drivers/irqchip/irq-gic-v3-its.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2013-2017 ARM Limited, All Rights Reserved.
   4  * Author: Marc Zyngier <marc.zyngier@arm.com>
   5  */
   6
   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/irqdomain.h>
  18 #include <linux/list.h>
  19 #include <linux/log2.h>
  20 #include <linux/memblock.h>
  21 #include <linux/mm.h>
  22 #include <linux/msi.h>
  23 #include <linux/of.h>
  24 #include <linux/of_address.h>
  25 #include <linux/of_irq.h>
  26 #include <linux/of_pci.h>
  27 #include <linux/of_platform.h>
  28 #include <linux/percpu.h>
  29 #include <linux/slab.h>
  30 #include <linux/syscore_ops.h>
  31
  32 #include <linux/irqchip.h>
  33 #include <linux/irqchip/arm-gic-v3.h>
  34 #include <linux/irqchip/arm-gic-v4.h>
  35
  36 #include <asm/cputype.h>
  37 #include <asm/exception.h>
  38
  39 #include "irq-gic-common.h"
  40
  41 #define ITS_FLAGS_CMDQ_NEEDS_FLUSHING           (1ULL << 0)
  42 #define ITS_FLAGS_WORKAROUND_CAVIUM_22375       (1ULL << 1)
  43 #define ITS_FLAGS_WORKAROUND_CAVIUM_23144       (1ULL << 2)
  44 #define ITS_FLAGS_SAVE_SUSPEND_STATE            (1ULL << 3)
  45
  46 #define RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING     (1 << 0)
  47 #define RDIST_FLAGS_RD_TABLES_PREALLOCATED      (1 << 1)
  48
  49 static u32 lpi_id_bits;
  50
  51 /*
  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).
  55  */
  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)
  59
  60 #define LPI_PROP_DEFAULT_PRIO   GICD_INT_DEF_PRI
  61
  62 /*
  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
  65  * CPU.
  66  */
  67 struct its_collection {
  68         u64                     target_address;
  69         u16                     col_id;
  70 };
  71
  72 /*
  73  * The ITS_BASER structure - contains memory information, cached
  74  * value of BASER register configuration and ITS page size.
  75  */
  76 struct its_baser {
  77         void            *base;
  78         u64             val;
  79         u32             order;
  80         u32             psz;
  81 };
  82
  83 struct its_device;
  84
  85 /*
  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.
  89  *
  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
  92  * list.
  93  */
  94 struct its_node {
  95         raw_spinlock_t          lock;
  96         struct mutex            dev_alloc_lock;
  97         struct list_head        entry;
  98         void __iomem            *base;
  99         phys_addr_t             phys_base;
 100         struct its_cmd_block    *cmd_base;
 101         struct its_cmd_block    *cmd_write;
 102         struct its_baser        tables[GITS_BASER_NR_REGS];
 103         struct its_collection   *collections;
 104         struct fwnode_handle    *fwnode_handle;
 105         u64                     (*get_msi_base)(struct its_device *its_dev);
 106         u64                     typer;
 107         u64                     cbaser_save;
 108         u32                     ctlr_save;
 109         struct list_head        its_device_list;
 110         u64                     flags;
 111         unsigned long           list_nr;
 112         int                     numa_node;
 113         unsigned int            msi_domain_flags;
 114         u32                     pre_its_base; /* for Socionext Synquacer */
 115         int                     vlpi_redist_offset;
 116 };
 117
 118 #define is_v4(its)              (!!((its)->typer & GITS_TYPER_VLPIS))
 119 #define device_ids(its)         (FIELD_GET(GITS_TYPER_DEVBITS, (its)->typer) + 1)
 120
 121 #define ITS_ITT_ALIGN           SZ_256
 122
 123 /* The maximum number of VPEID bits supported by VLPI commands */
 124 #define ITS_MAX_VPEID_BITS      (16)
 125 #define ITS_MAX_VPEID           (1 << (ITS_MAX_VPEID_BITS))
 126
 127 /* Convert page order to size in bytes */
 128 #define PAGE_ORDER_TO_SIZE(o)   (PAGE_SIZE << (o))
 129
 130 struct event_lpi_map {
 131         unsigned long           *lpi_map;
 132         u16                     *col_map;
 133         irq_hw_number_t         lpi_base;
 134         int                     nr_lpis;
 135         raw_spinlock_t          vlpi_lock;
 136         struct its_vm           *vm;
 137         struct its_vlpi_map     *vlpi_maps;
 138         int                     nr_vlpis;
 139 };
 140
 141 /*
 142  * The ITS view of a device - belongs to an ITS, owns an interrupt
 143  * translation table, and a list of interrupts.  If it some of its
 144  * LPIs are injected into a guest (GICv4), the event_map.vm field
 145  * indicates which one.
 146  */
 147 struct its_device {
 148         struct list_head        entry;
 149         struct its_node         *its;
 150         struct event_lpi_map    event_map;
 151         void                    *itt;
 152         u32                     nr_ites;
 153         u32                     device_id;
 154         bool                    shared;
 155 };
 156
 157 static struct {
 158         raw_spinlock_t          lock;
 159         struct its_device       *dev;
 160         struct its_vpe          **vpes;
 161         int                     next_victim;
 162 } vpe_proxy;
 163
 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;
 168
 169 static unsigned long its_list_map;
 170 static u16 vmovp_seq_num;
 171 static DEFINE_RAW_SPINLOCK(vmovp_lock);
 172
 173 static DEFINE_IDA(its_vpeid_ida);
 174
 175 #define gic_data_rdist()                (raw_cpu_ptr(gic_rdists->rdist))
 176 #define gic_data_rdist_cpu(cpu)         (per_cpu_ptr(gic_rdists->rdist, cpu))
 177 #define gic_data_rdist_rd_base()        (gic_data_rdist()->rd_base)
 178 #define gic_data_rdist_vlpi_base()      (gic_data_rdist_rd_base() + SZ_128K)
 179
 180 static u16 get_its_list(struct its_vm *vm)
 181 {
 182         struct its_node *its;
 183         unsigned long its_list = 0;
 184
 185         list_for_each_entry(its, &its_nodes, entry) {
 186                 if (!is_v4(its))
 187                         continue;
 188
 189                 if (vm->vlpi_count[its->list_nr])
 190                         __set_bit(its->list_nr, &its_list);
 191         }
 192
 193         return (u16)its_list;
 194 }
 195
 196 static inline u32 its_get_event_id(struct irq_data *d)
 197 {
 198         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
 199         return d->hwirq - its_dev->event_map.lpi_base;
 200 }
 201
 202 static struct its_collection *dev_event_to_col(struct its_device *its_dev,
 203                                                u32 event)
 204 {
 205         struct its_node *its = its_dev->its;
 206
 207         return its->collections + its_dev->event_map.col_map[event];
 208 }
 209
 210 static struct its_vlpi_map *dev_event_to_vlpi_map(struct its_device *its_dev,
 211                                                u32 event)
 212 {
 213         if (WARN_ON_ONCE(event >= its_dev->event_map.nr_lpis))
 214                 return NULL;
 215
 216         return &its_dev->event_map.vlpi_maps[event];
 217 }
 218
 219 static struct its_collection *irq_to_col(struct irq_data *d)
 220 {
 221         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
 222
 223         return dev_event_to_col(its_dev, its_get_event_id(d));
 224 }
 225
 226 static struct its_collection *valid_col(struct its_collection *col)
 227 {
 228         if (WARN_ON_ONCE(col->target_address & GENMASK_ULL(15, 0)))
 229                 return NULL;
 230
 231         return col;
 232 }
 233
 234 static struct its_vpe *valid_vpe(struct its_node *its, struct its_vpe *vpe)
 235 {
 236         if (valid_col(its->collections + vpe->col_idx))
 237                 return vpe;
 238
 239         return NULL;
 240 }
 241
 242 /*
 243  * ITS command descriptors - parameters to be encoded in a command
 244  * block.
 245  */
 246 struct its_cmd_desc {
 247         union {
 248                 struct {
 249                         struct its_device *dev;
 250                         u32 event_id;
 251                 } its_inv_cmd;
 252
 253                 struct {
 254                         struct its_device *dev;
 255                         u32 event_id;
 256                 } its_clear_cmd;
 257
 258                 struct {
 259                         struct its_device *dev;
 260                         u32 event_id;
 261                 } its_int_cmd;
 262
 263                 struct {
 264                         struct its_device *dev;
 265                         int valid;
 266                 } its_mapd_cmd;
 267
 268                 struct {
 269                         struct its_collection *col;
 270                         int valid;
 271                 } its_mapc_cmd;
 272
 273                 struct {
 274                         struct its_device *dev;
 275                         u32 phys_id;
 276                         u32 event_id;
 277                 } its_mapti_cmd;
 278
 279                 struct {
 280                         struct its_device *dev;
 281                         struct its_collection *col;
 282                         u32 event_id;
 283                 } its_movi_cmd;
 284
 285                 struct {
 286                         struct its_device *dev;
 287                         u32 event_id;
 288                 } its_discard_cmd;
 289
 290                 struct {
 291                         struct its_collection *col;
 292                 } its_invall_cmd;
 293
 294                 struct {
 295                         struct its_vpe *vpe;
 296                 } its_vinvall_cmd;
 297
 298                 struct {
 299                         struct its_vpe *vpe;
 300                         struct its_collection *col;
 301                         bool valid;
 302                 } its_vmapp_cmd;
 303
 304                 struct {
 305                         struct its_vpe *vpe;
 306                         struct its_device *dev;
 307                         u32 virt_id;
 308                         u32 event_id;
 309                         bool db_enabled;
 310                 } its_vmapti_cmd;
 311
 312                 struct {
 313                         struct its_vpe *vpe;
 314                         struct its_device *dev;
 315                         u32 event_id;
 316                         bool db_enabled;
 317                 } its_vmovi_cmd;
 318
 319                 struct {
 320                         struct its_vpe *vpe;
 321                         struct its_collection *col;
 322                         u16 seq_num;
 323                         u16 its_list;
 324                 } its_vmovp_cmd;
 325         };
 326 };
 327
 328 /*
 329  * The ITS command block, which is what the ITS actually parses.
 330  */
 331 struct its_cmd_block {
 332         union {
 333                 u64     raw_cmd[4];
 334                 __le64  raw_cmd_le[4];
 335         };
 336 };
 337
 338 #define ITS_CMD_QUEUE_SZ                SZ_64K
 339 #define ITS_CMD_QUEUE_NR_ENTRIES        (ITS_CMD_QUEUE_SZ / sizeof(struct its_cmd_block))
 340
 341 typedef struct its_collection *(*its_cmd_builder_t)(struct its_node *,
 342                                                     struct its_cmd_block *,
 343                                                     struct its_cmd_desc *);
 344
 345 typedef struct its_vpe *(*its_cmd_vbuilder_t)(struct its_node *,
 346                                               struct its_cmd_block *,
 347                                               struct its_cmd_desc *);
 348
 349 static void its_mask_encode(u64 *raw_cmd, u64 val, int h, int l)
 350 {
 351         u64 mask = GENMASK_ULL(h, l);
 352         *raw_cmd &= ~mask;
 353         *raw_cmd |= (val << l) & mask;
 354 }
 355
 356 static void its_encode_cmd(struct its_cmd_block *cmd, u8 cmd_nr)
 357 {
 358         its_mask_encode(&cmd->raw_cmd[0], cmd_nr, 7, 0);
 359 }
 360
 361 static void its_encode_devid(struct its_cmd_block *cmd, u32 devid)
 362 {
 363         its_mask_encode(&cmd->raw_cmd[0], devid, 63, 32);
 364 }
 365
 366 static void its_encode_event_id(struct its_cmd_block *cmd, u32 id)
 367 {
 368         its_mask_encode(&cmd->raw_cmd[1], id, 31, 0);
 369 }
 370
 371 static void its_encode_phys_id(struct its_cmd_block *cmd, u32 phys_id)
 372 {
 373         its_mask_encode(&cmd->raw_cmd[1], phys_id, 63, 32);
 374 }
 375
 376 static void its_encode_size(struct its_cmd_block *cmd, u8 size)
 377 {
 378         its_mask_encode(&cmd->raw_cmd[1], size, 4, 0);
 379 }
 380
 381 static void its_encode_itt(struct its_cmd_block *cmd, u64 itt_addr)
 382 {
 383         its_mask_encode(&cmd->raw_cmd[2], itt_addr >> 8, 51, 8);
 384 }
 385
 386 static void its_encode_valid(struct its_cmd_block *cmd, int valid)
 387 {
 388         its_mask_encode(&cmd->raw_cmd[2], !!valid, 63, 63);
 389 }
 390
 391 static void its_encode_target(struct its_cmd_block *cmd, u64 target_addr)
 392 {
 393         its_mask_encode(&cmd->raw_cmd[2], target_addr >> 16, 51, 16);
 394 }
 395
 396 static void its_encode_collection(struct its_cmd_block *cmd, u16 col)
 397 {
 398         its_mask_encode(&cmd->raw_cmd[2], col, 15, 0);
 399 }
 400
 401 static void its_encode_vpeid(struct its_cmd_block *cmd, u16 vpeid)
 402 {
 403         its_mask_encode(&cmd->raw_cmd[1], vpeid, 47, 32);
 404 }
 405
 406 static void its_encode_virt_id(struct its_cmd_block *cmd, u32 virt_id)
 407 {
 408         its_mask_encode(&cmd->raw_cmd[2], virt_id, 31, 0);
 409 }
 410
 411 static void its_encode_db_phys_id(struct its_cmd_block *cmd, u32 db_phys_id)
 412 {
 413         its_mask_encode(&cmd->raw_cmd[2], db_phys_id, 63, 32);
 414 }
 415
 416 static void its_encode_db_valid(struct its_cmd_block *cmd, bool db_valid)
 417 {
 418         its_mask_encode(&cmd->raw_cmd[2], db_valid, 0, 0);
 419 }
 420
 421 static void its_encode_seq_num(struct its_cmd_block *cmd, u16 seq_num)
 422 {
 423         its_mask_encode(&cmd->raw_cmd[0], seq_num, 47, 32);
 424 }
 425
 426 static void its_encode_its_list(struct its_cmd_block *cmd, u16 its_list)
 427 {
 428         its_mask_encode(&cmd->raw_cmd[1], its_list, 15, 0);
 429 }
 430
 431 static void its_encode_vpt_addr(struct its_cmd_block *cmd, u64 vpt_pa)
 432 {
 433         its_mask_encode(&cmd->raw_cmd[3], vpt_pa >> 16, 51, 16);
 434 }
 435
 436 static void its_encode_vpt_size(struct its_cmd_block *cmd, u8 vpt_size)
 437 {
 438         its_mask_encode(&cmd->raw_cmd[3], vpt_size, 4, 0);
 439 }
 440
 441 static inline void its_fixup_cmd(struct its_cmd_block *cmd)
 442 {
 443         /* Let's fixup BE commands */
 444         cmd->raw_cmd_le[0] = cpu_to_le64(cmd->raw_cmd[0]);
 445         cmd->raw_cmd_le[1] = cpu_to_le64(cmd->raw_cmd[1]);
 446         cmd->raw_cmd_le[2] = cpu_to_le64(cmd->raw_cmd[2]);
 447         cmd->raw_cmd_le[3] = cpu_to_le64(cmd->raw_cmd[3]);
 448 }
 449
 450 static struct its_collection *its_build_mapd_cmd(struct its_node *its,
 451                                                  struct its_cmd_block *cmd,
 452                                                  struct its_cmd_desc *desc)
 453 {
 454         unsigned long itt_addr;
 455         u8 size = ilog2(desc->its_mapd_cmd.dev->nr_ites);
 456
 457         itt_addr = virt_to_phys(desc->its_mapd_cmd.dev->itt);
 458         itt_addr = ALIGN(itt_addr, ITS_ITT_ALIGN);
 459
 460         its_encode_cmd(cmd, GITS_CMD_MAPD);
 461         its_encode_devid(cmd, desc->its_mapd_cmd.dev->device_id);
 462         its_encode_size(cmd, size - 1);
 463         its_encode_itt(cmd, itt_addr);
 464         its_encode_valid(cmd, desc->its_mapd_cmd.valid);
 465
 466         its_fixup_cmd(cmd);
 467
 468         return NULL;
 469 }
 470
 471 static struct its_collection *its_build_mapc_cmd(struct its_node *its,
 472                                                  struct its_cmd_block *cmd,
 473                                                  struct its_cmd_desc *desc)
 474 {
 475         its_encode_cmd(cmd, GITS_CMD_MAPC);
 476         its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
 477         its_encode_target(cmd, desc->its_mapc_cmd.col->target_address);
 478         its_encode_valid(cmd, desc->its_mapc_cmd.valid);
 479
 480         its_fixup_cmd(cmd);
 481
 482         return desc->its_mapc_cmd.col;
 483 }
 484
 485 static struct its_collection *its_build_mapti_cmd(struct its_node *its,
 486                                                   struct its_cmd_block *cmd,
 487                                                   struct its_cmd_desc *desc)
 488 {
 489         struct its_collection *col;
 490
 491         col = dev_event_to_col(desc->its_mapti_cmd.dev,
 492                                desc->its_mapti_cmd.event_id);
 493
 494         its_encode_cmd(cmd, GITS_CMD_MAPTI);
 495         its_encode_devid(cmd, desc->its_mapti_cmd.dev->device_id);
 496         its_encode_event_id(cmd, desc->its_mapti_cmd.event_id);
 497         its_encode_phys_id(cmd, desc->its_mapti_cmd.phys_id);
 498         its_encode_collection(cmd, col->col_id);
 499
 500         its_fixup_cmd(cmd);
 501
 502         return valid_col(col);
 503 }
 504
 505 static struct its_collection *its_build_movi_cmd(struct its_node *its,
 506                                                  struct its_cmd_block *cmd,
 507                                                  struct its_cmd_desc *desc)
 508 {
 509         struct its_collection *col;
 510
 511         col = dev_event_to_col(desc->its_movi_cmd.dev,
 512                                desc->its_movi_cmd.event_id);
 513
 514         its_encode_cmd(cmd, GITS_CMD_MOVI);
 515         its_encode_devid(cmd, desc->its_movi_cmd.dev->device_id);
 516         its_encode_event_id(cmd, desc->its_movi_cmd.event_id);
 517         its_encode_collection(cmd, desc->its_movi_cmd.col->col_id);
 518
 519         its_fixup_cmd(cmd);
 520
 521         return valid_col(col);
 522 }
 523
 524 static struct its_collection *its_build_discard_cmd(struct its_node *its,
 525                                                     struct its_cmd_block *cmd,
 526                                                     struct its_cmd_desc *desc)
 527 {
 528         struct its_collection *col;
 529
 530         col = dev_event_to_col(desc->its_discard_cmd.dev,
 531                                desc->its_discard_cmd.event_id);
 532
 533         its_encode_cmd(cmd, GITS_CMD_DISCARD);
 534         its_encode_devid(cmd, desc->its_discard_cmd.dev->device_id);
 535         its_encode_event_id(cmd, desc->its_discard_cmd.event_id);
 536
 537         its_fixup_cmd(cmd);
 538
 539         return valid_col(col);
 540 }
 541
 542 static struct its_collection *its_build_inv_cmd(struct its_node *its,
 543                                                 struct its_cmd_block *cmd,
 544                                                 struct its_cmd_desc *desc)
 545 {
 546         struct its_collection *col;
 547
 548         col = dev_event_to_col(desc->its_inv_cmd.dev,
 549                                desc->its_inv_cmd.event_id);
 550
 551         its_encode_cmd(cmd, GITS_CMD_INV);
 552         its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
 553         its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
 554
 555         its_fixup_cmd(cmd);
 556
 557         return valid_col(col);
 558 }
 559
 560 static struct its_collection *its_build_int_cmd(struct its_node *its,
 561                                                 struct its_cmd_block *cmd,
 562                                                 struct its_cmd_desc *desc)
 563 {
 564         struct its_collection *col;
 565
 566         col = dev_event_to_col(desc->its_int_cmd.dev,
 567                                desc->its_int_cmd.event_id);
 568
 569         its_encode_cmd(cmd, GITS_CMD_INT);
 570         its_encode_devid(cmd, desc->its_int_cmd.dev->device_id);
 571         its_encode_event_id(cmd, desc->its_int_cmd.event_id);
 572
 573         its_fixup_cmd(cmd);
 574
 575         return valid_col(col);
 576 }
 577
 578 static struct its_collection *its_build_clear_cmd(struct its_node *its,
 579                                                   struct its_cmd_block *cmd,
