1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2013-2017 ARM Limited, All Rights Reserved.
4 * Author: Marc Zyngier <marc.zyngier@arm.com>
7 #include <linux/acpi.h>
8 #include <linux/acpi_iort.h>
9 #include <linux/bitfield.h>
10 #include <linux/bitmap.h>
11 #include <linux/cpu.h>
12 #include <linux/crash_dump.h>
13 #include <linux/delay.h>
14 #include <linux/dma-iommu.h>
15 #include <linux/efi.h>
16 #include <linux/interrupt.h>
17 #include <linux/iopoll.h>
18 #include <linux/irqdomain.h>
19 #include <linux/list.h>
20 #include <linux/log2.h>
21 #include <linux/memblock.h>
23 #include <linux/msi.h>
25 #include <linux/of_address.h>
26 #include <linux/of_irq.h>
27 #include <linux/of_pci.h>
28 #include <linux/of_platform.h>
29 #include <linux/percpu.h>
30 #include <linux/slab.h>
31 #include <linux/syscore_ops.h>
33 #include <linux/irqchip.h>
34 #include <linux/irqchip/arm-gic-v3.h>
35 #include <linux/irqchip/arm-gic-v4.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)
44 #define ITS_FLAGS_WORKAROUND_CAVIUM_23144 (1ULL << 2)
46 #define RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING (1 << 0)
47 #define RDIST_FLAGS_RD_TABLES_PREALLOCATED (1 << 1)
49 static u32 lpi_id_bits;
52 * We allocate memory for PROPBASE to cover 2 ^ lpi_id_bits LPIs to
53 * deal with (one configuration byte per interrupt). PENDBASE has to
54 * be 64kB aligned (one bit per LPI, plus 8192 bits for SPI/PPI/SGI).
56 #define LPI_NRBITS lpi_id_bits
57 #define LPI_PROPBASE_SZ ALIGN(BIT(LPI_NRBITS), SZ_64K)
58 #define LPI_PENDBASE_SZ ALIGN(BIT(LPI_NRBITS) / 8, SZ_64K)
60 #define LPI_PROP_DEFAULT_PRIO GICD_INT_DEF_PRI
63 * Collection structure - just an ID, and a redistributor address to
64 * ping. We use one per CPU as a bag of interrupts assigned to this
67 struct its_collection {
73 * The ITS_BASER structure - contains memory information, cached
74 * value of BASER register configuration and ITS page size.
86 * The ITS structure - contains most of the infrastructure, with the
87 * top-level MSI domain, the command queue, the collections, and the
88 * list of devices writing to it.
90 * dev_alloc_lock has to be taken for device allocations, while the
91 * spinlock must be taken to parse data structures such as the device
96 struct mutex dev_alloc_lock;
97 struct list_head entry;
99 void __iomem *sgir_base;
100 phys_addr_t phys_base;
101 struct its_cmd_block *cmd_base;
102 struct its_cmd_block *cmd_write;
103 struct its_baser tables[GITS_BASER_NR_REGS];
104 struct its_collection *collections;
105 struct fwnode_handle *fwnode_handle;
106 u64 (*get_msi_base)(struct its_device *its_dev);
111 struct list_head its_device_list;
113 unsigned long list_nr;
115 unsigned int msi_domain_flags;
116 u32 pre_its_base; /* for Socionext Synquacer */
117 int vlpi_redist_offset;
120 #define is_v4(its) (!!((its)->typer & GITS_TYPER_VLPIS))
121 #define is_v4_1(its) (!!((its)->typer & GITS_TYPER_VMAPP))
122 #define device_ids(its) (FIELD_GET(GITS_TYPER_DEVBITS, (its)->typer) + 1)
124 #define ITS_ITT_ALIGN SZ_256
126 /* The maximum number of VPEID bits supported by VLPI commands */
127 #define ITS_MAX_VPEID_BITS \
130 if (gic_rdists->has_rvpeid && \
131 gic_rdists->gicd_typer2 & GICD_TYPER2_VIL) \
132 nvpeid = 1 + (gic_rdists->gicd_typer2 & \
137 #define ITS_MAX_VPEID (1 << (ITS_MAX_VPEID_BITS))
139 /* Convert page order to size in bytes */
140 #define PAGE_ORDER_TO_SIZE(o) (PAGE_SIZE << (o))
142 struct event_lpi_map {
143 unsigned long *lpi_map;
145 irq_hw_number_t lpi_base;
147 raw_spinlock_t vlpi_lock;
149 struct its_vlpi_map *vlpi_maps;
154 * The ITS view of a device - belongs to an ITS, owns an interrupt
155 * translation table, and a list of interrupts. If it some of its
156 * LPIs are injected into a guest (GICv4), the event_map.vm field
157 * indicates which one.
160 struct list_head entry;
161 struct its_node *its;
162 struct event_lpi_map event_map;
171 struct its_device *dev;
172 struct its_vpe **vpes;
176 struct cpu_lpi_count {
181 static DEFINE_PER_CPU(struct cpu_lpi_count, cpu_lpi_count);
183 static LIST_HEAD(its_nodes);
184 static DEFINE_RAW_SPINLOCK(its_lock);
185 static struct rdists *gic_rdists;
186 static struct irq_domain *its_parent;
188 static unsigned long its_list_map;
189 static u16 vmovp_seq_num;
190 static DEFINE_RAW_SPINLOCK(vmovp_lock);
192 static DEFINE_IDA(its_vpeid_ida);
194 #define gic_data_rdist() (raw_cpu_ptr(gic_rdists->rdist))
195 #define gic_data_rdist_cpu(cpu) (per_cpu_ptr(gic_rdists->rdist, cpu))
196 #define gic_data_rdist_rd_base() (gic_data_rdist()->rd_base)
197 #define gic_data_rdist_vlpi_base() (gic_data_rdist_rd_base() + SZ_128K)
200 * Skip ITSs that have no vLPIs mapped, unless we're on GICv4.1, as we
201 * always have vSGIs mapped.
203 static bool require_its_list_vmovp(struct its_vm *vm, struct its_node *its)
205 return (gic_rdists->has_rvpeid || vm->vlpi_count[its->list_nr]);
208 static u16 get_its_list(struct its_vm *vm)
210 struct its_node *its;
211 unsigned long its_list = 0;
213 list_for_each_entry(its, &its_nodes, entry) {
217 if (require_its_list_vmovp(vm, its))
218 __set_bit(its->list_nr, &its_list);
221 return (u16)its_list;
224 static inline u32 its_get_event_id(struct irq_data *d)
226 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
227 return d->hwirq - its_dev->event_map.lpi_base;
230 static struct its_collection *dev_event_to_col(struct its_device *its_dev,
233 struct its_node *its = its_dev->its;
235 return its->collections + its_dev->event_map.col_map[event];
238 static struct its_vlpi_map *dev_event_to_vlpi_map(struct its_device *its_dev,
241 if (WARN_ON_ONCE(event >= its_dev->event_map.nr_lpis))
244 return &its_dev->event_map.vlpi_maps[event];
247 static struct its_vlpi_map *get_vlpi_map(struct irq_data *d)
249 if (irqd_is_forwarded_to_vcpu(d)) {
250 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
251 u32 event = its_get_event_id(d);
253 return dev_event_to_vlpi_map(its_dev, event);
259 static int vpe_to_cpuid_lock(struct its_vpe *vpe, unsigned long *flags)
261 raw_spin_lock_irqsave(&vpe->vpe_lock, *flags);
265 static void vpe_to_cpuid_unlock(struct its_vpe *vpe, unsigned long flags)
267 raw_spin_unlock_irqrestore(&vpe->vpe_lock, flags);
270 static int irq_to_cpuid_lock(struct irq_data *d, unsigned long *flags)
272 struct its_vlpi_map *map = get_vlpi_map(d);
276 cpu = vpe_to_cpuid_lock(map->vpe, flags);
278 /* Physical LPIs are already locked via the irq_desc lock */
279 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
280 cpu = its_dev->event_map.col_map[its_get_event_id(d)];
281 /* Keep GCC quiet... */
288 static void irq_to_cpuid_unlock(struct irq_data *d, unsigned long flags)
290 struct its_vlpi_map *map = get_vlpi_map(d);
293 vpe_to_cpuid_unlock(map->vpe, flags);
296 static struct its_collection *valid_col(struct its_collection *col)
298 if (WARN_ON_ONCE(col->target_address & GENMASK_ULL(15, 0)))
304 static struct its_vpe *valid_vpe(struct its_node *its, struct its_vpe *vpe)
306 if (valid_col(its->collections + vpe->col_idx))
313 * ITS command descriptors - parameters to be encoded in a command
316 struct its_cmd_desc {
319 struct its_device *dev;
324 struct its_device *dev;
329 struct its_device *dev;
334 struct its_device *dev;
339 struct its_collection *col;
344 struct its_device *dev;
350 struct its_device *dev;
351 struct its_collection *col;
356 struct its_device *dev;
361 struct its_collection *col;
370 struct its_collection *col;
376 struct its_device *dev;
384 struct its_device *dev;
391 struct its_collection *col;
412 * The ITS command block, which is what the ITS actually parses.
414 struct its_cmd_block {
417 __le64 raw_cmd_le[4];
421 #define ITS_CMD_QUEUE_SZ SZ_64K
422 #define ITS_CMD_QUEUE_NR_ENTRIES (ITS_CMD_QUEUE_SZ / sizeof(struct its_cmd_block))
424 typedef struct its_collection *(*its_cmd_builder_t)(struct its_node *,
425 struct its_cmd_block *,
426 struct its_cmd_desc *);
428 typedef struct its_vpe *(*its_cmd_vbuilder_t)(struct its_node *,
429 struct its_cmd_block *,
430 struct its_cmd_desc *);
432 static void its_mask_encode(u64 *raw_cmd, u64 val, int h, int l)
434 u64 mask = GENMASK_ULL(h, l);
436 *raw_cmd |= (val << l) & mask;
439 static void its_encode_cmd(struct its_cmd_block *cmd, u8 cmd_nr)
441 its_mask_encode(&cmd->raw_cmd[0], cmd_nr, 7, 0);
444 static void its_encode_devid(struct its_cmd_block *cmd, u32 devid)
446 its_mask_encode(&cmd->raw_cmd[0], devid, 63, 32);
449 static void its_encode_event_id(struct its_cmd_block *cmd, u32 id)
451 its_mask_encode(&cmd->raw_cmd[1], id, 31, 0);
454 static void its_encode_phys_id(struct its_cmd_block *cmd, u32 phys_id)
456 its_mask_encode(&cmd->raw_cmd[1], phys_id, 63, 32);
459 static void its_encode_size(struct its_cmd_block *cmd, u8 size)
461 its_mask_encode(&cmd->raw_cmd[1], size, 4, 0);
464 static void its_encode_itt(struct its_cmd_block *cmd, u64 itt_addr)
466 its_mask_encode(&cmd->raw_cmd[2], itt_addr >> 8, 51, 8);
469 static void its_encode_valid(struct its_cmd_block *cmd, int valid)
471 its_mask_encode(&cmd->raw_cmd[2], !!valid, 63, 63);
474 static void its_encode_target(struct its_cmd_block *cmd, u64 target_addr)
476 its_mask_encode(&cmd->raw_cmd[2], target_addr >> 16, 51, 16);
479 static void its_encode_collection(struct its_cmd_block *cmd, u16 col)
481 its_mask_encode(&cmd->raw_cmd[2], col, 15, 0);
484 static void its_encode_vpeid(struct its_cmd_block *cmd, u16 vpeid)
486 its_mask_encode(&cmd->raw_cmd[1], vpeid, 47, 32);
489 static void its_encode_virt_id(struct its_cmd_block *cmd, u32 virt_id)
491 its_mask_encode(&cmd->raw_cmd[2], virt_id, 31, 0);
494 static void its_encode_db_phys_id(struct its_cmd_block *cmd, u32 db_phys_id)
496 its_mask_encode(&cmd->raw_cmd[2], db_phys_id, 63, 32);
499 static void its_encode_db_valid(struct its_cmd_block *cmd, bool db_valid)
501 its_mask_encode(&cmd->raw_cmd[2], db_valid, 0, 0);
504 static void its_encode_seq_num(struct its_cmd_block *cmd, u16 seq_num)
506 its_mask_encode(&cmd->raw_cmd[0], seq_num, 47, 32);
509 static void its_encode_its_list(struct its_cmd_block *cmd, u16 its_list)
511 its_mask_encode(&cmd->raw_cmd[1], its_list, 15, 0);
514 static void its_encode_vpt_addr(struct its_cmd_block *cmd, u64 vpt_pa)
516 its_mask_encode(&cmd->raw_cmd[3], vpt_pa >> 16, 51, 16);
519 static void its_encode_vpt_size(struct its_cmd_block *cmd, u8 vpt_size)
521 its_mask_encode(&cmd->raw_cmd[3], vpt_size, 4, 0);
524 static void its_encode_vconf_addr(struct its_cmd_block *cmd, u64 vconf_pa)
526 its_mask_encode(&cmd->raw_cmd[0], vconf_pa >> 16, 51, 16);
529 static void its_encode_alloc(struct its_cmd_block *cmd, bool alloc)
531 its_mask_encode(&cmd->raw_cmd[0], alloc, 8, 8);
534 static void its_encode_ptz(struct its_cmd_block *cmd, bool ptz)
536 its_mask_encode(&cmd->raw_cmd[0], ptz, 9, 9);
539 static void its_encode_vmapp_default_db(struct its_cmd_block *cmd,
542 its_mask_encode(&cmd->raw_cmd[1], vpe_db_lpi, 31, 0);
545 static void its_encode_vmovp_default_db(struct its_cmd_block *cmd,
548 its_mask_encode(&cmd->raw_cmd[3], vpe_db_lpi, 31, 0);
551 static void its_encode_db(struct its_cmd_block *cmd, bool db)
553 its_mask_encode(&cmd->raw_cmd[2], db, 63, 63);
556 static void its_encode_sgi_intid(struct its_cmd_block *cmd, u8 sgi)
558 its_mask_encode(&cmd->raw_cmd[0], sgi, 35, 32);
561 static void its_encode_sgi_priority(struct its_cmd_block *cmd, u8 prio)
563 its_mask_encode(&cmd->raw_cmd[0], prio >> 4, 23, 20);
566 static void its_encode_sgi_group(struct its_cmd_block *cmd, bool grp)
568 its_mask_encode(&cmd->raw_cmd[0], grp, 10, 10);
571 static void its_encode_sgi_clear(struct its_cmd_block *cmd, bool clr)
573 its_mask_encode(&cmd->raw_cmd[0], clr, 9, 9);
576 static void its_encode_sgi_enable(struct its_cmd_block *cmd, bool en)
578 its_mask_encode(&cmd->raw_cmd[0], en, 8, 8);
581 static inline void its_fixup_cmd(struct its_cmd_block *cmd)
583 /* Let's fixup BE commands */
584 cmd->raw_cmd_le[0] = cpu_to_le64(cmd->raw_cmd[0]);
585 cmd->raw_cmd_le[1] = cpu_to_le64(cmd->raw_cmd[1]);
586 cmd->raw_cmd_le[2] = cpu_to_le64(cmd->raw_cmd[2]);
587 cmd->raw_cmd_le[3] = cpu_to_le64(cmd->raw_cmd[3]);
590 static struct its_collection *its_build_mapd_cmd(struct its_node *its,
591 struct its_cmd_block *cmd,
592 struct its_cmd_desc *desc)
594 unsigned long itt_addr;
595 u8 size = ilog2(desc->its_mapd_cmd.dev->nr_ites);
597 itt_addr = virt_to_phys(desc->its_mapd_cmd.dev->itt);
598 itt_addr = ALIGN(itt_addr, ITS_ITT_ALIGN);
600 its_encode_cmd(cmd, GITS_CMD_MAPD);
601 its_encode_devid(cmd, desc->its_mapd_cmd.dev->device_id);
602 its_encode_size(cmd, size - 1);
603 its_encode_itt(cmd, itt_addr);
604 its_encode_valid(cmd, desc->its_mapd_cmd.valid);
611 static struct its_collection *its_build_mapc_cmd(struct its_node *its,
612 struct its_cmd_block *cmd,
613 struct its_cmd_desc *desc)
615 its_encode_cmd(cmd, GITS_CMD_MAPC);
616 its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
617 its_encode_target(cmd, desc->its_mapc_cmd.col->target_address);
618 its_encode_valid(cmd, desc->its_mapc_cmd.valid);
622 return desc->its_mapc_cmd.col;
625 static struct its_collection *its_build_mapti_cmd(struct its_node *its,
626 struct its_cmd_block *cmd,
627 struct its_cmd_desc *desc)
629 struct its_collection *col;
631 col = dev_event_to_col(desc->its_mapti_cmd.dev,
632 desc->its_mapti_cmd.event_id);
634 its_encode_cmd(cmd, GITS_CMD_MAPTI);
635 its_encode_devid(cmd, desc->its_mapti_cmd.dev->device_id);
636 its_encode_event_id(cmd, desc->its_mapti_cmd.event_id);
637 its_encode_phys_id(cmd, desc->its_mapti_cmd.phys_id);
638 its_encode_collection(cmd, col->col_id);
642 return valid_col(col);
645 static struct its_collection *its_build_movi_cmd(struct its_node *its,
646 struct its_cmd_block *cmd,
647 struct its_cmd_desc *desc)
649 struct its_collection *col;
651 col = dev_event_to_col(desc->its_movi_cmd.dev,
652 desc->its_movi_cmd.event_id);
654 its_encode_cmd(cmd, GITS_CMD_MOVI);
655 its_encode_devid(cmd, desc->its_movi_cmd.dev->device_id);
656 its_encode_event_id(cmd, desc->its_movi_cmd.event_id);
657 its_encode_collection(cmd, desc->its_movi_cmd.col->col_id);
661 return valid_col(col);
664 static struct its_collection *its_build_discard_cmd(struct its_node *its,
665 struct its_cmd_block *cmd,
666 struct its_cmd_desc *desc)
668 struct its_collection *col;
670 col = dev_event_to_col(desc->its_discard_cmd.dev,
671 desc->its_discard_cmd.event_id);
673 its_encode_cmd(cmd, GITS_CMD_DISCARD);
674 its_encode_devid(cmd, desc->its_discard_cmd.dev->device_id);
675 its_encode_event_id(cmd, desc->its_discard_cmd.event_id);
679 return valid_col(col);
682 static struct its_collection *its_build_inv_cmd(struct its_node *its,
683 struct its_cmd_block *cmd,
684 struct its_cmd_desc *desc)
686 struct its_collection *col;
688 col = dev_event_to_col(desc->its_inv_cmd.dev,
689 desc->its_inv_cmd.event_id);
691 its_encode_cmd(cmd, GITS_CMD_INV);
692 its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
693 its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
697 return valid_col(col);
700 static struct its_collection *its_build_int_cmd(struct its_node *its,
701 struct its_cmd_block *cmd,
702 struct its_cmd_desc *desc)
704 struct its_collection *col;
706 col = dev_event_to_col(desc->its_int_cmd.dev,
707 desc->its_int_cmd.event_id);
709 its_encode_cmd(cmd, GITS_CMD_INT);
710 its_encode_devid(cmd, desc->its_int_cmd.dev->device_id);
711 its_encode_event_id(cmd, desc->its_int_cmd.event_id);
715 return valid_col(col);
718 static struct its_collection *its_build_clear_cmd(struct its_node *its,
719 struct its_cmd_block *cmd,
720 struct its_cmd_desc *desc)
722 struct its_collection *col;
724 col = dev_event_to_col(desc->its_clear_cmd.dev,
725 desc->its_clear_cmd.event_id);
727 its_encode_cmd(cmd, GITS_CMD_CLEAR);
728 its_encode_devid(cmd, desc->its_clear_cmd.dev->device_id);
729 its_encode_event_id(cmd, desc->its_clear_cmd.event_id);
733 return valid_col(col);
736 static struct its_collection *its_build_invall_cmd(struct its_node *its,
737 struct its_cmd_block *cmd,
738 struct its_cmd_desc *desc)
740 its_encode_cmd(cmd, GITS_CMD_INVALL);
741 its_encode_collection(cmd, desc->its_invall_cmd.col->col_id);
745 return desc->its_invall_cmd.col;
748 static struct its_vpe *its_build_vinvall_cmd(struct its_node *its,
749 struct its_cmd_block *cmd,
750 struct its_cmd_desc *desc)
752 its_encode_cmd(cmd, GITS_CMD_VINVALL);
753 its_encode_vpeid(cmd, desc->its_vinvall_cmd.vpe->vpe_id);
757 return valid_vpe(its, desc->its_vinvall_cmd.vpe);
760 static struct its_vpe *its_build_vmapp_cmd(struct its_node *its,
761 struct its_cmd_block *cmd,
762 struct its_cmd_desc *desc)
764 unsigned long vpt_addr, vconf_addr;
768 its_encode_cmd(cmd, GITS_CMD_VMAPP);
769 its_encode_vpeid(cmd, desc->its_vmapp_cmd.vpe->vpe_id);
770 its_encode_valid(cmd, desc->its_vmapp_cmd.valid);
772 if (!desc->its_vmapp_cmd.valid) {
774 alloc = !atomic_dec_return(&desc->its_vmapp_cmd.vpe->vmapp_count);
775 its_encode_alloc(cmd, alloc);
781 vpt_addr = virt_to_phys(page_address(desc->its_vmapp_cmd.vpe->vpt_page));
782 target = desc->its_vmapp_cmd.col->target_address + its->vlpi_redist_offset;
784 its_encode_target(cmd, target);
785 its_encode_vpt_addr(cmd, vpt_addr);
786 its_encode_vpt_size(cmd, LPI_NRBITS - 1);
791 vconf_addr = virt_to_phys(page_address(desc->its_vmapp_cmd.vpe->its_vm->vprop_page));
793 alloc = !atomic_fetch_inc(&desc->its_vmapp_cmd.vpe->vmapp_count);
795 its_encode_alloc(cmd, alloc);
798 * GICv4.1 provides a way to get the VLPI state, which needs the vPE
799 * to be unmapped first, and in this case, we may remap the vPE
800 * back while the VPT is not empty. So we can't assume that the
801 * VPT is empty on map. This is why we never advertise PTZ.
