Loading...
1/*
2 * CCI cache coherent interconnect driver
3 *
4 * Copyright (C) 2013 ARM Ltd.
5 * Author: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
10 *
11 * This program is distributed "as is" WITHOUT ANY WARRANTY of any
12 * kind, whether express or implied; without even the implied warranty
13 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 */
16
17#include <linux/arm-cci.h>
18#include <linux/io.h>
19#include <linux/module.h>
20#include <linux/of_address.h>
21#include <linux/of_platform.h>
22#include <linux/platform_device.h>
23#include <linux/slab.h>
24
25#include <asm/cacheflush.h>
26#include <asm/smp_plat.h>
27
28static void __iomem *cci_ctrl_base __ro_after_init;
29static unsigned long cci_ctrl_phys __ro_after_init;
30
31#ifdef CONFIG_ARM_CCI400_PORT_CTRL
32struct cci_nb_ports {
33 unsigned int nb_ace;
34 unsigned int nb_ace_lite;
35};
36
37static const struct cci_nb_ports cci400_ports = {
38 .nb_ace = 2,
39 .nb_ace_lite = 3
40};
41
42#define CCI400_PORTS_DATA (&cci400_ports)
43#else
44#define CCI400_PORTS_DATA (NULL)
45#endif
46
47static const struct of_device_id arm_cci_matches[] = {
48#ifdef CONFIG_ARM_CCI400_COMMON
49 {.compatible = "arm,cci-400", .data = CCI400_PORTS_DATA },
50#endif
51#ifdef CONFIG_ARM_CCI5xx_PMU
52 { .compatible = "arm,cci-500", },
53 { .compatible = "arm,cci-550", },
54#endif
55 {},
56};
57
58static const struct of_dev_auxdata arm_cci_auxdata[] = {
59 OF_DEV_AUXDATA("arm,cci-400-pmu", 0, NULL, &cci_ctrl_base),
60 OF_DEV_AUXDATA("arm,cci-400-pmu,r0", 0, NULL, &cci_ctrl_base),
61 OF_DEV_AUXDATA("arm,cci-400-pmu,r1", 0, NULL, &cci_ctrl_base),
62 OF_DEV_AUXDATA("arm,cci-500-pmu,r0", 0, NULL, &cci_ctrl_base),
63 OF_DEV_AUXDATA("arm,cci-550-pmu,r0", 0, NULL, &cci_ctrl_base),
64 {}
65};
66
67#define DRIVER_NAME "ARM-CCI"
68
69static int cci_platform_probe(struct platform_device *pdev)
70{
71 if (!cci_probed())
72 return -ENODEV;
73
74 return of_platform_populate(pdev->dev.of_node, NULL,
75 arm_cci_auxdata, &pdev->dev);
76}
77
78static struct platform_driver cci_platform_driver = {
79 .driver = {
80 .name = DRIVER_NAME,
81 .of_match_table = arm_cci_matches,
82 },
83 .probe = cci_platform_probe,
84};
85
86static int __init cci_platform_init(void)
87{
88 return platform_driver_register(&cci_platform_driver);
89}
90
91#ifdef CONFIG_ARM_CCI400_PORT_CTRL
92
93#define CCI_PORT_CTRL 0x0
94#define CCI_CTRL_STATUS 0xc
95
96#define CCI_ENABLE_SNOOP_REQ 0x1
97#define CCI_ENABLE_DVM_REQ 0x2
98#define CCI_ENABLE_REQ (CCI_ENABLE_SNOOP_REQ | CCI_ENABLE_DVM_REQ)
99
100enum cci_ace_port_type {
101 ACE_INVALID_PORT = 0x0,
102 ACE_PORT,
103 ACE_LITE_PORT,
104};
105
106struct cci_ace_port {
107 void __iomem *base;
108 unsigned long phys;
109 enum cci_ace_port_type type;
110 struct device_node *dn;
111};
112
113static struct cci_ace_port *ports;
114static unsigned int nb_cci_ports;
115
116struct cpu_port {
117 u64 mpidr;
118 u32 port;
119};
120
121/*
122 * Use the port MSB as valid flag, shift can be made dynamic
123 * by computing number of bits required for port indexes.
124 * Code disabling CCI cpu ports runs with D-cache invalidated
125 * and SCTLR bit clear so data accesses must be kept to a minimum
126 * to improve performance; for now shift is left static to
127 * avoid one more data access while disabling the CCI port.
128 */
129#define PORT_VALID_SHIFT 31
130#define PORT_VALID (0x1 << PORT_VALID_SHIFT)
131
132static inline void init_cpu_port(struct cpu_port *port, u32 index, u64 mpidr)
133{
134 port->port = PORT_VALID | index;
135 port->mpidr = mpidr;
136}
137
138static inline bool cpu_port_is_valid(struct cpu_port *port)
139{
140 return !!(port->port & PORT_VALID);
141}
142
143static inline bool cpu_port_match(struct cpu_port *port, u64 mpidr)
144{
145 return port->mpidr == (mpidr & MPIDR_HWID_BITMASK);
146}
147
148static struct cpu_port cpu_port[NR_CPUS];
149
150/**
151 * __cci_ace_get_port - Function to retrieve the port index connected to
152 * a cpu or device.
153 *
154 * @dn: device node of the device to look-up
155 * @type: port type
156 *
157 * Return value:
158 * - CCI port index if success
159 * - -ENODEV if failure
160 */
161static int __cci_ace_get_port(struct device_node *dn, int type)
162{
163 int i;
164 bool ace_match;
165 struct device_node *cci_portn;
166
167 cci_portn = of_parse_phandle(dn, "cci-control-port", 0);
168 for (i = 0; i < nb_cci_ports; i++) {
169 ace_match = ports[i].type == type;
170 if (ace_match && cci_portn == ports[i].dn)
171 return i;
172 }
173 return -ENODEV;
174}
175
176int cci_ace_get_port(struct device_node *dn)
177{
178 return __cci_ace_get_port(dn, ACE_LITE_PORT);
179}
180EXPORT_SYMBOL_GPL(cci_ace_get_port);
181
182static void cci_ace_init_ports(void)
183{
184 int port, cpu;
185 struct device_node *cpun;
186
187 /*
188 * Port index look-up speeds up the function disabling ports by CPU,
189 * since the logical to port index mapping is done once and does
190 * not change after system boot.
191 * The stashed index array is initialized for all possible CPUs
192 * at probe time.
193 */
194 for_each_possible_cpu(cpu) {
195 /* too early to use cpu->of_node */
196 cpun = of_get_cpu_node(cpu, NULL);
197
198 if (WARN(!cpun, "Missing cpu device node\n"))
199 continue;
200
201 port = __cci_ace_get_port(cpun, ACE_PORT);
202 if (port < 0)
203 continue;
204
205 init_cpu_port(&cpu_port[cpu], port, cpu_logical_map(cpu));
206 }
207
208 for_each_possible_cpu(cpu) {
209 WARN(!cpu_port_is_valid(&cpu_port[cpu]),
210 "CPU %u does not have an associated CCI port\n",
211 cpu);
212 }
213}
214/*
215 * Functions to enable/disable a CCI interconnect slave port
216 *
217 * They are called by low-level power management code to disable slave
218 * interfaces snoops and DVM broadcast.
219 * Since they may execute with cache data allocation disabled and
220 * after the caches have been cleaned and invalidated the functions provide
221 * no explicit locking since they may run with D-cache disabled, so normal
222 * cacheable kernel locks based on ldrex/strex may not work.
223 * Locking has to be provided by BSP implementations to ensure proper
224 * operations.
225 */
226
227/**
228 * cci_port_control() - function to control a CCI port
229 *
230 * @port: index of the port to setup
231 * @enable: if true enables the port, if false disables it
232 */
233static void notrace cci_port_control(unsigned int port, bool enable)
234{
235 void __iomem *base = ports[port].base;
236
237 writel_relaxed(enable ? CCI_ENABLE_REQ : 0, base + CCI_PORT_CTRL);
238 /*
239 * This function is called from power down procedures
240 * and must not execute any instruction that might
241 * cause the processor to be put in a quiescent state
242 * (eg wfi). Hence, cpu_relax() can not be added to this
243 * read loop to optimize power, since it might hide possibly
244 * disruptive operations.
245 */
246 while (readl_relaxed(cci_ctrl_base + CCI_CTRL_STATUS) & 0x1)
247 ;
248}
249
250/**
251 * cci_disable_port_by_cpu() - function to disable a CCI port by CPU
252 * reference
253 *
254 * @mpidr: mpidr of the CPU whose CCI port should be disabled
255 *
256 * Disabling a CCI port for a CPU implies disabling the CCI port
257 * controlling that CPU cluster. Code disabling CPU CCI ports
258 * must make sure that the CPU running the code is the last active CPU
259 * in the cluster ie all other CPUs are quiescent in a low power state.
260 *
261 * Return:
262 * 0 on success
263 * -ENODEV on port look-up failure
264 */
265int notrace cci_disable_port_by_cpu(u64 mpidr)
266{
267 int cpu;
268 bool is_valid;
269 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
270 is_valid = cpu_port_is_valid(&cpu_port[cpu]);
271 if (is_valid && cpu_port_match(&cpu_port[cpu], mpidr)) {
272 cci_port_control(cpu_port[cpu].port, false);
273 return 0;
274 }
275 }
276 return -ENODEV;
277}
278EXPORT_SYMBOL_GPL(cci_disable_port_by_cpu);
279
280/**
281 * cci_enable_port_for_self() - enable a CCI port for calling CPU
282 *
283 * Enabling a CCI port for the calling CPU implies enabling the CCI
284 * port controlling that CPU's cluster. Caller must make sure that the
285 * CPU running the code is the first active CPU in the cluster and all
286 * other CPUs are quiescent in a low power state or waiting for this CPU
287 * to complete the CCI initialization.
288 *
289 * Because this is called when the MMU is still off and with no stack,
290 * the code must be position independent and ideally rely on callee
291 * clobbered registers only. To achieve this we must code this function
292 * entirely in assembler.
293 *
294 * On success this returns with the proper CCI port enabled. In case of
295 * any failure this never returns as the inability to enable the CCI is
296 * fatal and there is no possible recovery at this stage.