 580                                                   struct its_cmd_desc *desc)
 581 {
 582         struct its_collection *col;
 583
 584         col = dev_event_to_col(desc->its_clear_cmd.dev,
 585                                desc->its_clear_cmd.event_id);
 586
 587         its_encode_cmd(cmd, GITS_CMD_CLEAR);
 588         its_encode_devid(cmd, desc->its_clear_cmd.dev->device_id);
 589         its_encode_event_id(cmd, desc->its_clear_cmd.event_id);
 590
 591         its_fixup_cmd(cmd);
 592
 593         return valid_col(col);
 594 }
 595
 596 static struct its_collection *its_build_invall_cmd(struct its_node *its,
 597                                                    struct its_cmd_block *cmd,
 598                                                    struct its_cmd_desc *desc)
 599 {
 600         its_encode_cmd(cmd, GITS_CMD_INVALL);
 601         its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
 602
 603         its_fixup_cmd(cmd);
 604
 605         return NULL;
 606 }
 607
 608 static struct its_vpe *its_build_vinvall_cmd(struct its_node *its,
 609                                              struct its_cmd_block *cmd,
 610                                              struct its_cmd_desc *desc)
 611 {
 612         its_encode_cmd(cmd, GITS_CMD_VINVALL);
 613         its_encode_vpeid(cmd, desc->its_vinvall_cmd.vpe->vpe_id);
 614
 615         its_fixup_cmd(cmd);
 616
 617         return valid_vpe(its, desc->its_vinvall_cmd.vpe);
 618 }
 619
 620 static struct its_vpe *its_build_vmapp_cmd(struct its_node *its,
 621                                            struct its_cmd_block *cmd,
 622                                            struct its_cmd_desc *desc)
 623 {
 624         unsigned long vpt_addr;
 625         u64 target;
 626
 627         vpt_addr = virt_to_phys(page_address(desc->its_vmapp_cmd.vpe->vpt_page));
 628         target = desc->its_vmapp_cmd.col->target_address + its->vlpi_redist_offset;
 629
 630         its_encode_cmd(cmd, GITS_CMD_VMAPP);
 631         its_encode_vpeid(cmd, desc->its_vmapp_cmd.vpe->vpe_id);
 632         its_encode_valid(cmd, desc->its_vmapp_cmd.valid);
 633         its_encode_target(cmd, target);
 634         its_encode_vpt_addr(cmd, vpt_addr);
 635         its_encode_vpt_size(cmd, LPI_NRBITS - 1);
 636
 637         its_fixup_cmd(cmd);
 638
 639         return valid_vpe(its, desc->its_vmapp_cmd.vpe);
 640 }
 641
 642 static struct its_vpe *its_build_vmapti_cmd(struct its_node *its,
 643                                             struct its_cmd_block *cmd,
 644                                             struct its_cmd_desc *desc)
 645 {
 646         u32 db;
 647
 648         if (desc->its_vmapti_cmd.db_enabled)
 649                 db = desc->its_vmapti_cmd.vpe->vpe_db_lpi;
 650         else
 651                 db = 1023;
 652
 653         its_encode_cmd(cmd, GITS_CMD_VMAPTI);
 654         its_encode_devid(cmd, desc->its_vmapti_cmd.dev->device_id);
 655         its_encode_vpeid(cmd, desc->its_vmapti_cmd.vpe->vpe_id);
 656         its_encode_event_id(cmd, desc->its_vmapti_cmd.event_id);
 657         its_encode_db_phys_id(cmd, db);
 658         its_encode_virt_id(cmd, desc->its_vmapti_cmd.virt_id);
 659
 660         its_fixup_cmd(cmd);
 661
 662         return valid_vpe(its, desc->its_vmapti_cmd.vpe);
 663 }
 664
 665 static struct its_vpe *its_build_vmovi_cmd(struct its_node *its,
 666                                            struct its_cmd_block *cmd,
 667                                            struct its_cmd_desc *desc)
 668 {
 669         u32 db;
 670
 671         if (desc->its_vmovi_cmd.db_enabled)
 672                 db = desc->its_vmovi_cmd.vpe->vpe_db_lpi;
 673         else
 674                 db = 1023;
 675
 676         its_encode_cmd(cmd, GITS_CMD_VMOVI);
 677         its_encode_devid(cmd, desc->its_vmovi_cmd.dev->device_id);
 678         its_encode_vpeid(cmd, desc->its_vmovi_cmd.vpe->vpe_id);
 679         its_encode_event_id(cmd, desc->its_vmovi_cmd.event_id);
 680         its_encode_db_phys_id(cmd, db);
 681         its_encode_db_valid(cmd, true);
 682
 683         its_fixup_cmd(cmd);
 684
 685         return valid_vpe(its, desc->its_vmovi_cmd.vpe);
 686 }
 687
 688 static struct its_vpe *its_build_vmovp_cmd(struct its_node *its,
 689                                            struct its_cmd_block *cmd,
 690                                            struct its_cmd_desc *desc)
 691 {
 692         u64 target;
 693
 694         target = desc->its_vmovp_cmd.col->target_address + its->vlpi_redist_offset;
 695         its_encode_cmd(cmd, GITS_CMD_VMOVP);
 696         its_encode_seq_num(cmd, desc->its_vmovp_cmd.seq_num);
 697         its_encode_its_list(cmd, desc->its_vmovp_cmd.its_list);
 698         its_encode_vpeid(cmd, desc->its_vmovp_cmd.vpe->vpe_id);
 699         its_encode_target(cmd, target);
 700
 701         its_fixup_cmd(cmd);
 702
 703         return valid_vpe(its, desc->its_vmovp_cmd.vpe);
 704 }
 705
 706 static struct its_vpe *its_build_vinv_cmd(struct its_node *its,
 707                                           struct its_cmd_block *cmd,
 708                                           struct its_cmd_desc *desc)
 709 {
 710         struct its_vlpi_map *map;
 711
 712         map = dev_event_to_vlpi_map(desc->its_inv_cmd.dev,
 713                                     desc->its_inv_cmd.event_id);
 714
 715         its_encode_cmd(cmd, GITS_CMD_INV);
 716         its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
 717         its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
 718
 719         its_fixup_cmd(cmd);
 720
 721         return valid_vpe(its, map->vpe);
 722 }
 723
 724 static struct its_vpe *its_build_vint_cmd(struct its_node *its,
 725                                           struct its_cmd_block *cmd,
 726                                           struct its_cmd_desc *desc)
 727 {
 728         struct its_vlpi_map *map;
 729
 730         map = dev_event_to_vlpi_map(desc->its_int_cmd.dev,
 731                                     desc->its_int_cmd.event_id);
 732
 733         its_encode_cmd(cmd, GITS_CMD_INT);
 734         its_encode_devid(cmd, desc->its_int_cmd.dev->device_id);
 735         its_encode_event_id(cmd, desc->its_int_cmd.event_id);
 736
 737         its_fixup_cmd(cmd);
 738
 739         return valid_vpe(its, map->vpe);
 740 }
 741
 742 static struct its_vpe *its_build_vclear_cmd(struct its_node *its,
 743                                             struct its_cmd_block *cmd,
 744                                             struct its_cmd_desc *desc)
 745 {
 746         struct its_vlpi_map *map;
 747
 748         map = dev_event_to_vlpi_map(desc->its_clear_cmd.dev,
 749                                     desc->its_clear_cmd.event_id);
 750
 751         its_encode_cmd(cmd, GITS_CMD_CLEAR);
 752         its_encode_devid(cmd, desc->its_clear_cmd.dev->device_id);
 753         its_encode_event_id(cmd, desc->its_clear_cmd.event_id);
 754
 755         its_fixup_cmd(cmd);
 756
 757         return valid_vpe(its, map->vpe);
 758 }
 759
 760 static u64 its_cmd_ptr_to_offset(struct its_node *its,
 761                                  struct its_cmd_block *ptr)
 762 {
 763         return (ptr - its->cmd_base) * sizeof(*ptr);
 764 }
 765
 766 static int its_queue_full(struct its_node *its)
 767 {
 768         int widx;
 769         int ridx;
 770
 771         widx = its->cmd_write - its->cmd_base;
 772         ridx = readl_relaxed(its->base + GITS_CREADR) / sizeof(struct its_cmd_block);
 773
 774         /* This is incredibly unlikely to happen, unless the ITS locks up. */
 775         if (((widx + 1) % ITS_CMD_QUEUE_NR_ENTRIES) == ridx)
 776                 return 1;
 777
 778         return 0;
 779 }
 780
 781 static struct its_cmd_block *its_allocate_entry(struct its_node *its)
 782 {
 783         struct its_cmd_block *cmd;
 784         u32 count = 1000000;    /* 1s! */
 785
 786         while (its_queue_full(its)) {
 787                 count--;
 788                 if (!count) {
 789                         pr_err_ratelimited("ITS queue not draining\n");
 790                         return NULL;
 791                 }
 792                 cpu_relax();
 793                 udelay(1);
 794         }
 795
 796         cmd = its->cmd_write++;
 797
 798         /* Handle queue wrapping */
 799         if (its->cmd_write == (its->cmd_base + ITS_CMD_QUEUE_NR_ENTRIES))
 800                 its->cmd_write = its->cmd_base;
 801
 802         /* Clear command  */
 803         cmd->raw_cmd[0] = 0;
 804         cmd->raw_cmd[1] = 0;
 805         cmd->raw_cmd[2] = 0;
 806         cmd->raw_cmd[3] = 0;
 807
 808         return cmd;
 809 }
 810
 811 static struct its_cmd_block *its_post_commands(struct its_node *its)
 812 {
 813         u64 wr = its_cmd_ptr_to_offset(its, its->cmd_write);
 814
 815         writel_relaxed(wr, its->base + GITS_CWRITER);
 816
 817         return its->cmd_write;
 818 }
 819
 820 static void its_flush_cmd(struct its_node *its, struct its_cmd_block *cmd)
 821 {
 822         /*
 823          * Make sure the commands written to memory are observable by
 824          * the ITS.
 825          */
 826         if (its->flags & ITS_FLAGS_CMDQ_NEEDS_FLUSHING)
 827                 gic_flush_dcache_to_poc(cmd, sizeof(*cmd));
 828         else
 829                 dsb(ishst);
 830 }
 831
 832 static int its_wait_for_range_completion(struct its_node *its,
 833                                          u64    prev_idx,
 834                                          struct its_cmd_block *to)
 835 {
 836         u64 rd_idx, to_idx, linear_idx;
 837         u32 count = 1000000;    /* 1s! */
 838
 839         /* Linearize to_idx if the command set has wrapped around */
 840         to_idx = its_cmd_ptr_to_offset(its, to);
 841         if (to_idx < prev_idx)
 842                 to_idx += ITS_CMD_QUEUE_SZ;
 843
 844         linear_idx = prev_idx;
 845
 846         while (1) {
 847                 s64 delta;
 848
 849                 rd_idx = readl_relaxed(its->base + GITS_CREADR);
 850
 851                 /*
 852                  * Compute the read pointer progress, taking the
 853                  * potential wrap-around into account.
 854                  */
 855                 delta = rd_idx - prev_idx;
 856                 if (rd_idx < prev_idx)
 857                         delta += ITS_CMD_QUEUE_SZ;
 858
 859                 linear_idx += delta;
 860                 if (linear_idx >= to_idx)
 861                         break;
 862
 863                 count--;
 864                 if (!count) {
 865                         pr_err_ratelimited("ITS queue timeout (%llu %llu)\n",
 866                                            to_idx, linear_idx);
 867                         return -1;
 868                 }
 869                 prev_idx = rd_idx;
 870                 cpu_relax();
 871                 udelay(1);
 872         }
 873
 874         return 0;
 875 }
 876
 877 /* Warning, macro hell follows */
 878 #define BUILD_SINGLE_CMD_FUNC(name, buildtype, synctype, buildfn)       \
 879 void name(struct its_node *its,                                         \
 880           buildtype builder,                                            \
 881           struct its_cmd_desc *desc)                                    \
 882 {                                                                       \
 883         struct its_cmd_block *cmd, *sync_cmd, *next_cmd;                \
 884         synctype *sync_obj;                                             \
 885         unsigned long flags;                                            \
 886         u64 rd_idx;                                                     \
 887                                                                         \
 888         raw_spin_lock_irqsave(&its->lock, flags);                       \
 889                                                                         \
 890         cmd = its_allocate_entry(its);                                  \
 891         if (!cmd) {             /* We're soooooo screewed... */         \
 892                 raw_spin_unlock_irqrestore(&its->lock, flags);          \
 893                 return;                                                 \
 894         }                                                               \
 895         sync_obj = builder(its, cmd, desc);                             \
 896         its_flush_cmd(its, cmd);                                        \
 897                                                                         \
 898         if (sync_obj) {                                                 \
 899                 sync_cmd = its_allocate_entry(its);                     \
 900                 if (!sync_cmd)                                          \
 901                         goto post;                                      \
 902                                                                         \
 903                 buildfn(its, sync_cmd, sync_obj);                       \
 904                 its_flush_cmd(its, sync_cmd);                           \
 905         }                                                               \
 906                                                                         \
 907 post:                                                                   \
 908         rd_idx = readl_relaxed(its->base + GITS_CREADR);                \
 909         next_cmd = its_post_commands(its);                              \
 910         raw_spin_unlock_irqrestore(&its->lock, flags);                  \
 911                                                                         \
 912         if (its_wait_for_range_completion(its, rd_idx, next_cmd))       \
 913                 pr_err_ratelimited("ITS cmd %ps failed\n", builder);    \
 914 }
 915
 916 static void its_build_sync_cmd(struct its_node *its,
 917                                struct its_cmd_block *sync_cmd,
 918                                struct its_collection *sync_col)
 919 {
 920         its_encode_cmd(sync_cmd, GITS_CMD_SYNC);
 921         its_encode_target(sync_cmd, sync_col->target_address);
 922
 923         its_fixup_cmd(sync_cmd);
 924 }
 925
 926 static BUILD_SINGLE_CMD_FUNC(its_send_single_command, its_cmd_builder_t,
 927                              struct its_collection, its_build_sync_cmd)
 928
 929 static void its_build_vsync_cmd(struct its_node *its,
 930                                 struct its_cmd_block *sync_cmd,
 931                                 struct its_vpe *sync_vpe)
 932 {
 933         its_encode_cmd(sync_cmd, GITS_CMD_VSYNC);
 934         its_encode_vpeid(sync_cmd, sync_vpe->vpe_id);
 935
 936         its_fixup_cmd(sync_cmd);
 937 }
 938
 939 static BUILD_SINGLE_CMD_FUNC(its_send_single_vcommand, its_cmd_vbuilder_t,
 940                              struct its_vpe, its_build_vsync_cmd)
 941
 942 static void its_send_int(struct its_device *dev, u32 event_id)
 943 {
 944         struct its_cmd_desc desc;
 945
 946         desc.its_int_cmd.dev = dev;
 947         desc.its_int_cmd.event_id = event_id;
 948
 949         its_send_single_command(dev->its, its_build_int_cmd, &desc);
 950 }
 951
 952 static void its_send_clear(struct its_device *dev, u32 event_id)
 953 {
 954         struct its_cmd_desc desc;
 955
 956         desc.its_clear_cmd.dev = dev;
 957         desc.its_clear_cmd.event_id = event_id;
 958
 959         its_send_single_command(dev->its, its_build_clear_cmd, &desc);
 960 }
 961
 962 static void its_send_inv(struct its_device *dev, u32 event_id)
 963 {
 964         struct its_cmd_desc desc;
 965
 966         desc.its_inv_cmd.dev = dev;
 967         desc.its_inv_cmd.event_id = event_id;
 968
 969         its_send_single_command(dev->its, its_build_inv_cmd, &desc);
 970 }
 971
 972 static void its_send_mapd(struct its_device *dev, int valid)
 973 {
 974         struct its_cmd_desc desc;
 975
 976         desc.its_mapd_cmd.dev = dev;
 977         desc.its_mapd_cmd.valid = !!valid;
 978
 979         its_send_single_command(dev->its, its_build_mapd_cmd, &desc);
 980 }
 981
 982 static void its_send_mapc(struct its_node *its, struct its_collection *col,
 983                           int valid)
 984 {
 985         struct its_cmd_desc desc;
 986
 987         desc.its_mapc_cmd.col = col;
 988         desc.its_mapc_cmd.valid = !!valid;
 989
 990         its_send_single_command(its, its_build_mapc_cmd, &desc);
 991 }
 992
 993 static void its_send_mapti(struct its_device *dev, u32 irq_id, u32 id)
 994 {
 995         struct its_cmd_desc desc;
 996
 997         desc.its_mapti_cmd.dev = dev;
 998         desc.its_mapti_cmd.phys_id = irq_id;
 999         desc.its_mapti_cmd.event_id = id;
1000
1001         its_send_single_command(dev->its, its_build_mapti_cmd, &desc);
1002 }
1003
1004 static void its_send_movi(struct its_device *dev,
1005                           struct its_collection *col, u32 id)
1006 {
1007         struct its_cmd_desc desc;
1008
1009         desc.its_movi_cmd.dev = dev;
1010         desc.its_movi_cmd.col = col;
1011         desc.its_movi_cmd.event_id = id;
1012
1013         its_send_single_command(dev->its, its_build_movi_cmd, &desc);
1014 }
1015
1016 static void its_send_discard(struct its_device *dev, u32 id)
1017 {
1018         struct its_cmd_desc desc;
1019
1020         desc.its_discard_cmd.dev = dev;
1021         desc.its_discard_cmd.event_id = id;
1022
1023         its_send_single_command(dev->its, its_build_discard_cmd, &desc);
1024 }
1025
1026 static void its_send_invall(struct its_node *its, struct its_collection *col)
1027 {
1028         struct its_cmd_desc desc;
1029
1030         desc.its_invall_cmd.col = col;
1031
1032         its_send_single_command(its, its_build_invall_cmd, &desc);
1033 }
1034
1035 static void its_send_vmapti(struct its_device *dev, u32 id)
1036 {
1037         struct its_vlpi_map *map = dev_event_to_vlpi_map(dev, id);
1038         struct its_cmd_desc desc;
1039
1040         desc.its_vmapti_cmd.vpe = map->vpe;
1041         desc.its_vmapti_cmd.dev = dev;
1042         desc.its_vmapti_cmd.virt_id = map->vintid;
1043         desc.its_vmapti_cmd.event_id = id;
1044         desc.its_vmapti_cmd.db_enabled = map->db_enabled;
1045
1046         its_send_single_vcommand(dev->its, its_build_vmapti_cmd, &desc);
1047 }
1048
1049 static void its_send_vmovi(struct its_device *dev, u32 id)
1050 {
1051         struct its_vlpi_map *map = dev_event_to_vlpi_map(dev, id);
1052         struct its_cmd_desc desc;
1053
1054         desc.its_vmovi_cmd.vpe = map->vpe;
1055         desc.its_vmovi_cmd.dev = dev;
1056         desc.its_vmovi_cmd.event_id = id;
1057         desc.its_vmovi_cmd.db_enabled = map->db_enabled;
1058
1059         its_send_single_vcommand(dev->its, its_build_vmovi_cmd, &desc);
1060 }
1061
1062 static void its_send_vmapp(struct its_node *its,
1063                            struct its_vpe *vpe, bool valid)
1064 {
1065         struct its_cmd_desc desc;
1066
1067         desc.its_vmapp_cmd.vpe = vpe;
1068         desc.its_vmapp_cmd.valid = valid;
1069         desc.its_vmapp_cmd.col = &its->collections[vpe->col_idx];
1070
1071         its_send_single_vcommand(its, its_build_vmapp_cmd, &desc);
1072 }
1073
1074 static void its_send_vmovp(struct its_vpe *vpe)
1075 {
1076         struct its_cmd_desc desc = {};
1077         struct its_node *its;
1078         unsigned long flags;
1079         int col_id = vpe->col_idx;
1080
1081         desc.its_vmovp_cmd.vpe = vpe;
1082
1083         if (!its_list_map) {
1084                 its = list_first_entry(&its_nodes, struct its_node, entry);
1085                 desc.its_vmovp_cmd.col = &its->collections[col_id];
1086                 its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
1087                 return;
1088         }
1089
1090         /*
1091          * Yet another marvel of the architecture. If using the
1092          * its_list "feature", we need to make sure that all ITSs
1093          * receive all VMOVP commands in the same order. The only way
1094          * to guarantee this is to make vmovp a serialization point.
