Merge branch 'irq-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[muen/linux.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         struct page *page;
1750         void *base;
1751
1752 retry_alloc_baser:
1753         alloc_pages = (PAGE_ORDER_TO_SIZE(order) / psz);
1754         if (alloc_pages > GITS_BASER_PAGES_MAX) {
1755                 pr_warn("ITS@%pa: %s too large, reduce ITS pages %u->%u\n",
1756                         &its->phys_base, its_base_type_string[type],
1757                         alloc_pages, GITS_BASER_PAGES_MAX);
1758                 alloc_pages = GITS_BASER_PAGES_MAX;
1759                 order = get_order(GITS_BASER_PAGES_MAX * psz);
1760         }
1761
1762         page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO, order);
1763         if (!page)
1764                 return -ENOMEM;
1765
1766         base = (void *)page_address(page);
1767         baser_phys = virt_to_phys(base);
1768
1769         /* Check if the physical address of the memory is above 48bits */
1770         if (IS_ENABLED(CONFIG_ARM64_64K_PAGES) && (baser_phys >> 48)) {
1771
1772                 /* 52bit PA is supported only when PageSize=64K */
1773                 if (psz != SZ_64K) {
1774                         pr_err("ITS: no 52bit PA support when psz=%d\n", psz);
1775                         free_pages((unsigned long)base, order);
1776                         return -ENXIO;
1777                 }
1778
1779                 /* Convert 52bit PA to 48bit field */
1780                 baser_phys = GITS_BASER_PHYS_52_to_48(baser_phys);
1781         }
1782
1783 retry_baser:
1784         val = (baser_phys                                        |
1785                 (type << GITS_BASER_TYPE_SHIFT)                  |
1786                 ((esz - 1) << GITS_BASER_ENTRY_SIZE_SHIFT)       |
1787                 ((alloc_pages - 1) << GITS_BASER_PAGES_SHIFT)    |
1788                 cache                                            |
1789                 shr                                              |
1790                 GITS_BASER_VALID);
1791
1792         val |=  indirect ? GITS_BASER_INDIRECT : 0x0;
1793
1794         switch (psz) {
1795         case SZ_4K:
1796                 val |= GITS_BASER_PAGE_SIZE_4K;
1797                 break;
1798         case SZ_16K:
1799                 val |= GITS_BASER_PAGE_SIZE_16K;
1800                 break;
1801         case SZ_64K:
1802                 val |= GITS_BASER_PAGE_SIZE_64K;
1803                 break;
1804         }
1805
1806         its_write_baser(its, baser, val);
1807         tmp = baser->val;
1808
1809         if ((val ^ tmp) & GITS_BASER_SHAREABILITY_MASK) {
1810                 /*
1811                  * Shareability didn't stick. Just use
1812                  * whatever the read reported, which is likely
1813                  * to be the only thing this redistributor
1814                  * supports. If that's zero, make it
1815                  * non-cacheable as well.
1816                  */
1817                 shr = tmp & GITS_BASER_SHAREABILITY_MASK;
1818                 if (!shr) {
1819                         cache = GITS_BASER_nC;
1820                         gic_flush_dcache_to_poc(base, PAGE_ORDER_TO_SIZE(order));
1821                 }
1822                 goto retry_baser;
1823         }
1824
1825         if ((val ^ tmp) & GITS_BASER_PAGE_SIZE_MASK) {
1826                 /*
1827                  * Page size didn't stick. Let's try a smaller
1828                  * size and retry. If we reach 4K, then
1829                  * something is horribly wrong...
1830                  */
1831                 free_pages((unsigned long)base, order);
1832                 baser->base = NULL;
1833
1834                 switch (psz) {
1835                 case SZ_16K:
1836                         psz = SZ_4K;
1837                         goto retry_alloc_baser;
1838                 case SZ_64K:
1839                         psz = SZ_16K;
1840                         goto retry_alloc_baser;
1841                 }
1842         }
1843
1844         if (val != tmp) {
1845                 pr_err("ITS@%pa: %s doesn't stick: %llx %llx\n",
1846                        &its->phys_base, its_base_type_string[type],
1847                        val, tmp);
1848                 free_pages((unsigned long)base, order);
1849                 return -ENXIO;
1850         }
1851
1852         baser->order = order;
1853         baser->base = base;
1854         baser->psz = psz;
1855         tmp = indirect ? GITS_LVL1_ENTRY_SIZE : esz;
1856
1857         pr_info("ITS@%pa: allocated %d %s @%lx (%s, esz %d, psz %dK, shr %d)\n",
1858                 &its->phys_base, (int)(PAGE_ORDER_TO_SIZE(order) / (int)tmp),
1859                 its_base_type_string[type],
1860                 (unsigned long)virt_to_phys(base),
1861                 indirect ? "indirect" : "flat", (int)esz,
1862                 psz / SZ_1K, (int)shr >> GITS_BASER_SHAREABILITY_SHIFT);
1863
1864         return 0;
1865 }
1866
1867 static bool its_parse_indirect_baser(struct its_node *its,
1868                                      struct its_baser *baser,
1869                                      u32 psz, u32 *order, u32 ids)
1870 {
1871         u64 tmp = its_read_baser(its, baser);
1872         u64 type = GITS_BASER_TYPE(tmp);
1873         u64 esz = GITS_BASER_ENTRY_SIZE(tmp);
1874         u64 val = GITS_BASER_InnerShareable | GITS_BASER_RaWaWb;
1875         u32 new_order = *order;
1876         bool indirect = false;
1877
1878         /* No need to enable Indirection if memory requirement < (psz*2)bytes */
1879         if ((esz << ids) > (psz * 2)) {
1880                 /*
1881                  * Find out whether hw supports a single or two-level table by
1882                  * table by reading bit at offset '62' after writing '1' to it.
1883                  */
1884                 its_write_baser(its, baser, val | GITS_BASER_INDIRECT);
1885                 indirect = !!(baser->val & GITS_BASER_INDIRECT);
1886
1887                 if (indirect) {
1888                         /*
1889                          * The size of the lvl2 table is equal to ITS page size
1890                          * which is 'psz'. For computing lvl1 table size,
1891                          * subtract ID bits that sparse lvl2 table from 'ids'
1892                          * which is reported by ITS hardware times lvl1 table
1893                          * entry size.
1894                          */
1895                         ids -= ilog2(psz / (int)esz);
1896                         esz = GITS_LVL1_ENTRY_SIZE;
1897                 }
1898         }
1899
1900         /*
1901          * Allocate as many entries as required to fit the
1902          * range of device IDs that the ITS can grok... The ID
1903          * space being incredibly sparse, this results in a
1904          * massive waste of memory if two-level device table
1905          * feature is not supported by hardware.
