Merge branches 'clk-ingenic', 'clk-mtk-mux', 'clk-qcom-sdm845-pcie', 'clk-mtk-crit...
[linux-2.6-microblaze.git] / drivers / irqchip / irq-gic-v3-its.c
1 /*
2  * Copyright (C) 2013-2017 ARM Limited, All Rights Reserved.
3  * Author: Marc Zyngier <marc.zyngier@arm.com>
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License version 2 as
7  * published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
16  */
17
18 #include <linux/acpi.h>
19 #include <linux/acpi_iort.h>
20 #include <linux/bitmap.h>
21 #include <linux/cpu.h>
22 #include <linux/crash_dump.h>
23 #include <linux/delay.h>
24 #include <linux/dma-iommu.h>
25 #include <linux/efi.h>
26 #include <linux/interrupt.h>
27 #include <linux/irqdomain.h>
28 #include <linux/list.h>
29 #include <linux/list_sort.h>
30 #include <linux/log2.h>
31 #include <linux/memblock.h>
32 #include <linux/mm.h>
33 #include <linux/msi.h>
34 #include <linux/of.h>
35 #include <linux/of_address.h>
36 #include <linux/of_irq.h>
37 #include <linux/of_pci.h>
38 #include <linux/of_platform.h>
39 #include <linux/percpu.h>
40 #include <linux/slab.h>
41 #include <linux/syscore_ops.h>
42
43 #include <linux/irqchip.h>
44 #include <linux/irqchip/arm-gic-v3.h>
45 #include <linux/irqchip/arm-gic-v4.h>
46
47 #include <asm/cputype.h>
48 #include <asm/exception.h>
49
50 #include "irq-gic-common.h"
51
52 #define ITS_FLAGS_CMDQ_NEEDS_FLUSHING           (1ULL << 0)
53 #define ITS_FLAGS_WORKAROUND_CAVIUM_22375       (1ULL << 1)
54 #define ITS_FLAGS_WORKAROUND_CAVIUM_23144       (1ULL << 2)
55 #define ITS_FLAGS_SAVE_SUSPEND_STATE            (1ULL << 3)
56
57 #define RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING     (1 << 0)
58 #define RDIST_FLAGS_RD_TABLES_PREALLOCATED      (1 << 1)
59
60 static u32 lpi_id_bits;
61
62 /*
63  * We allocate memory for PROPBASE to cover 2 ^ lpi_id_bits LPIs to
64  * deal with (one configuration byte per interrupt). PENDBASE has to
65  * be 64kB aligned (one bit per LPI, plus 8192 bits for SPI/PPI/SGI).
66  */
67 #define LPI_NRBITS              lpi_id_bits
68 #define LPI_PROPBASE_SZ         ALIGN(BIT(LPI_NRBITS), SZ_64K)
69 #define LPI_PENDBASE_SZ         ALIGN(BIT(LPI_NRBITS) / 8, SZ_64K)
70
71 #define LPI_PROP_DEFAULT_PRIO   GICD_INT_DEF_PRI
72
73 /*
74  * Collection structure - just an ID, and a redistributor address to
75  * ping. We use one per CPU as a bag of interrupts assigned to this
76  * CPU.
77  */
78 struct its_collection {
79         u64                     target_address;
80         u16                     col_id;
81 };
82
83 /*
84  * The ITS_BASER structure - contains memory information, cached
85  * value of BASER register configuration and ITS page size.
86  */
87 struct its_baser {
88         void            *base;
89         u64             val;
90         u32             order;
91         u32             psz;
92 };
93
94 struct its_device;
95
96 /*
97  * The ITS structure - contains most of the infrastructure, with the
98  * top-level MSI domain, the command queue, the collections, and the
99  * list of devices writing to it.
100  *
101  * dev_alloc_lock has to be taken for device allocations, while the
102  * spinlock must be taken to parse data structures such as the device
103  * list.
104  */
105 struct its_node {
106         raw_spinlock_t          lock;
107         struct mutex            dev_alloc_lock;
108         struct list_head        entry;
109         void __iomem            *base;
110         phys_addr_t             phys_base;
111         struct its_cmd_block    *cmd_base;
112         struct its_cmd_block    *cmd_write;
113         struct its_baser        tables[GITS_BASER_NR_REGS];
114         struct its_collection   *collections;
115         struct fwnode_handle    *fwnode_handle;
116         u64                     (*get_msi_base)(struct its_device *its_dev);
117         u64                     cbaser_save;
118         u32                     ctlr_save;
119         struct list_head        its_device_list;
120         u64                     flags;
121         unsigned long           list_nr;
122         u32                     ite_size;
123         u32                     device_ids;
124         int                     numa_node;
125         unsigned int            msi_domain_flags;
126         u32                     pre_its_base; /* for Socionext Synquacer */
127         bool                    is_v4;
128         int                     vlpi_redist_offset;
129 };
130
131 #define ITS_ITT_ALIGN           SZ_256
132
133 /* The maximum number of VPEID bits supported by VLPI commands */
134 #define ITS_MAX_VPEID_BITS      (16)
135 #define ITS_MAX_VPEID           (1 << (ITS_MAX_VPEID_BITS))
136
137 /* Convert page order to size in bytes */
138 #define PAGE_ORDER_TO_SIZE(o)   (PAGE_SIZE << (o))
139
140 struct event_lpi_map {
141         unsigned long           *lpi_map;
142         u16                     *col_map;
143         irq_hw_number_t         lpi_base;
144         int                     nr_lpis;
145         struct mutex            vlpi_lock;
146         struct its_vm           *vm;
147         struct its_vlpi_map     *vlpi_maps;
148         int                     nr_vlpis;
149 };
150
151 /*
152  * The ITS view of a device - belongs to an ITS, owns an interrupt
153  * translation table, and a list of interrupts.  If it some of its
154  * LPIs are injected into a guest (GICv4), the event_map.vm field
155  * indicates which one.
156  */
157 struct its_device {
158         struct list_head        entry;
159         struct its_node         *its;
160         struct event_lpi_map    event_map;
161         void                    *itt;
162         u32                     nr_ites;
163         u32                     device_id;
164         bool                    shared;
165 };
166
167 static struct {
168         raw_spinlock_t          lock;
169         struct its_device       *dev;
170         struct its_vpe          **vpes;
171         int                     next_victim;
172 } vpe_proxy;
173
174 static LIST_HEAD(its_nodes);
175 static DEFINE_RAW_SPINLOCK(its_lock);
176 static struct rdists *gic_rdists;
177 static struct irq_domain *its_parent;
178
179 static unsigned long its_list_map;
180 static u16 vmovp_seq_num;
181 static DEFINE_RAW_SPINLOCK(vmovp_lock);
182
183 static DEFINE_IDA(its_vpeid_ida);
184
185 #define gic_data_rdist()                (raw_cpu_ptr(gic_rdists->rdist))
186 #define gic_data_rdist_cpu(cpu)         (per_cpu_ptr(gic_rdists->rdist, cpu))
187 #define gic_data_rdist_rd_base()        (gic_data_rdist()->rd_base)
188 #define gic_data_rdist_vlpi_base()      (gic_data_rdist_rd_base() + SZ_128K)
189
190 static struct its_collection *dev_event_to_col(struct its_device *its_dev,
191                                                u32 event)
192 {
193         struct its_node *its = its_dev->its;
194
195         return its->collections + its_dev->event_map.col_map[event];
196 }
197
198 static struct its_collection *valid_col(struct its_collection *col)
199 {
200         if (WARN_ON_ONCE(col->target_address & GENMASK_ULL(0, 15)))
201                 return NULL;
202
203         return col;
204 }
205
206 static struct its_vpe *valid_vpe(struct its_node *its, struct its_vpe *vpe)
207 {
208         if (valid_col(its->collections + vpe->col_idx))
209                 return vpe;
210
211         return NULL;
212 }
213
214 /*
215  * ITS command descriptors - parameters to be encoded in a command
216  * block.
217  */
218 struct its_cmd_desc {
219         union {
220                 struct {
221                         struct its_device *dev;
222                         u32 event_id;
223                 } its_inv_cmd;
224
225                 struct {
226                         struct its_device *dev;
227                         u32 event_id;
228                 } its_clear_cmd;
229
230                 struct {
231                         struct its_device *dev;
232                         u32 event_id;
233                 } its_int_cmd;
234
235                 struct {
236                         struct its_device *dev;
237                         int valid;
238                 } its_mapd_cmd;
239
240                 struct {
241                         struct its_collection *col;
242                         int valid;
243                 } its_mapc_cmd;
244
245                 struct {
246                         struct its_device *dev;
247                         u32 phys_id;
248                         u32 event_id;
249                 } its_mapti_cmd;
250
251                 struct {
252                         struct its_device *dev;
253                         struct its_collection *col;
254                         u32 event_id;
255                 } its_movi_cmd;
256
257                 struct {
258                         struct its_device *dev;
259                         u32 event_id;
260                 } its_discard_cmd;
261
262                 struct {
263                         struct its_collection *col;
264                 } its_invall_cmd;
265
266                 struct {
267                         struct its_vpe *vpe;
268                 } its_vinvall_cmd;
269
270                 struct {
271                         struct its_vpe *vpe;
272                         struct its_collection *col;
273                         bool valid;
274                 } its_vmapp_cmd;
275
276                 struct {
277                         struct its_vpe *vpe;
278                         struct its_device *dev;
279                         u32 virt_id;
280                         u32 event_id;
281                         bool db_enabled;
282                 } its_vmapti_cmd;
283
284                 struct {
285                         struct its_vpe *vpe;
286                         struct its_device *dev;
287                         u32 event_id;
288                         bool db_enabled;
289                 } its_vmovi_cmd;
290
291                 struct {
292                         struct its_vpe *vpe;
293                         struct its_collection *col;
294                         u16 seq_num;
295                         u16 its_list;
296                 } its_vmovp_cmd;
297         };
298 };
299
300 /*
301  * The ITS command block, which is what the ITS actually parses.
302  */
303 struct its_cmd_block {
304         u64     raw_cmd[4];
305 };
306
307 #define ITS_CMD_QUEUE_SZ                SZ_64K
308 #define ITS_CMD_QUEUE_NR_ENTRIES        (ITS_CMD_QUEUE_SZ / sizeof(struct its_cmd_block))
309
310 typedef struct its_collection *(*its_cmd_builder_t)(struct its_node *,
311                                                     struct its_cmd_block *,
312                                                     struct its_cmd_desc *);
313
314 typedef struct its_vpe *(*its_cmd_vbuilder_t)(struct its_node *,
315                                               struct its_cmd_block *,
316                                               struct its_cmd_desc *);
317
318 static void its_mask_encode(u64 *raw_cmd, u64 val, int h, int l)
319 {
320         u64 mask = GENMASK_ULL(h, l);
321         *raw_cmd &= ~mask;
322         *raw_cmd |= (val << l) & mask;
323 }
324
325 static void its_encode_cmd(struct its_cmd_block *cmd, u8 cmd_nr)
326 {
327         its_mask_encode(&cmd->raw_cmd[0], cmd_nr, 7, 0);
328 }
329
330 static void its_encode_devid(struct its_cmd_block *cmd, u32 devid)
331 {
332         its_mask_encode(&cmd->raw_cmd[0], devid, 63, 32);
333 }
334
335 static void its_encode_event_id(struct its_cmd_block *cmd, u32 id)
336 {
337         its_mask_encode(&cmd->raw_cmd[1], id, 31, 0);
338 }
339
340 static void its_encode_phys_id(struct its_cmd_block *cmd, u32 phys_id)
341 {
342         its_mask_encode(&cmd->raw_cmd[1], phys_id, 63, 32);
343 }
344
345 static void its_encode_size(struct its_cmd_block *cmd, u8 size)
346 {
347         its_mask_encode(&cmd->raw_cmd[1], size, 4, 0);
348 }
349
350 static void its_encode_itt(struct its_cmd_block *cmd, u64 itt_addr)
351 {
352         its_mask_encode(&cmd->raw_cmd[2], itt_addr >> 8, 51, 8);
353 }
354
355 static void its_encode_valid(struct its_cmd_block *cmd, int valid)
356 {
357         its_mask_encode(&cmd->raw_cmd[2], !!valid, 63, 63);
358 }
359
360 static void its_encode_target(struct its_cmd_block *cmd, u64 target_addr)
361 {
362         its_mask_encode(&cmd->raw_cmd[2], target_addr >> 16, 51, 16);
363 }
364
365 static void its_encode_collection(struct its_cmd_block *cmd, u16 col)
366 {
367         its_mask_encode(&cmd->raw_cmd[2], col, 15, 0);
368 }
369
370 static void its_encode_vpeid(struct its_cmd_block *cmd, u16 vpeid)
371 {
372         its_mask_encode(&cmd->raw_cmd[1], vpeid, 47, 32);
373 }
374
375 static void its_encode_virt_id(struct its_cmd_block *cmd, u32 virt_id)
376 {
377         its_mask_encode(&cmd->raw_cmd[2], virt_id, 31, 0);
378 }
379
380 static void its_encode_db_phys_id(struct its_cmd_block *cmd, u32 db_phys_id)
381 {
382         its_mask_encode(&cmd->raw_cmd[2], db_phys_id, 63, 32);
383 }
384
385 static void its_encode_db_valid(struct its_cmd_block *cmd, bool db_valid)
386 {
387         its_mask_encode(&cmd->raw_cmd[2], db_valid, 0, 0);
388 }
389
390 static void its_encode_seq_num(struct its_cmd_block *cmd, u16 seq_num)
391 {
392         its_mask_encode(&cmd->raw_cmd[0], seq_num, 47, 32);
393 }
394
395 static void its_encode_its_list(struct its_cmd_block *cmd, u16 its_list)
396 {
397         its_mask_encode(&cmd->raw_cmd[1], its_list, 15, 0);
398 }
399
400 static void its_encode_vpt_addr(struct its_cmd_block *cmd, u64 vpt_pa)
401 {
402         its_mask_encode(&cmd->raw_cmd[3], vpt_pa >> 16, 51, 16);
403 }
404
405 static void its_encode_vpt_size(struct its_cmd_block *cmd, u8 vpt_size)
406 {
407         its_mask_encode(&cmd->raw_cmd[3], vpt_size, 4, 0);
408 }
409
410 static inline void its_fixup_cmd(struct its_cmd_block *cmd)
411 {
412         /* Let's fixup BE commands */
413         cmd->raw_cmd[0] = cpu_to_le64(cmd->raw_cmd[0]);
414         cmd->raw_cmd[1] = cpu_to_le64(cmd->raw_cmd[1]);
415         cmd->raw_cmd[2] = cpu_to_le64(cmd->raw_cmd[2]);
416         cmd->raw_cmd[3] = cpu_to_le64(cmd->raw_cmd[3]);
417 }
418
419 static struct its_collection *its_build_mapd_cmd(struct its_node *its,
420                                                  struct its_cmd_block *cmd,
421                                                  struct its_cmd_desc *desc)
422 {
423         unsigned long itt_addr;
424         u8 size = ilog2(desc->its_mapd_cmd.dev->nr_ites);
425
426         itt_addr = virt_to_phys(desc->its_mapd_cmd.dev->itt);
427         itt_addr = ALIGN(itt_addr, ITS_ITT_ALIGN);
428
429         its_encode_cmd(cmd, GITS_CMD_MAPD);
430         its_encode_devid(cmd, desc->its_mapd_cmd.dev->device_id);
431         its_encode_size(cmd, size - 1);
432         its_encode_itt(cmd, itt_addr);
433         its_encode_valid(cmd, desc->its_mapd_cmd.valid);
434
435         its_fixup_cmd(cmd);
436
437         return NULL;
438 }
439
440 static struct its_collection *its_build_mapc_cmd(struct its_node *its,
441                                                  struct its_cmd_block *cmd,
442                                                  struct its_cmd_desc *desc)
443 {
444         its_encode_cmd(cmd, GITS_CMD_MAPC);
445         its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
446         its_encode_target(cmd, desc->its_mapc_cmd.col->target_address);
447         its_encode_valid(cmd, desc->its_mapc_cmd.valid);
448
449         its_fixup_cmd(cmd);
450
451         return desc->its_mapc_cmd.col;
452 }
453
454 static struct its_collection *its_build_mapti_cmd(struct its_node *its,
455                                                   struct its_cmd_block *cmd,
456                                                   struct its_cmd_desc *desc)
457 {
458         struct its_collection *col;
459
460         col = dev_event_to_col(desc->its_mapti_cmd.dev,
461                                desc->its_mapti_cmd.event_id);
462
463         its_encode_cmd(cmd, GITS_CMD_MAPTI);
464         its_encode_devid(cmd, desc->its_mapti_cmd.dev->device_id);
465         its_encode_event_id(cmd, desc->its_mapti_cmd.event_id);
466         its_encode_phys_id(cmd, desc->its_mapti_cmd.phys_id);
467         its_encode_collection(cmd, col->col_id);
468
469         its_fixup_cmd(cmd);
470
471         return valid_col(col);
472 }
473
474 static struct its_collection *its_build_movi_cmd(struct its_node *its,
475                                                  struct its_cmd_block *cmd,
476                                                  struct its_cmd_desc *desc)
477 {
478         struct its_collection *col;
479
480         col = dev_event_to_col(desc->its_movi_cmd.dev,
481                                desc->its_movi_cmd.event_id);
482
483         its_encode_cmd(cmd, GITS_CMD_MOVI);
484         its_encode_devid(cmd, desc->its_movi_cmd.dev->device_id);
485         its_encode_event_id(cmd, desc->its_movi_cmd.event_id);
486         its_encode_collection(cmd, desc->its_movi_cmd.col->col_id);
487
488         its_fixup_cmd(cmd);
489
490         return valid_col(col);
491 }
492
493 static struct its_collection *its_build_discard_cmd(struct its_node *its,
494                                                     struct its_cmd_block *cmd,
495                                                     struct its_cmd_desc *desc)
496 {
497         struct its_collection *col;
498
499         col = dev_event_to_col(desc->its_discard_cmd.dev,
500                                desc->its_discard_cmd.event_id);
501
502         its_encode_cmd(cmd, GITS_CMD_DISCARD);
503         its_encode_devid(cmd, desc->its_discard_cmd.dev->device_id);
504         its_encode_event_id(cmd, desc->its_discard_cmd.event_id);
505
506         its_fixup_cmd(cmd);
507
508         return valid_col(col);
509 }
510
511 static struct its_collection *its_build_inv_cmd(struct its_node *its,
512                                                 struct its_cmd_block *cmd,
513                                                 struct its_cmd_desc *desc)
514 {
515         struct its_collection *col;
516
517         col = dev_event_to_col(desc->its_inv_cmd.dev,
518                                desc->its_inv_cmd.event_id);
519
520         its_encode_cmd(cmd, GITS_CMD_INV);