297 */
298asmlinkage void __naked cci_enable_port_for_self(void)
299{
300 asm volatile ("\n"
301" .arch armv7-a\n"
302" mrc p15, 0, r0, c0, c0, 5 @ get MPIDR value \n"
303" and r0, r0, #"__stringify(MPIDR_HWID_BITMASK)" \n"
304" adr r1, 5f \n"
305" ldr r2, [r1] \n"
306" add r1, r1, r2 @ &cpu_port \n"
307" add ip, r1, %[sizeof_cpu_port] \n"
308
309 /* Loop over the cpu_port array looking for a matching MPIDR */
310"1: ldr r2, [r1, %[offsetof_cpu_port_mpidr_lsb]] \n"
311" cmp r2, r0 @ compare MPIDR \n"
312" bne 2f \n"
313
314 /* Found a match, now test port validity */
315" ldr r3, [r1, %[offsetof_cpu_port_port]] \n"
316" tst r3, #"__stringify(PORT_VALID)" \n"
317" bne 3f \n"
318
319 /* no match, loop with the next cpu_port entry */
320"2: add r1, r1, %[sizeof_struct_cpu_port] \n"
321" cmp r1, ip @ done? \n"
322" blo 1b \n"
323
324 /* CCI port not found -- cheaply try to stall this CPU */
325"cci_port_not_found: \n"
326" wfi \n"
327" wfe \n"
328" b cci_port_not_found \n"
329
330 /* Use matched port index to look up the corresponding ports entry */
331"3: bic r3, r3, #"__stringify(PORT_VALID)" \n"
332" adr r0, 6f \n"
333" ldmia r0, {r1, r2} \n"
334" sub r1, r1, r0 @ virt - phys \n"
335" ldr r0, [r0, r2] @ *(&ports) \n"
336" mov r2, %[sizeof_struct_ace_port] \n"
337" mla r0, r2, r3, r0 @ &ports[index] \n"
338" sub r0, r0, r1 @ virt_to_phys() \n"
339
340 /* Enable the CCI port */
341" ldr r0, [r0, %[offsetof_port_phys]] \n"
342" mov r3, %[cci_enable_req]\n"
343" str r3, [r0, #"__stringify(CCI_PORT_CTRL)"] \n"
344
345 /* poll the status reg for completion */
346" adr r1, 7f \n"
347" ldr r0, [r1] \n"
348" ldr r0, [r0, r1] @ cci_ctrl_base \n"
349"4: ldr r1, [r0, #"__stringify(CCI_CTRL_STATUS)"] \n"
350" tst r1, %[cci_control_status_bits] \n"
351" bne 4b \n"
352
353" mov r0, #0 \n"
354" bx lr \n"
355
356" .align 2 \n"
357"5: .word cpu_port - . \n"
358"6: .word . \n"
359" .word ports - 6b \n"
360"7: .word cci_ctrl_phys - . \n"
361 : :
362 [sizeof_cpu_port] "i" (sizeof(cpu_port)),
363 [cci_enable_req] "i" cpu_to_le32(CCI_ENABLE_REQ),
364 [cci_control_status_bits] "i" cpu_to_le32(1),
365#ifndef __ARMEB__
366 [offsetof_cpu_port_mpidr_lsb] "i" (offsetof(struct cpu_port, mpidr)),
367#else
368 [offsetof_cpu_port_mpidr_lsb] "i" (offsetof(struct cpu_port, mpidr)+4),
369#endif
370 [offsetof_cpu_port_port] "i" (offsetof(struct cpu_port, port)),
371 [sizeof_struct_cpu_port] "i" (sizeof(struct cpu_port)),
372 [sizeof_struct_ace_port] "i" (sizeof(struct cci_ace_port)),
373 [offsetof_port_phys] "i" (offsetof(struct cci_ace_port, phys)) );
374}
375
376/**
377 * __cci_control_port_by_device() - function to control a CCI port by device
378 * reference
379 *
380 * @dn: device node pointer of the device whose CCI port should be
381 * controlled
382 * @enable: if true enables the port, if false disables it
383 *
384 * Return:
385 * 0 on success
386 * -ENODEV on port look-up failure
387 */
388int notrace __cci_control_port_by_device(struct device_node *dn, bool enable)
389{
390 int port;
391
392 if (!dn)
393 return -ENODEV;
394
395 port = __cci_ace_get_port(dn, ACE_LITE_PORT);
396 if (WARN_ONCE(port < 0, "node %pOF ACE lite port look-up failure\n",
397 dn))
398 return -ENODEV;
399 cci_port_control(port, enable);
400 return 0;
401}
402EXPORT_SYMBOL_GPL(__cci_control_port_by_device);
403
404/**
405 * __cci_control_port_by_index() - function to control a CCI port by port index
406 *
407 * @port: port index previously retrieved with cci_ace_get_port()
408 * @enable: if true enables the port, if false disables it
409 *
410 * Return:
411 * 0 on success
412 * -ENODEV on port index out of range
413 * -EPERM if operation carried out on an ACE PORT
414 */
415int notrace __cci_control_port_by_index(u32 port, bool enable)
416{
417 if (port >= nb_cci_ports || ports[port].type == ACE_INVALID_PORT)
418 return -ENODEV;
419 /*
420 * CCI control for ports connected to CPUS is extremely fragile
421 * and must be made to go through a specific and controlled
422 * interface (ie cci_disable_port_by_cpu(); control by general purpose
423 * indexing is therefore disabled for ACE ports.
424 */
425 if (ports[port].type == ACE_PORT)
426 return -EPERM;
427
428 cci_port_control(port, enable);
429 return 0;
430}
431EXPORT_SYMBOL_GPL(__cci_control_port_by_index);
432
433static const struct of_device_id arm_cci_ctrl_if_matches[] = {
434 {.compatible = "arm,cci-400-ctrl-if", },
435 {},
436};
437
438static int cci_probe_ports(struct device_node *np)
439{
440 struct cci_nb_ports const *cci_config;
441 int ret, i, nb_ace = 0, nb_ace_lite = 0;
442 struct device_node *cp;
443 struct resource res;
444 const char *match_str;
445 bool is_ace;
446
447
448 cci_config = of_match_node(arm_cci_matches, np)->data;
449 if (!cci_config)
450 return -ENODEV;
451
452 nb_cci_ports = cci_config->nb_ace + cci_config->nb_ace_lite;
453
454 ports = kcalloc(nb_cci_ports, sizeof(*ports), GFP_KERNEL);
455 if (!ports)
456 return -ENOMEM;
457
458 for_each_available_child_of_node(np, cp) {
459 if (!of_match_node(arm_cci_ctrl_if_matches, cp))
460 continue;
461
462 i = nb_ace + nb_ace_lite;
463
464 if (i >= nb_cci_ports)
465 break;
466
467 if (of_property_read_string(cp, "interface-type",
468 &match_str)) {
469 WARN(1, "node %pOF missing interface-type property\n",
470 cp);
471 continue;
472 }
473 is_ace = strcmp(match_str, "ace") == 0;
474 if (!is_ace && strcmp(match_str, "ace-lite")) {
475 WARN(1, "node %pOF containing invalid interface-type property, skipping it\n",
476 cp);
477 continue;
478 }
479
480 ret = of_address_to_resource(cp, 0, &res);
481 if (!ret) {
482 ports[i].base = ioremap(res.start, resource_size(&res));
483 ports[i].phys = res.start;
484 }
485 if (ret || !ports[i].base) {
486 WARN(1, "unable to ioremap CCI port %d\n", i);
487 continue;
488 }
489
490 if (is_ace) {
491 if (WARN_ON(nb_ace >= cci_config->nb_ace))
492 continue;
493 ports[i].type = ACE_PORT;
494 ++nb_ace;
495 } else {
496 if (WARN_ON(nb_ace_lite >= cci_config->nb_ace_lite))
497 continue;
498 ports[i].type = ACE_LITE_PORT;
499 ++nb_ace_lite;
500 }
501 ports[i].dn = cp;
502 }
503
504 /*
505 * If there is no CCI port that is under kernel control
506 * return early and report probe status.
507 */
508 if (!nb_ace && !nb_ace_lite)
509 return -ENODEV;
510
511 /* initialize a stashed array of ACE ports to speed-up look-up */
512 cci_ace_init_ports();
513
514 /*
515 * Multi-cluster systems may need this data when non-coherent, during
516 * cluster power-up/power-down. Make sure it reaches main memory.
517 */
518 sync_cache_w(&cci_ctrl_base);
519 sync_cache_w(&cci_ctrl_phys);
520 sync_cache_w(&ports);
521 sync_cache_w(&cpu_port);
522 __sync_cache_range_w(ports, sizeof(*ports) * nb_cci_ports);
523 pr_info("ARM CCI driver probed\n");
524
525 return 0;
526}
527#else /* !CONFIG_ARM_CCI400_PORT_CTRL */
528static inline int cci_probe_ports(struct device_node *np)
529{
530 return 0;
531}
532#endif /* CONFIG_ARM_CCI400_PORT_CTRL */
533
534static int cci_probe(void)
535{
536 int ret;
537 struct device_node *np;
538 struct resource res;
539
540 np = of_find_matching_node(NULL, arm_cci_matches);
541 if (!of_device_is_available(np))
542 return -ENODEV;
543
544 ret = of_address_to_resource(np, 0, &res);
545 if (!ret) {
546 cci_ctrl_base = ioremap(res.start, resource_size(&res));
547 cci_ctrl_phys = res.start;
548 }
549 if (ret || !cci_ctrl_base) {
550 WARN(1, "unable to ioremap CCI ctrl\n");
551 return -ENXIO;
552 }
553
554 return cci_probe_ports(np);
555}
556
557static int cci_init_status = -EAGAIN;
558static DEFINE_MUTEX(cci_probing);
559
560static int cci_init(void)
561{
562 if (cci_init_status != -EAGAIN)
563 return cci_init_status;
564
565 mutex_lock(&cci_probing);
566 if (cci_init_status == -EAGAIN)
567 cci_init_status = cci_probe();
568 mutex_unlock(&cci_probing);
569 return cci_init_status;
570}
571
572/*
573 * To sort out early init calls ordering a helper function is provided to
574 * check if the CCI driver has beed initialized. Function check if the driver
575 * has been initialized, if not it calls the init function that probes
576 * the driver and updates the return value.
577 */
578bool cci_probed(void)
579{
580 return cci_init() == 0;
581}
582EXPORT_SYMBOL_GPL(cci_probed);
583
584early_initcall(cci_init);
585core_initcall(cci_platform_init);
586MODULE_LICENSE("GPL");
587MODULE_DESCRIPTION("ARM CCI support");
1/*
2 * CCI cache coherent interconnect driver
3 *
4 * Copyright (C) 2013 ARM Ltd.