1906          */
1907         new_order = max_t(u32, get_order(esz << ids), new_order);
1908         if (new_order >= MAX_ORDER) {
1909                 new_order = MAX_ORDER - 1;
1910                 ids = ilog2(PAGE_ORDER_TO_SIZE(new_order) / (int)esz);
1911                 pr_warn("ITS@%pa: %s Table too large, reduce ids %u->%u\n",
1912                         &its->phys_base, its_base_type_string[type],
1913                         its->device_ids, ids);
1914         }
1915
1916         *order = new_order;
1917
1918         return indirect;
1919 }
1920
1921 static void its_free_tables(struct its_node *its)
1922 {
1923         int i;
1924
1925         for (i = 0; i < GITS_BASER_NR_REGS; i++) {
1926                 if (its->tables[i].base) {
1927                         free_pages((unsigned long)its->tables[i].base,
1928                                    its->tables[i].order);
1929                         its->tables[i].base = NULL;
1930                 }
1931         }
1932 }
1933
1934 static int its_alloc_tables(struct its_node *its)
1935 {
1936         u64 shr = GITS_BASER_InnerShareable;
1937         u64 cache = GITS_BASER_RaWaWb;
1938         u32 psz = SZ_64K;
1939         int err, i;
1940
1941         if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_22375)
1942                 /* erratum 24313: ignore memory access type */
1943                 cache = GITS_BASER_nCnB;
1944
1945         for (i = 0; i < GITS_BASER_NR_REGS; i++) {
1946                 struct its_baser *baser = its->tables + i;
1947                 u64 val = its_read_baser(its, baser);
1948                 u64 type = GITS_BASER_TYPE(val);
1949                 u32 order = get_order(psz);
1950                 bool indirect = false;
1951
1952                 switch (type) {
1953                 case GITS_BASER_TYPE_NONE:
1954                         continue;
1955
1956                 case GITS_BASER_TYPE_DEVICE:
1957                         indirect = its_parse_indirect_baser(its, baser,
1958                                                             psz, &order,
1959                                                             its->device_ids);
1960                         break;
1961
1962                 case GITS_BASER_TYPE_VCPU:
1963                         indirect = its_parse_indirect_baser(its, baser,
1964                                                             psz, &order,
1965                                                             ITS_MAX_VPEID_BITS);
1966                         break;
1967                 }
1968
1969                 err = its_setup_baser(its, baser, cache, shr, psz, order, indirect);
1970                 if (err < 0) {
1971                         its_free_tables(its);
1972                         return err;
1973                 }
1974
1975                 /* Update settings which will be used for next BASERn */
1976                 psz = baser->psz;
1977                 cache = baser->val & GITS_BASER_CACHEABILITY_MASK;
1978                 shr = baser->val & GITS_BASER_SHAREABILITY_MASK;
1979         }
1980
1981         return 0;
1982 }
1983
1984 static int its_alloc_collections(struct its_node *its)
1985 {
1986         int i;
1987
1988         its->collections = kcalloc(nr_cpu_ids, sizeof(*its->collections),
1989                                    GFP_KERNEL);
1990         if (!its->collections)
1991                 return -ENOMEM;
1992
1993         for (i = 0; i < nr_cpu_ids; i++)
1994                 its->collections[i].target_address = ~0ULL;
1995
1996         return 0;
1997 }
1998
1999 static struct page *its_allocate_pending_table(gfp_t gfp_flags)
2000 {
2001         struct page *pend_page;
2002
2003         pend_page = alloc_pages(gfp_flags | __GFP_ZERO,
2004                                 get_order(LPI_PENDBASE_SZ));
2005         if (!pend_page)
2006                 return NULL;
2007
2008         /* Make sure the GIC will observe the zero-ed page */
2009         gic_flush_dcache_to_poc(page_address(pend_page), LPI_PENDBASE_SZ);
2010
2011         return pend_page;
2012 }
2013
2014 static void its_free_pending_table(struct page *pt)
2015 {
2016         free_pages((unsigned long)page_address(pt), get_order(LPI_PENDBASE_SZ));
2017 }
2018
2019 /*
2020  * Booting with kdump and LPIs enabled is generally fine. Any other
2021  * case is wrong in the absence of firmware/EFI support.
2022  */
2023 static bool enabled_lpis_allowed(void)
2024 {
2025         phys_addr_t addr;
2026         u64 val;
2027
2028         /* Check whether the property table is in a reserved region */
2029         val = gicr_read_propbaser(gic_data_rdist_rd_base() + GICR_PROPBASER);
2030         addr = val & GENMASK_ULL(51, 12);
2031
2032         return gic_check_reserved_range(addr, LPI_PROPBASE_SZ);
2033 }
2034
2035 static int __init allocate_lpi_tables(void)
2036 {
2037         u64 val;
2038         int err, cpu;
2039
2040         /*
2041          * If LPIs are enabled while we run this from the boot CPU,
2042          * flag the RD tables as pre-allocated if the stars do align.
2043          */
2044         val = readl_relaxed(gic_data_rdist_rd_base() + GICR_CTLR);
2045         if ((val & GICR_CTLR_ENABLE_LPIS) && enabled_lpis_allowed()) {
2046                 gic_rdists->flags |= (RDIST_FLAGS_RD_TABLES_PREALLOCATED |
2047                                       RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING);
2048                 pr_info("GICv3: Using preallocated redistributor tables\n");
2049         }
2050
2051         err = its_setup_lpi_prop_table();
2052         if (err)
2053                 return err;
2054
2055         /*
2056          * We allocate all the pending tables anyway, as we may have a
2057          * mix of RDs that have had LPIs enabled, and some that
2058          * don't. We'll free the unused ones as each CPU comes online.
2059          */
2060         for_each_possible_cpu(cpu) {
2061                 struct page *pend_page;
2062
2063                 pend_page = its_allocate_pending_table(GFP_NOWAIT);
2064                 if (!pend_page) {
2065                         pr_err("Failed to allocate PENDBASE for CPU%d\n", cpu);
2066                         return -ENOMEM;
2067                 }
2068
2069                 gic_data_rdist_cpu(cpu)->pend_page = pend_page;
2070         }
2071
2072         return 0;
2073 }
2074
2075 static u64 its_clear_vpend_valid(void __iomem *vlpi_base)
2076 {
2077         u32 count = 1000000;    /* 1s! */
2078         bool clean;
2079         u64 val;
2080
2081         val = gits_read_vpendbaser(vlpi_base + GICR_VPENDBASER);
2082         val &= ~GICR_VPENDBASER_Valid;
2083         gits_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
2084
2085         do {
2086                 val = gits_read_vpendbaser(vlpi_base + GICR_VPENDBASER);
2087                 clean = !(val & GICR_VPENDBASER_Dirty);
2088                 if (!clean) {
2089                         count--;
2090                         cpu_relax();
2091                         udelay(1);
2092                 }
2093         } while (!clean && count);
2094
2095         return val;
2096 }
2097
2098 static void its_cpu_init_lpis(void)
2099 {
2100         void __iomem *rbase = gic_data_rdist_rd_base();
2101         struct page *pend_page;
2102         phys_addr_t paddr;
2103         u64 val, tmp;
2104
2105         if (gic_data_rdist()->lpi_enabled)
2106                 return;
2107
2108         val = readl_relaxed(rbase + GICR_CTLR);
2109         if ((gic_rdists->flags & RDIST_FLAGS_RD_TABLES_PREALLOCATED) &&
2110             (val & GICR_CTLR_ENABLE_LPIS)) {
2111                 /*
2112                  * Check that we get the same property table on all
2113                  * RDs. If we don't, this is hopeless.
2114                  */
2115                 paddr = gicr_read_propbaser(rbase + GICR_PROPBASER);
2116                 paddr &= GENMASK_ULL(51, 12);
2117                 if (WARN_ON(gic_rdists->prop_table_pa != paddr))
2118                         add_taint(TAINT_CRAP, LOCKDEP_STILL_OK);
2119
2120                 paddr = gicr_read_pendbaser(rbase + GICR_PENDBASER);
2121                 paddr &= GENMASK_ULL(51, 16);
2122
2123                 WARN_ON(!gic_check_reserved_range(paddr, LPI_PENDBASE_SZ));
2124                 its_free_pending_table(gic_data_rdist()->pend_page);
2125                 gic_data_rdist()->pend_page = NULL;
2126
2127                 goto out;
2128         }
2129
2130         pend_page = gic_data_rdist()->pend_page;
2131         paddr = page_to_phys(pend_page);
2132         WARN_ON(gic_reserve_range(paddr, LPI_PENDBASE_SZ));
2133
2134         /* set PROPBASE */
2135         val = (gic_rdists->prop_table_pa |
2136                GICR_PROPBASER_InnerShareable |
2137                GICR_PROPBASER_RaWaWb |
2138                ((LPI_NRBITS - 1) & GICR_PROPBASER_IDBITS_MASK));
2139
2140         gicr_write_propbaser(val, rbase + GICR_PROPBASER);
2141         tmp = gicr_read_propbaser(rbase + GICR_PROPBASER);
2142
2143         if ((tmp ^ val) & GICR_PROPBASER_SHAREABILITY_MASK) {
2144                 if (!(tmp & GICR_PROPBASER_SHAREABILITY_MASK)) {
2145                         /*
2146                          * The HW reports non-shareable, we must
2147                          * remove the cacheability attributes as
2148                          * well.
