2 * Copyright (C) 2013-2017 ARM Limited, All Rights Reserved.
3 * Author: Marc Zyngier <marc.zyngier@arm.com>
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program. If not, see <http://www.gnu.org/licenses/>.
18 #include <linux/acpi.h>
19 #include <linux/acpi_iort.h>
20 #include <linux/bitmap.h>
21 #include <linux/cpu.h>
22 #include <linux/delay.h>
23 #include <linux/dma-iommu.h>
24 #include <linux/interrupt.h>
25 #include <linux/irqdomain.h>
26 #include <linux/list.h>
27 #include <linux/list_sort.h>
28 #include <linux/log2.h>
30 #include <linux/msi.h>
32 #include <linux/of_address.h>
33 #include <linux/of_irq.h>
34 #include <linux/of_pci.h>
35 #include <linux/of_platform.h>
36 #include <linux/percpu.h>
37 #include <linux/slab.h>
38 #include <linux/syscore_ops.h>
40 #include <linux/irqchip.h>
41 #include <linux/irqchip/arm-gic-v3.h>
42 #include <linux/irqchip/arm-gic-v4.h>
44 #include <asm/cputype.h>
45 #include <asm/exception.h>
47 #include "irq-gic-common.h"
49 #define ITS_FLAGS_CMDQ_NEEDS_FLUSHING (1ULL << 0)
50 #define ITS_FLAGS_WORKAROUND_CAVIUM_22375 (1ULL << 1)
51 #define ITS_FLAGS_WORKAROUND_CAVIUM_23144 (1ULL << 2)
52 #define ITS_FLAGS_SAVE_SUSPEND_STATE (1ULL << 3)
54 #define RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING (1 << 0)
56 static u32 lpi_id_bits;
59 * We allocate memory for PROPBASE to cover 2 ^ lpi_id_bits LPIs to
60 * deal with (one configuration byte per interrupt). PENDBASE has to
61 * be 64kB aligned (one bit per LPI, plus 8192 bits for SPI/PPI/SGI).
63 #define LPI_NRBITS lpi_id_bits
64 #define LPI_PROPBASE_SZ ALIGN(BIT(LPI_NRBITS), SZ_64K)
65 #define LPI_PENDBASE_SZ ALIGN(BIT(LPI_NRBITS) / 8, SZ_64K)
67 #define LPI_PROP_DEFAULT_PRIO 0xa0
70 * Collection structure - just an ID, and a redistributor address to
71 * ping. We use one per CPU as a bag of interrupts assigned to this
74 struct its_collection {
80 * The ITS_BASER structure - contains memory information, cached
81 * value of BASER register configuration and ITS page size.
93 * The ITS structure - contains most of the infrastructure, with the
94 * top-level MSI domain, the command queue, the collections, and the
95 * list of devices writing to it.
99 struct list_head entry;
101 phys_addr_t phys_base;
102 struct its_cmd_block *cmd_base;
103 struct its_cmd_block *cmd_write;
104 struct its_baser tables[GITS_BASER_NR_REGS];
105 struct its_collection *collections;
106 struct fwnode_handle *fwnode_handle;
107 u64 (*get_msi_base)(struct its_device *its_dev);
110 struct list_head its_device_list;
112 unsigned long list_nr;
116 unsigned int msi_domain_flags;
117 u32 pre_its_base; /* for Socionext Synquacer */
119 int vlpi_redist_offset;
122 #define ITS_ITT_ALIGN SZ_256
124 /* The maximum number of VPEID bits supported by VLPI commands */
125 #define ITS_MAX_VPEID_BITS (16)
126 #define ITS_MAX_VPEID (1 << (ITS_MAX_VPEID_BITS))
128 /* Convert page order to size in bytes */
129 #define PAGE_ORDER_TO_SIZE(o) (PAGE_SIZE << (o))
131 struct event_lpi_map {
132 unsigned long *lpi_map;
134 irq_hw_number_t lpi_base;
136 struct mutex vlpi_lock;
138 struct its_vlpi_map *vlpi_maps;
143 * The ITS view of a device - belongs to an ITS, owns an interrupt
144 * translation table, and a list of interrupts. If it some of its
145 * LPIs are injected into a guest (GICv4), the event_map.vm field
146 * indicates which one.
149 struct list_head entry;
150 struct its_node *its;
151 struct event_lpi_map event_map;
159 struct its_device *dev;
160 struct its_vpe **vpes;
164 static LIST_HEAD(its_nodes);
165 static DEFINE_RAW_SPINLOCK(its_lock);
166 static struct rdists *gic_rdists;
167 static struct irq_domain *its_parent;
169 static unsigned long its_list_map;
170 static u16 vmovp_seq_num;
171 static DEFINE_RAW_SPINLOCK(vmovp_lock);
173 static DEFINE_IDA(its_vpeid_ida);
175 #define gic_data_rdist() (raw_cpu_ptr(gic_rdists->rdist))
176 #define gic_data_rdist_rd_base() (gic_data_rdist()->rd_base)
177 #define gic_data_rdist_vlpi_base() (gic_data_rdist_rd_base() + SZ_128K)
179 static struct its_collection *dev_event_to_col(struct its_device *its_dev,
182 struct its_node *its = its_dev->its;
184 return its->collections + its_dev->event_map.col_map[event];
187 static struct its_collection *valid_col(struct its_collection *col)
189 if (WARN_ON_ONCE(col->target_address & GENMASK_ULL(0, 15)))
195 static struct its_vpe *valid_vpe(struct its_node *its, struct its_vpe *vpe)
197 if (valid_col(its->collections + vpe->col_idx))
204 * ITS command descriptors - parameters to be encoded in a command
207 struct its_cmd_desc {
210 struct its_device *dev;
215 struct its_device *dev;
220 struct its_device *dev;
225 struct its_device *dev;
230 struct its_collection *col;
235 struct its_device *dev;
241 struct its_device *dev;
242 struct its_collection *col;
247 struct its_device *dev;
252 struct its_collection *col;
261 struct its_collection *col;
267 struct its_device *dev;
275 struct its_device *dev;
282 struct its_collection *col;
290 * The ITS command block, which is what the ITS actually parses.
292 struct its_cmd_block {
296 #define ITS_CMD_QUEUE_SZ SZ_64K
297 #define ITS_CMD_QUEUE_NR_ENTRIES (ITS_CMD_QUEUE_SZ / sizeof(struct its_cmd_block))
299 typedef struct its_collection *(*its_cmd_builder_t)(struct its_node *,
300 struct its_cmd_block *,
301 struct its_cmd_desc *);
303 typedef struct its_vpe *(*its_cmd_vbuilder_t)(struct its_node *,
304 struct its_cmd_block *,
305 struct its_cmd_desc *);
307 static void its_mask_encode(u64 *raw_cmd, u64 val, int h, int l)
309 u64 mask = GENMASK_ULL(h, l);
311 *raw_cmd |= (val << l) & mask;
314 static void its_encode_cmd(struct its_cmd_block *cmd, u8 cmd_nr)
316 its_mask_encode(&cmd->raw_cmd[0], cmd_nr, 7, 0);
319 static void its_encode_devid(struct its_cmd_block *cmd, u32 devid)
321 its_mask_encode(&cmd->raw_cmd[0], devid, 63, 32);
324 static void its_encode_event_id(struct its_cmd_block *cmd, u32 id)
326 its_mask_encode(&cmd->raw_cmd[1], id, 31, 0);
329 static void its_encode_phys_id(struct its_cmd_block *cmd, u32 phys_id)
331 its_mask_encode(&cmd->raw_cmd[1], phys_id, 63, 32);
334 static void its_encode_size(struct its_cmd_block *cmd, u8 size)
336 its_mask_encode(&cmd->raw_cmd[1], size, 4, 0);
339 static void its_encode_itt(struct its_cmd_block *cmd, u64 itt_addr)
341 its_mask_encode(&cmd->raw_cmd[2], itt_addr >> 8, 51, 8);
344 static void its_encode_valid(struct its_cmd_block *cmd, int valid)
346 its_mask_encode(&cmd->raw_cmd[2], !!valid, 63, 63);
349 static void its_encode_target(struct its_cmd_block *cmd, u64 target_addr)
351 its_mask_encode(&cmd->raw_cmd[2], target_addr >> 16, 51, 16);
354 static void its_encode_collection(struct its_cmd_block *cmd, u16 col)
356 its_mask_encode(&cmd->raw_cmd[2], col, 15, 0);
359 static void its_encode_vpeid(struct its_cmd_block *cmd, u16 vpeid)
361 its_mask_encode(&cmd->raw_cmd[1], vpeid, 47, 32);
364 static void its_encode_virt_id(struct its_cmd_block *cmd, u32 virt_id)
366 its_mask_encode(&cmd->raw_cmd[2], virt_id, 31, 0);
369 static void its_encode_db_phys_id(struct its_cmd_block *cmd, u32 db_phys_id)
371 its_mask_encode(&cmd->raw_cmd[2], db_phys_id, 63, 32);
374 static void its_encode_db_valid(struct its_cmd_block *cmd, bool db_valid)
376 its_mask_encode(&cmd->raw_cmd[2], db_valid, 0, 0);
379 static void its_encode_seq_num(struct its_cmd_block *cmd, u16 seq_num)
381 its_mask_encode(&cmd->raw_cmd[0], seq_num, 47, 32);
384 static void its_encode_its_list(struct its_cmd_block *cmd, u16 its_list)
386 its_mask_encode(&cmd->raw_cmd[1], its_list, 15, 0);
389 static void its_encode_vpt_addr(struct its_cmd_block *cmd, u64 vpt_pa)
391 its_mask_encode(&cmd->raw_cmd[3], vpt_pa >> 16, 51, 16);
394 static void its_encode_vpt_size(struct its_cmd_block *cmd, u8 vpt_size)
396 its_mask_encode(&cmd->raw_cmd[3], vpt_size, 4, 0);
399 static inline void its_fixup_cmd(struct its_cmd_block *cmd)
401 /* Let's fixup BE commands */
402 cmd->raw_cmd[0] = cpu_to_le64(cmd->raw_cmd[0]);
403 cmd->raw_cmd[1] = cpu_to_le64(cmd->raw_cmd[1]);
404 cmd->raw_cmd[2] = cpu_to_le64(cmd->raw_cmd[2]);
405 cmd->raw_cmd[3] = cpu_to_le64(cmd->raw_cmd[3]);
408 static struct its_collection *its_build_mapd_cmd(struct its_node *its,
409 struct its_cmd_block *cmd,
410 struct its_cmd_desc *desc)
412 unsigned long itt_addr;
413 u8 size = ilog2(desc->its_mapd_cmd.dev->nr_ites);
415 itt_addr = virt_to_phys(desc->its_mapd_cmd.dev->itt);
416 itt_addr = ALIGN(itt_addr, ITS_ITT_ALIGN);
418 its_encode_cmd(cmd, GITS_CMD_MAPD);
419 its_encode_devid(cmd, desc->its_mapd_cmd.dev->device_id);
420 its_encode_size(cmd, size - 1);
421 its_encode_itt(cmd, itt_addr);
422 its_encode_valid(cmd, desc->its_mapd_cmd.valid);
429 static struct its_collection *its_build_mapc_cmd(struct its_node *its,
430 struct its_cmd_block *cmd,
431 struct its_cmd_desc *desc)
433 its_encode_cmd(cmd, GITS_CMD_MAPC);
434 its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
435 its_encode_target(cmd, desc->its_mapc_cmd.col->target_address);
436 its_encode_valid(cmd, desc->its_mapc_cmd.valid);
440 return desc->its_mapc_cmd.col;
443 static struct its_collection *its_build_mapti_cmd(struct its_node *its,
444 struct its_cmd_block *cmd,
445 struct its_cmd_desc *desc)
447 struct its_collection *col;
449 col = dev_event_to_col(desc->its_mapti_cmd.dev,
450 desc->its_mapti_cmd.event_id);
452 its_encode_cmd(cmd, GITS_CMD_MAPTI);
453 its_encode_devid(cmd, desc->its_mapti_cmd.dev->device_id);
454 its_encode_event_id(cmd, desc->its_mapti_cmd.event_id);
455 its_encode_phys_id(cmd, desc->its_mapti_cmd.phys_id);
456 its_encode_collection(cmd, col->col_id);
460 return valid_col(col);
463 static struct its_collection *its_build_movi_cmd(struct its_node *its,
464 struct its_cmd_block *cmd,
465 struct its_cmd_desc *desc)
467 struct its_collection *col;
469 col = dev_event_to_col(desc->its_movi_cmd.dev,
470 desc->its_movi_cmd.event_id);
472 its_encode_cmd(cmd, GITS_CMD_MOVI);
473 its_encode_devid(cmd, desc->its_movi_cmd.dev->device_id);
474 its_encode_event_id(cmd, desc->its_movi_cmd.event_id);
475 its_encode_collection(cmd, desc->its_movi_cmd.col->col_id);
479 return valid_col(col);
482 static struct its_collection *its_build_discard_cmd(struct its_node *its,
483 struct its_cmd_block *cmd,
484 struct its_cmd_desc *desc)
486 struct its_collection *col;
488 col = dev_event_to_col(desc->its_discard_cmd.dev,
489 desc->its_discard_cmd.event_id);
491 its_encode_cmd(cmd, GITS_CMD_DISCARD);
492 its_encode_devid(cmd, desc->its_discard_cmd.dev->device_id);
493 its_encode_event_id(cmd, desc->its_discard_cmd.event_id);
497 return valid_col(col);
500 static struct its_collection *its_build_inv_cmd(struct its_node *its,
501 struct its_cmd_block *cmd,
502 struct its_cmd_desc *desc)
504 struct its_collection *col;
506 col = dev_event_to_col(desc->its_inv_cmd.dev,
507 desc->its_inv_cmd.event_id);
509 its_encode_cmd(cmd, GITS_CMD_INV);
510 its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
511 its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
515 return valid_col(col);
518 static struct its_collection *its_build_int_cmd(struct its_node *its,
519 struct its_cmd_block *cmd,
520 struct its_cmd_desc *desc)