521         its_encode_devid(cmd, desc->its_inv_cmd.dev->device_id);
522         its_encode_event_id(cmd, desc->its_inv_cmd.event_id);
523
524         its_fixup_cmd(cmd);
525
526         return valid_col(col);
527 }
528
529 static struct its_collection *its_build_int_cmd(struct its_node *its,
530                                                 struct its_cmd_block *cmd,
531                                                 struct its_cmd_desc *desc)
532 {
533         struct its_collection *col;
534
535         col = dev_event_to_col(desc->its_int_cmd.dev,
536                                desc->its_int_cmd.event_id);
537
538         its_encode_cmd(cmd, GITS_CMD_INT);
539         its_encode_devid(cmd, desc->its_int_cmd.dev->device_id);
540         its_encode_event_id(cmd, desc->its_int_cmd.event_id);
541
542         its_fixup_cmd(cmd);
543
544         return valid_col(col);
545 }
546
547 static struct its_collection *its_build_clear_cmd(struct its_node *its,
548                                                   struct its_cmd_block *cmd,
549                                                   struct its_cmd_desc *desc)
550 {
551         struct its_collection *col;
552
553         col = dev_event_to_col(desc->its_clear_cmd.dev,
554                                desc->its_clear_cmd.event_id);
555
556         its_encode_cmd(cmd, GITS_CMD_CLEAR);
557         its_encode_devid(cmd, desc->its_clear_cmd.dev->device_id);
558         its_encode_event_id(cmd, desc->its_clear_cmd.event_id);
559
560         its_fixup_cmd(cmd);
561
562         return valid_col(col);
563 }
564
565 static struct its_collection *its_build_invall_cmd(struct its_node *its,
566                                                    struct its_cmd_block *cmd,
567                                                    struct its_cmd_desc *desc)
568 {
569         its_encode_cmd(cmd, GITS_CMD_INVALL);
570         its_encode_collection(cmd, desc->its_mapc_cmd.col->col_id);
571
572         its_fixup_cmd(cmd);
573
574         return NULL;
575 }
576
577 static struct its_vpe *its_build_vinvall_cmd(struct its_node *its,
578                                              struct its_cmd_block *cmd,
579                                              struct its_cmd_desc *desc)
580 {
581         its_encode_cmd(cmd, GITS_CMD_VINVALL);
582         its_encode_vpeid(cmd, desc->its_vinvall_cmd.vpe->vpe_id);
583
584         its_fixup_cmd(cmd);
585
586         return valid_vpe(its, desc->its_vinvall_cmd.vpe);
587 }
588
589 static struct its_vpe *its_build_vmapp_cmd(struct its_node *its,
590                                            struct its_cmd_block *cmd,
591                                            struct its_cmd_desc *desc)
592 {
593         unsigned long vpt_addr;
594         u64 target;
595
596         vpt_addr = virt_to_phys(page_address(desc->its_vmapp_cmd.vpe->vpt_page));
597         target = desc->its_vmapp_cmd.col->target_address + its->vlpi_redist_offset;
598
599         its_encode_cmd(cmd, GITS_CMD_VMAPP);
600         its_encode_vpeid(cmd, desc->its_vmapp_cmd.vpe->vpe_id);
601         its_encode_valid(cmd, desc->its_vmapp_cmd.valid);
602         its_encode_target(cmd, target);
603         its_encode_vpt_addr(cmd, vpt_addr);
604         its_encode_vpt_size(cmd, LPI_NRBITS - 1);
605
606         its_fixup_cmd(cmd);
607
608         return valid_vpe(its, desc->its_vmapp_cmd.vpe);
609 }
610
611 static struct its_vpe *its_build_vmapti_cmd(struct its_node *its,
612                                             struct its_cmd_block *cmd,
613                                             struct its_cmd_desc *desc)
614 {
615         u32 db;
616
617         if (desc->its_vmapti_cmd.db_enabled)
618                 db = desc->its_vmapti_cmd.vpe->vpe_db_lpi;
619         else
620                 db = 1023;
621
622         its_encode_cmd(cmd, GITS_CMD_VMAPTI);
623         its_encode_devid(cmd, desc->its_vmapti_cmd.dev->device_id);
624         its_encode_vpeid(cmd, desc->its_vmapti_cmd.vpe->vpe_id);
625         its_encode_event_id(cmd, desc->its_vmapti_cmd.event_id);
626         its_encode_db_phys_id(cmd, db);
627         its_encode_virt_id(cmd, desc->its_vmapti_cmd.virt_id);
628
629         its_fixup_cmd(cmd);
630
631         return valid_vpe(its, desc->its_vmapti_cmd.vpe);
632 }
633
634 static struct its_vpe *its_build_vmovi_cmd(struct its_node *its,
635                                            struct its_cmd_block *cmd,
636                                            struct its_cmd_desc *desc)
637 {
638         u32 db;
639
640         if (desc->its_vmovi_cmd.db_enabled)
641                 db = desc->its_vmovi_cmd.vpe->vpe_db_lpi;
642         else
643                 db = 1023;
644
645         its_encode_cmd(cmd, GITS_CMD_VMOVI);
646         its_encode_devid(cmd, desc->its_vmovi_cmd.dev->device_id);
647         its_encode_vpeid(cmd, desc->its_vmovi_cmd.vpe->vpe_id);
648         its_encode_event_id(cmd, desc->its_vmovi_cmd.event_id);
649         its_encode_db_phys_id(cmd, db);
650         its_encode_db_valid(cmd, true);
651
652         its_fixup_cmd(cmd);
653
654         return valid_vpe(its, desc->its_vmovi_cmd.vpe);
655 }
656
657 static struct its_vpe *its_build_vmovp_cmd(struct its_node *its,
658                                            struct its_cmd_block *cmd,
659                                            struct its_cmd_desc *desc)
660 {
661         u64 target;
662
663         target = desc->its_vmovp_cmd.col->target_address + its->vlpi_redist_offset;
664         its_encode_cmd(cmd, GITS_CMD_VMOVP);
665         its_encode_seq_num(cmd, desc->its_vmovp_cmd.seq_num);
666         its_encode_its_list(cmd, desc->its_vmovp_cmd.its_list);
667         its_encode_vpeid(cmd, desc->its_vmovp_cmd.vpe->vpe_id);
668         its_encode_target(cmd, target);
669
670         its_fixup_cmd(cmd);
671
672         return valid_vpe(its, desc->its_vmovp_cmd.vpe);
673 }
674
675 static u64 its_cmd_ptr_to_offset(struct its_node *its,
676                                  struct its_cmd_block *ptr)
677 {
678         return (ptr - its->cmd_base) * sizeof(*ptr);
679 }
680
681 static int its_queue_full(struct its_node *its)
682 {
683         int widx;
684         int ridx;
685
686         widx = its->cmd_write - its->cmd_base;
687         ridx = readl_relaxed(its->base + GITS_CREADR) / sizeof(struct its_cmd_block);
688
689         /* This is incredibly unlikely to happen, unless the ITS locks up. */
690         if (((widx + 1) % ITS_CMD_QUEUE_NR_ENTRIES) == ridx)
691                 return 1;
692
693         return 0;
694 }
695
696 static struct its_cmd_block *its_allocate_entry(struct its_node *its)
697 {
698         struct its_cmd_block *cmd;
699         u32 count = 1000000;    /* 1s! */
700
701         while (its_queue_full(its)) {
702                 count--;
703                 if (!count) {
704                         pr_err_ratelimited("ITS queue not draining\n");
705                         return NULL;
706                 }
707                 cpu_relax();
708                 udelay(1);
709         }
710
711         cmd = its->cmd_write++;
712
713         /* Handle queue wrapping */
714         if (its->cmd_write == (its->cmd_base + ITS_CMD_QUEUE_NR_ENTRIES))
715                 its->cmd_write = its->cmd_base;
716
717         /* Clear command  */
718         cmd->raw_cmd[0] = 0;
719         cmd->raw_cmd[1] = 0;
720         cmd->raw_cmd[2] = 0;
721         cmd->raw_cmd[3] = 0;
722
723         return cmd;
724 }
725
726 static struct its_cmd_block *its_post_commands(struct its_node *its)
727 {
728         u64 wr = its_cmd_ptr_to_offset(its, its->cmd_write);
729
730         writel_relaxed(wr, its->base + GITS_CWRITER);
731
732         return its->cmd_write;
733 }
734
735 static void its_flush_cmd(struct its_node *its, struct its_cmd_block *cmd)
736 {
737         /*
738          * Make sure the commands written to memory are observable by
739          * the ITS.
740          */
741         if (its->flags & ITS_FLAGS_CMDQ_NEEDS_FLUSHING)
742                 gic_flush_dcache_to_poc(cmd, sizeof(*cmd));
743         else
744                 dsb(ishst);
745 }
746
747 static int its_wait_for_range_completion(struct its_node *its,
748                                          struct its_cmd_block *from,
749                                          struct its_cmd_block *to)
750 {
751         u64 rd_idx, from_idx, to_idx;
752         u32 count = 1000000;    /* 1s! */
753
754         from_idx = its_cmd_ptr_to_offset(its, from);
755         to_idx = its_cmd_ptr_to_offset(its, to);
756
757         while (1) {
758                 rd_idx = readl_relaxed(its->base + GITS_CREADR);
759
760                 /* Direct case */
761                 if (from_idx < to_idx && rd_idx >= to_idx)
762                         break;
763
764                 /* Wrapped case */
765                 if (from_idx >= to_idx && rd_idx >= to_idx && rd_idx < from_idx)
766                         break;
767
768                 count--;
769                 if (!count) {
770                         pr_err_ratelimited("ITS queue timeout (%llu %llu %llu)\n",
771                                            from_idx, to_idx, rd_idx);
772                         return -1;
773                 }
774                 cpu_relax();
775                 udelay(1);
776         }
777
778         return 0;
779 }
780
781 /* Warning, macro hell follows */
782 #define BUILD_SINGLE_CMD_FUNC(name, buildtype, synctype, buildfn)       \
783 void name(struct its_node *its,                                         \
784           buildtype builder,                                            \
785           struct its_cmd_desc *desc)                                    \
786 {                                                                       \
787         struct its_cmd_block *cmd, *sync_cmd, *next_cmd;                \
788         synctype *sync_obj;                                             \
789         unsigned long flags;                                            \
790                                                                         \
791         raw_spin_lock_irqsave(&its->lock, flags);                       \
792                                                                         \
793         cmd = its_allocate_entry(its);                                  \
794         if (!cmd) {             /* We're soooooo screewed... */         \
795                 raw_spin_unlock_irqrestore(&its->lock, flags);          \
796                 return;                                                 \
797         }                                                               \
798         sync_obj = builder(its, cmd, desc);                             \
799         its_flush_cmd(its, cmd);                                        \
800                                                                         \
801         if (sync_obj) {                                                 \
802                 sync_cmd = its_allocate_entry(its);                     \
803                 if (!sync_cmd)                                          \
804                         goto post;                                      \
805                                                                         \
806                 buildfn(its, sync_cmd, sync_obj);                       \
807                 its_flush_cmd(its, sync_cmd);                           \
808         }                                                               \
809                                                                         \
810 post:                                                                   \
811         next_cmd = its_post_commands(its);                              \
812         raw_spin_unlock_irqrestore(&its->lock, flags);                  \
813                                                                         \
814         if (its_wait_for_range_completion(its, cmd, next_cmd))          \
815                 pr_err_ratelimited("ITS cmd %ps failed\n", builder);    \
816 }
817
818 static void its_build_sync_cmd(struct its_node *its,
819                                struct its_cmd_block *sync_cmd,
820                                struct its_collection *sync_col)
821 {
822         its_encode_cmd(sync_cmd, GITS_CMD_SYNC);
823         its_encode_target(sync_cmd, sync_col->target_address);
824
825         its_fixup_cmd(sync_cmd);
826 }
827
828 static BUILD_SINGLE_CMD_FUNC(its_send_single_command, its_cmd_builder_t,
829                              struct its_collection, its_build_sync_cmd)
830
831 static void its_build_vsync_cmd(struct its_node *its,
832                                 struct its_cmd_block *sync_cmd,
833                                 struct its_vpe *sync_vpe)
834 {
835         its_encode_cmd(sync_cmd, GITS_CMD_VSYNC);
836         its_encode_vpeid(sync_cmd, sync_vpe->vpe_id);
837
838         its_fixup_cmd(sync_cmd);
839 }
840
841 static BUILD_SINGLE_CMD_FUNC(its_send_single_vcommand, its_cmd_vbuilder_t,
842                              struct its_vpe, its_build_vsync_cmd)
843
844 static void its_send_int(struct its_device *dev, u32 event_id)
845 {
846         struct its_cmd_desc desc;
847
848         desc.its_int_cmd.dev = dev;
849         desc.its_int_cmd.event_id = event_id;
850
851         its_send_single_command(dev->its, its_build_int_cmd, &desc);
852 }
853
854 static void its_send_clear(struct its_device *dev, u32 event_id)
855 {
856         struct its_cmd_desc desc;
857
858         desc.its_clear_cmd.dev = dev;
859         desc.its_clear_cmd.event_id = event_id;
860
861         its_send_single_command(dev->its, its_build_clear_cmd, &desc);
862 }
863
864 static void its_send_inv(struct its_device *dev, u32 event_id)
865 {
866         struct its_cmd_desc desc;
867
868         desc.its_inv_cmd.dev = dev;
869         desc.its_inv_cmd.event_id = event_id;
870
871         its_send_single_command(dev->its, its_build_inv_cmd, &desc);
872 }
873
874 static void its_send_mapd(struct its_device *dev, int valid)
875 {
876         struct its_cmd_desc desc;
877
878         desc.its_mapd_cmd.dev = dev;
879         desc.its_mapd_cmd.valid = !!valid;
880
881         its_send_single_command(dev->its, its_build_mapd_cmd, &desc);
882 }
883
884 static void its_send_mapc(struct its_node *its, struct its_collection *col,
885                           int valid)
886 {
887         struct its_cmd_desc desc;
888
889         desc.its_mapc_cmd.col = col;
890         desc.its_mapc_cmd.valid = !!valid;
891
892         its_send_single_command(its, its_build_mapc_cmd, &desc);
893 }
894
895 static void its_send_mapti(struct its_device *dev, u32 irq_id, u32 id)
896 {
897         struct its_cmd_desc desc;
898
899         desc.its_mapti_cmd.dev = dev;
900         desc.its_mapti_cmd.phys_id = irq_id;
901         desc.its_mapti_cmd.event_id = id;
902
903         its_send_single_command(dev->its, its_build_mapti_cmd, &desc);
904 }
905
906 static void its_send_movi(struct its_device *dev,
907                           struct its_collection *col, u32 id)
908 {
909         struct its_cmd_desc desc;
910
911         desc.its_movi_cmd.dev = dev;
912         desc.its_movi_cmd.col = col;
913         desc.its_movi_cmd.event_id = id;
914
915         its_send_single_command(dev->its, its_build_movi_cmd, &desc);
916 }
917
918 static void its_send_discard(struct its_device *dev, u32 id)
919 {
920         struct its_cmd_desc desc;
921
922         desc.its_discard_cmd.dev = dev;
923         desc.its_discard_cmd.event_id = id;
924
925         its_send_single_command(dev->its, its_build_discard_cmd, &desc);
926 }
927
928 static void its_send_invall(struct its_node *its, struct its_collection *col)
929 {
930         struct its_cmd_desc desc;
931
932         desc.its_invall_cmd.col = col;
933
934         its_send_single_command(its, its_build_invall_cmd, &desc);
935 }
936
937 static void its_send_vmapti(struct its_device *dev, u32 id)
938 {
939         struct its_vlpi_map *map = &dev->event_map.vlpi_maps[id];
940         struct its_cmd_desc desc;
941
942         desc.its_vmapti_cmd.vpe = map->vpe;
943         desc.its_vmapti_cmd.dev = dev;
944         desc.its_vmapti_cmd.virt_id = map->vintid;
945         desc.its_vmapti_cmd.event_id = id;
946         desc.its_vmapti_cmd.db_enabled = map->db_enabled;
947
948         its_send_single_vcommand(dev->its, its_build_vmapti_cmd, &desc);
949 }
950
951 static void its_send_vmovi(struct its_device *dev, u32 id)
952 {
953         struct its_vlpi_map *map = &dev->event_map.vlpi_maps[id];
954         struct its_cmd_desc desc;
955
956         desc.its_vmovi_cmd.vpe = map->vpe;
957         desc.its_vmovi_cmd.dev = dev;
958         desc.its_vmovi_cmd.event_id = id;
959         desc.its_vmovi_cmd.db_enabled = map->db_enabled;
960
961         its_send_single_vcommand(dev->its, its_build_vmovi_cmd, &desc);
962 }
963
964 static void its_send_vmapp(struct its_node *its,
965                            struct its_vpe *vpe, bool valid)
966 {
967         struct its_cmd_desc desc;
968
969         desc.its_vmapp_cmd.vpe = vpe;
970         desc.its_vmapp_cmd.valid = valid;
971         desc.its_vmapp_cmd.col = &its->collections[vpe->col_idx];
972
973         its_send_single_vcommand(its, its_build_vmapp_cmd, &desc);
974 }
975
976 static void its_send_vmovp(struct its_vpe *vpe)
977 {
978         struct its_cmd_desc desc;
979         struct its_node *its;
980         unsigned long flags;
981         int col_id = vpe->col_idx;
982
983         desc.its_vmovp_cmd.vpe = vpe;
984         desc.its_vmovp_cmd.its_list = (u16)its_list_map;
985
986         if (!its_list_map) {
987                 its = list_first_entry(&its_nodes, struct its_node, entry);
988                 desc.its_vmovp_cmd.seq_num = 0;
989                 desc.its_vmovp_cmd.col = &its->collections[col_id];
990                 its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
991                 return;
992         }
993
994         /*
995          * Yet another marvel of the architecture. If using the
996          * its_list "feature", we need to make sure that all ITSs
997          * receive all VMOVP commands in the same order. The only way
998          * to guarantee this is to make vmovp a serialization point.