5 * Author: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
10 *
11 * This program is distributed "as is" WITHOUT ANY WARRANTY of any
12 * kind, whether express or implied; without even the implied warranty
13 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 */
16
17#include <linux/arm-cci.h>
18#include <linux/io.h>
19#include <linux/interrupt.h>
20#include <linux/module.h>
21#include <linux/of_address.h>
22#include <linux/of_irq.h>
23#include <linux/of_platform.h>
24#include <linux/perf_event.h>
25#include <linux/platform_device.h>
26#include <linux/slab.h>
27#include <linux/spinlock.h>
28
29#include <asm/cacheflush.h>
30#include <asm/smp_plat.h>
31
32static void __iomem *cci_ctrl_base;
33static unsigned long cci_ctrl_phys;
34
35#ifdef CONFIG_ARM_CCI400_PORT_CTRL
36struct cci_nb_ports {
37 unsigned int nb_ace;
38 unsigned int nb_ace_lite;
39};
40
41static const struct cci_nb_ports cci400_ports = {
42 .nb_ace = 2,
43 .nb_ace_lite = 3
44};
45
46#define CCI400_PORTS_DATA (&cci400_ports)
47#else
48#define CCI400_PORTS_DATA (NULL)
49#endif
50
51static const struct of_device_id arm_cci_matches[] = {
52#ifdef CONFIG_ARM_CCI400_COMMON
53 {.compatible = "arm,cci-400", .data = CCI400_PORTS_DATA },
54#endif
55#ifdef CONFIG_ARM_CCI5xx_PMU
56 { .compatible = "arm,cci-500", },
57 { .compatible = "arm,cci-550", },
58#endif
59 {},
60};
61
62#ifdef CONFIG_ARM_CCI_PMU
63
64#define DRIVER_NAME "ARM-CCI"
65#define DRIVER_NAME_PMU DRIVER_NAME " PMU"
66
67#define CCI_PMCR 0x0100
68#define CCI_PID2 0x0fe8
69
70#define CCI_PMCR_CEN 0x00000001
71#define CCI_PMCR_NCNT_MASK 0x0000f800
72#define CCI_PMCR_NCNT_SHIFT 11
73
74#define CCI_PID2_REV_MASK 0xf0
75#define CCI_PID2_REV_SHIFT 4
76
77#define CCI_PMU_EVT_SEL 0x000
78#define CCI_PMU_CNTR 0x004
79#define CCI_PMU_CNTR_CTRL 0x008
80#define CCI_PMU_OVRFLW 0x00c
81
82#define CCI_PMU_OVRFLW_FLAG 1
83
84#define CCI_PMU_CNTR_SIZE(model) ((model)->cntr_size)
85#define CCI_PMU_CNTR_BASE(model, idx) ((idx) * CCI_PMU_CNTR_SIZE(model))
86#define CCI_PMU_CNTR_MASK ((1ULL << 32) -1)
87#define CCI_PMU_CNTR_LAST(cci_pmu) (cci_pmu->num_cntrs - 1)
88
89#define CCI_PMU_MAX_HW_CNTRS(model) \
90 ((model)->num_hw_cntrs + (model)->fixed_hw_cntrs)
91
92/* Types of interfaces that can generate events */
93enum {
94 CCI_IF_SLAVE,
95 CCI_IF_MASTER,
96#ifdef CONFIG_ARM_CCI5xx_PMU
97 CCI_IF_GLOBAL,
98#endif
99 CCI_IF_MAX,
100};
101
102struct event_range {
103 u32 min;
104 u32 max;
105};
106
107struct cci_pmu_hw_events {
108 struct perf_event **events;
109 unsigned long *used_mask;
110 raw_spinlock_t pmu_lock;
111};
112
113struct cci_pmu;
114/*
115 * struct cci_pmu_model:
116 * @fixed_hw_cntrs - Number of fixed event counters
117 * @num_hw_cntrs - Maximum number of programmable event counters
118 * @cntr_size - Size of an event counter mapping
119 */
120struct cci_pmu_model {
121 char *name;
122 u32 fixed_hw_cntrs;
123 u32 num_hw_cntrs;
124 u32 cntr_size;
125 struct attribute **format_attrs;
126 struct attribute **event_attrs;
127 struct event_range event_ranges[CCI_IF_MAX];
128 int (*validate_hw_event)(struct cci_pmu *, unsigned long);
129 int (*get_event_idx)(struct cci_pmu *, struct cci_pmu_hw_events *, unsigned long);
130 void (*write_counters)(struct cci_pmu *, unsigned long *);
131};
132
133static struct cci_pmu_model cci_pmu_models[];
134
135struct cci_pmu {
136 void __iomem *base;
137 struct pmu pmu;
138 int nr_irqs;
139 int *irqs;
140 unsigned long active_irqs;
141 const struct cci_pmu_model *model;
142 struct cci_pmu_hw_events hw_events;
143 struct platform_device *plat_device;
144 int num_cntrs;
145 atomic_t active_events;
146 struct mutex reserve_mutex;
147 struct notifier_block cpu_nb;
148 cpumask_t cpus;
149};
150
151#define to_cci_pmu(c) (container_of(c, struct cci_pmu, pmu))
152
153enum cci_models {
154#ifdef CONFIG_ARM_CCI400_PMU
155 CCI400_R0,
156 CCI400_R1,
157#endif
158#ifdef CONFIG_ARM_CCI5xx_PMU
159 CCI500_R0,
160 CCI550_R0,
161#endif
162 CCI_MODEL_MAX
163};
164
165static void pmu_write_counters(struct cci_pmu *cci_pmu,
166 unsigned long *mask);
167static ssize_t cci_pmu_format_show(struct device *dev,
168 struct device_attribute *attr, char *buf);
169static ssize_t cci_pmu_event_show(struct device *dev,
170 struct device_attribute *attr, char *buf);
171
172#define CCI_EXT_ATTR_ENTRY(_name, _func, _config) \
173 &((struct dev_ext_attribute[]) { \
174 { __ATTR(_name, S_IRUGO, _func, NULL), (void *)_config } \
175 })[0].attr.attr
176
177#define CCI_FORMAT_EXT_ATTR_ENTRY(_name, _config) \
178 CCI_EXT_ATTR_ENTRY(_name, cci_pmu_format_show, (char *)_config)
179#define CCI_EVENT_EXT_ATTR_ENTRY(_name, _config) \
180 CCI_EXT_ATTR_ENTRY(_name, cci_pmu_event_show, (unsigned long)_config)
181
182/* CCI400 PMU Specific definitions */
183
184#ifdef CONFIG_ARM_CCI400_PMU
185
186/* Port ids */
187#define CCI400_PORT_S0 0
188#define CCI400_PORT_S1 1
189#define CCI400_PORT_S2 2
190#define CCI400_PORT_S3 3
191#define CCI400_PORT_S4 4
192#define CCI400_PORT_M0 5
193#define CCI400_PORT_M1 6
194#define CCI400_PORT_M2 7
195
196#define CCI400_R1_PX 5
197
198/*
199 * Instead of an event id to monitor CCI cycles, a dedicated counter is
200 * provided. Use 0xff to represent CCI cycles and hope that no future revisions
201 * make use of this event in hardware.
202 */
203enum cci400_perf_events {
204 CCI400_PMU_CYCLES = 0xff
205};
206
207#define CCI400_PMU_CYCLE_CNTR_IDX 0
208#define CCI400_PMU_CNTR0_IDX 1
209
210/*
211 * CCI PMU event id is an 8-bit value made of two parts - bits 7:5 for one of 8
212 * ports and bits 4:0 are event codes. There are different event codes
213 * associated with each port type.
214 *
215 * Additionally, the range of events associated with the port types changed
216 * between Rev0 and Rev1.
217 *
218 * The constants below define the range of valid codes for each port type for
219 * the different revisions and are used to validate the event to be monitored.
220 */
221
222#define CCI400_PMU_EVENT_MASK 0xffUL
223#define CCI400_PMU_EVENT_SOURCE_SHIFT 5
224#define CCI400_PMU_EVENT_SOURCE_MASK 0x7
225#define CCI400_PMU_EVENT_CODE_SHIFT 0
226#define CCI400_PMU_EVENT_CODE_MASK 0x1f
227#define CCI400_PMU_EVENT_SOURCE(event) \
228 ((event >> CCI400_PMU_EVENT_SOURCE_SHIFT) & \
229 CCI400_PMU_EVENT_SOURCE_MASK)
230#define CCI400_PMU_EVENT_CODE(event) \
231 ((event >> CCI400_PMU_EVENT_CODE_SHIFT) & CCI400_PMU_EVENT_CODE_MASK)
232
233#define CCI400_R0_SLAVE_PORT_MIN_EV 0x00
234#define CCI400_R0_SLAVE_PORT_MAX_EV 0x13
235#define CCI400_R0_MASTER_PORT_MIN_EV 0x14
236#define CCI400_R0_MASTER_PORT_MAX_EV 0x1a
237
238#define CCI400_R1_SLAVE_PORT_MIN_EV 0x00
239#define CCI400_R1_SLAVE_PORT_MAX_EV 0x14
240#define CCI400_R1_MASTER_PORT_MIN_EV 0x00
241#define CCI400_R1_MASTER_PORT_MAX_EV 0x11
242
243#define CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(_name, _config) \
244 CCI_EXT_ATTR_ENTRY(_name, cci400_pmu_cycle_event_show, \
245 (unsigned long)_config)
246
247static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
248 struct device_attribute *attr, char *buf);
249
250static struct attribute *cci400_pmu_format_attrs[] = {
251 CCI_FORMAT_EXT_ATTR_ENTRY(event, "config:0-4"),
252 CCI_FORMAT_EXT_ATTR_ENTRY(source, "config:5-7"),
253 NULL
254};
255
256static struct attribute *cci400_r0_pmu_event_attrs[] = {
257 /* Slave events */
258 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_any, 0x0),
259 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_device, 0x01),
260 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_normal_or_nonshareable, 0x2),
261 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_inner_or_outershareable, 0x3),
262 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maintenance, 0x4),
263 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_mem_barrier, 0x5),
264 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_sync_barrier, 0x6),
265 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
266 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg_sync, 0x8),
267 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_tt_full, 0x9),
268 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_last_hs_snoop, 0xA),
269 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall_rvalids_h_rready_l, 0xB),
270 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_any, 0xC),
271 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_device, 0xD),
272 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_normal_or_nonshareable, 0xE),
273 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_inner_or_outershare_wback_wclean, 0xF),
274 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_unique, 0x10),
275 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_line_unique, 0x11),
276 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_evict, 0x12),
277 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall_tt_full, 0x13),
278 /* Master events */
279 CCI_EVENT_EXT_ATTR_ENTRY(mi_retry_speculative_fetch, 0x14),
280 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_addr_hazard, 0x15),
281 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_id_hazard, 0x16),
282 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_tt_full, 0x17),
283 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_barrier_hazard, 0x18),
284 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_barrier_hazard, 0x19),
285 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_tt_full, 0x1A),
286 /* Special event for cycles counter */
287 CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(cycles, 0xff),
288 NULL
289};
290
291static struct attribute *cci400_r1_pmu_event_attrs[] = {
292 /* Slave events */
293 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_any, 0x0),
294 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_device, 0x01),
295 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_normal_or_nonshareable, 0x2),
296 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_inner_or_outershareable, 0x3),
297 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maintenance, 0x4),
298 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_mem_barrier, 0x5),
299 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_sync_barrier, 0x6),
300 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
301 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg_sync, 0x8),
302 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_tt_full, 0x9),
303 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_last_hs_snoop, 0xA),
304 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall_rvalids_h_rready_l, 0xB),
305 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_any, 0xC),
306 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_device, 0xD),
307 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_normal_or_nonshareable, 0xE),
308 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_inner_or_outershare_wback_wclean, 0xF),
309 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_unique, 0x10),
310 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_write_line_unique, 0x11),
311 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_evict, 0x12),
312 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall_tt_full, 0x13),
313 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_slave_id_hazard, 0x14),
314 /* Master events */
315 CCI_EVENT_EXT_ATTR_ENTRY(mi_retry_speculative_fetch, 0x0),
316 CCI_EVENT_EXT_ATTR_ENTRY(mi_stall_cycle_addr_hazard, 0x1),