522 struct its_collection *col;
524 col = dev_event_to_col(desc->its_int_cmd.dev,
525 desc->its_int_cmd.event_id);
527 its_encode_cmd(cmd, GITS_CMD_INT);
528 its_encode_devid(cmd, desc->its_int_cmd.dev->device_id);
529 its_encode_event_id(cmd, desc->its_int_cmd.event_id);
533 return valid_col(col);
536 static struct its_collection *its_build_clear_cmd(struct its_node *its,
537 struct its_cmd_block *cmd,
538 struct its_cmd_desc *desc)
540 struct its_collection *col;
542 col = dev_event_to_col(desc->its_clear_cmd.dev,
543 desc->its_clear_cmd.event_id);
545 its_encode_cmd(cmd, GITS_CMD_CLEAR);
546 its_encode_devid(cmd, desc->its_clear_cmd.dev->device_id);
547 its_encode_event_id(cmd, desc->its_clear_cmd.event_id);
551 return valid_col(col);
554 static struct its_collection *its_build_invall_cmd(struct its_node *its,
555 struct its_cmd_block *cmd,
556 struct its_cmd_desc *desc)
558 its_encode_cmd(cmd, GITS_CMD_INVALL);
559 its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
566 static struct its_vpe *its_build_vinvall_cmd(struct its_node *its,
567 struct its_cmd_block *cmd,
568 struct its_cmd_desc *desc)
570 its_encode_cmd(cmd, GITS_CMD_VINVALL);
571 its_encode_vpeid(cmd, desc->its_vinvall_cmd.vpe->vpe_id);
575 return valid_vpe(its, desc->its_vinvall_cmd.vpe);
578 static struct its_vpe *its_build_vmapp_cmd(struct its_node *its,
579 struct its_cmd_block *cmd,
580 struct its_cmd_desc *desc)
582 unsigned long vpt_addr;
585 vpt_addr = virt_to_phys(page_address(desc->its_vmapp_cmd.vpe->vpt_page));
586 target = desc->its_vmapp_cmd.col->target_address + its->vlpi_redist_offset;
588 its_encode_cmd(cmd, GITS_CMD_VMAPP);
589 its_encode_vpeid(cmd, desc->its_vmapp_cmd.vpe->vpe_id);
590 its_encode_valid(cmd, desc->its_vmapp_cmd.valid);
591 its_encode_target(cmd, target);
592 its_encode_vpt_addr(cmd, vpt_addr);
593 its_encode_vpt_size(cmd, LPI_NRBITS - 1);
597 return valid_vpe(its, desc->its_vmapp_cmd.vpe);
600 static struct its_vpe *its_build_vmapti_cmd(struct its_node *its,
601 struct its_cmd_block *cmd,
602 struct its_cmd_desc *desc)
606 if (desc->its_vmapti_cmd.db_enabled)
607 db = desc->its_vmapti_cmd.vpe->vpe_db_lpi;
611 its_encode_cmd(cmd, GITS_CMD_VMAPTI);
612 its_encode_devid(cmd, desc->its_vmapti_cmd.dev->device_id);
613 its_encode_vpeid(cmd, desc->its_vmapti_cmd.vpe->vpe_id);
614 its_encode_event_id(cmd, desc->its_vmapti_cmd.event_id);
615 its_encode_db_phys_id(cmd, db);
616 its_encode_virt_id(cmd, desc->its_vmapti_cmd.virt_id);
620 return valid_vpe(its, desc->its_vmapti_cmd.vpe);
623 static struct its_vpe *its_build_vmovi_cmd(struct its_node *its,
624 struct its_cmd_block *cmd,
625 struct its_cmd_desc *desc)
629 if (desc->its_vmovi_cmd.db_enabled)
630 db = desc->its_vmovi_cmd.vpe->vpe_db_lpi;
634 its_encode_cmd(cmd, GITS_CMD_VMOVI);
635 its_encode_devid(cmd, desc->its_vmovi_cmd.dev->device_id);
636 its_encode_vpeid(cmd, desc->its_vmovi_cmd.vpe->vpe_id);
637 its_encode_event_id(cmd, desc->its_vmovi_cmd.event_id);
638 its_encode_db_phys_id(cmd, db);
639 its_encode_db_valid(cmd, true);
643 return valid_vpe(its, desc->its_vmovi_cmd.vpe);
646 static struct its_vpe *its_build_vmovp_cmd(struct its_node *its,
647 struct its_cmd_block *cmd,
648 struct its_cmd_desc *desc)
652 target = desc->its_vmovp_cmd.col->target_address + its->vlpi_redist_offset;
653 its_encode_cmd(cmd, GITS_CMD_VMOVP);
654 its_encode_seq_num(cmd, desc->its_vmovp_cmd.seq_num);
655 its_encode_its_list(cmd, desc->its_vmovp_cmd.its_list);
656 its_encode_vpeid(cmd, desc->its_vmovp_cmd.vpe->vpe_id);
657 its_encode_target(cmd, target);
661 return valid_vpe(its, desc->its_vmovp_cmd.vpe);
664 static u64 its_cmd_ptr_to_offset(struct its_node *its,
665 struct its_cmd_block *ptr)
667 return (ptr - its->cmd_base) * sizeof(*ptr);
670 static int its_queue_full(struct its_node *its)
675 widx = its->cmd_write - its->cmd_base;
676 ridx = readl_relaxed(its->base + GITS_CREADR) / sizeof(struct its_cmd_block);
678 /* This is incredibly unlikely to happen, unless the ITS locks up. */
679 if (((widx + 1) % ITS_CMD_QUEUE_NR_ENTRIES) == ridx)
685 static struct its_cmd_block *its_allocate_entry(struct its_node *its)
687 struct its_cmd_block *cmd;
688 u32 count = 1000000; /* 1s! */
690 while (its_queue_full(its)) {
693 pr_err_ratelimited("ITS queue not draining\n");
700 cmd = its->cmd_write++;
702 /* Handle queue wrapping */
703 if (its->cmd_write == (its->cmd_base + ITS_CMD_QUEUE_NR_ENTRIES))
704 its->cmd_write = its->cmd_base;
715 static struct its_cmd_block *its_post_commands(struct its_node *its)
717 u64 wr = its_cmd_ptr_to_offset(its, its->cmd_write);
719 writel_relaxed(wr, its->base + GITS_CWRITER);
721 return its->cmd_write;
724 static void its_flush_cmd(struct its_node *its, struct its_cmd_block *cmd)
727 * Make sure the commands written to memory are observable by
730 if (its->flags & ITS_FLAGS_CMDQ_NEEDS_FLUSHING)
731 gic_flush_dcache_to_poc(cmd, sizeof(*cmd));
736 static int its_wait_for_range_completion(struct its_node *its,
737 struct its_cmd_block *from,
738 struct its_cmd_block *to)
740 u64 rd_idx, from_idx, to_idx;
741 u32 count = 1000000; /* 1s! */
743 from_idx = its_cmd_ptr_to_offset(its, from);
744 to_idx = its_cmd_ptr_to_offset(its, to);
747 rd_idx = readl_relaxed(its->base + GITS_CREADR);
750 if (from_idx < to_idx && rd_idx >= to_idx)
754 if (from_idx >= to_idx && rd_idx >= to_idx && rd_idx < from_idx)
759 pr_err_ratelimited("ITS queue timeout (%llu %llu %llu)\n",
760 from_idx, to_idx, rd_idx);
770 /* Warning, macro hell follows */
771 #define BUILD_SINGLE_CMD_FUNC(name, buildtype, synctype, buildfn) \
772 void name(struct its_node *its, \
774 struct its_cmd_desc *desc) \
776 struct its_cmd_block *cmd, *sync_cmd, *next_cmd; \
777 synctype *sync_obj; \
778 unsigned long flags; \
780 raw_spin_lock_irqsave(&its->lock, flags); \
782 cmd = its_allocate_entry(its); \
783 if (!cmd) { /* We're soooooo screewed... */ \
784 raw_spin_unlock_irqrestore(&its->lock, flags); \
787 sync_obj = builder(its, cmd, desc); \
788 its_flush_cmd(its, cmd); \
791 sync_cmd = its_allocate_entry(its); \
795 buildfn(its, sync_cmd, sync_obj); \
796 its_flush_cmd(its, sync_cmd); \
800 next_cmd = its_post_commands(its); \
801 raw_spin_unlock_irqrestore(&its->lock, flags); \
803 if (its_wait_for_range_completion(its, cmd, next_cmd)) \
804 pr_err_ratelimited("ITS cmd %ps failed\n", builder); \
807 static void its_build_sync_cmd(struct its_node *its,
808 struct its_cmd_block *sync_cmd,
809 struct its_collection *sync_col)
811 its_encode_cmd(sync_cmd, GITS_CMD_SYNC);
812 its_encode_target(sync_cmd, sync_col->target_address);
814 its_fixup_cmd(sync_cmd);
817 static BUILD_SINGLE_CMD_FUNC(its_send_single_command, its_cmd_builder_t,
818 struct its_collection, its_build_sync_cmd)
820 static void its_build_vsync_cmd(struct its_node *its,
821 struct its_cmd_block *sync_cmd,
822 struct its_vpe *sync_vpe)
824 its_encode_cmd(sync_cmd, GITS_CMD_VSYNC);
825 its_encode_vpeid(sync_cmd, sync_vpe->vpe_id);
827 its_fixup_cmd(sync_cmd);
830 static BUILD_SINGLE_CMD_FUNC(its_send_single_vcommand, its_cmd_vbuilder_t,
831 struct its_vpe, its_build_vsync_cmd)
833 static void its_send_int(struct its_device *dev, u32 event_id)
835 struct its_cmd_desc desc;
837 desc.its_int_cmd.dev = dev;
838 desc.its_int_cmd.event_id = event_id;
840 its_send_single_command(dev->its, its_build_int_cmd, &desc);
843 static void its_send_clear(struct its_device *dev, u32 event_id)
845 struct its_cmd_desc desc;
847 desc.its_clear_cmd.dev = dev;
848 desc.its_clear_cmd.event_id = event_id;
850 its_send_single_command(dev->its, its_build_clear_cmd, &desc);
853 static void its_send_inv(struct its_device *dev, u32 event_id)
855 struct its_cmd_desc desc;
857 desc.its_inv_cmd.dev = dev;
858 desc.its_inv_cmd.event_id = event_id;
860 its_send_single_command(dev->its, its_build_inv_cmd, &desc);
863 static void its_send_mapd(struct its_device *dev, int valid)
865 struct its_cmd_desc desc;
867 desc.its_mapd_cmd.dev = dev;
868 desc.its_mapd_cmd.valid = !!valid;
870 its_send_single_command(dev->its, its_build_mapd_cmd, &desc);
873 static void its_send_mapc(struct its_node *its, struct its_collection *col,
876 struct its_cmd_desc desc;
878 desc.its_mapc_cmd.col = col;
879 desc.its_mapc_cmd.valid = !!valid;
881 its_send_single_command(its, its_build_mapc_cmd, &desc);
884 static void its_send_mapti(struct its_device *dev, u32 irq_id, u32 id)
886 struct its_cmd_desc desc;
888 desc.its_mapti_cmd.dev = dev;
889 desc.its_mapti_cmd.phys_id = irq_id;
890 desc.its_mapti_cmd.event_id = id;
892 its_send_single_command(dev->its, its_build_mapti_cmd, &desc);
895 static void its_send_movi(struct its_device *dev,
896 struct its_collection *col, u32 id)
898 struct its_cmd_desc desc;
900 desc.its_movi_cmd.dev = dev;
901 desc.its_movi_cmd.col = col;
902 desc.its_movi_cmd.event_id = id;
904 its_send_single_command(dev->its, its_build_movi_cmd, &desc);
907 static void its_send_discard(struct its_device *dev, u32 id)
909 struct its_cmd_desc desc;
911 desc.its_discard_cmd.dev = dev;
912 desc.its_discard_cmd.event_id = id;
914 its_send_single_command(dev->its, its_build_discard_cmd, &desc);
917 static void its_send_invall(struct its_node *its, struct its_collection *col)
919 struct its_cmd_desc desc;
921 desc.its_invall_cmd.col = col;
923 its_send_single_command(its, its_build_invall_cmd, &desc);
926 static void its_send_vmapti(struct its_device *dev, u32 id)
928 struct its_vlpi_map *map = &dev->event_map.vlpi_maps[id];
929 struct its_cmd_desc desc;
931 desc.its_vmapti_cmd.vpe = map->vpe;
932 desc.its_vmapti_cmd.dev = dev;
933 desc.its_vmapti_cmd.virt_id = map->vintid;
934 desc.its_vmapti_cmd.event_id = id;
935 desc.its_vmapti_cmd.db_enabled = map->db_enabled;
937 its_send_single_vcommand(dev->its, its_build_vmapti_cmd, &desc);
940 static void its_send_vmovi(struct its_device *dev, u32 id)
942 struct its_vlpi_map *map = &dev->event_map.vlpi_maps[id];
943 struct its_cmd_desc desc;
945 desc.its_vmovi_cmd.vpe = map->vpe;
946 desc.its_vmovi_cmd.dev = dev;
947 desc.its_vmovi_cmd.event_id = id;
948 desc.its_vmovi_cmd.db_enabled = map->db_enabled;
950 its_send_single_vcommand(dev->its, its_build_vmovi_cmd, &desc);
953 static void its_send_vmapp(struct its_node *its,
954 struct its_vpe *vpe, bool valid)
956 struct its_cmd_desc desc;
958 desc.its_vmapp_cmd.vpe = vpe;
959 desc.its_vmapp_cmd.valid = valid;
960 desc.its_vmapp_cmd.col = &its->collections[vpe->col_idx];
962 its_send_single_vcommand(its, its_build_vmapp_cmd, &desc);
965 static void its_send_vmovp(struct its_vpe *vpe)
967 struct its_cmd_desc desc;
968 struct its_node *its;
970 int col_id = vpe->col_idx;
972 desc.its_vmovp_cmd.vpe = vpe;
973 desc.its_vmovp_cmd.its_list = (u16)its_list_map;
976 its = list_first_entry(&its_nodes, struct its_node, entry);
977 desc.its_vmovp_cmd.seq_num = 0;
978 desc.its_vmovp_cmd.col = &its->collections[col_id];
979 its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
984 * Yet another marvel of the architecture. If using the
985 * its_list "feature", we need to make sure that all ITSs
986 * receive all VMOVP commands in the same order. The only way
987 * to guarantee this is to make vmovp a serialization point.