2149                          */
2150                         val &= ~(GICR_PROPBASER_SHAREABILITY_MASK |
2151                                  GICR_PROPBASER_CACHEABILITY_MASK);
2152                         val |= GICR_PROPBASER_nC;
2153                         gicr_write_propbaser(val, rbase + GICR_PROPBASER);
2154                 }
2155                 pr_info_once("GIC: using cache flushing for LPI property table\n");
2156                 gic_rdists->flags |= RDIST_FLAGS_PROPBASE_NEEDS_FLUSHING;
2157         }
2158
2159         /* set PENDBASE */
2160         val = (page_to_phys(pend_page) |
2161                GICR_PENDBASER_InnerShareable |
2162                GICR_PENDBASER_RaWaWb);
2163
2164         gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
2165         tmp = gicr_read_pendbaser(rbase + GICR_PENDBASER);
2166
2167         if (!(tmp & GICR_PENDBASER_SHAREABILITY_MASK)) {
2168                 /*
2169                  * The HW reports non-shareable, we must remove the
2170                  * cacheability attributes as well.
2171                  */
2172                 val &= ~(GICR_PENDBASER_SHAREABILITY_MASK |
2173                          GICR_PENDBASER_CACHEABILITY_MASK);
2174                 val |= GICR_PENDBASER_nC;
2175                 gicr_write_pendbaser(val, rbase + GICR_PENDBASER);
2176         }
2177
2178         /* Enable LPIs */
2179         val = readl_relaxed(rbase + GICR_CTLR);
2180         val |= GICR_CTLR_ENABLE_LPIS;
2181         writel_relaxed(val, rbase + GICR_CTLR);
2182
2183         if (gic_rdists->has_vlpis) {
2184                 void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2185
2186                 /*
2187                  * It's possible for CPU to receive VLPIs before it is
2188                  * sheduled as a vPE, especially for the first CPU, and the
2189                  * VLPI with INTID larger than 2^(IDbits+1) will be considered
2190                  * as out of range and dropped by GIC.
2191                  * So we initialize IDbits to known value to avoid VLPI drop.
2192                  */
2193                 val = (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
2194                 pr_debug("GICv4: CPU%d: Init IDbits to 0x%llx for GICR_VPROPBASER\n",
2195                         smp_processor_id(), val);
2196                 gits_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2197
2198                 /*
2199                  * Also clear Valid bit of GICR_VPENDBASER, in case some
2200                  * ancient programming gets left in and has possibility of
2201                  * corrupting memory.
2202                  */
2203                 val = its_clear_vpend_valid(vlpi_base);
2204                 WARN_ON(val & GICR_VPENDBASER_Dirty);
2205         }
2206
2207         /* Make sure the GIC has seen the above */
2208         dsb(sy);
2209 out:
2210         gic_data_rdist()->lpi_enabled = true;
2211         pr_info("GICv3: CPU%d: using %s LPI pending table @%pa\n",
2212                 smp_processor_id(),
2213                 gic_data_rdist()->pend_page ? "allocated" : "reserved",
2214                 &paddr);
2215 }
2216
2217 static void its_cpu_init_collection(struct its_node *its)
2218 {
2219         int cpu = smp_processor_id();
2220         u64 target;
2221
2222         /* avoid cross node collections and its mapping */
2223         if (its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) {
2224                 struct device_node *cpu_node;
2225
2226                 cpu_node = of_get_cpu_node(cpu, NULL);
2227                 if (its->numa_node != NUMA_NO_NODE &&
2228                         its->numa_node != of_node_to_nid(cpu_node))
2229                         return;
2230         }
2231
2232         /*
2233          * We now have to bind each collection to its target
2234          * redistributor.
2235          */
2236         if (gic_read_typer(its->base + GITS_TYPER) & GITS_TYPER_PTA) {
2237                 /*
2238                  * This ITS wants the physical address of the
2239                  * redistributor.
2240                  */
2241                 target = gic_data_rdist()->phys_base;
2242         } else {
2243                 /* This ITS wants a linear CPU number. */
2244                 target = gic_read_typer(gic_data_rdist_rd_base() + GICR_TYPER);
2245                 target = GICR_TYPER_CPU_NUMBER(target) << 16;
2246         }
2247
2248         /* Perform collection mapping */
2249         its->collections[cpu].target_address = target;
2250         its->collections[cpu].col_id = cpu;
2251
2252         its_send_mapc(its, &its->collections[cpu], 1);
2253         its_send_invall(its, &its->collections[cpu]);
2254 }
2255
2256 static void its_cpu_init_collections(void)
2257 {
2258         struct its_node *its;
2259
2260         raw_spin_lock(&its_lock);
2261
2262         list_for_each_entry(its, &its_nodes, entry)
2263                 its_cpu_init_collection(its);
2264
2265         raw_spin_unlock(&its_lock);
2266 }
2267
2268 static struct its_device *its_find_device(struct its_node *its, u32 dev_id)
2269 {
2270         struct its_device *its_dev = NULL, *tmp;
2271         unsigned long flags;
2272
2273         raw_spin_lock_irqsave(&its->lock, flags);
2274
2275         list_for_each_entry(tmp, &its->its_device_list, entry) {
2276                 if (tmp->device_id == dev_id) {
2277                         its_dev = tmp;
2278                         break;
2279                 }
2280         }
2281
2282         raw_spin_unlock_irqrestore(&its->lock, flags);
2283
2284         return its_dev;
2285 }
2286
2287 static struct its_baser *its_get_baser(struct its_node *its, u32 type)
2288 {
2289         int i;
2290
2291         for (i = 0; i < GITS_BASER_NR_REGS; i++) {
2292                 if (GITS_BASER_TYPE(its->tables[i].val) == type)
2293                         return &its->tables[i];
2294         }
2295
2296         return NULL;
2297 }
2298
2299 static bool its_alloc_table_entry(struct its_node *its,
2300                                   struct its_baser *baser, u32 id)
2301 {
2302         struct page *page;
2303         u32 esz, idx;
2304         __le64 *table;
2305
2306         /* Don't allow device id that exceeds single, flat table limit */
2307         esz = GITS_BASER_ENTRY_SIZE(baser->val);
2308         if (!(baser->val & GITS_BASER_INDIRECT))
2309                 return (id < (PAGE_ORDER_TO_SIZE(baser->order) / esz));
2310
2311         /* Compute 1st level table index & check if that exceeds table limit */
2312         idx = id >> ilog2(baser->psz / esz);
2313         if (idx >= (PAGE_ORDER_TO_SIZE(baser->order) / GITS_LVL1_ENTRY_SIZE))
2314                 return false;
2315
2316         table = baser->base;
2317
2318         /* Allocate memory for 2nd level table */
2319         if (!table[idx]) {
2320                 page = alloc_pages_node(its->numa_node, GFP_KERNEL | __GFP_ZERO,
2321                                         get_order(baser->psz));
2322                 if (!page)
2323                         return false;
2324
2325                 /* Flush Lvl2 table to PoC if hw doesn't support coherency */
2326                 if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
2327                         gic_flush_dcache_to_poc(page_address(page), baser->psz);
2328
2329                 table[idx] = cpu_to_le64(page_to_phys(page) | GITS_BASER_VALID);
2330
2331                 /* Flush Lvl1 entry to PoC if hw doesn't support coherency */
2332                 if (!(baser->val & GITS_BASER_SHAREABILITY_MASK))
2333                         gic_flush_dcache_to_poc(table + idx, GITS_LVL1_ENTRY_SIZE);
2334
2335                 /* Ensure updated table contents are visible to ITS hardware */
2336                 dsb(sy);
2337         }
2338
2339         return true;
2340 }
2341
2342 static bool its_alloc_device_table(struct its_node *its, u32 dev_id)
2343 {
2344         struct its_baser *baser;
2345
2346         baser = its_get_baser(its, GITS_BASER_TYPE_DEVICE);
2347
2348         /* Don't allow device id that exceeds ITS hardware limit */
2349         if (!baser)
2350                 return (ilog2(dev_id) < its->device_ids);
2351
2352         return its_alloc_table_entry(its, baser, dev_id);
2353 }
2354
2355 static bool its_alloc_vpe_table(u32 vpe_id)
2356 {
2357         struct its_node *its;
2358
2359         /*
2360          * Make sure the L2 tables are allocated on *all* v4 ITSs. We
2361          * could try and only do it on ITSs corresponding to devices
2362          * that have interrupts targeted at this VPE, but the
2363          * complexity becomes crazy (and you have tons of memory
2364          * anyway, right?).