317 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_master_id_hazard, 0x2),
318 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_hi_prio_rtq_full, 0x3),
319 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_barrier_hazard, 0x4),
320 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_barrier_hazard, 0x5),
321 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_wtq_full, 0x6),
322 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_low_prio_rtq_full, 0x7),
323 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_mid_prio_rtq_full, 0x8),
324 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn0, 0x9),
325 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn1, 0xA),
326 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn2, 0xB),
327 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall_qvn_vn3, 0xC),
328 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn0, 0xD),
329 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn1, 0xE),
330 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn2, 0xF),
331 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall_qvn_vn3, 0x10),
332 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_unique_or_line_unique_addr_hazard, 0x11),
333 /* Special event for cycles counter */
334 CCI400_CYCLE_EVENT_EXT_ATTR_ENTRY(cycles, 0xff),
335 NULL
336};
337
338static ssize_t cci400_pmu_cycle_event_show(struct device *dev,
339 struct device_attribute *attr, char *buf)
340{
341 struct dev_ext_attribute *eattr = container_of(attr,
342 struct dev_ext_attribute, attr);
343 return snprintf(buf, PAGE_SIZE, "config=0x%lx\n", (unsigned long)eattr->var);
344}
345
346static int cci400_get_event_idx(struct cci_pmu *cci_pmu,
347 struct cci_pmu_hw_events *hw,
348 unsigned long cci_event)
349{
350 int idx;
351
352 /* cycles event idx is fixed */
353 if (cci_event == CCI400_PMU_CYCLES) {
354 if (test_and_set_bit(CCI400_PMU_CYCLE_CNTR_IDX, hw->used_mask))
355 return -EAGAIN;
356
357 return CCI400_PMU_CYCLE_CNTR_IDX;
358 }
359
360 for (idx = CCI400_PMU_CNTR0_IDX; idx <= CCI_PMU_CNTR_LAST(cci_pmu); ++idx)
361 if (!test_and_set_bit(idx, hw->used_mask))
362 return idx;
363
364 /* No counters available */
365 return -EAGAIN;
366}
367
368static int cci400_validate_hw_event(struct cci_pmu *cci_pmu, unsigned long hw_event)
369{
370 u8 ev_source = CCI400_PMU_EVENT_SOURCE(hw_event);
371 u8 ev_code = CCI400_PMU_EVENT_CODE(hw_event);
372 int if_type;
373
374 if (hw_event & ~CCI400_PMU_EVENT_MASK)
375 return -ENOENT;
376
377 if (hw_event == CCI400_PMU_CYCLES)
378 return hw_event;
379
380 switch (ev_source) {
381 case CCI400_PORT_S0:
382 case CCI400_PORT_S1:
383 case CCI400_PORT_S2:
384 case CCI400_PORT_S3:
385 case CCI400_PORT_S4:
386 /* Slave Interface */
387 if_type = CCI_IF_SLAVE;
388 break;
389 case CCI400_PORT_M0:
390 case CCI400_PORT_M1:
391 case CCI400_PORT_M2:
392 /* Master Interface */
393 if_type = CCI_IF_MASTER;
394 break;
395 default:
396 return -ENOENT;
397 }
398
399 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
400 ev_code <= cci_pmu->model->event_ranges[if_type].max)
401 return hw_event;
402
403 return -ENOENT;
404}
405
406static int probe_cci400_revision(void)
407{
408 int rev;
409 rev = readl_relaxed(cci_ctrl_base + CCI_PID2) & CCI_PID2_REV_MASK;
410 rev >>= CCI_PID2_REV_SHIFT;
411
412 if (rev < CCI400_R1_PX)
413 return CCI400_R0;
414 else
415 return CCI400_R1;
416}
417
418static const struct cci_pmu_model *probe_cci_model(struct platform_device *pdev)
419{
420 if (platform_has_secure_cci_access())
421 return &cci_pmu_models[probe_cci400_revision()];
422 return NULL;
423}
424#else /* !CONFIG_ARM_CCI400_PMU */
425static inline struct cci_pmu_model *probe_cci_model(struct platform_device *pdev)
426{
427 return NULL;
428}
429#endif /* CONFIG_ARM_CCI400_PMU */
430
431#ifdef CONFIG_ARM_CCI5xx_PMU
432
433/*
434 * CCI5xx PMU event id is an 9-bit value made of two parts.
435 * bits [8:5] - Source for the event
436 * bits [4:0] - Event code (specific to type of interface)
437 *
438 *
439 */
440
441/* Port ids */
442#define CCI5xx_PORT_S0 0x0
443#define CCI5xx_PORT_S1 0x1
444#define CCI5xx_PORT_S2 0x2
445#define CCI5xx_PORT_S3 0x3
446#define CCI5xx_PORT_S4 0x4
447#define CCI5xx_PORT_S5 0x5
448#define CCI5xx_PORT_S6 0x6
449
450#define CCI5xx_PORT_M0 0x8
451#define CCI5xx_PORT_M1 0x9
452#define CCI5xx_PORT_M2 0xa
453#define CCI5xx_PORT_M3 0xb
454#define CCI5xx_PORT_M4 0xc
455#define CCI5xx_PORT_M5 0xd
456#define CCI5xx_PORT_M6 0xe
457
458#define CCI5xx_PORT_GLOBAL 0xf
459
460#define CCI5xx_PMU_EVENT_MASK 0x1ffUL
461#define CCI5xx_PMU_EVENT_SOURCE_SHIFT 0x5
462#define CCI5xx_PMU_EVENT_SOURCE_MASK 0xf
463#define CCI5xx_PMU_EVENT_CODE_SHIFT 0x0
464#define CCI5xx_PMU_EVENT_CODE_MASK 0x1f
465
466#define CCI5xx_PMU_EVENT_SOURCE(event) \
467 ((event >> CCI5xx_PMU_EVENT_SOURCE_SHIFT) & CCI5xx_PMU_EVENT_SOURCE_MASK)
468#define CCI5xx_PMU_EVENT_CODE(event) \
469 ((event >> CCI5xx_PMU_EVENT_CODE_SHIFT) & CCI5xx_PMU_EVENT_CODE_MASK)
470
471#define CCI5xx_SLAVE_PORT_MIN_EV 0x00
472#define CCI5xx_SLAVE_PORT_MAX_EV 0x1f
473#define CCI5xx_MASTER_PORT_MIN_EV 0x00
474#define CCI5xx_MASTER_PORT_MAX_EV 0x06
475#define CCI5xx_GLOBAL_PORT_MIN_EV 0x00
476#define CCI5xx_GLOBAL_PORT_MAX_EV 0x0f
477
478
479#define CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(_name, _config) \
480 CCI_EXT_ATTR_ENTRY(_name, cci5xx_pmu_global_event_show, \
481 (unsigned long) _config)
482
483static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
484 struct device_attribute *attr, char *buf);
485
486static struct attribute *cci5xx_pmu_format_attrs[] = {
487 CCI_FORMAT_EXT_ATTR_ENTRY(event, "config:0-4"),
488 CCI_FORMAT_EXT_ATTR_ENTRY(source, "config:5-8"),
489 NULL,
490};
491
492static struct attribute *cci5xx_pmu_event_attrs[] = {
493 /* Slave events */
494 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_arvalid, 0x0),
495 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_dev, 0x1),
496 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_nonshareable, 0x2),
497 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_shareable_non_alloc, 0x3),
498 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_shareable_alloc, 0x4),
499 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_invalidate, 0x5),
500 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_cache_maint, 0x6),
501 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_dvm_msg, 0x7),
502 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_rval, 0x8),
503 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_hs_rlast_snoop, 0x9),
504 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_hs_awalid, 0xA),
505 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_dev, 0xB),
506 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_non_shareable, 0xC),
507 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wb, 0xD),
508 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wlu, 0xE),
509 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_share_wunique, 0xF),
510 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_evict, 0x10),
511 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_wrevict, 0x11),
512 CCI_EVENT_EXT_ATTR_ENTRY(si_w_data_beat, 0x12),
513 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_acvalid, 0x13),
514 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_read, 0x14),
515 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_clean, 0x15),
516 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_data_transfer_low, 0x16),
517 CCI_EVENT_EXT_ATTR_ENTRY(si_rrq_stall_arvalid, 0x17),
518 CCI_EVENT_EXT_ATTR_ENTRY(si_r_data_stall, 0x18),
519 CCI_EVENT_EXT_ATTR_ENTRY(si_wrq_stall, 0x19),
520 CCI_EVENT_EXT_ATTR_ENTRY(si_w_data_stall, 0x1A),
521 CCI_EVENT_EXT_ATTR_ENTRY(si_w_resp_stall, 0x1B),
522 CCI_EVENT_EXT_ATTR_ENTRY(si_srq_stall, 0x1C),
523 CCI_EVENT_EXT_ATTR_ENTRY(si_s_data_stall, 0x1D),
524 CCI_EVENT_EXT_ATTR_ENTRY(si_rq_stall_ot_limit, 0x1E),
525 CCI_EVENT_EXT_ATTR_ENTRY(si_r_stall_arbit, 0x1F),
526
527 /* Master events */
528 CCI_EVENT_EXT_ATTR_ENTRY(mi_r_data_beat_any, 0x0),
529 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_data_beat_any, 0x1),
530 CCI_EVENT_EXT_ATTR_ENTRY(mi_rrq_stall, 0x2),
531 CCI_EVENT_EXT_ATTR_ENTRY(mi_r_data_stall, 0x3),
532 CCI_EVENT_EXT_ATTR_ENTRY(mi_wrq_stall, 0x4),
533 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_data_stall, 0x5),
534 CCI_EVENT_EXT_ATTR_ENTRY(mi_w_resp_stall, 0x6),
535
536 /* Global events */
537 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_0_1, 0x0),
538 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_2_3, 0x1),
539 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_4_5, 0x2),
540 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_filter_bank_6_7, 0x3),
541 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_0_1, 0x4),
542 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_2_3, 0x5),
543 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_4_5, 0x6),
544 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_access_miss_filter_bank_6_7, 0x7),
545 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_back_invalidation, 0x8),
546 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_stall_alloc_busy, 0x9),
547 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_stall_tt_full, 0xA),
548 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_wrq, 0xB),
549 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_cd_hs, 0xC),
550 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_rq_stall_addr_hazard, 0xD),
551 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snopp_rq_stall_tt_full, 0xE),
552 CCI5xx_GLOBAL_EVENT_EXT_ATTR_ENTRY(cci_snoop_rq_tzmp1_prot, 0xF),
553 NULL
554};
555
556static ssize_t cci5xx_pmu_global_event_show(struct device *dev,
557 struct device_attribute *attr, char *buf)
558{
559 struct dev_ext_attribute *eattr = container_of(attr,
560 struct dev_ext_attribute, attr);
561 /* Global events have single fixed source code */
562 return snprintf(buf, PAGE_SIZE, "event=0x%lx,source=0x%x\n",
563 (unsigned long)eattr->var, CCI5xx_PORT_GLOBAL);
564}
565
566/*
567 * CCI500 provides 8 independent event counters that can count
568 * any of the events available.
569 * CCI500 PMU event source ids
570 * 0x0-0x6 - Slave interfaces
571 * 0x8-0xD - Master interfaces
572 * 0xf - Global Events
573 * 0x7,0xe - Reserved
574 */
575static int cci500_validate_hw_event(struct cci_pmu *cci_pmu,
576 unsigned long hw_event)
577{
578 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
579 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
580 int if_type;
581
582 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
583 return -ENOENT;
584
585 switch (ev_source) {
586 case CCI5xx_PORT_S0:
587 case CCI5xx_PORT_S1:
588 case CCI5xx_PORT_S2:
589 case CCI5xx_PORT_S3:
590 case CCI5xx_PORT_S4:
591 case CCI5xx_PORT_S5:
592 case CCI5xx_PORT_S6:
593 if_type = CCI_IF_SLAVE;
594 break;
595 case CCI5xx_PORT_M0:
596 case CCI5xx_PORT_M1:
597 case CCI5xx_PORT_M2:
598 case CCI5xx_PORT_M3:
599 case CCI5xx_PORT_M4:
600 case CCI5xx_PORT_M5:
601 if_type = CCI_IF_MASTER;
602 break;
603 case CCI5xx_PORT_GLOBAL:
604 if_type = CCI_IF_GLOBAL;
605 break;
606 default:
607 return -ENOENT;
608 }
609
610 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
611 ev_code <= cci_pmu->model->event_ranges[if_type].max)
612 return hw_event;
613
614 return -ENOENT;
615}
616
617/*
618 * CCI550 provides 8 independent event counters that can count
619 * any of the events available.