991 raw_spin_lock_irqsave(&vmovp_lock, flags);
993 desc.its_vmovp_cmd.seq_num = vmovp_seq_num++;
996 list_for_each_entry(its, &its_nodes, entry) {
1000 if (!vpe->its_vm->vlpi_count[its->list_nr])
1003 desc.its_vmovp_cmd.col = &its->collections[col_id];
1004 its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
1007 raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1010 static void its_send_vinvall(struct its_node *its, struct its_vpe *vpe)
1012 struct its_cmd_desc desc;
1014 desc.its_vinvall_cmd.vpe = vpe;
1015 its_send_single_vcommand(its, its_build_vinvall_cmd, &desc);
1019 * irqchip functions - assumes MSI, mostly.
1022 static inline u32 its_get_event_id(struct irq_data *d)
1024 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1025 return d->hwirq - its_dev->event_map.lpi_base;
1028 static void lpi_write_config(struct irq_data *d, u8 clr, u8 set)
1030 irq_hw_number_t hwirq;
1031 struct page *prop_page;
1034 if (irqd_is_forwarded_to_vcpu(d)) {
1035 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1036 u32 event = its_get_event_id(d);
1037 struct its_vlpi_map *map;
1039 prop_page = its_dev->event_map.vm->vprop_page;
1040 map = &its_dev->event_map.vlpi_maps[event];
1041 hwirq = map->vintid;
1043 /* Remember the updated property */
1044 map->properties &= ~clr;
1045 map->properties |= set | LPI_PROP_GROUP1;
1047 prop_page = gic_rdists->prop_page;
1051 cfg = page_address(prop_page) + hwirq - 8192;
1053 *cfg |= set | LPI_PROP_GROUP1;
1056 * Make the above write visible to the redistributors.
1057 * And yes, we're flushing exactly: One. Single. Byte.
1060 if (gic_rdists->flags & RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING)
1061 gic_flush_dcache_to_poc(cfg, sizeof(*cfg));
1066 static void lpi_update_config(struct irq_data *d, u8 clr, u8 set)
1068 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1070 lpi_write_config(d, clr, set);
1071 its_send_inv(its_dev, its_get_event_id(d));
1074 static void its_vlpi_set_doorbell(struct irq_data *d, bool enable)
1076 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1077 u32 event = its_get_event_id(d);
1079 if (its_dev->event_map.vlpi_maps[event].db_enabled == enable)
1082 its_dev->event_map.vlpi_maps[event].db_enabled = enable;
1085 * More fun with the architecture:
1087 * Ideally, we'd issue a VMAPTI to set the doorbell to its LPI
1088 * value or to 1023, depending on the enable bit. But that
1089 * would be issueing a mapping for an /existing/ DevID+EventID
1090 * pair, which is UNPREDICTABLE. Instead, let's issue a VMOVI
1091 * to the /same/ vPE, using this opportunity to adjust the
1092 * doorbell. Mouahahahaha. We loves it, Precious.
1094 its_send_vmovi(its_dev, event);
1097 static void its_mask_irq(struct irq_data *d)
1099 if (irqd_is_forwarded_to_vcpu(d))
1100 its_vlpi_set_doorbell(d, false);
1102 lpi_update_config(d, LPI_PROP_ENABLED, 0);
1105 static void its_unmask_irq(struct irq_data *d)
1107 if (irqd_is_forwarded_to_vcpu(d))
1108 its_vlpi_set_doorbell(d, true);
1110 lpi_update_config(d, 0, LPI_PROP_ENABLED);
1113 static int its_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
1117 const struct cpumask *cpu_mask = cpu_online_mask;
1118 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1119 struct its_collection *target_col;
1120 u32 id = its_get_event_id(d);
1122 /* A forwarded interrupt should use irq_set_vcpu_affinity */
1123 if (irqd_is_forwarded_to_vcpu(d))
1126 /* lpi cannot be routed to a redistributor that is on a foreign node */
1127 if (its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) {
1128 if (its_dev->its->numa_node >= 0) {
1129 cpu_mask = cpumask_of_node(its_dev->its->numa_node);
1130 if (!cpumask_intersects(mask_val, cpu_mask))
1135 cpu = cpumask_any_and(mask_val, cpu_mask);
1137 if (cpu >= nr_cpu_ids)
1140 /* don't set the affinity when the target cpu is same as current one */
1141 if (cpu != its_dev->event_map.col_map[id]) {
1142 target_col = &its_dev->its->collections[cpu];
1143 its_send_movi(its_dev, target_col, id);
1144 its_dev->event_map.col_map[id] = cpu;
1145 irq_data_update_effective_affinity(d, cpumask_of(cpu));
1148 return IRQ_SET_MASK_OK_DONE;
1151 static u64 its_irq_get_msi_base(struct its_device *its_dev)
1153 struct its_node *its = its_dev->its;
1155 return its->phys_base + GITS_TRANSLATER;
1158 static void its_irq_compose_msi_msg(struct irq_data *d, struct msi_msg *msg)
1160 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1161 struct its_node *its;
1165 addr = its->get_msi_base(its_dev);
1167 msg->address_lo = lower_32_bits(addr);
1168 msg->address_hi = upper_32_bits(addr);
1169 msg->data = its_get_event_id(d);
1171 iommu_dma_map_msi_msg(d->irq, msg);
1174 static int its_irq_set_irqchip_state(struct irq_data *d,
1175 enum irqchip_irq_state which,
1178 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1179 u32 event = its_get_event_id(d);
1181 if (which != IRQCHIP_STATE_PENDING)
1185 its_send_int(its_dev, event);
1187 its_send_clear(its_dev, event);
1192 static void its_map_vm(struct its_node *its, struct its_vm *vm)
1194 unsigned long flags;
1196 /* Not using the ITS list? Everything is always mapped. */
1200 raw_spin_lock_irqsave(&vmovp_lock, flags);
1203 * If the VM wasn't mapped yet, iterate over the vpes and get
1206 vm->vlpi_count[its->list_nr]++;
1208 if (vm->vlpi_count[its->list_nr] == 1) {
1211 for (i = 0; i < vm->nr_vpes; i++) {
1212 struct its_vpe *vpe = vm->vpes[i];
1213 struct irq_data *d = irq_get_irq_data(vpe->irq);
1215 /* Map the VPE to the first possible CPU */
1216 vpe->col_idx = cpumask_first(cpu_online_mask);
1217 its_send_vmapp(its, vpe, true);
1218 its_send_vinvall(its, vpe);
1219 irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
1223 raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1226 static void its_unmap_vm(struct its_node *its, struct its_vm *vm)
1228 unsigned long flags;
1230 /* Not using the ITS list? Everything is always mapped. */
1234 raw_spin_lock_irqsave(&vmovp_lock, flags);
1236 if (!--vm->vlpi_count[its->list_nr]) {
1239 for (i = 0; i < vm->nr_vpes; i++)
1240 its_send_vmapp(its, vm->vpes[i], false);
1243 raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1246 static int its_vlpi_map(struct irq_data *d, struct its_cmd_info *info)
1248 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1249 u32 event = its_get_event_id(d);
1255 mutex_lock(&its_dev->event_map.vlpi_lock);
1257 if (!its_dev->event_map.vm) {
1258 struct its_vlpi_map *maps;
1260 maps = kcalloc(its_dev->event_map.nr_lpis, sizeof(*maps),
1267 its_dev->event_map.vm = info->map->vm;
1268 its_dev->event_map.vlpi_maps = maps;
1269 } else if (its_dev->event_map.vm != info->map->vm) {
1274 /* Get our private copy of the mapping information */
1275 its_dev->event_map.vlpi_maps[event] = *info->map;
1277 if (irqd_is_forwarded_to_vcpu(d)) {
1278 /* Already mapped, move it around */
1279 its_send_vmovi(its_dev, event);
1281 /* Ensure all the VPEs are mapped on this ITS */
1282 its_map_vm(its_dev->its, info->map->vm);
1285 * Flag the interrupt as forwarded so that we can
1286 * start poking the virtual property table.
1288 irqd_set_forwarded_to_vcpu(d);
1290 /* Write out the property to the prop table */
1291 lpi_write_config(d, 0xff, info->map->properties);
1293 /* Drop the physical mapping */
1294 its_send_discard(its_dev, event);
1296 /* and install the virtual one */
1297 its_send_vmapti(its_dev, event);
1299 /* Increment the number of VLPIs */
1300 its_dev->event_map.nr_vlpis++;
1304 mutex_unlock(&its_dev->event_map.vlpi_lock);
1308 static int its_vlpi_get(struct irq_data *d, struct its_cmd_info *info)
1310 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1311 u32 event = its_get_event_id(d);
1314 mutex_lock(&its_dev->event_map.vlpi_lock);
1316 if (!its_dev->event_map.vm ||
1317 !its_dev->event_map.vlpi_maps[event].vm) {
1322 /* Copy our mapping information to the incoming request */
1323 *info->map = its_dev->event_map.vlpi_maps[event];
1326 mutex_unlock(&its_dev->event_map.vlpi_lock);
1330 static int its_vlpi_unmap(struct irq_data *d)
1332 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1333 u32 event = its_get_event_id(d);
1336 mutex_lock(&its_dev->event_map.vlpi_lock);
1338 if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d)) {
1343 /* Drop the virtual mapping */
1344 its_send_discard(its_dev, event);
1346 /* and restore the physical one */
1347 irqd_clr_forwarded_to_vcpu(d);
1348 its_send_mapti(its_dev, d->hwirq, event);
1349 lpi_update_config(d, 0xff, (LPI_PROP_DEFAULT_PRIO |
1353 /* Potentially unmap the VM from this ITS */
1354 its_unmap_vm(its_dev->its, its_dev->event_map.vm);
1357 * Drop the refcount and make the device available again if
1358 * this was the last VLPI.
1360 if (!--its_dev->event_map.nr_vlpis) {
1361 its_dev->event_map.vm = NULL;
1362 kfree(its_dev->event_map.vlpi_maps);
1366 mutex_unlock(&its_dev->event_map.vlpi_lock);
1370 static int its_vlpi_prop_update(struct irq_data *d, struct its_cmd_info *info)
1372 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1374 if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d))
1377 if (info->cmd_type == PROP_UPDATE_AND_INV_VLPI)
1378 lpi_update_config(d, 0xff, info->config);
1380 lpi_write_config(d, 0xff, info->config);
1381 its_vlpi_set_doorbell(d, !!(info->config & LPI_PROP_ENABLED));
1386 static int its_irq_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
1388 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1389 struct its_cmd_info *info = vcpu_info;
1392 if (!its_dev->its->is_v4)
1395 /* Unmap request? */
1397 return its_vlpi_unmap(d);
1399 switch (info->cmd_type) {
1401 return its_vlpi_map(d, info);
1404 return its_vlpi_get(d, info);
1406 case PROP_UPDATE_VLPI:
1407 case PROP_UPDATE_AND_INV_VLPI:
1408 return its_vlpi_prop_update(d, info);
1415 static struct irq_chip its_irq_chip = {
1417 .irq_mask = its_mask_irq,
1418 .irq_unmask = its_unmask_irq,
1419 .irq_eoi = irq_chip_eoi_parent,
1420 .irq_set_affinity = its_set_affinity,
1421 .irq_compose_msi_msg = its_irq_compose_msi_msg,
1422 .irq_set_irqchip_state = its_irq_set_irqchip_state,
1423 .irq_set_vcpu_affinity = its_irq_set_vcpu_affinity,
1428 * How we allocate LPIs:
1430 * lpi_range_list contains ranges of LPIs that are to available to
1431 * allocate from. To allocate LPIs, just pick the first range that
1432 * fits the required allocation, and reduce it by the required
1433 * amount. Once empty, remove the range from the list.