1095          *
1096          * Wall <-- Head.
1097          */
1098         raw_spin_lock_irqsave(&vmovp_lock, flags);
1099
1100         desc.its_vmovp_cmd.seq_num = vmovp_seq_num++;
1101         desc.its_vmovp_cmd.its_list = get_its_list(vpe->its_vm);
1102
1103         /* Emit VMOVPs */
1104         list_for_each_entry(its, &its_nodes, entry) {
1105                 if (!is_v4(its))
1106                         continue;
1107
1108                 if (!vpe->its_vm->vlpi_count[its->list_nr])
1109                         continue;
1110
1111                 desc.its_vmovp_cmd.col = &its->collections[col_id];
1112                 its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
1113         }
1114
1115         raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1116 }
1117
1118 static void its_send_vinvall(struct its_node *its, struct its_vpe *vpe)
1119 {
1120         struct its_cmd_desc desc;
1121
1122         desc.its_vinvall_cmd.vpe = vpe;
1123         its_send_single_vcommand(its, its_build_vinvall_cmd, &desc);
1124 }
1125
1126 static void its_send_vinv(struct its_device *dev, u32 event_id)
1127 {
1128         struct its_cmd_desc desc;
1129
1130         /*
1131          * There is no real VINV command. This is just a normal INV,
1132          * with a VSYNC instead of a SYNC.
1133          */
1134         desc.its_inv_cmd.dev = dev;
1135         desc.its_inv_cmd.event_id = event_id;
1136
1137         its_send_single_vcommand(dev->its, its_build_vinv_cmd, &desc);
1138 }
1139
1140 static void its_send_vint(struct its_device *dev, u32 event_id)
1141 {
1142         struct its_cmd_desc desc;
1143
1144         /*
1145          * There is no real VINT command. This is just a normal INT,
1146          * with a VSYNC instead of a SYNC.
1147          */
1148         desc.its_int_cmd.dev = dev;
1149         desc.its_int_cmd.event_id = event_id;
1150
1151         its_send_single_vcommand(dev->its, its_build_vint_cmd, &desc);
1152 }
1153
1154 static void its_send_vclear(struct its_device *dev, u32 event_id)
1155 {
1156         struct its_cmd_desc desc;
1157
1158         /*
1159          * There is no real VCLEAR command. This is just a normal CLEAR,
1160          * with a VSYNC instead of a SYNC.
1161          */
1162         desc.its_clear_cmd.dev = dev;
1163         desc.its_clear_cmd.event_id = event_id;
1164
1165         its_send_single_vcommand(dev->its, its_build_vclear_cmd, &desc);
1166 }
1167
1168 /*
1169  * irqchip functions - assumes MSI, mostly.
1170  */
1171 static struct its_vlpi_map *get_vlpi_map(struct irq_data *d)
1172 {
1173         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1174         u32 event = its_get_event_id(d);
1175
1176         if (!irqd_is_forwarded_to_vcpu(d))
1177                 return NULL;
1178
1179         return dev_event_to_vlpi_map(its_dev, event);
1180 }
1181
1182 static void lpi_write_config(struct irq_data *d, u8 clr, u8 set)
1183 {
1184         struct its_vlpi_map *map = get_vlpi_map(d);
1185         irq_hw_number_t hwirq;
1186         void *va;
1187         u8 *cfg;
1188
1189         if (map) {
1190                 va = page_address(map->vm->vprop_page);
1191                 hwirq = map->vintid;
1192
1193                 /* Remember the updated property */
1194                 map->properties &= ~clr;
1195                 map->properties |= set | LPI_PROP_GROUP1;
1196         } else {
1197                 va = gic_rdists->prop_table_va;
1198                 hwirq = d->hwirq;
1199         }
1200
1201         cfg = va + hwirq - 8192;
1202         *cfg &= ~clr;
1203         *cfg |= set | LPI_PROP_GROUP1;
1204
1205         /*
1206          * Make the above write visible to the redistributors.
1207          * And yes, we're flushing exactly: One. Single. Byte.
1208          * Humpf...
1209          */
1210         if (gic_rdists->flags & RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING)
1211                 gic_flush_dcache_to_poc(cfg, sizeof(*cfg));
1212         else
1213                 dsb(ishst);
1214 }
1215
1216 static void wait_for_syncr(void __iomem *rdbase)
1217 {
1218         while (gic_read_lpir(rdbase + GICR_SYNCR) & 1)
1219                 cpu_relax();
1220 }
1221
1222 static void direct_lpi_inv(struct irq_data *d)
1223 {
1224         struct its_collection *col;
1225         void __iomem *rdbase;
1226
1227         /* Target the redistributor this LPI is currently routed to */
1228         col = irq_to_col(d);
1229         rdbase = per_cpu_ptr(gic_rdists->rdist, col->col_id)->rd_base;
1230         gic_write_lpir(d->hwirq, rdbase + GICR_INVLPIR);
1231
1232         wait_for_syncr(rdbase);
1233 }
1234
1235 static void lpi_update_config(struct irq_data *d, u8 clr, u8 set)
1236 {
1237         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1238
1239         lpi_write_config(d, clr, set);
1240         if (gic_rdists->has_direct_lpi && !irqd_is_forwarded_to_vcpu(d))
1241                 direct_lpi_inv(d);
1242         else if (!irqd_is_forwarded_to_vcpu(d))
1243                 its_send_inv(its_dev, its_get_event_id(d));
1244         else
1245                 its_send_vinv(its_dev, its_get_event_id(d));
1246 }
1247
1248 static void its_vlpi_set_doorbell(struct irq_data *d, bool enable)
1249 {
1250         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1251         u32 event = its_get_event_id(d);
1252         struct its_vlpi_map *map;
1253
1254         map = dev_event_to_vlpi_map(its_dev, event);
1255
1256         if (map->db_enabled == enable)
1257                 return;
1258
1259         map->db_enabled = enable;
1260
1261         /*
1262          * More fun with the architecture:
1263          *
1264          * Ideally, we'd issue a VMAPTI to set the doorbell to its LPI
1265          * value or to 1023, depending on the enable bit. But that
1266          * would be issueing a mapping for an /existing/ DevID+EventID
1267          * pair, which is UNPREDICTABLE. Instead, let's issue a VMOVI
1268          * to the /same/ vPE, using this opportunity to adjust the
1269          * doorbell. Mouahahahaha. We loves it, Precious.
1270          */
1271         its_send_vmovi(its_dev, event);
1272 }
1273
1274 static void its_mask_irq(struct irq_data *d)
1275 {
1276         if (irqd_is_forwarded_to_vcpu(d))
1277                 its_vlpi_set_doorbell(d, false);
1278
1279         lpi_update_config(d, LPI_PROP_ENABLED, 0);
1280 }
1281
1282 static void its_unmask_irq(struct irq_data *d)
1283 {
1284         if (irqd_is_forwarded_to_vcpu(d))
1285                 its_vlpi_set_doorbell(d, true);
1286
1287         lpi_update_config(d, 0, LPI_PROP_ENABLED);
1288 }
1289
1290 static int its_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
1291                             bool force)
1292 {
1293         unsigned int cpu;
1294         const struct cpumask *cpu_mask = cpu_online_mask;
1295         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1296         struct its_collection *target_col;
1297         u32 id = its_get_event_id(d);
1298
1299         /* A forwarded interrupt should use irq_set_vcpu_affinity */
1300         if (irqd_is_forwarded_to_vcpu(d))
1301                 return -EINVAL;
1302
1303        /* lpi cannot be routed to a redistributor that is on a foreign node */
1304         if (its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) {
1305                 if (its_dev->its->numa_node >= 0) {
1306                         cpu_mask = cpumask_of_node(its_dev->its->numa_node);
1307                         if (!cpumask_intersects(mask_val, cpu_mask))
1308                                 return -EINVAL;
1309                 }
1310         }
1311
1312         cpu = cpumask_any_and(mask_val, cpu_mask);
1313
1314         if (cpu >= nr_cpu_ids)
1315                 return -EINVAL;
1316
1317         /* don't set the affinity when the target cpu is same as current one */
1318         if (cpu != its_dev->event_map.col_map[id]) {
1319                 target_col = &its_dev->its->collections[cpu];
1320                 its_send_movi(its_dev, target_col, id);
1321                 its_dev->event_map.col_map[id] = cpu;
1322                 irq_data_update_effective_affinity(d, cpumask_of(cpu));
1323         }
1324
1325         return IRQ_SET_MASK_OK_DONE;
1326 }
1327
1328 static u64 its_irq_get_msi_base(struct its_device *its_dev)
1329 {
1330         struct its_node *its = its_dev->its;
1331
1332         return its->phys_base + GITS_TRANSLATER;
1333 }
1334
1335 static void its_irq_compose_msi_msg(struct irq_data *d, struct msi_msg *msg)
1336 {
1337         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1338         struct its_node *its;
1339         u64 addr;
1340
1341         its = its_dev->its;
1342         addr = its->get_msi_base(its_dev);
1343
1344         msg->address_lo         = lower_32_bits(addr);
1345         msg->address_hi         = upper_32_bits(addr);
1346         msg->data               = its_get_event_id(d);
1347
1348         iommu_dma_compose_msi_msg(irq_data_get_msi_desc(d), msg);
1349 }
1350
1351 static int its_irq_set_irqchip_state(struct irq_data *d,
1352                                      enum irqchip_irq_state which,
1353                                      bool state)
1354 {
1355         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1356         u32 event = its_get_event_id(d);
1357
1358         if (which != IRQCHIP_STATE_PENDING)
1359                 return -EINVAL;
1360
1361         if (irqd_is_forwarded_to_vcpu(d)) {
1362                 if (state)
1363                         its_send_vint(its_dev, event);
1364                 else
1365                         its_send_vclear(its_dev, event);
1366         } else {
1367                 if (state)
1368                         its_send_int(its_dev, event);
1369                 else
1370                         its_send_clear(its_dev, event);
1371         }
1372
1373         return 0;
1374 }
1375
1376 static void its_map_vm(struct its_node *its, struct its_vm *vm)
1377 {
1378         unsigned long flags;
1379
1380         /* Not using the ITS list? Everything is always mapped. */
1381         if (!its_list_map)
1382                 return;
1383
1384         raw_spin_lock_irqsave(&vmovp_lock, flags);
1385
1386         /*
1387          * If the VM wasn't mapped yet, iterate over the vpes and get
1388          * them mapped now.
1389          */
1390         vm->vlpi_count[its->list_nr]++;
1391
1392         if (vm->vlpi_count[its->list_nr] == 1) {
1393                 int i;
1394
1395                 for (i = 0; i < vm->nr_vpes; i++) {
1396                         struct its_vpe *vpe = vm->vpes[i];
1397                         struct irq_data *d = irq_get_irq_data(vpe->irq);
1398
1399                         /* Map the VPE to the first possible CPU */
1400                         vpe->col_idx = cpumask_first(cpu_online_mask);
1401                         its_send_vmapp(its, vpe, true);
1402                         its_send_vinvall(its, vpe);
1403                         irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
1404                 }
1405         }
1406
1407         raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1408 }
1409
1410 static void its_unmap_vm(struct its_node *its, struct its_vm *vm)
1411 {
1412         unsigned long flags;
1413
1414         /* Not using the ITS list? Everything is always mapped. */
1415         if (!its_list_map)
1416                 return;
1417
1418         raw_spin_lock_irqsave(&vmovp_lock, flags);
1419
1420         if (!--vm->vlpi_count[its->list_nr]) {
1421                 int i;
1422
1423                 for (i = 0; i < vm->nr_vpes; i++)
1424                         its_send_vmapp(its, vm->vpes[i], false);
1425         }
1426
1427         raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1428 }
1429
1430 static int its_vlpi_map(struct irq_data *d, struct its_cmd_info *info)
1431 {
1432         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1433         u32 event = its_get_event_id(d);
1434         int ret = 0;
1435
1436         if (!info->map)
1437                 return -EINVAL;
1438
1439         raw_spin_lock(&its_dev->event_map.vlpi_lock);
1440
1441         if (!its_dev->event_map.vm) {
1442                 struct its_vlpi_map *maps;
1443
1444                 maps = kcalloc(its_dev->event_map.nr_lpis, sizeof(*maps),
1445                                GFP_ATOMIC);
1446                 if (!maps) {
1447                         ret = -ENOMEM;
1448                         goto out;
1449                 }
1450
1451                 its_dev->event_map.vm = info->map->vm;
1452                 its_dev->event_map.vlpi_maps = maps;
1453         } else if (its_dev->event_map.vm != info->map->vm) {
1454                 ret = -EINVAL;
1455                 goto out;
1456         }
1457
1458         /* Get our private copy of the mapping information */
1459         its_dev->event_map.vlpi_maps[event] = *info->map;
1460
1461         if (irqd_is_forwarded_to_vcpu(d)) {
1462                 /* Already mapped, move it around */
1463                 its_send_vmovi(its_dev, event);
1464         } else {
1465                 /* Ensure all the VPEs are mapped on this ITS */
1466                 its_map_vm(its_dev->its, info->map->vm);
1467
1468                 /*
1469                  * Flag the interrupt as forwarded so that we can
1470                  * start poking the virtual property table.
1471                  */
1472                 irqd_set_forwarded_to_vcpu(d);
1473
1474                 /* Write out the property to the prop table */
1475                 lpi_write_config(d, 0xff, info->map->properties);
1476
1477                 /* Drop the physical mapping */
1478                 its_send_discard(its_dev, event);
1479
1480                 /* and install the virtual one */
1481                 its_send_vmapti(its_dev, event);
1482
1483                 /* Increment the number of VLPIs */
1484                 its_dev->event_map.nr_vlpis++;
1485         }
1486
1487 out:
1488         raw_spin_unlock(&its_dev->event_map.vlpi_lock);
1489         return ret;
1490 }
1491
1492 static int its_vlpi_get(struct irq_data *d, struct its_cmd_info *info)
1493 {
1494         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1495         struct its_vlpi_map *map;
1496         int ret = 0;
1497
1498         raw_spin_lock(&its_dev->event_map.vlpi_lock);
1499
1500         map = get_vlpi_map(d);
1501
1502         if (!its_dev->event_map.vm || !map) {
1503                 ret = -EINVAL;
1504                 goto out;
1505         }
1506
1507         /* Copy our mapping information to the incoming request */
1508         *info->map = *map;
1509
1510 out:
1511         raw_spin_unlock(&its_dev->event_map.vlpi_lock);
1512         return ret;
1513 }
1514
1515 static int its_vlpi_unmap(struct irq_data *d)
1516 {
1517         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1518         u32 event = its_get_event_id(d);
1519         int ret = 0;
1520
1521         raw_spin_lock(&its_dev->event_map.vlpi_lock);
1522
1523         if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d)) {
1524                 ret = -EINVAL;
1525                 goto out;
1526         }
1527
1528         /* Drop the virtual mapping */
1529         its_send_discard(its_dev, event);
1530
1531         /* and restore the physical one */
1532         irqd_clr_forwarded_to_vcpu(d);
1533         its_send_mapti(its_dev, d->hwirq, event);
1534         lpi_update_config(d, 0xff, (LPI_PROP_DEFAULT_PRIO |
1535                                     LPI_PROP_ENABLED |
1536                                     LPI_PROP_GROUP1));
1537
1538         /* Potentially unmap the VM from this ITS */
1539         its_unmap_vm(its_dev->its, its_dev->event_map.vm);
1540
1541         /*
1542          * Drop the refcount and make the device available again if
1543          * this was the last VLPI.