999          *
1000          * Wall <-- Head.
1001          */
1002         raw_spin_lock_irqsave(&vmovp_lock, flags);
1003
1004         desc.its_vmovp_cmd.seq_num = vmovp_seq_num++;
1005
1006         /* Emit VMOVPs */
1007         list_for_each_entry(its, &its_nodes, entry) {
1008                 if (!its->is_v4)
1009                         continue;
1010
1011                 if (!vpe->its_vm->vlpi_count[its->list_nr])
1012                         continue;
1013
1014                 desc.its_vmovp_cmd.col = &its->collections[col_id];
1015                 its_send_single_vcommand(its, its_build_vmovp_cmd, &desc);
1016         }
1017
1018         raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1019 }
1020
1021 static void its_send_vinvall(struct its_node *its, struct its_vpe *vpe)
1022 {
1023         struct its_cmd_desc desc;
1024
1025         desc.its_vinvall_cmd.vpe = vpe;
1026         its_send_single_vcommand(its, its_build_vinvall_cmd, &desc);
1027 }
1028
1029 /*
1030  * irqchip functions - assumes MSI, mostly.
1031  */
1032
1033 static inline u32 its_get_event_id(struct irq_data *d)
1034 {
1035         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1036         return d->hwirq - its_dev->event_map.lpi_base;
1037 }
1038
1039 static void lpi_write_config(struct irq_data *d, u8 clr, u8 set)
1040 {
1041         irq_hw_number_t hwirq;
1042         void *va;
1043         u8 *cfg;
1044
1045         if (irqd_is_forwarded_to_vcpu(d)) {
1046                 struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1047                 u32 event = its_get_event_id(d);
1048                 struct its_vlpi_map *map;
1049
1050                 va = page_address(its_dev->event_map.vm->vprop_page);
1051                 map = &its_dev->event_map.vlpi_maps[event];
1052                 hwirq = map->vintid;
1053
1054                 /* Remember the updated property */
1055                 map->properties &= ~clr;
1056                 map->properties |= set | LPI_PROP_GROUP1;
1057         } else {
1058                 va = gic_rdists->prop_table_va;
1059                 hwirq = d->hwirq;
1060         }
1061
1062         cfg = va + hwirq - 8192;
1063         *cfg &= ~clr;
1064         *cfg |= set | LPI_PROP_GROUP1;
1065
1066         /*
1067          * Make the above write visible to the redistributors.
1068          * And yes, we're flushing exactly: One. Single. Byte.
1069          * Humpf...
1070          */
1071         if (gic_rdists->flags & RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING)
1072                 gic_flush_dcache_to_poc(cfg, sizeof(*cfg));
1073         else
1074                 dsb(ishst);
1075 }
1076
1077 static void lpi_update_config(struct irq_data *d, u8 clr, u8 set)
1078 {
1079         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1080
1081         lpi_write_config(d, clr, set);
1082         its_send_inv(its_dev, its_get_event_id(d));
1083 }
1084
1085 static void its_vlpi_set_doorbell(struct irq_data *d, bool enable)
1086 {
1087         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1088         u32 event = its_get_event_id(d);
1089
1090         if (its_dev->event_map.vlpi_maps[event].db_enabled == enable)
1091                 return;
1092
1093         its_dev->event_map.vlpi_maps[event].db_enabled = enable;
1094
1095         /*
1096          * More fun with the architecture:
1097          *
1098          * Ideally, we'd issue a VMAPTI to set the doorbell to its LPI
1099          * value or to 1023, depending on the enable bit. But that
1100          * would be issueing a mapping for an /existing/ DevID+EventID
1101          * pair, which is UNPREDICTABLE. Instead, let's issue a VMOVI
1102          * to the /same/ vPE, using this opportunity to adjust the
1103          * doorbell. Mouahahahaha. We loves it, Precious.
1104          */
1105         its_send_vmovi(its_dev, event);
1106 }
1107
1108 static void its_mask_irq(struct irq_data *d)
1109 {
1110         if (irqd_is_forwarded_to_vcpu(d))
1111                 its_vlpi_set_doorbell(d, false);
1112
1113         lpi_update_config(d, LPI_PROP_ENABLED, 0);
1114 }
1115
1116 static void its_unmask_irq(struct irq_data *d)
1117 {
1118         if (irqd_is_forwarded_to_vcpu(d))
1119                 its_vlpi_set_doorbell(d, true);
1120
1121         lpi_update_config(d, 0, LPI_PROP_ENABLED);
1122 }
1123
1124 static int its_set_affinity(struct irq_data *d, const struct cpumask *mask_val,
1125                             bool force)
1126 {
1127         unsigned int cpu;
1128         const struct cpumask *cpu_mask = cpu_online_mask;
1129         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1130         struct its_collection *target_col;
1131         u32 id = its_get_event_id(d);
1132
1133         /* A forwarded interrupt should use irq_set_vcpu_affinity */
1134         if (irqd_is_forwarded_to_vcpu(d))
1135                 return -EINVAL;
1136
1137        /* lpi cannot be routed to a redistributor that is on a foreign node */
1138         if (its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) {
1139                 if (its_dev->its->numa_node >= 0) {
1140                         cpu_mask = cpumask_of_node(its_dev->its->numa_node);
1141                         if (!cpumask_intersects(mask_val, cpu_mask))
1142                                 return -EINVAL;
1143                 }
1144         }
1145
1146         cpu = cpumask_any_and(mask_val, cpu_mask);
1147
1148         if (cpu >= nr_cpu_ids)
1149                 return -EINVAL;
1150
1151         /* don't set the affinity when the target cpu is same as current one */
1152         if (cpu != its_dev->event_map.col_map[id]) {
1153                 target_col = &its_dev->its->collections[cpu];
1154                 its_send_movi(its_dev, target_col, id);
1155                 its_dev->event_map.col_map[id] = cpu;
1156                 irq_data_update_effective_affinity(d, cpumask_of(cpu));
1157         }
1158
1159         return IRQ_SET_MASK_OK_DONE;
1160 }
1161
1162 static u64 its_irq_get_msi_base(struct its_device *its_dev)
1163 {
1164         struct its_node *its = its_dev->its;
1165
1166         return its->phys_base + GITS_TRANSLATER;
1167 }
1168
1169 static void its_irq_compose_msi_msg(struct irq_data *d, struct msi_msg *msg)
1170 {
1171         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1172         struct its_node *its;
1173         u64 addr;
1174
1175         its = its_dev->its;
1176         addr = its->get_msi_base(its_dev);
1177
1178         msg->address_lo         = lower_32_bits(addr);
1179         msg->address_hi         = upper_32_bits(addr);
1180         msg->data               = its_get_event_id(d);
1181
1182         iommu_dma_map_msi_msg(d->irq, msg);
1183 }
1184
1185 static int its_irq_set_irqchip_state(struct irq_data *d,
1186                                      enum irqchip_irq_state which,
1187                                      bool state)
1188 {
1189         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1190         u32 event = its_get_event_id(d);
1191
1192         if (which != IRQCHIP_STATE_PENDING)
1193                 return -EINVAL;
1194
1195         if (state)
1196                 its_send_int(its_dev, event);
1197         else
1198                 its_send_clear(its_dev, event);
1199
1200         return 0;
1201 }
1202
1203 static void its_map_vm(struct its_node *its, struct its_vm *vm)
1204 {
1205         unsigned long flags;
1206
1207         /* Not using the ITS list? Everything is always mapped. */
1208         if (!its_list_map)
1209                 return;
1210
1211         raw_spin_lock_irqsave(&vmovp_lock, flags);
1212
1213         /*
1214          * If the VM wasn't mapped yet, iterate over the vpes and get
1215          * them mapped now.
1216          */
1217         vm->vlpi_count[its->list_nr]++;
1218
1219         if (vm->vlpi_count[its->list_nr] == 1) {
1220                 int i;
1221
1222                 for (i = 0; i < vm->nr_vpes; i++) {
1223                         struct its_vpe *vpe = vm->vpes[i];
1224                         struct irq_data *d = irq_get_irq_data(vpe->irq);
1225
1226                         /* Map the VPE to the first possible CPU */
1227                         vpe->col_idx = cpumask_first(cpu_online_mask);
1228                         its_send_vmapp(its, vpe, true);
1229                         its_send_vinvall(its, vpe);
1230                         irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
1231                 }
1232         }
1233
1234         raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1235 }
1236
1237 static void its_unmap_vm(struct its_node *its, struct its_vm *vm)
1238 {
1239         unsigned long flags;
1240
1241         /* Not using the ITS list? Everything is always mapped. */
1242         if (!its_list_map)
1243                 return;
1244
1245         raw_spin_lock_irqsave(&vmovp_lock, flags);
1246
1247         if (!--vm->vlpi_count[its->list_nr]) {
1248                 int i;
1249
1250                 for (i = 0; i < vm->nr_vpes; i++)
1251                         its_send_vmapp(its, vm->vpes[i], false);
1252         }
1253
1254         raw_spin_unlock_irqrestore(&vmovp_lock, flags);
1255 }
1256
1257 static int its_vlpi_map(struct irq_data *d, struct its_cmd_info *info)
1258 {
1259         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1260         u32 event = its_get_event_id(d);
1261         int ret = 0;
1262
1263         if (!info->map)
1264                 return -EINVAL;
1265
1266         mutex_lock(&its_dev->event_map.vlpi_lock);
1267
1268         if (!its_dev->event_map.vm) {
1269                 struct its_vlpi_map *maps;
1270
1271                 maps = kcalloc(its_dev->event_map.nr_lpis, sizeof(*maps),
1272                                GFP_KERNEL);
1273                 if (!maps) {
1274                         ret = -ENOMEM;
1275                         goto out;
1276                 }
1277
1278                 its_dev->event_map.vm = info->map->vm;
1279                 its_dev->event_map.vlpi_maps = maps;
1280         } else if (its_dev->event_map.vm != info->map->vm) {
1281                 ret = -EINVAL;
1282                 goto out;
1283         }
1284
1285         /* Get our private copy of the mapping information */
1286         its_dev->event_map.vlpi_maps[event] = *info->map;
1287
1288         if (irqd_is_forwarded_to_vcpu(d)) {
1289                 /* Already mapped, move it around */
1290                 its_send_vmovi(its_dev, event);
1291         } else {
1292                 /* Ensure all the VPEs are mapped on this ITS */
1293                 its_map_vm(its_dev->its, info->map->vm);
1294
1295                 /*
1296                  * Flag the interrupt as forwarded so that we can
1297                  * start poking the virtual property table.
1298                  */
1299                 irqd_set_forwarded_to_vcpu(d);
1300
1301                 /* Write out the property to the prop table */
1302                 lpi_write_config(d, 0xff, info->map->properties);
1303
1304                 /* Drop the physical mapping */
1305                 its_send_discard(its_dev, event);
1306
1307                 /* and install the virtual one */
1308                 its_send_vmapti(its_dev, event);
1309
1310                 /* Increment the number of VLPIs */
1311                 its_dev->event_map.nr_vlpis++;
1312         }
1313
1314 out:
1315         mutex_unlock(&its_dev->event_map.vlpi_lock);
1316         return ret;
1317 }
1318
1319 static int its_vlpi_get(struct irq_data *d, struct its_cmd_info *info)
1320 {
1321         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1322         u32 event = its_get_event_id(d);
1323         int ret = 0;
1324
1325         mutex_lock(&its_dev->event_map.vlpi_lock);
1326
1327         if (!its_dev->event_map.vm ||
1328             !its_dev->event_map.vlpi_maps[event].vm) {
1329                 ret = -EINVAL;
1330                 goto out;
1331         }
1332
1333         /* Copy our mapping information to the incoming request */
1334         *info->map = its_dev->event_map.vlpi_maps[event];
1335
1336 out:
1337         mutex_unlock(&its_dev->event_map.vlpi_lock);
1338         return ret;
1339 }
1340
1341 static int its_vlpi_unmap(struct irq_data *d)
1342 {
1343         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1344         u32 event = its_get_event_id(d);
1345         int ret = 0;
1346
1347         mutex_lock(&its_dev->event_map.vlpi_lock);
1348
1349         if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d)) {
1350                 ret = -EINVAL;
1351                 goto out;
1352         }
1353
1354         /* Drop the virtual mapping */
1355         its_send_discard(its_dev, event);
1356
1357         /* and restore the physical one */
1358         irqd_clr_forwarded_to_vcpu(d);
1359         its_send_mapti(its_dev, d->hwirq, event);
1360         lpi_update_config(d, 0xff, (LPI_PROP_DEFAULT_PRIO |
1361                                     LPI_PROP_ENABLED |
1362                                     LPI_PROP_GROUP1));
1363
1364         /* Potentially unmap the VM from this ITS */
1365         its_unmap_vm(its_dev->its, its_dev->event_map.vm);
1366
1367         /*
1368          * Drop the refcount and make the device available again if
1369          * this was the last VLPI.
1370          */
1371         if (!--its_dev->event_map.nr_vlpis) {
1372                 its_dev->event_map.vm = NULL;
1373                 kfree(its_dev->event_map.vlpi_maps);
1374         }
1375
1376 out:
1377         mutex_unlock(&its_dev->event_map.vlpi_lock);
1378         return ret;
1379 }
1380
1381 static int its_vlpi_prop_update(struct irq_data *d, struct its_cmd_info *info)
1382 {
1383         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1384
1385         if (!its_dev->event_map.vm || !irqd_is_forwarded_to_vcpu(d))
1386                 return -EINVAL;
1387
1388         if (info->cmd_type == PROP_UPDATE_AND_INV_VLPI)
1389                 lpi_update_config(d, 0xff, info->config);
1390         else
1391                 lpi_write_config(d, 0xff, info->config);
1392         its_vlpi_set_doorbell(d, !!(info->config & LPI_PROP_ENABLED));
1393
1394         return 0;
1395 }
1396
1397 static int its_irq_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
1398 {
1399         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
1400         struct its_cmd_info *info = vcpu_info;
1401
1402         /* Need a v4 ITS */
1403         if (!its_dev->its->is_v4)
1404                 return -EINVAL;
1405
1406         /* Unmap request? */
1407         if (!info)
1408                 return its_vlpi_unmap(d);
1409
1410         switch (info->cmd_type) {
1411         case MAP_VLPI:
1412                 return its_vlpi_map(d, info);
1413
1414         case GET_VLPI:
1415                 return its_vlpi_get(d, info);
1416
1417         case PROP_UPDATE_VLPI:
1418         case PROP_UPDATE_AND_INV_VLPI:
1419                 return its_vlpi_prop_update(d, info);
1420
1421         default:
1422                 return -EINVAL;
1423         }
1424 }
1425
1426 static struct irq_chip its_irq_chip = {
1427         .name                   = "ITS",
1428         .irq_mask               = its_mask_irq,
1429         .irq_unmask             = its_unmask_irq,
1430         .irq_eoi                = irq_chip_eoi_parent,
1431         .irq_set_affinity       = its_set_affinity,
1432         .irq_compose_msi_msg    = its_irq_compose_msi_msg,
1433         .irq_set_irqchip_state  = its_irq_set_irqchip_state,
1434         .irq_set_vcpu_affinity  = its_irq_set_vcpu_affinity,
1435 };
1436
1437
1438 /*
1439  * How we allocate LPIs:
1440  *
1441  * lpi_range_list contains ranges of LPIs that are to available to
1442  * allocate from. To allocate LPIs, just pick the first range that
1443  * fits the required allocation, and reduce it by the required
1444  * amount. Once empty, remove the range from the list.
1445  *
1446  * To free a range of LPIs, add a free range to the list, sort it and
1447  * merge the result if the new range happens to be adjacent to an
1448  * already free block.
1449  *
1450  * The consequence of the above is that allocation is cost is low, but
1451  * freeing is expensive. We assumes that freeing rarely occurs.
1452  */
1453 #define ITS_MAX_LPI_NRBITS      16 /* 64K LPIs */
1454
1455 static DEFINE_MUTEX(lpi_range_lock);
1456 static LIST_HEAD(lpi_range_list);
1457
1458 struct lpi_range {
1459         struct list_head        entry;
1460         u32                     base_id;
1461         u32                     span;
1462 };
1463
1464 static struct lpi_range *mk_lpi_range(u32 base, u32 span)
1465 {
1466         struct lpi_range *range;
1467
1468         range = kzalloc(sizeof(*range), GFP_KERNEL);
1469         if (range) {
1470                 INIT_LIST_HEAD(&range->entry);
1471                 range->base_id = base;
1472                 range->span = span;
1473         }
1474
1475         return range;
1476 }
1477
1478 static int lpi_range_cmp(void *priv, struct list_head *a, struct list_head *b)
1479 {
1480         struct lpi_range *ra, *rb;
1481
1482         ra = container_of(a, struct lpi_range, entry);
1483         rb = container_of(b, struct lpi_range, entry);
1484
1485         return rb->base_id - ra->base_id;
1486 }
1487
1488 static void merge_lpi_ranges(void)
1489 {
1490         struct lpi_range *range, *tmp;
1491
1492         list_for_each_entry_safe(range, tmp, &lpi_range_list, entry) {
1493                 if (!list_is_last(&range->entry, &lpi_range_list) &&
1494                     (tmp->base_id == (range->base_id + range->span))) {
1495                         tmp->base_id = range->base_id;
1496                         tmp->span += range->span;
1497                         list_del(&range->entry);
1498                         kfree(range);
1499                 }
1500         }
1501 }
1502
1503 static int alloc_lpi_range(u32 nr_lpis, u32 *base)
1504 {
1505         struct lpi_range *range, *tmp;
1506         int err = -ENOSPC;
1507
1508         mutex_lock(&lpi_range_lock);
1509
1510         list_for_each_entry_safe(range, tmp, &lpi_range_list, entry) {
1511                 if (range->span >= nr_lpis) {
1512                         *base = range->base_id;
1513                         range->base_id += nr_lpis;
1514                         range->span -= nr_lpis;
1515
1516                         if (range->span == 0) {
1517                                 list_del(&range->entry);
1518                                 kfree(range);
1519                         }
1520
1521                         err = 0;
1522                         break;
1523                 }
1524         }
1525
1526         mutex_unlock(&lpi_range_lock);
1527
1528         pr_debug("ITS: alloc %u:%u\n", *base, nr_lpis);
1529         return err;
1530 }
1531
1532 static int free_lpi_range(u32 base, u32 nr_lpis)
1533 {
1534         struct lpi_range *new;
1535         int err = 0;
1536
1537         mutex_lock(&lpi_range_lock);
1538
1539         new = mk_lpi_range(base, nr_lpis);
1540         if (!new) {
1541                 err = -ENOMEM;
1542                 goto out;
1543         }
1544
1545         list_add(&new->entry, &lpi_range_list);
1546         list_sort(NULL, &lpi_range_list, lpi_range_cmp);
1547         merge_lpi_ranges();
1548 out:
1549         mutex_unlock(&lpi_range_lock);
1550         return err;
1551 }
1552
1553 static int __init its_lpi_init(u32 id_bits)
1554 {
1555         u32 lpis = (1UL << id_bits) - 8192;
1556         u32 numlpis;
1557         int err;
1558
1559         numlpis = 1UL << GICD_TYPER_NUM_LPIS(gic_rdists->gicd_typer);
1560
1561         if (numlpis > 2 && !WARN_ON(numlpis > lpis)) {
1562                 lpis = numlpis;
1563                 pr_info("ITS: Using hypervisor restricted LPI range [%u]\n",
1564                         lpis);
1565         }
1566
1567         /*
1568          * Initializing the allocator is just the same as freeing the
1569          * full range of LPIs.