620 * CCI550 PMU event source ids
621 * 0x0-0x6 - Slave interfaces
622 * 0x8-0xe - Master interfaces
623 * 0xf - Global Events
624 * 0x7 - Reserved
625 */
626static int cci550_validate_hw_event(struct cci_pmu *cci_pmu,
627 unsigned long hw_event)
628{
629 u32 ev_source = CCI5xx_PMU_EVENT_SOURCE(hw_event);
630 u32 ev_code = CCI5xx_PMU_EVENT_CODE(hw_event);
631 int if_type;
632
633 if (hw_event & ~CCI5xx_PMU_EVENT_MASK)
634 return -ENOENT;
635
636 switch (ev_source) {
637 case CCI5xx_PORT_S0:
638 case CCI5xx_PORT_S1:
639 case CCI5xx_PORT_S2:
640 case CCI5xx_PORT_S3:
641 case CCI5xx_PORT_S4:
642 case CCI5xx_PORT_S5:
643 case CCI5xx_PORT_S6:
644 if_type = CCI_IF_SLAVE;
645 break;
646 case CCI5xx_PORT_M0:
647 case CCI5xx_PORT_M1:
648 case CCI5xx_PORT_M2:
649 case CCI5xx_PORT_M3:
650 case CCI5xx_PORT_M4:
651 case CCI5xx_PORT_M5:
652 case CCI5xx_PORT_M6:
653 if_type = CCI_IF_MASTER;
654 break;
655 case CCI5xx_PORT_GLOBAL:
656 if_type = CCI_IF_GLOBAL;
657 break;
658 default:
659 return -ENOENT;
660 }
661
662 if (ev_code >= cci_pmu->model->event_ranges[if_type].min &&
663 ev_code <= cci_pmu->model->event_ranges[if_type].max)
664 return hw_event;
665
666 return -ENOENT;
667}
668
669#endif /* CONFIG_ARM_CCI5xx_PMU */
670
671/*
672 * Program the CCI PMU counters which have PERF_HES_ARCH set
673 * with the event period and mark them ready before we enable
674 * PMU.
675 */
676static void cci_pmu_sync_counters(struct cci_pmu *cci_pmu)
677{
678 int i;
679 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
680
681 DECLARE_BITMAP(mask, cci_pmu->num_cntrs);
682
683 bitmap_zero(mask, cci_pmu->num_cntrs);
684 for_each_set_bit(i, cci_pmu->hw_events.used_mask, cci_pmu->num_cntrs) {
685 struct perf_event *event = cci_hw->events[i];
686
687 if (WARN_ON(!event))
688 continue;
689
690 /* Leave the events which are not counting */
691 if (event->hw.state & PERF_HES_STOPPED)
692 continue;
693 if (event->hw.state & PERF_HES_ARCH) {
694 set_bit(i, mask);
695 event->hw.state &= ~PERF_HES_ARCH;
696 }
697 }
698
699 pmu_write_counters(cci_pmu, mask);
700}
701
702/* Should be called with cci_pmu->hw_events->pmu_lock held */
703static void __cci_pmu_enable_nosync(struct cci_pmu *cci_pmu)
704{
705 u32 val;
706
707 /* Enable all the PMU counters. */
708 val = readl_relaxed(cci_ctrl_base + CCI_PMCR) | CCI_PMCR_CEN;
709 writel(val, cci_ctrl_base + CCI_PMCR);
710}
711
712/* Should be called with cci_pmu->hw_events->pmu_lock held */
713static void __cci_pmu_enable_sync(struct cci_pmu *cci_pmu)
714{
715 cci_pmu_sync_counters(cci_pmu);
716 __cci_pmu_enable_nosync(cci_pmu);
717}
718
719/* Should be called with cci_pmu->hw_events->pmu_lock held */
720static void __cci_pmu_disable(void)
721{
722 u32 val;
723
724 /* Disable all the PMU counters. */
725 val = readl_relaxed(cci_ctrl_base + CCI_PMCR) & ~CCI_PMCR_CEN;
726 writel(val, cci_ctrl_base + CCI_PMCR);
727}
728
729static ssize_t cci_pmu_format_show(struct device *dev,
730 struct device_attribute *attr, char *buf)
731{
732 struct dev_ext_attribute *eattr = container_of(attr,
733 struct dev_ext_attribute, attr);
734 return snprintf(buf, PAGE_SIZE, "%s\n", (char *)eattr->var);
735}
736
737static ssize_t cci_pmu_event_show(struct device *dev,
738 struct device_attribute *attr, char *buf)
739{
740 struct dev_ext_attribute *eattr = container_of(attr,
741 struct dev_ext_attribute, attr);
742 /* source parameter is mandatory for normal PMU events */
743 return snprintf(buf, PAGE_SIZE, "source=?,event=0x%lx\n",
744 (unsigned long)eattr->var);
745}
746
747static int pmu_is_valid_counter(struct cci_pmu *cci_pmu, int idx)
748{
749 return 0 <= idx && idx <= CCI_PMU_CNTR_LAST(cci_pmu);
750}
751
752static u32 pmu_read_register(struct cci_pmu *cci_pmu, int idx, unsigned int offset)
753{
754 return readl_relaxed(cci_pmu->base +
755 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
756}
757
758static void pmu_write_register(struct cci_pmu *cci_pmu, u32 value,
759 int idx, unsigned int offset)
760{
761 writel_relaxed(value, cci_pmu->base +
762 CCI_PMU_CNTR_BASE(cci_pmu->model, idx) + offset);
763}
764
765static void pmu_disable_counter(struct cci_pmu *cci_pmu, int idx)
766{
767 pmu_write_register(cci_pmu, 0, idx, CCI_PMU_CNTR_CTRL);
768}
769
770static void pmu_enable_counter(struct cci_pmu *cci_pmu, int idx)
771{
772 pmu_write_register(cci_pmu, 1, idx, CCI_PMU_CNTR_CTRL);
773}
774
775static bool __maybe_unused
776pmu_counter_is_enabled(struct cci_pmu *cci_pmu, int idx)
777{
778 return (pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR_CTRL) & 0x1) != 0;
779}
780
781static void pmu_set_event(struct cci_pmu *cci_pmu, int idx, unsigned long event)
782{
783 pmu_write_register(cci_pmu, event, idx, CCI_PMU_EVT_SEL);
784}
785
786/*
787 * For all counters on the CCI-PMU, disable any 'enabled' counters,
788 * saving the changed counters in the mask, so that we can restore
789 * it later using pmu_restore_counters. The mask is private to the
790 * caller. We cannot rely on the used_mask maintained by the CCI_PMU
791 * as it only tells us if the counter is assigned to perf_event or not.
792 * The state of the perf_event cannot be locked by the PMU layer, hence
793 * we check the individual counter status (which can be locked by
794 * cci_pm->hw_events->pmu_lock).
795 *
796 * @mask should be initialised to empty by the caller.
797 */
798static void __maybe_unused
799pmu_save_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
800{
801 int i;
802
803 for (i = 0; i < cci_pmu->num_cntrs; i++) {
804 if (pmu_counter_is_enabled(cci_pmu, i)) {
805 set_bit(i, mask);
806 pmu_disable_counter(cci_pmu, i);
807 }
808 }
809}
810
811/*
812 * Restore the status of the counters. Reversal of the pmu_save_counters().
813 * For each counter set in the mask, enable the counter back.
814 */
815static void __maybe_unused
816pmu_restore_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
817{
818 int i;
819
820 for_each_set_bit(i, mask, cci_pmu->num_cntrs)
821 pmu_enable_counter(cci_pmu, i);
822}
823
824/*
825 * Returns the number of programmable counters actually implemented
826 * by the cci
827 */
828static u32 pmu_get_max_counters(void)
829{
830 return (readl_relaxed(cci_ctrl_base + CCI_PMCR) &
831 CCI_PMCR_NCNT_MASK) >> CCI_PMCR_NCNT_SHIFT;
832}
833
834static int pmu_get_event_idx(struct cci_pmu_hw_events *hw, struct perf_event *event)
835{
836 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
837 unsigned long cci_event = event->hw.config_base;
838 int idx;
839
840 if (cci_pmu->model->get_event_idx)
841 return cci_pmu->model->get_event_idx(cci_pmu, hw, cci_event);
842
843 /* Generic code to find an unused idx from the mask */
844 for(idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++)
845 if (!test_and_set_bit(idx, hw->used_mask))
846 return idx;
847
848 /* No counters available */
849 return -EAGAIN;
850}
851
852static int pmu_map_event(struct perf_event *event)
853{
854 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
855
856 if (event->attr.type < PERF_TYPE_MAX ||
857 !cci_pmu->model->validate_hw_event)
858 return -ENOENT;
859
860 return cci_pmu->model->validate_hw_event(cci_pmu, event->attr.config);
861}
862
863static int pmu_request_irq(struct cci_pmu *cci_pmu, irq_handler_t handler)
864{
865 int i;
866 struct platform_device *pmu_device = cci_pmu->plat_device;
867
868 if (unlikely(!pmu_device))
869 return -ENODEV;
870
871 if (cci_pmu->nr_irqs < 1) {
872 dev_err(&pmu_device->dev, "no irqs for CCI PMUs defined\n");
873 return -ENODEV;
874 }
875
876 /*
877 * Register all available CCI PMU interrupts. In the interrupt handler
878 * we iterate over the counters checking for interrupt source (the
879 * overflowing counter) and clear it.
880 *
881 * This should allow handling of non-unique interrupt for the counters.
882 */
883 for (i = 0; i < cci_pmu->nr_irqs; i++) {
884 int err = request_irq(cci_pmu->irqs[i], handler, IRQF_SHARED,
885 "arm-cci-pmu", cci_pmu);
886 if (err) {
887 dev_err(&pmu_device->dev, "unable to request IRQ%d for ARM CCI PMU counters\n",
888 cci_pmu->irqs[i]);
889 return err;
890 }
891
892 set_bit(i, &cci_pmu->active_irqs);
893 }
894
895 return 0;
896}
897
898static void pmu_free_irq(struct cci_pmu *cci_pmu)
899{
900 int i;
901
902 for (i = 0; i < cci_pmu->nr_irqs; i++) {
903 if (!test_and_clear_bit(i, &cci_pmu->active_irqs))
904 continue;
905
906 free_irq(cci_pmu->irqs[i], cci_pmu);
907 }
908}
909
910static u32 pmu_read_counter(struct perf_event *event)
911{
912 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
913 struct hw_perf_event *hw_counter = &event->hw;
914 int idx = hw_counter->idx;
915 u32 value;
916
917 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
918 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
919 return 0;
920 }
921 value = pmu_read_register(cci_pmu, idx, CCI_PMU_CNTR);
922
923 return value;
924}
925
926static void pmu_write_counter(struct cci_pmu *cci_pmu, u32 value, int idx)
927{
928 pmu_write_register(cci_pmu, value, idx, CCI_PMU_CNTR);
929}
930
931static void __pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
932{
933 int i;
934 struct cci_pmu_hw_events *cci_hw = &cci_pmu->hw_events;
935
936 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
937 struct perf_event *event = cci_hw->events[i];
938
939 if (WARN_ON(!event))
940 continue;
941 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
942 }
943}
944
945static void pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
946{
947 if (cci_pmu->model->write_counters)
948 cci_pmu->model->write_counters(cci_pmu, mask);
949 else
950 __pmu_write_counters(cci_pmu, mask);
951}
952
953#ifdef CONFIG_ARM_CCI5xx_PMU
954
955/*
956 * CCI-500/CCI-550 has advanced power saving policies, which could gate the
957 * clocks to the PMU counters, which makes the writes to them ineffective.