2365          */
2366         list_for_each_entry(its, &its_nodes, entry) {
2367                 struct its_baser *baser;
2368
2369                 if (!its->is_v4)
2370                         continue;
2371
2372                 baser = its_get_baser(its, GITS_BASER_TYPE_VCPU);
2373                 if (!baser)
2374                         return false;
2375
2376                 if (!its_alloc_table_entry(its, baser, vpe_id))
2377                         return false;
2378         }
2379
2380         return true;
2381 }
2382
2383 static struct its_device *its_create_device(struct its_node *its, u32 dev_id,
2384                                             int nvecs, bool alloc_lpis)
2385 {
2386         struct its_device *dev;
2387         unsigned long *lpi_map = NULL;
2388         unsigned long flags;
2389         u16 *col_map = NULL;
2390         void *itt;
2391         int lpi_base;
2392         int nr_lpis;
2393         int nr_ites;
2394         int sz;
2395
2396         if (!its_alloc_device_table(its, dev_id))
2397                 return NULL;
2398
2399         if (WARN_ON(!is_power_of_2(nvecs)))
2400                 nvecs = roundup_pow_of_two(nvecs);
2401
2402         dev = kzalloc(sizeof(*dev), GFP_KERNEL);
2403         /*
2404          * Even if the device wants a single LPI, the ITT must be
2405          * sized as a power of two (and you need at least one bit...).
2406          */
2407         nr_ites = max(2, nvecs);
2408         sz = nr_ites * its->ite_size;
2409         sz = max(sz, ITS_ITT_ALIGN) + ITS_ITT_ALIGN - 1;
2410         itt = kzalloc_node(sz, GFP_KERNEL, its->numa_node);
2411         if (alloc_lpis) {
2412                 lpi_map = its_lpi_alloc(nvecs, &lpi_base, &nr_lpis);
2413                 if (lpi_map)
2414                         col_map = kcalloc(nr_lpis, sizeof(*col_map),
2415                                           GFP_KERNEL);
2416         } else {
2417                 col_map = kcalloc(nr_ites, sizeof(*col_map), GFP_KERNEL);
2418                 nr_lpis = 0;
2419                 lpi_base = 0;
2420         }
2421
2422         if (!dev || !itt ||  !col_map || (!lpi_map && alloc_lpis)) {
2423                 kfree(dev);
2424                 kfree(itt);
2425                 kfree(lpi_map);
2426                 kfree(col_map);
2427                 return NULL;
2428         }
2429
2430         gic_flush_dcache_to_poc(itt, sz);
2431
2432         dev->its = its;
2433         dev->itt = itt;
2434         dev->nr_ites = nr_ites;
2435         dev->event_map.lpi_map = lpi_map;
2436         dev->event_map.col_map = col_map;
2437         dev->event_map.lpi_base = lpi_base;
2438         dev->event_map.nr_lpis = nr_lpis;
2439         mutex_init(&dev->event_map.vlpi_lock);
2440         dev->device_id = dev_id;
2441         INIT_LIST_HEAD(&dev->entry);
2442
2443         raw_spin_lock_irqsave(&its->lock, flags);
2444         list_add(&dev->entry, &its->its_device_list);
2445         raw_spin_unlock_irqrestore(&its->lock, flags);
2446
2447         /* Map device to its ITT */
2448         its_send_mapd(dev, 1);
2449
2450         return dev;
2451 }
2452
2453 static void its_free_device(struct its_device *its_dev)
2454 {
2455         unsigned long flags;
2456
2457         raw_spin_lock_irqsave(&its_dev->its->lock, flags);
2458         list_del(&its_dev->entry);
2459         raw_spin_unlock_irqrestore(&its_dev->its->lock, flags);
2460         kfree(its_dev->itt);
2461         kfree(its_dev);
2462 }
2463
2464 static int its_alloc_device_irq(struct its_device *dev, int nvecs, irq_hw_number_t *hwirq)
2465 {
2466         int idx;
2467
2468         idx = bitmap_find_free_region(dev->event_map.lpi_map,
2469                                       dev->event_map.nr_lpis,
2470                                       get_count_order(nvecs));
2471         if (idx < 0)
2472                 return -ENOSPC;
2473
2474         *hwirq = dev->event_map.lpi_base + idx;
2475         set_bit(idx, dev->event_map.lpi_map);
2476
2477         return 0;
2478 }
2479
2480 static int its_msi_prepare(struct irq_domain *domain, struct device *dev,
2481                            int nvec, msi_alloc_info_t *info)
2482 {
2483         struct its_node *its;
2484         struct its_device *its_dev;
2485         struct msi_domain_info *msi_info;
2486         u32 dev_id;
2487         int err = 0;
2488
2489         /*
2490          * We ignore "dev" entierely, and rely on the dev_id that has
2491          * been passed via the scratchpad. This limits this domain's
2492          * usefulness to upper layers that definitely know that they
2493          * are built on top of the ITS.
2494          */
2495         dev_id = info->scratchpad[0].ul;
2496
2497         msi_info = msi_get_domain_info(domain);
2498         its = msi_info->data;
2499
2500         if (!gic_rdists->has_direct_lpi &&
2501             vpe_proxy.dev &&
2502             vpe_proxy.dev->its == its &&
2503             dev_id == vpe_proxy.dev->device_id) {
2504                 /* Bad luck. Get yourself a better implementation */
2505                 WARN_ONCE(1, "DevId %x clashes with GICv4 VPE proxy device\n",
2506                           dev_id);
2507                 return -EINVAL;
2508         }
2509
2510         mutex_lock(&its->dev_alloc_lock);
2511         its_dev = its_find_device(its, dev_id);
2512         if (its_dev) {
2513                 /*
2514                  * We already have seen this ID, probably through
2515                  * another alias (PCI bridge of some sort). No need to
2516                  * create the device.