1544          */
1545         if (!--its_dev->event_map.nr_vlpis) {
1546                 its_dev->event_map.vm = NULL;
1547                 kfree(its_dev->event_map.vlpi_maps);
1548         }
1549
1550 out:
1551         raw_spin_unlock(&its_dev->event_map.vlpi_lock);
1552         return ret;
1553 }
1554
1555 static int its_vlpi_prop_update(struct irq_data *d, struct its_cmd_info *info)
1556 {
1557         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1558
1559         if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d))
1560                 return -EINVAL;
1561
1562         if (info->cmd_type == PROP_UPDATE_AND_INV_VLPI)
1563                 lpi_update_config(d, 0xff, info->config);
1564         else
1565                 lpi_write_config(d, 0xff, info->config);
1566         its_vlpi_set_doorbell(d, !!(info->config & LPI_PROP_ENABLED));
1567
1568         return 0;
1569 }
1570
1571 static int its_irq_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
1572 {
1573         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1574         struct its_cmd_info *info = vcpu_info;
1575
1576         /* Need a v4 ITS */
1577         if (!is_v4(its_dev->its))
1578                 return -EINVAL;
1579
1580         /* Unmap request? */
1581         if (!info)
1582                 return its_vlpi_unmap(d);
1583
1584         switch (info->cmd_type) {
1585         case MAP_VLPI:
1586                 return its_vlpi_map(d, info);
1587
1588         case GET_VLPI:
1589                 return its_vlpi_get(d, info);
1590
1591         case PROP_UPDATE_VLPI:
1592         case PROP_UPDATE_AND_INV_VLPI:
1593                 return its_vlpi_prop_update(d, info);
1594
1595         default:
1596                 return -EINVAL;
1597         }
1598 }
1599
1600 static struct irq_chip its_irq_chip = {
1601         .name                   = "ITS",
1602         .irq_mask               = its_mask_irq,
1603         .irq_unmask             = its_unmask_irq,
1604         .irq_eoi                = irq_chip_eoi_parent,
1605         .irq_set_affinity       = its_set_affinity,
1606         .irq_compose_msi_msg    = its_irq_compose_msi_msg,
1607         .irq_set_irqchip_state  = its_irq_set_irqchip_state,
1608         .irq_set_vcpu_affinity  = its_irq_set_vcpu_affinity,
1609 };
1610
1611
1612 /*
1613  * How we allocate LPIs:
1614  *
1615  * lpi_range_list contains ranges of LPIs that are to available to
1616  * allocate from. To allocate LPIs, just pick the first range that
1617  * fits the required allocation, and reduce it by the required
1618  * amount. Once empty, remove the range from the list.
1619  *
1620  * To free a range of LPIs, add a free range to the list, sort it and
1621  * merge the result if the new range happens to be adjacent to an
1622  * already free block.
1623  *
1624  * The consequence of the above is that allocation is cost is low, but
1625  * freeing is expensive. We assumes that freeing rarely occurs.
1626  */
1627 #define ITS_MAX_LPI_NRBITS      16 /* 64K LPIs */
1628
1629 static DEFINE_MUTEX(lpi_range_lock);
1630 static LIST_HEAD(lpi_range_list);
1631
1632 struct lpi_range {
1633         struct list_head        entry;
1634         u32                     base_id;
1635         u32                     span;
1636 };
1637
1638 static struct lpi_range *mk_lpi_range(u32 base, u32 span)
1639 {
1640         struct lpi_range *range;
1641
1642         range = kmalloc(sizeof(*range), GFP_KERNEL);
1643         if (range) {
1644                 range->base_id = base;
1645                 range->span = span;
1646         }
1647
1648         return range;
1649 }
1650
1651 static int alloc_lpi_range(u32 nr_lpis, u32 *base)
1652 {
1653         struct lpi_range *range, *tmp;
1654         int err = -ENOSPC;
1655
1656         mutex_lock(&lpi_range_lock);
1657
1658         list_for_each_entry_safe(range, tmp, &lpi_range_list, entry) {
1659                 if (range->span >= nr_lpis) {
1660                         *base = range->base_id;
1661                         range->base_id += nr_lpis;
1662                         range->span -= nr_lpis;
1663
1664                         if (range->span == 0) {
1665                                 list_del(&range->entry);
1666                                 kfree(range);
1667                         }
1668
1669                         err = 0;
1670                         break;
1671                 }
1672         }
1673
1674         mutex_unlock(&lpi_range_lock);
1675
1676         pr_debug("ITS: alloc %u:%u\n", *base, nr_lpis);
1677         return err;
1678 }
1679
1680 static void merge_lpi_ranges(struct lpi_range *a, struct lpi_range *b)
1681 {
1682         if (&a->entry == &lpi_range_list || &b->entry == &lpi_range_list)
1683                 return;
1684         if (a->base_id + a->span != b->base_id)
1685                 return;
1686         b->base_id = a->base_id;
1687         b->span += a->span;
1688         list_del(&a->entry);
1689         kfree(a);
1690 }
1691
1692 static int free_lpi_range(u32 base, u32 nr_lpis)
1693 {
1694         struct lpi_range *new, *old;
1695
1696         new = mk_lpi_range(base, nr_lpis);
1697         if (!new)
1698                 return -ENOMEM;
1699
1700         mutex_lock(&lpi_range_lock);
1701
1702         list_for_each_entry_reverse(old, &lpi_range_list, entry) {
1703                 if (old->base_id < base)
1704                         break;
1705         }
1706         /*
1707          * old is the last element with ->base_id smaller than base,
1708          * so new goes right after it. If there are no elements with
1709          * ->base_id smaller than base, &old->entry ends up pointing
1710          * at the head of the list, and inserting new it the start of
1711          * the list is the right thing to do in that case as well.
1712          */
1713         list_add(&new->entry, &old->entry);
1714         /*
1715          * Now check if we can merge with the preceding and/or
1716          * following ranges.
1717          */
1718         merge_lpi_ranges(old, new);
1719         merge_lpi_ranges(new, list_next_entry(new, entry));
1720
1721         mutex_unlock(&lpi_range_lock);
1722         return 0;
1723 }
1724
1725 static int __init its_lpi_init(u32 id_bits)
1726 {
1727         u32 lpis = (1UL << id_bits) - 8192;
1728         u32 numlpis;
1729         int err;
1730
1731         numlpis = 1UL << GICD_TYPER_NUM_LPIS(gic_rdists->gicd_typer);
1732
1733         if (numlpis > 2 && !WARN_ON(numlpis > lpis)) {
1734                 lpis = numlpis;
1735                 pr_info("ITS: Using hypervisor restricted LPI range [%u]\n",
1736                         lpis);
1737         }
1738
1739         /*
1740          * Initializing the allocator is just the same as freeing the
1741          * full range of LPIs.
1742          */
1743         err = free_lpi_range(8192, lpis);
1744         pr_debug("ITS: Allocator initialized for %u LPIs\n", lpis);
1745         return err;
1746 }
1747
1748 static unsigned long *its_lpi_alloc(int nr_irqs, u32 *base, int *nr_ids)
1749 {
1750         unsigned long *bitmap = NULL;
1751         int err = 0;
1752
1753         do {
1754                 err = alloc_lpi_range(nr_irqs, base);
1755                 if (!err)
1756                         break;
1757
1758                 nr_irqs /= 2;
1759         } while (nr_irqs > 0);
1760
1761         if (!nr_irqs)
1762                 err = -ENOSPC;
1763
1764         if (err)
1765                 goto out;
1766
1767         bitmap = kcalloc(BITS_TO_LONGS(nr_irqs), sizeof (long), GFP_ATOMIC);
1768         if (!bitmap)
1769                 goto out;
1770
1771         *nr_ids = nr_irqs;
1772
1773 out:
1774         if (!bitmap)
1775                 *base = *nr_ids = 0;
1776
1777         return bitmap;
1778 }
1779
1780 static void its_lpi_free(unsigned long *bitmap, u32 base, u32 nr_ids)
1781 {
1782         WARN_ON(free_lpi_range(base, nr_ids));
1783         kfree(bitmap);
1784 }
1785
1786 static void gic_reset_prop_table(void *va)
1787 {
1788         /* Priority 0xa0, Group-1, disabled */
1789         memset(va, LPI_PROP_DEFAULT_PRIO | LPI_PROP_GROUP1, LPI_PROPBASE_SZ);
1790
1791         /* Make sure the GIC will observe the written configuration */
1792         gic_flush_dcache_to_poc(va, LPI_PROPBASE_SZ);
1793 }
1794
1795 static struct page *its_allocate_prop_table(gfp_t gfp_flags)
1796 {
1797         struct page *prop_page;
1798
1799         prop_page = alloc_pages(gfp_flags, get_order(LPI_PROPBASE_SZ));
1800         if (!prop_page)
1801                 return NULL;
1802
1803         gic_reset_prop_table(page_address(prop_page));
1804
1805         return prop_page;
1806 }
1807
1808 static void its_free_prop_table(struct page *prop_page)
1809 {
1810         free_pages((unsigned long)page_address(prop_page),
1811                    get_order(LPI_PROPBASE_SZ));
1812 }
1813
1814 static bool gic_check_reserved_range(phys_addr_t addr, unsigned long size)
1815 {
1816         phys_addr_t start, end, addr_end;
1817         u64 i;
1818
1819         /*
1820          * We don't bother checking for a kdump kernel as by
1821          * construction, the LPI tables are out of this kernel's
1822          * memory map.
1823          */
1824         if (is_kdump_kernel())
1825                 return true;
1826
1827         addr_end = addr + size - 1;
1828
1829         for_each_reserved_mem_region(i, &start, &end) {
1830                 if (addr >= start && addr_end <= end)
1831                         return true;
1832         }
1833
1834         /* Not found, not a good sign... */
1835         pr_warn("GICv3: Expected reserved range [%pa:%pa], not found\n",
1836                 &addr, &addr_end);
1837         add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
1838         return false;
1839 }
1840
1841 static int gic_reserve_range(phys_addr_t addr, unsigned long size)
1842 {
1843         if (efi_enabled(EFI_CONFIG_TABLES))
1844                 return efi_mem_reserve_persistent(addr, size);
1845
1846         return 0;
1847 }
1848
1849 static int __init its_setup_lpi_prop_table(void)
1850 {
1851         if (gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED) {
1852                 u64 val;
1853
1854                 val = gicr_read_propbaser(gic_data_rdist_rd_base() + GICR_PROPBASER);
1855                 lpi_id_bits = (val & GICR_PROPBASER_IDBITS_MASK) + 1;
1856
1857                 gic_rdists->prop_table_pa = val & GENMASK_ULL(51, 12);
1858                 gic_rdists->prop_table_va = memremap(gic_rdists->prop_table_pa,
1859                                                      LPI_PROPBASE_SZ,
1860                                                      MEMREMAP_WB);
1861                 gic_reset_prop_table(gic_rdists->prop_table_va);
1862         } else {
1863                 struct page *page;
1864
1865                 lpi_id_bits = min_t(u32,
1866                                     GICD_TYPER_ID_BITS(gic_rdists->gicd_typer),
1867                                     ITS_MAX_LPI_NRBITS);
1868                 page = its_allocate_prop_table(GFP_NOWAIT);
1869                 if (!page) {
1870                         pr_err("Failed to allocate PROPBASE\n");
1871                         return -ENOMEM;
1872                 }
1873
1874                 gic_rdists->prop_table_pa = page_to_phys(page);
1875                 gic_rdists->prop_table_va = page_address(page);
1876                 WARN_ON(gic_reserve_range(gic_rdists->prop_table_pa,
1877                                           LPI_PROPBASE_SZ));
1878         }
1879
1880         pr_info("GICv3: using LPI property table @%pa\n",
1881                 &gic_rdists->prop_table_pa);
1882
1883         return its_lpi_init(lpi_id_bits);
1884 }
1885
1886 static const char *its_base_type_string[] = {
1887         [GITS_BASER_TYPE_DEVICE]        = "Devices",
1888         [GITS_BASER_TYPE_VCPU]          = "Virtual CPUs",
1889         [GITS_BASER_TYPE_RESERVED3]     = "Reserved (3)",
1890         [GITS_BASER_TYPE_COLLECTION]    = "Interrupt Collections",
1891         [GITS_BASER_TYPE_RESERVED5]     = "Reserved (5)",
1892         [GITS_BASER_TYPE_RESERVED6]     = "Reserved (6)",
1893         [GITS_BASER_TYPE_RESERVED7]     = "Reserved (7)",
1894 };
1895
1896 static u64 its_read_baser(struct its_node *its, struct its_baser *baser)
1897 {
1898         u32 idx = baser - its->tables;
1899
1900         return gits_read_baser(its->base + GITS_BASER + (idx << 3));
1901 }
1902
1903 static void its_write_baser(struct its_node *its, struct its_baser *baser,
1904                             u64 val)
1905 {
1906         u32 idx = baser - its->tables;
1907
1908         gits_write_baser(val, its->base + GITS_BASER + (idx << 3));
1909         baser->val = its_read_baser(its, baser);
1910 }
1911
1912 static int its_setup_baser(struct its_node *its, struct its_baser *baser,
1913                            u64 cache, u64 shr, u32 psz, u32 order,
1914                            bool indirect)
1915 {
1916         u64 val = its_read_baser(its, baser);
1917         u64 esz = GITS_BASER_ENTRY_SIZE(val);
1918         u64 type = GITS_BASER_TYPE(val);
1919         u64 baser_phys, tmp;
1920         u32 alloc_pages;
1921         struct page *page;
1922         void *base;
1923
1924 retry_alloc_baser:
1925         alloc_pages = (PAGE_ORDER_TO_SIZE(order) / psz);
1926         if (alloc_pages > GITS_BASER_PAGES_MAX) {
1927                 pr_warn("ITS@%pa: %s too large, reduce ITS pages %u->%u\n",
1928                         &its->phys_base, its_base_type_string[type],
1929                         alloc_pages, GITS_BASER_PAGES_MAX);
1930                 alloc_pages = GITS_BASER_PAGES_MAX;
1931                 order = get_order(GITS_BASER_PAGES_MAX * psz);
1932         }
1933
1934         page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO, order);
1935         if (!page)
1936                 return -ENOMEM;
1937
1938         base = (void *)page_address(page);
1939         baser_phys = virt_to_phys(base);
1940
1941         /* Check if the physical address of the memory is above 48bits */
1942         if (IS_ENABLED(CONFIG_ARM64_64K_PAGES) && (baser_phys >> 48)) {
1943
1944                 /* 52bit PA is supported only when PageSize=64K */
1945                 if (psz != SZ_64K) {
1946                         pr_err("ITS: no 52bit PA support when psz=%d\n", psz);
1947                         free_pages((unsigned long)base, order);
1948                         return -ENXIO;
1949                 }
1950
1951                 /* Convert 52bit PA to 48bit field */
1952                 baser_phys = GITS_BASER_PHYS_52_to_48(baser_phys);
1953         }
1954
1955 retry_baser:
1956         val = (baser_phys                                        |
1957                 (type << GITS_BASER_TYPE_SHIFT)                  |
1958                 ((esz - 1) << GITS_BASER_ENTRY_SIZE_SHIFT)       |
1959                 ((alloc_pages - 1) << GITS_BASER_PAGES_SHIFT)    |
1960                 cache                                            |
1961                 shr                                              |
1962                 GITS_BASER_VALID);
1963
1964         val |=  indirect ? GITS_BASER_INDIRECT : 0x0;
1965
1966         switch (psz) {
1967         case SZ_4K:
1968                 val |= GITS_BASER_PAGE_SIZE_4K;
1969                 break;
1970         case SZ_16K:
1971                 val |= GITS_BASER_PAGE_SIZE_16K;
1972                 break;
1973         case SZ_64K:
1974                 val |= GITS_BASER_PAGE_SIZE_64K;
1975                 break;
1976         }
1977
1978         its_write_baser(its, baser, val);
1979         tmp = baser->val;
1980
1981         if ((val ^ tmp) & GITS_BASER_SHAREABILITY_MASK) {
1982                 /*
1983                  * Shareability didn't stick. Just use
1984                  * whatever the read reported, which is likely
1985                  * to be the only thing this redistributor
1986                  * supports. If that's zero, make it
1987                  * non-cacheable as well.
1988                  */
1989                 shr = tmp & GITS_BASER_SHAREABILITY_MASK;
1990                 if (!shr) {
1991                         cache = GITS_BASER_nC;
1992                         gic_flush_dcache_to_poc(base, PAGE_ORDER_TO_SIZE(order));
1993                 }
1994                 goto retry_baser;
1995         }
1996
1997         if ((val ^ tmp) & GITS_BASER_PAGE_SIZE_MASK) {
1998                 /*
1999                  * Page size didn't stick. Let's try a smaller
2000                  * size and retry. If we reach 4K, then
2001                  * something is horribly wrong...
2002                  */
2003                 free_pages((unsigned long)base, order);
2004                 baser->base = NULL;
2005
2006                 switch (psz) {
2007                 case SZ_16K:
2008                         psz = SZ_4K;
2009                         goto retry_alloc_baser;
2010                 case SZ_64K:
2011                         psz = SZ_16K;
2012                         goto retry_alloc_baser;
2013                 }
2014         }
2015
2016         if (val != tmp) {
2017                 pr_err("ITS@%pa: %s doesn't stick: %llx %llx\n",
2018                        &its->phys_base, its_base_type_string[type],
2019                        val, tmp);
2020                 free_pages((unsigned long)base, order);
2021                 return -ENXIO;
2022         }
2023
2024         baser->order = order;
2025         baser->base = base;
2026         baser->psz = psz;
2027         tmp = indirect ? GITS_LVL1_ENTRY_SIZE : esz;
2028
2029         pr_info("ITS@%pa: allocated %d %s @%lx (%s, esz %d, psz %dK, shr %d)\n",
2030                 &its->phys_base, (int)(PAGE_ORDER_TO_SIZE(order) / (int)tmp),
2031                 its_base_type_string[type],
2032                 (unsigned long)virt_to_phys(base),
2033                 indirect ? "indirect" : "flat", (int)esz,
2034                 psz / SZ_1K, (int)shr >> GITS_BASER_SHAREABILITY_SHIFT);
2035
2036         return 0;
2037 }
2038
2039 static bool its_parse_indirect_baser(struct its_node *its,
2040                                      struct its_baser *baser,
2041                                      u32 psz, u32 *order, u32 ids)
2042 {
2043         u64 tmp = its_read_baser(its, baser);
2044         u64 type = GITS_BASER_TYPE(tmp);
2045         u64 esz = GITS_BASER_ENTRY_SIZE(tmp);
2046         u64 val = GITS_BASER_InnerShareable | GITS_BASER_RaWaWb;
2047         u32 new_order = *order;
2048         bool indirect = false;
2049
2050         /* No need to enable Indirection if memory requirement < (psz*2)bytes */
2051         if ((esz << ids) > (psz * 2)) {
2052                 /*
2053                  * Find out whether hw supports a single or two-level table by
2054                  * table by reading bit at offset '62' after writing '1' to it.
2055                  */
2056                 its_write_baser(its, baser, val | GITS_BASER_INDIRECT);
2057                 indirect = !!(baser->val & GITS_BASER_INDIRECT);
2058
2059                 if (indirect) {
2060                         /*
2061                          * The size of the lvl2 table is equal to ITS page size
2062                          * which is 'psz'. For computing lvl1 table size,
2063                          * subtract ID bits that sparse lvl2 table from 'ids'
2064                          * which is reported by ITS hardware times lvl1 table
2065                          * entry size.
2066                          */
2067                         ids -= ilog2(psz / (int)esz);
2068                         esz = GITS_LVL1_ENTRY_SIZE;
2069                 }
2070         }
2071
2072         /*
2073          * Allocate as many entries as required to fit the
2074          * range of device IDs that the ITS can grok... The ID
2075          * space being incredibly sparse, this results in a
2076          * massive waste of memory if two-level device table
2077          * feature is not supported by hardware.