1570          */
1571         err = free_lpi_range(8192, lpis);
1572         pr_debug("ITS: Allocator initialized for %u LPIs\n", lpis);
1573         return err;
1574 }
1575
1576 static unsigned long *its_lpi_alloc(int nr_irqs, u32 *base, int *nr_ids)
1577 {
1578         unsigned long *bitmap = NULL;
1579         int err = 0;
1580
1581         do {
1582                 err = alloc_lpi_range(nr_irqs, base);
1583                 if (!err)
1584                         break;
1585
1586                 nr_irqs /= 2;
1587         } while (nr_irqs > 0);
1588
1589         if (!nr_irqs)
1590                 err = -ENOSPC;
1591
1592         if (err)
1593                 goto out;
1594
1595         bitmap = kcalloc(BITS_TO_LONGS(nr_irqs), sizeof (long), GFP_ATOMIC);
1596         if (!bitmap)
1597                 goto out;
1598
1599         *nr_ids = nr_irqs;
1600
1601 out:
1602         if (!bitmap)
1603                 *base = *nr_ids = 0;
1604
1605         return bitmap;
1606 }
1607
1608 static void its_lpi_free(unsigned long *bitmap, u32 base, u32 nr_ids)
1609 {
1610         WARN_ON(free_lpi_range(base, nr_ids));
1611         kfree(bitmap);
1612 }
1613
1614 static void gic_reset_prop_table(void *va)
1615 {
1616         /* Priority 0xa0, Group-1, disabled */
1617         memset(va, LPI_PROP_DEFAULT_PRIO | LPI_PROP_GROUP1, LPI_PROPBASE_SZ);
1618
1619         /* Make sure the GIC will observe the written configuration */
1620         gic_flush_dcache_to_poc(va, LPI_PROPBASE_SZ);
1621 }
1622
1623 static struct page *its_allocate_prop_table(gfp_t gfp_flags)
1624 {
1625         struct page *prop_page;
1626
1627         prop_page = alloc_pages(gfp_flags, get_order(LPI_PROPBASE_SZ));
1628         if (!prop_page)
1629                 return NULL;
1630
1631         gic_reset_prop_table(page_address(prop_page));
1632
1633         return prop_page;
1634 }
1635
1636 static void its_free_prop_table(struct page *prop_page)
1637 {
1638         free_pages((unsigned long)page_address(prop_page),
1639                    get_order(LPI_PROPBASE_SZ));
1640 }
1641
1642 static bool gic_check_reserved_range(phys_addr_t addr, unsigned long size)
1643 {
1644         phys_addr_t start, end, addr_end;
1645         u64 i;
1646
1647         /*
1648          * We don't bother checking for a kdump kernel as by
1649          * construction, the LPI tables are out of this kernel's
1650          * memory map.
1651          */
1652         if (is_kdump_kernel())
1653                 return true;
1654
1655         addr_end = addr + size - 1;
1656
1657         for_each_reserved_mem_region(i, &start, &end) {
1658                 if (addr >= start && addr_end <= end)
1659                         return true;
1660         }
1661
1662         /* Not found, not a good sign... */
1663         pr_warn("GICv3: Expected reserved range [%pa:%pa], not found\n",
1664                 &addr, &addr_end);
1665         add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
1666         return false;
1667 }
1668
1669 static int gic_reserve_range(phys_addr_t addr, unsigned long size)
1670 {
1671         if (efi_enabled(EFI_CONFIG_TABLES))
1672                 return efi_mem_reserve_persistent(addr, size);
1673
1674         return 0;
1675 }
1676
1677 static int __init its_setup_lpi_prop_table(void)
1678 {
1679         if (gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED) {
1680                 u64 val;
1681
1682                 val = gicr_read_propbaser(gic_data_rdist_rd_base() + GICR_PROPBASER);
1683                 lpi_id_bits = (val & GICR_PROPBASER_IDBITS_MASK) + 1;
1684
1685                 gic_rdists->prop_table_pa = val & GENMASK_ULL(51, 12);
1686                 gic_rdists->prop_table_va = memremap(gic_rdists->prop_table_pa,
1687                                                      LPI_PROPBASE_SZ,
1688                                                      MEMREMAP_WB);
1689                 gic_reset_prop_table(gic_rdists->prop_table_va);
1690         } else {
1691                 struct page *page;
1692
1693                 lpi_id_bits = min_t(u32,
1694                                     GICD_TYPER_ID_BITS(gic_rdists->gicd_typer),
1695                                     ITS_MAX_LPI_NRBITS);
1696                 page = its_allocate_prop_table(GFP_NOWAIT);
1697                 if (!page) {
1698                         pr_err("Failed to allocate PROPBASE\n");
1699                         return -ENOMEM;
1700                 }
1701
1702                 gic_rdists->prop_table_pa = page_to_phys(page);
1703                 gic_rdists->prop_table_va = page_address(page);
1704                 WARN_ON(gic_reserve_range(gic_rdists->prop_table_pa,
1705                                           LPI_PROPBASE_SZ));
1706         }
1707
1708         pr_info("GICv3: using LPI property table @%pa\n",
1709                 &gic_rdists->prop_table_pa);
1710
1711         return its_lpi_init(lpi_id_bits);
1712 }
1713
1714 static const char *its_base_type_string[] = {
1715         [GITS_BASER_TYPE_DEVICE]        = "Devices",
1716         [GITS_BASER_TYPE_VCPU]          = "Virtual CPUs",
1717         [GITS_BASER_TYPE_RESERVED3]     = "Reserved (3)",
1718         [GITS_BASER_TYPE_COLLECTION]    = "Interrupt Collections",
1719         [GITS_BASER_TYPE_RESERVED5]     = "Reserved (5)",
1720         [GITS_BASER_TYPE_RESERVED6]     = "Reserved (6)",
1721         [GITS_BASER_TYPE_RESERVED7]     = "Reserved (7)",
1722 };
1723
1724 static u64 its_read_baser(struct its_node *its, struct its_baser *baser)
1725 {
1726         u32 idx = baser - its->tables;
1727
1728         return gits_read_baser(its->base + GITS_BASER + (idx << 3));
1729 }
1730
1731 static void its_write_baser(struct its_node *its, struct its_baser *baser,
1732                             u64 val)
1733 {
1734         u32 idx = baser - its->tables;
1735
1736         gits_write_baser(val, its->base + GITS_BASER + (idx << 3));
1737         baser->val = its_read_baser(its, baser);
1738 }
1739
1740 static int its_setup_baser(struct its_node *its, struct its_baser *baser,
1741                            u64 cache, u64 shr, u32 psz, u32 order,
1742                            bool indirect)
1743 {
1744         u64 val = its_read_baser(its, baser);
1745         u64 esz = GITS_BASER_ENTRY_SIZE(val);
1746         u64 type = GITS_BASER_TYPE(val);
1747         u64 baser_phys, tmp;
1748         u32 alloc_pages;
1749         void *base;
1750
1751 retry_alloc_baser:
1752         alloc_pages = (PAGE_ORDER_TO_SIZE(order) / psz);
1753         if (alloc_pages > GITS_BASER_PAGES_MAX) {
1754                 pr_warn("ITS@%pa: %s too large, reduce ITS pages %u->%u\n",
1755                         &its->phys_base, its_base_type_string[type],
1756                         alloc_pages, GITS_BASER_PAGES_MAX);
1757                 alloc_pages = GITS_BASER_PAGES_MAX;
1758                 order = get_order(GITS_BASER_PAGES_MAX * psz);
1759         }
1760
1761         base = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, order);
1762         if (!base)
1763                 return -ENOMEM;
1764
1765         baser_phys = virt_to_phys(base);
1766
1767         /* Check if the physical address of the memory is above 48bits */
1768         if (IS_ENABLED(CONFIG_ARM64_64K_PAGES) && (baser_phys >> 48)) {
1769
1770                 /* 52bit PA is supported only when PageSize=64K */
1771                 if (psz != SZ_64K) {
1772                         pr_err("ITS: no 52bit PA support when psz=%d\n", psz);
1773                         free_pages((unsigned long)base, order);
1774                         return -ENXIO;
1775                 }
1776
1777                 /* Convert 52bit PA to 48bit field */
1778                 baser_phys = GITS_BASER_PHYS_52_to_48(baser_phys);
1779         }
1780
1781 retry_baser:
1782         val = (baser_phys                                        |
1783                 (type << GITS_BASER_TYPE_SHIFT)                  |
1784                 ((esz - 1) << GITS_BASER_ENTRY_SIZE_SHIFT)       |
1785                 ((alloc_pages - 1) << GITS_BASER_PAGES_SHIFT)    |
1786                 cache                                            |
1787                 shr                                              |
1788                 GITS_BASER_VALID);
1789
1790         val |=  indirect ? GITS_BASER_INDIRECT : 0x0;
1791
1792         switch (psz) {
1793         case SZ_4K:
1794                 val |= GITS_BASER_PAGE_SIZE_4K;
1795                 break;
1796         case SZ_16K:
1797                 val |= GITS_BASER_PAGE_SIZE_16K;
1798                 break;
1799         case SZ_64K:
1800                 val |= GITS_BASER_PAGE_SIZE_64K;
1801                 break;
1802         }
1803
1804         its_write_baser(its, baser, val);
1805         tmp = baser->val;
1806
1807         if ((val ^ tmp) & GITS_BASER_SHAREABILITY_MASK) {
1808                 /*
1809                  * Shareability didn't stick. Just use
1810                  * whatever the read reported, which is likely
1811                  * to be the only thing this redistributor
1812                  * supports. If that's zero, make it
1813                  * non-cacheable as well.
1814                  */
1815                 shr = tmp & GITS_BASER_SHAREABILITY_MASK;
1816                 if (!shr) {
1817                         cache = GITS_BASER_nC;
1818                         gic_flush_dcache_to_poc(base, PAGE_ORDER_TO_SIZE(order));
1819                 }
1820                 goto retry_baser;
1821         }
1822
1823         if ((val ^ tmp) & GITS_BASER_PAGE_SIZE_MASK) {
1824                 /*
1825                  * Page size didn't stick. Let's try a smaller
1826                  * size and retry. If we reach 4K, then
1827                  * something is horribly wrong...
1828                  */
1829                 free_pages((unsigned long)base, order);
1830                 baser->base = NULL;
1831
1832                 switch (psz) {
1833                 case SZ_16K:
1834                         psz = SZ_4K;
1835                         goto retry_alloc_baser;
1836                 case SZ_64K:
1837                         psz = SZ_16K;
1838                         goto retry_alloc_baser;
1839                 }
1840         }
1841
1842         if (val != tmp) {
1843                 pr_err("ITS@%pa: %s doesn't stick: %llx %llx\n",
1844                        &its->phys_base, its_base_type_string[type],
1845                        val, tmp);
1846                 free_pages((unsigned long)base, order);
1847                 return -ENXIO;
1848         }
1849
1850         baser->order = order;
1851         baser->base = base;
1852         baser->psz = psz;
1853         tmp = indirect ? GITS_LVL1_ENTRY_SIZE : esz;
1854
1855         pr_info("ITS@%pa: allocated %d %s @%lx (%s, esz %d, psz %dK, shr %d)\n",
1856                 &its->phys_base, (int)(PAGE_ORDER_TO_SIZE(order) / (int)tmp),
1857                 its_base_type_string[type],
1858                 (unsigned long)virt_to_phys(base),
1859                 indirect ? "indirect" : "flat", (int)esz,
1860                 psz / SZ_1K, (int)shr >> GITS_BASER_SHAREABILITY_SHIFT);
1861
1862         return 0;
1863 }
1864
1865 static bool its_parse_indirect_baser(struct its_node *its,
1866                                      struct its_baser *baser,
1867                                      u32 psz, u32 *order, u32 ids)
1868 {
1869         u64 tmp = its_read_baser(its, baser);
1870         u64 type = GITS_BASER_TYPE(tmp);
1871         u64 esz = GITS_BASER_ENTRY_SIZE(tmp);
1872         u64 val = GITS_BASER_InnerShareable | GITS_BASER_RaWaWb;
1873         u32 new_order = *order;
1874         bool indirect = false;
1875
1876         /* No need to enable Indirection if memory requirement < (psz*2)bytes */
1877         if ((esz << ids) > (psz * 2)) {
1878                 /*
1879                  * Find out whether hw supports a single or two-level table by
1880                  * table by reading bit at offset '62' after writing '1' to it.
1881                  */
1882                 its_write_baser(its, baser, val | GITS_BASER_INDIRECT);
1883                 indirect = !!(baser->val & GITS_BASER_INDIRECT);
1884
1885                 if (indirect) {
1886                         /*
1887                          * The size of the lvl2 table is equal to ITS page size
1888                          * which is 'psz'. For computing lvl1 table size,
1889                          * subtract ID bits that sparse lvl2 table from 'ids'
1890                          * which is reported by ITS hardware times lvl1 table
1891                          * entry size.
1892                          */
1893                         ids -= ilog2(psz / (int)esz);
1894                         esz = GITS_LVL1_ENTRY_SIZE;
1895                 }
1896         }
1897
1898         /*
1899          * Allocate as many entries as required to fit the
1900          * range of device IDs that the ITS can grok... The ID
1901          * space being incredibly sparse, this results in a
1902          * massive waste of memory if two-level device table
1903          * feature is not supported by hardware.
1904          */
1905         new_order = max_t(u32, get_order(esz << ids), new_order);
1906         if (new_order >= MAX_ORDER) {
1907                 new_order = MAX_ORDER - 1;
1908                 ids = ilog2(PAGE_ORDER_TO_SIZE(new_order) / (int)esz);
1909                 pr_warn("ITS@%pa: %s Table too large, reduce ids %u->%u\n",
1910                         &its->phys_base, its_base_type_string[type],
1911                         its->device_ids, ids);
1912         }
1913
1914         *order = new_order;
1915
1916         return indirect;
1917 }
1918
1919 static void its_free_tables(struct its_node *its)
1920 {
1921         int i;
1922
1923         for (i = 0; i < GITS_BASER_NR_REGS; i++) {
1924                 if (its->tables[i].base) {
1925                         free_pages((unsigned long)its->tables[i].base,
1926                                    its->tables[i].order);
1927                         its->tables[i].base = NULL;
1928                 }
1929         }
1930 }
1931
1932 static int its_alloc_tables(struct its_node *its)
1933 {
1934         u64 shr = GITS_BASER_InnerShareable;
1935         u64 cache = GITS_BASER_RaWaWb;
1936         u32 psz = SZ_64K;
1937         int err, i;
1938
1939         if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_22375)
1940                 /* erratum 24313: ignore memory access type */
1941                 cache = GITS_BASER_nCnB;
1942
1943         for (i = 0; i < GITS_BASER_NR_REGS; i++) {
1944                 struct its_baser *baser = its->tables + i;
1945                 u64 val = its_read_baser(its, baser);
1946                 u64 type = GITS_BASER_TYPE(val);
1947                 u32 order = get_order(psz);
1948                 bool indirect = false;
1949
1950                 switch (type) {
1951                 case GITS_BASER_TYPE_NONE:
1952                         continue;
1953
1954                 case GITS_BASER_TYPE_DEVICE:
1955                         indirect = its_parse_indirect_baser(its, baser,
1956                                                             psz, &order,
1957                                                             its->device_ids);
1958                 case GITS_BASER_TYPE_VCPU:
1959                         indirect = its_parse_indirect_baser(its, baser,
1960                                                             psz, &order,
1961                                                             ITS_MAX_VPEID_BITS);
1962                         break;
1963                 }
1964
1965                 err = its_setup_baser(its, baser, cache, shr, psz, order, indirect);
1966                 if (err < 0) {
1967                         its_free_tables(its);
1968                         return err;
1969                 }
1970
1971                 /* Update settings which will be used for next BASERn */
1972                 psz = baser->psz;
1973                 cache = baser->val & GITS_BASER_CACHEABILITY_MASK;
1974                 shr = baser->val & GITS_BASER_SHAREABILITY_MASK;
1975         }
1976
1977         return 0;
1978 }
1979
1980 static int its_alloc_collections(struct its_node *its)
1981 {
1982         int i;
1983
1984         its->collections = kcalloc(nr_cpu_ids, sizeof(*its->collections),
1985                                    GFP_KERNEL);
1986         if (!its->collections)
1987                 return -ENOMEM;
1988
1989         for (i = 0; i < nr_cpu_ids; i++)
1990                 its->collections[i].target_address = ~0ULL;
1991
1992         return 0;
1993 }
1994
1995 static struct page *its_allocate_pending_table(gfp_t gfp_flags)
1996 {
1997         struct page *pend_page;
1998
1999         pend_page = alloc_pages(gfp_flags | __GFP_ZERO,
2000                                 get_order(LPI_PENDBASE_SZ));
2001         if (!pend_page)
2002                 return NULL;
2003
2004         /* Make sure the GIC will observe the zero-ed page */
2005         gic_flush_dcache_to_poc(page_address(pend_page), LPI_PENDBASE_SZ);
2006
2007         return pend_page;
2008 }
2009
2010 static void its_free_pending_table(struct page *pt)
2011 {
2012         free_pages((unsigned long)page_address(pt), get_order(LPI_PENDBASE_SZ));
2013 }
2014
2015 /*
2016  * Booting with kdump and LPIs enabled is generally fine. Any other
2017  * case is wrong in the absence of firmware/EFI support.
