirq);
1306 irq_domain_set_hwirq_and_chip(domain, virq + i,
1307 hwirq, &its_irq_chip, its_dev);
1308 pr_debug("ID:%d pID:%d vID:%d\n",
1309 (int)(hwirq - its_dev->event_map.lpi_base),
1310 (int) hwirq, virq + i);
1316 static void its_irq_domain_activate(struct irq_domain *domain,
1319 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1320 u32 event = its_get_event_id(d);
1322 /* Bind the LPI to the first possible CPU */
1323 its_dev->event_map.col_map[event] = cpumask_first(cpu_online_mask);
1325 /* Map the GIC IRQ and event to the device */
1326 its_send_mapvi(its_dev, d->hwirq, event);
1329 static void its_irq_domain_deactivate(struct irq_domain *domain,
1332 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1333 u32 event = its_get_event_id(d);
1335 /* Stop the delivery of interrupts */
1336 its_send_discard(its_dev, event);
1339 static void its_irq_domain_free(struct irq_domain *domain, unsigned int virq,
1340 unsigned int nr_irqs)
1342 struct irq_data *d = irq_domain_get_irq_data(domain, virq);
1343 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1346 for (i = 0; i < nr_irqs; i++) {
1347 struct irq_data *data = irq_domain_get_irq_data(domain,
1349 u32 event = its_get_event_id(data);
1351 /* Mark interrupt index as unused */
1352 clear_bit(event, its_dev->event_map.lpi_map);
1354 /* Nuke the entry in the domain */
1355 irq_domain_reset_irq_data(data);
1358 /* If all interrupts have been freed, start mopping the floor */
1359 if (bitmap_empty(its_dev->event_map.lpi_map,
1360 its_dev->event_map.nr_lpis)) {
1361 its_lpi_free(&its_dev->event_map);
1363 /* Unmap device/itt */
1364 its_send_mapd(its_dev, 0);
1365 its_free_device(its_dev);
1368 irq_domain_free_irqs_parent(domain, virq, nr_irqs);
1371 static const struct irq_domain_ops its_domain_ops = {
1372 .alloc = its_irq_domain_alloc,
1373 .free = its_irq_domain_free,
1374 .activate = its_irq_domain_activate,
1375 .deactivate = its_irq_domain_deactivate,
1378 static int its_force_quiescent(void __iomem *base)
1380 u32 count = 1000000; /* 1s */
1383 val = readl_relaxed(base + GITS_CTLR);
1384 if (val & GITS_CTLR_QUIESCENT)
1387 /* Disable the generation of all interrupts to this ITS */
1388 val &= ~GITS_CTLR_ENABLE;
1389 writel_relaxed(val, base + GITS_CTLR);
1391 /* Poll GITS_CTLR and wait until ITS becomes quiescent */
1393 val = readl_relaxed(base + GITS_CTLR);
1394 if (val & GITS_CTLR_QUIESCENT)
1406 static void __maybe_unused its_enable_quirk_cavium_22375(void *data)
1408 struct its_node *its = data;
1410 its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_22375;
1413 static const struct gic_quirk its_quirks[] = {
1414 #ifdef CONFIG_CAVIUM_ERRATUM_22375
1416 .desc = "ITS: Cavium errata 22375, 24313",
1417 .iidr = 0xa100034c, /* ThunderX pass 1.x */
1419 .init = its_enable_quirk_cavium_22375,
1426 static void its_enable_quirks(struct its_node *its)
1428 u32 iidr = readl_relaxed(its->base + GITS_IIDR);
1430 gic_enable_quirks(iidr, its_quirks, its);
1433 static int its_probe(struct device_node *node, struct irq_domain *parent)
1435 struct resource res;
1436 struct its_node *its;
1437 void __iomem *its_base;
1438 struct irq_domain *inner_domain;
1443 err = of_address_to_resource(node, 0, &res);
1445 pr_warn("%s: no regs?\n", node->full_name);
1449 its_base = ioremap(res.start, resource_size(&res));