2078          */
2079         new_order = max_t(u32, get_order(esz << ids), new_order);
2080         if (new_order >= MAX_ORDER) {
2081                 new_order = MAX_ORDER - 1;
2082                 ids = ilog2(PAGE_ORDER_TO_SIZE(new_order) / (int)esz);
2083                 pr_warn("ITS@%pa: %s Table too large, reduce ids %llu->%u\n",
2084                         &its->phys_base, its_base_type_string[type],
2085                         device_ids(its), ids);
2086         }
2087
2088         *order = new_order;
2089
2090         return indirect;
2091 }
2092
2093 static void its_free_tables(struct its_node *its)
2094 {
2095         int i;
2096
2097         for (i = 0; i < GITS_BASER_NR_REGS; i++) {
2098                 if (its->tables[i].base) {
2099                         free_pages((unsigned long)its->tables[i].base,
2100                                    its->tables[i].order);
2101                         its->tables[i].base = NULL;
2102                 }
2103         }
2104 }
2105
2106 static int its_alloc_tables(struct its_node *its)
2107 {
2108         u64 shr = GITS_BASER_InnerShareable;
2109         u64 cache = GITS_BASER_RaWaWb;
2110         u32 psz = SZ_64K;
2111         int err, i;
2112
2113         if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_22375)
2114                 /* erratum 24313: ignore memory access type */
2115                 cache = GITS_BASER_nCnB;
2116
2117         for (i = 0; i < GITS_BASER_NR_REGS; i++) {
2118                 struct its_baser *baser = its->tables + i;
2119                 u64 val = its_read_baser(its, baser);
2120                 u64 type = GITS_BASER_TYPE(val);
2121                 u32 order = get_order(psz);
2122                 bool indirect = false;
2123
2124                 switch (type) {
2125                 case GITS_BASER_TYPE_NONE:
2126                         continue;
2127
2128                 case GITS_BASER_TYPE_DEVICE:
2129                         indirect = its_parse_indirect_baser(its, baser,
2130                                                             psz, &order,
2131                                                             device_ids(its));
2132                         break;
2133
2134                 case GITS_BASER_TYPE_VCPU:
2135                         indirect = its_parse_indirect_baser(its, baser,
2136                                                             psz, &order,
2137                                                             ITS_MAX_VPEID_BITS);
2138                         break;
2139                 }
2140
2141                 err = its_setup_baser(its, baser, cache, shr, psz, order, indirect);
2142                 if (err < 0) {
2143                         its_free_tables(its);
2144                         return err;
2145                 }
2146
2147                 /* Update settings which will be used for next BASERn */
2148                 psz = baser->psz;
2149                 cache = baser->val & GITS_BASER_CACHEABILITY_MASK;
2150                 shr = baser->val & GITS_BASER_SHAREABILITY_MASK;
2151         }
2152
2153         return 0;
2154 }
2155
2156 static int its_alloc_collections(struct its_node *its)
2157 {
2158         int i;
2159
2160         its->collections = kcalloc(nr_cpu_ids, sizeof(*its->collections),
2161                                    GFP_KERNEL);
2162         if (!its->collections)
2163                 return -ENOMEM;
2164
2165         for (i = 0; i < nr_cpu_ids; i++)
2166                 its->collections[i].target_address = ~0ULL;
2167
2168         return 0;
2169 }
2170
2171 static struct page *its_allocate_pending_table(gfp_t gfp_flags)
2172 {
2173         struct page *pend_page;
2174
2175         pend_page = alloc_pages(gfp_flags | __GFP_ZERO,
2176                                 get_order(LPI_PENDBASE_SZ));
2177         if (!pend_page)
2178                 return NULL;
2179
2180         /* Make sure the GIC will observe the zero-ed page */
2181         gic_flush_dcache_to_poc(page_address(pend_page), LPI_PENDBASE_SZ);
2182
2183         return pend_page;
2184 }
2185
2186 static void its_free_pending_table(struct page *pt)
2187 {
2188         free_pages((unsigned long)page_address(pt), get_order(LPI_PENDBASE_SZ));
2189 }
2190
2191 /*
2192  * Booting with kdump and LPIs enabled is generally fine. Any other
2193  * case is wrong in the absence of firmware/EFI support.
2194  */
2195 static bool enabled_lpis_allowed(void)
2196 {
2197         phys_addr_t addr;
2198         u64 val;
2199
2200         /* Check whether the property table is in a reserved region */
2201         val = gicr_read_propbaser(gic_data_rdist_rd_base() + GICR_PROPBASER);
2202         addr = val & GENMASK_ULL(51, 12);
2203
2204         return gic_check_reserved_range(addr, LPI_PROPBASE_SZ);
2205 }
2206
2207 static int __init allocate_lpi_tables(void)
2208 {
2209         u64 val;
2210         int err, cpu;
2211
2212         /*
2213          * If LPIs are enabled while we run this from the boot CPU,
2214          * flag the RD tables as pre-allocated if the stars do align.
2215          */
2216         val = readl_relaxed(gic_data_rdist_rd_base() + GICR_CTLR);
2217         if ((val & GICR_CTLR_ENABLE_LPIS) && enabled_lpis_allowed()) {
2218                 gic_rdists->flags |= (RDIST_FLAGS_RD_TABLES_PREALLOCATED |
2219                                       RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING);
2220                 pr_info("GICv3: Using preallocated redistributor tables\n");
2221         }
2222
2223         err = its_setup_lpi_prop_table();
2224         if (err)
2225                 return err;
2226
2227         /*
2228          * We allocate all the pending tables anyway, as we may have a
2229          * mix of RDs that have had LPIs enabled, and some that
2230          * don't. We'll free the unused ones as each CPU comes online.
2231          */
2232         for_each_possible_cpu(cpu) {
2233                 struct page *pend_page;
2234
2235                 pend_page = its_allocate_pending_table(GFP_NOWAIT);
2236                 if (!pend_page) {
2237                         pr_err("Failed to allocate PENDBASE for CPU%d\n", cpu);
2238                         return -ENOMEM;
2239                 }
2240
2241                 gic_data_rdist_cpu(cpu)->pend_page = pend_page;
2242         }
2243
2244         return 0;
2245 }
2246
2247 static u64 its_clear_vpend_valid(void __iomem *vlpi_base)
2248 {
2249         u32 count = 1000000;    /* 1s! */
2250         bool clean;
2251         u64 val;
2252
2253         val = gits_read_vpendbaser(vlpi_base + GICR_VPENDBASER);
2254         val &= ~GICR_VPENDBASER_Valid;
2255         gits_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
2256
2257         do {
2258                 val = gits_read_vpendbaser(vlpi_base + GICR_VPENDBASER);
2259                 clean = !(val & GICR_VPENDBASER_Dirty);
2260                 if (!clean) {
2261                         count--;
2262                         cpu_relax();
2263                         udelay(1);
2264                 }
2265         } while (!clean && count);
2266
2267         return val;
2268 }
2269
2270 static void its_cpu_init_lpis(void)
2271 {
2272         void __iomem *rbase = gic_data_rdist_rd_base();
2273         struct page *pend_page;
2274         phys_addr_t paddr;
2275         u64 val, tmp;
2276
2277         if (gic_data_rdist()->lpi_enabled)
2278                 return;
2279
2280         val = readl_relaxed(rbase + GICR_CTLR);
2281         if ((gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED) &&
2282             (val & GICR_CTLR_ENABLE_LPIS)) {
2283                 /*
2284                  * Check that we get the same property table on all
2285                  * RDs. If we don't, this is hopeless.
2286                  */
2287                 paddr = gicr_read_propbaser(rbase + GICR_PROPBASER);
2288                 paddr &= GENMASK_ULL(51, 12);
2289                 if (WARN_ON(gic_rdists->prop_table_pa != paddr))
2290                         add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
2291
2292                 paddr = gicr_read_pendbaser(rbase + GICR_PENDBASER);
2293                 paddr &= GENMASK_ULL(51, 16);
2294
2295                 WARN_ON(!gic_check_reserved_range(paddr, LPI_PENDBASE_SZ));
2296                 its_free_pending_table(gic_data_rdist()->pend_page);
2297                 gic_data_rdist()->pend_page = NULL;
2298
2299                 goto out;
2300         }
2301
2302         pend_page = gic_data_rdist()->pend_page;
2303         paddr = page_to_phys(pend_page);
2304         WARN_ON(gic_reserve_range(paddr, LPI_PENDBASE_SZ));
2305
2306         /* set PROPBASE */
2307         val = (gic_rdists->prop_table_pa |
2308                GICR_PROPBASER_InnerShareable |
2309                GICR_PROPBASER_RaWaWb |
2310                ((LPI_NRBITS - 1) & GICR_PROPBASER_IDBITS_MASK));
2311
2312         gicr_write_propbaser(val, rbase + GICR_PROPBASER);
2313         tmp = gicr_read_propbaser(rbase + GICR_PROPBASER);
2314
2315         if ((tmp ^ val) & GICR_PROPBASER_SHAREABILITY_MASK) {
2316                 if (!(tmp & GICR_PROPBASER_SHAREABILITY_MASK)) {
2317                         /*
2318                          * The HW reports non-shareable, we must
2319                          * remove the cacheability attributes as
2320                          * well.
2321                          */
2322                         val &= ~(GICR_PROPBASER_SHAREABILITY_MASK |
2323                                  GICR_PROPBASER_CACHEABILITY_MASK);
2324                         val |= GICR_PROPBASER_nC;
2325                         gicr_write_propbaser(val, rbase + GICR_PROPBASER);
2326                 }
2327                 pr_info_once("GIC: using cache flushing for LPI property table\n");
2328                 gic_rdists->flags |= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING;
2329         }
2330
2331         /* set PENDBASE */
2332         val = (page_to_phys(pend_page) |
2333                GICR_PENDBASER_InnerShareable |
2334                GICR_PENDBASER_RaWaWb);
2335
2336         gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
2337         tmp = gicr_read_pendbaser(rbase + GICR_PENDBASER);
2338
2339         if (!(tmp & GICR_PENDBASER_SHAREABILITY_MASK)) {
2340                 /*
2341                  * The HW reports non-shareable, we must remove the
2342                  * cacheability attributes as well.
2343                  */
2344                 val &= ~(GICR_PENDBASER_SHAREABILITY_MASK |
2345                          GICR_PENDBASER_CACHEABILITY_MASK);
2346                 val |= GICR_PENDBASER_nC;
2347                 gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
2348         }
2349
2350         /* Enable LPIs */
2351         val = readl_relaxed(rbase + GICR_CTLR);
2352         val |= GICR_CTLR_ENABLE_LPIS;
2353         writel_relaxed(val, rbase + GICR_CTLR);
2354
2355         if (gic_rdists->has_vlpis) {
2356                 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2357
2358                 /*
2359                  * It's possible for CPU to receive VLPIs before it is
2360                  * sheduled as a vPE, especially for the first CPU, and the
2361                  * VLPI with INTID larger than 2^(IDbits+1) will be considered
2362                  * as out of range and dropped by GIC.
2363                  * So we initialize IDbits to known value to avoid VLPI drop.
2364                  */
2365                 val = (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
2366                 pr_debug("GICv4: CPU%d: Init IDbits to 0x%llx for GICR_VPROPBASER\n",
2367                         smp_processor_id(), val);
2368                 gits_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2369
2370                 /*
2371                  * Also clear Valid bit of GICR_VPENDBASER, in case some
2372                  * ancient programming gets left in and has possibility of
2373                  * corrupting memory.
2374                  */
2375                 val = its_clear_vpend_valid(vlpi_base);
2376                 WARN_ON(val & GICR_VPENDBASER_Dirty);
2377         }
2378
2379         /* Make sure the GIC has seen the above */
2380         dsb(sy);
2381 out:
2382         gic_data_rdist()->lpi_enabled = true;
2383         pr_info("GICv3: CPU%d: using %s LPI pending table @%pa\n",
2384                 smp_processor_id(),
2385                 gic_data_rdist()->pend_page ? "allocated" : "reserved",
2386                 &paddr);
2387 }
2388
2389 static void its_cpu_init_collection(struct its_node *its)
2390 {
2391         int cpu = smp_processor_id();
2392         u64 target;
2393
2394         /* avoid cross node collections and its mapping */
2395         if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) {
2396                 struct device_node *cpu_node;
2397
2398                 cpu_node = of_get_cpu_node(cpu, NULL);
2399                 if (its->numa_node != NUMA_NO_NODE &&
2400                         its->numa_node != of_node_to_nid(cpu_node))
2401                         return;
2402         }
2403
2404         /*
2405          * We now have to bind each collection to its target
2406          * redistributor.
2407          */
2408         if (gic_read_typer(its->base + GITS_TYPER) & GITS_TYPER_PTA) {
2409                 /*
2410                  * This ITS wants the physical address of the
2411                  * redistributor.
2412                  */
2413                 target = gic_data_rdist()->phys_base;
2414         } else {
2415                 /* This ITS wants a linear CPU number. */
2416                 target = gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER);
2417                 target = GICR_TYPER_CPU_NUMBER(target) << 16;
2418         }
2419
2420         /* Perform collection mapping */
2421         its->collections[cpu].target_address = target;
2422         its->collections[cpu].col_id = cpu;
2423
2424         its_send_mapc(its, &its->collections[cpu], 1);
2425         its_send_invall(its, &its->collections[cpu]);
2426 }
2427
2428 static void its_cpu_init_collections(void)
2429 {
2430         struct its_node *its;
2431
2432         raw_spin_lock(&its_lock);
2433
2434         list_for_each_entry(its, &its_nodes, entry)
2435                 its_cpu_init_collection(its);
2436
2437         raw_spin_unlock(&its_lock);
2438 }
2439
2440 static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
2441 {
2442         struct its_device *its_dev = NULL, *tmp;
2443         unsigned long flags;
2444
2445         raw_spin_lock_irqsave(&its->lock, flags);
2446
2447         list_for_each_entry(tmp, &its->its_device_list, entry) {
2448                 if (tmp->device_id == dev_id) {
2449                         its_dev = tmp;
2450                         break;
2451                 }
2452         }
2453
2454         raw_spin_unlock_irqrestore(&its->lock, flags);
2455
2456         return its_dev;
2457 }
2458
2459 static struct its_baser *its_get_baser(struct its_node *its, u32 type)
2460 {
2461         int i;
2462
2463         for (i = 0; i < GITS_BASER_NR_REGS; i++) {
2464                 if (GITS_BASER_TYPE(its->tables[i].val) == type)
2465                         return &its->tables[i];
2466         }
2467
2468         return NULL;
2469 }
2470
2471 static bool its_alloc_table_entry(struct its_node *its,
2472                                   struct its_baser *baser, u32 id)
2473 {
2474         struct page *page;
2475         u32 esz, idx;
2476         __le64 *table;
2477
2478         /* Don't allow device id that exceeds single, flat table limit */
2479         esz = GITS_BASER_ENTRY_SIZE(baser->val);
2480         if (!(baser->val & GITS_BASER_INDIRECT))
2481                 return (id < (PAGE_ORDER_TO_SIZE(baser->order) / esz));
2482
2483         /* Compute 1st level table index & check if that exceeds table limit */
2484         idx = id >> ilog2(baser->psz / esz);
2485         if (idx >= (PAGE_ORDER_TO_SIZE(baser->order) / GITS_LVL1_ENTRY_SIZE))
2486                 return false;
2487
2488         table = baser->base;
2489
2490         /* Allocate memory for 2nd level table */
2491         if (!table[idx]) {
2492                 page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO,
2493                                         get_order(baser->psz));
2494                 if (!page)
2495                         return false;
2496
2497                 /* Flush Lvl2 table to PoC if hw doesn't support coherency */
2498                 if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
2499                         gic_flush_dcache_to_poc(page_address(page), baser->psz);
2500
2501                 table[idx] = cpu_to_le64(page_to_phys(page) | GITS_BASER_VALID);
2502
2503                 /* Flush Lvl1 entry to PoC if hw doesn't support coherency */
2504                 if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
2505                         gic_flush_dcache_to_poc(table + idx, GITS_LVL1_ENTRY_SIZE);
2506
2507                 /* Ensure updated table contents are visible to ITS hardware */
2508                 dsb(sy);
2509         }
2510
2511         return true;
2512 }
2513
2514 static bool its_alloc_device_table(struct its_node *its, u32 dev_id)
2515 {
2516         struct its_baser *baser;
2517
2518         baser = its_get_baser(its, GITS_BASER_TYPE_DEVICE);
2519
2520         /* Don't allow device id that exceeds ITS hardware limit */
2521         if (!baser)
2522                 return (ilog2(dev_id) < device_ids(its));
2523
2524         return its_alloc_table_entry(its, baser, dev_id);
2525 }
2526
2527 static bool its_alloc_vpe_table(u32 vpe_id)
2528 {
2529         struct its_node *its;
2530
2531         /*
2532          * Make sure the L2 tables are allocated on *all* v4 ITSs. We
2533          * could try and only do it on ITSs corresponding to devices
2534          * that have interrupts targeted at this VPE, but the
2535          * complexity becomes crazy (and you have tons of memory
2536          * anyway, right?).
2537          */
2538         list_for_each_entry(its, &its_nodes, entry) {
2539                 struct its_baser *baser;
2540
2541                 if (!is_v4(its))
2542                         continue;
2543
2544                 baser = its_get_baser(its, GITS_BASER_TYPE_VCPU);
2545                 if (!baser)
2546                         return false;
2547
2548                 if (!its_alloc_table_entry(its, baser, vpe_id))
2549                         return false;
2550         }
2551
2552         return true;
2553 }
2554
2555 static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
2556                                             int nvecs, bool alloc_lpis)
2557 {
2558         struct its_device *dev;
2559         unsigned long *lpi_map = NULL;
2560         unsigned long flags;
2561         u16 *col_map = NULL;
2562         void *itt;
2563         int lpi_base;
2564         int nr_lpis;
2565         int nr_ites;
2566         int sz;
2567
2568         if (!its_alloc_device_table(its, dev_id))
2569                 return NULL;
2570
2571         if (WARN_ON(!is_power_of_2(nvecs)))
2572                 nvecs = roundup_pow_of_two(nvecs);
2573
2574         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2575         /*
2576          * Even if the device wants a single LPI, the ITT must be
2577          * sized as a power of two (and you need at least one bit...).