2517                  */
2518                 its_dev->shared = true;
2519                 pr_debug("Reusing ITT for devID %x\n", dev_id);
2520                 goto out;
2521         }
2522
2523         its_dev = its_create_device(its, dev_id, nvec, true);
2524         if (!its_dev) {
2525                 err = -ENOMEM;
2526                 goto out;
2527         }
2528
2529         pr_debug("ITT %d entries, %d bits\n", nvec, ilog2(nvec));
2530 out:
2531         mutex_unlock(&its->dev_alloc_lock);
2532         info->scratchpad[0].ptr = its_dev;
2533         return err;
2534 }
2535
2536 static struct msi_domain_ops its_msi_domain_ops = {
2537         .msi_prepare    = its_msi_prepare,
2538 };
2539
2540 static int its_irq_gic_domain_alloc(struct irq_domain *domain,
2541                                     unsigned int virq,
2542                                     irq_hw_number_t hwirq)
2543 {
2544         struct irq_fwspec fwspec;
2545
2546         if (irq_domain_get_of_node(domain->parent)) {
2547                 fwspec.fwnode = domain->parent->fwnode;
2548                 fwspec.param_count = 3;
2549                 fwspec.param[0] = GIC_IRQ_TYPE_LPI;
2550                 fwspec.param[1] = hwirq;
2551                 fwspec.param[2] = IRQ_TYPE_EDGE_RISING;
2552         } else if (is_fwnode_irqchip(domain->parent->fwnode)) {
2553                 fwspec.fwnode = domain->parent->fwnode;
2554                 fwspec.param_count = 2;
2555                 fwspec.param[0] = hwirq;
2556                 fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
2557         } else {
2558                 return -EINVAL;
2559         }
2560
2561         return irq_domain_alloc_irqs_parent(domain, virq, 1, &fwspec);
2562 }
2563
2564 static int its_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
2565                                 unsigned int nr_irqs, void *args)
2566 {
2567         msi_alloc_info_t *info = args;
2568         struct its_device *its_dev = info->scratchpad[0].ptr;
2569         irq_hw_number_t hwirq;
2570         int err;
2571         int i;
2572
2573         err = its_alloc_device_irq(its_dev, nr_irqs, &hwirq);
2574         if (err)
2575                 return err;
2576
2577         for (i = 0; i < nr_irqs; i++) {
2578                 err = its_irq_gic_domain_alloc(domain, virq + i, hwirq + i);
2579                 if (err)
2580                         return err;
2581
2582                 irq_domain_set_hwirq_and_chip(domain, virq + i,
2583                                               hwirq + i, &its_irq_chip, its_dev);
2584                 irqd_set_single_target(irq_desc_get_irq_data(irq_to_desc(virq + i)));
2585                 pr_debug("ID:%d pID:%d vID:%d\n",
2586                          (int)(hwirq + i - its_dev->event_map.lpi_base),
2587                          (int)(hwirq + i), virq + i);
2588         }
2589
2590         return 0;
2591 }
2592
2593 static int its_irq_domain_activate(struct irq_domain *domain,
2594                                    struct irq_data *d, bool reserve)
2595 {
2596         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2597         u32 event = its_get_event_id(d);
2598         const struct cpumask *cpu_mask = cpu_online_mask;
2599         int cpu;
2600
2601         /* get the cpu_mask of local node */
2602         if (its_dev->its->numa_node >= 0)
2603                 cpu_mask = cpumask_of_node(its_dev->its->numa_node);
2604
2605         /* Bind the LPI to the first possible CPU */
2606         cpu = cpumask_first_and(cpu_mask, cpu_online_mask);
2607         if (cpu >= nr_cpu_ids) {
2608                 if (its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144)
2609                         return -EINVAL;
2610
2611                 cpu = cpumask_first(cpu_online_mask);
2612         }
2613
2614         its_dev->event_map.col_map[event] = cpu;
2615         irq_data_update_effective_affinity(d, cpumask_of(cpu));
2616
2617         /* Map the GIC IRQ and event to the device */
2618         its_send_mapti(its_dev, d->hwirq, event);
2619         return 0;
2620 }
2621
2622 static void its_irq_domain_deactivate(struct irq_domain *domain,
2623                                       struct irq_data *d)
2624 {
2625         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2626         u32 event = its_get_event_id(d);
2627
2628         /* Stop the delivery of interrupts */
2629         its_send_discard(its_dev, event);
2630 }
2631
2632 static void its_irq_domain_free(struct irq_domain *domain, unsigned int virq,
2633                                 unsigned int nr_irqs)
2634 {
2635         struct irq_data *d = irq_domain_get_irq_data(domain, virq);
2636         struct its_device *its_dev = irq_data_get_irq_chip_data(d);
2637         struct its_node *its = its_dev->its;
2638         int i;
2639
2640         for (i = 0; i < nr_irqs; i++) {
2641                 struct irq_data *data = irq_domain_get_irq_data(domain,
2642                                                                 virq + i);
2643                 u32 event = its_get_event_id(data);
2644
2645                 /* Mark interrupt index as unused */
2646                 clear_bit(event, its_dev->event_map.lpi_map);
2647
2648                 /* Nuke the entry in the domain */
2649                 irq_domain_reset_irq_data(data);
2650         }
2651
2652         mutex_lock(&its->dev_alloc_lock);
2653
2654         /*
2655          * If all interrupts have been freed, start mopping the
2656          * floor. This is conditionned on the device not being shared.
2657          */
2658         if (!its_dev->shared &&
2659             bitmap_empty(its_dev->event_map.lpi_map,
2660                          its_dev->event_map.nr_lpis)) {
2661                 its_lpi_free(its_dev->event_map.lpi_map,
2662                              its_dev->event_map.lpi_base,
2663                              its_dev->event_map.nr_lpis);
2664                 kfree(its_dev->event_map.col_map);
2665
2666                 /* Unmap device/itt */
2667                 its_send_mapd(its_dev, 0);
2668                 its_free_device(its_dev);
2669         }
2670
2671         mutex_unlock(&its->dev_alloc_lock);
2672
2673         irq_domain_free_irqs_parent(domain, virq, nr_irqs);
2674 }
2675
2676 static const struct irq_domain_ops its_domain_ops = {
2677         .alloc                  = its_irq_domain_alloc,
2678         .free                   = its_irq_domain_free,
2679         .activate               = its_irq_domain_activate,
2680         .deactivate             = its_irq_domain_deactivate,
2681 };
2682
2683 /*
2684  * This is insane.
2685  *
2686  * If a GICv4 doesn't implement Direct LPIs (which is extremely
2687  * likely), the only way to perform an invalidate is to use a fake
2688  * device to issue an INV command, implying that the LPI has first
2689  * been mapped to some event on that device. Since this is not exactly
2690  * cheap, we try to keep that mapping around as long as possible, and
2691  * only issue an UNMAP if we're short on available slots.
2692  *
2693  * Broken by design(tm).
2694  */
2695 static void its_vpe_db_proxy_unmap_locked(struct its_vpe *vpe)
2696 {
2697         /* Already unmapped? */
2698         if (vpe->vpe_proxy_event == -1)
2699                 return;
2700
2701         its_send_discard(vpe_proxy.dev, vpe->vpe_proxy_event);
2702         vpe_proxy.vpes[vpe->vpe_proxy_event] = NULL;
2703
2704         /*
2705          * We don't track empty slots at all, so let's move the
2706          * next_victim pointer if we can quickly reuse that slot
2707          * instead of nuking an existing entry. Not clear that this is
2708          * always a win though, and this might just generate a ripple
2709          * effect... Let's just hope VPEs don't migrate too often.