803 its_encode_ptz(cmd, false);
804 its_encode_vconf_addr(cmd, vconf_addr);
805 its_encode_vmapp_default_db(cmd, desc->its_vmapp_cmd.vpe->vpe_db_lpi);
810 return valid_vpe(its, desc->its_vmapp_cmd.vpe);
813 static struct its_vpe *its_build_vmapti_cmd(struct its_node *its,
814 struct its_cmd_block *cmd,
815 struct its_cmd_desc *desc)
819 if (!is_v4_1(its) && desc->its_vmapti_cmd.db_enabled)
820 db = desc->its_vmapti_cmd.vpe->vpe_db_lpi;
824 its_encode_cmd(cmd, GITS_CMD_VMAPTI);
825 its_encode_devid(cmd, desc->its_vmapti_cmd.dev->device_id);
826 its_encode_vpeid(cmd, desc->its_vmapti_cmd.vpe->vpe_id);
827 its_encode_event_id(cmd, desc->its_vmapti_cmd.event_id);
828 its_encode_db_phys_id(cmd, db);
829 its_encode_virt_id(cmd, desc->its_vmapti_cmd.virt_id);
833 return valid_vpe(its, desc->its_vmapti_cmd.vpe);
836 static struct its_vpe *its_build_vmovi_cmd(struct its_node *its,
837 struct its_cmd_block *cmd,
838 struct its_cmd_desc *desc)
842 if (!is_v4_1(its) && desc->its_vmovi_cmd.db_enabled)
843 db = desc->its_vmovi_cmd.vpe->vpe_db_lpi;
847 its_encode_cmd(cmd, GITS_CMD_VMOVI);
848 its_encode_devid(cmd, desc->its_vmovi_cmd.dev->device_id);
849 its_encode_vpeid(cmd, desc->its_vmovi_cmd.vpe->vpe_id);
850 its_encode_event_id(cmd, desc->its_vmovi_cmd.event_id);
851 its_encode_db_phys_id(cmd, db);
852 its_encode_db_valid(cmd, true);
856 return valid_vpe(its, desc->its_vmovi_cmd.vpe);
859 static struct its_vpe *its_build_vmovp_cmd(struct its_node *its,
860 struct its_cmd_block *cmd,
861 struct its_cmd_desc *desc)
865 target = desc->its_vmovp_cmd.col->target_address + its->vlpi_redist_offset;
866 its_encode_cmd(cmd, GITS_CMD_VMOVP);
867 its_encode_seq_num(cmd, desc->its_vmovp_cmd.seq_num);
868 its_encode_its_list(cmd, desc->its_vmovp_cmd.its_list);
869 its_encode_vpeid(cmd, desc->its_vmovp_cmd.vpe->vpe_id);
870 its_encode_target(cmd, target);
873 its_encode_db(cmd, true);
874 its_encode_vmovp_default_db(cmd, desc->its_vmovp_cmd.vpe->vpe_db_lpi);
879 return valid_vpe(its, desc->its_vmovp_cmd.vpe);
882 static struct its_vpe *its_build_vinv_cmd(struct its_node *its,
883 struct its_cmd_block *cmd,
884 struct its_cmd_desc *desc)
886 struct its_vlpi_map *map;
888 map = dev_event_to_vlpi_map(desc->its_inv_cmd.dev,
889 desc->its_inv_cmd.event_id);
891 its_encode_cmd(cmd, GITS_CMD_INV);
892 its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
893 its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
897 return valid_vpe(its, map->vpe);
900 static struct its_vpe *its_build_vint_cmd(struct its_node *its,
901 struct its_cmd_block *cmd,
902 struct its_cmd_desc *desc)
904 struct its_vlpi_map *map;
906 map = dev_event_to_vlpi_map(desc->its_int_cmd.dev,
907 desc->its_int_cmd.event_id);
909 its_encode_cmd(cmd, GITS_CMD_INT);
910 its_encode_devid(cmd, desc->its_int_cmd.dev->device_id);
911 its_encode_event_id(cmd, desc->its_int_cmd.event_id);
915 return valid_vpe(its, map->vpe);
918 static struct its_vpe *its_build_vclear_cmd(struct its_node *its,
919 struct its_cmd_block *cmd,
920 struct its_cmd_desc *desc)
922 struct its_vlpi_map *map;
924 map = dev_event_to_vlpi_map(desc->its_clear_cmd.dev,
925 desc->its_clear_cmd.event_id);
927 its_encode_cmd(cmd, GITS_CMD_CLEAR);
928 its_encode_devid(cmd, desc->its_clear_cmd.dev->device_id);
929 its_encode_event_id(cmd, desc->its_clear_cmd.event_id);
933 return valid_vpe(its, map->vpe);
936 static struct its_vpe *its_build_invdb_cmd(struct its_node *its,
937 struct its_cmd_block *cmd,
938 struct its_cmd_desc *desc)
940 if (WARN_ON(!is_v4_1(its)))
943 its_encode_cmd(cmd, GITS_CMD_INVDB);
944 its_encode_vpeid(cmd, desc->its_invdb_cmd.vpe->vpe_id);
948 return valid_vpe(its, desc->its_invdb_cmd.vpe);
951 static struct its_vpe *its_build_vsgi_cmd(struct its_node *its,
952 struct its_cmd_block *cmd,
953 struct its_cmd_desc *desc)
955 if (WARN_ON(!is_v4_1(its)))
958 its_encode_cmd(cmd, GITS_CMD_VSGI);
959 its_encode_vpeid(cmd, desc->its_vsgi_cmd.vpe->vpe_id);
960 its_encode_sgi_intid(cmd, desc->its_vsgi_cmd.sgi);
961 its_encode_sgi_priority(cmd, desc->its_vsgi_cmd.priority);
962 its_encode_sgi_group(cmd, desc->its_vsgi_cmd.group);
963 its_encode_sgi_clear(cmd, desc->its_vsgi_cmd.clear);
964 its_encode_sgi_enable(cmd, desc->its_vsgi_cmd.enable);
968 return valid_vpe(its, desc->its_vsgi_cmd.vpe);
971 static u64 its_cmd_ptr_to_offset(struct its_node *its,
972 struct its_cmd_block *ptr)
974 return (ptr - its->cmd_base) * sizeof(*ptr);
977 static int its_queue_full(struct its_node *its)
982 widx = its->cmd_write - its->cmd_base;
983 ridx = readl_relaxed(its->base + GITS_CREADR) / sizeof(struct its_cmd_block);
985 /* This is incredibly unlikely to happen, unless the ITS locks up. */
986 if (((widx + 1) % ITS_CMD_QUEUE_NR_ENTRIES) == ridx)
992 static struct its_cmd_block *its_allocate_entry(struct its_node *its)
994 struct its_cmd_block *cmd;
995 u32 count = 1000000; /* 1s! */
997 while (its_queue_full(its)) {
1000 pr_err_ratelimited("ITS queue not draining\n");
1007 cmd = its->cmd_write++;
1009 /* Handle queue wrapping */
1010 if (its->cmd_write == (its->cmd_base + ITS_CMD_QUEUE_NR_ENTRIES))
1011 its->cmd_write = its->cmd_base;
1014 cmd->raw_cmd[0] = 0;
1015 cmd->raw_cmd[1] = 0;
1016 cmd->raw_cmd[2] = 0;
1017 cmd->raw_cmd[3] = 0;
1022 static struct its_cmd_block *its_post_commands(struct its_node *its)
1024 u64 wr = its_cmd_ptr_to_offset(its, its->cmd_write);
1026 writel_relaxed(wr, its->base + GITS_CWRITER);
1028 return its->cmd_write;
1031 static void its_flush_cmd(struct its_node *its, struct its_cmd_block *cmd)
1034 * Make sure the commands written to memory are observable by
1037 if (its->flags & ITS_FLAGS_CMDQ_NEEDS_FLUSHING)
1038 gic_flush_dcache_to_poc(cmd, sizeof(*cmd));
1043 static int its_wait_for_range_completion(struct its_node *its,
1045 struct its_cmd_block *to)
1047 u64 rd_idx, to_idx, linear_idx;
1048 u32 count = 1000000; /* 1s! */
1050 /* Linearize to_idx if the command set has wrapped around */
1051 to_idx = its_cmd_ptr_to_offset(its, to);
1052 if (to_idx < prev_idx)
1053 to_idx += ITS_CMD_QUEUE_SZ;
1055 linear_idx = prev_idx;
1060 rd_idx = readl_relaxed(its->base + GITS_CREADR);
1063 * Compute the read pointer progress, taking the
1064 * potential wrap-around into account.
1066 delta = rd_idx - prev_idx;
1067 if (rd_idx < prev_idx)
1068 delta += ITS_CMD_QUEUE_SZ;
1070 linear_idx += delta;
1071 if (linear_idx >= to_idx)
1076 pr_err_ratelimited("ITS queue timeout (%llu %llu)\n",
1077 to_idx, linear_idx);
1088 /* Warning, macro hell follows */
1089 #define BUILD_SINGLE_CMD_FUNC(name, buildtype, synctype, buildfn) \
1090 void name(struct its_node *its, \
1091 buildtype builder, \
1092 struct its_cmd_desc *desc) \
1094 struct its_cmd_block *cmd, *sync_cmd, *next_cmd; \
1095 synctype *sync_obj; \
1096 unsigned long flags; \
1099 raw_spin_lock_irqsave(&its->lock, flags); \
1101 cmd = its_allocate_entry(its); \
1102 if (!cmd) { /* We're soooooo screewed... */ \
1103 raw_spin_unlock_irqrestore(&its->lock, flags); \
1106 sync_obj = builder(its, cmd, desc); \
1107 its_flush_cmd(its, cmd); \
1110 sync_cmd = its_allocate_entry(its); \
1114 buildfn(its, sync_cmd, sync_obj); \
1115 its_flush_cmd(its, sync_cmd); \
1119 rd_idx = readl_relaxed(its->base + GITS_CREADR); \
1120 next_cmd = its_post_commands(its); \
1121 raw_spin_unlock_irqrestore(&its->lock, flags); \
1123 if (its_wait_for_range_completion(its, rd_idx, next_cmd)) \
1124 pr_err_ratelimited("ITS cmd %ps failed\n", builder); \
1127 static void its_build_sync_cmd(struct its_node *its,
1128 struct its_cmd_block *sync_cmd,
1129 struct its_collection *sync_col)
1131 its_encode_cmd(sync_cmd, GITS_CMD_SYNC);
1132 its_encode_target(sync_cmd, sync_col->target_address);
1134 its_fixup_cmd(sync_cmd);
1137 static BUILD_SINGLE_CMD_FUNC(its_send_single_command, its_cmd_builder_t,
1138 struct its_collection, its_build_sync_cmd)
1140 static void its_build_vsync_cmd(struct its_node *its,
1141 struct its_cmd_block *sync_cmd,
1142 struct its_vpe *sync_vpe)
1144 its_encode_cmd(sync_cmd, GITS_CMD_VSYNC);
1145 its_encode_vpeid(sync_cmd, sync_vpe->vpe_id);
1147 its_fixup_cmd(sync_cmd);
1150 static BUILD_SINGLE_CMD_FUNC(its_send_single_vcommand, its_cmd_vbuilder_t,
1151 struct its_vpe, its_build_vsync_cmd)
1153 static void its_send_int(struct its_device *dev, u32 event_id)
1155 struct its_cmd_desc desc;
1157 desc.its_int_cmd.dev = dev;
1158 desc.its_int_cmd.event_id = event_id;
1160 its_send_single_command(dev->its, its_build_int_cmd, &desc);
1163 static void its_send_clear(struct its_device *dev, u32 event_id)
1165 struct its_cmd_desc desc;
1167 desc.its_clear_cmd.dev = dev;
1168 desc.its_clear_cmd.event_id = event_id;
1170 its_send_single_command(dev->its, its_build_clear_cmd, &desc);
1173 static void its_send_inv(struct its_device *dev, u32 event_id)
1175 struct its_cmd_desc desc;
1177 desc.its_inv_cmd.dev = dev;
1178 desc.its_inv_cmd.event_id = event_id;
1180 its_send_single_command(dev->its, its_build_inv_cmd, &desc);
1183 static void its_send_mapd(struct its_device *dev, int valid)
1185 struct its_cmd_desc desc;
1187 desc.its_mapd_cmd.dev = dev;
1188 desc.its_mapd_cmd.valid = !!valid;
1190 its_send_single_command(dev->its, its_build_mapd_cmd, &desc);
1193 static void its_send_mapc(struct its_node *its, struct its_collection *col,
1196 struct its_cmd_desc desc;
1198 desc.its_mapc_cmd.col = col;
1199 desc.its_mapc_cmd.valid = !!valid;
1201 its_send_single_command(its, its_build_mapc_cmd, &desc);
1204 static void its_send_mapti(struct its_device *dev, u32 irq_id, u32 id)
1206 struct its_cmd_desc desc;
1208 desc.its_mapti_cmd.dev = dev;
1209 desc.its_mapti_cmd.phys_id = irq_id;
1210 desc.its_mapti_cmd.event_id = id;
1212 its_send_single_command(dev->its, its_build_mapti_cmd, &desc);
1215 static void its_send_movi(struct its_device *dev,
1216 struct its_collection *col, u32 id)
1218 struct its_cmd_desc desc;
1220 desc.its_movi_cmd.dev = dev;
1221 desc.its_movi_cmd.col = col;
1222 desc.its_movi_cmd.event_id = id;
1224 its_send_single_command(dev->its, its_build_movi_cmd, &desc);
1227 static void its_send_discard(struct its_device *dev, u32 id)
1229 struct its_cmd_desc desc;
1231 desc.its_discard_cmd.dev = dev;
1232 desc.its_discard_cmd.event_id = id;
1234 its_send_single_command(dev->its, its_build_discard_cmd, &desc);
1237 static void its_send_invall(struct its_node *its, struct its_collection *col)
1239 struct its_cmd_desc desc;
1241 desc.its_invall_cmd.col = col;
1243 its_send_single_command(its, its_build_invall_cmd, &desc);
1246 static void its_send_vmapti(struct its_device *dev, u32 id)
1248 struct its_vlpi_map *map = dev_event_to_vlpi_map(dev, id);
1249 struct its_cmd_desc desc;
1251 desc.its_vmapti_cmd.vpe = map->vpe;
1252 desc.its_vmapti_cmd.dev = dev;
1253 desc.its_vmapti_cmd.virt_id = map->vintid;
1254 desc.its_vmapti_cmd.event_id = id;
1255 desc.its_vmapti_cmd.db_enabled = map->db_enabled;
1257 its_send_single_vcommand(dev->its, its_build_vmapti_cmd, &desc);
1260 static void its_send_vmovi(struct its_device *dev, u32 id)
1262 struct its_vlpi_map *map = dev_event_to_vlpi_map(dev, id);
1263 struct its_cmd_desc desc;
1265 desc.its_vmovi_cmd.vpe = map->vpe;
1266 desc.its_vmovi_cmd.dev = dev;
1267 desc.its_vmovi_cmd.event_id = id;
1268 desc.its_vmovi_cmd.db_enabled = map->db_enabled;
1270 its_send_single_vcommand(dev->its, its_build_vmovi_cmd, &desc);
1273 static void its_send_vmapp(struct its_node *its,
1274 struct its_vpe *vpe, bool valid)
1276 struct its_cmd_desc desc;
1278 desc.its_vmapp_cmd.vpe = vpe;
1279 desc.its_vmapp_cmd.valid = valid;
1280 desc.its_vmapp_cmd.col = &its->collections[vpe->col_idx];
1282 its_send_single_vcommand(its, its_build_vmapp_cmd, &desc);
1285 static void its_send_vmovp(struct its_vpe *vpe)
1287 struct its_cmd_desc desc = {};
1288 struct its_node *its;
1289 unsigned long flags;
1290 int col_id = vpe->col_idx;
1292 desc.its_vmovp_cmd.vpe = vpe;
1294 if (!its_list_map) {
1295 its = list_first_entry(&its_nodes, struct its_node, entry);
1296 desc.its_vmovp_cmd.col = &its->collections[col_id];
1297 its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
1302 * Yet another marvel of the architecture. If using the
1303 * its_list "feature", we need to make sure that all ITSs
1304 * receive all VMOVP commands in the same order. The only way
1305 * to guarantee this is to make vmovp a serialization point.
1309 raw_spin_lock_irqsave(&vmovp_lock, flags);
1311 desc.its_vmovp_cmd.seq_num = vmovp_seq_num++;
1312 desc.its_vmovp_cmd.its_list = get_its_list(vpe->its_vm);
1315 list_for_each_entry(its, &its_nodes, entry) {
1319 if (!require_its_list_vmovp(vpe->its_vm, its))
1322 desc.its_vmovp_cmd.col = &its->collections[col_id];
1323 its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
1326 raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1329 static void its_send_vinvall(struct its_node *its, struct its_vpe *vpe)
1331 struct its_cmd_desc desc;
1333 desc.its_vinvall_cmd.vpe = vpe;
1334 its_send_single_vcommand(its, its_build_vinvall_cmd, &desc);
1337 static void its_send_vinv(struct its_device *dev, u32 event_id)
1339 struct its_cmd_desc desc;
1342 * There is no real VINV command. This is just a normal INV,
1343 * with a VSYNC instead of a SYNC.