1435 * To free a range of LPIs, add a free range to the list, sort it and
1436 * merge the result if the new range happens to be adjacent to an
1437 * already free block.
1439 * The consequence of the above is that allocation is cost is low, but
1440 * freeing is expensive. We assumes that freeing rarely occurs.
1443 static DEFINE_MUTEX(lpi_range_lock);
1444 static LIST_HEAD(lpi_range_list);
1447 struct list_head entry;
1452 static struct lpi_range *mk_lpi_range(u32 base, u32 span)
1454 struct lpi_range *range;
1456 range = kzalloc(sizeof(*range), GFP_KERNEL);
1458 INIT_LIST_HEAD(&range->entry);
1459 range->base_id = base;
1466 static int lpi_range_cmp(void *priv, struct list_head *a, struct list_head *b)
1468 struct lpi_range *ra, *rb;
1470 ra = container_of(a, struct lpi_range, entry);
1471 rb = container_of(b, struct lpi_range, entry);
1473 return rb->base_id - ra->base_id;
1476 static void merge_lpi_ranges(void)
1478 struct lpi_range *range, *tmp;
1480 list_for_each_entry_safe(range, tmp, &lpi_range_list, entry) {
1481 if (!list_is_last(&range->entry, &lpi_range_list) &&
1482 (tmp->base_id == (range->base_id + range->span))) {
1483 tmp->base_id = range->base_id;
1484 tmp->span += range->span;
1485 list_del(&range->entry);
1491 static int alloc_lpi_range(u32 nr_lpis, u32 *base)
1493 struct lpi_range *range, *tmp;
1496 mutex_lock(&lpi_range_lock);
1498 list_for_each_entry_safe(range, tmp, &lpi_range_list, entry) {
1499 if (range->span >= nr_lpis) {
1500 *base = range->base_id;
1501 range->base_id += nr_lpis;
1502 range->span -= nr_lpis;
1504 if (range->span == 0) {
1505 list_del(&range->entry);
1514 mutex_unlock(&lpi_range_lock);
1516 pr_debug("ITS: alloc %u:%u\n", *base, nr_lpis);
1520 static int free_lpi_range(u32 base, u32 nr_lpis)
1522 struct lpi_range *new;
1525 mutex_lock(&lpi_range_lock);
1527 new = mk_lpi_range(base, nr_lpis);
1533 list_add(&new->entry, &lpi_range_list);
1534 list_sort(NULL, &lpi_range_list, lpi_range_cmp);
1537 mutex_unlock(&lpi_range_lock);
1541 static int __init its_lpi_init(u32 id_bits)
1543 u32 lpis = (1UL << id_bits) - 8192;
1547 numlpis = 1UL << GICD_TYPER_NUM_LPIS(gic_rdists->gicd_typer);
1549 if (numlpis > 2 && !WARN_ON(numlpis > lpis)) {
1551 pr_info("ITS: Using hypervisor restricted LPI range [%u]\n",
1556 * Initializing the allocator is just the same as freeing the
1557 * full range of LPIs.
1559 err = free_lpi_range(8192, lpis);
1560 pr_debug("ITS: Allocator initialized for %u LPIs\n", lpis);
1564 static unsigned long *its_lpi_alloc(int nr_irqs, u32 *base, int *nr_ids)
1566 unsigned long *bitmap = NULL;
1570 err = alloc_lpi_range(nr_irqs, base);
1575 } while (nr_irqs > 0);
1580 bitmap = kcalloc(BITS_TO_LONGS(nr_irqs), sizeof (long), GFP_ATOMIC);
1588 *base = *nr_ids = 0;
1593 static void its_lpi_free(unsigned long *bitmap, u32 base, u32 nr_ids)
1595 WARN_ON(free_lpi_range(base, nr_ids));
1599 static struct page *its_allocate_prop_table(gfp_t gfp_flags)
1601 struct page *prop_page;
1603 prop_page = alloc_pages(gfp_flags, get_order(LPI_PROPBASE_SZ));
1607 /* Priority 0xa0, Group-1, disabled */
1608 memset(page_address(prop_page),
1609 LPI_PROP_DEFAULT_PRIO | LPI_PROP_GROUP1,
1612 /* Make sure the GIC will observe the written configuration */
1613 gic_flush_dcache_to_poc(page_address(prop_page), LPI_PROPBASE_SZ);
1618 static void its_free_prop_table(struct page *prop_page)
1620 free_pages((unsigned long)page_address(prop_page),
1621 get_order(LPI_PROPBASE_SZ));
1624 static int __init its_alloc_lpi_tables(void)
1628 lpi_id_bits = GICD_TYPER_ID_BITS(gic_rdists->gicd_typer);
1629 gic_rdists->prop_page = its_allocate_prop_table(GFP_NOWAIT);
1630 if (!gic_rdists->prop_page) {
1631 pr_err("Failed to allocate PROPBASE\n");
1635 paddr = page_to_phys(gic_rdists->prop_page);
1636 pr_info("GIC: using LPI property table @%pa\n", &paddr);
1638 return its_lpi_init(lpi_id_bits);
1641 static const char *its_base_type_string[] = {
1642 [GITS_BASER_TYPE_DEVICE] = "Devices",
1643 [GITS_BASER_TYPE_VCPU] = "Virtual CPUs",
1644 [GITS_BASER_TYPE_RESERVED3] = "Reserved (3)",
1645 [GITS_BASER_TYPE_COLLECTION] = "Interrupt Collections",
1646 [GITS_BASER_TYPE_RESERVED5] = "Reserved (5)",
1647 [GITS_BASER_TYPE_RESERVED6] = "Reserved (6)",
1648 [GITS_BASER_TYPE_RESERVED7] = "Reserved (7)",
1651 static u64 its_read_baser(struct its_node *its, struct its_baser *baser)
1653 u32 idx = baser - its->tables;
1655 return gits_read_baser(its->base + GITS_BASER + (idx << 3));
1658 static void its_write_baser(struct its_node *its, struct its_baser *baser,
1661 u32 idx = baser - its->tables;
1663 gits_write_baser(val, its->base + GITS_BASER + (idx << 3));
1664 baser->val = its_read_baser(its, baser);
1667 static int its_setup_baser(struct its_node *its, struct its_baser *baser,
1668 u64 cache, u64 shr, u32 psz, u32 order,
1671 u64 val = its_read_baser(its, baser);
1672 u64 esz = GITS_BASER_ENTRY_SIZE(val);
1673 u64 type = GITS_BASER_TYPE(val);
1674 u64 baser_phys, tmp;
1679 alloc_pages = (PAGE_ORDER_TO_SIZE(order) / psz);
1680 if (alloc_pages > GITS_BASER_PAGES_MAX) {
1681 pr_warn("ITS@%pa: %s too large, reduce ITS pages %u->%u\n",
1682 &its->phys_base, its_base_type_string[type],
1683 alloc_pages, GITS_BASER_PAGES_MAX);
1684 alloc_pages = GITS_BASER_PAGES_MAX;
1685 order = get_order(GITS_BASER_PAGES_MAX * psz);
1688 base = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, order);
1692 baser_phys = virt_to_phys(base);
1694 /* Check if the physical address of the memory is above 48bits */
1695 if (IS_ENABLED(CONFIG_ARM64_64K_PAGES) && (baser_phys >> 48)) {
1697 /* 52bit PA is supported only when PageSize=64K */
1698 if (psz != SZ_64K) {
1699 pr_err("ITS: no 52bit PA support when psz=%d\n", psz);
1700 free_pages((unsigned long)base, order);
1704 /* Convert 52bit PA to 48bit field */
1705 baser_phys = GITS_BASER_PHYS_52_to_48(baser_phys);
1710 (type << GITS_BASER_TYPE_SHIFT) |
1711 ((esz - 1) << GITS_BASER_ENTRY_SIZE_SHIFT) |
1712 ((alloc_pages - 1) << GITS_BASER_PAGES_SHIFT) |
1717 val |= indirect ? GITS_BASER_INDIRECT : 0x0;
1721 val |= GITS_BASER_PAGE_SIZE_4K;
1724 val |= GITS_BASER_PAGE_SIZE_16K;
1727 val |= GITS_BASER_PAGE_SIZE_64K;
1731 its_write_baser(its, baser, val);
1734 if ((val ^ tmp) & GITS_BASER_SHAREABILITY_MASK) {
1736 * Shareability didn't stick. Just use
1737 * whatever the read reported, which is likely
1738 * to be the only thing this redistributor
1739 * supports. If that's zero, make it
1740 * non-cacheable as well.
1742 shr = tmp & GITS_BASER_SHAREABILITY_MASK;
1744 cache = GITS_BASER_nC;
1745 gic_flush_dcache_to_poc(base, PAGE_ORDER_TO_SIZE(order));
1750 if ((val ^ tmp) & GITS_BASER_PAGE_SIZE_MASK) {
1752 * Page size didn't stick. Let's try a smaller
1753 * size and retry. If we reach 4K, then
1754 * something is horribly wrong...
1756 free_pages((unsigned long)base, order);
1762 goto retry_alloc_baser;
1765 goto retry_alloc_baser;
1770 pr_err("ITS@%pa: %s doesn't stick: %llx %llx\n",
1771 &its->phys_base, its_base_type_string[type],
1773 free_pages((unsigned long)base, order);
1777 baser->order = order;
1780 tmp = indirect ? GITS_LVL1_ENTRY_SIZE : esz;
1782 pr_info("ITS@%pa: allocated %d %s @%lx (%s, esz %d, psz %dK, shr %d)\n",
1783 &its->phys_base, (int)(PAGE_ORDER_TO_SIZE(order) / (int)tmp),
1784 its_base_type_string[type],
1785 (unsigned long)virt_to_phys(base),
1786 indirect ? "indirect" : "flat", (int)esz,
1787 psz / SZ_1K, (int)shr >> GITS_BASER_SHAREABILITY_SHIFT);
1792 static bool its_parse_indirect_baser(struct its_node *its,
1793 struct its_baser *baser,
1794 u32 psz, u32 *order, u32 ids)
1796 u64 tmp = its_read_baser(its, baser);
1797 u64 type = GITS_BASER_TYPE(tmp);
1798 u64 esz = GITS_BASER_ENTRY_SIZE(tmp);
1799 u64 val = GITS_BASER_InnerShareable | GITS_BASER_RaWaWb;
1800 u32 new_order = *order;
1801 bool indirect = false;
1803 /* No need to enable Indirection if memory requirement < (psz*2)bytes */
1804 if ((esz << ids) > (psz * 2)) {
1806 * Find out whether hw supports a single or two-level table by
1807 * table by reading bit at offset '62' after writing '1' to it.
1809 its_write_baser(its, baser, val | GITS_BASER_INDIRECT);
1810 indirect = !!(baser->val & GITS_BASER_INDIRECT);
1814 * The size of the lvl2 table is equal to ITS page size
1815 * which is 'psz'. For computing lvl1 table size,
1816 * subtract ID bits that sparse lvl2 table from 'ids'
1817 * which is reported by ITS hardware times lvl1 table
1820 ids -= ilog2(psz / (int)esz);
1821 esz = GITS_LVL1_ENTRY_SIZE;
1826 * Allocate as many entries as required to fit the
1827 * range of device IDs that the ITS can grok... The ID
1828 * space being incredibly sparse, this results in a
1829 * massive waste of memory if two-level device table
1830 * feature is not supported by hardware.
1832 new_order = max_t(u32, get_order(esz << ids), new_order);
1833 if (new_order >= MAX_ORDER) {
1834 new_order = MAX_ORDER - 1;
1835 ids = ilog2(PAGE_ORDER_TO_SIZE(new_order) / (int)esz);
1836 pr_warn("ITS@%pa: %s Table too large, reduce ids %u->%u\n",
1837 &its->phys_base, its_base_type_string[type],
1838 its->device_ids, ids);
1846 static void its_free_tables(struct its_node *its)
1850 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
1851 if (its->tables[i].base) {
1852 free_pages((unsigned long)its->tables[i].base,
1853 its->tables[i].order);
1854 its->tables[i].base = NULL;
1859 static int its_alloc_tables(struct its_node *its)
1861 u64 shr = GITS_BASER_InnerShareable;
1862 u64 cache = GITS_BASER_RaWaWb;
1866 if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_22375)
1867 /* erratum 24313: ignore memory access type */
1868 cache = GITS_BASER_nCnB;
1870 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
1871 struct its_baser *baser = its->tables + i;
1872 u64 val = its_read_baser(its, baser);
1873 u64 type = GITS_BASER_TYPE(val);
1874 u32 order = get_order(psz);
1875 bool indirect = false;
1878 case GITS_BASER_TYPE_NONE:
1881 case GITS_BASER_TYPE_DEVICE:
1882 indirect = its_parse_indirect_baser(its, baser,
1885 case GITS_BASER_TYPE_VCPU:
1886 indirect = its_parse_indirect_baser(its, baser,
1888 ITS_MAX_VPEID_BITS);
1892 err = its_setup_baser(its, baser, cache, shr, psz, order, indirect);
1894 its_free_tables(its);
1898 /* Update settings which will be used for next BASERn */
1900 cache = baser->val & GITS_BASER_CACHEABILITY_MASK;
1901 shr = baser->val & GITS_BASER_SHAREABILITY_MASK;
1907 static int its_alloc_collections(struct its_node *its)
1911 its->collections = kcalloc(nr_cpu_ids, sizeof(*its->collections),
1913 if (!its->collections)
1916 for (i = 0; i < nr_cpu_ids; i++)
1917 its->collections[i].target_address = ~0ULL;
1922 static struct page *its_allocate_pending_table(gfp_t gfp_flags)
1924 struct page *pend_page;
1926 * The pending pages have to be at least 64kB aligned,
1927 * hence the 'max(LPI_PENDBASE_SZ, SZ_64K)' below.