2578          */
2579         nr_ites = max(2, nvecs);
2580         sz = nr_ites * (FIELD_GET(GITS_TYPER_ITT_ENTRY_SIZE, its->typer) + 1);
2581         sz = max(sz, ITS_ITT_ALIGN) + ITS_ITT_ALIGN - 1;
2582         itt = kzalloc_node(sz, GFP_KERNEL, its->numa_node);
2583         if (alloc_lpis) {
2584                 lpi_map = its_lpi_alloc(nvecs, &lpi_base, &nr_lpis);
2585                 if (lpi_map)
2586                         col_map = kcalloc(nr_lpis, sizeof(*col_map),
2587                                           GFP_KERNEL);
2588         } else {
2589                 col_map = kcalloc(nr_ites, sizeof(*col_map), GFP_KERNEL);
2590                 nr_lpis = 0;
2591                 lpi_base = 0;
2592         }
2593
2594         if (!dev || !itt ||  !col_map || (!lpi_map && alloc_lpis)) {
2595                 kfree(dev);
2596                 kfree(itt);
2597                 kfree(lpi_map);
2598                 kfree(col_map);
2599                 return NULL;
2600         }
2601
2602         gic_flush_dcache_to_poc(itt, sz);
2603
2604         dev->its = its;
2605         dev->itt = itt;
2606         dev->nr_ites = nr_ites;
2607         dev->event_map.lpi_map = lpi_map;
2608         dev->event_map.col_map = col_map;
2609         dev->event_map.lpi_base = lpi_base;
2610         dev->event_map.nr_lpis = nr_lpis;
2611         raw_spin_lock_init(&dev->event_map.vlpi_lock);
2612         dev->device_id = dev_id;
2613         INIT_LIST_HEAD(&dev->entry);
2614
2615         raw_spin_lock_irqsave(&its->lock, flags);
2616         list_add(&dev->entry, &its->its_device_list);
2617         raw_spin_unlock_irqrestore(&its->lock, flags);
2618
2619         /* Map device to its ITT */
2620         its_send_mapd(dev, 1);
2621
2622         return dev;
2623 }
2624
2625 static void its_free_device(struct its_device *its_dev)
2626 {
2627         unsigned long flags;
2628
2629         raw_spin_lock_irqsave(&its_dev->its->lock, flags);
2630         list_del(&its_dev->entry);
2631         raw_spin_unlock_irqrestore(&its_dev->its->lock, flags);
2632         kfree(its_dev->event_map.col_map);
2633         kfree(its_dev->itt);
2634         kfree(its_dev);
2635 }
2636
2637 static int its_alloc_device_irq(struct its_device *dev, int nvecs, irq_hw_number_t *hwirq)
2638 {
2639         int idx;
2640
2641         /* Find a free LPI region in lpi_map and allocate them. */
2642         idx = bitmap_find_free_region(dev->event_map.lpi_map,
2643                                       dev->event_map.nr_lpis,
2644                                       get_count_order(nvecs));
2645         if (idx < 0)
2646                 return -ENOSPC;
2647
2648         *hwirq = dev->event_map.lpi_base + idx;
2649
2650         return 0;
2651 }
2652
2653 static int its_msi_prepare(struct irq_domain *domain, struct device *dev,
2654                            int nvec, msi_alloc_info_t *info)
2655 {
2656         struct its_node *its;
2657         struct its_device *its_dev;
2658         struct msi_domain_info *msi_info;
2659         u32 dev_id;
2660         int err = 0;
2661
2662         /*
2663          * We ignore "dev" entirely, and rely on the dev_id that has
2664          * been passed via the scratchpad. This limits this domain's
2665          * usefulness to upper layers that definitely know that they
2666          * are built on top of the ITS.
2667          */
2668         dev_id = info->scratchpad[0].ul;
2669
2670         msi_info = msi_get_domain_info(domain);
2671         its = msi_info->data;
2672
2673         if (!gic_rdists->has_direct_lpi &&
2674             vpe_proxy.dev &&
2675             vpe_proxy.dev->its == its &&
2676             dev_id == vpe_proxy.dev->device_id) {
2677                 /* Bad luck. Get yourself a better implementation */
2678                 WARN_ONCE(1, "DevId %x clashes with GICv4 VPE proxy device\n",
2679                           dev_id);
2680                 return -EINVAL;
2681         }
2682
2683         mutex_lock(&its->dev_alloc_lock);
2684         its_dev = its_find_device(its, dev_id);
2685         if (its_dev) {
2686                 /*
2687                  * We already have seen this ID, probably through
2688                  * another alias (PCI bridge of some sort). No need to
2689                  * create the device.
2690                  */
2691                 its_dev->shared = true;
2692                 pr_debug("Reusing ITT for devID %x\n", dev_id);
2693                 goto out;
2694         }
2695
2696         its_dev = its_create_device(its, dev_id, nvec, true);
2697         if (!its_dev) {
2698                 err = -ENOMEM;
2699                 goto out;
2700         }
2701
2702         pr_debug("ITT %d entries, %d bits\n", nvec, ilog2(nvec));
2703 out:
2704         mutex_unlock(&its->dev_alloc_lock);
2705         info->scratchpad[0].ptr = its_dev;
2706         return err;
2707 }
2708
2709 static struct msi_domain_ops its_msi_domain_ops = {
2710         .msi_prepare    = its_msi_prepare,
2711 };
2712
2713 static int its_irq_gic_domain_alloc(struct irq_domain *domain,
2714                                     unsigned int virq,
2715                                     irq_hw_number_t hwirq)
2716 {
2717         struct irq_fwspec fwspec;
2718
2719         if (irq_domain_get_of_node(domain->parent)) {
2720                 fwspec.fwnode = domain->parent->fwnode;
2721                 fwspec.param_count = 3;
2722                 fwspec.param[0] = GIC_IRQ_TYPE_LPI;
2723                 fwspec.param[1] = hwirq;
2724                 fwspec.param[2] = IRQ_TYPE_EDGE_RISING;
2725         } else if (is_fwnode_irqchip(domain->parent->fwnode)) {
2726                 fwspec.fwnode = domain->parent->fwnode;
2727                 fwspec.param_count = 2;
2728                 fwspec.param[0] = hwirq;
2729                 fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
2730         } else {
2731                 return -EINVAL;
2732         }
2733
2734         return irq_domain_alloc_irqs_parent(domain, virq, 1, &fwspec);
2735 }
2736
2737 static int its_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
2738                                 unsigned int nr_irqs, void *args)
2739 {
2740         msi_alloc_info_t *info = args;
2741         struct its_device *its_dev = info->scratchpad[0].ptr;
2742         struct its_node *its = its_dev->its;
2743         irq_hw_number_t hwirq;
2744         int err;
2745         int i;
2746
2747         err = its_alloc_device_irq(its_dev, nr_irqs, &hwirq);
2748         if (err)
2749                 return err;
2750
2751         err = iommu_dma_prepare_msi(info->desc, its->get_msi_base(its_dev));
2752         if (err)
2753                 return err;
2754
2755         for (i = 0; i < nr_irqs; i++) {
2756                 err = its_irq_gic_domain_alloc(domain, virq + i, hwirq + i);
2757                 if (err)
2758                         return err;
2759
2760                 irq_domain_set_hwirq_and_chip(domain, virq + i,
2761                                               hwirq + i, &its_irq_chip, its_dev);
2762                 irqd_set_single_target(irq_desc_get_irq_data(irq_to_desc(virq + i)));
2763                 pr_debug("ID:%d pID:%d vID:%d\n",
2764                          (int)(hwirq + i - its_dev->event_map.lpi_base),
2765                          (int)(hwirq + i), virq + i);
2766         }
2767
2768         return 0;
2769 }
2770
2771 static int its_irq_domain_activate(struct irq_domain *domain,
2772                                    struct irq_data *d, bool reserve)
2773 {
2774         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2775         u32 event = its_get_event_id(d);
2776         const struct cpumask *cpu_mask = cpu_online_mask;
2777         int cpu;
2778
2779         /* get the cpu_mask of local node */
2780         if (its_dev->its->numa_node >= 0)
2781                 cpu_mask = cpumask_of_node(its_dev->its->numa_node);
2782
2783         /* Bind the LPI to the first possible CPU */
2784         cpu = cpumask_first_and(cpu_mask, cpu_online_mask);
2785         if (cpu >= nr_cpu_ids) {
2786                 if (its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144)
2787                         return -EINVAL;
2788
2789                 cpu = cpumask_first(cpu_online_mask);
2790         }
2791
2792         its_dev->event_map.col_map[event] = cpu;
2793         irq_data_update_effective_affinity(d, cpumask_of(cpu));
2794
2795         /* Map the GIC IRQ and event to the device */
2796         its_send_mapti(its_dev, d->hwirq, event);
2797         return 0;
2798 }
2799
2800 static void its_irq_domain_deactivate(struct irq_domain *domain,
2801                                       struct irq_data *d)
2802 {
2803         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2804         u32 event = its_get_event_id(d);
2805
2806         /* Stop the delivery of interrupts */
2807         its_send_discard(its_dev, event);
2808 }
2809
2810 static void its_irq_domain_free(struct irq_domain *domain, unsigned int virq,
2811                                 unsigned int nr_irqs)
2812 {
2813         struct irq_data *d = irq_domain_get_irq_data(domain, virq);
2814         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2815         struct its_node *its = its_dev->its;
2816         int i;
2817
2818         bitmap_release_region(its_dev->event_map.lpi_map,
2819                               its_get_event_id(irq_domain_get_irq_data(domain, virq)),
2820                               get_count_order(nr_irqs));
2821
2822         for (i = 0; i < nr_irqs; i++) {
2823                 struct irq_data *data = irq_domain_get_irq_data(domain,
2824                                                                 virq + i);
2825                 /* Nuke the entry in the domain */
2826                 irq_domain_reset_irq_data(data);
2827         }
2828
2829         mutex_lock(&its->dev_alloc_lock);
2830
2831         /*
2832          * If all interrupts have been freed, start mopping the
2833          * floor. This is conditionned on the device not being shared.
2834          */
2835         if (!its_dev->shared &&
2836             bitmap_empty(its_dev->event_map.lpi_map,
2837                          its_dev->event_map.nr_lpis)) {
2838                 its_lpi_free(its_dev->event_map.lpi_map,
2839                              its_dev->event_map.lpi_base,
2840                              its_dev->event_map.nr_lpis);
2841
2842                 /* Unmap device/itt */
2843                 its_send_mapd(its_dev, 0);
2844                 its_free_device(its_dev);
2845         }
2846
2847         mutex_unlock(&its->dev_alloc_lock);
2848
2849         irq_domain_free_irqs_parent(domain, virq, nr_irqs);
2850 }
2851
2852 static const struct irq_domain_ops its_domain_ops = {
2853         .alloc                  = its_irq_domain_alloc,
2854         .free                   = its_irq_domain_free,
2855         .activate               = its_irq_domain_activate,
2856         .deactivate             = its_irq_domain_deactivate,
2857 };
2858
2859 /*
2860  * This is insane.
2861  *
2862  * If a GICv4 doesn't implement Direct LPIs (which is extremely
2863  * likely), the only way to perform an invalidate is to use a fake
2864  * device to issue an INV command, implying that the LPI has first
2865  * been mapped to some event on that device. Since this is not exactly
2866  * cheap, we try to keep that mapping around as long as possible, and
2867  * only issue an UNMAP if we're short on available slots.
2868  *
2869  * Broken by design(tm).
2870  */
2871 static void its_vpe_db_proxy_unmap_locked(struct its_vpe *vpe)
2872 {
2873         /* Already unmapped? */
2874         if (vpe->vpe_proxy_event == -1)
2875                 return;
2876
2877         its_send_discard(vpe_proxy.dev, vpe->vpe_proxy_event);
2878         vpe_proxy.vpes[vpe->vpe_proxy_event] = NULL;
2879
2880         /*
2881          * We don't track empty slots at all, so let's move the
2882          * next_victim pointer if we can quickly reuse that slot
2883          * instead of nuking an existing entry. Not clear that this is
2884          * always a win though, and this might just generate a ripple
2885          * effect... Let's just hope VPEs don't migrate too often.
2886          */
2887         if (vpe_proxy.vpes[vpe_proxy.next_victim])
2888                 vpe_proxy.next_victim = vpe->vpe_proxy_event;
2889
2890         vpe->vpe_proxy_event = -1;
2891 }
2892
2893 static void its_vpe_db_proxy_unmap(struct its_vpe *vpe)
2894 {
2895         if (!gic_rdists->has_direct_lpi) {
2896                 unsigned long flags;
2897
2898                 raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2899                 its_vpe_db_proxy_unmap_locked(vpe);
2900                 raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2901         }
2902 }
2903
2904 static void its_vpe_db_proxy_map_locked(struct its_vpe *vpe)
2905 {
2906         /* Already mapped? */
2907         if (vpe->vpe_proxy_event != -1)
2908                 return;
2909
2910         /* This slot was already allocated. Kick the other VPE out. */
2911         if (vpe_proxy.vpes[vpe_proxy.next_victim])
2912                 its_vpe_db_proxy_unmap_locked(vpe_proxy.vpes[vpe_proxy.next_victim]);
2913
2914         /* Map the new VPE instead */
2915         vpe_proxy.vpes[vpe_proxy.next_victim] = vpe;
2916         vpe->vpe_proxy_event = vpe_proxy.next_victim;
2917         vpe_proxy.next_victim = (vpe_proxy.next_victim + 1) % vpe_proxy.dev->nr_ites;
2918
2919         vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = vpe->col_idx;
2920         its_send_mapti(vpe_proxy.dev, vpe->vpe_db_lpi, vpe->vpe_proxy_event);
2921 }
2922
2923 static void its_vpe_db_proxy_move(struct its_vpe *vpe, int from, int to)
2924 {
2925         unsigned long flags;
2926         struct its_collection *target_col;
2927
2928         if (gic_rdists->has_direct_lpi) {
2929                 void __iomem *rdbase;
2930
2931                 rdbase = per_cpu_ptr(gic_rdists->rdist, from)->rd_base;
2932                 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
2933                 wait_for_syncr(rdbase);
2934
2935                 return;
2936         }
2937
2938         raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2939
2940         its_vpe_db_proxy_map_locked(vpe);
2941
2942         target_col = &vpe_proxy.dev->its->collections[to];
2943         its_send_movi(vpe_proxy.dev, target_col, vpe->vpe_proxy_event);
2944         vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = to;
2945
2946         raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2947 }
2948
2949 static int its_vpe_set_affinity(struct irq_data *d,
2950                                 const struct cpumask *mask_val,
2951                                 bool force)
2952 {
2953         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2954         int cpu = cpumask_first(mask_val);
2955
2956         /*
2957          * Changing affinity is mega expensive, so let's be as lazy as
2958          * we can and only do it if we really have to. Also, if mapped
2959          * into the proxy device, we need to move the doorbell
2960          * interrupt to its new location.
2961          */
2962         if (vpe->col_idx != cpu) {
2963                 int from = vpe->col_idx;
2964
2965                 vpe->col_idx = cpu;
2966                 its_send_vmovp(vpe);
2967                 its_vpe_db_proxy_move(vpe, from, cpu);
2968         }
2969
2970         irq_data_update_effective_affinity(d, cpumask_of(cpu));
2971
2972         return IRQ_SET_MASK_OK_DONE;
2973 }
2974
2975 static void its_vpe_schedule(struct its_vpe *vpe)
2976 {
2977         void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2978         u64 val;
2979
2980         /* Schedule the VPE */
2981         val  = virt_to_phys(page_address(vpe->its_vm->vprop_page)) &
2982                 GENMASK_ULL(51, 12);
2983         val |= (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
2984         val |= GICR_VPROPBASER_RaWb;
2985         val |= GICR_VPROPBASER_InnerShareable;
2986         gits_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2987
2988         val  = virt_to_phys(page_address(vpe->vpt_page)) &
2989                 GENMASK_ULL(51, 16);
2990         val |= GICR_VPENDBASER_RaWaWb;
2991         val |= GICR_VPENDBASER_NonShareable;
2992         /*
2993          * There is no good way of finding out if the pending table is
2994          * empty as we can race against the doorbell interrupt very
2995          * easily. So in the end, vpe->pending_last is only an
2996          * indication that the vcpu has something pending, not one
2997          * that the pending table is empty. A good implementation
2998          * would be able to read its coarse map pretty quickly anyway,
2999          * making this a tolerable issue.
3000          */
3001         val |= GICR_VPENDBASER_PendingLast;
3002         val |= vpe->idai ? GICR_VPENDBASER_IDAI : 0;
3003         val |= GICR_VPENDBASER_Valid;
3004         gits_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
3005 }
3006
3007 static void its_vpe_deschedule(struct its_vpe *vpe)
3008 {
3009         void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
3010         u64 val;
3011
3012         val = its_clear_vpend_valid(vlpi_base);
3013
3014         if (unlikely(val & GICR_VPENDBASER_Dirty)) {
3015                 pr_err_ratelimited("ITS virtual pending table not cleaning\n");
3016                 vpe->idai = false;
3017                 vpe->pending_last = true;
3018         } else {
3019                 vpe->idai = !!(val & GICR_VPENDBASER_IDAI);
3020                 vpe->pending_last = !!(val & GICR_VPENDBASER_PendingLast);
3021         }
3022 }
3023
3024 static void its_vpe_invall(struct its_vpe *vpe)
3025 {
3026         struct its_node *its;
3027
3028         list_for_each_entry(its, &its_nodes, entry) {
3029                 if (!is_v4(its))
3030                         continue;
3031
3032                 if (its_list_map && !vpe->its_vm->vlpi_count[its->list_nr])
3033                         continue;
3034
3035                 /*
3036                  * Sending a VINVALL to a single ITS is enough, as all
3037                  * we need is to reach the redistributors.
3038                  */
3039                 its_send_vinvall(its, vpe);
3040                 return;
3041         }
3042 }
3043
3044 static int its_vpe_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
3045 {
3046         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3047         struct its_cmd_info *info = vcpu_info;
3048
3049         switch (info->cmd_type) {
3050         case SCHEDULE_VPE:
3051                 its_vpe_schedule(vpe);
3052                 return 0;
3053
3054         case DESCHEDULE_VPE:
3055                 its_vpe_deschedule(vpe);
3056                 return 0;
3057
3058         case INVALL_VPE:
3059                 its_vpe_invall(vpe);
3060                 return 0;
3061
3062         default:
3063                 return -EINVAL;
3064         }
3065 }
3066
3067 static void its_vpe_send_cmd(struct its_vpe *vpe,
3068                              void (*cmd)(struct its_device *, u32))
3069 {
3070         unsigned long flags;
3071
3072         raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
3073
3074         its_vpe_db_proxy_map_locked(vpe);
3075         cmd(vpe_proxy.dev, vpe->vpe_proxy_event);
3076
3077         raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
3078 }
3079
3080 static void its_vpe_send_inv(struct irq_data *d)
3081 {
3082         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3083
3084         if (gic_rdists->has_direct_lpi) {
3085                 void __iomem *rdbase;
3086
3087                 /* Target the redistributor this VPE is currently known on */
3088                 rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
3089                 gic_write_lpir(d->parent_data->hwirq, rdbase + GICR_INVLPIR);
3090                 wait_for_syncr(rdbase);
3091         } else {
3092                 its_vpe_send_cmd(vpe, its_send_inv);
3093         }
3094 }
3095
3096 static void its_vpe_mask_irq(struct irq_data *d)
3097 {
3098         /*
3099          * We need to unmask the LPI, which is described by the parent
3100          * irq_data. Instead of calling into the parent (which won't
3101          * exactly do the right thing, let's simply use the
3102          * parent_data pointer. Yes, I'm naughty.