1345 desc.its_inv_cmd.dev = dev;
1346 desc.its_inv_cmd.event_id = event_id;
1348 its_send_single_vcommand(dev->its, its_build_vinv_cmd, &desc);
1351 static void its_send_vint(struct its_device *dev, u32 event_id)
1353 struct its_cmd_desc desc;
1356 * There is no real VINT command. This is just a normal INT,
1357 * with a VSYNC instead of a SYNC.
1359 desc.its_int_cmd.dev = dev;
1360 desc.its_int_cmd.event_id = event_id;
1362 its_send_single_vcommand(dev->its, its_build_vint_cmd, &desc);
1365 static void its_send_vclear(struct its_device *dev, u32 event_id)
1367 struct its_cmd_desc desc;
1370 * There is no real VCLEAR command. This is just a normal CLEAR,
1371 * with a VSYNC instead of a SYNC.
1373 desc.its_clear_cmd.dev = dev;
1374 desc.its_clear_cmd.event_id = event_id;
1376 its_send_single_vcommand(dev->its, its_build_vclear_cmd, &desc);
1379 static void its_send_invdb(struct its_node *its, struct its_vpe *vpe)
1381 struct its_cmd_desc desc;
1383 desc.its_invdb_cmd.vpe = vpe;
1384 its_send_single_vcommand(its, its_build_invdb_cmd, &desc);
1388 * irqchip functions - assumes MSI, mostly.
1390 static void lpi_write_config(struct irq_data *d, u8 clr, u8 set)
1392 struct its_vlpi_map *map = get_vlpi_map(d);
1393 irq_hw_number_t hwirq;
1398 va = page_address(map->vm->vprop_page);
1399 hwirq = map->vintid;
1401 /* Remember the updated property */
1402 map->properties &= ~clr;
1403 map->properties |= set | LPI_PROP_GROUP1;
1405 va = gic_rdists->prop_table_va;
1409 cfg = va + hwirq - 8192;
1411 *cfg |= set | LPI_PROP_GROUP1;
1414 * Make the above write visible to the redistributors.
1415 * And yes, we're flushing exactly: One. Single. Byte.
1418 if (gic_rdists->flags & RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING)
1419 gic_flush_dcache_to_poc(cfg, sizeof(*cfg));
1424 static void wait_for_syncr(void __iomem *rdbase)
1426 while (readl_relaxed(rdbase + GICR_SYNCR) & 1)
1430 static void direct_lpi_inv(struct irq_data *d)
1432 struct its_vlpi_map *map = get_vlpi_map(d);
1433 void __iomem *rdbase;
1434 unsigned long flags;
1439 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1441 WARN_ON(!is_v4_1(its_dev->its));
1443 val = GICR_INVLPIR_V;
1444 val |= FIELD_PREP(GICR_INVLPIR_VPEID, map->vpe->vpe_id);
1445 val |= FIELD_PREP(GICR_INVLPIR_INTID, map->vintid);
1450 /* Target the redistributor this LPI is currently routed to */
1451 cpu = irq_to_cpuid_lock(d, &flags);
1452 raw_spin_lock(&gic_data_rdist_cpu(cpu)->rd_lock);
1453 rdbase = per_cpu_ptr(gic_rdists->rdist, cpu)->rd_base;
1454 gic_write_lpir(val, rdbase + GICR_INVLPIR);
1456 wait_for_syncr(rdbase);
1457 raw_spin_unlock(&gic_data_rdist_cpu(cpu)->rd_lock);
1458 irq_to_cpuid_unlock(d, flags);
1461 static void lpi_update_config(struct irq_data *d, u8 clr, u8 set)
1463 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1465 lpi_write_config(d, clr, set);
1466 if (gic_rdists->has_direct_lpi &&
1467 (is_v4_1(its_dev->its) || !irqd_is_forwarded_to_vcpu(d)))
1469 else if (!irqd_is_forwarded_to_vcpu(d))
1470 its_send_inv(its_dev, its_get_event_id(d));
1472 its_send_vinv(its_dev, its_get_event_id(d));
1475 static void its_vlpi_set_doorbell(struct irq_data *d, bool enable)
1477 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1478 u32 event = its_get_event_id(d);
1479 struct its_vlpi_map *map;
1482 * GICv4.1 does away with the per-LPI nonsense, nothing to do
1485 if (is_v4_1(its_dev->its))
1488 map = dev_event_to_vlpi_map(its_dev, event);
1490 if (map->db_enabled == enable)
1493 map->db_enabled = enable;
1496 * More fun with the architecture:
1498 * Ideally, we'd issue a VMAPTI to set the doorbell to its LPI
1499 * value or to 1023, depending on the enable bit. But that
1500 * would be issuing a mapping for an /existing/ DevID+EventID
1501 * pair, which is UNPREDICTABLE. Instead, let's issue a VMOVI
1502 * to the /same/ vPE, using this opportunity to adjust the
1503 * doorbell. Mouahahahaha. We loves it, Precious.
1505 its_send_vmovi(its_dev, event);
1508 static void its_mask_irq(struct irq_data *d)
1510 if (irqd_is_forwarded_to_vcpu(d))
1511 its_vlpi_set_doorbell(d, false);
1513 lpi_update_config(d, LPI_PROP_ENABLED, 0);
1516 static void its_unmask_irq(struct irq_data *d)
1518 if (irqd_is_forwarded_to_vcpu(d))
1519 its_vlpi_set_doorbell(d, true);
1521 lpi_update_config(d, 0, LPI_PROP_ENABLED);
1524 static __maybe_unused u32 its_read_lpi_count(struct irq_data *d, int cpu)
1526 if (irqd_affinity_is_managed(d))
1527 return atomic_read(&per_cpu_ptr(&cpu_lpi_count, cpu)->managed);
1529 return atomic_read(&per_cpu_ptr(&cpu_lpi_count, cpu)->unmanaged);
1532 static void its_inc_lpi_count(struct irq_data *d, int cpu)
1534 if (irqd_affinity_is_managed(d))
1535 atomic_inc(&per_cpu_ptr(&cpu_lpi_count, cpu)->managed);
1537 atomic_inc(&per_cpu_ptr(&cpu_lpi_count, cpu)->unmanaged);
1540 static void its_dec_lpi_count(struct irq_data *d, int cpu)
1542 if (irqd_affinity_is_managed(d))
1543 atomic_dec(&per_cpu_ptr(&cpu_lpi_count, cpu)->managed);
1545 atomic_dec(&per_cpu_ptr(&cpu_lpi_count, cpu)->unmanaged);
1548 static unsigned int cpumask_pick_least_loaded(struct irq_data *d,
1549 const struct cpumask *cpu_mask)
1551 unsigned int cpu = nr_cpu_ids, tmp;
1552 int count = S32_MAX;
1554 for_each_cpu(tmp, cpu_mask) {
1555 int this_count = its_read_lpi_count(d, tmp);
1556 if (this_count < count) {
1566 * As suggested by Thomas Gleixner in:
1567 * https://lore.kernel.org/r/87h80q2aoc.fsf@nanos.tec.linutronix.de
1569 static int its_select_cpu(struct irq_data *d,
1570 const struct cpumask *aff_mask)
1572 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1573 cpumask_var_t tmpmask;
1576 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
1579 node = its_dev->its->numa_node;
1581 if (!irqd_affinity_is_managed(d)) {
1582 /* First try the NUMA node */
1583 if (node != NUMA_NO_NODE) {
1585 * Try the intersection of the affinity mask and the
1586 * node mask (and the online mask, just to be safe).
1588 cpumask_and(tmpmask, cpumask_of_node(node), aff_mask);
1589 cpumask_and(tmpmask, tmpmask, cpu_online_mask);
1592 * Ideally, we would check if the mask is empty, and
1593 * try again on the full node here.
1595 * But it turns out that the way ACPI describes the
1596 * affinity for ITSs only deals about memory, and
1597 * not target CPUs, so it cannot describe a single
1598 * ITS placed next to two NUMA nodes.
1600 * Instead, just fallback on the online mask. This
1601 * diverges from Thomas' suggestion above.
1603 cpu = cpumask_pick_least_loaded(d, tmpmask);
1604 if (cpu < nr_cpu_ids)
1607 /* If we can't cross sockets, give up */
1608 if ((its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144))
1611 /* If the above failed, expand the search */
1614 /* Try the intersection of the affinity and online masks */
1615 cpumask_and(tmpmask, aff_mask, cpu_online_mask);
1617 /* If that doesn't fly, the online mask is the last resort */
1618 if (cpumask_empty(tmpmask))
1619 cpumask_copy(tmpmask, cpu_online_mask);
1621 cpu = cpumask_pick_least_loaded(d, tmpmask);
1623 cpumask_and(tmpmask, irq_data_get_affinity_mask(d), cpu_online_mask);
1625 /* If we cannot cross sockets, limit the search to that node */
1626 if ((its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) &&
1627 node != NUMA_NO_NODE)
1628 cpumask_and(tmpmask, tmpmask, cpumask_of_node(node));
1630 cpu = cpumask_pick_least_loaded(d, tmpmask);
1633 free_cpumask_var(tmpmask);
1635 pr_debug("IRQ%d -> %*pbl CPU%d\n", d->irq, cpumask_pr_args(aff_mask), cpu);
1639 static int its_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
1642 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1643 struct its_collection *target_col;
1644 u32 id = its_get_event_id(d);
1647 /* A forwarded interrupt should use irq_set_vcpu_affinity */
1648 if (irqd_is_forwarded_to_vcpu(d))
1651 prev_cpu = its_dev->event_map.col_map[id];
1652 its_dec_lpi_count(d, prev_cpu);
1655 cpu = its_select_cpu(d, mask_val);
1657 cpu = cpumask_pick_least_loaded(d, mask_val);
1659 if (cpu < 0 || cpu >= nr_cpu_ids)
1662 /* don't set the affinity when the target cpu is same as current one */
1663 if (cpu != prev_cpu) {
1664 target_col = &its_dev->its->collections[cpu];
1665 its_send_movi(its_dev, target_col, id);
1666 its_dev->event_map.col_map[id] = cpu;
1667 irq_data_update_effective_affinity(d, cpumask_of(cpu));
1670 its_inc_lpi_count(d, cpu);
1672 return IRQ_SET_MASK_OK_DONE;
1675 its_inc_lpi_count(d, prev_cpu);
1679 static u64 its_irq_get_msi_base(struct its_device *its_dev)
1681 struct its_node *its = its_dev->its;
1683 return its->phys_base + GITS_TRANSLATER;
1686 static void its_irq_compose_msi_msg(struct irq_data *d, struct msi_msg *msg)
1688 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1689 struct its_node *its;
1693 addr = its->get_msi_base(its_dev);
1695 msg->address_lo = lower_32_bits(addr);
1696 msg->address_hi = upper_32_bits(addr);
1697 msg->data = its_get_event_id(d);
1699 iommu_dma_compose_msi_msg(irq_data_get_msi_desc(d), msg);
1702 static int its_irq_set_irqchip_state(struct irq_data *d,
1703 enum irqchip_irq_state which,
1706 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1707 u32 event = its_get_event_id(d);
1709 if (which != IRQCHIP_STATE_PENDING)
1712 if (irqd_is_forwarded_to_vcpu(d)) {
1714 its_send_vint(its_dev, event);
1716 its_send_vclear(its_dev, event);
1719 its_send_int(its_dev, event);
1721 its_send_clear(its_dev, event);
1727 static int its_irq_retrigger(struct irq_data *d)
1729 return !its_irq_set_irqchip_state(d, IRQCHIP_STATE_PENDING, true);
1733 * Two favourable cases:
1735 * (a) Either we have a GICv4.1, and all vPEs have to be mapped at all times
1738 * (b) Or the ITSs do not use a list map, meaning that VMOVP is cheap enough
1739 * and we're better off mapping all VPEs always
1741 * If neither (a) nor (b) is true, then we map vPEs on demand.
1744 static bool gic_requires_eager_mapping(void)
1746 if (!its_list_map || gic_rdists->has_rvpeid)
1752 static void its_map_vm(struct its_node *its, struct its_vm *vm)
1754 unsigned long flags;
1756 if (gic_requires_eager_mapping())
1759 raw_spin_lock_irqsave(&vmovp_lock, flags);
1762 * If the VM wasn't mapped yet, iterate over the vpes and get
1765 vm->vlpi_count[its->list_nr]++;
1767 if (vm->vlpi_count[its->list_nr] == 1) {
1770 for (i = 0; i < vm->nr_vpes; i++) {
1771 struct its_vpe *vpe = vm->vpes[i];
1772 struct irq_data *d = irq_get_irq_data(vpe->irq);
1774 /* Map the VPE to the first possible CPU */
1775 vpe->col_idx = cpumask_first(cpu_online_mask);
1776 its_send_vmapp(its, vpe, true);
1777 its_send_vinvall(its, vpe);
1778 irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
1782 raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1785 static void its_unmap_vm(struct its_node *its, struct its_vm *vm)
1787 unsigned long flags;
1789 /* Not using the ITS list? Everything is always mapped. */
1790 if (gic_requires_eager_mapping())
1793 raw_spin_lock_irqsave(&vmovp_lock, flags);
1795 if (!--vm->vlpi_count[its->list_nr]) {
1798 for (i = 0; i < vm->nr_vpes; i++)
1799 its_send_vmapp(its, vm->vpes[i], false);
1802 raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1805 static int its_vlpi_map(struct irq_data *d, struct its_cmd_info *info)
1807 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1808 u32 event = its_get_event_id(d);
1814 raw_spin_lock(&its_dev->event_map.vlpi_lock);
1816 if (!its_dev->event_map.vm) {
1817 struct its_vlpi_map *maps;
1819 maps = kcalloc(its_dev->event_map.nr_lpis, sizeof(*maps),
1826 its_dev->event_map.vm = info->map->vm;
1827 its_dev->event_map.vlpi_maps = maps;
1828 } else if (its_dev->event_map.vm != info->map->vm) {
1833 /* Get our private copy of the mapping information */
1834 its_dev->event_map.vlpi_maps[event] = *info->map;
1836 if (irqd_is_forwarded_to_vcpu(d)) {
1837 /* Already mapped, move it around */
1838 its_send_vmovi(its_dev, event);
1840 /* Ensure all the VPEs are mapped on this ITS */
1841 its_map_vm(its_dev->its, info->map->vm);
1844 * Flag the interrupt as forwarded so that we can
1845 * start poking the virtual property table.
1847 irqd_set_forwarded_to_vcpu(d);
1849 /* Write out the property to the prop table */
1850 lpi_write_config(d, 0xff, info->map->properties);
1852 /* Drop the physical mapping */
1853 its_send_discard(its_dev, event);
1855 /* and install the virtual one */
1856 its_send_vmapti(its_dev, event);
1858 /* Increment the number of VLPIs */
1859 its_dev->event_map.nr_vlpis++;
1863 raw_spin_unlock(&its_dev->event_map.vlpi_lock);
1867 static int its_vlpi_get(struct irq_data *d, struct its_cmd_info *info)
1869 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1870 struct its_vlpi_map *map;
1873 raw_spin_lock(&its_dev->event_map.vlpi_lock);
1875 map = get_vlpi_map(d);
1877 if (!its_dev->event_map.vm || !map) {
1882 /* Copy our mapping information to the incoming request */
1886 raw_spin_unlock(&its_dev->event_map.vlpi_lock);
1890 static int its_vlpi_unmap(struct irq_data *d)
1892 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1893 u32 event = its_get_event_id(d);
1896 raw_spin_lock(&its_dev->event_map.vlpi_lock);
1898 if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d)) {
1903 /* Drop the virtual mapping */
1904 its_send_discard(its_dev, event);
1906 /* and restore the physical one */
1907 irqd_clr_forwarded_to_vcpu(d);
1908 its_send_mapti(its_dev, d->hwirq, event);
1909 lpi_update_config(d, 0xff, (LPI_PROP_DEFAULT_PRIO |
1913 /* Potentially unmap the VM from this ITS */
1914 its_unmap_vm(its_dev->its, its_dev->event_map.vm);
1917 * Drop the refcount and make the device available again if
1918 * this was the last VLPI.
1920 if (!--its_dev->event_map.nr_vlpis) {
1921 its_dev->event_map.vm = NULL;
1922 kfree(its_dev->event_map.vlpi_maps);
1926 raw_spin_unlock(&its_dev->event_map.vlpi_lock);
1930 static int its_vlpi_prop_update(struct irq_data *d, struct its_cmd_info *info)
1932 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1934 if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d))
1937 if (info->cmd_type == PROP_UPDATE_AND_INV_VLPI)
1938 lpi_update_config(d, 0xff, info->config);
1940 lpi_write_config(d, 0xff, info->config);
1941 its_vlpi_set_doorbell(d, !!(info->config & LPI_PROP_ENABLED));
1946 static int its_irq_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
1948 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1949 struct its_cmd_info *info = vcpu_info;
1952 if (!is_v4(its_dev->its))
1955 /* Unmap request? */
1957 return its_vlpi_unmap(d);
1959 switch (info->cmd_type) {
1961 return its_vlpi_map(d, info);
1964 return its_vlpi_get(d, info);
1966 case PROP_UPDATE_VLPI:
1967 case PROP_UPDATE_AND_INV_VLPI:
1968 return its_vlpi_prop_update(d, info);
1975 static struct irq_chip its_irq_chip = {
1977 .irq_mask = its_mask_irq,
1978 .irq_unmask = its_unmask_irq,
1979 .irq_eoi = irq_chip_eoi_parent,
1980 .irq_set_affinity = its_set_affinity,
1981 .irq_compose_msi_msg = its_irq_compose_msi_msg,
1982 .irq_set_irqchip_state = its_irq_set_irqchip_state,
1983 .irq_retrigger = its_irq_retrigger,
1984 .irq_set_vcpu_affinity = its_irq_set_vcpu_affinity,
1989 * How we allocate LPIs:
1991 * lpi_range_list contains ranges of LPIs that are to available to
1992 * allocate from. To allocate LPIs, just pick the first range that
1993 * fits the required allocation, and reduce it by the required
1994 * amount. Once empty, remove the range from the list.
1996 * To free a range of LPIs, add a free range to the list, sort it and
1997 * merge the result if the new range happens to be adjacent to an
1998 * already free block.
2000 * The consequence of the above is that allocation is cost is low, but
2001 * freeing is expensive. We assumes that freeing rarely occurs.
2003 #define ITS_MAX_LPI_NRBITS 16 /* 64K LPIs */
2005 static DEFINE_MUTEX(lpi_range_lock);
2006 static LIST_HEAD(lpi_range_list);
2009 struct list_head entry;
2014 static struct lpi_range *mk_lpi_range(u32 base, u32 span)
2016 struct lpi_range *range;
2018 range = kmalloc(sizeof(*range), GFP_KERNEL);
2020 range->base_id = base;
2027 static int alloc_lpi_range(u32 nr_lpis, u32 *base)
2029 struct lpi_range *range, *tmp;
2032 mutex_lock(&lpi_range_lock);
2034 list_for_each_entry_safe(range, tmp, &lpi_range_list, entry) {
2035 if (range->span >= nr_lpis) {
2036 *base = range->base_id;
2037 range->base_id += nr_lpis;
2038 range->span -= nr_lpis;
2040 if (range->span == 0) {
2041 list_del(&range->entry);
2050 mutex_unlock(&lpi_range_lock);
2052 pr_debug("ITS: alloc %u:%u\n", *base, nr_lpis);
2056 static void merge_lpi_ranges(struct lpi_range *a, struct lpi_range *b)
2058 if (&a->entry == &lpi_range_list || &b->entry == &lpi_range_list)
2060 if (a->base_id + a->span != b->base_id)
2062 b->base_id = a->base_id;
2064 list_del(&a->entry);
2068 static int free_lpi_range(u32 base, u32 nr_lpis)
2070 struct lpi_range *new, *old;
2072 new = mk_lpi_range(base, nr_lpis);
2076 mutex_lock(&lpi_range_lock);
2078 list_for_each_entry_reverse(old, &lpi_range_list, entry) {
2079 if (old->base_id < base)
2083 * old is the last element with ->base_id smaller than base,
2084 * so new goes right after it. If there are no elements with
2085 * ->base_id smaller than base, &old->entry ends up pointing
2086 * at the head of the list, and inserting new it the start of
2087 * the list is the right thing to do in that case as well.