1929 pend_page = alloc_pages(gfp_flags | __GFP_ZERO,
1930 get_order(max_t(u32, LPI_PENDBASE_SZ, SZ_64K)));
1934 /* Make sure the GIC will observe the zero-ed page */
1935 gic_flush_dcache_to_poc(page_address(pend_page), LPI_PENDBASE_SZ);
1940 static void its_free_pending_table(struct page *pt)
1942 free_pages((unsigned long)page_address(pt),
1943 get_order(max_t(u32, LPI_PENDBASE_SZ, SZ_64K)));
1946 static void its_cpu_init_lpis(void)
1948 void __iomem *rbase = gic_data_rdist_rd_base();
1949 struct page *pend_page;
1952 /* If we didn't allocate the pending table yet, do it now */
1953 pend_page = gic_data_rdist()->pend_page;
1957 pend_page = its_allocate_pending_table(GFP_NOWAIT);
1959 pr_err("Failed to allocate PENDBASE for CPU%d\n",
1960 smp_processor_id());
1964 paddr = page_to_phys(pend_page);
1965 pr_info("CPU%d: using LPI pending table @%pa\n",
1966 smp_processor_id(), &paddr);
1967 gic_data_rdist()->pend_page = pend_page;
1971 val = (page_to_phys(gic_rdists->prop_page) |
1972 GICR_PROPBASER_InnerShareable |
1973 GICR_PROPBASER_RaWaWb |
1974 ((LPI_NRBITS - 1) & GICR_PROPBASER_IDBITS_MASK));
1976 gicr_write_propbaser(val, rbase + GICR_PROPBASER);
1977 tmp = gicr_read_propbaser(rbase + GICR_PROPBASER);
1979 if ((tmp ^ val) & GICR_PROPBASER_SHAREABILITY_MASK) {
1980 if (!(tmp & GICR_PROPBASER_SHAREABILITY_MASK)) {
1982 * The HW reports non-shareable, we must
1983 * remove the cacheability attributes as
1986 val &= ~(GICR_PROPBASER_SHAREABILITY_MASK |
1987 GICR_PROPBASER_CACHEABILITY_MASK);
1988 val |= GICR_PROPBASER_nC;
1989 gicr_write_propbaser(val, rbase + GICR_PROPBASER);
1991 pr_info_once("GIC: using cache flushing for LPI property table\n");
1992 gic_rdists->flags |= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING;
1996 val = (page_to_phys(pend_page) |
1997 GICR_PENDBASER_InnerShareable |
1998 GICR_PENDBASER_RaWaWb);
2000 gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
2001 tmp = gicr_read_pendbaser(rbase + GICR_PENDBASER);
2003 if (!(tmp & GICR_PENDBASER_SHAREABILITY_MASK)) {
2005 * The HW reports non-shareable, we must remove the
2006 * cacheability attributes as well.
2008 val &= ~(GICR_PENDBASER_SHAREABILITY_MASK |
2009 GICR_PENDBASER_CACHEABILITY_MASK);
2010 val |= GICR_PENDBASER_nC;
2011 gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
2015 val = readl_relaxed(rbase + GICR_CTLR);
2016 val |= GICR_CTLR_ENABLE_LPIS;
2017 writel_relaxed(val, rbase + GICR_CTLR);
2019 /* Make sure the GIC has seen the above */
2023 static void its_cpu_init_collection(struct its_node *its)
2025 int cpu = smp_processor_id();
2028 /* avoid cross node collections and its mapping */
2029 if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) {
2030 struct device_node *cpu_node;
2032 cpu_node = of_get_cpu_node(cpu, NULL);
2033 if (its->numa_node != NUMA_NO_NODE &&
2034 its->numa_node != of_node_to_nid(cpu_node))
2039 * We now have to bind each collection to its target
2042 if (gic_read_typer(its->base + GITS_TYPER) & GITS_TYPER_PTA) {
2044 * This ITS wants the physical address of the
2047 target = gic_data_rdist()->phys_base;
2049 /* This ITS wants a linear CPU number. */
2050 target = gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER);
2051 target = GICR_TYPER_CPU_NUMBER(target) << 16;
2054 /* Perform collection mapping */
2055 its->collections[cpu].target_address = target;
2056 its->collections[cpu].col_id = cpu;
2058 its_send_mapc(its, &its->collections[cpu], 1);
2059 its_send_invall(its, &its->collections[cpu]);
2062 static void its_cpu_init_collections(void)
2064 struct its_node *its;
2066 raw_spin_lock(&its_lock);
2068 list_for_each_entry(its, &its_nodes, entry)
2069 its_cpu_init_collection(its);
2071 raw_spin_unlock(&its_lock);
2074 static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
2076 struct its_device *its_dev = NULL, *tmp;
2077 unsigned long flags;
2079 raw_spin_lock_irqsave(&its->lock, flags);
2081 list_for_each_entry(tmp, &its->its_device_list, entry) {
2082 if (tmp->device_id == dev_id) {
2088 raw_spin_unlock_irqrestore(&its->lock, flags);
2093 static struct its_baser *its_get_baser(struct its_node *its, u32 type)
2097 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
2098 if (GITS_BASER_TYPE(its->tables[i].val) == type)
2099 return &its->tables[i];
2105 static bool its_alloc_table_entry(struct its_baser *baser, u32 id)
2111 /* Don't allow device id that exceeds single, flat table limit */
2112 esz = GITS_BASER_ENTRY_SIZE(baser->val);
2113 if (!(baser->val & GITS_BASER_INDIRECT))
2114 return (id < (PAGE_ORDER_TO_SIZE(baser->order) / esz));
2116 /* Compute 1st level table index & check if that exceeds table limit */
2117 idx = id >> ilog2(baser->psz / esz);
2118 if (idx >= (PAGE_ORDER_TO_SIZE(baser->order) / GITS_LVL1_ENTRY_SIZE))
2121 table = baser->base;
2123 /* Allocate memory for 2nd level table */
2125 page = alloc_pages(GFP_KERNEL | __GFP_ZERO, get_order(baser->psz));
2129 /* Flush Lvl2 table to PoC if hw doesn't support coherency */
2130 if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
2131 gic_flush_dcache_to_poc(page_address(page), baser->psz);
2133 table[idx] = cpu_to_le64(page_to_phys(page) | GITS_BASER_VALID);
2135 /* Flush Lvl1 entry to PoC if hw doesn't support coherency */
2136 if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
2137 gic_flush_dcache_to_poc(table + idx, GITS_LVL1_ENTRY_SIZE);
2139 /* Ensure updated table contents are visible to ITS hardware */
2146 static bool its_alloc_device_table(struct its_node *its, u32 dev_id)
2148 struct its_baser *baser;
2150 baser = its_get_baser(its, GITS_BASER_TYPE_DEVICE);
2152 /* Don't allow device id that exceeds ITS hardware limit */
2154 return (ilog2(dev_id) < its->device_ids);
2156 return its_alloc_table_entry(baser, dev_id);
2159 static bool its_alloc_vpe_table(u32 vpe_id)
2161 struct its_node *its;
2164 * Make sure the L2 tables are allocated on *all* v4 ITSs. We
2165 * could try and only do it on ITSs corresponding to devices
2166 * that have interrupts targeted at this VPE, but the
2167 * complexity becomes crazy (and you have tons of memory
2170 list_for_each_entry(its, &its_nodes, entry) {
2171 struct its_baser *baser;
2176 baser = its_get_baser(its, GITS_BASER_TYPE_VCPU);
2180 if (!its_alloc_table_entry(baser, vpe_id))
2187 static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
2188 int nvecs, bool alloc_lpis)
2190 struct its_device *dev;
2191 unsigned long *lpi_map = NULL;
2192 unsigned long flags;
2193 u16 *col_map = NULL;
2200 if (!its_alloc_device_table(its, dev_id))
2203 if (WARN_ON(!is_power_of_2(nvecs)))
2204 nvecs = roundup_pow_of_two(nvecs);
2206 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2208 * Even if the device wants a single LPI, the ITT must be
2209 * sized as a power of two (and you need at least one bit...).
2211 nr_ites = max(2, nvecs);
2212 sz = nr_ites * its->ite_size;
2213 sz = max(sz, ITS_ITT_ALIGN) + ITS_ITT_ALIGN - 1;
2214 itt = kzalloc(sz, GFP_KERNEL);
2216 lpi_map = its_lpi_alloc(nvecs, &lpi_base, &nr_lpis);
2218 col_map = kcalloc(nr_lpis, sizeof(*col_map),
2221 col_map = kcalloc(nr_ites, sizeof(*col_map), GFP_KERNEL);
2226 if (!dev || !itt || !col_map || (!lpi_map && alloc_lpis)) {
2234 gic_flush_dcache_to_poc(itt, sz);
2238 dev->nr_ites = nr_ites;
2239 dev->event_map.lpi_map = lpi_map;
2240 dev->event_map.col_map = col_map;
2241 dev->event_map.lpi_base = lpi_base;
2242 dev->event_map.nr_lpis = nr_lpis;
2243 mutex_init(&dev->event_map.vlpi_lock);
2244 dev->device_id = dev_id;
2245 INIT_LIST_HEAD(&dev->entry);
2247 raw_spin_lock_irqsave(&its->lock, flags);
2248 list_add(&dev->entry, &its->its_device_list);
2249 raw_spin_unlock_irqrestore(&its->lock, flags);
2251 /* Map device to its ITT */
2252 its_send_mapd(dev, 1);
2257 static void its_free_device(struct its_device *its_dev)
2259 unsigned long flags;
2261 raw_spin_lock_irqsave(&its_dev->its->lock, flags);
2262 list_del(&its_dev->entry);
2263 raw_spin_unlock_irqrestore(&its_dev->its->lock, flags);
2264 kfree(its_dev->itt);
2268 static int its_alloc_device_irq(struct its_device *dev, irq_hw_number_t *hwirq)
2272 idx = find_first_zero_bit(dev->event_map.lpi_map,
2273 dev->event_map.nr_lpis);
2274 if (idx == dev->event_map.nr_lpis)
2277 *hwirq = dev->event_map.lpi_base + idx;
2278 set_bit(idx, dev->event_map.lpi_map);
2283 static int its_msi_prepare(struct irq_domain *domain, struct device *dev,
2284 int nvec, msi_alloc_info_t *info)
2286 struct its_node *its;
2287 struct its_device *its_dev;
2288 struct msi_domain_info *msi_info;
2292 * We ignore "dev" entierely, and rely on the dev_id that has
2293 * been passed via the scratchpad. This limits this domain's
2294 * usefulness to upper layers that definitely know that they
2295 * are built on top of the ITS.
2297 dev_id = info->scratchpad[0].ul;
2299 msi_info = msi_get_domain_info(domain);
2300 its = msi_info->data;
2302 if (!gic_rdists->has_direct_lpi &&
2304 vpe_proxy.dev->its == its &&
2305 dev_id == vpe_proxy.dev->device_id) {
2306 /* Bad luck. Get yourself a better implementation */
2307 WARN_ONCE(1, "DevId %x clashes with GICv4 VPE proxy device\n",
2312 its_dev = its_find_device(its, dev_id);
2315 * We already have seen this ID, probably through
2316 * another alias (PCI bridge of some sort). No need to
2317 * create the device.