2710          */
2711         if (vpe_proxy.vpes[vpe_proxy.next_victim])
2712                 vpe_proxy.next_victim = vpe->vpe_proxy_event;
2713
2714         vpe->vpe_proxy_event = -1;
2715 }
2716
2717 static void its_vpe_db_proxy_unmap(struct its_vpe *vpe)
2718 {
2719         if (!gic_rdists->has_direct_lpi) {
2720                 unsigned long flags;
2721
2722                 raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2723                 its_vpe_db_proxy_unmap_locked(vpe);
2724                 raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2725         }
2726 }
2727
2728 static void its_vpe_db_proxy_map_locked(struct its_vpe *vpe)
2729 {
2730         /* Already mapped? */
2731         if (vpe->vpe_proxy_event != -1)
2732                 return;
2733
2734         /* This slot was already allocated. Kick the other VPE out. */
2735         if (vpe_proxy.vpes[vpe_proxy.next_victim])
2736                 its_vpe_db_proxy_unmap_locked(vpe_proxy.vpes[vpe_proxy.next_victim]);
2737
2738         /* Map the new VPE instead */
2739         vpe_proxy.vpes[vpe_proxy.next_victim] = vpe;
2740         vpe->vpe_proxy_event = vpe_proxy.next_victim;
2741         vpe_proxy.next_victim = (vpe_proxy.next_victim + 1) % vpe_proxy.dev->nr_ites;
2742
2743         vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = vpe->col_idx;
2744         its_send_mapti(vpe_proxy.dev, vpe->vpe_db_lpi, vpe->vpe_proxy_event);
2745 }
2746
2747 static void its_vpe_db_proxy_move(struct its_vpe *vpe, int from, int to)
2748 {
2749         unsigned long flags;
2750         struct its_collection *target_col;
2751
2752         if (gic_rdists->has_direct_lpi) {
2753                 void __iomem *rdbase;
2754
2755                 rdbase = per_cpu_ptr(gic_rdists->rdist, from)->rd_base;
2756                 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
2757                 while (gic_read_lpir(rdbase + GICR_SYNCR) & 1)
2758                         cpu_relax();
2759
2760                 return;
2761         }
2762
2763         raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2764
2765         its_vpe_db_proxy_map_locked(vpe);
2766
2767         target_col = &vpe_proxy.dev->its->collections[to];
2768         its_send_movi(vpe_proxy.dev, target_col, vpe->vpe_proxy_event);
2769         vpe_proxy.dev->event_map.col_map[vpe->vpe_proxy_event] = to;
2770
2771         raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2772 }
2773
2774 static int its_vpe_set_affinity(struct irq_data *d,
2775                                 const struct cpumask *mask_val,
2776                                 bool force)
2777 {
2778         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2779         int cpu = cpumask_first(mask_val);
2780
2781         /*
2782          * Changing affinity is mega expensive, so let's be as lazy as
2783          * we can and only do it if we really have to. Also, if mapped
2784          * into the proxy device, we need to move the doorbell
2785          * interrupt to its new location.
2786          */
2787         if (vpe->col_idx != cpu) {
2788                 int from = vpe->col_idx;
2789
2790                 vpe->col_idx = cpu;
2791                 its_send_vmovp(vpe);
2792                 its_vpe_db_proxy_move(vpe, from, cpu);
2793         }
2794
2795         irq_data_update_effective_affinity(d, cpumask_of(cpu));
2796
2797         return IRQ_SET_MASK_OK_DONE;
2798 }
2799
2800 static void its_vpe_schedule(struct its_vpe *vpe)
2801 {
2802         void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2803         u64 val;
2804
2805         /* Schedule the VPE */
2806         val  = virt_to_phys(page_address(vpe->its_vm->vprop_page)) &
2807                 GENMASK_ULL(51, 12);
2808         val |= (LPI_NRBITS - 1) & GICR_VPROPBASER_IDBITS_MASK;
2809         val |= GICR_VPROPBASER_RaWb;
2810         val |= GICR_VPROPBASER_InnerShareable;
2811         gits_write_vpropbaser(val, vlpi_base + GICR_VPROPBASER);
2812
2813         val  = virt_to_phys(page_address(vpe->vpt_page)) &
2814                 GENMASK_ULL(51, 16);
2815         val |= GICR_VPENDBASER_RaWaWb;
2816         val |= GICR_VPENDBASER_NonShareable;
2817         /*
2818          * There is no good way of finding out if the pending table is
2819          * empty as we can race against the doorbell interrupt very
2820          * easily. So in the end, vpe->pending_last is only an
2821          * indication that the vcpu has something pending, not one
2822          * that the pending table is empty. A good implementation
2823          * would be able to read its coarse map pretty quickly anyway,
2824          * making this a tolerable issue.
2825          */
2826         val |= GICR_VPENDBASER_PendingLast;
2827         val |= vpe->idai ? GICR_VPENDBASER_IDAI : 0;
2828         val |= GICR_VPENDBASER_Valid;
2829         gits_write_vpendbaser(val, vlpi_base + GICR_VPENDBASER);
2830 }
2831
2832 static void its_vpe_deschedule(struct its_vpe *vpe)
2833 {
2834         void __iomem *vlpi_base = gic_data_rdist_vlpi_base();
2835         u64 val;
2836
2837         val = its_clear_vpend_valid(vlpi_base);
2838
2839         if (unlikely(val & GICR_VPENDBASER_Dirty)) {
2840                 pr_err_ratelimited("ITS virtual pending table not cleaning\n");
2841                 vpe->idai = false;
2842                 vpe->pending_last = true;
2843         } else {
2844                 vpe->idai = !!(val & GICR_VPENDBASER_IDAI);
2845                 vpe->pending_last = !!(val & GICR_VPENDBASER_PendingLast);
2846         }
2847 }
2848
2849 static void its_vpe_invall(struct its_vpe *vpe)
2850 {
2851         struct its_node *its;
2852
2853         list_for_each_entry(its, &its_nodes, entry) {
2854                 if (!its->is_v4)
2855                         continue;
2856
2857                 if (its_list_map && !vpe->its_vm->vlpi_count[its->list_nr])
2858                         continue;
2859
2860                 /*
2861                  * Sending a VINVALL to a single ITS is enough, as all
2862                  * we need is to reach the redistributors.
2863                  */
2864                 its_send_vinvall(its, vpe);
2865                 return;
2866         }
2867 }
2868
2869 static int its_vpe_set_vcpu_affinity(struct irq_data *d, void *vcpu_info)
2870 {
2871         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2872         struct its_cmd_info *info = vcpu_info;
2873
2874         switch (info->cmd_type) {
2875         case SCHEDULE_VPE:
2876                 its_vpe_schedule(vpe);
2877                 return 0;
2878
2879         case DESCHEDULE_VPE:
2880                 its_vpe_deschedule(vpe);
2881                 return 0;
2882
2883         case INVALL_VPE:
2884                 its_vpe_invall(vpe);
2885                 return 0;
2886
2887         default:
2888                 return -EINVAL;
2889         }
2890 }
2891
2892 static void its_vpe_send_cmd(struct its_vpe *vpe,
2893                              void (*cmd)(struct its_device *, u32))
2894 {
2895         unsigned long flags;
2896
2897         raw_spin_lock_irqsave(&vpe_proxy.lock, flags);
2898
2899         its_vpe_db_proxy_map_locked(vpe);
2900         cmd(vpe_proxy.dev, vpe->vpe_proxy_event);
2901
2902         raw_spin_unlock_irqrestore(&vpe_proxy.lock, flags);
2903 }
2904
2905 static void its_vpe_send_inv(struct irq_data *d)
2906 {
2907         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2908
2909         if (gic_rdists->has_direct_lpi) {
2910                 void __iomem *rdbase;
2911
2912                 rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
2913                 gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_INVLPIR);
2914                 while (gic_read_lpir(rdbase + GICR_SYNCR) & 1)
2915                         cpu_relax();
2916         } else {
2917                 its_vpe_send_cmd(vpe, its_send_inv);
2918         }
2919 }
2920
2921 static void its_vpe_mask_irq(struct irq_data *d)
2922 {
2923         /*
2924          * We need to unmask the LPI, which is described by the parent
2925          * irq_data. Instead of calling into the parent (which won't
2926          * exactly do the right thing, let's simply use the
2927          * parent_data pointer. Yes, I'm naughty.