2089 list_add(&new->entry, &old->entry);
2091 * Now check if we can merge with the preceding and/or
2094 merge_lpi_ranges(old, new);
2095 merge_lpi_ranges(new, list_next_entry(new, entry));
2097 mutex_unlock(&lpi_range_lock);
2101 static int __init its_lpi_init(u32 id_bits)
2103 u32 lpis = (1UL << id_bits) - 8192;
2107 numlpis = 1UL << GICD_TYPER_NUM_LPIS(gic_rdists->gicd_typer);
2109 if (numlpis > 2 && !WARN_ON(numlpis > lpis)) {
2111 pr_info("ITS: Using hypervisor restricted LPI range [%u]\n",
2116 * Initializing the allocator is just the same as freeing the
2117 * full range of LPIs.
2119 err = free_lpi_range(8192, lpis);
2120 pr_debug("ITS: Allocator initialized for %u LPIs\n", lpis);
2124 static unsigned long *its_lpi_alloc(int nr_irqs, u32 *base, int *nr_ids)
2126 unsigned long *bitmap = NULL;
2130 err = alloc_lpi_range(nr_irqs, base);
2135 } while (nr_irqs > 0);
2143 bitmap = bitmap_zalloc(nr_irqs, GFP_ATOMIC);
2151 *base = *nr_ids = 0;
2156 static void its_lpi_free(unsigned long *bitmap, u32 base, u32 nr_ids)
2158 WARN_ON(free_lpi_range(base, nr_ids));
2159 bitmap_free(bitmap);
2162 static void gic_reset_prop_table(void *va)
2164 /* Priority 0xa0, Group-1, disabled */
2165 memset(va, LPI_PROP_DEFAULT_PRIO | LPI_PROP_GROUP1, LPI_PROPBASE_SZ);
2167 /* Make sure the GIC will observe the written configuration */
2168 gic_flush_dcache_to_poc(va, LPI_PROPBASE_SZ);
2171 static struct page *its_allocate_prop_table(gfp_t gfp_flags)
2173 struct page *prop_page;
2175 prop_page = alloc_pages(gfp_flags, get_order(LPI_PROPBASE_SZ));
2179 gic_reset_prop_table(page_address(prop_page));
2184 static void its_free_prop_table(struct page *prop_page)
2186 free_pages((unsigned long)page_address(prop_page),
2187 get_order(LPI_PROPBASE_SZ));
2190 static bool gic_check_reserved_range(phys_addr_t addr, unsigned long size)
2192 phys_addr_t start, end, addr_end;
2196 * We don't bother checking for a kdump kernel as by
2197 * construction, the LPI tables are out of this kernel's
2200 if (is_kdump_kernel())
2203 addr_end = addr + size - 1;
2205 for_each_reserved_mem_range(i, &start, &end) {
2206 if (addr >= start && addr_end <= end)
2210 /* Not found, not a good sign... */
2211 pr_warn("GICv3: Expected reserved range [%pa:%pa], not found\n",
2213 add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
2217 static int gic_reserve_range(phys_addr_t addr, unsigned long size)
2219 if (efi_enabled(EFI_CONFIG_TABLES))
2220 return efi_mem_reserve_persistent(addr, size);
2225 static int __init its_setup_lpi_prop_table(void)
2227 if (gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED) {
2230 val = gicr_read_propbaser(gic_data_rdist_rd_base() + GICR_PROPBASER);
2231 lpi_id_bits = (val & GICR_PROPBASER_IDBITS_MASK) + 1;
2233 gic_rdists->prop_table_pa = val & GENMASK_ULL(51, 12);
2234 gic_rdists->prop_table_va = memremap(gic_rdists->prop_table_pa,
2237 gic_reset_prop_table(gic_rdists->prop_table_va);
2241 lpi_id_bits = min_t(u32,
2242 GICD_TYPER_ID_BITS(gic_rdists->gicd_typer),
2243 ITS_MAX_LPI_NRBITS);
2244 page = its_allocate_prop_table(GFP_NOWAIT);
2246 pr_err("Failed to allocate PROPBASE\n");
2250 gic_rdists->prop_table_pa = page_to_phys(page);
2251 gic_rdists->prop_table_va = page_address(page);
2252 WARN_ON(gic_reserve_range(gic_rdists->prop_table_pa,
2256 pr_info("GICv3: using LPI property table @%pa\n",
2257 &gic_rdists->prop_table_pa);
2259 return its_lpi_init(lpi_id_bits);
2262 static const char *its_base_type_string[] = {
2263 [GITS_BASER_TYPE_DEVICE] = "Devices",
2264 [GITS_BASER_TYPE_VCPU] = "Virtual CPUs",
2265 [GITS_BASER_TYPE_RESERVED3] = "Reserved (3)",
2266 [GITS_BASER_TYPE_COLLECTION] = "Interrupt Collections",
2267 [GITS_BASER_TYPE_RESERVED5] = "Reserved (5)",
2268 [GITS_BASER_TYPE_RESERVED6] = "Reserved (6)",
2269 [GITS_BASER_TYPE_RESERVED7] = "Reserved (7)",
2272 static u64 its_read_baser(struct its_node *its, struct its_baser *baser)
2274 u32 idx = baser - its->tables;
2276 return gits_read_baser(its->base + GITS_BASER + (idx << 3));
2279 static void its_write_baser(struct its_node *its, struct its_baser *baser,
2282 u32 idx = baser - its->tables;
2284 gits_write_baser(val, its->base + GITS_BASER + (idx << 3));
2285 baser->val = its_read_baser(its, baser);
2288 static int its_setup_baser(struct its_node *its, struct its_baser *baser,
2289 u64 cache, u64 shr, u32 order, bool indirect)
2291 u64 val = its_read_baser(its, baser);
2292 u64 esz = GITS_BASER_ENTRY_SIZE(val);
2293 u64 type = GITS_BASER_TYPE(val);
2294 u64 baser_phys, tmp;
2295 u32 alloc_pages, psz;
2300 alloc_pages = (PAGE_ORDER_TO_SIZE(order) / psz);
2301 if (alloc_pages > GITS_BASER_PAGES_MAX) {
2302 pr_warn("ITS@%pa: %s too large, reduce ITS pages %u->%u\n",
2303 &its->phys_base, its_base_type_string[type],
2304 alloc_pages, GITS_BASER_PAGES_MAX);
2305 alloc_pages = GITS_BASER_PAGES_MAX;
2306 order = get_order(GITS_BASER_PAGES_MAX * psz);
2309 page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO, order);
2313 base = (void *)page_address(page);
2314 baser_phys = virt_to_phys(base);
2316 /* Check if the physical address of the memory is above 48bits */
2317 if (IS_ENABLED(CONFIG_ARM64_64K_PAGES) && (baser_phys >> 48)) {
2319 /* 52bit PA is supported only when PageSize=64K */
2320 if (psz != SZ_64K) {
2321 pr_err("ITS: no 52bit PA support when psz=%d\n", psz);
2322 free_pages((unsigned long)base, order);
2326 /* Convert 52bit PA to 48bit field */
2327 baser_phys = GITS_BASER_PHYS_52_to_48(baser_phys);
2332 (type << GITS_BASER_TYPE_SHIFT) |
2333 ((esz - 1) << GITS_BASER_ENTRY_SIZE_SHIFT) |
2334 ((alloc_pages - 1) << GITS_BASER_PAGES_SHIFT) |
2339 val |= indirect ? GITS_BASER_INDIRECT : 0x0;
2343 val |= GITS_BASER_PAGE_SIZE_4K;
2346 val |= GITS_BASER_PAGE_SIZE_16K;
2349 val |= GITS_BASER_PAGE_SIZE_64K;
2353 its_write_baser(its, baser, val);
2356 if ((val ^ tmp) & GITS_BASER_SHAREABILITY_MASK) {
2358 * Shareability didn't stick. Just use
2359 * whatever the read reported, which is likely
2360 * to be the only thing this redistributor
2361 * supports. If that's zero, make it
2362 * non-cacheable as well.
2364 shr = tmp & GITS_BASER_SHAREABILITY_MASK;
2366 cache = GITS_BASER_nC;
2367 gic_flush_dcache_to_poc(base, PAGE_ORDER_TO_SIZE(order));
2373 pr_err("ITS@%pa: %s doesn't stick: %llx %llx\n",
2374 &its->phys_base, its_base_type_string[type],
2376 free_pages((unsigned long)base, order);
2380 baser->order = order;
2383 tmp = indirect ? GITS_LVL1_ENTRY_SIZE : esz;
2385 pr_info("ITS@%pa: allocated %d %s @%lx (%s, esz %d, psz %dK, shr %d)\n",
2386 &its->phys_base, (int)(PAGE_ORDER_TO_SIZE(order) / (int)tmp),
2387 its_base_type_string[type],
2388 (unsigned long)virt_to_phys(base),
2389 indirect ? "indirect" : "flat", (int)esz,
2390 psz / SZ_1K, (int)shr >> GITS_BASER_SHAREABILITY_SHIFT);
2395 static bool its_parse_indirect_baser(struct its_node *its,
2396 struct its_baser *baser,
2397 u32 *order, u32 ids)
2399 u64 tmp = its_read_baser(its, baser);
2400 u64 type = GITS_BASER_TYPE(tmp);
2401 u64 esz = GITS_BASER_ENTRY_SIZE(tmp);
2402 u64 val = GITS_BASER_InnerShareable | GITS_BASER_RaWaWb;
2403 u32 new_order = *order;
2404 u32 psz = baser->psz;
2405 bool indirect = false;
2407 /* No need to enable Indirection if memory requirement < (psz*2)bytes */
2408 if ((esz << ids) > (psz * 2)) {
2410 * Find out whether hw supports a single or two-level table by
2411 * table by reading bit at offset '62' after writing '1' to it.
2413 its_write_baser(its, baser, val | GITS_BASER_INDIRECT);
2414 indirect = !!(baser->val & GITS_BASER_INDIRECT);
2418 * The size of the lvl2 table is equal to ITS page size
2419 * which is 'psz'. For computing lvl1 table size,
2420 * subtract ID bits that sparse lvl2 table from 'ids'
2421 * which is reported by ITS hardware times lvl1 table
2424 ids -= ilog2(psz / (int)esz);
2425 esz = GITS_LVL1_ENTRY_SIZE;
2430 * Allocate as many entries as required to fit the
2431 * range of device IDs that the ITS can grok... The ID
2432 * space being incredibly sparse, this results in a
2433 * massive waste of memory if two-level device table
2434 * feature is not supported by hardware.
2436 new_order = max_t(u32, get_order(esz << ids), new_order);
2437 if (new_order >= MAX_ORDER) {
2438 new_order = MAX_ORDER - 1;
2439 ids = ilog2(PAGE_ORDER_TO_SIZE(new_order) / (int)esz);
2440 pr_warn("ITS@%pa: %s Table too large, reduce ids %llu->%u\n",
2441 &its->phys_base, its_base_type_string[type],
2442 device_ids(its), ids);
2450 static u32 compute_common_aff(u64 val)
2454 aff = FIELD_GET(GICR_TYPER_AFFINITY, val);
2455 clpiaff = FIELD_GET(GICR_TYPER_COMMON_LPI_AFF, val);
2457 return aff & ~(GENMASK(31, 0) >> (clpiaff * 8));
2460 static u32 compute_its_aff(struct its_node *its)
2466 * Reencode the ITS SVPET and MPIDR as a GICR_TYPER, and compute
2467 * the resulting affinity. We then use that to see if this match
2470 svpet = FIELD_GET(GITS_TYPER_SVPET, its->typer);
2471 val = FIELD_PREP(GICR_TYPER_COMMON_LPI_AFF, svpet);
2472 val |= FIELD_PREP(GICR_TYPER_AFFINITY, its->mpidr);
2473 return compute_common_aff(val);
2476 static struct its_node *find_sibling_its(struct its_node *cur_its)
2478 struct its_node *its;
2481 if (!FIELD_GET(GITS_TYPER_SVPET, cur_its->typer))
2484 aff = compute_its_aff(cur_its);
2486 list_for_each_entry(its, &its_nodes, entry) {
2489 if (!is_v4_1(its) || its == cur_its)
2492 if (!FIELD_GET(GITS_TYPER_SVPET, its->typer))
2495 if (aff != compute_its_aff(its))
2498 /* GICv4.1 guarantees that the vPE table is GITS_BASER2 */
2499 baser = its->tables[2].val;
2500 if (!(baser & GITS_BASER_VALID))
2509 static void its_free_tables(struct its_node *its)
2513 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
2514 if (its->tables[i].base) {
2515 free_pages((unsigned long)its->tables[i].base,
2516 its->tables[i].order);
2517 its->tables[i].base = NULL;
2522 static int its_probe_baser_psz(struct its_node *its, struct its_baser *baser)
2529 val = its_read_baser(its, baser);
2530 val &= ~GITS_BASER_PAGE_SIZE_MASK;
2534 gpsz = GITS_BASER_PAGE_SIZE_64K;
2537 gpsz = GITS_BASER_PAGE_SIZE_16K;
2541 gpsz = GITS_BASER_PAGE_SIZE_4K;
2545 gpsz >>= GITS_BASER_PAGE_SIZE_SHIFT;
2547 val |= FIELD_PREP(GITS_BASER_PAGE_SIZE_MASK, gpsz);
2548 its_write_baser(its, baser, val);
2550 if (FIELD_GET(GITS_BASER_PAGE_SIZE_MASK, baser->val) == gpsz)
2570 static int its_alloc_tables(struct its_node *its)
2572 u64 shr = GITS_BASER_InnerShareable;
2573 u64 cache = GITS_BASER_RaWaWb;
2576 if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_22375)
2577 /* erratum 24313: ignore memory access type */
2578 cache = GITS_BASER_nCnB;
2580 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
2581 struct its_baser *baser = its->tables + i;
2582 u64 val = its_read_baser(its, baser);
2583 u64 type = GITS_BASER_TYPE(val);
2584 bool indirect = false;
2587 if (type == GITS_BASER_TYPE_NONE)
2590 if (its_probe_baser_psz(its, baser)) {
2591 its_free_tables(its);
2595 order = get_order(baser->psz);
2598 case GITS_BASER_TYPE_DEVICE:
2599 indirect = its_parse_indirect_baser(its, baser, &order,
2603 case GITS_BASER_TYPE_VCPU:
2605 struct its_node *sibling;
2608 if ((sibling = find_sibling_its(its))) {
2609 *baser = sibling->tables[2];
2610 its_write_baser(its, baser, baser->val);
2615 indirect = its_parse_indirect_baser(its, baser, &order,
2616 ITS_MAX_VPEID_BITS);
2620 err = its_setup_baser(its, baser, cache, shr, order, indirect);
2622 its_free_tables(its);
2626 /* Update settings which will be used for next BASERn */
2627 cache = baser->val & GITS_BASER_CACHEABILITY_MASK;
2628 shr = baser->val & GITS_BASER_SHAREABILITY_MASK;
2634 static u64 inherit_vpe_l1_table_from_its(void)
2636 struct its_node *its;
2640 val = gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER);
2641 aff = compute_common_aff(val);
2643 list_for_each_entry(its, &its_nodes, entry) {
2649 if (!FIELD_GET(GITS_TYPER_SVPET, its->typer))
2652 if (aff != compute_its_aff(its))
2655 /* GICv4.1 guarantees that the vPE table is GITS_BASER2 */
2656 baser = its->tables[2].val;
2657 if (!(baser & GITS_BASER_VALID))
2660 /* We have a winner! */
2661 gic_data_rdist()->vpe_l1_base = its->tables[2].base;
2663 val = GICR_VPROPBASER_4_1_VALID;
2664 if (baser & GITS_BASER_INDIRECT)
2665 val |= GICR_VPROPBASER_4_1_INDIRECT;
2666 val |= FIELD_PREP(GICR_VPROPBASER_4_1_PAGE_SIZE,
2667 FIELD_GET(GITS_BASER_PAGE_SIZE_MASK, baser));
2668 switch (FIELD_GET(GITS_BASER_PAGE_SIZE_MASK, baser)) {
2669 case GIC_PAGE_SIZE_64K:
2670 addr = GITS_BASER_ADDR_48_to_52(baser);
2673 addr = baser & GENMASK_ULL(47, 12);
2676 val |= FIELD_PREP(GICR_VPROPBASER_4_1_ADDR, addr >> 12);
2677 val |= FIELD_PREP(GICR_VPROPBASER_SHAREABILITY_MASK,
2678 FIELD_GET(GITS_BASER_SHAREABILITY_MASK, baser));
2679 val |= FIELD_PREP(GICR_VPROPBASER_INNER_CACHEABILITY_MASK,
2680 FIELD_GET(GITS_BASER_INNER_CACHEABILITY_MASK, baser));
2681 val |= FIELD_PREP(GICR_VPROPBASER_4_1_SIZE, GITS_BASER_NR_PAGES(baser) - 1);
2689 static u64 inherit_vpe_l1_table_from_rd(cpumask_t **mask)
2695 val = gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER);
2696 aff = compute_common_aff(val);
2698 for_each_possible_cpu(cpu) {
2699 void __iomem *base = gic_data_rdist_cpu(cpu)->rd_base;
2701 if (!base || cpu == smp_processor_id())
2704 val = gic_read_typer(base + GICR_TYPER);
2705 if (aff != compute_common_aff(val))
2709 * At this point, we have a victim. This particular CPU
2710 * has already booted, and has an affinity that matches
2711 * ours wrt CommonLPIAff. Let's use its own VPROPBASER.
2712 * Make sure we don't write the Z bit in that case.
2714 val = gicr_read_vpropbaser(base + SZ_128K + GICR_VPROPBASER);
2715 val &= ~GICR_VPROPBASER_4_1_Z;
2717 gic_data_rdist()->vpe_l1_base = gic_data_rdist_cpu(cpu)->vpe_l1_base;
2718 *mask = gic_data_rdist_cpu(cpu)->vpe_table_mask;
2726 static bool allocate_vpe_l2_table(int cpu, u32 id)
2728 void __iomem *base = gic_data_rdist_cpu(cpu)->rd_base;
2729 unsigned int psz, esz, idx, npg, gpsz;
2734 if (!gic_rdists->has_rvpeid)
2737 /* Skip non-present CPUs */
2741 val = gicr_read_vpropbaser(base + SZ_128K + GICR_VPROPBASER);
2743 esz = FIELD_GET(GICR_VPROPBASER_4_1_ENTRY_SIZE, val) + 1;
2744 gpsz = FIELD_GET(GICR_VPROPBASER_4_1_PAGE_SIZE, val);
2745 npg = FIELD_GET(GICR_VPROPBASER_4_1_SIZE, val) + 1;
2751 case GIC_PAGE_SIZE_4K:
2754 case GIC_PAGE_SIZE_16K:
2757 case GIC_PAGE_SIZE_64K:
2762 /* Don't allow vpe_id that exceeds single, flat table limit */
2763 if (!(val & GICR_VPROPBASER_4_1_INDIRECT))
2764 return (id < (npg * psz / (esz * SZ_8)));
2766 /* Compute 1st level table index & check if that exceeds table limit */
2767 idx = id >> ilog2(psz / (esz * SZ_8));
2768 if (idx >= (npg * psz / GITS_LVL1_ENTRY_SIZE))
2771 table = gic_data_rdist_cpu(cpu)->vpe_l1_base;
2773 /* Allocate memory for 2nd level table */
2775 page = alloc_pages(GFP_KERNEL | __GFP_ZERO, get_order(psz));
2779 /* Flush Lvl2 table to PoC if hw doesn't support coherency */
2780 if (!(val & GICR_VPROPBASER_SHAREABILITY_MASK))
2781 gic_flush_dcache_to_poc(page_address(page), psz);
2783 table[idx] = cpu_to_le64(page_to_phys(page) | GITS_BASER_VALID);
2785 /* Flush Lvl1 entry to PoC if hw doesn't support coherency */
2786 if (!(val & GICR_VPROPBASER_SHAREABILITY_MASK))
2787 gic_flush_dcache_to_poc(table + idx, GITS_LVL1_ENTRY_SIZE);
2789 /* Ensure updated table contents are visible to RD hardware */
2796 static int allocate_vpe_l1_table(void)
2798 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2799 u64 val, gpsz, npg, pa;
2800 unsigned int psz = SZ_64K;
2801 unsigned int np, epp, esz;
2804 if (!gic_rdists->has_rvpeid)
2808 * if VPENDBASER.Valid is set, disable any previously programmed
2809 * VPE by setting PendingLast while clearing Valid. This has the
2810 * effect of making sure no doorbell will be generated and we can
2811 * then safely clear VPROPBASER.Valid.