2319 pr_debug("Reusing ITT for devID %x\n", dev_id);
2323 its_dev = its_create_device(its, dev_id, nvec, true);
2327 pr_debug("ITT %d entries, %d bits\n", nvec, ilog2(nvec));
2329 info->scratchpad[0].ptr = its_dev;
2333 static struct msi_domain_ops its_msi_domain_ops = {
2334 .msi_prepare = its_msi_prepare,
2337 static int its_irq_gic_domain_alloc(struct irq_domain *domain,
2339 irq_hw_number_t hwirq)
2341 struct irq_fwspec fwspec;
2343 if (irq_domain_get_of_node(domain->parent)) {
2344 fwspec.fwnode = domain->parent->fwnode;
2345 fwspec.param_count = 3;
2346 fwspec.param[0] = GIC_IRQ_TYPE_LPI;
2347 fwspec.param[1] = hwirq;
2348 fwspec.param[2] = IRQ_TYPE_EDGE_RISING;
2349 } else if (is_fwnode_irqchip(domain->parent->fwnode)) {
2350 fwspec.fwnode = domain->parent->fwnode;
2351 fwspec.param_count = 2;
2352 fwspec.param[0] = hwirq;
2353 fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
2358 return irq_domain_alloc_irqs_parent(domain, virq, 1, &fwspec);
2361 static int its_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
2362 unsigned int nr_irqs, void *args)
2364 msi_alloc_info_t *info = args;
2365 struct its_device *its_dev = info->scratchpad[0].ptr;
2366 irq_hw_number_t hwirq;
2370 for (i = 0; i < nr_irqs; i++) {
2371 err = its_alloc_device_irq(its_dev, &hwirq);
2375 err = its_irq_gic_domain_alloc(domain, virq + i, hwirq);
2379 irq_domain_set_hwirq_and_chip(domain, virq + i,
2380 hwirq, &its_irq_chip, its_dev);
2381 irqd_set_single_target(irq_desc_get_irq_data(irq_to_desc(virq + i)));
2382 pr_debug("ID:%d pID:%d vID:%d\n",
2383 (int)(hwirq - its_dev->event_map.lpi_base),
2384 (int) hwirq, virq + i);
2390 static int its_irq_domain_activate(struct irq_domain *domain,
2391 struct irq_data *d, bool reserve)
2393 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2394 u32 event = its_get_event_id(d);
2395 const struct cpumask *cpu_mask = cpu_online_mask;
2398 /* get the cpu_mask of local node */
2399 if (its_dev->its->numa_node >= 0)
2400 cpu_mask = cpumask_of_node(its_dev->its->numa_node);
2402 /* Bind the LPI to the first possible CPU */
2403 cpu = cpumask_first_and(cpu_mask, cpu_online_mask);
2404 if (cpu >= nr_cpu_ids) {
2405 if (its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144)
2408 cpu = cpumask_first(cpu_online_mask);
2411 its_dev->event_map.col_map[event] = cpu;
2412 irq_data_update_effective_affinity(d, cpumask_of(cpu));
2414 /* Map the GIC IRQ and event to the device */
2415 its_send_mapti(its_dev, d->hwirq, event);
2419 static void its_irq_domain_deactivate(struct irq_domain *domain,
2422 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2423 u32 event = its_get_event_id(d);
2425 /* Stop the delivery of interrupts */
2426 its_send_discard(its_dev, event);
2429 static void its_irq_domain_free(struct irq_domain *domain, unsigned int virq,
2430 unsigned int nr_irqs)
2432 struct irq_data *d = irq_domain_get_irq_data(domain, virq);
2433 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2436 for (i = 0; i < nr_irqs; i++) {
2437 struct irq_data *data = irq_domain_get_irq_data(domain,
2439 u32 event = its_get_event_id(data);
2441 /* Mark interrupt index as unused */
2442 clear_bit(event, its_dev->event_map.lpi_map);
2444 /* Nuke the entry in the domain */
2445 irq_domain_reset_irq_data(data);
2448 /* If all interrupts have been freed, start mopping the floor */
2449 if (bitmap_empty(its_dev->event_map.lpi_map,
2450 its_dev->event_map.nr_lpis)) {
2451 its_lpi_free(its_dev->event_map.lpi_map,
2452 its_dev->event_map.lpi_base,
2453 its_dev->event_map.nr_lpis);
2454 kfree(its_dev->event_map.col_map);
2456 /* Unmap device/itt */
2457 its_send_mapd(its_dev, 0);
2458 its_free_device(its_dev);
2461 irq_domain_free_irqs_parent(domain, virq, nr_irqs);
2464 static const struct irq_domain_ops its_domain_ops = {
2465 .alloc = its_irq_domain_alloc,
2466 .free = its_irq_domain_free,
2467 .activate = its_irq_domain_activate,
2468 .deactivate = its_irq_domain_deactivate,
2474 * If a GICv4 doesn't implement Direct LPIs (which is extremely
2475 * likely), the only way to perform an invalidate is to use a fake
2476 * device to issue an INV command, implying that the LPI has first
2477 * been mapped to some event on that device. Since this is not exactly
2478 * cheap, we try to keep that mapping around as long as possible, and
2479 * only issue an UNMAP if we're short on available slots.
2481 * Broken by design(tm).
2483 static void its_vpe_db_proxy_unmap_locked(struct its_vpe *vpe)
2485 /* Already unmapped? */
2486 if (vpe->vpe_proxy_event == -1)
2489 its_send_discard(vpe_proxy.dev, vpe->vpe_proxy_event);
2490 vpe_proxy.vpes[vpe->vpe_proxy_event] = NULL;
2493 * We don't track empty slots at all, so let's move the
2494 * next_victim pointer if we can quickly reuse that slot
2495 * instead of nuking an existing entry. Not clear that this is
2496 * always a win though, and this might just generate a ripple
2497 * effect... Let's just hope VPEs don't migrate too often.
2499 if (vpe_proxy.vpes[vpe_proxy.next_victim])
2500 vpe_proxy.next_victim = vpe->vpe_proxy_event;
2502 vpe->vpe_proxy_event = -1;
2505 static void its_vpe_db_proxy_unmap(struct its_vpe *vpe)
2507 if (!gic_rdists->has_direct_lpi) {
2508 unsigned long flags;
2510 raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2511 its_vpe_db_proxy_unmap_locked(vpe);
2512 raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2516 static void its_vpe_db_proxy_map_locked(struct its_vpe *vpe)
2518 /* Already mapped? */
2519 if (vpe->vpe_proxy_event != -1)
2522 /* This slot was already allocated. Kick the other VPE out. */
2523 if (vpe_proxy.vpes[vpe_proxy.next_victim])
2524 its_vpe_db_proxy_unmap_locked(vpe_proxy.vpes[vpe_proxy.next_victim]);
2526 /* Map the new VPE instead */
2527 vpe_proxy.vpes[vpe_proxy.next_victim] = vpe;
2528 vpe->vpe_proxy_event = vpe_proxy.next_victim;
2529 vpe_proxy.next_victim = (vpe_proxy.next_victim + 1) % vpe_proxy.dev->nr_ites;
2531 vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = vpe->col_idx;
2532 its_send_mapti(vpe_proxy.dev, vpe->vpe_db_lpi, vpe->vpe_proxy_event);
2535 static void its_vpe_db_proxy_move(struct its_vpe *vpe, int from, int to)
2537 unsigned long flags;
2538 struct its_collection *target_col;
2540 if (gic_rdists->has_direct_lpi) {
2541 void __iomem *rdbase;
2543 rdbase = per_cpu_ptr(gic_rdists->rdist, from)->rd_base;
2544 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
2545 while (gic_read_lpir(rdbase + GICR_SYNCR) & 1)
2551 raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2553 its_vpe_db_proxy_map_locked(vpe);
2555 target_col = &vpe_proxy.dev->its->collections[to];
2556 its_send_movi(vpe_proxy.dev, target_col, vpe->vpe_proxy_event);
2557 vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = to;
2559 raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2562 static int its_vpe_set_affinity(struct irq_data *d,
2563 const struct cpumask *mask_val,
2566 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2567 int cpu = cpumask_first(mask_val);
2570 * Changing affinity is mega expensive, so let's be as lazy as
2571 * we can and only do it if we really have to. Also, if mapped
2572 * into the proxy device, we need to move the doorbell
2573 * interrupt to its new location.
2575 if (vpe->col_idx != cpu) {
2576 int from = vpe->col_idx;
2579 its_send_vmovp(vpe);
2580 its_vpe_db_proxy_move(vpe, from, cpu);
2583 irq_data_update_effective_affinity(d, cpumask_of(cpu));
2585 return IRQ_SET_MASK_OK_DONE;
2588 static void its_vpe_schedule(struct its_vpe *vpe)
2590 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2593 /* Schedule the VPE */
2594 val = virt_to_phys(page_address(vpe->its_vm->vprop_page)) &
2595 GENMASK_ULL(51, 12);
2596 val |= (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
2597 val |= GICR_VPROPBASER_RaWb;
2598 val |= GICR_VPROPBASER_InnerShareable;
2599 gits_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2601 val = virt_to_phys(page_address(vpe->vpt_page)) &
2602 GENMASK_ULL(51, 16);
2603 val |= GICR_VPENDBASER_RaWaWb;
2604 val |= GICR_VPENDBASER_NonShareable;
2606 * There is no good way of finding out if the pending table is
2607 * empty as we can race against the doorbell interrupt very
2608 * easily. So in the end, vpe->pending_last is only an
2609 * indication that the vcpu has something pending, not one
2610 * that the pending table is empty. A good implementation
2611 * would be able to read its coarse map pretty quickly anyway,
2612 * making this a tolerable issue.
2614 val |= GICR_VPENDBASER_PendingLast;
2615 val |= vpe->idai ? GICR_VPENDBASER_IDAI : 0;
2616 val |= GICR_VPENDBASER_Valid;
2617 gits_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
2620 static void its_vpe_deschedule(struct its_vpe *vpe)
2622 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2623 u32 count = 1000000; /* 1s! */
2627 /* We're being scheduled out */
2628 val = gits_read_vpendbaser(vlpi_base + GICR_VPENDBASER);
2629 val &= ~GICR_VPENDBASER_Valid;
2630 gits_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
2633 val = gits_read_vpendbaser(vlpi_base + GICR_VPENDBASER);
2634 clean = !(val & GICR_VPENDBASER_Dirty);
2640 } while (!clean && count);
2642 if (unlikely(!clean && !count)) {
2643 pr_err_ratelimited("ITS virtual pending table not cleaning\n");
2645 vpe->pending_last = true;
2647 vpe->idai = !!(val & GICR_VPENDBASER_IDAI);
2648 vpe->pending_last = !!(val & GICR_VPENDBASER_PendingLast);
2652 static void its_vpe_invall(struct its_vpe *vpe)
2654 struct its_node *its;
2656 list_for_each_entry(its, &its_nodes, entry) {
2660 if (its_list_map && !vpe->its_vm->vlpi_count[its->list_nr])
2664 * Sending a VINVALL to a single ITS is enough, as all
2665 * we need is to reach the redistributors.
2667 its_send_vinvall(its, vpe);
2672 static int its_vpe_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
2674 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2675 struct its_cmd_info *info = vcpu_info;
2677 switch (info->cmd_type) {
2679 its_vpe_schedule(vpe);
2682 case DESCHEDULE_VPE:
2683 its_vpe_deschedule(vpe);
2687 its_vpe_invall(vpe);
2695 static void its_vpe_send_cmd(struct its_vpe *vpe,
2696 void (*cmd)(struct its_device *, u32))
2698 unsigned long flags;
2700 raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2702 its_vpe_db_proxy_map_locked(vpe);
2703 cmd(vpe_proxy.dev, vpe->vpe_proxy_event);
2705 raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2708 static void its_vpe_send_inv(struct irq_data *d)
2710 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2712 if (gic_rdists->has_direct_lpi) {
2713 void __iomem *rdbase;
2715 rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
2716 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_INVLPIR);
2717 while (gic_read_lpir(rdbase + GICR_SYNCR) & 1)
2720 its_vpe_send_cmd(vpe, its_send_inv);
2724 static void its_vpe_mask_irq(struct irq_data *d)
2727 * We need to unmask the LPI, which is described by the parent
2728 * irq_data. Instead of calling into the parent (which won't
2729 * exactly do the right thing, let's simply use the
2730 * parent_data pointer. Yes, I'm naughty.