2018  */
2019 static bool enabled_lpis_allowed(void)
2020 {
2021         phys_addr_t addr;
2022         u64 val;
2023
2024         /* Check whether the property table is in a reserved region */
2025         val = gicr_read_propbaser(gic_data_rdist_rd_base() + GICR_PROPBASER);
2026         addr = val & GENMASK_ULL(51, 12);
2027
2028         return gic_check_reserved_range(addr, LPI_PROPBASE_SZ);
2029 }
2030
2031 static int __init allocate_lpi_tables(void)
2032 {
2033         u64 val;
2034         int err, cpu;
2035
2036         /*
2037          * If LPIs are enabled while we run this from the boot CPU,
2038          * flag the RD tables as pre-allocated if the stars do align.
2039          */
2040         val = readl_relaxed(gic_data_rdist_rd_base() + GICR_CTLR);
2041         if ((val & GICR_CTLR_ENABLE_LPIS) && enabled_lpis_allowed()) {
2042                 gic_rdists->flags |= (RDIST_FLAGS_RD_TABLES_PREALLOCATED |
2043                                       RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING);
2044                 pr_info("GICv3: Using preallocated redistributor tables\n");
2045         }
2046
2047         err = its_setup_lpi_prop_table();
2048         if (err)
2049                 return err;
2050
2051         /*
2052          * We allocate all the pending tables anyway, as we may have a
2053          * mix of RDs that have had LPIs enabled, and some that
2054          * don't. We'll free the unused ones as each CPU comes online.
2055          */
2056         for_each_possible_cpu(cpu) {
2057                 struct page *pend_page;
2058
2059                 pend_page = its_allocate_pending_table(GFP_NOWAIT);
2060                 if (!pend_page) {
2061                         pr_err("Failed to allocate PENDBASE for CPU%d\n", cpu);
2062                         return -ENOMEM;
2063                 }
2064
2065                 gic_data_rdist_cpu(cpu)->pend_page = pend_page;
2066         }
2067
2068         return 0;
2069 }
2070
2071 static u64 its_clear_vpend_valid(void __iomem *vlpi_base)
2072 {
2073         u32 count = 1000000;    /* 1s! */
2074         bool clean;
2075         u64 val;
2076
2077         val = gits_read_vpendbaser(vlpi_base + GICR_VPENDBASER);
2078         val &= ~GICR_VPENDBASER_Valid;
2079         gits_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
2080
2081         do {
2082                 val = gits_read_vpendbaser(vlpi_base + GICR_VPENDBASER);
2083                 clean = !(val & GICR_VPENDBASER_Dirty);
2084                 if (!clean) {
2085                         count--;
2086                         cpu_relax();
2087                         udelay(1);
2088                 }
2089         } while (!clean && count);
2090
2091         return val;
2092 }
2093
2094 static void its_cpu_init_lpis(void)
2095 {
2096         void __iomem *rbase = gic_data_rdist_rd_base();
2097         struct page *pend_page;
2098         phys_addr_t paddr;
2099         u64 val, tmp;
2100
2101         if (gic_data_rdist()->lpi_enabled)
2102                 return;
2103
2104         val = readl_relaxed(rbase + GICR_CTLR);
2105         if ((gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED) &&
2106             (val & GICR_CTLR_ENABLE_LPIS)) {
2107                 /*
2108                  * Check that we get the same property table on all
2109                  * RDs. If we don't, this is hopeless.
2110                  */
2111                 paddr = gicr_read_propbaser(rbase + GICR_PROPBASER);
2112                 paddr &= GENMASK_ULL(51, 12);
2113                 if (WARN_ON(gic_rdists->prop_table_pa != paddr))
2114                         add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
2115
2116                 paddr = gicr_read_pendbaser(rbase + GICR_PENDBASER);
2117                 paddr &= GENMASK_ULL(51, 16);
2118
2119                 WARN_ON(!gic_check_reserved_range(paddr, LPI_PENDBASE_SZ));
2120                 its_free_pending_table(gic_data_rdist()->pend_page);
2121                 gic_data_rdist()->pend_page = NULL;
2122
2123                 goto out;
2124         }
2125
2126         pend_page = gic_data_rdist()->pend_page;
2127         paddr = page_to_phys(pend_page);
2128         WARN_ON(gic_reserve_range(paddr, LPI_PENDBASE_SZ));
2129
2130         /* set PROPBASE */
2131         val = (gic_rdists->prop_table_pa |
2132                GICR_PROPBASER_InnerShareable |
2133                GICR_PROPBASER_RaWaWb |
2134                ((LPI_NRBITS - 1) & GICR_PROPBASER_IDBITS_MASK));
2135
2136         gicr_write_propbaser(val, rbase + GICR_PROPBASER);
2137         tmp = gicr_read_propbaser(rbase + GICR_PROPBASER);
2138
2139         if ((tmp ^ val) & GICR_PROPBASER_SHAREABILITY_MASK) {
2140                 if (!(tmp & GICR_PROPBASER_SHAREABILITY_MASK)) {
2141                         /*
2142                          * The HW reports non-shareable, we must
2143                          * remove the cacheability attributes as
2144                          * well.
2145                          */
2146                         val &= ~(GICR_PROPBASER_SHAREABILITY_MASK |
2147                                  GICR_PROPBASER_CACHEABILITY_MASK);
2148                         val |= GICR_PROPBASER_nC;
2149                         gicr_write_propbaser(val, rbase + GICR_PROPBASER);
2150                 }
2151                 pr_info_once("GIC: using cache flushing for LPI property table\n");
2152                 gic_rdists->flags |= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING;
2153         }
2154
2155         /* set PENDBASE */
2156         val = (page_to_phys(pend_page) |
2157                GICR_PENDBASER_InnerShareable |
2158                GICR_PENDBASER_RaWaWb);
2159
2160         gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
2161         tmp = gicr_read_pendbaser(rbase + GICR_PENDBASER);
2162
2163         if (!(tmp & GICR_PENDBASER_SHAREABILITY_MASK)) {
2164                 /*
2165                  * The HW reports non-shareable, we must remove the
2166                  * cacheability attributes as well.
2167                  */
2168                 val &= ~(GICR_PENDBASER_SHAREABILITY_MASK |
2169                          GICR_PENDBASER_CACHEABILITY_MASK);
2170                 val |= GICR_PENDBASER_nC;
2171                 gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
2172         }
2173
2174         /* Enable LPIs */
2175         val = readl_relaxed(rbase + GICR_CTLR);
2176         val |= GICR_CTLR_ENABLE_LPIS;
2177         writel_relaxed(val, rbase + GICR_CTLR);
2178
2179         if (gic_rdists->has_vlpis) {
2180                 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2181
2182                 /*
2183                  * It's possible for CPU to receive VLPIs before it is
2184                  * sheduled as a vPE, especially for the first CPU, and the
2185                  * VLPI with INTID larger than 2^(IDbits+1) will be considered
2186                  * as out of range and dropped by GIC.
2187                  * So we initialize IDbits to known value to avoid VLPI drop.
2188                  */
2189                 val = (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
2190                 pr_debug("GICv4: CPU%d: Init IDbits to 0x%llx for GICR_VPROPBASER\n",
2191                         smp_processor_id(), val);
2192                 gits_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2193
2194                 /*
2195                  * Also clear Valid bit of GICR_VPENDBASER, in case some
2196                  * ancient programming gets left in and has possibility of
2197                  * corrupting memory.
2198                  */
2199                 val = its_clear_vpend_valid(vlpi_base);
2200                 WARN_ON(val & GICR_VPENDBASER_Dirty);
2201         }
2202
2203         /* Make sure the GIC has seen the above */
2204         dsb(sy);
2205 out:
2206         gic_data_rdist()->lpi_enabled = true;
2207         pr_info("GICv3: CPU%d: using %s LPI pending table @%pa\n",
2208                 smp_processor_id(),
2209                 gic_data_rdist()->pend_page ? "allocated" : "reserved",
2210                 &paddr);
2211 }
2212
2213 static void its_cpu_init_collection(struct its_node *its)
2214 {
2215         int cpu = smp_processor_id();
2216         u64 target;
2217
2218         /* avoid cross node collections and its mapping */
2219         if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) {
2220                 struct device_node *cpu_node;
2221
2222                 cpu_node = of_get_cpu_node(cpu, NULL);
2223                 if (its->numa_node != NUMA_NO_NODE &&
2224                         its->numa_node != of_node_to_nid(cpu_node))
2225                         return;
2226         }
2227
2228         /*
2229          * We now have to bind each collection to its target
2230          * redistributor.
2231          */
2232         if (gic_read_typer(its->base + GITS_TYPER) & GITS_TYPER_PTA) {
2233                 /*
2234                  * This ITS wants the physical address of the
2235                  * redistributor.
2236                  */
2237                 target = gic_data_rdist()->phys_base;
2238         } else {
2239                 /* This ITS wants a linear CPU number. */
2240                 target = gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER);
2241                 target = GICR_TYPER_CPU_NUMBER(target) << 16;
2242         }
2243
2244         /* Perform collection mapping */
2245         its->collections[cpu].target_address = target;
2246         its->collections[cpu].col_id = cpu;
2247
2248         its_send_mapc(its, &its->collections[cpu], 1);
2249         its_send_invall(its, &its->collections[cpu]);
2250 }
2251
2252 static void its_cpu_init_collections(void)
2253 {
2254         struct its_node *its;
2255
2256         raw_spin_lock(&its_lock);
2257
2258         list_for_each_entry(its, &its_nodes, entry)
2259                 its_cpu_init_collection(its);
2260
2261         raw_spin_unlock(&its_lock);
2262 }
2263
2264 static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
2265 {
2266         struct its_device *its_dev = NULL, *tmp;
2267         unsigned long flags;
2268
2269         raw_spin_lock_irqsave(&its->lock, flags);
2270
2271         list_for_each_entry(tmp, &its->its_device_list, entry) {
2272                 if (tmp->device_id == dev_id) {
2273                         its_dev = tmp;
2274                         break;
2275                 }
2276         }
2277
2278         raw_spin_unlock_irqrestore(&its->lock, flags);
2279
2280         return its_dev;
2281 }
2282
2283 static struct its_baser *its_get_baser(struct its_node *its, u32 type)
2284 {
2285         int i;
2286
2287         for (i = 0; i < GITS_BASER_NR_REGS; i++) {
2288                 if (GITS_BASER_TYPE(its->tables[i].val) == type)
2289                         return &its->tables[i];
2290         }
2291
2292         return NULL;
2293 }
2294
2295 static bool its_alloc_table_entry(struct its_baser *baser, u32 id)
2296 {
2297         struct page *page;
2298         u32 esz, idx;
2299         __le64 *table;
2300
2301         /* Don't allow device id that exceeds single, flat table limit */
2302         esz = GITS_BASER_ENTRY_SIZE(baser->val);
2303         if (!(baser->val & GITS_BASER_INDIRECT))
2304                 return (id < (PAGE_ORDER_TO_SIZE(baser->order) / esz));
2305
2306         /* Compute 1st level table index & check if that exceeds table limit */
2307         idx = id >> ilog2(baser->psz / esz);
2308         if (idx >= (PAGE_ORDER_TO_SIZE(baser->order) / GITS_LVL1_ENTRY_SIZE))
2309                 return false;
2310
2311         table = baser->base;
2312
2313         /* Allocate memory for 2nd level table */
2314         if (!table[idx]) {
2315                 page = alloc_pages(GFP_KERNEL | __GFP_ZERO, get_order(baser->psz));
2316                 if (!page)
2317                         return false;
2318
2319                 /* Flush Lvl2 table to PoC if hw doesn't support coherency */
2320                 if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
2321                         gic_flush_dcache_to_poc(page_address(page), baser->psz);
2322
2323                 table[idx] = cpu_to_le64(page_to_phys(page) | GITS_BASER_VALID);
2324
2325                 /* Flush Lvl1 entry to PoC if hw doesn't support coherency */
2326                 if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
2327                         gic_flush_dcache_to_poc(table + idx, GITS_LVL1_ENTRY_SIZE);
2328
2329                 /* Ensure updated table contents are visible to ITS hardware */
2330                 dsb(sy);
2331         }
2332
2333         return true;
2334 }
2335
2336 static bool its_alloc_device_table(struct its_node *its, u32 dev_id)
2337 {
2338         struct its_baser *baser;
2339
2340         baser = its_get_baser(its, GITS_BASER_TYPE_DEVICE);
2341
2342         /* Don't allow device id that exceeds ITS hardware limit */
2343         if (!baser)
2344                 return (ilog2(dev_id) < its->device_ids);
2345
2346         return its_alloc_table_entry(baser, dev_id);
2347 }
2348
2349 static bool its_alloc_vpe_table(u32 vpe_id)
2350 {
2351         struct its_node *its;
2352
2353         /*
2354          * Make sure the L2 tables are allocated on *all* v4 ITSs. We
2355          * could try and only do it on ITSs corresponding to devices
2356          * that have interrupts targeted at this VPE, but the
2357          * complexity becomes crazy (and you have tons of memory
2358          * anyway, right?).
2359          */
2360         list_for_each_entry(its, &its_nodes, entry) {
2361                 struct its_baser *baser;
2362
2363                 if (!its->is_v4)
2364                         continue;
2365
2366                 baser = its_get_baser(its, GITS_BASER_TYPE_VCPU);
2367                 if (!baser)
2368                         return false;
2369
2370                 if (!its_alloc_table_entry(baser, vpe_id))
2371                         return false;
2372         }
2373
2374         return true;
2375 }
2376
2377 static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
2378                                             int nvecs, bool alloc_lpis)
2379 {
2380         struct its_device *dev;
2381         unsigned long *lpi_map = NULL;
2382         unsigned long flags;
2383         u16 *col_map = NULL;
2384         void *itt;
2385         int lpi_base;
2386         int nr_lpis;
2387         int nr_ites;
2388         int sz;
2389
2390         if (!its_alloc_device_table(its, dev_id))
2391                 return NULL;
2392
2393         if (WARN_ON(!is_power_of_2(nvecs)))
2394                 nvecs = roundup_pow_of_two(nvecs);
2395
2396         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2397         /*
2398          * Even if the device wants a single LPI, the ITT must be
2399          * sized as a power of two (and you need at least one bit...).
2400          */
2401         nr_ites = max(2, nvecs);
2402         sz = nr_ites * its->ite_size;
2403         sz = max(sz, ITS_ITT_ALIGN) + ITS_ITT_ALIGN - 1;
2404         itt = kzalloc(sz, GFP_KERNEL);
2405         if (alloc_lpis) {
2406                 lpi_map = its_lpi_alloc(nvecs, &lpi_base, &nr_lpis);
2407                 if (lpi_map)
2408                         col_map = kcalloc(nr_lpis, sizeof(*col_map),
2409                                           GFP_KERNEL);
2410         } else {
2411                 col_map = kcalloc(nr_ites, sizeof(*col_map), GFP_KERNEL);
2412                 nr_lpis = 0;
2413                 lpi_base = 0;
2414         }
2415
2416         if (!dev || !itt ||  !col_map || (!lpi_map && alloc_lpis)) {
2417                 kfree(dev);
2418                 kfree(itt);
2419                 kfree(lpi_map);
2420                 kfree(col_map);
2421                 return NULL;
2422         }
2423
2424         gic_flush_dcache_to_poc(itt, sz);
2425
2426         dev->its = its;
2427         dev->itt = itt;
2428         dev->nr_ites = nr_ites;
2429         dev->event_map.lpi_map = lpi_map;
2430         dev->event_map.col_map = col_map;
2431         dev->event_map.lpi_base = lpi_base;
2432         dev->event_map.nr_lpis = nr_lpis;
2433         mutex_init(&dev->event_map.vlpi_lock);
2434         dev->device_id = dev_id;
2435         INIT_LIST_HEAD(&dev->entry);
2436
2437         raw_spin_lock_irqsave(&its->lock, flags);
2438         list_add(&dev->entry, &its->its_device_list);
2439         raw_spin_unlock_irqrestore(&its->lock, flags);
2440
2441         /* Map device to its ITT */
2442         its_send_mapd(dev, 1);
2443
2444         return dev;
2445 }
2446
2447 static void its_free_device(struct its_device *its_dev)
2448 {
2449         unsigned long flags;
2450
2451         raw_spin_lock_irqsave(&its_dev->its->lock, flags);
2452         list_del(&its_dev->entry);
2453         raw_spin_unlock_irqrestore(&its_dev->its->lock, flags);
2454         kfree(its_dev->itt);
2455         kfree(its_dev);
2456 }
2457
2458 static int its_alloc_device_irq(struct its_device *dev, int nvecs, irq_hw_number_t *hwirq)
2459 {
2460         int idx;
2461
2462         idx = bitmap_find_free_region(dev->event_map.lpi_map,
2463                                       dev->event_map.nr_lpis,
2464                                       get_count_order(nvecs));
2465         if (idx < 0)
2466                 return -ENOSPC;
2467
2468         *hwirq = dev->event_map.lpi_base + idx;
2469         set_bit(idx, dev->event_map.lpi_map);
2470
2471         return 0;
2472 }
2473
2474 static int its_msi_prepare(struct irq_domain *domain, struct device *dev,
2475                            int nvec, msi_alloc_info_t *info)
2476 {
2477         struct its_node *its;
2478         struct its_device *its_dev;
2479         struct msi_domain_info *msi_info;
2480         u32 dev_id;
2481         int err = 0;
2482
2483         /*
2484          * We ignore "dev" entierely, and rely on the dev_id that has
2485          * been passed via the scratchpad. This limits this domain's
2486          * usefulness to upper layers that definitely know that they
2487          * are built on top of the ITS.