2813 if (gicr_read_vpendbaser(vlpi_base + GICR_VPENDBASER) & GICR_VPENDBASER_Valid)
2814 gicr_write_vpendbaser(GICR_VPENDBASER_PendingLast,
2815 vlpi_base + GICR_VPENDBASER);
2818 * If we can inherit the configuration from another RD, let's do
2819 * so. Otherwise, we have to go through the allocation process. We
2820 * assume that all RDs have the exact same requirements, as
2821 * nothing will work otherwise.
2823 val = inherit_vpe_l1_table_from_rd(&gic_data_rdist()->vpe_table_mask);
2824 if (val & GICR_VPROPBASER_4_1_VALID)
2827 gic_data_rdist()->vpe_table_mask = kzalloc(sizeof(cpumask_t), GFP_ATOMIC);
2828 if (!gic_data_rdist()->vpe_table_mask)
2831 val = inherit_vpe_l1_table_from_its();
2832 if (val & GICR_VPROPBASER_4_1_VALID)
2835 /* First probe the page size */
2836 val = FIELD_PREP(GICR_VPROPBASER_4_1_PAGE_SIZE, GIC_PAGE_SIZE_64K);
2837 gicr_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2838 val = gicr_read_vpropbaser(vlpi_base + GICR_VPROPBASER);
2839 gpsz = FIELD_GET(GICR_VPROPBASER_4_1_PAGE_SIZE, val);
2840 esz = FIELD_GET(GICR_VPROPBASER_4_1_ENTRY_SIZE, val);
2844 gpsz = GIC_PAGE_SIZE_4K;
2846 case GIC_PAGE_SIZE_4K:
2849 case GIC_PAGE_SIZE_16K:
2852 case GIC_PAGE_SIZE_64K:
2858 * Start populating the register from scratch, including RO fields
2859 * (which we want to print in debug cases...)
2862 val |= FIELD_PREP(GICR_VPROPBASER_4_1_PAGE_SIZE, gpsz);
2863 val |= FIELD_PREP(GICR_VPROPBASER_4_1_ENTRY_SIZE, esz);
2865 /* How many entries per GIC page? */
2867 epp = psz / (esz * SZ_8);
2870 * If we need more than just a single L1 page, flag the table
2871 * as indirect and compute the number of required L1 pages.
2873 if (epp < ITS_MAX_VPEID) {
2876 val |= GICR_VPROPBASER_4_1_INDIRECT;
2878 /* Number of L2 pages required to cover the VPEID space */
2879 nl2 = DIV_ROUND_UP(ITS_MAX_VPEID, epp);
2881 /* Number of L1 pages to point to the L2 pages */
2882 npg = DIV_ROUND_UP(nl2 * SZ_8, psz);
2887 val |= FIELD_PREP(GICR_VPROPBASER_4_1_SIZE, npg - 1);
2889 /* Right, that's the number of CPU pages we need for L1 */
2890 np = DIV_ROUND_UP(npg * psz, PAGE_SIZE);
2892 pr_debug("np = %d, npg = %lld, psz = %d, epp = %d, esz = %d\n",
2893 np, npg, psz, epp, esz);
2894 page = alloc_pages(GFP_ATOMIC | __GFP_ZERO, get_order(np * PAGE_SIZE));
2898 gic_data_rdist()->vpe_l1_base = page_address(page);
2899 pa = virt_to_phys(page_address(page));
2900 WARN_ON(!IS_ALIGNED(pa, psz));
2902 val |= FIELD_PREP(GICR_VPROPBASER_4_1_ADDR, pa >> 12);
2903 val |= GICR_VPROPBASER_RaWb;
2904 val |= GICR_VPROPBASER_InnerShareable;
2905 val |= GICR_VPROPBASER_4_1_Z;
2906 val |= GICR_VPROPBASER_4_1_VALID;
2909 gicr_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2910 cpumask_set_cpu(smp_processor_id(), gic_data_rdist()->vpe_table_mask);
2912 pr_debug("CPU%d: VPROPBASER = %llx %*pbl\n",
2913 smp_processor_id(), val,
2914 cpumask_pr_args(gic_data_rdist()->vpe_table_mask));
2919 static int its_alloc_collections(struct its_node *its)
2923 its->collections = kcalloc(nr_cpu_ids, sizeof(*its->collections),
2925 if (!its->collections)
2928 for (i = 0; i < nr_cpu_ids; i++)
2929 its->collections[i].target_address = ~0ULL;
2934 static struct page *its_allocate_pending_table(gfp_t gfp_flags)
2936 struct page *pend_page;
2938 pend_page = alloc_pages(gfp_flags | __GFP_ZERO,
2939 get_order(LPI_PENDBASE_SZ));
2943 /* Make sure the GIC will observe the zero-ed page */
2944 gic_flush_dcache_to_poc(page_address(pend_page), LPI_PENDBASE_SZ);
2949 static void its_free_pending_table(struct page *pt)
2951 free_pages((unsigned long)page_address(pt), get_order(LPI_PENDBASE_SZ));
2955 * Booting with kdump and LPIs enabled is generally fine. Any other
2956 * case is wrong in the absence of firmware/EFI support.
2958 static bool enabled_lpis_allowed(void)
2963 /* Check whether the property table is in a reserved region */
2964 val = gicr_read_propbaser(gic_data_rdist_rd_base() + GICR_PROPBASER);
2965 addr = val & GENMASK_ULL(51, 12);
2967 return gic_check_reserved_range(addr, LPI_PROPBASE_SZ);
2970 static int __init allocate_lpi_tables(void)
2976 * If LPIs are enabled while we run this from the boot CPU,
2977 * flag the RD tables as pre-allocated if the stars do align.
2979 val = readl_relaxed(gic_data_rdist_rd_base() + GICR_CTLR);
2980 if ((val & GICR_CTLR_ENABLE_LPIS) && enabled_lpis_allowed()) {
2981 gic_rdists->flags |= (RDIST_FLAGS_RD_TABLES_PREALLOCATED |
2982 RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING);
2983 pr_info("GICv3: Using preallocated redistributor tables\n");
2986 err = its_setup_lpi_prop_table();
2991 * We allocate all the pending tables anyway, as we may have a
2992 * mix of RDs that have had LPIs enabled, and some that
2993 * don't. We'll free the unused ones as each CPU comes online.
2995 for_each_possible_cpu(cpu) {
2996 struct page *pend_page;
2998 pend_page = its_allocate_pending_table(GFP_NOWAIT);
3000 pr_err("Failed to allocate PENDBASE for CPU%d\n", cpu);
3004 gic_data_rdist_cpu(cpu)->pend_page = pend_page;
3010 static u64 read_vpend_dirty_clear(void __iomem *vlpi_base)
3012 u32 count = 1000000; /* 1s! */
3017 val = gicr_read_vpendbaser(vlpi_base + GICR_VPENDBASER);
3018 clean = !(val & GICR_VPENDBASER_Dirty);
3024 } while (!clean && count);
3026 if (unlikely(!clean))
3027 pr_err_ratelimited("ITS virtual pending table not cleaning\n");
3032 static u64 its_clear_vpend_valid(void __iomem *vlpi_base, u64 clr, u64 set)
3036 /* Make sure we wait until the RD is done with the initial scan */
3037 val = read_vpend_dirty_clear(vlpi_base);
3038 val &= ~GICR_VPENDBASER_Valid;
3041 gicr_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
3043 val = read_vpend_dirty_clear(vlpi_base);
3044 if (unlikely(val & GICR_VPENDBASER_Dirty))
3045 val |= GICR_VPENDBASER_PendingLast;
3050 static void its_cpu_init_lpis(void)
3052 void __iomem *rbase = gic_data_rdist_rd_base();
3053 struct page *pend_page;
3057 if (gic_data_rdist()->lpi_enabled)
3060 val = readl_relaxed(rbase + GICR_CTLR);
3061 if ((gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED) &&
3062 (val & GICR_CTLR_ENABLE_LPIS)) {
3064 * Check that we get the same property table on all
3065 * RDs. If we don't, this is hopeless.
3067 paddr = gicr_read_propbaser(rbase + GICR_PROPBASER);
3068 paddr &= GENMASK_ULL(51, 12);
3069 if (WARN_ON(gic_rdists->prop_table_pa != paddr))
3070 add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
3072 paddr = gicr_read_pendbaser(rbase + GICR_PENDBASER);
3073 paddr &= GENMASK_ULL(51, 16);
3075 WARN_ON(!gic_check_reserved_range(paddr, LPI_PENDBASE_SZ));
3076 its_free_pending_table(gic_data_rdist()->pend_page);
3077 gic_data_rdist()->pend_page = NULL;
3082 pend_page = gic_data_rdist()->pend_page;
3083 paddr = page_to_phys(pend_page);
3084 WARN_ON(gic_reserve_range(paddr, LPI_PENDBASE_SZ));
3087 val = (gic_rdists->prop_table_pa |
3088 GICR_PROPBASER_InnerShareable |
3089 GICR_PROPBASER_RaWaWb |
3090 ((LPI_NRBITS - 1) & GICR_PROPBASER_IDBITS_MASK));
3092 gicr_write_propbaser(val, rbase + GICR_PROPBASER);
3093 tmp = gicr_read_propbaser(rbase + GICR_PROPBASER);
3095 if ((tmp ^ val) & GICR_PROPBASER_SHAREABILITY_MASK) {
3096 if (!(tmp & GICR_PROPBASER_SHAREABILITY_MASK)) {
3098 * The HW reports non-shareable, we must
3099 * remove the cacheability attributes as
3102 val &= ~(GICR_PROPBASER_SHAREABILITY_MASK |
3103 GICR_PROPBASER_CACHEABILITY_MASK);
3104 val |= GICR_PROPBASER_nC;
3105 gicr_write_propbaser(val, rbase + GICR_PROPBASER);
3107 pr_info_once("GIC: using cache flushing for LPI property table\n");
3108 gic_rdists->flags |= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING;
3112 val = (page_to_phys(pend_page) |
3113 GICR_PENDBASER_InnerShareable |
3114 GICR_PENDBASER_RaWaWb);
3116 gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
3117 tmp = gicr_read_pendbaser(rbase + GICR_PENDBASER);
3119 if (!(tmp & GICR_PENDBASER_SHAREABILITY_MASK)) {
3121 * The HW reports non-shareable, we must remove the
3122 * cacheability attributes as well.
3124 val &= ~(GICR_PENDBASER_SHAREABILITY_MASK |
3125 GICR_PENDBASER_CACHEABILITY_MASK);
3126 val |= GICR_PENDBASER_nC;
3127 gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
3131 val = readl_relaxed(rbase + GICR_CTLR);
3132 val |= GICR_CTLR_ENABLE_LPIS;
3133 writel_relaxed(val, rbase + GICR_CTLR);
3135 if (gic_rdists->has_vlpis && !gic_rdists->has_rvpeid) {
3136 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
3139 * It's possible for CPU to receive VLPIs before it is
3140 * scheduled as a vPE, especially for the first CPU, and the
3141 * VLPI with INTID larger than 2^(IDbits+1) will be considered
3142 * as out of range and dropped by GIC.
3143 * So we initialize IDbits to known value to avoid VLPI drop.
3145 val = (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
3146 pr_debug("GICv4: CPU%d: Init IDbits to 0x%llx for GICR_VPROPBASER\n",
3147 smp_processor_id(), val);
3148 gicr_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
3151 * Also clear Valid bit of GICR_VPENDBASER, in case some
3152 * ancient programming gets left in and has possibility of
3153 * corrupting memory.
3155 val = its_clear_vpend_valid(vlpi_base, 0, 0);
3158 if (allocate_vpe_l1_table()) {
3160 * If the allocation has failed, we're in massive trouble.
3161 * Disable direct injection, and pray that no VM was
3162 * already running...
3164 gic_rdists->has_rvpeid = false;
3165 gic_rdists->has_vlpis = false;
3168 /* Make sure the GIC has seen the above */
3171 gic_data_rdist()->lpi_enabled = true;
3172 pr_info("GICv3: CPU%d: using %s LPI pending table @%pa\n",
3174 gic_data_rdist()->pend_page ? "allocated" : "reserved",
3178 static void its_cpu_init_collection(struct its_node *its)
3180 int cpu = smp_processor_id();
3183 /* avoid cross node collections and its mapping */
3184 if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) {
3185 struct device_node *cpu_node;
3187 cpu_node = of_get_cpu_node(cpu, NULL);
3188 if (its->numa_node != NUMA_NO_NODE &&
3189 its->numa_node != of_node_to_nid(cpu_node))
3194 * We now have to bind each collection to its target
3197 if (gic_read_typer(its->base + GITS_TYPER) & GITS_TYPER_PTA) {
3199 * This ITS wants the physical address of the
3202 target = gic_data_rdist()->phys_base;
3204 /* This ITS wants a linear CPU number. */
3205 target = gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER);
3206 target = GICR_TYPER_CPU_NUMBER(target) << 16;
3209 /* Perform collection mapping */
3210 its->collections[cpu].target_address = target;
3211 its->collections[cpu].col_id = cpu;
3213 its_send_mapc(its, &its->collections[cpu], 1);
3214 its_send_invall(its, &its->collections[cpu]);
3217 static void its_cpu_init_collections(void)
3219 struct its_node *its;
3221 raw_spin_lock(&its_lock);
3223 list_for_each_entry(its, &its_nodes, entry)
3224 its_cpu_init_collection(its);
3226 raw_spin_unlock(&its_lock);
3229 static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
3231 struct its_device *its_dev = NULL, *tmp;
3232 unsigned long flags;
3234 raw_spin_lock_irqsave(&its->lock, flags);
3236 list_for_each_entry(tmp, &its->its_device_list, entry) {
3237 if (tmp->device_id == dev_id) {
3243 raw_spin_unlock_irqrestore(&its->lock, flags);
3248 static struct its_baser *its_get_baser(struct its_node *its, u32 type)
3252 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
3253 if (GITS_BASER_TYPE(its->tables[i].val) == type)
3254 return &its->tables[i];
3260 static bool its_alloc_table_entry(struct its_node *its,
3261 struct its_baser *baser, u32 id)
3267 /* Don't allow device id that exceeds single, flat table limit */
3268 esz = GITS_BASER_ENTRY_SIZE(baser->val);
3269 if (!(baser->val & GITS_BASER_INDIRECT))
3270 return (id < (PAGE_ORDER_TO_SIZE(baser->order) / esz));
3272 /* Compute 1st level table index & check if that exceeds table limit */
3273 idx = id >> ilog2(baser->psz / esz);
3274 if (idx >= (PAGE_ORDER_TO_SIZE(baser->order) / GITS_LVL1_ENTRY_SIZE))
3277 table = baser->base;
3279 /* Allocate memory for 2nd level table */
3281 page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO,
3282 get_order(baser->psz));
3286 /* Flush Lvl2 table to PoC if hw doesn't support coherency */
3287 if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
3288 gic_flush_dcache_to_poc(page_address(page), baser->psz);
3290 table[idx] = cpu_to_le64(page_to_phys(page) | GITS_BASER_VALID);
3292 /* Flush Lvl1 entry to PoC if hw doesn't support coherency */
3293 if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
3294 gic_flush_dcache_to_poc(table + idx, GITS_LVL1_ENTRY_SIZE);
3296 /* Ensure updated table contents are visible to ITS hardware */
3303 static bool its_alloc_device_table(struct its_node *its, u32 dev_id)
3305 struct its_baser *baser;
3307 baser = its_get_baser(its, GITS_BASER_TYPE_DEVICE);
3309 /* Don't allow device id that exceeds ITS hardware limit */
3311 return (ilog2(dev_id) < device_ids(its));
3313 return its_alloc_table_entry(its, baser, dev_id);
3316 static bool its_alloc_vpe_table(u32 vpe_id)
3318 struct its_node *its;
3322 * Make sure the L2 tables are allocated on *all* v4 ITSs. We
3323 * could try and only do it on ITSs corresponding to devices
3324 * that have interrupts targeted at this VPE, but the
3325 * complexity becomes crazy (and you have tons of memory
3328 list_for_each_entry(its, &its_nodes, entry) {
3329 struct its_baser *baser;
3334 baser = its_get_baser(its, GITS_BASER_TYPE_VCPU);
3338 if (!its_alloc_table_entry(its, baser, vpe_id))
3342 /* Non v4.1? No need to iterate RDs and go back early. */
3343 if (!gic_rdists->has_rvpeid)
3347 * Make sure the L2 tables are allocated for all copies of
3348 * the L1 table on *all* v4.1 RDs.
3350 for_each_possible_cpu(cpu) {
3351 if (!allocate_vpe_l2_table(cpu, vpe_id))
3358 static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
3359 int nvecs, bool alloc_lpis)
3361 struct its_device *dev;
3362 unsigned long *lpi_map = NULL;
3363 unsigned long flags;
3364 u16 *col_map = NULL;
3371 if (!its_alloc_device_table(its, dev_id))
3374 if (WARN_ON(!is_power_of_2(nvecs)))
3375 nvecs = roundup_pow_of_two(nvecs);
3377 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
3379 * Even if the device wants a single LPI, the ITT must be
3380 * sized as a power of two (and you need at least one bit...).