958 * The only way to write to those counters is when the global counters
959 * are enabled and the particular counter is enabled.
960 *
961 * So we do the following :
962 *
963 * 1) Disable all the PMU counters, saving their current state
964 * 2) Enable the global PMU profiling, now that all counters are
965 * disabled.
966 *
967 * For each counter to be programmed, repeat steps 3-7:
968 *
969 * 3) Write an invalid event code to the event control register for the
970 counter, so that the counters are not modified.
971 * 4) Enable the counter control for the counter.
972 * 5) Set the counter value
973 * 6) Disable the counter
974 * 7) Restore the event in the target counter
975 *
976 * 8) Disable the global PMU.
977 * 9) Restore the status of the rest of the counters.
978 *
979 * We choose an event which for CCI-5xx is guaranteed not to count.
980 * We use the highest possible event code (0x1f) for the master interface 0.
981 */
982#define CCI5xx_INVALID_EVENT ((CCI5xx_PORT_M0 << CCI5xx_PMU_EVENT_SOURCE_SHIFT) | \
983 (CCI5xx_PMU_EVENT_CODE_MASK << CCI5xx_PMU_EVENT_CODE_SHIFT))
984static void cci5xx_pmu_write_counters(struct cci_pmu *cci_pmu, unsigned long *mask)
985{
986 int i;
987 DECLARE_BITMAP(saved_mask, cci_pmu->num_cntrs);
988
989 bitmap_zero(saved_mask, cci_pmu->num_cntrs);
990 pmu_save_counters(cci_pmu, saved_mask);
991
992 /*
993 * Now that all the counters are disabled, we can safely turn the PMU on,
994 * without syncing the status of the counters
995 */
996 __cci_pmu_enable_nosync(cci_pmu);
997
998 for_each_set_bit(i, mask, cci_pmu->num_cntrs) {
999 struct perf_event *event = cci_pmu->hw_events.events[i];
1000
1001 if (WARN_ON(!event))
1002 continue;
1003
1004 pmu_set_event(cci_pmu, i, CCI5xx_INVALID_EVENT);
1005 pmu_enable_counter(cci_pmu, i);
1006 pmu_write_counter(cci_pmu, local64_read(&event->hw.prev_count), i);
1007 pmu_disable_counter(cci_pmu, i);
1008 pmu_set_event(cci_pmu, i, event->hw.config_base);
1009 }
1010
1011 __cci_pmu_disable();
1012
1013 pmu_restore_counters(cci_pmu, saved_mask);
1014}
1015
1016#endif /* CONFIG_ARM_CCI5xx_PMU */
1017
1018static u64 pmu_event_update(struct perf_event *event)
1019{
1020 struct hw_perf_event *hwc = &event->hw;
1021 u64 delta, prev_raw_count, new_raw_count;
1022
1023 do {
1024 prev_raw_count = local64_read(&hwc->prev_count);
1025 new_raw_count = pmu_read_counter(event);
1026 } while (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
1027 new_raw_count) != prev_raw_count);
1028
1029 delta = (new_raw_count - prev_raw_count) & CCI_PMU_CNTR_MASK;
1030
1031 local64_add(delta, &event->count);
1032
1033 return new_raw_count;
1034}
1035
1036static void pmu_read(struct perf_event *event)
1037{
1038 pmu_event_update(event);
1039}
1040
1041static void pmu_event_set_period(struct perf_event *event)
1042{
1043 struct hw_perf_event *hwc = &event->hw;
1044 /*
1045 * The CCI PMU counters have a period of 2^32. To account for the
1046 * possiblity of extreme interrupt latency we program for a period of
1047 * half that. Hopefully we can handle the interrupt before another 2^31
1048 * events occur and the counter overtakes its previous value.
1049 */
1050 u64 val = 1ULL << 31;
1051 local64_set(&hwc->prev_count, val);
1052
1053 /*
1054 * CCI PMU uses PERF_HES_ARCH to keep track of the counters, whose
1055 * values needs to be sync-ed with the s/w state before the PMU is
1056 * enabled.
1057 * Mark this counter for sync.
1058 */
1059 hwc->state |= PERF_HES_ARCH;
1060}
1061
1062static irqreturn_t pmu_handle_irq(int irq_num, void *dev)
1063{
1064 unsigned long flags;
1065 struct cci_pmu *cci_pmu = dev;
1066 struct cci_pmu_hw_events *events = &cci_pmu->hw_events;
1067 int idx, handled = IRQ_NONE;
1068
1069 raw_spin_lock_irqsave(&events->pmu_lock, flags);
1070
1071 /* Disable the PMU while we walk through the counters */
1072 __cci_pmu_disable();
1073 /*
1074 * Iterate over counters and update the corresponding perf events.
1075 * This should work regardless of whether we have per-counter overflow
1076 * interrupt or a combined overflow interrupt.
1077 */
1078 for (idx = 0; idx <= CCI_PMU_CNTR_LAST(cci_pmu); idx++) {
1079 struct perf_event *event = events->events[idx];
1080
1081 if (!event)
1082 continue;
1083
1084 /* Did this counter overflow? */
1085 if (!(pmu_read_register(cci_pmu, idx, CCI_PMU_OVRFLW) &
1086 CCI_PMU_OVRFLW_FLAG))
1087 continue;
1088
1089 pmu_write_register(cci_pmu, CCI_PMU_OVRFLW_FLAG, idx,
1090 CCI_PMU_OVRFLW);
1091
1092 pmu_event_update(event);
1093 pmu_event_set_period(event);
1094 handled = IRQ_HANDLED;
1095 }
1096
1097 /* Enable the PMU and sync possibly overflowed counters */
1098 __cci_pmu_enable_sync(cci_pmu);
1099 raw_spin_unlock_irqrestore(&events->pmu_lock, flags);
1100
1101 return IRQ_RETVAL(handled);
1102}
1103
1104static int cci_pmu_get_hw(struct cci_pmu *cci_pmu)
1105{
1106 int ret = pmu_request_irq(cci_pmu, pmu_handle_irq);
1107 if (ret) {
1108 pmu_free_irq(cci_pmu);
1109 return ret;
1110 }
1111 return 0;
1112}
1113
1114static void cci_pmu_put_hw(struct cci_pmu *cci_pmu)
1115{
1116 pmu_free_irq(cci_pmu);
1117}
1118
1119static void hw_perf_event_destroy(struct perf_event *event)
1120{
1121 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1122 atomic_t *active_events = &cci_pmu->active_events;
1123 struct mutex *reserve_mutex = &cci_pmu->reserve_mutex;
1124
1125 if (atomic_dec_and_mutex_lock(active_events, reserve_mutex)) {
1126 cci_pmu_put_hw(cci_pmu);
1127 mutex_unlock(reserve_mutex);
1128 }
1129}
1130
1131static void cci_pmu_enable(struct pmu *pmu)
1132{
1133 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1134 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1135 int enabled = bitmap_weight(hw_events->used_mask, cci_pmu->num_cntrs);
1136 unsigned long flags;
1137
1138 if (!enabled)
1139 return;
1140
1141 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1142 __cci_pmu_enable_sync(cci_pmu);
1143 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1144
1145}
1146
1147static void cci_pmu_disable(struct pmu *pmu)
1148{
1149 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1150 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1151 unsigned long flags;
1152
1153 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1154 __cci_pmu_disable();
1155 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1156}
1157
1158/*
1159 * Check if the idx represents a non-programmable counter.
1160 * All the fixed event counters are mapped before the programmable
1161 * counters.
1162 */
1163static bool pmu_fixed_hw_idx(struct cci_pmu *cci_pmu, int idx)
1164{
1165 return (idx >= 0) && (idx < cci_pmu->model->fixed_hw_cntrs);
1166}
1167
1168static void cci_pmu_start(struct perf_event *event, int pmu_flags)
1169{
1170 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1171 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1172 struct hw_perf_event *hwc = &event->hw;
1173 int idx = hwc->idx;
1174 unsigned long flags;
1175
1176 /*
1177 * To handle interrupt latency, we always reprogram the period
1178 * regardlesss of PERF_EF_RELOAD.
1179 */
1180 if (pmu_flags & PERF_EF_RELOAD)
1181 WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
1182
1183 hwc->state = 0;
1184
1185 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1186 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1187 return;
1188 }
1189
1190 raw_spin_lock_irqsave(&hw_events->pmu_lock, flags);
1191
1192 /* Configure the counter unless you are counting a fixed event */
1193 if (!pmu_fixed_hw_idx(cci_pmu, idx))
1194 pmu_set_event(cci_pmu, idx, hwc->config_base);
1195
1196 pmu_event_set_period(event);
1197 pmu_enable_counter(cci_pmu, idx);
1198
1199 raw_spin_unlock_irqrestore(&hw_events->pmu_lock, flags);
1200}
1201
1202static void cci_pmu_stop(struct perf_event *event, int pmu_flags)
1203{
1204 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1205 struct hw_perf_event *hwc = &event->hw;
1206 int idx = hwc->idx;
1207
1208 if (hwc->state & PERF_HES_STOPPED)
1209 return;
1210
1211 if (unlikely(!pmu_is_valid_counter(cci_pmu, idx))) {
1212 dev_err(&cci_pmu->plat_device->dev, "Invalid CCI PMU counter %d\n", idx);
1213 return;
1214 }
1215
1216 /*
1217 * We always reprogram the counter, so ignore PERF_EF_UPDATE. See
1218 * cci_pmu_start()
1219 */
1220 pmu_disable_counter(cci_pmu, idx);
1221 pmu_event_update(event);
1222 hwc->state |= PERF_HES_STOPPED | PERF_HES_UPTODATE;
1223}
1224
1225static int cci_pmu_add(struct perf_event *event, int flags)
1226{
1227 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1228 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1229 struct hw_perf_event *hwc = &event->hw;
1230 int idx;
1231 int err = 0;
1232
1233 perf_pmu_disable(event->pmu);
1234
1235 /* If we don't have a space for the counter then finish early. */
1236 idx = pmu_get_event_idx(hw_events, event);
1237 if (idx < 0) {
1238 err = idx;
1239 goto out;
1240 }
1241
1242 event->hw.idx = idx;
1243 hw_events->events[idx] = event;
1244
1245 hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
1246 if (flags & PERF_EF_START)
1247 cci_pmu_start(event, PERF_EF_RELOAD);
1248
1249 /* Propagate our changes to the userspace mapping. */
1250 perf_event_update_userpage(event);
1251
1252out:
1253 perf_pmu_enable(event->pmu);
1254 return err;
1255}
1256
1257static void cci_pmu_del(struct perf_event *event, int flags)
1258{
1259 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1260 struct cci_pmu_hw_events *hw_events = &cci_pmu->hw_events;
1261 struct hw_perf_event *hwc = &event->hw;
1262 int idx = hwc->idx;
1263
1264 cci_pmu_stop(event, PERF_EF_UPDATE);
1265 hw_events->events[idx] = NULL;
1266 clear_bit(idx, hw_events->used_mask);
1267
1268 perf_event_update_userpage(event);
1269}
1270
1271static int
1272validate_event(struct pmu *cci_pmu,
1273 struct cci_pmu_hw_events *hw_events,
1274 struct perf_event *event)
1275{
1276 if (is_software_event(event))
1277 return 1;
1278
1279 /*
1280 * Reject groups spanning multiple HW PMUs (e.g. CPU + CCI). The
1281 * core perf code won't check that the pmu->ctx == leader->ctx
1282 * until after pmu->event_init(event).