2488          */
2489         dev_id = info->scratchpad[0].ul;
2490
2491         msi_info = msi_get_domain_info(domain);
2492         its = msi_info->data;
2493
2494         if (!gic_rdists->has_direct_lpi &&
2495             vpe_proxy.dev &&
2496             vpe_proxy.dev->its == its &&
2497             dev_id == vpe_proxy.dev->device_id) {
2498                 /* Bad luck. Get yourself a better implementation */
2499                 WARN_ONCE(1, "DevId %x clashes with GICv4 VPE proxy device\n",
2500                           dev_id);
2501                 return -EINVAL;
2502         }
2503
2504         mutex_lock(&its->dev_alloc_lock);
2505         its_dev = its_find_device(its, dev_id);
2506         if (its_dev) {
2507                 /*
2508                  * We already have seen this ID, probably through
2509                  * another alias (PCI bridge of some sort). No need to
2510                  * create the device.
2511                  */
2512                 its_dev->shared = true;
2513                 pr_debug("Reusing ITT for devID %x\n", dev_id);
2514                 goto out;
2515         }
2516
2517         its_dev = its_create_device(its, dev_id, nvec, true);
2518         if (!its_dev) {
2519                 err = -ENOMEM;
2520                 goto out;
2521         }
2522
2523         pr_debug("ITT %d entries, %d bits\n", nvec, ilog2(nvec));
2524 out:
2525         mutex_unlock(&its->dev_alloc_lock);
2526         info->scratchpad[0].ptr = its_dev;
2527         return err;
2528 }
2529
2530 static struct msi_domain_ops its_msi_domain_ops = {
2531         .msi_prepare    = its_msi_prepare,
2532 };
2533
2534 static int its_irq_gic_domain_alloc(struct irq_domain *domain,
2535                                     unsigned int virq,
2536                                     irq_hw_number_t hwirq)
2537 {
2538         struct irq_fwspec fwspec;
2539
2540         if (irq_domain_get_of_node(domain->parent)) {
2541                 fwspec.fwnode = domain->parent->fwnode;
2542                 fwspec.param_count = 3;
2543                 fwspec.param[0] = GIC_IRQ_TYPE_LPI;
2544                 fwspec.param[1] = hwirq;
2545                 fwspec.param[2] = IRQ_TYPE_EDGE_RISING;
2546         } else if (is_fwnode_irqchip(domain->parent->fwnode)) {
2547                 fwspec.fwnode = domain->parent->fwnode;
2548                 fwspec.param_count = 2;
2549                 fwspec.param[0] = hwirq;
2550                 fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
2551         } else {
2552                 return -EINVAL;
2553         }
2554
2555         return irq_domain_alloc_irqs_parent(domain, virq, 1, &fwspec);
2556 }
2557
2558 static int its_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
2559                                 unsigned int nr_irqs, void *args)
2560 {
2561         msi_alloc_info_t *info = args;
2562         struct its_device *its_dev = info->scratchpad[0].ptr;
2563         irq_hw_number_t hwirq;
2564         int err;
2565         int i;
2566
2567         err = its_alloc_device_irq(its_dev, nr_irqs, &hwirq);
2568         if (err)
2569                 return err;
2570
2571         for (i = 0; i < nr_irqs; i++) {
2572                 err = its_irq_gic_domain_alloc(domain, virq + i, hwirq + i);
2573                 if (err)
2574                         return err;
2575
2576                 irq_domain_set_hwirq_and_chip(domain, virq + i,
2577                                               hwirq + i, &its_irq_chip, its_dev);
2578                 irqd_set_single_target(irq_desc_get_irq_data(irq_to_desc(virq + i)));
2579                 pr_debug("ID:%d pID:%d vID:%d\n",
2580                          (int)(hwirq + i - its_dev->event_map.lpi_base),
2581                          (int)(hwirq + i), virq + i);
2582         }
2583
2584         return 0;
2585 }
2586
2587 static int its_irq_domain_activate(struct irq_domain *domain,
2588                                    struct irq_data *d, bool reserve)
2589 {
2590         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2591         u32 event = its_get_event_id(d);
2592         const struct cpumask *cpu_mask = cpu_online_mask;
2593         int cpu;
2594
2595         /* get the cpu_mask of local node */
2596         if (its_dev->its->numa_node >= 0)
2597                 cpu_mask = cpumask_of_node(its_dev->its->numa_node);
2598
2599         /* Bind the LPI to the first possible CPU */
2600         cpu = cpumask_first_and(cpu_mask, cpu_online_mask);
2601         if (cpu >= nr_cpu_ids) {
2602                 if (its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144)
2603                         return -EINVAL;
2604
2605                 cpu = cpumask_first(cpu_online_mask);
2606         }
2607
2608         its_dev->event_map.col_map[event] = cpu;
2609         irq_data_update_effective_affinity(d, cpumask_of(cpu));
2610
2611         /* Map the GIC IRQ and event to the device */
2612         its_send_mapti(its_dev, d->hwirq, event);
2613         return 0;
2614 }
2615
2616 static void its_irq_domain_deactivate(struct irq_domain *domain,
2617                                       struct irq_data *d)
2618 {
2619         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2620         u32 event = its_get_event_id(d);
2621
2622         /* Stop the delivery of interrupts */
2623         its_send_discard(its_dev, event);
2624 }
2625
2626 static void its_irq_domain_free(struct irq_domain *domain, unsigned int virq,
2627                                 unsigned int nr_irqs)
2628 {
2629         struct irq_data *d = irq_domain_get_irq_data(domain, virq);
2630         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2631         struct its_node *its = its_dev->its;
2632         int i;
2633
2634         for (i = 0; i < nr_irqs; i++) {
2635                 struct irq_data *data = irq_domain_get_irq_data(domain,
2636                                                                 virq + i);
2637                 u32 event = its_get_event_id(data);
2638
2639                 /* Mark interrupt index as unused */
2640                 clear_bit(event, its_dev->event_map.lpi_map);
2641
2642                 /* Nuke the entry in the domain */
2643                 irq_domain_reset_irq_data(data);
2644         }
2645
2646         mutex_lock(&its->dev_alloc_lock);
2647
2648         /*
2649          * If all interrupts have been freed, start mopping the
2650          * floor. This is conditionned on the device not being shared.
2651          */
2652         if (!its_dev->shared &&
2653             bitmap_empty(its_dev->event_map.lpi_map,
2654                          its_dev->event_map.nr_lpis)) {
2655                 its_lpi_free(its_dev->event_map.lpi_map,
2656                              its_dev->event_map.lpi_base,
2657                              its_dev->event_map.nr_lpis);
2658                 kfree(its_dev->event_map.col_map);
2659
2660                 /* Unmap device/itt */
2661                 its_send_mapd(its_dev, 0);
2662                 its_free_device(its_dev);
2663         }
2664
2665         mutex_unlock(&its->dev_alloc_lock);
2666
2667         irq_domain_free_irqs_parent(domain, virq, nr_irqs);
2668 }
2669
2670 static const struct irq_domain_ops its_domain_ops = {
2671         .alloc                  = its_irq_domain_alloc,
2672         .free                   = its_irq_domain_free,
2673         .activate               = its_irq_domain_activate,
2674         .deactivate             = its_irq_domain_deactivate,
2675 };
2676
2677 /*
2678  * This is insane.
2679  *
2680  * If a GICv4 doesn't implement Direct LPIs (which is extremely
2681  * likely), the only way to perform an invalidate is to use a fake
2682  * device to issue an INV command, implying that the LPI has first
2683  * been mapped to some event on that device. Since this is not exactly
2684  * cheap, we try to keep that mapping around as long as possible, and
2685  * only issue an UNMAP if we're short on available slots.
2686  *
2687  * Broken by design(tm).
2688  */
2689 static void its_vpe_db_proxy_unmap_locked(struct its_vpe *vpe)
2690 {
2691         /* Already unmapped? */
2692         if (vpe->vpe_proxy_event == -1)
2693                 return;
2694
2695         its_send_discard(vpe_proxy.dev, vpe->vpe_proxy_event);
2696         vpe_proxy.vpes[vpe->vpe_proxy_event] = NULL;
2697
2698         /*
2699          * We don't track empty slots at all, so let's move the
2700          * next_victim pointer if we can quickly reuse that slot
2701          * instead of nuking an existing entry. Not clear that this is
2702          * always a win though, and this might just generate a ripple
2703          * effect... Let's just hope VPEs don't migrate too often.
2704          */
2705         if (vpe_proxy.vpes[vpe_proxy.next_victim])
2706                 vpe_proxy.next_victim = vpe->vpe_proxy_event;
2707
2708         vpe->vpe_proxy_event = -1;
2709 }
2710
2711 static void its_vpe_db_proxy_unmap(struct its_vpe *vpe)
2712 {
2713         if (!gic_rdists->has_direct_lpi) {
2714                 unsigned long flags;
2715
2716                 raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2717                 its_vpe_db_proxy_unmap_locked(vpe);
2718                 raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2719         }
2720 }
2721
2722 static void its_vpe_db_proxy_map_locked(struct its_vpe *vpe)
2723 {
2724         /* Already mapped? */
2725         if (vpe->vpe_proxy_event != -1)
2726                 return;
2727
2728         /* This slot was already allocated. Kick the other VPE out. */
2729         if (vpe_proxy.vpes[vpe_proxy.next_victim])
2730                 its_vpe_db_proxy_unmap_locked(vpe_proxy.vpes[vpe_proxy.next_victim]);
2731
2732         /* Map the new VPE instead */
2733         vpe_proxy.vpes[vpe_proxy.next_victim] = vpe;
2734         vpe->vpe_proxy_event = vpe_proxy.next_victim;
2735         vpe_proxy.next_victim = (vpe_proxy.next_victim + 1) % vpe_proxy.dev->nr_ites;
2736
2737         vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = vpe->col_idx;
2738         its_send_mapti(vpe_proxy.dev, vpe->vpe_db_lpi, vpe->vpe_proxy_event);
2739 }
2740
2741 static void its_vpe_db_proxy_move(struct its_vpe *vpe, int from, int to)
2742 {
2743         unsigned long flags;
2744         struct its_collection *target_col;
2745
2746         if (gic_rdists->has_direct_lpi) {
2747                 void __iomem *rdbase;
2748
2749                 rdbase = per_cpu_ptr(gic_rdists->rdist, from)->rd_base;
2750                 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
2751                 while (gic_read_lpir(rdbase + GICR_SYNCR) & 1)
2752                         cpu_relax();
2753
2754                 return;
2755         }
2756
2757         raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2758
2759         its_vpe_db_proxy_map_locked(vpe);
2760
2761         target_col = &vpe_proxy.dev->its->collections[to];
2762         its_send_movi(vpe_proxy.dev, target_col, vpe->vpe_proxy_event);
2763         vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = to;
2764
2765         raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2766 }
2767
2768 static int its_vpe_set_affinity(struct irq_data *d,
2769                                 const struct cpumask *mask_val,
2770                                 bool force)
2771 {
2772         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2773         int cpu = cpumask_first(mask_val);
2774
2775         /*
2776          * Changing affinity is mega expensive, so let's be as lazy as
2777          * we can and only do it if we really have to. Also, if mapped
2778          * into the proxy device, we need to move the doorbell
2779          * interrupt to its new location.
2780          */
2781         if (vpe->col_idx != cpu) {
2782                 int from = vpe->col_idx;
2783
2784                 vpe->col_idx = cpu;
2785                 its_send_vmovp(vpe);
2786                 its_vpe_db_proxy_move(vpe, from, cpu);
2787         }
2788
2789         irq_data_update_effective_affinity(d, cpumask_of(cpu));
2790
2791         return IRQ_SET_MASK_OK_DONE;
2792 }
2793
2794 static void its_vpe_schedule(struct its_vpe *vpe)
2795 {
2796         void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2797         u64 val;
2798
2799         /* Schedule the VPE */
2800         val  = virt_to_phys(page_address(vpe->its_vm->vprop_page)) &
2801                 GENMASK_ULL(51, 12);
2802         val |= (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
2803         val |= GICR_VPROPBASER_RaWb;
2804         val |= GICR_VPROPBASER_InnerShareable;
2805         gits_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2806
2807         val  = virt_to_phys(page_address(vpe->vpt_page)) &
2808                 GENMASK_ULL(51, 16);
2809         val |= GICR_VPENDBASER_RaWaWb;
2810         val |= GICR_VPENDBASER_NonShareable;
2811         /*
2812          * There is no good way of finding out if the pending table is
2813          * empty as we can race against the doorbell interrupt very
2814          * easily. So in the end, vpe->pending_last is only an
2815          * indication that the vcpu has something pending, not one
2816          * that the pending table is empty. A good implementation
2817          * would be able to read its coarse map pretty quickly anyway,
2818          * making this a tolerable issue.
2819          */
2820         val |= GICR_VPENDBASER_PendingLast;
2821         val |= vpe->idai ? GICR_VPENDBASER_IDAI : 0;
2822         val |= GICR_VPENDBASER_Valid;
2823         gits_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
2824 }
2825
2826 static void its_vpe_deschedule(struct its_vpe *vpe)
2827 {
2828         void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2829         u64 val;
2830
2831         val = its_clear_vpend_valid(vlpi_base);
2832
2833         if (unlikely(val & GICR_VPENDBASER_Dirty)) {
2834                 pr_err_ratelimited("ITS virtual pending table not cleaning\n");
2835                 vpe->idai = false;
2836                 vpe->pending_last = true;
2837         } else {
2838                 vpe->idai = !!(val & GICR_VPENDBASER_IDAI);
2839                 vpe->pending_last = !!(val & GICR_VPENDBASER_PendingLast);
2840         }
2841 }
2842
2843 static void its_vpe_invall(struct its_vpe *vpe)
2844 {
2845         struct its_node *its;
2846
2847         list_for_each_entry(its, &its_nodes, entry) {
2848                 if (!its->is_v4)
2849                         continue;
2850
2851                 if (its_list_map && !vpe->its_vm->vlpi_count[its->list_nr])
2852                         continue;
2853
2854                 /*
2855                  * Sending a VINVALL to a single ITS is enough, as all
2856                  * we need is to reach the redistributors.
2857                  */
2858                 its_send_vinvall(its, vpe);
2859                 return;
2860         }
2861 }
2862
2863 static int its_vpe_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
2864 {
2865         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2866         struct its_cmd_info *info = vcpu_info;
2867
2868         switch (info->cmd_type) {
2869         case SCHEDULE_VPE:
2870                 its_vpe_schedule(vpe);
2871                 return 0;
2872
2873         case DESCHEDULE_VPE:
2874                 its_vpe_deschedule(vpe);
2875                 return 0;
2876
2877         case INVALL_VPE:
2878                 its_vpe_invall(vpe);
2879                 return 0;
2880
2881         default:
2882                 return -EINVAL;
2883         }
2884 }
2885
2886 static void its_vpe_send_cmd(struct its_vpe *vpe,
2887                              void (*cmd)(struct its_device *, u32))
2888 {
2889         unsigned long flags;
2890
2891         raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2892
2893         its_vpe_db_proxy_map_locked(vpe);
2894         cmd(vpe_proxy.dev, vpe->vpe_proxy_event);
2895
2896         raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2897 }
2898
2899 static void its_vpe_send_inv(struct irq_data *d)
2900 {
2901         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2902
2903         if (gic_rdists->has_direct_lpi) {
2904                 void __iomem *rdbase;
2905
2906                 rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
2907                 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_INVLPIR);
2908                 while (gic_read_lpir(rdbase + GICR_SYNCR) & 1)
2909                         cpu_relax();
2910         } else {
2911                 its_vpe_send_cmd(vpe, its_send_inv);
2912         }
2913 }
2914
2915 static void its_vpe_mask_irq(struct irq_data *d)
2916 {
2917         /*
2918          * We need to unmask the LPI, which is described by the parent
2919          * irq_data. Instead of calling into the parent (which won't
2920          * exactly do the right thing, let's simply use the
2921          * parent_data pointer. Yes, I'm naughty.