3382 nr_ites = max(2, nvecs);
3383 sz = nr_ites * (FIELD_GET(GITS_TYPER_ITT_ENTRY_SIZE, its->typer) + 1);
3384 sz = max(sz, ITS_ITT_ALIGN) + ITS_ITT_ALIGN - 1;
3385 itt = kzalloc_node(sz, GFP_KERNEL, its->numa_node);
3387 lpi_map = its_lpi_alloc(nvecs, &lpi_base, &nr_lpis);
3389 col_map = kcalloc(nr_lpis, sizeof(*col_map),
3392 col_map = kcalloc(nr_ites, sizeof(*col_map), GFP_KERNEL);
3397 if (!dev || !itt || !col_map || (!lpi_map && alloc_lpis)) {
3400 bitmap_free(lpi_map);
3405 gic_flush_dcache_to_poc(itt, sz);
3409 dev->nr_ites = nr_ites;
3410 dev->event_map.lpi_map = lpi_map;
3411 dev->event_map.col_map = col_map;
3412 dev->event_map.lpi_base = lpi_base;
3413 dev->event_map.nr_lpis = nr_lpis;
3414 raw_spin_lock_init(&dev->event_map.vlpi_lock);
3415 dev->device_id = dev_id;
3416 INIT_LIST_HEAD(&dev->entry);
3418 raw_spin_lock_irqsave(&its->lock, flags);
3419 list_add(&dev->entry, &its->its_device_list);
3420 raw_spin_unlock_irqrestore(&its->lock, flags);
3422 /* Map device to its ITT */
3423 its_send_mapd(dev, 1);
3428 static void its_free_device(struct its_device *its_dev)
3430 unsigned long flags;
3432 raw_spin_lock_irqsave(&its_dev->its->lock, flags);
3433 list_del(&its_dev->entry);
3434 raw_spin_unlock_irqrestore(&its_dev->its->lock, flags);
3435 kfree(its_dev->event_map.col_map);
3436 kfree(its_dev->itt);
3440 static int its_alloc_device_irq(struct its_device *dev, int nvecs, irq_hw_number_t *hwirq)
3444 /* Find a free LPI region in lpi_map and allocate them. */
3445 idx = bitmap_find_free_region(dev->event_map.lpi_map,
3446 dev->event_map.nr_lpis,
3447 get_count_order(nvecs));
3451 *hwirq = dev->event_map.lpi_base + idx;
3456 static int its_msi_prepare(struct irq_domain *domain, struct device *dev,
3457 int nvec, msi_alloc_info_t *info)
3459 struct its_node *its;
3460 struct its_device *its_dev;
3461 struct msi_domain_info *msi_info;
3466 * We ignore "dev" entirely, and rely on the dev_id that has
3467 * been passed via the scratchpad. This limits this domain's
3468 * usefulness to upper layers that definitely know that they
3469 * are built on top of the ITS.
3471 dev_id = info->scratchpad[0].ul;
3473 msi_info = msi_get_domain_info(domain);
3474 its = msi_info->data;
3476 if (!gic_rdists->has_direct_lpi &&
3478 vpe_proxy.dev->its == its &&
3479 dev_id == vpe_proxy.dev->device_id) {
3480 /* Bad luck. Get yourself a better implementation */
3481 WARN_ONCE(1, "DevId %x clashes with GICv4 VPE proxy device\n",
3486 mutex_lock(&its->dev_alloc_lock);
3487 its_dev = its_find_device(its, dev_id);
3490 * We already have seen this ID, probably through
3491 * another alias (PCI bridge of some sort). No need to
3492 * create the device.
3494 its_dev->shared = true;
3495 pr_debug("Reusing ITT for devID %x\n", dev_id);
3499 its_dev = its_create_device(its, dev_id, nvec, true);
3505 if (info->flags & MSI_ALLOC_FLAGS_PROXY_DEVICE)
3506 its_dev->shared = true;
3508 pr_debug("ITT %d entries, %d bits\n", nvec, ilog2(nvec));
3510 mutex_unlock(&its->dev_alloc_lock);
3511 info->scratchpad[0].ptr = its_dev;
3515 static struct msi_domain_ops its_msi_domain_ops = {
3516 .msi_prepare = its_msi_prepare,
3519 static int its_irq_gic_domain_alloc(struct irq_domain *domain,
3521 irq_hw_number_t hwirq)
3523 struct irq_fwspec fwspec;
3525 if (irq_domain_get_of_node(domain->parent)) {
3526 fwspec.fwnode = domain->parent->fwnode;
3527 fwspec.param_count = 3;
3528 fwspec.param[0] = GIC_IRQ_TYPE_LPI;
3529 fwspec.param[1] = hwirq;
3530 fwspec.param[2] = IRQ_TYPE_EDGE_RISING;
3531 } else if (is_fwnode_irqchip(domain->parent->fwnode)) {
3532 fwspec.fwnode = domain->parent->fwnode;
3533 fwspec.param_count = 2;
3534 fwspec.param[0] = hwirq;
3535 fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
3540 return irq_domain_alloc_irqs_parent(domain, virq, 1, &fwspec);
3543 static int its_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
3544 unsigned int nr_irqs, void *args)
3546 msi_alloc_info_t *info = args;
3547 struct its_device *its_dev = info->scratchpad[0].ptr;
3548 struct its_node *its = its_dev->its;
3549 struct irq_data *irqd;
3550 irq_hw_number_t hwirq;
3554 err = its_alloc_device_irq(its_dev, nr_irqs, &hwirq);
3558 err = iommu_dma_prepare_msi(info->desc, its->get_msi_base(its_dev));
3562 for (i = 0; i < nr_irqs; i++) {
3563 err = its_irq_gic_domain_alloc(domain, virq + i, hwirq + i);
3567 irq_domain_set_hwirq_and_chip(domain, virq + i,
3568 hwirq + i, &its_irq_chip, its_dev);
3569 irqd = irq_get_irq_data(virq + i);
3570 irqd_set_single_target(irqd);
3571 irqd_set_affinity_on_activate(irqd);
3572 pr_debug("ID:%d pID:%d vID:%d\n",
3573 (int)(hwirq + i - its_dev->event_map.lpi_base),
3574 (int)(hwirq + i), virq + i);
3580 static int its_irq_domain_activate(struct irq_domain *domain,
3581 struct irq_data *d, bool reserve)
3583 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
3584 u32 event = its_get_event_id(d);
3587 cpu = its_select_cpu(d, cpu_online_mask);
3588 if (cpu < 0 || cpu >= nr_cpu_ids)
3591 its_inc_lpi_count(d, cpu);
3592 its_dev->event_map.col_map[event] = cpu;
3593 irq_data_update_effective_affinity(d, cpumask_of(cpu));
3595 /* Map the GIC IRQ and event to the device */
3596 its_send_mapti(its_dev, d->hwirq, event);
3600 static void its_irq_domain_deactivate(struct irq_domain *domain,
3603 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
3604 u32 event = its_get_event_id(d);
3606 its_dec_lpi_count(d, its_dev->event_map.col_map[event]);
3607 /* Stop the delivery of interrupts */
3608 its_send_discard(its_dev, event);
3611 static void its_irq_domain_free(struct irq_domain *domain, unsigned int virq,
3612 unsigned int nr_irqs)
3614 struct irq_data *d = irq_domain_get_irq_data(domain, virq);
3615 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
3616 struct its_node *its = its_dev->its;
3619 bitmap_release_region(its_dev->event_map.lpi_map,
3620 its_get_event_id(irq_domain_get_irq_data(domain, virq)),
3621 get_count_order(nr_irqs));
3623 for (i = 0; i < nr_irqs; i++) {
3624 struct irq_data *data = irq_domain_get_irq_data(domain,
3626 /* Nuke the entry in the domain */
3627 irq_domain_reset_irq_data(data);
3630 mutex_lock(&its->dev_alloc_lock);
3633 * If all interrupts have been freed, start mopping the
3634 * floor. This is conditioned on the device not being shared.
3636 if (!its_dev->shared &&
3637 bitmap_empty(its_dev->event_map.lpi_map,
3638 its_dev->event_map.nr_lpis)) {
3639 its_lpi_free(its_dev->event_map.lpi_map,
3640 its_dev->event_map.lpi_base,
3641 its_dev->event_map.nr_lpis);
3643 /* Unmap device/itt */
3644 its_send_mapd(its_dev, 0);
3645 its_free_device(its_dev);
3648 mutex_unlock(&its->dev_alloc_lock);
3650 irq_domain_free_irqs_parent(domain, virq, nr_irqs);
3653 static const struct irq_domain_ops its_domain_ops = {
3654 .alloc = its_irq_domain_alloc,
3655 .free = its_irq_domain_free,
3656 .activate = its_irq_domain_activate,
3657 .deactivate = its_irq_domain_deactivate,
3663 * If a GICv4.0 doesn't implement Direct LPIs (which is extremely
3664 * likely), the only way to perform an invalidate is to use a fake
3665 * device to issue an INV command, implying that the LPI has first
3666 * been mapped to some event on that device. Since this is not exactly
3667 * cheap, we try to keep that mapping around as long as possible, and
3668 * only issue an UNMAP if we're short on available slots.
3670 * Broken by design(tm).
3672 * GICv4.1, on the other hand, mandates that we're able to invalidate
3673 * by writing to a MMIO register. It doesn't implement the whole of
3674 * DirectLPI, but that's good enough. And most of the time, we don't
3675 * even have to invalidate anything, as the redistributor can be told
3676 * whether to generate a doorbell or not (we thus leave it enabled,
3679 static void its_vpe_db_proxy_unmap_locked(struct its_vpe *vpe)
3681 /* GICv4.1 doesn't use a proxy, so nothing to do here */
3682 if (gic_rdists->has_rvpeid)
3685 /* Already unmapped? */
3686 if (vpe->vpe_proxy_event == -1)
3689 its_send_discard(vpe_proxy.dev, vpe->vpe_proxy_event);
3690 vpe_proxy.vpes[vpe->vpe_proxy_event] = NULL;
3693 * We don't track empty slots at all, so let's move the
3694 * next_victim pointer if we can quickly reuse that slot
3695 * instead of nuking an existing entry. Not clear that this is
3696 * always a win though, and this might just generate a ripple
3697 * effect... Let's just hope VPEs don't migrate too often.
3699 if (vpe_proxy.vpes[vpe_proxy.next_victim])
3700 vpe_proxy.next_victim = vpe->vpe_proxy_event;
3702 vpe->vpe_proxy_event = -1;
3705 static void its_vpe_db_proxy_unmap(struct its_vpe *vpe)
3707 /* GICv4.1 doesn't use a proxy, so nothing to do here */
3708 if (gic_rdists->has_rvpeid)
3711 if (!gic_rdists->has_direct_lpi) {
3712 unsigned long flags;
3714 raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
3715 its_vpe_db_proxy_unmap_locked(vpe);
3716 raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
3720 static void its_vpe_db_proxy_map_locked(struct its_vpe *vpe)
3722 /* GICv4.1 doesn't use a proxy, so nothing to do here */
3723 if (gic_rdists->has_rvpeid)
3726 /* Already mapped? */
3727 if (vpe->vpe_proxy_event != -1)
3730 /* This slot was already allocated. Kick the other VPE out. */
3731 if (vpe_proxy.vpes[vpe_proxy.next_victim])
3732 its_vpe_db_proxy_unmap_locked(vpe_proxy.vpes[vpe_proxy.next_victim]);
3734 /* Map the new VPE instead */
3735 vpe_proxy.vpes[vpe_proxy.next_victim] = vpe;
3736 vpe->vpe_proxy_event = vpe_proxy.next_victim;
3737 vpe_proxy.next_victim = (vpe_proxy.next_victim + 1) % vpe_proxy.dev->nr_ites;
3739 vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = vpe->col_idx;
3740 its_send_mapti(vpe_proxy.dev, vpe->vpe_db_lpi, vpe->vpe_proxy_event);
3743 static void its_vpe_db_proxy_move(struct its_vpe *vpe, int from, int to)
3745 unsigned long flags;
3746 struct its_collection *target_col;
3748 /* GICv4.1 doesn't use a proxy, so nothing to do here */
3749 if (gic_rdists->has_rvpeid)
3752 if (gic_rdists->has_direct_lpi) {
3753 void __iomem *rdbase;
3755 rdbase = per_cpu_ptr(gic_rdists->rdist, from)->rd_base;
3756 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
3757 wait_for_syncr(rdbase);
3762 raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
3764 its_vpe_db_proxy_map_locked(vpe);
3766 target_col = &vpe_proxy.dev->its->collections[to];
3767 its_send_movi(vpe_proxy.dev, target_col, vpe->vpe_proxy_event);
3768 vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = to;
3770 raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
3773 static int its_vpe_set_affinity(struct irq_data *d,
3774 const struct cpumask *mask_val,
3777 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3778 int from, cpu = cpumask_first(mask_val);
3779 unsigned long flags;
3782 * Changing affinity is mega expensive, so let's be as lazy as
3783 * we can and only do it if we really have to. Also, if mapped
3784 * into the proxy device, we need to move the doorbell
3785 * interrupt to its new location.
3787 * Another thing is that changing the affinity of a vPE affects
3788 * *other interrupts* such as all the vLPIs that are routed to
3789 * this vPE. This means that the irq_desc lock is not enough to
3790 * protect us, and that we must ensure nobody samples vpe->col_idx
3791 * during the update, hence the lock below which must also be
3792 * taken on any vLPI handling path that evaluates vpe->col_idx.
3794 from = vpe_to_cpuid_lock(vpe, &flags);
3801 * GICv4.1 allows us to skip VMOVP if moving to a cpu whose RD
3802 * is sharing its VPE table with the current one.
3804 if (gic_data_rdist_cpu(cpu)->vpe_table_mask &&
3805 cpumask_test_cpu(from, gic_data_rdist_cpu(cpu)->vpe_table_mask))
3808 its_send_vmovp(vpe);
3809 its_vpe_db_proxy_move(vpe, from, cpu);
3812 irq_data_update_effective_affinity(d, cpumask_of(cpu));
3813 vpe_to_cpuid_unlock(vpe, flags);
3815 return IRQ_SET_MASK_OK_DONE;
3818 static void its_wait_vpt_parse_complete(void)
3820 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
3823 if (!gic_rdists->has_vpend_valid_dirty)
3826 WARN_ON_ONCE(readq_relaxed_poll_timeout_atomic(vlpi_base + GICR_VPENDBASER,
3828 !(val & GICR_VPENDBASER_Dirty),
3832 static void its_vpe_schedule(struct its_vpe *vpe)
3834 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
3837 /* Schedule the VPE */
3838 val = virt_to_phys(page_address(vpe->its_vm->vprop_page)) &
3839 GENMASK_ULL(51, 12);
3840 val |= (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
3841 val |= GICR_VPROPBASER_RaWb;
3842 val |= GICR_VPROPBASER_InnerShareable;
3843 gicr_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
3845 val = virt_to_phys(page_address(vpe->vpt_page)) &
3846 GENMASK_ULL(51, 16);
3847 val |= GICR_VPENDBASER_RaWaWb;
3848 val |= GICR_VPENDBASER_InnerShareable;
3850 * There is no good way of finding out if the pending table is
3851 * empty as we can race against the doorbell interrupt very
3852 * easily. So in the end, vpe->pending_last is only an
3853 * indication that the vcpu has something pending, not one
3854 * that the pending table is empty. A good implementation
3855 * would be able to read its coarse map pretty quickly anyway,
3856 * making this a tolerable issue.
3858 val |= GICR_VPENDBASER_PendingLast;
3859 val |= vpe->idai ? GICR_VPENDBASER_IDAI : 0;
3860 val |= GICR_VPENDBASER_Valid;
3861 gicr_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
3864 static void its_vpe_deschedule(struct its_vpe *vpe)
3866 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
3869 val = its_clear_vpend_valid(vlpi_base, 0, 0);
3871 vpe->idai = !!(val & GICR_VPENDBASER_IDAI);
3872 vpe->pending_last = !!(val & GICR_VPENDBASER_PendingLast);
3875 static void its_vpe_invall(struct its_vpe *vpe)
3877 struct its_node *its;
3879 list_for_each_entry(its, &its_nodes, entry) {
3883 if (its_list_map && !vpe->its_vm->vlpi_count[its->list_nr])
3887 * Sending a VINVALL to a single ITS is enough, as all
3888 * we need is to reach the redistributors.
3890 its_send_vinvall(its, vpe);
3895 static int its_vpe_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
3897 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3898 struct its_cmd_info *info = vcpu_info;
3900 switch (info->cmd_type) {
3902 its_vpe_schedule(vpe);
3905 case DESCHEDULE_VPE:
3906 its_vpe_deschedule(vpe);
3910 its_wait_vpt_parse_complete();
3914 its_vpe_invall(vpe);
3922 static void its_vpe_send_cmd(struct its_vpe *vpe,
3923 void (*cmd)(struct its_device *, u32))
3925 unsigned long flags;
3927 raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
3929 its_vpe_db_proxy_map_locked(vpe);
3930 cmd(vpe_proxy.dev, vpe->vpe_proxy_event);
3932 raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
3935 static void its_vpe_send_inv(struct irq_data *d)
3937 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3939 if (gic_rdists->has_direct_lpi) {
3940 void __iomem *rdbase;
3942 /* Target the redistributor this VPE is currently known on */
3943 raw_spin_lock(&gic_data_rdist_cpu(vpe->col_idx)->rd_lock);
3944 rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
3945 gic_write_lpir(d->parent_data->hwirq, rdbase + GICR_INVLPIR);
3946 wait_for_syncr(rdbase);
3947 raw_spin_unlock(&gic_data_rdist_cpu(vpe->col_idx)->rd_lock);
3949 its_vpe_send_cmd(vpe, its_send_inv);
3953 static void its_vpe_mask_irq(struct irq_data *d)
3956 * We need to unmask the LPI, which is described by the parent
3957 * irq_data. Instead of calling into the parent (which won't
3958 * exactly do the right thing, let's simply use the
3959 * parent_data pointer. Yes, I'm naughty.
3961 lpi_write_config(d->parent_data, LPI_PROP_ENABLED, 0);
3962 its_vpe_send_inv(d);
3965 static void its_vpe_unmask_irq(struct irq_data *d)
3967 /* Same hack as above... */
3968 lpi_write_config(d->parent_data, 0, LPI_PROP_ENABLED);
3969 its_vpe_send_inv(d);
3972 static int its_vpe_set_irqchip_state(struct irq_data *d,
3973 enum irqchip_irq_state which,
3976 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3978 if (which != IRQCHIP_STATE_PENDING)
3981 if (gic_rdists->has_direct_lpi) {
3982 void __iomem *rdbase;
3984 rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
3986 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_SETLPIR);
3988 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
3989 wait_for_syncr(rdbase);
3993 its_vpe_send_cmd(vpe, its_send_int);
3995 its_vpe_send_cmd(vpe, its_send_clear);
4001 static int its_vpe_retrigger(struct irq_data *d)
4003 return !its_vpe_set_irqchip_state(d, IRQCHIP_STATE_PENDING, true);
4006 static struct irq_chip its_vpe_irq_chip = {
4007 .name = "GICv4-vpe",
4008 .irq_mask = its_vpe_mask_irq,
4009 .irq_unmask = its_vpe_unmask_irq,
4010 .irq_eoi = irq_chip_eoi_parent,
4011 .irq_set_affinity = its_vpe_set_affinity,
4012 .irq_retrigger = its_vpe_retrigger,
4013 .irq_set_irqchip_state = its_vpe_set_irqchip_state,
4014 .irq_set_vcpu_affinity = its_vpe_set_vcpu_affinity,
4017 static struct its_node *find_4_1_its(void)
4019 static struct its_node *its = NULL;
4022 list_for_each_entry(its, &its_nodes, entry) {
4034 static void its_vpe_4_1_send_inv(struct irq_data *d)
4036 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4037 struct its_node *its;
4040 * GICv4.1 wants doorbells to be invalidated using the
4041 * INVDB command in order to be broadcast to all RDs. Send
4042 * it to the first valid ITS, and let the HW do its magic.
4044 its = find_4_1_its();
4046 its_send_invdb(its, vpe);
4049 static void its_vpe_4_1_mask_irq(struct irq_data *d)
4051 lpi_write_config(d->parent_data, LPI_PROP_ENABLED, 0);
4052 its_vpe_4_1_send_inv(d);
4055 static void its_vpe_4_1_unmask_irq(struct irq_data *d)
4057 lpi_write_config(d->parent_data, 0, LPI_PROP_ENABLED);
4058 its_vpe_4_1_send_inv(d);
4061 static void its_vpe_4_1_schedule(struct its_vpe *vpe,
4062 struct its_cmd_info *info)
4064 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
4067 /* Schedule the VPE */
4068 val |= GICR_VPENDBASER_Valid;
4069 val |= info->g0en ? GICR_VPENDBASER_4_1_VGRP0EN : 0;
4070 val |= info->g1en ? GICR_VPENDBASER_4_1_VGRP1EN : 0;
4071 val |= FIELD_PREP(GICR_VPENDBASER_4_1_VPEID, vpe->vpe_id);
4073 gicr_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
4076 static void its_vpe_4_1_deschedule(struct its_vpe *vpe,
4077 struct its_cmd_info *info)
4079 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
4083 unsigned long flags;
4086 * vPE is going to block: make the vPE non-resident with
4087 * PendingLast clear and DB set. The GIC guarantees that if
4088 * we read-back PendingLast clear, then a doorbell will be
4089 * delivered when an interrupt comes.