2732 lpi_write_config(d->parent_data, LPI_PROP_ENABLED, 0);
2733 its_vpe_send_inv(d);
2736 static void its_vpe_unmask_irq(struct irq_data *d)
2738 /* Same hack as above... */
2739 lpi_write_config(d->parent_data, 0, LPI_PROP_ENABLED);
2740 its_vpe_send_inv(d);
2743 static int its_vpe_set_irqchip_state(struct irq_data *d,
2744 enum irqchip_irq_state which,
2747 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2749 if (which != IRQCHIP_STATE_PENDING)
2752 if (gic_rdists->has_direct_lpi) {
2753 void __iomem *rdbase;
2755 rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
2757 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_SETLPIR);
2759 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
2760 while (gic_read_lpir(rdbase + GICR_SYNCR) & 1)
2765 its_vpe_send_cmd(vpe, its_send_int);
2767 its_vpe_send_cmd(vpe, its_send_clear);
2773 static struct irq_chip its_vpe_irq_chip = {
2774 .name = "GICv4-vpe",
2775 .irq_mask = its_vpe_mask_irq,
2776 .irq_unmask = its_vpe_unmask_irq,
2777 .irq_eoi = irq_chip_eoi_parent,
2778 .irq_set_affinity = its_vpe_set_affinity,
2779 .irq_set_irqchip_state = its_vpe_set_irqchip_state,
2780 .irq_set_vcpu_affinity = its_vpe_set_vcpu_affinity,
2783 static int its_vpe_id_alloc(void)
2785 return ida_simple_get(&its_vpeid_ida, 0, ITS_MAX_VPEID, GFP_KERNEL);
2788 static void its_vpe_id_free(u16 id)
2790 ida_simple_remove(&its_vpeid_ida, id);
2793 static int its_vpe_init(struct its_vpe *vpe)
2795 struct page *vpt_page;
2798 /* Allocate vpe_id */
2799 vpe_id = its_vpe_id_alloc();
2804 vpt_page = its_allocate_pending_table(GFP_KERNEL);
2806 its_vpe_id_free(vpe_id);
2810 if (!its_alloc_vpe_table(vpe_id)) {
2811 its_vpe_id_free(vpe_id);
2812 its_free_pending_table(vpe->vpt_page);
2816 vpe->vpe_id = vpe_id;
2817 vpe->vpt_page = vpt_page;
2818 vpe->vpe_proxy_event = -1;
2823 static void its_vpe_teardown(struct its_vpe *vpe)
2825 its_vpe_db_proxy_unmap(vpe);
2826 its_vpe_id_free(vpe->vpe_id);
2827 its_free_pending_table(vpe->vpt_page);
2830 static void its_vpe_irq_domain_free(struct irq_domain *domain,
2832 unsigned int nr_irqs)
2834 struct its_vm *vm = domain->host_data;
2837 irq_domain_free_irqs_parent(domain, virq, nr_irqs);
2839 for (i = 0; i < nr_irqs; i++) {
2840 struct irq_data *data = irq_domain_get_irq_data(domain,
2842 struct its_vpe *vpe = irq_data_get_irq_chip_data(data);
2844 BUG_ON(vm != vpe->its_vm);
2846 clear_bit(data->hwirq, vm->db_bitmap);
2847 its_vpe_teardown(vpe);
2848 irq_domain_reset_irq_data(data);
2851 if (bitmap_empty(vm->db_bitmap, vm->nr_db_lpis)) {
2852 its_lpi_free(vm->db_bitmap, vm->db_lpi_base, vm->nr_db_lpis);
2853 its_free_prop_table(vm->vprop_page);
2857 static int its_vpe_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
2858 unsigned int nr_irqs, void *args)
2860 struct its_vm *vm = args;
2861 unsigned long *bitmap;
2862 struct page *vprop_page;
2863 int base, nr_ids, i, err = 0;
2867 bitmap = its_lpi_alloc(roundup_pow_of_two(nr_irqs), &base, &nr_ids);
2871 if (nr_ids < nr_irqs) {
2872 its_lpi_free(bitmap, base, nr_ids);
2876 vprop_page = its_allocate_prop_table(GFP_KERNEL);
2878 its_lpi_free(bitmap, base, nr_ids);
2882 vm->db_bitmap = bitmap;
2883 vm->db_lpi_base = base;
2884 vm->nr_db_lpis = nr_ids;
2885 vm->vprop_page = vprop_page;
2887 for (i = 0; i < nr_irqs; i++) {
2888 vm->vpes[i]->vpe_db_lpi = base + i;
2889 err = its_vpe_init(vm->vpes[i]);
2892 err = its_irq_gic_domain_alloc(domain, virq + i,
2893 vm->vpes[i]->vpe_db_lpi);
2896 irq_domain_set_hwirq_and_chip(domain, virq + i, i,
2897 &its_vpe_irq_chip, vm->vpes[i]);
2903 its_vpe_irq_domain_free(domain, virq, i - 1);
2905 its_lpi_free(bitmap, base, nr_ids);
2906 its_free_prop_table(vprop_page);
2912 static int its_vpe_irq_domain_activate(struct irq_domain *domain,
2913 struct irq_data *d, bool reserve)
2915 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2916 struct its_node *its;
2918 /* If we use the list map, we issue VMAPP on demand... */
2922 /* Map the VPE to the first possible CPU */
2923 vpe->col_idx = cpumask_first(cpu_online_mask);
2925 list_for_each_entry(its, &its_nodes, entry) {
2929 its_send_vmapp(its, vpe, true);
2930 its_send_vinvall(its, vpe);
2933 irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
2938 static void its_vpe_irq_domain_deactivate(struct irq_domain *domain,
2941 struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2942 struct its_node *its;
2945 * If we use the list map, we unmap the VPE once no VLPIs are
2946 * associated with the VM.
2951 list_for_each_entry(its, &its_nodes, entry) {
2955 its_send_vmapp(its, vpe, false);
2959 static const struct irq_domain_ops its_vpe_domain_ops = {
2960 .alloc = its_vpe_irq_domain_alloc,
2961 .free = its_vpe_irq_domain_free,
2962 .activate = its_vpe_irq_domain_activate,
2963 .deactivate = its_vpe_irq_domain_deactivate,
2966 static int its_force_quiescent(void __iomem *base)
2968 u32 count = 1000000; /* 1s */
2971 val = readl_relaxed(base + GITS_CTLR);
2973 * GIC architecture specification requires the ITS to be both
2974 * disabled and quiescent for writes to GITS_BASER<n> or
2975 * GITS_CBASER to not have UNPREDICTABLE results.
2977 if ((val & GITS_CTLR_QUIESCENT) && !(val & GITS_CTLR_ENABLE))
2980 /* Disable the generation of all interrupts to this ITS */
2981 val &= ~(GITS_CTLR_ENABLE | GITS_CTLR_ImDe);
2982 writel_relaxed(val, base + GITS_CTLR);
2984 /* Poll GITS_CTLR and wait until ITS becomes quiescent */
2986 val = readl_relaxed(base + GITS_CTLR);
2987 if (val & GITS_CTLR_QUIESCENT)
2999 static bool __maybe_unused its_enable_quirk_cavium_22375(void *data)
3001 struct its_node *its = data;
3003 /* erratum 22375: only alloc 8MB table size */
3004 its->device_ids = 0x14; /* 20 bits, 8MB */
3005 its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_22375;
3010 static bool __maybe_unused its_enable_quirk_cavium_23144(void *data)
3012 struct its_node *its = data;
3014 its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_23144;
3019 static bool __maybe_unused its_enable_quirk_qdf2400_e0065(void *data)
3021 struct its_node *its = data;
3023 /* On QDF2400, the size of the ITE is 16Bytes */
3029 static u64 its_irq_get_msi_base_pre_its(struct its_device *its_dev)
3031 struct its_node *its = its_dev->its;
3034 * The Socionext Synquacer SoC has a so-called 'pre-ITS',
3035 * which maps 32-bit writes targeted at a separate window of
3036 * size '4 << device_id_bits' onto writes to GITS_TRANSLATER
3037 * with device ID taken from bits [device_id_bits + 1:2] of
3038 * the window offset.
3040 return its->pre_its_base + (its_dev->device_id << 2);
3043 static bool __maybe_unused its_enable_quirk_socionext_synquacer(void *data)
3045 struct its_node *its = data;
3046 u32 pre_its_window[2];
3049 if (!fwnode_property_read_u32_array(its->fwnode_handle,
3050 "socionext,synquacer-pre-its",
3052 ARRAY_SIZE(pre_its_window))) {
3054 its->pre_its_base = pre_its_window[0];
3055 its->get_msi_base = its_irq_get_msi_base_pre_its;
3057 ids = ilog2(pre_its_window[1]) - 2;
3058 if (its->device_ids > ids)
3059 its->device_ids = ids;
3061 /* the pre-ITS breaks isolation, so disable MSI remapping */
3062 its->msi_domain_flags &= ~IRQ_DOMAIN_FLAG_MSI_REMAP;
3068 static bool __maybe_unused its_enable_quirk_hip07_161600802(void *data)
3070 struct its_node *its = data;
3073 * Hip07 insists on using the wrong address for the VLPI
3074 * page. Trick it into doing the right thing...
3076 its->vlpi_redist_offset = SZ_128K;
3080 static const struct gic_quirk its_quirks[] = {
3081 #ifdef CONFIG_CAVIUM_ERRATUM_22375
3083 .desc = "ITS: Cavium errata 22375, 24313",
3084 .iidr = 0xa100034c, /* ThunderX pass 1.x */
3086 .init = its_enable_quirk_cavium_22375,
3089 #ifdef CONFIG_CAVIUM_ERRATUM_23144
3091 .desc = "ITS: Cavium erratum 23144",
3092 .iidr = 0xa100034c, /* ThunderX pass 1.x */
3094 .init = its_enable_quirk_cavium_23144,
3097 #ifdef CONFIG_QCOM_QDF2400_ERRATUM_0065
3099 .desc = "ITS: QDF2400 erratum 0065",
3100 .iidr = 0x00001070, /* QDF2400 ITS rev 1.x */
3102 .init = its_enable_quirk_qdf2400_e0065,
3105 #ifdef CONFIG_SOCIONEXT_SYNQUACER_PREITS
3108 * The Socionext Synquacer SoC incorporates ARM's own GIC-500
3109 * implementation, but with a 'pre-ITS' added that requires
3110 * special handling in software.
3112 .desc = "ITS: Socionext Synquacer pre-ITS",
3115 .init = its_enable_quirk_socionext_synquacer,
3118 #ifdef CONFIG_HISILICON_ERRATUM_161600802
3120 .desc = "ITS: Hip07 erratum 161600802",
3123 .init = its_enable_quirk_hip07_161600802,
3130 static void its_enable_quirks(struct its_node *its)
3132 u32 iidr = readl_relaxed(its->base + GITS_IIDR);
3134 gic_enable_quirks(iidr, its_quirks, its);
3137 static int its_save_disable(void)
3139 struct its_node *its;
3142 raw_spin_lock(&its_lock);
3143 list_for_each_entry(its, &its_nodes, entry) {
3146 if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3150 its->ctlr_save = readl_relaxed(base + GITS_CTLR);
3151 err = its_force_quiescent(base);
3153 pr_err("ITS@%pa: failed to quiesce: %d\n",
3154 &its->phys_base, err);
3155 writel_relaxed(its->ctlr_save, base + GITS_CTLR);
3159 its->cbaser_save = gits_read_cbaser(base + GITS_CBASER);
3164 list_for_each_entry_continue_reverse(its, &its_nodes, entry) {
3167 if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3171 writel_relaxed(its->ctlr_save, base + GITS_CTLR);
3174 raw_spin_unlock(&its_lock);
3179 static void its_restore_enable(void)
3181 struct its_node *its;
3184 raw_spin_lock(&its_lock);
3185 list_for_each_entry(its, &its_nodes, entry) {
3189 if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3195 * Make sure that the ITS is disabled. If it fails to quiesce,
3196 * don't restore it since writing to CBASER or BASER<n>
3197 * registers is undefined according to the GIC v3 ITS
3200 ret = its_force_quiescent(base);
3202 pr_err("ITS@%pa: failed to quiesce on resume: %d\n",
3203 &its->phys_base, ret);
3207 gits_write_cbaser(its->cbaser_save, base + GITS_CBASER);
3210 * Writing CBASER resets CREADR to 0, so make CWRITER and
3211 * cmd_write line up with it.
3213 its->cmd_write = its->cmd_base;
3214 gits_write_cwriter(0, base + GITS_CWRITER);
3216 /* Restore GITS_BASER from the value cache. */
3217 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
3218 struct its_baser *baser = &its->tables[i];
3220 if (!(baser->val & GITS_BASER_VALID))
3223 its_write_baser(its, baser, baser->val);
3225 writel_relaxed(its->ctlr_save, base + GITS_CTLR);
3228 * Reinit the collection if it's stored in the ITS. This is
3229 * indicated by the col_id being less than the HCC field.
3230 * CID < HCC as specified in the GIC v3 Documentation.