3103          */
3104         lpi_write_config(d->parent_data, LPI_PROP_ENABLED, 0);
3105         its_vpe_send_inv(d);
3106 }
3107
3108 static void its_vpe_unmask_irq(struct irq_data *d)
3109 {
3110         /* Same hack as above... */
3111         lpi_write_config(d->parent_data, 0, LPI_PROP_ENABLED);
3112         its_vpe_send_inv(d);
3113 }
3114
3115 static int its_vpe_set_irqchip_state(struct irq_data *d,
3116                                      enum irqchip_irq_state which,
3117                                      bool state)
3118 {
3119         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3120
3121         if (which != IRQCHIP_STATE_PENDING)
3122                 return -EINVAL;
3123
3124         if (gic_rdists->has_direct_lpi) {
3125                 void __iomem *rdbase;
3126
3127                 rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
3128                 if (state) {
3129                         gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_SETLPIR);
3130                 } else {
3131                         gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
3132                         wait_for_syncr(rdbase);
3133                 }
3134         } else {
3135                 if (state)
3136                         its_vpe_send_cmd(vpe, its_send_int);
3137                 else
3138                         its_vpe_send_cmd(vpe, its_send_clear);
3139         }
3140
3141         return 0;
3142 }
3143
3144 static struct irq_chip its_vpe_irq_chip = {
3145         .name                   = "GICv4-vpe",
3146         .irq_mask               = its_vpe_mask_irq,
3147         .irq_unmask             = its_vpe_unmask_irq,
3148         .irq_eoi                = irq_chip_eoi_parent,
3149         .irq_set_affinity       = its_vpe_set_affinity,
3150         .irq_set_irqchip_state  = its_vpe_set_irqchip_state,
3151         .irq_set_vcpu_affinity  = its_vpe_set_vcpu_affinity,
3152 };
3153
3154 static int its_vpe_id_alloc(void)
3155 {
3156         return ida_simple_get(&its_vpeid_ida, 0, ITS_MAX_VPEID, GFP_KERNEL);
3157 }
3158
3159 static void its_vpe_id_free(u16 id)
3160 {
3161         ida_simple_remove(&its_vpeid_ida, id);
3162 }
3163
3164 static int its_vpe_init(struct its_vpe *vpe)
3165 {
3166         struct page *vpt_page;
3167         int vpe_id;
3168
3169         /* Allocate vpe_id */
3170         vpe_id = its_vpe_id_alloc();
3171         if (vpe_id < 0)
3172                 return vpe_id;
3173
3174         /* Allocate VPT */
3175         vpt_page = its_allocate_pending_table(GFP_KERNEL);
3176         if (!vpt_page) {
3177                 its_vpe_id_free(vpe_id);
3178                 return -ENOMEM;
3179         }
3180
3181         if (!its_alloc_vpe_table(vpe_id)) {
3182                 its_vpe_id_free(vpe_id);
3183                 its_free_pending_table(vpt_page);
3184                 return -ENOMEM;
3185         }
3186
3187         vpe->vpe_id = vpe_id;
3188         vpe->vpt_page = vpt_page;
3189         vpe->vpe_proxy_event = -1;
3190
3191         return 0;
3192 }
3193
3194 static void its_vpe_teardown(struct its_vpe *vpe)
3195 {
3196         its_vpe_db_proxy_unmap(vpe);
3197         its_vpe_id_free(vpe->vpe_id);
3198         its_free_pending_table(vpe->vpt_page);
3199 }
3200
3201 static void its_vpe_irq_domain_free(struct irq_domain *domain,
3202                                     unsigned int virq,
3203                                     unsigned int nr_irqs)
3204 {
3205         struct its_vm *vm = domain->host_data;
3206         int i;
3207
3208         irq_domain_free_irqs_parent(domain, virq, nr_irqs);
3209
3210         for (i = 0; i < nr_irqs; i++) {
3211                 struct irq_data *data = irq_domain_get_irq_data(domain,
3212                                                                 virq + i);
3213                 struct its_vpe *vpe = irq_data_get_irq_chip_data(data);
3214
3215                 BUG_ON(vm != vpe->its_vm);
3216
3217                 clear_bit(data->hwirq, vm->db_bitmap);
3218                 its_vpe_teardown(vpe);
3219                 irq_domain_reset_irq_data(data);
3220         }
3221
3222         if (bitmap_empty(vm->db_bitmap, vm->nr_db_lpis)) {
3223                 its_lpi_free(vm->db_bitmap, vm->db_lpi_base, vm->nr_db_lpis);
3224                 its_free_prop_table(vm->vprop_page);
3225         }
3226 }
3227
3228 static int its_vpe_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
3229                                     unsigned int nr_irqs, void *args)
3230 {
3231         struct its_vm *vm = args;
3232         unsigned long *bitmap;
3233         struct page *vprop_page;
3234         int base, nr_ids, i, err = 0;
3235
3236         BUG_ON(!vm);
3237
3238         bitmap = its_lpi_alloc(roundup_pow_of_two(nr_irqs), &base, &nr_ids);
3239         if (!bitmap)
3240                 return -ENOMEM;
3241
3242         if (nr_ids < nr_irqs) {
3243                 its_lpi_free(bitmap, base, nr_ids);
3244                 return -ENOMEM;
3245         }
3246
3247         vprop_page = its_allocate_prop_table(GFP_KERNEL);
3248         if (!vprop_page) {
3249                 its_lpi_free(bitmap, base, nr_ids);
3250                 return -ENOMEM;
3251         }
3252
3253         vm->db_bitmap = bitmap;
3254         vm->db_lpi_base = base;
3255         vm->nr_db_lpis = nr_ids;
3256         vm->vprop_page = vprop_page;
3257
3258         for (i = 0; i < nr_irqs; i++) {
3259                 vm->vpes[i]->vpe_db_lpi = base + i;
3260                 err = its_vpe_init(vm->vpes[i]);
3261                 if (err)
3262                         break;
3263                 err = its_irq_gic_domain_alloc(domain, virq + i,
3264                                                vm->vpes[i]->vpe_db_lpi);
3265                 if (err)
3266                         break;
3267                 irq_domain_set_hwirq_and_chip(domain, virq + i, i,
3268                                               &its_vpe_irq_chip, vm->vpes[i]);
3269                 set_bit(i, bitmap);
3270         }
3271
3272         if (err) {
3273                 if (i > 0)
3274                         its_vpe_irq_domain_free(domain, virq, i - 1);
3275
3276                 its_lpi_free(bitmap, base, nr_ids);
3277                 its_free_prop_table(vprop_page);
3278         }
3279
3280         return err;
3281 }
3282
3283 static int its_vpe_irq_domain_activate(struct irq_domain *domain,
3284                                        struct irq_data *d, bool reserve)
3285 {
3286         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3287         struct its_node *its;
3288
3289         /* If we use the list map, we issue VMAPP on demand... */
3290         if (its_list_map)
3291                 return 0;
3292
3293         /* Map the VPE to the first possible CPU */
3294         vpe->col_idx = cpumask_first(cpu_online_mask);
3295
3296         list_for_each_entry(its, &its_nodes, entry) {
3297                 if (!is_v4(its))
3298                         continue;
3299
3300                 its_send_vmapp(its, vpe, true);
3301                 its_send_vinvall(its, vpe);
3302         }
3303
3304         irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
3305
3306         return 0;
3307 }
3308
3309 static void its_vpe_irq_domain_deactivate(struct irq_domain *domain,
3310                                           struct irq_data *d)
3311 {
3312         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3313         struct its_node *its;
3314
3315         /*
3316          * If we use the list map, we unmap the VPE once no VLPIs are
3317          * associated with the VM.
3318          */
3319         if (its_list_map)
3320                 return;
3321
3322         list_for_each_entry(its, &its_nodes, entry) {
3323                 if (!is_v4(its))
3324                         continue;
3325
3326                 its_send_vmapp(its, vpe, false);
3327         }
3328 }
3329
3330 static const struct irq_domain_ops its_vpe_domain_ops = {
3331         .alloc                  = its_vpe_irq_domain_alloc,
3332         .free                   = its_vpe_irq_domain_free,
3333         .activate               = its_vpe_irq_domain_activate,
3334         .deactivate             = its_vpe_irq_domain_deactivate,
3335 };
3336
3337 static int its_force_quiescent(void __iomem *base)
3338 {
3339         u32 count = 1000000;    /* 1s */
3340         u32 val;
3341
3342         val = readl_relaxed(base + GITS_CTLR);
3343         /*
3344          * GIC architecture specification requires the ITS to be both
3345          * disabled and quiescent for writes to GITS_BASER<n> or
3346          * GITS_CBASER to not have UNPREDICTABLE results.
3347          */
3348         if ((val & GITS_CTLR_QUIESCENT) && !(val & GITS_CTLR_ENABLE))
3349                 return 0;
3350
3351         /* Disable the generation of all interrupts to this ITS */
3352         val &= ~(GITS_CTLR_ENABLE | GITS_CTLR_ImDe);
3353         writel_relaxed(val, base + GITS_CTLR);
3354
3355         /* Poll GITS_CTLR and wait until ITS becomes quiescent */
3356         while (1) {
3357                 val = readl_relaxed(base + GITS_CTLR);
3358                 if (val & GITS_CTLR_QUIESCENT)
3359                         return 0;
3360
3361                 count--;
3362                 if (!count)
3363                         return -EBUSY;
3364
3365                 cpu_relax();
3366                 udelay(1);
3367         }
3368 }
3369
3370 static bool __maybe_unused its_enable_quirk_cavium_22375(void *data)
3371 {
3372         struct its_node *its = data;
3373
3374         /* erratum 22375: only alloc 8MB table size (20 bits) */
3375         its->typer &= ~GITS_TYPER_DEVBITS;
3376         its->typer |= FIELD_PREP(GITS_TYPER_DEVBITS, 20 - 1);
3377         its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_22375;
3378
3379         return true;
3380 }
3381
3382 static bool __maybe_unused its_enable_quirk_cavium_23144(void *data)
3383 {
3384         struct its_node *its = data;
3385
3386         its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_23144;
3387
3388         return true;
3389 }
3390
3391 static bool __maybe_unused its_enable_quirk_qdf2400_e0065(void *data)
3392 {
3393         struct its_node *its = data;
3394
3395         /* On QDF2400, the size of the ITE is 16Bytes */
3396         its->typer &= ~GITS_TYPER_ITT_ENTRY_SIZE;
3397         its->typer |= FIELD_PREP(GITS_TYPER_ITT_ENTRY_SIZE, 16 - 1);
3398
3399         return true;
3400 }
3401
3402 static u64 its_irq_get_msi_base_pre_its(struct its_device *its_dev)
3403 {
3404         struct its_node *its = its_dev->its;
3405
3406         /*
3407          * The Socionext Synquacer SoC has a so-called 'pre-ITS',
3408          * which maps 32-bit writes targeted at a separate window of
3409          * size '4 << device_id_bits' onto writes to GITS_TRANSLATER
3410          * with device ID taken from bits [device_id_bits + 1:2] of
3411          * the window offset.
3412          */
3413         return its->pre_its_base + (its_dev->device_id << 2);
3414 }
3415
3416 static bool __maybe_unused its_enable_quirk_socionext_synquacer(void *data)
3417 {
3418         struct its_node *its = data;
3419         u32 pre_its_window[2];
3420         u32 ids;
3421
3422         if (!fwnode_property_read_u32_array(its->fwnode_handle,
3423                                            "socionext,synquacer-pre-its",
3424                                            pre_its_window,
3425                                            ARRAY_SIZE(pre_its_window))) {
3426
3427                 its->pre_its_base = pre_its_window[0];
3428                 its->get_msi_base = its_irq_get_msi_base_pre_its;
3429
3430                 ids = ilog2(pre_its_window[1]) - 2;
3431                 if (device_ids(its) > ids) {
3432                         its->typer &= ~GITS_TYPER_DEVBITS;
3433                         its->typer |= FIELD_PREP(GITS_TYPER_DEVBITS, ids - 1);
3434                 }
3435
3436                 /* the pre-ITS breaks isolation, so disable MSI remapping */
3437                 its->msi_domain_flags &= ~IRQ_DOMAIN_FLAG_MSI_REMAP;
3438                 return true;
3439         }
3440         return false;
3441 }
3442
3443 static bool __maybe_unused its_enable_quirk_hip07_161600802(void *data)
3444 {
3445         struct its_node *its = data;
3446
3447         /*
3448          * Hip07 insists on using the wrong address for the VLPI
3449          * page. Trick it into doing the right thing...
3450          */
3451         its->vlpi_redist_offset = SZ_128K;
3452         return true;
3453 }
3454
3455 static const struct gic_quirk its_quirks[] = {
3456 #ifdef CONFIG_CAVIUM_ERRATUM_22375
3457         {
3458                 .desc   = "ITS: Cavium errata 22375, 24313",
3459                 .iidr   = 0xa100034c,   /* ThunderX pass 1.x */
3460                 .mask   = 0xffff0fff,
3461                 .init   = its_enable_quirk_cavium_22375,
3462         },
3463 #endif
3464 #ifdef CONFIG_CAVIUM_ERRATUM_23144
3465         {
3466                 .desc   = "ITS: Cavium erratum 23144",
3467                 .iidr   = 0xa100034c,   /* ThunderX pass 1.x */
3468                 .mask   = 0xffff0fff,
3469                 .init   = its_enable_quirk_cavium_23144,
3470         },
3471 #endif
3472 #ifdef CONFIG_QCOM_QDF2400_ERRATUM_0065
3473         {
3474                 .desc   = "ITS: QDF2400 erratum 0065",
3475                 .iidr   = 0x00001070, /* QDF2400 ITS rev 1.x */
3476                 .mask   = 0xffffffff,
3477                 .init   = its_enable_quirk_qdf2400_e0065,
3478         },
3479 #endif
3480 #ifdef CONFIG_SOCIONEXT_SYNQUACER_PREITS
3481         {
3482                 /*
3483                  * The Socionext Synquacer SoC incorporates ARM's own GIC-500
3484                  * implementation, but with a 'pre-ITS' added that requires
3485                  * special handling in software.
3486                  */
3487                 .desc   = "ITS: Socionext Synquacer pre-ITS",
3488                 .iidr   = 0x0001143b,
3489                 .mask   = 0xffffffff,
3490                 .init   = its_enable_quirk_socionext_synquacer,
3491         },
3492 #endif
3493 #ifdef CONFIG_HISILICON_ERRATUM_161600802
3494         {
3495                 .desc   = "ITS: Hip07 erratum 161600802",
3496                 .iidr   = 0x00000004,
3497                 .mask   = 0xffffffff,
3498                 .init   = its_enable_quirk_hip07_161600802,
3499         },
3500 #endif
3501         {
3502         }
3503 };
3504
3505 static void its_enable_quirks(struct its_node *its)
3506 {
3507         u32 iidr = readl_relaxed(its->base + GITS_IIDR);
3508
3509         gic_enable_quirks(iidr, its_quirks, its);
3510 }
3511
3512 static int its_save_disable(void)
3513 {
3514         struct its_node *its;
3515         int err = 0;
3516
3517         raw_spin_lock(&its_lock);
3518         list_for_each_entry(its, &its_nodes, entry) {
3519                 void __iomem *base;
3520
3521                 if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3522                         continue;
3523
3524                 base = its->base;
3525                 its->ctlr_save = readl_relaxed(base + GITS_CTLR);
3526                 err = its_force_quiescent(base);
3527                 if (err) {
3528                         pr_err("ITS@%pa: failed to quiesce: %d\n",
3529                                &its->phys_base, err);
3530                         writel_relaxed(its->ctlr_save, base + GITS_CTLR);
3531                         goto err;
3532                 }
3533
3534                 its->cbaser_save = gits_read_cbaser(base + GITS_CBASER);
3535         }
3536
3537 err:
3538         if (err) {
3539                 list_for_each_entry_continue_reverse(its, &its_nodes, entry) {
3540                         void __iomem *base;
3541
3542                         if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3543                                 continue;
3544
3545                         base = its->base;
3546                         writel_relaxed(its->ctlr_save, base + GITS_CTLR);
3547                 }
3548         }
3549         raw_spin_unlock(&its_lock);
3550
3551         return err;
3552 }
3553
3554 static void its_restore_enable(void)
3555 {
3556         struct its_node *its;
3557         int ret;
3558
3559         raw_spin_lock(&its_lock);
3560         list_for_each_entry(its, &its_nodes, entry) {
3561                 void __iomem *base;
3562                 int i;
3563
3564                 if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3565                         continue;
3566
3567                 base = its->base;
3568
3569                 /*
3570                  * Make sure that the ITS is disabled. If it fails to quiesce,
3571                  * don't restore it since writing to CBASER or BASER<n>
3572                  * registers is undefined according to the GIC v3 ITS
3573                  * Specification.
3574                  */
3575                 ret = its_force_quiescent(base);
3576                 if (ret) {
3577                         pr_err("ITS@%pa: failed to quiesce on resume: %d\n",
3578                                &its->phys_base, ret);
3579                         continue;
3580                 }
3581
3582                 gits_write_cbaser(its->cbaser_save, base + GITS_CBASER);
3583
3584                 /*
3585                  * Writing CBASER resets CREADR to 0, so make CWRITER and
3586                  * cmd_write line up with it.
3587                  */
3588                 its->cmd_write = its->cmd_base;
3589                 gits_write_cwriter(0, base + GITS_CWRITER);
3590
3591                 /* Restore GITS_BASER from the value cache. */
3592                 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
3593                         struct its_baser *baser = &its->tables[i];
3594
3595                         if (!(baser->val & GITS_BASER_VALID))
3596                                 continue;
3597
3598                         its_write_baser(its, baser, baser->val);
3599                 }
3600                 writel_relaxed(its->ctlr_save, base + GITS_CTLR);
3601
3602                 /*
3603                  * Reinit the collection if it's stored in the ITS. This is
3604                  * indicated by the col_id being less than the HCC field.
3605                  * CID < HCC as specified in the GIC v3 Documentation.
3606                  */
3607                 if (its->collections[smp_processor_id()].col_id <
3608                     GITS_TYPER_HCC(gic_read_typer(base + GITS_TYPER)))
3609                         its_cpu_init_collection(its);
3610         }
3611         raw_spin_unlock(&its_lock);
3612 }
3613
3614 static struct syscore_ops its_syscore_ops = {
3615         .suspend = its_save_disable,
3616         .resume = its_restore_enable,
3617 };
3618
3619 static int its_init_domain(struct fwnode_handle *handle, struct its_node *its)
3620 {
3621         struct irq_domain *inner_domain;
3622         struct msi_domain_info *info;
3623
3624         info = kzalloc(sizeof(*info), GFP_KERNEL);
3625         if (!info)
3626                 return -ENOMEM;
3627
3628         inner_domain = irq_domain_create_tree(handle, &its_domain_ops, its);
3629         if (!inner_domain) {
3630                 kfree(info);
3631                 return -ENOMEM;
3632         }
3633
3634         inner_domain->parent = its_parent;
3635         irq_domain_update_bus_token(inner_domain, DOMAIN_BUS_NEXUS);
3636         inner_domain->flags |= its->msi_domain_flags;
3637         info->ops = &its_msi_domain_ops;
3638         info->data = its;
3639         inner_domain->host_data = info;
3640
3641         return 0;
3642 }
3643
3644 static int its_init_vpe_domain(void)
3645 {
3646         struct its_node *its;
3647         u32 devid;
3648         int entries;
3649
3650         if (gic_rdists->has_direct_lpi) {
3651                 pr_info("ITS: Using DirectLPI for VPE invalidation\n");
3652                 return 0;
3653         }
3654
3655         /* Any ITS will do, even if not v4 */
3656         its = list_first_entry(&its_nodes, struct its_node, entry);
3657
3658         entries = roundup_pow_of_two(nr_cpu_ids);
3659         vpe_proxy.vpes = kcalloc(entries, sizeof(*vpe_proxy.vpes),
3660                                  GFP_KERNEL);
3661         if (!vpe_proxy.vpes) {
3662                 pr_err("ITS: Can't allocate GICv4 proxy device array\n");
3663                 return -ENOMEM;
3664         }
3665
3666         /* Use the last possible DevID */
3667         devid = GENMASK(device_ids(its) - 1, 0);
3668         vpe_proxy.dev = its_create_device(its, devid, entries, false);
3669         if (!vpe_proxy.dev) {
3670                 kfree(vpe_proxy.vpes);
3671                 pr_err("ITS: Can't allocate GICv4 proxy device\n");
3672                 return -ENOMEM;
3673         }
3674
3675         BUG_ON(entries > vpe_proxy.dev->nr_ites);
3676
3677         raw_spin_lock_init(&vpe_proxy.lock);
3678         vpe_proxy.next_victim = 0;
3679         pr_info("ITS: Allocated DevID %x as GICv4 proxy device (%d slots)\n",
3680                 devid, vpe_proxy.dev->nr_ites);
3681
3682         return 0;
3683 }
3684
3685 static int __init its_compute_its_list_map(struct resource *res,
3686                                            void __iomem *its_base)
3687 {
3688         int its_number;
3689         u32 ctlr;
3690
3691         /*
3692          * This is assumed to be done early enough that we're
3693          * guaranteed to be single-threaded, hence no
3694          * locking. Should this change, we should address
3695          * this.