4091 * Note the locking to deal with the concurrent update of
4092 * pending_last from the doorbell interrupt handler that can
4095 raw_spin_lock_irqsave(&vpe->vpe_lock, flags);
4096 val = its_clear_vpend_valid(vlpi_base,
4097 GICR_VPENDBASER_PendingLast,
4098 GICR_VPENDBASER_4_1_DB);
4099 vpe->pending_last = !!(val & GICR_VPENDBASER_PendingLast);
4100 raw_spin_unlock_irqrestore(&vpe->vpe_lock, flags);
4103 * We're not blocking, so just make the vPE non-resident
4104 * with PendingLast set, indicating that we'll be back.
4106 val = its_clear_vpend_valid(vlpi_base,
4108 GICR_VPENDBASER_PendingLast);
4109 vpe->pending_last = true;
4113 static void its_vpe_4_1_invall(struct its_vpe *vpe)
4115 void __iomem *rdbase;
4116 unsigned long flags;
4120 val = GICR_INVALLR_V;
4121 val |= FIELD_PREP(GICR_INVALLR_VPEID, vpe->vpe_id);
4123 /* Target the redistributor this vPE is currently known on */
4124 cpu = vpe_to_cpuid_lock(vpe, &flags);
4125 raw_spin_lock(&gic_data_rdist_cpu(cpu)->rd_lock);
4126 rdbase = per_cpu_ptr(gic_rdists->rdist, cpu)->rd_base;
4127 gic_write_lpir(val, rdbase + GICR_INVALLR);
4129 wait_for_syncr(rdbase);
4130 raw_spin_unlock(&gic_data_rdist_cpu(cpu)->rd_lock);
4131 vpe_to_cpuid_unlock(vpe, flags);
4134 static int its_vpe_4_1_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
4136 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4137 struct its_cmd_info *info = vcpu_info;
4139 switch (info->cmd_type) {
4141 its_vpe_4_1_schedule(vpe, info);
4144 case DESCHEDULE_VPE:
4145 its_vpe_4_1_deschedule(vpe, info);
4149 its_wait_vpt_parse_complete();
4153 its_vpe_4_1_invall(vpe);
4161 static struct irq_chip its_vpe_4_1_irq_chip = {
4162 .name = "GICv4.1-vpe",
4163 .irq_mask = its_vpe_4_1_mask_irq,
4164 .irq_unmask = its_vpe_4_1_unmask_irq,
4165 .irq_eoi = irq_chip_eoi_parent,
4166 .irq_set_affinity = its_vpe_set_affinity,
4167 .irq_set_vcpu_affinity = its_vpe_4_1_set_vcpu_affinity,
4170 static void its_configure_sgi(struct irq_data *d, bool clear)
4172 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4173 struct its_cmd_desc desc;
4175 desc.its_vsgi_cmd.vpe = vpe;
4176 desc.its_vsgi_cmd.sgi = d->hwirq;
4177 desc.its_vsgi_cmd.priority = vpe->sgi_config[d->hwirq].priority;
4178 desc.its_vsgi_cmd.enable = vpe->sgi_config[d->hwirq].enabled;
4179 desc.its_vsgi_cmd.group = vpe->sgi_config[d->hwirq].group;
4180 desc.its_vsgi_cmd.clear = clear;
4183 * GICv4.1 allows us to send VSGI commands to any ITS as long as the
4184 * destination VPE is mapped there. Since we map them eagerly at
4185 * activation time, we're pretty sure the first GICv4.1 ITS will do.
4187 its_send_single_vcommand(find_4_1_its(), its_build_vsgi_cmd, &desc);
4190 static void its_sgi_mask_irq(struct irq_data *d)
4192 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4194 vpe->sgi_config[d->hwirq].enabled = false;
4195 its_configure_sgi(d, false);
4198 static void its_sgi_unmask_irq(struct irq_data *d)
4200 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4202 vpe->sgi_config[d->hwirq].enabled = true;
4203 its_configure_sgi(d, false);
4206 static int its_sgi_set_affinity(struct irq_data *d,
4207 const struct cpumask *mask_val,
4211 * There is no notion of affinity for virtual SGIs, at least
4212 * not on the host (since they can only be targeting a vPE).
4213 * Tell the kernel we've done whatever it asked for.
4215 irq_data_update_effective_affinity(d, mask_val);
4216 return IRQ_SET_MASK_OK;
4219 static int its_sgi_set_irqchip_state(struct irq_data *d,
4220 enum irqchip_irq_state which,
4223 if (which != IRQCHIP_STATE_PENDING)
4227 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4228 struct its_node *its = find_4_1_its();
4231 val = FIELD_PREP(GITS_SGIR_VPEID, vpe->vpe_id);
4232 val |= FIELD_PREP(GITS_SGIR_VINTID, d->hwirq);
4233 writeq_relaxed(val, its->sgir_base + GITS_SGIR - SZ_128K);
4235 its_configure_sgi(d, true);
4241 static int its_sgi_get_irqchip_state(struct irq_data *d,
4242 enum irqchip_irq_state which, bool *val)
4244 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4246 unsigned long flags;
4247 u32 count = 1000000; /* 1s! */
4251 if (which != IRQCHIP_STATE_PENDING)
4255 * Locking galore! We can race against two different events:
4257 * - Concurrent vPE affinity change: we must make sure it cannot
4258 * happen, or we'll talk to the wrong redistributor. This is
4259 * identical to what happens with vLPIs.
4261 * - Concurrent VSGIPENDR access: As it involves accessing two
4262 * MMIO registers, this must be made atomic one way or another.
4264 cpu = vpe_to_cpuid_lock(vpe, &flags);
4265 raw_spin_lock(&gic_data_rdist_cpu(cpu)->rd_lock);
4266 base = gic_data_rdist_cpu(cpu)->rd_base + SZ_128K;
4267 writel_relaxed(vpe->vpe_id, base + GICR_VSGIR);
4269 status = readl_relaxed(base + GICR_VSGIPENDR);
4270 if (!(status & GICR_VSGIPENDR_BUSY))
4275 pr_err_ratelimited("Unable to get SGI status\n");
4283 raw_spin_unlock(&gic_data_rdist_cpu(cpu)->rd_lock);
4284 vpe_to_cpuid_unlock(vpe, flags);
4289 *val = !!(status & (1 << d->hwirq));
4294 static int its_sgi_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
4296 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4297 struct its_cmd_info *info = vcpu_info;
4299 switch (info->cmd_type) {
4300 case PROP_UPDATE_VSGI:
4301 vpe->sgi_config[d->hwirq].priority = info->priority;
4302 vpe->sgi_config[d->hwirq].group = info->group;
4303 its_configure_sgi(d, false);
4311 static struct irq_chip its_sgi_irq_chip = {
4312 .name = "GICv4.1-sgi",
4313 .irq_mask = its_sgi_mask_irq,
4314 .irq_unmask = its_sgi_unmask_irq,
4315 .irq_set_affinity = its_sgi_set_affinity,
4316 .irq_set_irqchip_state = its_sgi_set_irqchip_state,
4317 .irq_get_irqchip_state = its_sgi_get_irqchip_state,
4318 .irq_set_vcpu_affinity = its_sgi_set_vcpu_affinity,
4321 static int its_sgi_irq_domain_alloc(struct irq_domain *domain,
4322 unsigned int virq, unsigned int nr_irqs,
4325 struct its_vpe *vpe = args;
4328 /* Yes, we do want 16 SGIs */
4329 WARN_ON(nr_irqs != 16);
4331 for (i = 0; i < 16; i++) {
4332 vpe->sgi_config[i].priority = 0;
4333 vpe->sgi_config[i].enabled = false;
4334 vpe->sgi_config[i].group = false;
4336 irq_domain_set_hwirq_and_chip(domain, virq + i, i,
4337 &its_sgi_irq_chip, vpe);
4338 irq_set_status_flags(virq + i, IRQ_DISABLE_UNLAZY);
4344 static void its_sgi_irq_domain_free(struct irq_domain *domain,
4346 unsigned int nr_irqs)
4351 static int its_sgi_irq_domain_activate(struct irq_domain *domain,
4352 struct irq_data *d, bool reserve)
4354 /* Write out the initial SGI configuration */
4355 its_configure_sgi(d, false);
4359 static void its_sgi_irq_domain_deactivate(struct irq_domain *domain,
4362 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4365 * The VSGI command is awkward:
4367 * - To change the configuration, CLEAR must be set to false,
4368 * leaving the pending bit unchanged.
4369 * - To clear the pending bit, CLEAR must be set to true, leaving
4370 * the configuration unchanged.
4372 * You just can't do both at once, hence the two commands below.
4374 vpe->sgi_config[d->hwirq].enabled = false;
4375 its_configure_sgi(d, false);
4376 its_configure_sgi(d, true);
4379 static const struct irq_domain_ops its_sgi_domain_ops = {
4380 .alloc = its_sgi_irq_domain_alloc,
4381 .free = its_sgi_irq_domain_free,
4382 .activate = its_sgi_irq_domain_activate,
4383 .deactivate = its_sgi_irq_domain_deactivate,
4386 static int its_vpe_id_alloc(void)
4388 return ida_simple_get(&its_vpeid_ida, 0, ITS_MAX_VPEID, GFP_KERNEL);
4391 static void its_vpe_id_free(u16 id)
4393 ida_simple_remove(&its_vpeid_ida, id);
4396 static int its_vpe_init(struct its_vpe *vpe)
4398 struct page *vpt_page;
4401 /* Allocate vpe_id */
4402 vpe_id = its_vpe_id_alloc();
4407 vpt_page = its_allocate_pending_table(GFP_KERNEL);
4409 its_vpe_id_free(vpe_id);
4413 if (!its_alloc_vpe_table(vpe_id)) {
4414 its_vpe_id_free(vpe_id);
4415 its_free_pending_table(vpt_page);
4419 raw_spin_lock_init(&vpe->vpe_lock);
4420 vpe->vpe_id = vpe_id;
4421 vpe->vpt_page = vpt_page;
4422 if (gic_rdists->has_rvpeid)
4423 atomic_set(&vpe->vmapp_count, 0);
4425 vpe->vpe_proxy_event = -1;
4430 static void its_vpe_teardown(struct its_vpe *vpe)
4432 its_vpe_db_proxy_unmap(vpe);
4433 its_vpe_id_free(vpe->vpe_id);
4434 its_free_pending_table(vpe->vpt_page);
4437 static void its_vpe_irq_domain_free(struct irq_domain *domain,
4439 unsigned int nr_irqs)
4441 struct its_vm *vm = domain->host_data;
4444 irq_domain_free_irqs_parent(domain, virq, nr_irqs);
4446 for (i = 0; i < nr_irqs; i++) {
4447 struct irq_data *data = irq_domain_get_irq_data(domain,
4449 struct its_vpe *vpe = irq_data_get_irq_chip_data(data);
4451 BUG_ON(vm != vpe->its_vm);
4453 clear_bit(data->hwirq, vm->db_bitmap);
4454 its_vpe_teardown(vpe);
4455 irq_domain_reset_irq_data(data);
4458 if (bitmap_empty(vm->db_bitmap, vm->nr_db_lpis)) {
4459 its_lpi_free(vm->db_bitmap, vm->db_lpi_base, vm->nr_db_lpis);
4460 its_free_prop_table(vm->vprop_page);
4464 static int its_vpe_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
4465 unsigned int nr_irqs, void *args)
4467 struct irq_chip *irqchip = &its_vpe_irq_chip;
4468 struct its_vm *vm = args;
4469 unsigned long *bitmap;
4470 struct page *vprop_page;
4471 int base, nr_ids, i, err = 0;
4475 bitmap = its_lpi_alloc(roundup_pow_of_two(nr_irqs), &base, &nr_ids);
4479 if (nr_ids < nr_irqs) {
4480 its_lpi_free(bitmap, base, nr_ids);
4484 vprop_page = its_allocate_prop_table(GFP_KERNEL);
4486 its_lpi_free(bitmap, base, nr_ids);
4490 vm->db_bitmap = bitmap;
4491 vm->db_lpi_base = base;
4492 vm->nr_db_lpis = nr_ids;
4493 vm->vprop_page = vprop_page;
4495 if (gic_rdists->has_rvpeid)
4496 irqchip = &its_vpe_4_1_irq_chip;
4498 for (i = 0; i < nr_irqs; i++) {
4499 vm->vpes[i]->vpe_db_lpi = base + i;
4500 err = its_vpe_init(vm->vpes[i]);
4503 err = its_irq_gic_domain_alloc(domain, virq + i,
4504 vm->vpes[i]->vpe_db_lpi);
4507 irq_domain_set_hwirq_and_chip(domain, virq + i, i,
4508 irqchip, vm->vpes[i]);
4514 its_vpe_irq_domain_free(domain, virq, i);
4516 its_lpi_free(bitmap, base, nr_ids);
4517 its_free_prop_table(vprop_page);
4523 static int its_vpe_irq_domain_activate(struct irq_domain *domain,
4524 struct irq_data *d, bool reserve)
4526 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4527 struct its_node *its;
4530 * If we use the list map, we issue VMAPP on demand... Unless
4531 * we're on a GICv4.1 and we eagerly map the VPE on all ITSs
4532 * so that VSGIs can work.
4534 if (!gic_requires_eager_mapping())
4537 /* Map the VPE to the first possible CPU */
4538 vpe->col_idx = cpumask_first(cpu_online_mask);
4540 list_for_each_entry(its, &its_nodes, entry) {
4544 its_send_vmapp(its, vpe, true);
4545 its_send_vinvall(its, vpe);
4548 irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
4553 static void its_vpe_irq_domain_deactivate(struct irq_domain *domain,
4556 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
4557 struct its_node *its;
4560 * If we use the list map on GICv4.0, we unmap the VPE once no
4561 * VLPIs are associated with the VM.
4563 if (!gic_requires_eager_mapping())
4566 list_for_each_entry(its, &its_nodes, entry) {
4570 its_send_vmapp(its, vpe, false);
4574 * There may be a direct read to the VPT after unmapping the
4575 * vPE, to guarantee the validity of this, we make the VPT
4576 * memory coherent with the CPU caches here.
4578 if (find_4_1_its() && !atomic_read(&vpe->vmapp_count))
4579 gic_flush_dcache_to_poc(page_address(vpe->vpt_page),
4583 static const struct irq_domain_ops its_vpe_domain_ops = {
4584 .alloc = its_vpe_irq_domain_alloc,
4585 .free = its_vpe_irq_domain_free,
4586 .activate = its_vpe_irq_domain_activate,
4587 .deactivate = its_vpe_irq_domain_deactivate,
4590 static int its_force_quiescent(void __iomem *base)
4592 u32 count = 1000000; /* 1s */
4595 val = readl_relaxed(base + GITS_CTLR);
4597 * GIC architecture specification requires the ITS to be both
4598 * disabled and quiescent for writes to GITS_BASER<n> or
4599 * GITS_CBASER to not have UNPREDICTABLE results.
4601 if ((val & GITS_CTLR_QUIESCENT) && !(val & GITS_CTLR_ENABLE))
4604 /* Disable the generation of all interrupts to this ITS */
4605 val &= ~(GITS_CTLR_ENABLE | GITS_CTLR_ImDe);
4606 writel_relaxed(val, base + GITS_CTLR);
4608 /* Poll GITS_CTLR and wait until ITS becomes quiescent */
4610 val = readl_relaxed(base + GITS_CTLR);
4611 if (val & GITS_CTLR_QUIESCENT)
4623 static bool __maybe_unused its_enable_quirk_cavium_22375(void *data)
4625 struct its_node *its = data;
4627 /* erratum 22375: only alloc 8MB table size (20 bits) */
4628 its->typer &= ~GITS_TYPER_DEVBITS;
4629 its->typer |= FIELD_PREP(GITS_TYPER_DEVBITS, 20 - 1);
4630 its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_22375;
4635 static bool __maybe_unused its_enable_quirk_cavium_23144(void *data)
4637 struct its_node *its = data;
4639 its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_23144;
4644 static bool __maybe_unused its_enable_quirk_qdf2400_e0065(void *data)
4646 struct its_node *its = data;
4648 /* On QDF2400, the size of the ITE is 16Bytes */
4649 its->typer &= ~GITS_TYPER_ITT_ENTRY_SIZE;
4650 its->typer |= FIELD_PREP(GITS_TYPER_ITT_ENTRY_SIZE, 16 - 1);
4655 static u64 its_irq_get_msi_base_pre_its(struct its_device *its_dev)
4657 struct its_node *its = its_dev->its;
4660 * The Socionext Synquacer SoC has a so-called 'pre-ITS',
4661 * which maps 32-bit writes targeted at a separate window of
4662 * size '4 << device_id_bits' onto writes to GITS_TRANSLATER
4663 * with device ID taken from bits [device_id_bits + 1:2] of
4664 * the window offset.
4666 return its->pre_its_base + (its_dev->device_id << 2);
4669 static bool __maybe_unused its_enable_quirk_socionext_synquacer(void *data)
4671 struct its_node *its = data;
4672 u32 pre_its_window[2];
4675 if (!fwnode_property_read_u32_array(its->fwnode_handle,
4676 "socionext,synquacer-pre-its",
4678 ARRAY_SIZE(pre_its_window))) {
4680 its->pre_its_base = pre_its_window[0];
4681 its->get_msi_base = its_irq_get_msi_base_pre_its;
4683 ids = ilog2(pre_its_window[1]) - 2;
4684 if (device_ids(its) > ids) {
4685 its->typer &= ~GITS_TYPER_DEVBITS;
4686 its->typer |= FIELD_PREP(GITS_TYPER_DEVBITS, ids - 1);
4689 /* the pre-ITS breaks isolation, so disable MSI remapping */
4690 its->msi_domain_flags &= ~IRQ_DOMAIN_FLAG_MSI_REMAP;
4696 static bool __maybe_unused its_enable_quirk_hip07_161600802(void *data)
4698 struct its_node *its = data;
4701 * Hip07 insists on using the wrong address for the VLPI
4702 * page. Trick it into doing the right thing...
4704 its->vlpi_redist_offset = SZ_128K;
4708 static const struct gic_quirk its_quirks[] = {
4709 #ifdef CONFIG_CAVIUM_ERRATUM_22375
4711 .desc = "ITS: Cavium errata 22375, 24313",
4712 .iidr = 0xa100034c, /* ThunderX pass 1.x */
4714 .init = its_enable_quirk_cavium_22375,
4717 #ifdef CONFIG_CAVIUM_ERRATUM_23144
4719 .desc = "ITS: Cavium erratum 23144",
4720 .iidr = 0xa100034c, /* ThunderX pass 1.x */
4722 .init = its_enable_quirk_cavium_23144,
4725 #ifdef CONFIG_QCOM_QDF2400_ERRATUM_0065
4727 .desc = "ITS: QDF2400 erratum 0065",
4728 .iidr = 0x00001070, /* QDF2400 ITS rev 1.x */
4730 .init = its_enable_quirk_qdf2400_e0065,
4733 #ifdef CONFIG_SOCIONEXT_SYNQUACER_PREITS
4736 * The Socionext Synquacer SoC incorporates ARM's own GIC-500
4737 * implementation, but with a 'pre-ITS' added that requires
4738 * special handling in software.