2928          */
2929         lpi_write_config(d->parent_data, LPI_PROP_ENABLED, 0);
2930         its_vpe_send_inv(d);
2931 }
2932
2933 static void its_vpe_unmask_irq(struct irq_data *d)
2934 {
2935         /* Same hack as above... */
2936         lpi_write_config(d->parent_data, 0, LPI_PROP_ENABLED);
2937         its_vpe_send_inv(d);
2938 }
2939
2940 static int its_vpe_set_irqchip_state(struct irq_data *d,
2941                                      enum irqchip_irq_state which,
2942                                      bool state)
2943 {
2944         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
2945
2946         if (which != IRQCHIP_STATE_PENDING)
2947                 return -EINVAL;
2948
2949         if (gic_rdists->has_direct_lpi) {
2950                 void __iomem *rdbase;
2951
2952                 rdbase = per_cpu_ptr(gic_rdists->rdist, vpe->col_idx)->rd_base;
2953                 if (state) {
2954                         gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_SETLPIR);
2955                 } else {
2956                         gic_write_lpir(vpe->vpe_db_lpi, rdbase + GICR_CLRLPIR);
2957                         while (gic_read_lpir(rdbase + GICR_SYNCR) & 1)
2958                                 cpu_relax();
2959                 }
2960         } else {
2961                 if (state)
2962                         its_vpe_send_cmd(vpe, its_send_int);
2963                 else
2964                         its_vpe_send_cmd(vpe, its_send_clear);
2965         }
2966
2967         return 0;
2968 }
2969
2970 static struct irq_chip its_vpe_irq_chip = {
2971         .name                   = "GICv4-vpe",
2972         .irq_mask               = its_vpe_mask_irq,
2973         .irq_unmask             = its_vpe_unmask_irq,
2974         .irq_eoi                = irq_chip_eoi_parent,
2975         .irq_set_affinity       = its_vpe_set_affinity,
2976         .irq_set_irqchip_state  = its_vpe_set_irqchip_state,
2977         .irq_set_vcpu_affinity  = its_vpe_set_vcpu_affinity,
2978 };
2979
2980 static int its_vpe_id_alloc(void)
2981 {
2982         return ida_simple_get(&its_vpeid_ida, 0, ITS_MAX_VPEID, GFP_KERNEL);
2983 }
2984
2985 static void its_vpe_id_free(u16 id)
2986 {
2987         ida_simple_remove(&its_vpeid_ida, id);
2988 }
2989
2990 static int its_vpe_init(struct its_vpe *vpe)
2991 {
2992         struct page *vpt_page;
2993         int vpe_id;
2994
2995         /* Allocate vpe_id */
2996         vpe_id = its_vpe_id_alloc();
2997         if (vpe_id < 0)
2998                 return vpe_id;
2999
3000         /* Allocate VPT */
3001         vpt_page = its_allocate_pending_table(GFP_KERNEL);
3002         if (!vpt_page) {
3003                 its_vpe_id_free(vpe_id);
3004                 return -ENOMEM;
3005         }
3006
3007         if (!its_alloc_vpe_table(vpe_id)) {
3008                 its_vpe_id_free(vpe_id);
3009                 its_free_pending_table(vpe->vpt_page);
3010                 return -ENOMEM;
3011         }
3012
3013         vpe->vpe_id = vpe_id;
3014         vpe->vpt_page = vpt_page;
3015         vpe->vpe_proxy_event = -1;
3016
3017         return 0;
3018 }
3019
3020 static void its_vpe_teardown(struct its_vpe *vpe)
3021 {
3022         its_vpe_db_proxy_unmap(vpe);
3023         its_vpe_id_free(vpe->vpe_id);
3024         its_free_pending_table(vpe->vpt_page);
3025 }
3026
3027 static void its_vpe_irq_domain_free(struct irq_domain *domain,
3028                                     unsigned int virq,
3029                                     unsigned int nr_irqs)
3030 {
3031         struct its_vm *vm = domain->host_data;
3032         int i;
3033
3034         irq_domain_free_irqs_parent(domain, virq, nr_irqs);
3035
3036         for (i = 0; i < nr_irqs; i++) {
3037                 struct irq_data *data = irq_domain_get_irq_data(domain,
3038                                                                 virq + i);
3039                 struct its_vpe *vpe = irq_data_get_irq_chip_data(data);
3040
3041                 BUG_ON(vm != vpe->its_vm);
3042
3043                 clear_bit(data->hwirq, vm->db_bitmap);
3044                 its_vpe_teardown(vpe);
3045                 irq_domain_reset_irq_data(data);
3046         }
3047
3048         if (bitmap_empty(vm->db_bitmap, vm->nr_db_lpis)) {
3049                 its_lpi_free(vm->db_bitmap, vm->db_lpi_base, vm->nr_db_lpis);
3050                 its_free_prop_table(vm->vprop_page);
3051         }
3052 }
3053
3054 static int its_vpe_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
3055                                     unsigned int nr_irqs, void *args)
3056 {
3057         struct its_vm *vm = args;
3058         unsigned long *bitmap;
3059         struct page *vprop_page;
3060         int base, nr_ids, i, err = 0;
3061
3062         BUG_ON(!vm);
3063
3064         bitmap = its_lpi_alloc(roundup_pow_of_two(nr_irqs), &base, &nr_ids);
3065         if (!bitmap)
3066                 return -ENOMEM;
3067
3068         if (nr_ids < nr_irqs) {
3069                 its_lpi_free(bitmap, base, nr_ids);
3070                 return -ENOMEM;
3071         }
3072
3073         vprop_page = its_allocate_prop_table(GFP_KERNEL);
3074         if (!vprop_page) {
3075                 its_lpi_free(bitmap, base, nr_ids);
3076                 return -ENOMEM;
3077         }
3078
3079         vm->db_bitmap = bitmap;
3080         vm->db_lpi_base = base;
3081         vm->nr_db_lpis = nr_ids;
3082         vm->vprop_page = vprop_page;
3083
3084         for (i = 0; i < nr_irqs; i++) {
3085                 vm->vpes[i]->vpe_db_lpi = base + i;
3086                 err = its_vpe_init(vm->vpes[i]);
3087                 if (err)
3088                         break;
3089                 err = its_irq_gic_domain_alloc(domain, virq + i,
3090                                                vm->vpes[i]->vpe_db_lpi);
3091                 if (err)
3092                         break;
3093                 irq_domain_set_hwirq_and_chip(domain, virq + i, i,
3094                                               &its_vpe_irq_chip, vm->vpes[i]);
3095                 set_bit(i, bitmap);
3096         }
3097
3098         if (err) {
3099                 if (i > 0)
3100                         its_vpe_irq_domain_free(domain, virq, i - 1);
3101
3102                 its_lpi_free(bitmap, base, nr_ids);
3103                 its_free_prop_table(vprop_page);
3104         }
3105
3106         return err;
3107 }
3108
3109 static int its_vpe_irq_domain_activate(struct irq_domain *domain,
3110                                        struct irq_data *d, bool reserve)
3111 {
3112         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3113         struct its_node *its;
3114
3115         /* If we use the list map, we issue VMAPP on demand... */
3116         if (its_list_map)
3117                 return 0;
3118
3119         /* Map the VPE to the first possible CPU */
3120         vpe->col_idx = cpumask_first(cpu_online_mask);
3121
3122         list_for_each_entry(its, &its_nodes, entry) {
3123                 if (!its->is_v4)
3124                         continue;
3125
3126                 its_send_vmapp(its, vpe, true);
3127                 its_send_vinvall(its, vpe);
3128         }
3129
3130         irq_data_update_effective_affinity(d, cpumask_of(vpe->col_idx));
3131
3132         return 0;
3133 }
3134
3135 static void its_vpe_irq_domain_deactivate(struct irq_domain *domain,
3136                                           struct irq_data *d)
3137 {
3138         struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
3139         struct its_node *its;
3140
3141         /*
3142          * If we use the list map, we unmap the VPE once no VLPIs are
3143          * associated with the VM.