1451 pr_warn("%s: unable to map registers\n", node->full_name);
1455 val = readl_relaxed(its_base + GITS_PIDR2) & GIC_PIDR2_ARCH_MASK;
1456 if (val != 0x30 && val != 0x40) {
1457 pr_warn("%s: no ITS detected, giving up\n", node->full_name);
1462 err = its_force_quiescent(its_base);
1464 pr_warn("%s: failed to quiesce, giving up\n",
1469 pr_info("ITS: %s\n", node->full_name);
1471 its = kzalloc(sizeof(*its), GFP_KERNEL);
1477 raw_spin_lock_init(&its->lock);
1478 INIT_LIST_HEAD(&its->entry);
1479 INIT_LIST_HEAD(&its->its_device_list);
1480 its->base = its_base;
1481 its->phys_base = res.start;
1482 its->ite_size = ((readl_relaxed(its_base + GITS_TYPER) >> 4) & 0xf) + 1;
1484 its->cmd_base = kzalloc(ITS_CMD_QUEUE_SZ, GFP_KERNEL);
1485 if (!its->cmd_base) {
1489 its->cmd_write = its->cmd_base;
1491 its_enable_quirks(its);
1493 err = its_alloc_tables(node->full_name, its);
1497 err = its_alloc_collections(its);
1499 goto out_free_tables;
1501 baser = (virt_to_phys(its->cmd_base) |
1503 GITS_CBASER_InnerShareable |
1504 (ITS_CMD_QUEUE_SZ / SZ_4K - 1) |
1507 writeq_relaxed(baser, its->base + GITS_CBASER);
1508 tmp = readq_relaxed(its->base + GITS_CBASER);
1510 if ((tmp ^ baser) & GITS_CBASER_SHAREABILITY_MASK) {
1511 if (!(tmp & GITS_CBASER_SHAREABILITY_MASK)) {
1513 * The HW reports non-shareable, we must
1514 * remove the cacheability attributes as
1517 baser &= ~(GITS_CBASER_SHAREABILITY_MASK |
1518 GITS_CBASER_CACHEABILITY_MASK);
1519 baser |= GITS_CBASER_nC;
1520 writeq_relaxed(baser, its->base + GITS_CBASER);
1522 pr_info("ITS: using cache flushing for cmd queue\n");
1523 its->flags |= ITS_FLAGS_CMDQ_NEEDS_FLUSHING;
1526 writeq_relaxed(0, its->base + GITS_CWRITER);
1527 writel_relaxed(GITS_CTLR_ENABLE, its->base + GITS_CTLR);
1529 if (of_property_read_bool(node, "msi-controller")) {
1530 struct msi_domain_info *info;
1532 info = kzalloc(sizeof(*info), GFP_KERNEL);
1535 goto out_free_tables;
1538 inner_domain = irq_domain_add_tree(node, &its_domain_ops, its);
1539 if (!inner_domain) {
1542 goto out_free_tables;
1545 inner_domain->parent = parent;
1546 inner_domain->bus_token = DOMAIN_BUS_NEXUS;
1547 info->ops = &its_msi_domain_ops;
1549 inner_domain->host_data = info;
1552 spin_lock(&its_lock);
1553 list_add(&its->entry, &its_nodes);
1554 spin_unlock(&its_lock);
1559 its_free_tables(its);
1561 kfree(its->cmd_base);
1566 pr_err("ITS: failed probing %s (%d)\n", node->full_name, err);
1570 static bool gic_rdists_supports_plpis(void)
1572 return !!(readl_relaxed(gic_data_rdist_rd_base() + GICR_TYPER) & GICR_TYPER_PLPIS);
1575 int its_cpu_init(void)
1577 if (!list_empty(&its_nodes)) {
1578 if (!gic_rdists_supports_plpis()) {
1579 pr_info("CPU%d: LPIs not supported\n", smp_processor_id());
1582 its_cpu_init_lpis();
1583 its_cpu_init_collection();
1589 static struct of_device_id its_device_id[] = {
1590 { .compatible = "arm,gic-v3-its", },
1594 int its_init(struct device_node *node, struct rdists *rdists,
1595 struct irq_domain *parent_domain)
1597 struct device_node *np;
1599 for (np = of_find_matching_node(node, its_device_id); np;
1600 np = of_find_matching_node(np, its_device_id)) {
1601 its_probe(np, parent_domain);
1604 if (list_empty(&its_nodes)) {
1605 pr_warn("ITS: No ITS available, not enabling LPIs\n");
1609 gic_rdists = rdists;
1610 gic_root_node = node;
1612 its_alloc_lpi_tables();
1613 its_lpi_init(rdists->id_bits);