4740 .desc = "ITS: Socionext Synquacer pre-ITS",
4743 .init = its_enable_quirk_socionext_synquacer,
4746 #ifdef CONFIG_HISILICON_ERRATUM_161600802
4748 .desc = "ITS: Hip07 erratum 161600802",
4751 .init = its_enable_quirk_hip07_161600802,
4758 static void its_enable_quirks(struct its_node *its)
4760 u32 iidr = readl_relaxed(its->base + GITS_IIDR);
4762 gic_enable_quirks(iidr, its_quirks, its);
4765 static int its_save_disable(void)
4767 struct its_node *its;
4770 raw_spin_lock(&its_lock);
4771 list_for_each_entry(its, &its_nodes, entry) {
4775 its->ctlr_save = readl_relaxed(base + GITS_CTLR);
4776 err = its_force_quiescent(base);
4778 pr_err("ITS@%pa: failed to quiesce: %d\n",
4779 &its->phys_base, err);
4780 writel_relaxed(its->ctlr_save, base + GITS_CTLR);
4784 its->cbaser_save = gits_read_cbaser(base + GITS_CBASER);
4789 list_for_each_entry_continue_reverse(its, &its_nodes, entry) {
4793 writel_relaxed(its->ctlr_save, base + GITS_CTLR);
4796 raw_spin_unlock(&its_lock);
4801 static void its_restore_enable(void)
4803 struct its_node *its;
4806 raw_spin_lock(&its_lock);
4807 list_for_each_entry(its, &its_nodes, entry) {
4814 * Make sure that the ITS is disabled. If it fails to quiesce,
4815 * don't restore it since writing to CBASER or BASER<n>
4816 * registers is undefined according to the GIC v3 ITS
4819 * Firmware resuming with the ITS enabled is terminally broken.
4821 WARN_ON(readl_relaxed(base + GITS_CTLR) & GITS_CTLR_ENABLE);
4822 ret = its_force_quiescent(base);
4824 pr_err("ITS@%pa: failed to quiesce on resume: %d\n",
4825 &its->phys_base, ret);
4829 gits_write_cbaser(its->cbaser_save, base + GITS_CBASER);
4832 * Writing CBASER resets CREADR to 0, so make CWRITER and
4833 * cmd_write line up with it.
4835 its->cmd_write = its->cmd_base;
4836 gits_write_cwriter(0, base + GITS_CWRITER);
4838 /* Restore GITS_BASER from the value cache. */
4839 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
4840 struct its_baser *baser = &its->tables[i];
4842 if (!(baser->val & GITS_BASER_VALID))
4845 its_write_baser(its, baser, baser->val);
4847 writel_relaxed(its->ctlr_save, base + GITS_CTLR);
4850 * Reinit the collection if it's stored in the ITS. This is
4851 * indicated by the col_id being less than the HCC field.
4852 * CID < HCC as specified in the GIC v3 Documentation.
4854 if (its->collections[smp_processor_id()].col_id <
4855 GITS_TYPER_HCC(gic_read_typer(base + GITS_TYPER)))
4856 its_cpu_init_collection(its);
4858 raw_spin_unlock(&its_lock);
4861 static struct syscore_ops its_syscore_ops = {
4862 .suspend = its_save_disable,
4863 .resume = its_restore_enable,
4866 static int its_init_domain(struct fwnode_handle *handle, struct its_node *its)
4868 struct irq_domain *inner_domain;
4869 struct msi_domain_info *info;
4871 info = kzalloc(sizeof(*info), GFP_KERNEL);
4875 inner_domain = irq_domain_create_tree(handle, &its_domain_ops, its);
4876 if (!inner_domain) {
4881 inner_domain->parent = its_parent;
4882 irq_domain_update_bus_token(inner_domain, DOMAIN_BUS_NEXUS);
4883 inner_domain->flags |= its->msi_domain_flags;
4884 info->ops = &its_msi_domain_ops;
4886 inner_domain->host_data = info;
4891 static int its_init_vpe_domain(void)
4893 struct its_node *its;
4897 if (gic_rdists->has_direct_lpi) {
4898 pr_info("ITS: Using DirectLPI for VPE invalidation\n");
4902 /* Any ITS will do, even if not v4 */
4903 its = list_first_entry(&its_nodes, struct its_node, entry);
4905 entries = roundup_pow_of_two(nr_cpu_ids);
4906 vpe_proxy.vpes = kcalloc(entries, sizeof(*vpe_proxy.vpes),
4908 if (!vpe_proxy.vpes)
4911 /* Use the last possible DevID */
4912 devid = GENMASK(device_ids(its) - 1, 0);
4913 vpe_proxy.dev = its_create_device(its, devid, entries, false);
4914 if (!vpe_proxy.dev) {
4915 kfree(vpe_proxy.vpes);
4916 pr_err("ITS: Can't allocate GICv4 proxy device\n");
4920 BUG_ON(entries > vpe_proxy.dev->nr_ites);
4922 raw_spin_lock_init(&vpe_proxy.lock);
4923 vpe_proxy.next_victim = 0;
4924 pr_info("ITS: Allocated DevID %x as GICv4 proxy device (%d slots)\n",
4925 devid, vpe_proxy.dev->nr_ites);
4930 static int __init its_compute_its_list_map(struct resource *res,
4931 void __iomem *its_base)
4937 * This is assumed to be done early enough that we're
4938 * guaranteed to be single-threaded, hence no
4939 * locking. Should this change, we should address
4942 its_number = find_first_zero_bit(&its_list_map, GICv4_ITS_LIST_MAX);
4943 if (its_number >= GICv4_ITS_LIST_MAX) {
4944 pr_err("ITS@%pa: No ITSList entry available!\n",
4949 ctlr = readl_relaxed(its_base + GITS_CTLR);
4950 ctlr &= ~GITS_CTLR_ITS_NUMBER;
4951 ctlr |= its_number << GITS_CTLR_ITS_NUMBER_SHIFT;
4952 writel_relaxed(ctlr, its_base + GITS_CTLR);
4953 ctlr = readl_relaxed(its_base + GITS_CTLR);
4954 if ((ctlr & GITS_CTLR_ITS_NUMBER) != (its_number << GITS_CTLR_ITS_NUMBER_SHIFT)) {
4955 its_number = ctlr & GITS_CTLR_ITS_NUMBER;
4956 its_number >>= GITS_CTLR_ITS_NUMBER_SHIFT;
4959 if (test_and_set_bit(its_number, &its_list_map)) {
4960 pr_err("ITS@%pa: Duplicate ITSList entry %d\n",
4961 &res->start, its_number);
4968 static int __init its_probe_one(struct resource *res,
4969 struct fwnode_handle *handle, int numa_node)
4971 struct its_node *its;
4972 void __iomem *its_base;
4974 u64 baser, tmp, typer;
4978 its_base = ioremap(res->start, SZ_64K);
4980 pr_warn("ITS@%pa: Unable to map ITS registers\n", &res->start);
4984 val = readl_relaxed(its_base + GITS_PIDR2) & GIC_PIDR2_ARCH_MASK;
4985 if (val != 0x30 && val != 0x40) {
4986 pr_warn("ITS@%pa: No ITS detected, giving up\n", &res->start);
4991 err = its_force_quiescent(its_base);
4993 pr_warn("ITS@%pa: Failed to quiesce, giving up\n", &res->start);
4997 pr_info("ITS %pR\n", res);
4999 its = kzalloc(sizeof(*its), GFP_KERNEL);
5005 raw_spin_lock_init(&its->lock);
5006 mutex_init(&its->dev_alloc_lock);
5007 INIT_LIST_HEAD(&its->entry);
5008 INIT_LIST_HEAD(&its->its_device_list);
5009 typer = gic_read_typer(its_base + GITS_TYPER);
5011 its->base = its_base;
5012 its->phys_base = res->start;
5014 if (!(typer & GITS_TYPER_VMOVP)) {
5015 err = its_compute_its_list_map(res, its_base);
5021 pr_info("ITS@%pa: Using ITS number %d\n",
5024 pr_info("ITS@%pa: Single VMOVP capable\n", &res->start);
5028 u32 svpet = FIELD_GET(GITS_TYPER_SVPET, typer);
5030 its->sgir_base = ioremap(res->start + SZ_128K, SZ_64K);
5031 if (!its->sgir_base) {
5036 its->mpidr = readl_relaxed(its_base + GITS_MPIDR);
5038 pr_info("ITS@%pa: Using GICv4.1 mode %08x %08x\n",
5039 &res->start, its->mpidr, svpet);
5043 its->numa_node = numa_node;
5045 page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO,
5046 get_order(ITS_CMD_QUEUE_SZ));
5049 goto out_unmap_sgir;
5051 its->cmd_base = (void *)page_address(page);
5052 its->cmd_write = its->cmd_base;
5053 its->fwnode_handle = handle;
5054 its->get_msi_base = its_irq_get_msi_base;
5055 its->msi_domain_flags = IRQ_DOMAIN_FLAG_MSI_REMAP;
5057 its_enable_quirks(its);
5059 err = its_alloc_tables(its);
5063 err = its_alloc_collections(its);
5065 goto out_free_tables;
5067 baser = (virt_to_phys(its->cmd_base) |
5068 GITS_CBASER_RaWaWb |
5069 GITS_CBASER_InnerShareable |
5070 (ITS_CMD_QUEUE_SZ / SZ_4K - 1) |
5073 gits_write_cbaser(baser, its->base + GITS_CBASER);
5074 tmp = gits_read_cbaser(its->base + GITS_CBASER);
5076 if ((tmp ^ baser) & GITS_CBASER_SHAREABILITY_MASK) {
5077 if (!(tmp & GITS_CBASER_SHAREABILITY_MASK)) {
5079 * The HW reports non-shareable, we must
5080 * remove the cacheability attributes as
5083 baser &= ~(GITS_CBASER_SHAREABILITY_MASK |
5084 GITS_CBASER_CACHEABILITY_MASK);
5085 baser |= GITS_CBASER_nC;
5086 gits_write_cbaser(baser, its->base + GITS_CBASER);
5088 pr_info("ITS: using cache flushing for cmd queue\n");
5089 its->flags |= ITS_FLAGS_CMDQ_NEEDS_FLUSHING;
5092 gits_write_cwriter(0, its->base + GITS_CWRITER);
5093 ctlr = readl_relaxed(its->base + GITS_CTLR);
5094 ctlr |= GITS_CTLR_ENABLE;
5096 ctlr |= GITS_CTLR_ImDe;
5097 writel_relaxed(ctlr, its->base + GITS_CTLR);
5099 err = its_init_domain(handle, its);
5101 goto out_free_tables;
5103 raw_spin_lock(&its_lock);
5104 list_add(&its->entry, &its_nodes);
5105 raw_spin_unlock(&its_lock);
5110 its_free_tables(its);
5112 free_pages((unsigned long)its->cmd_base, get_order(ITS_CMD_QUEUE_SZ));
5115 iounmap(its->sgir_base);
5120 pr_err("ITS@%pa: failed probing (%d)\n", &res->start, err);
5124 static bool gic_rdists_supports_plpis(void)
5126 return !!(gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER) & GICR_TYPER_PLPIS);
5129 static int redist_disable_lpis(void)
5131 void __iomem *rbase = gic_data_rdist_rd_base();
5132 u64 timeout = USEC_PER_SEC;
5135 if (!gic_rdists_supports_plpis()) {
5136 pr_info("CPU%d: LPIs not supported\n", smp_processor_id());
5140 val = readl_relaxed(rbase + GICR_CTLR);
5141 if (!(val & GICR_CTLR_ENABLE_LPIS))
5145 * If coming via a CPU hotplug event, we don't need to disable
5146 * LPIs before trying to re-enable them. They are already
5147 * configured and all is well in the world.
5149 * If running with preallocated tables, there is nothing to do.
5151 if (gic_data_rdist()->lpi_enabled ||
5152 (gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED))
5156 * From that point on, we only try to do some damage control.
5158 pr_warn("GICv3: CPU%d: Booted with LPIs enabled, memory probably corrupted\n",
5159 smp_processor_id());
5160 add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
5163 val &= ~GICR_CTLR_ENABLE_LPIS;
5164 writel_relaxed(val, rbase + GICR_CTLR);
5166 /* Make sure any change to GICR_CTLR is observable by the GIC */
5170 * Software must observe RWP==0 after clearing GICR_CTLR.EnableLPIs
5171 * from 1 to 0 before programming GICR_PEND{PROP}BASER registers.
5172 * Error out if we time out waiting for RWP to clear.
5174 while (readl_relaxed(rbase + GICR_CTLR) & GICR_CTLR_RWP) {
5176 pr_err("CPU%d: Timeout while disabling LPIs\n",
5177 smp_processor_id());
5185 * After it has been written to 1, it is IMPLEMENTATION
5186 * DEFINED whether GICR_CTLR.EnableLPI becomes RES1 or can be
5187 * cleared to 0. Error out if clearing the bit failed.
5189 if (readl_relaxed(rbase + GICR_CTLR) & GICR_CTLR_ENABLE_LPIS) {
5190 pr_err("CPU%d: Failed to disable LPIs\n", smp_processor_id());
5197 int its_cpu_init(void)
5199 if (!list_empty(&its_nodes)) {
5202 ret = redist_disable_lpis();
5206 its_cpu_init_lpis();
5207 its_cpu_init_collections();
5213 static const struct of_device_id its_device_id[] = {
5214 { .compatible = "arm,gic-v3-its", },
5218 static int __init its_of_probe(struct device_node *node)
5220 struct device_node *np;
5221 struct resource res;
5223 for (np = of_find_matching_node(node, its_device_id); np;
5224 np = of_find_matching_node(np, its_device_id)) {
5225 if (!of_device_is_available(np))
5227 if (!of_property_read_bool(np, "msi-controller")) {
5228 pr_warn("%pOF: no msi-controller property, ITS ignored\n",
5233 if (of_address_to_resource(np, 0, &res)) {
5234 pr_warn("%pOF: no regs?\n", np);
5238 its_probe_one(&res, &np->fwnode, of_node_to_nid(np));
5245 #define ACPI_GICV3_ITS_MEM_SIZE (SZ_128K)
5247 #ifdef CONFIG_ACPI_NUMA
5248 struct its_srat_map {
5255 static struct its_srat_map *its_srat_maps __initdata;
5256 static int its_in_srat __initdata;
5258 static int __init acpi_get_its_numa_node(u32 its_id)
5262 for (i = 0; i < its_in_srat; i++) {
5263 if (its_id == its_srat_maps[i].its_id)
5264 return its_srat_maps[i].numa_node;
5266 return NUMA_NO_NODE;
5269 static int __init gic_acpi_match_srat_its(union acpi_subtable_headers *header,
5270 const unsigned long end)
5275 static int __init gic_acpi_parse_srat_its(union acpi_subtable_headers *header,
5276 const unsigned long end)
5279 struct acpi_srat_gic_its_affinity *its_affinity;
5281 its_affinity = (struct acpi_srat_gic_its_affinity *)header;
5285 if (its_affinity->header.length < sizeof(*its_affinity)) {
5286 pr_err("SRAT: Invalid header length %d in ITS affinity\n",
5287 its_affinity->header.length);
5292 * Note that in theory a new proximity node could be created by this
5293 * entry as it is an SRAT resource allocation structure.
5294 * We do not currently support doing so.
5296 node = pxm_to_node(its_affinity->proximity_domain);
5298 if (node == NUMA_NO_NODE || node >= MAX_NUMNODES) {
5299 pr_err("SRAT: Invalid NUMA node %d in ITS affinity\n", node);
5303 its_srat_maps[its_in_srat].numa_node = node;
5304 its_srat_maps[its_in_srat].its_id = its_affinity->its_id;
5306 pr_info("SRAT: PXM %d -> ITS %d -> Node %d\n",
5307 its_affinity->proximity_domain, its_affinity->its_id, node);
5312 static void __init acpi_table_parse_srat_its(void)
5316 count = acpi_table_parse_entries(ACPI_SIG_SRAT,
5317 sizeof(struct acpi_table_srat),
5318 ACPI_SRAT_TYPE_GIC_ITS_AFFINITY,
5319 gic_acpi_match_srat_its, 0);
5323 its_srat_maps = kmalloc_array(count, sizeof(struct its_srat_map),
5328 acpi_table_parse_entries(ACPI_SIG_SRAT,
5329 sizeof(struct acpi_table_srat),
5330 ACPI_SRAT_TYPE_GIC_ITS_AFFINITY,
5331 gic_acpi_parse_srat_its, 0);
5334 /* free the its_srat_maps after ITS probing */
5335 static void __init acpi_its_srat_maps_free(void)
5337 kfree(its_srat_maps);
5340 static void __init acpi_table_parse_srat_its(void) { }
5341 static int __init acpi_get_its_numa_node(u32 its_id) { return NUMA_NO_NODE; }
5342 static void __init acpi_its_srat_maps_free(void) { }
5345 static int __init gic_acpi_parse_madt_its(union acpi_subtable_headers *header,
5346 const unsigned long end)
5348 struct acpi_madt_generic_translator *its_entry;
5349 struct fwnode_handle *dom_handle;
5350 struct resource res;
5353 its_entry = (struct acpi_madt_generic_translator *)header;
5354 memset(&res, 0, sizeof(res));
5355 res.start = its_entry->base_address;
5356 res.end = its_entry->base_address + ACPI_GICV3_ITS_MEM_SIZE - 1;
5357 res.flags = IORESOURCE_MEM;
5359 dom_handle = irq_domain_alloc_fwnode(&res.start);
5361 pr_err("ITS@%pa: Unable to allocate GICv3 ITS domain token\n",
5366 err = iort_register_domain_token(its_entry->translation_id, res.start,
5369 pr_err("ITS@%pa: Unable to register GICv3 ITS domain token (ITS ID %d) to IORT\n",
5370 &res.start, its_entry->translation_id);
5374 err = its_probe_one(&res, dom_handle,
5375 acpi_get_its_numa_node(its_entry->translation_id));
5379 iort_deregister_domain_token(its_entry->translation_id);
5381 irq_domain_free_fwnode(dom_handle);
5385 static void __init its_acpi_probe(void)
5387 acpi_table_parse_srat_its();
5388 acpi_table_parse_madt(ACPI_MADT_TYPE_GENERIC_TRANSLATOR,
5389 gic_acpi_parse_madt_its, 0);
5390 acpi_its_srat_maps_free();
5393 static void __init its_acpi_probe(void) { }
5396 int __init its_init(struct fwnode_handle *handle, struct rdists *rdists,
5397 struct irq_domain *parent_domain)
5399 struct device_node *of_node;
5400 struct its_node *its;
5401 bool has_v4 = false;
5402 bool has_v4_1 = false;
5405 gic_rdists = rdists;
5407 its_parent = parent_domain;
5408 of_node = to_of_node(handle);
5410 its_of_probe(of_node);
5414 if (list_empty(&its_nodes)) {
5415 pr_warn("ITS: No ITS available, not enabling LPIs\n");
5419 err = allocate_lpi_tables();
5423 list_for_each_entry(its, &its_nodes, entry) {
5424 has_v4 |= is_v4(its);
5425 has_v4_1 |= is_v4_1(its);
5428 /* Don't bother with inconsistent systems */
5429 if (WARN_ON(!has_v4_1 && rdists->has_rvpeid))
5430 rdists->has_rvpeid = false;
5432 if (has_v4 & rdists->has_vlpis) {
5433 const struct irq_domain_ops *sgi_ops;
5436 sgi_ops = &its_sgi_domain_ops;
5440 if (its_init_vpe_domain() ||
5441 its_init_v4(parent_domain, &its_vpe_domain_ops, sgi_ops)) {
5442 rdists->has_vlpis = false;
5443 pr_err("ITS: Disabling GICv4 support\n");
5447 register_syscore_ops(&its_syscore_ops);