2922          */
2923         lpi_write_config(d->parent_data, LPI_PROP_ENABLED, 0);
2924         its_vpe_send_inv(d);
2925 }
2926
2927 static void its_vpe_unmask_irq(struct irq_data *d)
2928 {
2929         /* Same hack as above... */
2930         lpi_write_config(d->parent_data, 0, LPI_PROP_ENABLED);
2931         its_vpe_send_inv(d);
2932 }
2933
2934 static int its_vpe_set_irqchip_state(struct irq_data *d,
2935                                      enum irqchip_irq_state which,
2936                                      bool state)
2937 {
2938         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2939
2940         if (which != IRQCHIP_STATE_PENDING)
2941                 return -EINVAL;
2942
2943         if (gic_rdists->has_direct_lpi) {
2944                 void __iomem *rdbase;
2945
2946                 rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
2947                 if (state) {
2948                         gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_SETLPIR);
2949                 } else {
2950                         gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
2951                         while (gic_read_lpir(rdbase + GICR_SYNCR) & 1)
2952                                 cpu_relax();
2953                 }
2954         } else {
2955                 if (state)
2956                         its_vpe_send_cmd(vpe, its_send_int);
2957                 else
2958                         its_vpe_send_cmd(vpe, its_send_clear);
2959         }
2960
2961         return 0;
2962 }
2963
2964 static struct irq_chip its_vpe_irq_chip = {
2965         .name                   = "GICv4-vpe",
2966         .irq_mask               = its_vpe_mask_irq,
2967         .irq_unmask             = its_vpe_unmask_irq,
2968         .irq_eoi                = irq_chip_eoi_parent,
2969         .irq_set_affinity       = its_vpe_set_affinity,
2970         .irq_set_irqchip_state  = its_vpe_set_irqchip_state,
2971         .irq_set_vcpu_affinity  = its_vpe_set_vcpu_affinity,
2972 };
2973
2974 static int its_vpe_id_alloc(void)
2975 {
2976         return ida_simple_get(&its_vpeid_ida, 0, ITS_MAX_VPEID, GFP_KERNEL);
2977 }
2978
2979 static void its_vpe_id_free(u16 id)
2980 {
2981         ida_simple_remove(&its_vpeid_ida, id);
2982 }
2983
2984 static int its_vpe_init(struct its_vpe *vpe)
2985 {
2986         struct page *vpt_page;
2987         int vpe_id;
2988
2989         /* Allocate vpe_id */
2990         vpe_id = its_vpe_id_alloc();
2991         if (vpe_id < 0)
2992                 return vpe_id;
2993
2994         /* Allocate VPT */
2995         vpt_page = its_allocate_pending_table(GFP_KERNEL);
2996         if (!vpt_page) {
2997                 its_vpe_id_free(vpe_id);
2998                 return -ENOMEM;
2999         }
3000
3001         if (!its_alloc_vpe_table(vpe_id)) {
3002                 its_vpe_id_free(vpe_id);
3003                 its_free_pending_table(vpe->vpt_page);
3004                 return -ENOMEM;
3005         }
3006
3007         vpe->vpe_id = vpe_id;
3008         vpe->vpt_page = vpt_page;
3009         vpe->vpe_proxy_event = -1;
3010
3011         return 0;
3012 }
3013
3014 static void its_vpe_teardown(struct its_vpe *vpe)
3015 {
3016         its_vpe_db_proxy_unmap(vpe);
3017         its_vpe_id_free(vpe->vpe_id);
3018         its_free_pending_table(vpe->vpt_page);
3019 }
3020
3021 static void its_vpe_irq_domain_free(struct irq_domain *domain,
3022                                     unsigned int virq,
3023                                     unsigned int nr_irqs)
3024 {
3025         struct its_vm *vm = domain->host_data;
3026         int i;
3027
3028         irq_domain_free_irqs_parent(domain, virq, nr_irqs);
3029
3030         for (i = 0; i < nr_irqs; i++) {
3031                 struct irq_data *data = irq_domain_get_irq_data(domain,
3032                                                                 virq + i);
3033                 struct its_vpe *vpe = irq_data_get_irq_chip_data(data);
3034
3035                 BUG_ON(vm != vpe->its_vm);
3036
3037                 clear_bit(data->hwirq, vm->db_bitmap);
3038                 its_vpe_teardown(vpe);
3039                 irq_domain_reset_irq_data(data);
3040         }
3041
3042         if (bitmap_empty(vm->db_bitmap, vm->nr_db_lpis)) {
3043                 its_lpi_free(vm->db_bitmap, vm->db_lpi_base, vm->nr_db_lpis);
3044                 its_free_prop_table(vm->vprop_page);
3045         }
3046 }
3047
3048 static int its_vpe_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
3049                                     unsigned int nr_irqs, void *args)
3050 {
3051         struct its_vm *vm = args;
3052         unsigned long *bitmap;
3053         struct page *vprop_page;
3054         int base, nr_ids, i, err = 0;
3055
3056         BUG_ON(!vm);
3057
3058         bitmap = its_lpi_alloc(roundup_pow_of_two(nr_irqs), &base, &nr_ids);
3059         if (!bitmap)
3060                 return -ENOMEM;
3061
3062         if (nr_ids < nr_irqs) {
3063                 its_lpi_free(bitmap, base, nr_ids);
3064                 return -ENOMEM;
3065         }
3066
3067         vprop_page = its_allocate_prop_table(GFP_KERNEL);
3068         if (!vprop_page) {
3069                 its_lpi_free(bitmap, base, nr_ids);
3070                 return -ENOMEM;
3071         }
3072
3073         vm->db_bitmap = bitmap;
3074         vm->db_lpi_base = base;
3075         vm->nr_db_lpis = nr_ids;
3076         vm->vprop_page = vprop_page;
3077
3078         for (i = 0; i < nr_irqs; i++) {
3079                 vm->vpes[i]->vpe_db_lpi = base + i;
3080                 err = its_vpe_init(vm->vpes[i]);
3081                 if (err)
3082                         break;
3083                 err = its_irq_gic_domain_alloc(domain, virq + i,
3084                                                vm->vpes[i]->vpe_db_lpi);
3085                 if (err)
3086                         break;
3087                 irq_domain_set_hwirq_and_chip(domain, virq + i, i,
3088                                               &its_vpe_irq_chip, vm->vpes[i]);
3089                 set_bit(i, bitmap);
3090         }
3091
3092         if (err) {
3093                 if (i > 0)
3094                         its_vpe_irq_domain_free(domain, virq, i - 1);
3095
3096                 its_lpi_free(bitmap, base, nr_ids);
3097                 its_free_prop_table(vprop_page);
3098         }
3099
3100         return err;
3101 }
3102
3103 static int its_vpe_irq_domain_activate(struct irq_domain *domain,
3104                                        struct irq_data *d, bool reserve)
3105 {
3106         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3107         struct its_node *its;
3108
3109         /* If we use the list map, we issue VMAPP on demand... */
3110         if (its_list_map)
3111                 return 0;
3112
3113         /* Map the VPE to the first possible CPU */
3114         vpe->col_idx = cpumask_first(cpu_online_mask);
3115
3116         list_for_each_entry(its, &its_nodes, entry) {
3117                 if (!its->is_v4)
3118                         continue;
3119
3120                 its_send_vmapp(its, vpe, true);
3121                 its_send_vinvall(its, vpe);
3122         }
3123
3124         irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
3125
3126         return 0;
3127 }
3128
3129 static void its_vpe_irq_domain_deactivate(struct irq_domain *domain,
3130                                           struct irq_data *d)
3131 {
3132         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3133         struct its_node *its;
3134
3135         /*
3136          * If we use the list map, we unmap the VPE once no VLPIs are
3137          * associated with the VM.
3138          */
3139         if (its_list_map)
3140                 return;
3141
3142         list_for_each_entry(its, &its_nodes, entry) {
3143                 if (!its->is_v4)
3144                         continue;
3145
3146                 its_send_vmapp(its, vpe, false);
3147         }
3148 }
3149
3150 static const struct irq_domain_ops its_vpe_domain_ops = {
3151         .alloc                  = its_vpe_irq_domain_alloc,
3152         .free                   = its_vpe_irq_domain_free,
3153         .activate               = its_vpe_irq_domain_activate,
3154         .deactivate             = its_vpe_irq_domain_deactivate,
3155 };
3156
3157 static int its_force_quiescent(void __iomem *base)
3158 {
3159         u32 count = 1000000;    /* 1s */
3160         u32 val;
3161
3162         val = readl_relaxed(base + GITS_CTLR);
3163         /*
3164          * GIC architecture specification requires the ITS to be both
3165          * disabled and quiescent for writes to GITS_BASER<n> or
3166          * GITS_CBASER to not have UNPREDICTABLE results.
3167          */
3168         if ((val & GITS_CTLR_QUIESCENT) && !(val & GITS_CTLR_ENABLE))
3169                 return 0;
3170
3171         /* Disable the generation of all interrupts to this ITS */
3172         val &= ~(GITS_CTLR_ENABLE | GITS_CTLR_ImDe);
3173         writel_relaxed(val, base + GITS_CTLR);
3174
3175         /* Poll GITS_CTLR and wait until ITS becomes quiescent */
3176         while (1) {
3177                 val = readl_relaxed(base + GITS_CTLR);
3178                 if (val & GITS_CTLR_QUIESCENT)
3179                         return 0;
3180
3181                 count--;
3182                 if (!count)
3183                         return -EBUSY;
3184
3185                 cpu_relax();
3186                 udelay(1);
3187         }
3188 }
3189
3190 static bool __maybe_unused its_enable_quirk_cavium_22375(void *data)
3191 {
3192         struct its_node *its = data;
3193
3194         /* erratum 22375: only alloc 8MB table size */
3195         its->device_ids = 0x14;         /* 20 bits, 8MB */
3196         its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_22375;
3197
3198         return true;
3199 }
3200
3201 static bool __maybe_unused its_enable_quirk_cavium_23144(void *data)
3202 {
3203         struct its_node *its = data;
3204
3205         its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_23144;
3206
3207         return true;
3208 }
3209
3210 static bool __maybe_unused its_enable_quirk_qdf2400_e0065(void *data)
3211 {
3212         struct its_node *its = data;
3213
3214         /* On QDF2400, the size of the ITE is 16Bytes */
3215         its->ite_size = 16;
3216
3217         return true;
3218 }
3219
3220 static u64 its_irq_get_msi_base_pre_its(struct its_device *its_dev)
3221 {
3222         struct its_node *its = its_dev->its;
3223
3224         /*
3225          * The Socionext Synquacer SoC has a so-called 'pre-ITS',
3226          * which maps 32-bit writes targeted at a separate window of
3227          * size '4 << device_id_bits' onto writes to GITS_TRANSLATER
3228          * with device ID taken from bits [device_id_bits + 1:2] of
3229          * the window offset.
3230          */
3231         return its->pre_its_base + (its_dev->device_id << 2);
3232 }
3233
3234 static bool __maybe_unused its_enable_quirk_socionext_synquacer(void *data)
3235 {
3236         struct its_node *its = data;
3237         u32 pre_its_window[2];
3238         u32 ids;
3239
3240         if (!fwnode_property_read_u32_array(its->fwnode_handle,
3241                                            "socionext,synquacer-pre-its",
3242                                            pre_its_window,
3243                                            ARRAY_SIZE(pre_its_window))) {
3244
3245                 its->pre_its_base = pre_its_window[0];
3246                 its->get_msi_base = its_irq_get_msi_base_pre_its;
3247
3248                 ids = ilog2(pre_its_window[1]) - 2;
3249                 if (its->device_ids > ids)
3250                         its->device_ids = ids;
3251
3252                 /* the pre-ITS breaks isolation, so disable MSI remapping */
3253                 its->msi_domain_flags &= ~IRQ_DOMAIN_FLAG_MSI_REMAP;
3254                 return true;
3255         }
3256         return false;
3257 }
3258
3259 static bool __maybe_unused its_enable_quirk_hip07_161600802(void *data)
3260 {
3261         struct its_node *its = data;
3262
3263         /*
3264          * Hip07 insists on using the wrong address for the VLPI
3265          * page. Trick it into doing the right thing...
3266          */
3267         its->vlpi_redist_offset = SZ_128K;
3268         return true;
3269 }
3270
3271 static const struct gic_quirk its_quirks[] = {
3272 #ifdef CONFIG_CAVIUM_ERRATUM_22375
3273         {
3274                 .desc   = "ITS: Cavium errata 22375, 24313",
3275                 .iidr   = 0xa100034c,   /* ThunderX pass 1.x */
3276                 .mask   = 0xffff0fff,
3277                 .init   = its_enable_quirk_cavium_22375,
3278         },
3279 #endif
3280 #ifdef CONFIG_CAVIUM_ERRATUM_23144
3281         {
3282                 .desc   = "ITS: Cavium erratum 23144",
3283                 .iidr   = 0xa100034c,   /* ThunderX pass 1.x */
3284                 .mask   = 0xffff0fff,
3285                 .init   = its_enable_quirk_cavium_23144,
3286         },
3287 #endif
3288 #ifdef CONFIG_QCOM_QDF2400_ERRATUM_0065
3289         {
3290                 .desc   = "ITS: QDF2400 erratum 0065",
3291                 .iidr   = 0x00001070, /* QDF2400 ITS rev 1.x */
3292                 .mask   = 0xffffffff,
3293                 .init   = its_enable_quirk_qdf2400_e0065,
3294         },
3295 #endif
3296 #ifdef CONFIG_SOCIONEXT_SYNQUACER_PREITS
3297         {
3298                 /*
3299                  * The Socionext Synquacer SoC incorporates ARM's own GIC-500
3300                  * implementation, but with a 'pre-ITS' added that requires
3301                  * special handling in software.
3302                  */
3303                 .desc   = "ITS: Socionext Synquacer pre-ITS",
3304                 .iidr   = 0x0001143b,
3305                 .mask   = 0xffffffff,
3306                 .init   = its_enable_quirk_socionext_synquacer,
3307         },
3308 #endif
3309 #ifdef CONFIG_HISILICON_ERRATUM_161600802
3310         {
3311                 .desc   = "ITS: Hip07 erratum 161600802",
3312                 .iidr   = 0x00000004,
3313                 .mask   = 0xffffffff,
3314                 .init   = its_enable_quirk_hip07_161600802,
3315         },
3316 #endif
3317         {
3318         }
3319 };
3320
3321 static void its_enable_quirks(struct its_node *its)
3322 {
3323         u32 iidr = readl_relaxed(its->base + GITS_IIDR);
3324
3325         gic_enable_quirks(iidr, its_quirks, its);
3326 }
3327
3328 static int its_save_disable(void)
3329 {
3330         struct its_node *its;
3331         int err = 0;
3332
3333         raw_spin_lock(&its_lock);
3334         list_for_each_entry(its, &its_nodes, entry) {
3335                 void __iomem *base;
3336
3337                 if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3338                         continue;
3339
3340                 base = its->base;
3341                 its->ctlr_save = readl_relaxed(base + GITS_CTLR);
3342                 err = its_force_quiescent(base);
3343                 if (err) {
3344                         pr_err("ITS@%pa: failed to quiesce: %d\n",
3345                                &its->phys_base, err);
3346                         writel_relaxed(its->ctlr_save, base + GITS_CTLR);
3347                         goto err;
3348                 }
3349
3350                 its->cbaser_save = gits_read_cbaser(base + GITS_CBASER);
3351         }
3352
3353 err:
3354         if (err) {
3355                 list_for_each_entry_continue_reverse(its, &its_nodes, entry) {
3356                         void __iomem *base;
3357
3358                         if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3359                                 continue;
3360
3361                         base = its->base;
3362                         writel_relaxed(its->ctlr_save, base + GITS_CTLR);
3363                 }
3364         }
3365         raw_spin_unlock(&its_lock);
3366
3367         return err;
3368 }
3369
3370 static void its_restore_enable(void)
3371 {
3372         struct its_node *its;
3373         int ret;
3374
3375         raw_spin_lock(&its_lock);
3376         list_for_each_entry(its, &its_nodes, entry) {
3377                 void __iomem *base;
3378                 int i;
3379
3380                 if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3381                         continue;
3382
3383                 base = its->base;
3384
3385                 /*
3386                  * Make sure that the ITS is disabled. If it fails to quiesce,
3387                  * don't restore it since writing to CBASER or BASER<n>
3388                  * registers is undefined according to the GIC v3 ITS
3389                  * Specification.
3390                  */
3391                 ret = its_force_quiescent(base);
3392                 if (ret) {
3393                         pr_err("ITS@%pa: failed to quiesce on resume: %d\n",
3394                                &its->phys_base, ret);
3395                         continue;
3396                 }
3397
3398                 gits_write_cbaser(its->cbaser_save, base + GITS_CBASER);
3399
3400                 /*
3401                  * Writing CBASER resets CREADR to 0, so make CWRITER and
3402                  * cmd_write line up with it.
3403                  */
3404                 its->cmd_write = its->cmd_base;
3405                 gits_write_cwriter(0, base + GITS_CWRITER);
3406
3407                 /* Restore GITS_BASER from the value cache. */
3408                 for (i = 0; i < GITS_BASER_NR_REGS; i++) {
3409                         struct its_baser *baser = &its->tables[i];
3410
3411                         if (!(baser->val & GITS_BASER_VALID))
3412                                 continue;
3413
3414                         its_write_baser(its, baser, baser->val);
3415                 }
3416                 writel_relaxed(its->ctlr_save, base + GITS_CTLR);
3417
3418                 /*
3419                  * Reinit the collection if it's stored in the ITS. This is
3420                  * indicated by the col_id being less than the HCC field.
3421                  * CID < HCC as specified in the GIC v3 Documentation.
3422                  */
3423                 if (its->collections[smp_processor_id()].col_id <
3424                     GITS_TYPER_HCC(gic_read_typer(base + GITS_TYPER)))
3425                         its_cpu_init_collection(its);
3426         }
3427         raw_spin_unlock(&its_lock);
3428 }
3429
3430 static struct syscore_ops its_syscore_ops = {
3431         .suspend = its_save_disable,
3432         .resume = its_restore_enable,
3433 };
3434
3435 static int its_init_domain(struct fwnode_handle *handle, struct its_node *its)
3436 {
3437         struct irq_domain *inner_domain;
3438         struct msi_domain_info *info;
3439
3440         info = kzalloc(sizeof(*info), GFP_KERNEL);
3441         if (!info)
3442                 return -ENOMEM;
3443
3444         inner_domain = irq_domain_create_tree(handle, &its_domain_ops, its);
3445         if (!inner_domain) {
3446                 kfree(info);
3447                 return -ENOMEM;
3448         }
3449
3450         inner_domain->parent = its_parent;
3451         irq_domain_update_bus_token(inner_domain, DOMAIN_BUS_NEXUS);
3452         inner_domain->flags |= its->msi_domain_flags;
3453         info->ops = &its_msi_domain_ops;
3454         info->data = its;
3455         inner_domain->host_data = info;
3456
3457         return 0;
3458 }
3459
3460 static int its_init_vpe_domain(void)
3461 {
3462         struct its_node *its;
3463         u32 devid;
3464         int entries;
3465
3466         if (gic_rdists->has_direct_lpi) {
3467                 pr_info("ITS: Using DirectLPI for VPE invalidation\n");
3468                 return 0;
3469         }
3470
3471         /* Any ITS will do, even if not v4 */
3472         its = list_first_entry(&its_nodes, struct its_node, entry);
3473
3474         entries = roundup_pow_of_two(nr_cpu_ids);
3475         vpe_proxy.vpes = kcalloc(entries, sizeof(*vpe_proxy.vpes),
3476                                  GFP_KERNEL);
3477         if (!vpe_proxy.vpes) {
3478                 pr_err("ITS: Can't allocate GICv4 proxy device array\n");
3479                 return -ENOMEM;
3480         }
3481
3482         /* Use the last possible DevID */
3483         devid = GENMASK(its->device_ids - 1, 0);
3484         vpe_proxy.dev = its_create_device(its, devid, entries, false);
3485         if (!vpe_proxy.dev) {
3486                 kfree(vpe_proxy.vpes);
3487                 pr_err("ITS: Can't allocate GICv4 proxy device\n");
3488                 return -ENOMEM;
3489         }
3490
3491         BUG_ON(entries > vpe_proxy.dev->nr_ites);
3492
3493         raw_spin_lock_init(&vpe_proxy.lock);
3494         vpe_proxy.next_victim = 0;
3495         pr_info("ITS: Allocated DevID %x as GICv4 proxy device (%d slots)\n",
3496                 devid, vpe_proxy.dev->nr_ites);
3497
3498         return 0;
3499 }
3500
3501 static int __init its_compute_its_list_map(struct resource *res,
3502                                            void __iomem *its_base)
3503 {
3504         int its_number;
3505         u32 ctlr;
3506
3507         /*
3508          * This is assumed to be done early enough that we're
3509          * guaranteed to be single-threaded, hence no
3510          * locking. Should this change, we should address
3511          * this.