1283 */
1284 if (event->pmu != cci_pmu)
1285 return 0;
1286
1287 if (event->state < PERF_EVENT_STATE_OFF)
1288 return 1;
1289
1290 if (event->state == PERF_EVENT_STATE_OFF && !event->attr.enable_on_exec)
1291 return 1;
1292
1293 return pmu_get_event_idx(hw_events, event) >= 0;
1294}
1295
1296static int
1297validate_group(struct perf_event *event)
1298{
1299 struct perf_event *sibling, *leader = event->group_leader;
1300 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1301 unsigned long mask[BITS_TO_LONGS(cci_pmu->num_cntrs)];
1302 struct cci_pmu_hw_events fake_pmu = {
1303 /*
1304 * Initialise the fake PMU. We only need to populate the
1305 * used_mask for the purposes of validation.
1306 */
1307 .used_mask = mask,
1308 };
1309 memset(mask, 0, BITS_TO_LONGS(cci_pmu->num_cntrs) * sizeof(unsigned long));
1310
1311 if (!validate_event(event->pmu, &fake_pmu, leader))
1312 return -EINVAL;
1313
1314 list_for_each_entry(sibling, &leader->sibling_list, group_entry) {
1315 if (!validate_event(event->pmu, &fake_pmu, sibling))
1316 return -EINVAL;
1317 }
1318
1319 if (!validate_event(event->pmu, &fake_pmu, event))
1320 return -EINVAL;
1321
1322 return 0;
1323}
1324
1325static int
1326__hw_perf_event_init(struct perf_event *event)
1327{
1328 struct hw_perf_event *hwc = &event->hw;
1329 int mapping;
1330
1331 mapping = pmu_map_event(event);
1332
1333 if (mapping < 0) {
1334 pr_debug("event %x:%llx not supported\n", event->attr.type,
1335 event->attr.config);
1336 return mapping;
1337 }
1338
1339 /*
1340 * We don't assign an index until we actually place the event onto
1341 * hardware. Use -1 to signify that we haven't decided where to put it
1342 * yet.
1343 */
1344 hwc->idx = -1;
1345 hwc->config_base = 0;
1346 hwc->config = 0;
1347 hwc->event_base = 0;
1348
1349 /*
1350 * Store the event encoding into the config_base field.
1351 */
1352 hwc->config_base |= (unsigned long)mapping;
1353
1354 /*
1355 * Limit the sample_period to half of the counter width. That way, the
1356 * new counter value is far less likely to overtake the previous one
1357 * unless you have some serious IRQ latency issues.
1358 */
1359 hwc->sample_period = CCI_PMU_CNTR_MASK >> 1;
1360 hwc->last_period = hwc->sample_period;
1361 local64_set(&hwc->period_left, hwc->sample_period);
1362
1363 if (event->group_leader != event) {
1364 if (validate_group(event) != 0)
1365 return -EINVAL;
1366 }
1367
1368 return 0;
1369}
1370
1371static int cci_pmu_event_init(struct perf_event *event)
1372{
1373 struct cci_pmu *cci_pmu = to_cci_pmu(event->pmu);
1374 atomic_t *active_events = &cci_pmu->active_events;
1375 int err = 0;
1376 int cpu;
1377
1378 if (event->attr.type != event->pmu->type)
1379 return -ENOENT;
1380
1381 /* Shared by all CPUs, no meaningful state to sample */
1382 if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
1383 return -EOPNOTSUPP;
1384
1385 /* We have no filtering of any kind */
1386 if (event->attr.exclude_user ||
1387 event->attr.exclude_kernel ||
1388 event->attr.exclude_hv ||
1389 event->attr.exclude_idle ||
1390 event->attr.exclude_host ||
1391 event->attr.exclude_guest)
1392 return -EINVAL;
1393
1394 /*
1395 * Following the example set by other "uncore" PMUs, we accept any CPU
1396 * and rewrite its affinity dynamically rather than having perf core
1397 * handle cpu == -1 and pid == -1 for this case.
1398 *
1399 * The perf core will pin online CPUs for the duration of this call and
1400 * the event being installed into its context, so the PMU's CPU can't
1401 * change under our feet.
1402 */
1403 cpu = cpumask_first(&cci_pmu->cpus);
1404 if (event->cpu < 0 || cpu < 0)
1405 return -EINVAL;
1406 event->cpu = cpu;
1407
1408 event->destroy = hw_perf_event_destroy;
1409 if (!atomic_inc_not_zero(active_events)) {
1410 mutex_lock(&cci_pmu->reserve_mutex);
1411 if (atomic_read(active_events) == 0)
1412 err = cci_pmu_get_hw(cci_pmu);
1413 if (!err)
1414 atomic_inc(active_events);
1415 mutex_unlock(&cci_pmu->reserve_mutex);
1416 }
1417 if (err)
1418 return err;
1419
1420 err = __hw_perf_event_init(event);
1421 if (err)
1422 hw_perf_event_destroy(event);
1423
1424 return err;
1425}
1426
1427static ssize_t pmu_cpumask_attr_show(struct device *dev,
1428 struct device_attribute *attr, char *buf)
1429{
1430 struct pmu *pmu = dev_get_drvdata(dev);
1431 struct cci_pmu *cci_pmu = to_cci_pmu(pmu);
1432
1433 int n = scnprintf(buf, PAGE_SIZE - 1, "%*pbl",
1434 cpumask_pr_args(&cci_pmu->cpus));
1435 buf[n++] = '\n';
1436 buf[n] = '\0';
1437 return n;
1438}
1439
1440static struct device_attribute pmu_cpumask_attr =
1441 __ATTR(cpumask, S_IRUGO, pmu_cpumask_attr_show, NULL);
1442
1443static struct attribute *pmu_attrs[] = {
1444 &pmu_cpumask_attr.attr,
1445 NULL,
1446};
1447
1448static struct attribute_group pmu_attr_group = {
1449 .attrs = pmu_attrs,
1450};
1451
1452static struct attribute_group pmu_format_attr_group = {
1453 .name = "format",
1454 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1455};
1456
1457static struct attribute_group pmu_event_attr_group = {
1458 .name = "events",
1459 .attrs = NULL, /* Filled in cci_pmu_init_attrs */
1460};
1461
1462static const struct attribute_group *pmu_attr_groups[] = {
1463 &pmu_attr_group,
1464 &pmu_format_attr_group,
1465 &pmu_event_attr_group,
1466 NULL
1467};
1468
1469static int cci_pmu_init(struct cci_pmu *cci_pmu, struct platform_device *pdev)
1470{
1471 const struct cci_pmu_model *model = cci_pmu->model;
1472 char *name = model->name;
1473 u32 num_cntrs;
1474
1475 pmu_event_attr_group.attrs = model->event_attrs;
1476 pmu_format_attr_group.attrs = model->format_attrs;
1477
1478 cci_pmu->pmu = (struct pmu) {
1479 .name = cci_pmu->model->name,
1480 .task_ctx_nr = perf_invalid_context,
1481 .pmu_enable = cci_pmu_enable,
1482 .pmu_disable = cci_pmu_disable,
1483 .event_init = cci_pmu_event_init,
1484 .add = cci_pmu_add,
1485 .del = cci_pmu_del,
1486 .start = cci_pmu_start,
1487 .stop = cci_pmu_stop,
1488 .read = pmu_read,
1489 .attr_groups = pmu_attr_groups,
1490 };
1491
1492 cci_pmu->plat_device = pdev;
1493 num_cntrs = pmu_get_max_counters();
1494 if (num_cntrs > cci_pmu->model->num_hw_cntrs) {
1495 dev_warn(&pdev->dev,
1496 "PMU implements more counters(%d) than supported by"
1497 " the model(%d), truncated.",
1498 num_cntrs, cci_pmu->model->num_hw_cntrs);
1499 num_cntrs = cci_pmu->model->num_hw_cntrs;
1500 }
1501 cci_pmu->num_cntrs = num_cntrs + cci_pmu->model->fixed_hw_cntrs;
1502
1503 return perf_pmu_register(&cci_pmu->pmu, name, -1);
1504}
1505
1506static int cci_pmu_cpu_notifier(struct notifier_block *self,
1507 unsigned long action, void *hcpu)
1508{
1509 struct cci_pmu *cci_pmu = container_of(self,
1510 struct cci_pmu, cpu_nb);
1511 unsigned int cpu = (long)hcpu;
1512 unsigned int target;
1513
1514 switch (action & ~CPU_TASKS_FROZEN) {
1515 case CPU_DOWN_PREPARE:
1516 if (!cpumask_test_and_clear_cpu(cpu, &cci_pmu->cpus))
1517 break;
1518 target = cpumask_any_but(cpu_online_mask, cpu);
1519 if (target >= nr_cpu_ids) // UP, last CPU
1520 break;
1521 /*
1522 * TODO: migrate context once core races on event->ctx have
1523 * been fixed.