3696          */
3697         its_number = find_first_zero_bit(&its_list_map, GICv4_ITS_LIST_MAX);
3698         if (its_number >= GICv4_ITS_LIST_MAX) {
3699                 pr_err("ITS@%pa: No ITSList entry available!\n",
3700                        &res->start);
3701                 return -EINVAL;
3702         }
3703
3704         ctlr = readl_relaxed(its_base + GITS_CTLR);
3705         ctlr &= ~GITS_CTLR_ITS_NUMBER;
3706         ctlr |= its_number << GITS_CTLR_ITS_NUMBER_SHIFT;
3707         writel_relaxed(ctlr, its_base + GITS_CTLR);
3708         ctlr = readl_relaxed(its_base + GITS_CTLR);
3709         if ((ctlr & GITS_CTLR_ITS_NUMBER) != (its_number << GITS_CTLR_ITS_NUMBER_SHIFT)) {
3710                 its_number = ctlr & GITS_CTLR_ITS_NUMBER;
3711                 its_number >>= GITS_CTLR_ITS_NUMBER_SHIFT;
3712         }
3713
3714         if (test_and_set_bit(its_number, &its_list_map)) {
3715                 pr_err("ITS@%pa: Duplicate ITSList entry %d\n",
3716                        &res->start, its_number);
3717                 return -EINVAL;
3718         }
3719
3720         return its_number;
3721 }
3722
3723 static int __init its_probe_one(struct resource *res,
3724                                 struct fwnode_handle *handle, int numa_node)
3725 {
3726         struct its_node *its;
3727         void __iomem *its_base;
3728         u32 val, ctlr;
3729         u64 baser, tmp, typer;
3730         struct page *page;
3731         int err;
3732
3733         its_base = ioremap(res->start, resource_size(res));
3734         if (!its_base) {
3735                 pr_warn("ITS@%pa: Unable to map ITS registers\n", &res->start);
3736                 return -ENOMEM;
3737         }
3738
3739         val = readl_relaxed(its_base + GITS_PIDR2) & GIC_PIDR2_ARCH_MASK;
3740         if (val != 0x30 && val != 0x40) {
3741                 pr_warn("ITS@%pa: No ITS detected, giving up\n", &res->start);
3742                 err = -ENODEV;
3743                 goto out_unmap;
3744         }
3745
3746         err = its_force_quiescent(its_base);
3747         if (err) {
3748                 pr_warn("ITS@%pa: Failed to quiesce, giving up\n", &res->start);
3749                 goto out_unmap;
3750         }
3751
3752         pr_info("ITS %pR\n", res);
3753
3754         its = kzalloc(sizeof(*its), GFP_KERNEL);
3755         if (!its) {
3756                 err = -ENOMEM;
3757                 goto out_unmap;
3758         }
3759
3760         raw_spin_lock_init(&its->lock);
3761         mutex_init(&its->dev_alloc_lock);
3762         INIT_LIST_HEAD(&its->entry);
3763         INIT_LIST_HEAD(&its->its_device_list);
3764         typer = gic_read_typer(its_base + GITS_TYPER);
3765         its->typer = typer;
3766         its->base = its_base;
3767         its->phys_base = res->start;
3768         if (is_v4(its)) {
3769                 if (!(typer & GITS_TYPER_VMOVP)) {
3770                         err = its_compute_its_list_map(res, its_base);
3771                         if (err < 0)
3772                                 goto out_free_its;
3773
3774                         its->list_nr = err;
3775
3776                         pr_info("ITS@%pa: Using ITS number %d\n",
3777                                 &res->start, err);
3778                 } else {
3779                         pr_info("ITS@%pa: Single VMOVP capable\n", &res->start);
3780                 }
3781         }
3782
3783         its->numa_node = numa_node;
3784
3785         page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO,
3786                                 get_order(ITS_CMD_QUEUE_SZ));
3787         if (!page) {
3788                 err = -ENOMEM;
3789                 goto out_free_its;
3790         }
3791         its->cmd_base = (void *)page_address(page);
3792         its->cmd_write = its->cmd_base;
3793         its->fwnode_handle = handle;
3794         its->get_msi_base = its_irq_get_msi_base;
3795         its->msi_domain_flags = IRQ_DOMAIN_FLAG_MSI_REMAP;
3796
3797         its_enable_quirks(its);
3798
3799         err = its_alloc_tables(its);
3800         if (err)
3801                 goto out_free_cmd;
3802
3803         err = its_alloc_collections(its);
3804         if (err)
3805                 goto out_free_tables;
3806
3807         baser = (virt_to_phys(its->cmd_base)    |
3808                  GITS_CBASER_RaWaWb             |
3809                  GITS_CBASER_InnerShareable     |
3810                  (ITS_CMD_QUEUE_SZ / SZ_4K - 1) |
3811                  GITS_CBASER_VALID);
3812
3813         gits_write_cbaser(baser, its->base + GITS_CBASER);
3814         tmp = gits_read_cbaser(its->base + GITS_CBASER);
3815
3816         if ((tmp ^ baser) & GITS_CBASER_SHAREABILITY_MASK) {
3817                 if (!(tmp & GITS_CBASER_SHAREABILITY_MASK)) {
3818                         /*
3819                          * The HW reports non-shareable, we must
3820                          * remove the cacheability attributes as
3821                          * well.
3822                          */
3823                         baser &= ~(GITS_CBASER_SHAREABILITY_MASK |
3824                                    GITS_CBASER_CACHEABILITY_MASK);
3825                         baser |= GITS_CBASER_nC;
3826                         gits_write_cbaser(baser, its->base + GITS_CBASER);
3827                 }
3828                 pr_info("ITS: using cache flushing for cmd queue\n");
3829                 its->flags |= ITS_FLAGS_CMDQ_NEEDS_FLUSHING;
3830         }
3831
3832         gits_write_cwriter(0, its->base + GITS_CWRITER);
3833         ctlr = readl_relaxed(its->base + GITS_CTLR);
3834         ctlr |= GITS_CTLR_ENABLE;
3835         if (is_v4(its))
3836                 ctlr |= GITS_CTLR_ImDe;
3837         writel_relaxed(ctlr, its->base + GITS_CTLR);
3838
3839         if (GITS_TYPER_HCC(typer))
3840                 its->flags |= ITS_FLAGS_SAVE_SUSPEND_STATE;
3841
3842         err = its_init_domain(handle, its);
3843         if (err)
3844                 goto out_free_tables;
3845
3846         raw_spin_lock(&its_lock);
3847         list_add(&its->entry, &its_nodes);
3848         raw_spin_unlock(&its_lock);
3849
3850         return 0;
3851
3852 out_free_tables:
3853         its_free_tables(its);
3854 out_free_cmd:
3855         free_pages((unsigned long)its->cmd_base, get_order(ITS_CMD_QUEUE_SZ));
3856 out_free_its:
3857         kfree(its);
3858 out_unmap:
3859         iounmap(its_base);
3860         pr_err("ITS@%pa: failed probing (%d)\n", &res->start, err);
3861         return err;
3862 }
3863
3864 static bool gic_rdists_supports_plpis(void)
3865 {
3866         return !!(gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER) & GICR_TYPER_PLPIS);
3867 }
3868
3869 static int redist_disable_lpis(void)
3870 {
3871         void __iomem *rbase = gic_data_rdist_rd_base();
3872         u64 timeout = USEC_PER_SEC;
3873         u64 val;
3874
3875         if (!gic_rdists_supports_plpis()) {
3876                 pr_info("CPU%d: LPIs not supported\n", smp_processor_id());
3877                 return -ENXIO;
3878         }
3879
3880         val = readl_relaxed(rbase + GICR_CTLR);
3881         if (!(val & GICR_CTLR_ENABLE_LPIS))
3882                 return 0;
3883
3884         /*
3885          * If coming via a CPU hotplug event, we don't need to disable
3886          * LPIs before trying to re-enable them. They are already
3887          * configured and all is well in the world.
3888          *
3889          * If running with preallocated tables, there is nothing to do.
3890          */
3891         if (gic_data_rdist()->lpi_enabled ||
3892             (gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED))
3893                 return 0;
3894
3895         /*
3896          * From that point on, we only try to do some damage control.
3897          */
3898         pr_warn("GICv3: CPU%d: Booted with LPIs enabled, memory probably corrupted\n",
3899                 smp_processor_id());
3900         add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
3901
3902         /* Disable LPIs */
3903         val &= ~GICR_CTLR_ENABLE_LPIS;
3904         writel_relaxed(val, rbase + GICR_CTLR);
3905
3906         /* Make sure any change to GICR_CTLR is observable by the GIC */
3907         dsb(sy);
3908
3909         /*
3910          * Software must observe RWP==0 after clearing GICR_CTLR.EnableLPIs
3911          * from 1 to 0 before programming GICR_PEND{PROP}BASER registers.
3912          * Error out if we time out waiting for RWP to clear.
3913          */
3914         while (readl_relaxed(rbase + GICR_CTLR) & GICR_CTLR_RWP) {
3915                 if (!timeout) {
3916                         pr_err("CPU%d: Timeout while disabling LPIs\n",
3917                                smp_processor_id());
3918                         return -ETIMEDOUT;
3919                 }
3920                 udelay(1);
3921                 timeout--;
3922         }
3923
3924         /*
3925          * After it has been written to 1, it is IMPLEMENTATION
3926          * DEFINED whether GICR_CTLR.EnableLPI becomes RES1 or can be
3927          * cleared to 0. Error out if clearing the bit failed.
3928          */
3929         if (readl_relaxed(rbase + GICR_CTLR) & GICR_CTLR_ENABLE_LPIS) {
3930                 pr_err("CPU%d: Failed to disable LPIs\n", smp_processor_id());
3931                 return -EBUSY;
3932         }
3933
3934         return 0;
3935 }
3936
3937 int its_cpu_init(void)
3938 {
3939         if (!list_empty(&its_nodes)) {
3940                 int ret;
3941
3942                 ret = redist_disable_lpis();
3943                 if (ret)
3944                         return ret;
3945
3946                 its_cpu_init_lpis();
3947                 its_cpu_init_collections();
3948         }
3949
3950         return 0;
3951 }
3952
3953 static const struct of_device_id its_device_id[] = {
3954         {       .compatible     = "arm,gic-v3-its",     },
3955         {},
3956 };
3957
3958 static int __init its_of_probe(struct device_node *node)
3959 {
3960         struct device_node *np;
3961         struct resource res;
3962
3963         for (np = of_find_matching_node(node, its_device_id); np;
3964              np = of_find_matching_node(np, its_device_id)) {
3965                 if (!of_device_is_available(np))
3966                         continue;
3967                 if (!of_property_read_bool(np, "msi-controller")) {
3968                         pr_warn("%pOF: no msi-controller property, ITS ignored\n",
3969                                 np);
3970                         continue;
3971                 }
3972
3973                 if (of_address_to_resource(np, 0, &res)) {
3974                         pr_warn("%pOF: no regs?\n", np);
3975                         continue;
3976                 }
3977
3978                 its_probe_one(&res, &np->fwnode, of_node_to_nid(np));
3979         }
3980         return 0;
3981 }
3982
3983 #ifdef CONFIG_ACPI
3984
3985 #define ACPI_GICV3_ITS_MEM_SIZE (SZ_128K)
3986
3987 #ifdef CONFIG_ACPI_NUMA
3988 struct its_srat_map {
3989         /* numa node id */
3990         u32     numa_node;
3991         /* GIC ITS ID */
3992         u32     its_id;
3993 };
3994
3995 static struct its_srat_map *its_srat_maps __initdata;
3996 static int its_in_srat __initdata;
3997
3998 static int __init acpi_get_its_numa_node(u32 its_id)
3999 {
4000         int i;
4001
4002         for (i = 0; i < its_in_srat; i++) {
4003                 if (its_id == its_srat_maps[i].its_id)
4004                         return its_srat_maps[i].numa_node;
4005         }
4006         return NUMA_NO_NODE;
4007 }
4008
4009 static int __init gic_acpi_match_srat_its(union acpi_subtable_headers *header,
4010                                           const unsigned long end)
4011 {
4012         return 0;
4013 }
4014
4015 static int __init gic_acpi_parse_srat_its(union acpi_subtable_headers *header,
4016                          const unsigned long end)
4017 {
4018         int node;
4019         struct acpi_srat_gic_its_affinity *its_affinity;
4020
4021         its_affinity = (struct acpi_srat_gic_its_affinity *)header;
4022         if (!its_affinity)
4023                 return -EINVAL;
4024
4025         if (its_affinity->header.length < sizeof(*its_affinity)) {
4026                 pr_err("SRAT: Invalid header length %d in ITS affinity\n",
4027                         its_affinity->header.length);
4028                 return -EINVAL;
4029         }
4030
4031         node = acpi_map_pxm_to_node(its_affinity->proximity_domain);
4032
4033         if (node == NUMA_NO_NODE || node >= MAX_NUMNODES) {
4034                 pr_err("SRAT: Invalid NUMA node %d in ITS affinity\n", node);
4035                 return 0;
4036         }
4037
4038         its_srat_maps[its_in_srat].numa_node = node;
4039         its_srat_maps[its_in_srat].its_id = its_affinity->its_id;
4040         its_in_srat++;
4041         pr_info("SRAT: PXM %d -> ITS %d -> Node %d\n",
4042                 its_affinity->proximity_domain, its_affinity->its_id, node);
4043
4044         return 0;
4045 }
4046
4047 static void __init acpi_table_parse_srat_its(void)
4048 {
4049         int count;
4050
4051         count = acpi_table_parse_entries(ACPI_SIG_SRAT,
4052                         sizeof(struct acpi_table_srat),
4053                         ACPI_SRAT_TYPE_GIC_ITS_AFFINITY,
4054                         gic_acpi_match_srat_its, 0);
4055         if (count <= 0)
4056                 return;
4057
4058         its_srat_maps = kmalloc_array(count, sizeof(struct its_srat_map),
4059                                       GFP_KERNEL);
4060         if (!its_srat_maps) {
4061                 pr_warn("SRAT: Failed to allocate memory for its_srat_maps!\n");
4062                 return;
4063         }
4064
4065         acpi_table_parse_entries(ACPI_SIG_SRAT,
4066                         sizeof(struct acpi_table_srat),
4067                         ACPI_SRAT_TYPE_GIC_ITS_AFFINITY,
4068                         gic_acpi_parse_srat_its, 0);
4069 }
4070
4071 /* free the its_srat_maps after ITS probing */
4072 static void __init acpi_its_srat_maps_free(void)
4073 {
4074         kfree(its_srat_maps);
4075 }
4076 #else
4077 static void __init acpi_table_parse_srat_its(void)      { }
4078 static int __init acpi_get_its_numa_node(u32 its_id) { return NUMA_NO_NODE; }
4079 static void __init acpi_its_srat_maps_free(void) { }
4080 #endif
4081
4082 static int __init gic_acpi_parse_madt_its(union acpi_subtable_headers *header,
4083                                           const unsigned long end)
4084 {
4085         struct acpi_madt_generic_translator *its_entry;
4086         struct fwnode_handle *dom_handle;
4087         struct resource res;
4088         int err;
4089
4090         its_entry = (struct acpi_madt_generic_translator *)header;
4091         memset(&res, 0, sizeof(res));
4092         res.start = its_entry->base_address;
4093         res.end = its_entry->base_address + ACPI_GICV3_ITS_MEM_SIZE - 1;
4094         res.flags = IORESOURCE_MEM;
4095
4096         dom_handle = irq_domain_alloc_fwnode(&res.start);
4097         if (!dom_handle) {
4098                 pr_err("ITS@%pa: Unable to allocate GICv3 ITS domain token\n",
4099                        &res.start);
4100                 return -ENOMEM;
4101         }
4102
4103         err = iort_register_domain_token(its_entry->translation_id, res.start,
4104                                          dom_handle);
4105         if (err) {
4106                 pr_err("ITS@%pa: Unable to register GICv3 ITS domain token (ITS ID %d) to IORT\n",
4107                        &res.start, its_entry->translation_id);
4108                 goto dom_err;
4109         }
4110
4111         err = its_probe_one(&res, dom_handle,
4112                         acpi_get_its_numa_node(its_entry->translation_id));
4113         if (!err)
4114                 return 0;
4115
4116         iort_deregister_domain_token(its_entry->translation_id);
4117 dom_err:
4118         irq_domain_free_fwnode(dom_handle);
4119         return err;
4120 }
4121
4122 static void __init its_acpi_probe(void)
4123 {
4124         acpi_table_parse_srat_its();
4125         acpi_table_parse_madt(ACPI_MADT_TYPE_GENERIC_TRANSLATOR,
4126                               gic_acpi_parse_madt_its, 0);
4127         acpi_its_srat_maps_free();
4128 }
4129 #else
4130 static void __init its_acpi_probe(void) { }
4131 #endif
4132
4133 int __init its_init(struct fwnode_handle *handle, struct rdists *rdists,
4134                     struct irq_domain *parent_domain)
4135 {
4136         struct device_node *of_node;
4137         struct its_node *its;
4138         bool has_v4 = false;
4139         int err;
4140
4141         its_parent = parent_domain;
4142         of_node = to_of_node(handle);
4143         if (of_node)
4144                 its_of_probe(of_node);
4145         else
4146                 its_acpi_probe();
4147
4148         if (list_empty(&its_nodes)) {
4149                 pr_warn("ITS: No ITS available, not enabling LPIs\n");
4150                 return -ENXIO;
4151         }
4152
4153         gic_rdists = rdists;
4154
4155         err = allocate_lpi_tables();
4156         if (err)
4157                 return err;
4158
4159         list_for_each_entry(its, &its_nodes, entry)
4160                 has_v4 |= is_v4(its);
4161
4162         if (has_v4 & rdists->has_vlpis) {
4163                 if (its_init_vpe_domain() ||
4164                     its_init_v4(parent_domain, &its_vpe_domain_ops)) {
4165                         rdists->has_vlpis = false;
4166                         pr_err("ITS: Disabling GICv4 support\n");
4167                 }
4168         }
4169
4170         register_syscore_ops(&its_syscore_ops);
4171
4172         return 0;
4173 }