3144          */
3145         if (its_list_map)
3146                 return;
3147
3148         list_for_each_entry(its, &its_nodes, entry) {
3149                 if (!its->is_v4)
3150                         continue;
3151
3152                 its_send_vmapp(its, vpe, false);
3153         }
3154 }
3155
3156 static const struct irq_domain_ops its_vpe_domain_ops = {
3157         .alloc                  = its_vpe_irq_domain_alloc,
3158         .free                   = its_vpe_irq_domain_free,
3159         .activate               = its_vpe_irq_domain_activate,
3160         .deactivate             = its_vpe_irq_domain_deactivate,
3161 };
3162
3163 static int its_force_quiescent(void __iomem *base)
3164 {
3165         u32 count = 1000000;    /* 1s */
3166         u32 val;
3167
3168         val = readl_relaxed(base + GITS_CTLR);
3169         /*
3170          * GIC architecture specification requires the ITS to be both
3171          * disabled and quiescent for writes to GITS_BASER<n> or
3172          * GITS_CBASER to not have UNPREDICTABLE results.
3173          */
3174         if ((val & GITS_CTLR_QUIESCENT) && !(val & GITS_CTLR_ENABLE))
3175                 return 0;
3176
3177         /* Disable the generation of all interrupts to this ITS */
3178         val &= ~(GITS_CTLR_ENABLE | GITS_CTLR_ImDe);
3179         writel_relaxed(val, base + GITS_CTLR);
3180
3181         /* Poll GITS_CTLR and wait until ITS becomes quiescent */
3182         while (1) {
3183                 val = readl_relaxed(base + GITS_CTLR);
3184                 if (val & GITS_CTLR_QUIESCENT)
3185                         return 0;
3186
3187                 count--;
3188                 if (!count)
3189                         return -EBUSY;
3190
3191                 cpu_relax();
3192                 udelay(1);
3193         }
3194 }
3195
3196 static bool __maybe_unused its_enable_quirk_cavium_22375(void *data)
3197 {
3198         struct its_node *its = data;
3199
3200         /* erratum 22375: only alloc 8MB table size */
3201         its->device_ids = 0x14;         /* 20 bits, 8MB */
3202         its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_22375;
3203
3204         return true;
3205 }
3206
3207 static bool __maybe_unused its_enable_quirk_cavium_23144(void *data)
3208 {
3209         struct its_node *its = data;
3210
3211         its->flags |= ITS_FLAGS_WORKAROUND_CAVIUM_23144;
3212
3213         return true;
3214 }
3215
3216 static bool __maybe_unused its_enable_quirk_qdf2400_e0065(void *data)
3217 {
3218         struct its_node *its = data;
3219
3220         /* On QDF2400, the size of the ITE is 16Bytes */
3221         its->ite_size = 16;
3222
3223         return true;
3224 }
3225
3226 static u64 its_irq_get_msi_base_pre_its(struct its_device *its_dev)
3227 {
3228         struct its_node *its = its_dev->its;
3229
3230         /*
3231          * The Socionext Synquacer SoC has a so-called 'pre-ITS',
3232          * which maps 32-bit writes targeted at a separate window of
3233          * size '4 << device_id_bits' onto writes to GITS_TRANSLATER
3234          * with device ID taken from bits [device_id_bits + 1:2] of
3235          * the window offset.
3236          */
3237         return its->pre_its_base + (its_dev->device_id << 2);
3238 }
3239
3240 static bool __maybe_unused its_enable_quirk_socionext_synquacer(void *data)
3241 {
3242         struct its_node *its = data;
3243         u32 pre_its_window[2];
3244         u32 ids;
3245
3246         if (!fwnode_property_read_u32_array(its->fwnode_handle,
3247                                            "socionext,synquacer-pre-its",
3248                                            pre_its_window,
3249                                            ARRAY_SIZE(pre_its_window))) {
3250
3251                 its->pre_its_base = pre_its_window[0];
3252                 its->get_msi_base = its_irq_get_msi_base_pre_its;
3253
3254                 ids = ilog2(pre_its_window[1]) - 2;
3255                 if (its->device_ids > ids)
3256                         its->device_ids = ids;
3257
3258                 /* the pre-ITS breaks isolation, so disable MSI remapping */
3259                 its->msi_domain_flags &= ~IRQ_DOMAIN_FLAG_MSI_REMAP;
3260                 return true;
3261         }
3262         return false;
3263 }
3264
3265 static bool __maybe_unused its_enable_quirk_hip07_161600802(void *data)
3266 {
3267         struct its_node *its = data;
3268
3269         /*
3270          * Hip07 insists on using the wrong address for the VLPI
3271          * page. Trick it into doing the right thing...
3272          */
3273         its->vlpi_redist_offset = SZ_128K;
3274         return true;
3275 }
3276
3277 static const struct gic_quirk its_quirks[] = {
3278 #ifdef CONFIG_CAVIUM_ERRATUM_22375
3279         {
3280                 .desc   = "ITS: Cavium errata 22375, 24313",
3281                 .iidr   = 0xa100034c,   /* ThunderX pass 1.x */
3282                 .mask   = 0xffff0fff,
3283                 .init   = its_enable_quirk_cavium_22375,
3284         },
3285 #endif
3286 #ifdef CONFIG_CAVIUM_ERRATUM_23144
3287         {
3288                 .desc   = "ITS: Cavium erratum 23144",
3289                 .iidr   = 0xa100034c,   /* ThunderX pass 1.x */
3290                 .mask   = 0xffff0fff,
3291                 .init   = its_enable_quirk_cavium_23144,
3292         },
3293 #endif
3294 #ifdef CONFIG_QCOM_QDF2400_ERRATUM_0065
3295         {
3296                 .desc   = "ITS: QDF2400 erratum 0065",
3297                 .iidr   = 0x00001070, /* QDF2400 ITS rev 1.x */
3298                 .mask   = 0xffffffff,
3299                 .init   = its_enable_quirk_qdf2400_e0065,
3300         },
3301 #endif
3302 #ifdef CONFIG_SOCIONEXT_SYNQUACER_PREITS
3303         {
3304                 /*
3305                  * The Socionext Synquacer SoC incorporates ARM's own GIC-500
3306                  * implementation, but with a 'pre-ITS' added that requires
3307                  * special handling in software.
3308                  */
3309                 .desc   = "ITS: Socionext Synquacer pre-ITS",
3310                 .iidr   = 0x0001143b,
3311                 .mask   = 0xffffffff,
3312                 .init   = its_enable_quirk_socionext_synquacer,
3313         },
3314 #endif
3315 #ifdef CONFIG_HISILICON_ERRATUM_161600802
3316         {
3317                 .desc   = "ITS: Hip07 erratum 161600802",
3318                 .iidr   = 0x00000004,
3319                 .mask   = 0xffffffff,
3320                 .init   = its_enable_quirk_hip07_161600802,
3321         },
3322 #endif
3323         {
3324         }
3325 };
3326
3327 static void its_enable_quirks(struct its_node *its)
3328 {
3329         u32 iidr = readl_relaxed(its->base + GITS_IIDR);
3330
3331         gic_enable_quirks(iidr, its_quirks, its);
3332 }
3333
3334 static int its_save_disable(void)
3335 {
3336         struct its_node *its;
3337         int err = 0;
3338
3339         raw_spin_lock(&its_lock);
3340         list_for_each_entry(its, &its_nodes, entry) {
3341                 void __iomem *base;
3342
3343                 if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3344                         continue;
3345
3346                 base = its->base;
3347                 its->ctlr_save = readl_relaxed(base + GITS_CTLR);
3348                 err = its_force_quiescent(base);
3349                 if (err) {
3350                         pr_err("ITS@%pa: failed to quiesce: %d\n",
3351                                &its->phys_base, err);
3352                         writel_relaxed(its->ctlr_save, base + GITS_CTLR);
3353                         goto err;
3354                 }
3355
3356                 its->cbaser_save = gits_read_cbaser(base + GITS_CBASER);
3357         }
3358
3359 err:
3360         if (err) {
3361                 list_for_each_entry_continue_reverse(its, &its_nodes, entry) {
3362                         void __iomem *base;
3363
3364                         if (!(its->flags & ITS_FLAGS_SAVE_SUSPEND_STATE))
3365                                 continue;
3366
3367                         base = its->base;
3368                         writel_relaxed(its->ctlr_sa