3512          */
3513         its_number = find_first_zero_bit(&its_list_map, GICv4_ITS_LIST_MAX);
3514         if (its_number >= GICv4_ITS_LIST_MAX) {
3515                 pr_err("ITS@%pa: No ITSList entry available!\n",
3516                        &res->start);
3517                 return -EINVAL;
3518         }
3519
3520         ctlr = readl_relaxed(its_base + GITS_CTLR);
3521         ctlr &= ~GITS_CTLR_ITS_NUMBER;
3522         ctlr |= its_number << GITS_CTLR_ITS_NUMBER_SHIFT;
3523         writel_relaxed(ctlr, its_base + GITS_CTLR);
3524         ctlr = readl_relaxed(its_base + GITS_CTLR);
3525         if ((ctlr & GITS_CTLR_ITS_NUMBER) != (its_number << GITS_CTLR_ITS_NUMBER_SHIFT)) {
3526                 its_number = ctlr & GITS_CTLR_ITS_NUMBER;
3527                 its_number >>= GITS_CTLR_ITS_NUMBER_SHIFT;
3528         }
3529
3530         if (test_and_set_bit(its_number, &its_list_map)) {
3531                 pr_err("ITS@%pa: Duplicate ITSList entry %d\n",
3532                        &res->start, its_number);
3533                 return -EINVAL;
3534         }
3535
3536         return its_number;
3537 }
3538
3539 static int __init its_probe_one(struct resource *res,
3540                                 struct fwnode_handle *handle, int numa_node)
3541 {
3542         struct its_node *its;
3543         void __iomem *its_base;
3544         u32 val, ctlr;
3545         u64 baser, tmp, typer;
3546         int err;
3547
3548         its_base = ioremap(res->start, resource_size(res));
3549         if (!its_base) {
3550                 pr_warn("ITS@%pa: Unable to map ITS registers\n", &res->start);
3551                 return -ENOMEM;
3552         }
3553
3554         val = readl_relaxed(its_base + GITS_PIDR2) & GIC_PIDR2_ARCH_MASK;
3555         if (val != 0x30 && val != 0x40) {
3556                 pr_warn("ITS@%pa: No ITS detected, giving up\n", &res->start);
3557                 err = -ENODEV;
3558                 goto out_unmap;
3559         }
3560
3561         err = its_force_quiescent(its_base);
3562         if (err) {
3563                 pr_warn("ITS@%pa: Failed to quiesce, giving up\n", &res->start);
3564                 goto out_unmap;
3565         }
3566
3567         pr_info("ITS %pR\n", res);
3568
3569         its = kzalloc(sizeof(*its), GFP_KERNEL);
3570         if (!its) {
3571                 err = -ENOMEM;
3572                 goto out_unmap;
3573         }
3574
3575         raw_spin_lock_init(&its->lock);
3576         mutex_init(&its->dev_alloc_lock);
3577         INIT_LIST_HEAD(&its->entry);
3578         INIT_LIST_HEAD(&its->its_device_list);
3579         typer = gic_read_typer(its_base + GITS_TYPER);
3580         its->base = its_base;
3581         its->phys_base = res->start;
3582         its->ite_size = GITS_TYPER_ITT_ENTRY_SIZE(typer);
3583         its->device_ids = GITS_TYPER_DEVBITS(typer);
3584         its->is_v4 = !!(typer & GITS_TYPER_VLPIS);
3585         if (its->is_v4) {
3586                 if (!(typer & GITS_TYPER_VMOVP)) {
3587                         err = its_compute_its_list_map(res, its_base);
3588                         if (err < 0)
3589                                 goto out_free_its;
3590
3591                         its->list_nr = err;
3592
3593                         pr_info("ITS@%pa: Using ITS number %d\n",
3594                                 &res->start, err);
3595                 } else {
3596                         pr_info("ITS@%pa: Single VMOVP capable\n", &res->start);
3597                 }
3598         }
3599
3600         its->numa_node = numa_node;
3601
3602         its->cmd_base = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
3603                                                 get_order(ITS_CMD_QUEUE_SZ));
3604         if (!its->cmd_base) {
3605                 err = -ENOMEM;
3606                 goto out_free_its;
3607         }
3608         its->cmd_write = its->cmd_base;
3609         its->fwnode_handle = handle;
3610         its->get_msi_base = its_irq_get_msi_base;
3611         its->msi_domain_flags = IRQ_DOMAIN_FLAG_MSI_REMAP;
3612
3613         its_enable_quirks(its);
3614
3615         err = its_alloc_tables(its);
3616         if (err)
3617                 goto out_free_cmd;
3618
3619         err = its_alloc_collections(its);
3620         if (err)
3621                 goto out_free_tables;
3622
3623         baser = (virt_to_phys(its->cmd_base)    |
3624                  GITS_CBASER_RaWaWb             |
3625                  GITS_CBASER_InnerShareable     |
3626                  (ITS_CMD_QUEUE_SZ / SZ_4K - 1) |
3627                  GITS_CBASER_VALID);
3628
3629         gits_write_cbaser(baser, its->base + GITS_CBASER);
3630         tmp = gits_read_cbaser(its->base + GITS_CBASER);
3631
3632         if ((tmp ^ baser) & GITS_CBASER_SHAREABILITY_MASK) {
3633                 if (!(tmp & GITS_CBASER_SHAREABILITY_MASK)) {
3634                         /*
3635                          * The HW reports non-shareable, we must
3636                          * remove the cacheability attributes as
3637                          * well.
3638                          */
3639                         baser &= ~(GITS_CBASER_SHAREABILITY_MASK |
3640                                    GITS_CBASER_CACHEABILITY_MASK);
3641                         baser |= GITS_CBASER_nC;
3642                         gits_write_cbaser(baser, its->base + GITS_CBASER);
3643                 }
3644                 pr_info("ITS: using cache flushing for cmd queue\n");
3645                 its->flags |= ITS_FLAGS_CMDQ_NEEDS_FLUSHING;
3646         }
3647
3648         gits_write_cwriter(0, its->base + GITS_CWRITER);
3649         ctlr = readl_relaxed(its->base + GITS_CTLR);
3650         ctlr |= GITS_CTLR_ENABLE;
3651         if (its->is_v4)
3652                 ctlr |= GITS_CTLR_ImDe;
3653         writel_relaxed(ctlr, its->base + GITS_CTLR);
3654
3655         if (GITS_TYPER_HCC(typer))
3656                 its->flags |= ITS_FLAGS_SAVE_SUSPEND_STATE;
3657
3658         err = its_init_domain(handle, its);
3659         if (err)
3660                 goto out_free_tables;
3661
3662         raw_spin_lock(&its_lock);
3663         list_add(&its->entry, &its_nodes);
3664         raw_spin_unlock(&its_lock);
3665
3666         return 0;
3667
3668 out_free_tables:
3669         its_free_tables(its);
3670 out_free_cmd:
3671         free_pages((unsigned long)its->cmd_base, get_order(ITS_CMD_QUEUE_SZ));
3672 out_free_its:
3673         kfree(its);
3674 out_unmap:
3675         iounmap(its_base);
3676         pr_err("ITS@%pa: failed probing (%d)\n", &res->start, err);
3677         return err;
3678 }
3679
3680 static bool gic_rdists_supports_plpis(void)
3681 {
3682         return !!(gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER) & GICR_TYPER_PLPIS);
3683 }
3684
3685 static int redist_disable_lpis(void)
3686 {
3687         void __iomem *rbase = gic_data_rdist_rd_base();
3688         u64 timeout = USEC_PER_SEC;
3689         u64 val;
3690
3691         if (!gic_rdists_supports_plpis()) {
3692                 pr_info("CPU%d: LPIs not supported\n", smp_processor_id());
3693                 return -ENXIO;
3694         }
3695
3696         val = readl_relaxed(rbase + GICR_CTLR);
3697         if (!(val & GICR_CTLR_ENABLE_LPIS))
3698                 return 0;
3699
3700         /*
3701          * If coming via a CPU hotplug event, we don't need to disable
3702          * LPIs before trying to re-enable them. They are already
3703          * configured and all is well in the world.
3704          *
3705          * If running with preallocated tables, there is nothing to do.
3706          */
3707         if (gic_data_rdist()->lpi_enabled ||
3708             (gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED))
3709                 return 0;
3710
3711         /*
3712          * From that point on, we only try to do some damage control.
3713          */
3714         pr_warn("GICv3: CPU%d: Booted with LPIs enabled, memory probably corrupted\n",
3715                 smp_processor_id());
3716         add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
3717
3718         /* Disable LPIs */
3719         val &= ~GICR_CTLR_ENABLE_LPIS;
3720         writel_relaxed(val, rbase + GICR_CTLR);
3721
3722         /* Make sure any change to GICR_CTLR is observable by the GIC */
3723         dsb(sy);
3724
3725         /*
3726          * Software must observe RWP==0 after clearing GICR_CTLR.EnableLPIs
3727          * from 1 to 0 before programming GICR_PEND{PROP}BASER registers.
3728          * Error out if we time out waiting for RWP to clear.
3729          */
3730         while (readl_relaxed(rbase + GICR_CTLR) & GICR_CTLR_RWP) {
3731                 if (!timeout) {
3732                         pr_err("CPU%d: Timeout while disabling LPIs\n",
3733                                smp_processor_id());
3734                         return -ETIMEDOUT;
3735                 }
3736                 udelay(1);
3737                 timeout--;
3738         }
3739
3740         /*
3741          * After it has been written to 1, it is IMPLEMENTATION
3742          * DEFINED whether GICR_CTLR.EnableLPI becomes RES1 or can be
3743          * cleared to 0. Error out if clearing the bit failed.
3744          */
3745         if (readl_relaxed(rbase + GICR_CTLR) & GICR_CTLR_ENABLE_LPIS) {
3746                 pr_err("CPU%d: Failed to disable LPIs\n", smp_processor_id());
3747                 return -EBUSY;
3748         }
3749
3750         return 0;
3751 }
3752
3753 int its_cpu_init(void)
3754 {
3755         if (!list_empty(&its_nodes)) {
3756                 int ret;
3757
3758                 ret = redist_disable_lpis();
3759                 if (ret)
3760                         return ret;
3761
3762                 its_cpu_init_lpis();
3763                 its_cpu_init_collections();
3764         }
3765
3766         return 0;
3767 }
3768
3769 static const struct of_device_id its_device_id[] = {
3770         {       .compatible     = "arm,gic-v3-its",     },
3771         {},
3772 };
3773
3774 static int __init its_of_probe(struct device_node *node)
3775 {
3776         struct device_node *np;
3777         struct resource res;
3778
3779         for (np = of_find_matching_node(node, its_device_id); np;
3780              np = of_find_matching_node(np, its_device_id)) {
3781                 if (!of_device_is_available(np))
3782                         continue;
3783                 if (!of_property_read_bool(np, "msi-controller")) {
3784                         pr_warn("%pOF: no msi-controller property, ITS ignored\n",
3785                                 np);
3786                         continue;
3787                 }
3788
3789                 if (of_address_to_resource(np, 0, &res)) {
3790                         pr_warn("%pOF: no regs?\n", np);
3791                         continue;
3792                 }
3793
3794                 its_probe_one(&res, &np->fwnode, of_node_to_nid(np));
3795         }
3796         return 0;
3797 }
3798
3799 #ifdef CONFIG_ACPI
3800
3801 #define ACPI_GICV3_ITS_MEM_SIZE (SZ_128K)
3802
3803 #ifdef CONFIG_ACPI_NUMA
3804 struct its_srat_map {
3805         /* numa node id */
3806         u32     numa_node;
3807         /* GIC ITS ID */
3808         u32     its_id;
3809 };
3810
3811 static struct its_srat_map *its_srat_maps __initdata;
3812 static int its_in_srat __initdata;
3813
3814 static int __init acpi_get_its_numa_node(u32 its_id)
3815 {
3816         int i;
3817
3818         for (i = 0; i < its_in_srat; i++) {
3819                 if (its_id == its_srat_maps[i].its_id)
3820                         return its_srat_maps[i].numa_node;
3821         }
3822         return NUMA_NO_NODE;
3823 }
3824
3825 static int __init gic_acpi_match_srat_its(struct acpi_subtable_header *header,
3826                                           const unsigned long end)
3827 {
3828         return 0;
3829 }
3830
3831 static int __init gic_acpi_parse_srat_its(struct acpi_subtable_header *header,
3832                          const unsigned long end)
3833 {
3834         int node;
3835         struct acpi_srat_gic_its_affinity *its_affinity;
3836
3837         its_affinity = (struct acpi_srat_gic_its_affinity *)header;
3838         if (!its_affinity)
3839                 return -EINVAL;
3840
3841         if (its_affinity->header.length < sizeof(*its_affinity)) {
3842                 pr_err("SRAT: Invalid header length %d in ITS affinity\n",
3843                         its_affinity->header.length);
3844                 return -EINVAL;
3845         }
3846
3847         node = acpi_map_pxm_to_node(its_affinity->proximity_domain);
3848
3849         if (node == NUMA_NO_NODE || node >= MAX_NUMNODES) {
3850                 pr_err("SRAT: Invalid NUMA node %d in ITS affinity\n", node);
3851                 return 0;
3852         }
3853
3854         its_srat_maps[its_in_srat].numa_node = node;
3855         its_srat_maps[its_in_srat].its_id = its_affinity->its_id;
3856         its_in_srat++;
3857         pr_info("SRAT: PXM %d -> ITS %d -> Node %d\n",
3858                 its_affinity->proximity_domain, its_affinity->its_id, node);
3859
3860         return 0;
3861 }
3862
3863 static void __init acpi_table_parse_srat_its(void)
3864 {
3865         int count;
3866
3867         count = acpi_table_parse_entries(ACPI_SIG_SRAT,
3868                         sizeof(struct acpi_table_srat),
3869                         ACPI_SRAT_TYPE_GIC_ITS_AFFINITY,
3870                         gic_acpi_match_srat_its, 0);
3871         if (count <= 0)
3872                 return;
3873
3874         its_srat_maps = kmalloc_array(count, sizeof(struct its_srat_map),
3875                                       GFP_KERNEL);
3876         if (!its_srat_maps) {
3877                 pr_warn("SRAT: Failed to allocate memory for its_srat_maps!\n");
3878                 return;
3879         }
3880
3881         acpi_table_parse_entries(ACPI_SIG_SRAT,
3882                         sizeof(struct acpi_table_srat),
3883                         ACPI_SRAT_TYPE_GIC_ITS_AFFINITY,
3884                         gic_acpi_parse_srat_its, 0);
3885 }
3886
3887 /* free the its_srat_maps after ITS probing */
3888 static void __init acpi_its_srat_maps_free(void)
3889 {
3890         kfree(its_srat_maps);
3891 }
3892 #else
3893 static void __init acpi_table_parse_srat_its(void)      { }
3894 static int __init acpi_get_its_numa_node(u32 its_id) { return NUMA_NO_NODE; }
3895 static void __init acpi_its_srat_maps_free(void) { }
3896 #endif
3897
3898 static int __init gic_acpi_parse_madt_its(struct acpi_subtable_header *header,
3899                                           const unsigned long end)
3900 {
3901         struct acpi_madt_generic_translator *its_entry;
3902         struct fwnode_handle *dom_handle;
3903         struct resource res;
3904         int err;
3905
3906         its_entry = (struct acpi_madt_generic_translator *)header;
3907         memset(&res, 0, sizeof(res));
3908         res.start = its_entry->base_address;
3909         res.end = its_entry->base_address + ACPI_GICV3_ITS_MEM_SIZE - 1;
3910         res.flags = IORESOURCE_MEM;
3911
3912         dom_handle = irq_domain_alloc_fwnode((void *)its_entry->base_address);
3913         if (!dom_handle) {
3914                 pr_err("ITS@%pa: Unable to allocate GICv3 ITS domain token\n",
3915                        &res.start);
3916                 return -ENOMEM;
3917         }
3918
3919         err = iort_register_domain_token(its_entry->translation_id, res.start,
3920                                          dom_handle);
3921         if (err) {
3922                 pr_err("ITS@%pa: Unable to register GICv3 ITS domain token (ITS ID %d) to IORT\n",
3923                        &res.start, its_entry->translation_id);
3924                 goto dom_err;
3925         }
3926
3927         err = its_probe_one(&res, dom_handle,
3928                         acpi_get_its_numa_node(its_entry->translation_id));
3929         if (!err)
3930                 return 0;
3931
3932         iort_deregister_domain_token(its_entry->translation_id);
3933 dom_err:
3934         irq_domain_free_fwnode(dom_handle);
3935         return err;
3936 }
3937
3938 static void __init its_acpi_probe(void)
3939 {
3940         acpi_table_parse_srat_its();
3941         acpi_table_parse_madt(ACPI_MADT_TYPE_GENERIC_TRANSLATOR,
3942                               gic_acpi_parse_madt_its, 0);
3943         acpi_its_srat_maps_free();
3944 }
3945 #else
3946 static void __init its_acpi_probe(void) { }
3947 #endif
3948
3949 int __init its_init(struct fwnode_handle *handle, struct rdists *rdists,
3950                     struct irq_domain *parent_domain)
3951 {
3952         struct device_node *of_node;
3953         struct its_node *its;
3954         bool has_v4 = false;
3955         int err;
3956
3957         its_parent = parent_domain;
3958         of_node = to_of_node(handle);
3959         if (of_node)
3960                 its_of_probe(of_node);
3961         else
3962                 its_acpi_probe();
3963
3964         if (list_empty(&its_nodes)) {
3965                 pr_warn("ITS: No ITS available, not enabling LPIs\n");
3966                 return -ENXIO;
3967         }
3968
3969         gic_rdists = rdists;
3970
3971         err = allocate_lpi_tables();
3972         if (err)
3973                 return err;
3974
3975         list_for_each_entry(its, &its_nodes, entry)
3976                 has_v4 |= its->is_v4;
3977
3978         if (has_v4 & rdists->has_vlpis) {
3979                 if (its_init_vpe_domain() ||
3980                     its_init_v4(parent_domain, &its_vpe_domain_ops)) {
3981                         rdists->has_vlpis = false;
3982                         pr_err("ITS: Disabling GICv4 support\n");
3983                 }
3984         }
3985
3986         register_syscore_ops(&its_syscore_ops);
3987
3988         return 0;
3989 }