3232 if (its->collections[smp_processor_id()].col_id <
3233 GITS_TYPER_HCC(gic_read_typer(base + GITS_TYPER)))
3234 its_cpu_init_collection(its);
3236 raw_spin_unlock(&its_lock);
3239 static struct syscore_ops its_syscore_ops = {
3240 .suspend = its_save_disable,
3241 .resume = its_restore_enable,
3244 static int its_init_domain(struct fwnode_handle *handle, struct its_node *its)
3246 struct irq_domain *inner_domain;
3247 struct msi_domain_info *info;
3249 info = kzalloc(sizeof(*info), GFP_KERNEL);
3253 inner_domain = irq_domain_create_tree(handle, &its_domain_ops, its);
3254 if (!inner_domain) {
3259 inner_domain->parent = its_parent;
3260 irq_domain_update_bus_token(inner_domain, DOMAIN_BUS_NEXUS);
3261 inner_domain->flags |= its->msi_domain_flags;
3262 info->ops = &its_msi_domain_ops;
3264 inner_domain->host_data = info;
3269 static int its_init_vpe_domain(void)
3271 struct its_node *its;
3275 if (gic_rdists->has_direct_lpi) {
3276 pr_info("ITS: Using DirectLPI for VPE invalidation\n");
3280 /* Any ITS will do, even if not v4 */
3281 its = list_first_entry(&its_nodes, struct its_node, entry);
3283 entries = roundup_pow_of_two(nr_cpu_ids);
3284 vpe_proxy.vpes = kcalloc(entries, sizeof(*vpe_proxy.vpes),
3286 if (!vpe_proxy.vpes) {
3287 pr_err("ITS: Can't allocate GICv4 proxy device array\n");
3291 /* Use the last possible DevID */
3292 devid = GENMASK(its->device_ids - 1, 0);
3293 vpe_proxy.dev = its_create_device(its, devid, entries, false);
3294 if (!vpe_proxy.dev) {
3295 kfree(vpe_proxy.vpes);
3296 pr_err("ITS: Can't allocate GICv4 proxy device\n");
3300 BUG_ON(entries > vpe_proxy.dev->nr_ites);
3302 raw_spin_lock_init(&vpe_proxy.lock);
3303 vpe_proxy.next_victim = 0;
3304 pr_info("ITS: Allocated DevID %x as GICv4 proxy device (%d slots)\n",
3305 devid, vpe_proxy.dev->nr_ites);
3310 static int __init its_compute_its_list_map(struct resource *res,
3311 void __iomem *its_base)
3317 * This is assumed to be done early enough that we're
3318 * guaranteed to be single-threaded, hence no
3319 * locking. Should this change, we should address
3322 its_number = find_first_zero_bit(&its_list_map, GICv4_ITS_LIST_MAX);
3323 if (its_number >= GICv4_ITS_LIST_MAX) {
3324 pr_err("ITS@%pa: No ITSList entry available!\n",
3329 ctlr = readl_relaxed(its_base + GITS_CTLR);
3330 ctlr &= ~GITS_CTLR_ITS_NUMBER;
3331 ctlr |= its_number << GITS_CTLR_ITS_NUMBER_SHIFT;
3332 writel_relaxed(ctlr, its_base + GITS_CTLR);
3333 ctlr = readl_relaxed(its_base + GITS_CTLR);
3334 if ((ctlr & GITS_CTLR_ITS_NUMBER) != (its_number << GITS_CTLR_ITS_NUMBER_SHIFT)) {
3335 its_number = ctlr & GITS_CTLR_ITS_NUMBER;
3336 its_number >>= GITS_CTLR_ITS_NUMBER_SHIFT;
3339 if (test_and_set_bit(its_number, &its_list_map)) {
3340 pr_err("ITS@%pa: Duplicate ITSList entry %d\n",
3341 &res->start, its_number);
3348 static int __init its_probe_one(struct resource *res,
3349 struct fwnode_handle *handle, int numa_node)
3351 struct its_node *its;
3352 void __iomem *its_base;
3354 u64 baser, tmp, typer;
3357 its_base = ioremap(res->start, resource_size(res));
3359 pr_warn("ITS@%pa: Unable to map ITS registers\n", &res->start);
3363 val = readl_relaxed(its_base + GITS_PIDR2) & GIC_PIDR2_ARCH_MASK;
3364 if (val != 0x30 && val != 0x40) {
3365 pr_warn("ITS@%pa: No ITS detected, giving up\n", &res->start);
3370 err = its_force_quiescent(its_base);
3372 pr_warn("ITS@%pa: Failed to quiesce, giving up\n", &res->start);
3376 pr_info("ITS %pR\n", res);
3378 its = kzalloc(sizeof(*its), GFP_KERNEL);
3384 raw_spin_lock_init(&its->lock);
3385 INIT_LIST_HEAD(&its->entry);
3386 INIT_LIST_HEAD(&its->its_device_list);
3387 typer = gic_read_typer(its_base + GITS_TYPER);
3388 its->base = its_base;
3389 its->phys_base = res->start;
3390 its->ite_size = GITS_TYPER_ITT_ENTRY_SIZE(typer);
3391 its->device_ids = GITS_TYPER_DEVBITS(typer);
3392 its->is_v4 = !!(typer & GITS_TYPER_VLPIS);
3394 if (!(typer & GITS_TYPER_VMOVP)) {
3395 err = its_compute_its_list_map(res, its_base);
3401 pr_info("ITS@%pa: Using ITS number %d\n",
3404 pr_info("ITS@%pa: Single VMOVP capable\n", &res->start);
3408 its->numa_node = numa_node;
3410 its->cmd_base = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
3411 get_order(ITS_CMD_QUEUE_SZ));
3412 if (!its->cmd_base) {
3416 its->cmd_write = its->cmd_base;
3417 its->fwnode_handle = handle;
3418 its->get_msi_base = its_irq_get_msi_base;
3419 its->msi_domain_flags = IRQ_DOMAIN_FLAG_MSI_REMAP;
3421 its_enable_quirks(its);
3423 err = its_alloc_tables(its);
3427 err = its_alloc_collections(its);
3429 goto out_free_tables;
3431 baser = (virt_to_phys(its->cmd_base) |
3432 GITS_CBASER_RaWaWb |
3433 GITS_CBASER_InnerShareable |
3434 (ITS_CMD_QUEUE_SZ / SZ_4K - 1) |
3437 gits_write_cbaser(baser, its->base + GITS_CBASER);
3438 tmp = gits_read_cbaser(its->base + GITS_CBASER);
3440 if ((tmp ^ baser) & GITS_CBASER_SHAREABILITY_MASK) {
3441 if (!(tmp & GITS_CBASER_SHAREABILITY_MASK)) {
3443 * The HW reports non-shareable, we must
3444 * remove the cacheability attributes as
3447 baser &= ~(GITS_CBASER_SHAREABILITY_MASK |
3448 GITS_CBASER_CACHEABILITY_MASK);
3449 baser |= GITS_CBASER_nC;
3450 gits_write_cbaser(baser, its->base + GITS_CBASER);
3452 pr_info("ITS: using cache flushing for cmd queue\n");
3453 its->flags |= ITS_FLAGS_CMDQ_NEEDS_FLUSHING;
3456 gits_write_cwriter(0, its->base + GITS_CWRITER);
3457 ctlr = readl_relaxed(its->base + GITS_CTLR);
3458 ctlr |= GITS_CTLR_ENABLE;
3460 ctlr |= GITS_CTLR_ImDe;
3461 writel_relaxed(ctlr, its->base + GITS_CTLR);
3463 if (GITS_TYPER_HCC(typer))
3464 its->flags |= ITS_FLAGS_SAVE_SUSPEND_STATE;
3466 err = its_init_domain(handle, its);
3468 goto out_free_tables;
3470 raw_spin_lock(&its_lock);
3471 list_add(&its->entry, &its_nodes);
3472 raw_spin_unlock(&its_lock);
3477 its_free_tables(its);
3479 free_pages((unsigned long)its->cmd_base, get_order(ITS_CMD_QUEUE_SZ));
3484 pr_err("ITS@%pa: failed probing (%d)\n", &res->start, err);
3488 static bool gic_rdists_supports_plpis(void)
3490 return !!(gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER) & GICR_TYPER_PLPIS);
3493 static int redist_disable_lpis(void)
3495 void __iomem *rbase = gic_data_rdist_rd_base();
3496 u64 timeout = USEC_PER_SEC;
3500 * If coming via a CPU hotplug event, we don't need to disable
3501 * LPIs before trying to re-enable them. They are already
3502 * configured and all is well in the world. Detect this case
3503 * by checking the allocation of the pending table for the
3506 if (gic_data_rdist()->pend_page)
3509 if (!gic_rdists_supports_plpis()) {
3510 pr_info("CPU%d: LPIs not supported\n", smp_processor_id());
3514 val = readl_relaxed(rbase + GICR_CTLR);
3515 if (!(val & GICR_CTLR_ENABLE_LPIS))
3518 pr_warn("CPU%d: Booted with LPIs enabled, memory probably corrupted\n",
3519 smp_processor_id());
3520 add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
3523 val &= ~GICR_CTLR_ENABLE_LPIS;
3524 writel_relaxed(val, rbase + GICR_CTLR);
3526 /* Make sure any change to GICR_CTLR is observable by the GIC */
3530 * Software must observe RWP==0 after clearing GICR_CTLR.EnableLPIs
3531 * from 1 to 0 before programming GICR_PEND{PROP}BASER registers.
3532 * Error out if we time out waiting for RWP to clear.
3534 while (readl_relaxed(rbase + GICR_CTLR) & GICR_CTLR_RWP) {
3536 pr_err("CPU%d: Timeout while disabling LPIs\n",
3537 smp_processor_id());
3545 * After it has been written to 1, it is IMPLEMENTATION
3546 * DEFINED whether GICR_CTLR.EnableLPI becomes RES1 or can be
3547 * cleared to 0. Error out if clearing the bit failed.
3549 if (readl_relaxed(rbase + GICR_CTLR) & GICR_CTLR_ENABLE_LPIS) {
3550 pr_err("CPU%d: Failed to disable LPIs\n", smp_processor_id());
3557 int its_cpu_init(void)
3559 if (!list_empty(&its_nodes)) {
3562 ret = redist_disable_lpis();
3566 its_cpu_init_lpis();
3567 its_cpu_init_collections();
3573 static const struct of_device_id its_device_id[] = {
3574 { .compatible = "arm,gic-v3-its", },
3578 static int __init its_of_probe(struct device_node *node)
3580 struct device_node *np;
3581 struct resource res;
3583 for (np = of_find_matching_node(node, its_device_id); np;
3584 np = of_find_matching_node(np, its_device_id)) {
3585 if (!of_device_is_available(np))
3587 if (!of_property_read_bool(np, "msi-controller")) {
3588 pr_warn("%pOF: no msi-controller property, ITS ignored\n",
3593 if (of_address_to_resource(np, 0, &res)) {
3594 pr_warn("%pOF: no regs?\n", np);
3598 its_probe_one(&res, &np->fwnode, of_node_to_nid(np));
3605 #define ACPI_GICV3_ITS_MEM_SIZE (SZ_128K)
3607 #ifdef CONFIG_ACPI_NUMA
3608 struct its_srat_map {
3615 static struct its_srat_map *its_srat_maps __initdata;
3616 static int its_in_srat __initdata;
3618 static int __init acpi_get_its_numa_node(u32 its_id)
3622 for (i = 0; i < its_in_srat; i++) {
3623 if (its_id == its_srat_maps[i].its_id)
3624 return its_srat_maps[i].numa_node;
3626 return NUMA_NO_NODE;
3629 static int __init gic_acpi_match_srat_its(struct acpi_subtable_header *header,
3630 const unsigned long end)
3635 static int __init gic_acpi_parse_srat_its(struct acpi_subtable_header *header,
3636 const unsigned long end)
3639 struct acpi_srat_gic_its_affinity *its_affinity;
3641 its_affinity = (struct acpi_srat_gic_its_affinity *)header;
3645 if (its_affinity->header.length < sizeof(*its_affinity)) {
3646 pr_err("SRAT: Invalid header length %d in ITS affinity\n",
3647 its_affinity->header.length);
3651 node = acpi_map_pxm_to_node(its_affinity->proximity_domain);
3653 if (node == NUMA_NO_NODE || node >= MAX_NUMNODES) {
3654 pr_err("SRAT: Invalid NUMA node %d in ITS affinity\n", node);
3658 its_srat_maps[its_in_srat].numa_node = node;
3659 its_srat_maps[its_in_srat].its_id = its_affinity->its_id;
3661 pr_info("SRAT: PXM %d -> ITS %d -> Node %d\n",
3662 its_affinity->proximity_domain, its_affinity->its_id, node);
3667 static void __init acpi_table_parse_srat_its(void)
3671 count = acpi_table_parse_entries(ACPI_SIG_SRAT,
3672 sizeof(struct acpi_table_srat),
3673 ACPI_SRAT_TYPE_GIC_ITS_AFFINITY,
3674 gic_acpi_match_srat_its, 0);
3678 its_srat_maps = kmalloc_array(count, sizeof(struct its_srat_map),
3680 if (!its_srat_maps) {
3681 pr_warn("SRAT: Failed to allocate memory for its_srat_maps!\n");
3685 acpi_table_parse_entries(ACPI_SIG_SRAT,
3686 sizeof(struct acpi_table_srat),
3687 ACPI_SRAT_TYPE_GIC_ITS_AFFINITY,
3688 gic_acpi_parse_srat_its, 0);
3691 /* free the its_srat_maps after ITS probing */
3692 static void __init acpi_its_srat_maps_free(void)
3694 kfree(its_srat_maps);
3697 static void __init acpi_table_parse_srat_its(void) { }
3698 static int __init acpi_get_its_numa_node(u32 its_id) { return NUMA_NO_NODE; }
3699 static void __init acpi_its_srat_maps_free(void) { }
3702 static int __init gic_acpi_parse_madt_its(struct acpi_subtable_header *header,
3703 const unsigned long end)
3705 struct acpi_madt_generic_translator *its_entry;
3706 struct fwnode_handle *dom_handle;
3707 struct resource res;
3710 its_entry = (struct acpi_madt_generic_translator *)header;
3711 memset(&res, 0, sizeof(res));
3712 res.start = its_entry->base_address;
3713 res.end = its_entry->base_address + ACPI_GICV3_ITS_MEM_SIZE - 1;
3714 res.flags = IORESOURCE_MEM;
3716 dom_handle = irq_domain_alloc_fwnode((void *)its_entry->base_address);
3718 pr_err("ITS@%pa: Unable to allocate GICv3 ITS domain token\n",
3723 err = iort_register_domain_token(its_entry->translation_id, res.start,
3726 pr_err("ITS@%pa: Unable to register GICv3 ITS domain token (ITS ID %d) to IORT\n",
3727 &res.start, its_entry->translation_id);
3731 err = its_probe_one(&res, dom_handle,
3732 acpi_get_its_numa_node(its_entry->translation_id));
3736 iort_deregister_domain_token(its_entry->translation_id);
3738 irq_domain_free_fwnode(dom_handle);
3742 static void __init its_acpi_probe(void)
3744 acpi_table_parse_srat_its();
3745 acpi_table_parse_madt(ACPI_MADT_TYPE_GENERIC_TRANSLATOR,
3746 gic_acpi_parse_madt_its, 0);
3747 acpi_its_srat_maps_free();
3750 static void __init its_acpi_probe(void) { }
3753 int __init its_init(struct fwnode_handle *handle, struct rdists *rdists,
3754 struct irq_domain *parent_domain)
3756 struct device_node *of_node;
3757 struct its_node *its;
3758 bool has_v4 = false;
3761 its_parent = parent_domain;
3762 of_node = to_of_node(handle);
3764 its_of_probe(of_node);
3768 if (list_empty(&its_nodes)) {
3769 pr_warn("ITS: No ITS available, not enabling LPIs\n");
3773 gic_rdists = rdists;
3774 err = its_alloc_lpi_tables();
3778 list_for_each_entry(its, &its_nodes, entry)
3779 has_v4 |= its->is_v4;
3781 if (has_v4 & rdists->has_vlpis) {
3782 if (its_init_vpe_domain() ||
3783 its_init_v4(parent_domain, &its_vpe_domain_ops)) {
3784 rdists->has_vlpis = false;
3785 pr_err("ITS: Disabling GICv4 support\n");
3789 register_syscore_ops(&its_syscore_ops);