1524 */
1525 cpumask_set_cpu(target, &cci_pmu->cpus);
1526 default:
1527 break;
1528 }
1529
1530 return NOTIFY_OK;
1531}
1532
1533static struct cci_pmu_model cci_pmu_models[] = {
1534#ifdef CONFIG_ARM_CCI400_PMU
1535 [CCI400_R0] = {
1536 .name = "CCI_400",
1537 .fixed_hw_cntrs = 1, /* Cycle counter */
1538 .num_hw_cntrs = 4,
1539 .cntr_size = SZ_4K,
1540 .format_attrs = cci400_pmu_format_attrs,
1541 .event_attrs = cci400_r0_pmu_event_attrs,
1542 .event_ranges = {
1543 [CCI_IF_SLAVE] = {
1544 CCI400_R0_SLAVE_PORT_MIN_EV,
1545 CCI400_R0_SLAVE_PORT_MAX_EV,
1546 },
1547 [CCI_IF_MASTER] = {
1548 CCI400_R0_MASTER_PORT_MIN_EV,
1549 CCI400_R0_MASTER_PORT_MAX_EV,
1550 },
1551 },
1552 .validate_hw_event = cci400_validate_hw_event,
1553 .get_event_idx = cci400_get_event_idx,
1554 },
1555 [CCI400_R1] = {
1556 .name = "CCI_400_r1",
1557 .fixed_hw_cntrs = 1, /* Cycle counter */
1558 .num_hw_cntrs = 4,
1559 .cntr_size = SZ_4K,
1560 .format_attrs = cci400_pmu_format_attrs,
1561 .event_attrs = cci400_r1_pmu_event_attrs,
1562 .event_ranges = {
1563 [CCI_IF_SLAVE] = {
1564 CCI400_R1_SLAVE_PORT_MIN_EV,
1565 CCI400_R1_SLAVE_PORT_MAX_EV,
1566 },
1567 [CCI_IF_MASTER] = {
1568 CCI400_R1_MASTER_PORT_MIN_EV,
1569 CCI400_R1_MASTER_PORT_MAX_EV,
1570 },
1571 },
1572 .validate_hw_event = cci400_validate_hw_event,
1573 .get_event_idx = cci400_get_event_idx,
1574 },
1575#endif
1576#ifdef CONFIG_ARM_CCI5xx_PMU
1577 [CCI500_R0] = {
1578 .name = "CCI_500",
1579 .fixed_hw_cntrs = 0,
1580 .num_hw_cntrs = 8,
1581 .cntr_size = SZ_64K,
1582 .format_attrs = cci5xx_pmu_format_attrs,
1583 .event_attrs = cci5xx_pmu_event_attrs,
1584 .event_ranges = {
1585 [CCI_IF_SLAVE] = {
1586 CCI5xx_SLAVE_PORT_MIN_EV,
1587 CCI5xx_SLAVE_PORT_MAX_EV,
1588 },
1589 [CCI_IF_MASTER] = {
1590 CCI5xx_MASTER_PORT_MIN_EV,
1591 CCI5xx_MASTER_PORT_MAX_EV,
1592 },
1593 [CCI_IF_GLOBAL] = {
1594 CCI5xx_GLOBAL_PORT_MIN_EV,
1595 CCI5xx_GLOBAL_PORT_MAX_EV,
1596 },
1597 },
1598 .validate_hw_event = cci500_validate_hw_event,
1599 .write_counters = cci5xx_pmu_write_counters,
1600 },
1601 [CCI550_R0] = {
1602 .name = "CCI_550",
1603 .fixed_hw_cntrs = 0,
1604 .num_hw_cntrs = 8,
1605 .cntr_size = SZ_64K,
1606 .format_attrs = cci5xx_pmu_format_attrs,
1607 .event_attrs = cci5xx_pmu_event_attrs,
1608 .event_ranges = {
1609 [CCI_IF_SLAVE] = {
1610 CCI5xx_SLAVE_PORT_MIN_EV,
1611 CCI5xx_SLAVE_PORT_MAX_EV,
1612 },
1613 [CCI_IF_MASTER] = {
1614 CCI5xx_MASTER_PORT_MIN_EV,
1615 CCI5xx_MASTER_PORT_MAX_EV,
1616 },
1617 [CCI_IF_GLOBAL] = {
1618 CCI5xx_GLOBAL_PORT_MIN_EV,
1619 CCI5xx_GLOBAL_PORT_MAX_EV,
1620 },
1621 },
1622 .validate_hw_event = cci550_validate_hw_event,
1623 .write_counters = cci5xx_pmu_write_counters,
1624 },
1625#endif
1626};
1627
1628static const struct of_device_id arm_cci_pmu_matches[] = {
1629#ifdef CONFIG_ARM_CCI400_PMU
1630 {
1631 .compatible = "arm,cci-400-pmu",
1632 .data = NULL,
1633 },
1634 {
1635 .compatible = "arm,cci-400-pmu,r0",
1636 .data = &cci_pmu_models[CCI400_R0],
1637 },
1638 {
1639 .compatible = "arm,cci-400-pmu,r1",
1640 .data = &cci_pmu_models[CCI400_R1],
1641 },
1642#endif
1643#ifdef CONFIG_ARM_CCI5xx_PMU
1644 {
1645 .compatible = "arm,cci-500-pmu,r0",
1646 .data = &cci_pmu_models[CCI500_R0],
1647 },
1648 {
1649 .compatible = "arm,cci-550-pmu,r0",
1650 .data = &cci_pmu_models[CCI550_R0],
1651 },
1652#endif
1653 {},
1654};
1655
1656static inline const struct cci_pmu_model *get_cci_model(struct platform_device *pdev)
1657{
1658 const struct of_device_id *match = of_match_node(arm_cci_pmu_matches,
1659 pdev->dev.of_node);
1660 if (!match)
1661 return NULL;
1662 if (match->data)
1663 return match->data;
1664
1665 dev_warn(&pdev->dev, "DEPRECATED compatible property,"
1666 "requires secure access to CCI registers");
1667 return probe_cci_model(pdev);
1668}
1669
1670static bool is_duplicate_irq(int irq, int *irqs, int nr_irqs)
1671{
1672 int i;
1673
1674 for (i = 0; i < nr_irqs; i++)
1675 if (irq == irqs[i])
1676 return true;
1677
1678 return false;
1679}
1680
1681static struct cci_pmu *cci_pmu_alloc(struct platform_device *pdev)
1682{
1683 struct cci_pmu *cci_pmu;
1684 const struct cci_pmu_model *model;
1685
1686 /*
1687 * All allocations are devm_* hence we don't have to free
1688 * them explicitly on an error, as it would end up in driver
1689 * detach.
1690 */
1691 model = get_cci_model(pdev);
1692 if (!model) {
1693 dev_warn(&pdev->dev, "CCI PMU version not supported\n");
1694 return ERR_PTR(-ENODEV);
1695 }
1696
1697 cci_pmu = devm_kzalloc(&pdev->dev, sizeof(*cci_pmu), GFP_KERNEL);
1698 if (!cci_pmu)
1699 return ERR_PTR(-ENOMEM);
1700
1701 cci_pmu->model = model;
1702 cci_pmu->irqs = devm_kcalloc(&pdev->dev, CCI_PMU_MAX_HW_CNTRS(model),
1703 sizeof(*cci_pmu->irqs), GFP_KERNEL);
1704 if (!cci_pmu->irqs)
1705 return ERR_PTR(-ENOMEM);
1706 cci_pmu->hw_events.events = devm_kcalloc(&pdev->dev,
1707 CCI_PMU_MAX_HW_CNTRS(model),
1708 sizeof(*cci_pmu->hw_events.events),
1709 GFP_KERNEL);
1710 if (!cci_pmu->hw_events.events)
1711 return ERR_PTR(-ENOMEM);
1712 cci_pmu->hw_events.used_mask = devm_kcalloc(&pdev->dev,
1713 BITS_TO_LONGS(CCI_PMU_MAX_HW_CNTRS(model)),
1714 sizeof(*cci_pmu->hw_events.used_mask),
1715 GFP_KERNEL);
1716 if (!cci_pmu->hw_events.used_mask)
1717 return ERR_PTR(-ENOMEM);
1718
1719 return cci_pmu;
1720}
1721
1722
1723static int cci_pmu_probe(struct platform_device *pdev)
1724{
1725 struct resource *res;
1726 struct cci_pmu *cci_pmu;
1727 int i, ret, irq;
1728
1729 cci_pmu = cci_pmu_alloc(pdev);
1730 if (IS_ERR(cci_pmu))
1731 return PTR_ERR(cci_pmu);
1732
1733 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1734 cci_pmu->base = devm_ioremap_resource(&pdev->dev, res);
1735 if (IS_ERR(cci_pmu->base))
1736 return -ENOMEM;
1737
1738 /*
1739 * CCI PMU has one overflow interrupt per counter; but some may be tied
1740 * together to a common interrupt.
1741 */
1742 cci_pmu->nr_irqs = 0;
1743 for (i = 0; i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model); i++) {
1744 irq = platform_get_irq(pdev, i);
1745 if (irq < 0)
1746 break;
1747
1748 if (is_duplicate_irq(irq, cci_pmu->irqs, cci_pmu->nr_irqs))
1749 continue;
1750
1751 cci_pmu->irqs[cci_pmu->nr_irqs++] = irq;
1752 }
1753
1754 /*
1755 * Ensure that the device tree has as many interrupts as the number
1756 * of counters.
1757 */
1758 if (i < CCI_PMU_MAX_HW_CNTRS(cci_pmu->model)) {
1759 dev_warn(&pdev->dev, "In-correct number of interrupts: %d, should be %d\n",
1760 i, CCI_PMU_MAX_HW_CNTRS(cci_pmu->model));
1761 return -EINVAL;
1762 }
1763
1764 raw_spin_lock_init(&cci_pmu->hw_events.pmu_lock);
1765 mutex_init(&cci_pmu->reserve_mutex);
1766 atomic_set(&cci_pmu->active_events, 0);
1767 cpumask_set_cpu(smp_processor_id(), &cci_pmu->cpus);
1768
1769 cci_pmu->cpu_nb = (struct notifier_block) {
1770 .notifier_call = cci_pmu_cpu_notifier,
1771 /*
1772 * to migrate uncore events, our notifier should be executed
1773 * before perf core's notifier.
1774 */
1775 .priority = CPU_PRI_PERF + 1,
1776 };
1777
1778 ret = register_cpu_notifier(&cci_pmu->cpu_nb);
1779 if (ret)
1780 return ret;
1781
1782 ret = cci_pmu_init(cci_pmu, pdev);
1783 if (ret) {
1784 unregister_cpu_notifier(&cci_pmu->cpu_nb);
1785 return ret;
1786 }
1787
1788 pr_info("ARM %s PMU driver probed", cci_pmu->model->name);
1789 return 0;
1790}
1791
1792static int cci_platform_probe(struct platform_device *pdev)
1793{
1794 if (!cci_probed())
1795 return -ENODEV;
1796
1797 return of_platform_populate(pdev->dev.of_node, NULL, NULL, &pdev->dev);
1798}
1799
1800static struct platform_driver cci_pmu_driver = {
1801 .driver = {
1802 .name = DRIVER_NAME_PMU,
1803 .of_match_table = arm_cci_pmu_matches,
1804 },
1805 .probe = cci_pmu_probe,
1806};
1807
1808static struct platform_driver cci_platform_driver = {
1809 .driver = {
1810 .name = DRIVER_NAME,
1811 .of_match_table = arm_cci_matches,
1812 },
1813 .probe = cci_platform_probe,
1814};
1815
1816static int __init cci_platform_init(void)
1817{
1818 int ret;
1819
1820 ret = platform_driver_register(&cci_pmu_driver);
1821 if (ret)
1822 return ret;
1823
1824 return platform_driver_register(&cci_platform_driver);
1825}
1826
1827#else /* !CONFIG_ARM_CCI_PMU */
1828
1829static int __init cci_platform_init(void)
1830{
1831 return 0;
1832}
1833
1834#endif /* CONFIG_ARM_CCI_PMU */
1835
1836#ifdef CONFIG_ARM_CCI400_PORT_CTRL
1837
1838#define CCI_PORT_CTRL 0x0
1839#define CCI_CTRL_STATUS 0xc
1840
1841#define CCI_ENABLE_SNOOP_REQ 0x1
1842#define CCI_ENABLE_DVM_REQ 0x2
1843#define CCI_ENABLE_REQ (CCI_ENABLE_SNOOP_REQ | CCI_ENABLE_DVM_REQ)
1844
1845enum cci_ace_port_type {
1846 ACE_INVALID_PORT = 0x0,
1847 ACE_PORT,
1848 ACE_LITE_PORT,
1849};
1850
1851struct cci_ace_port {
1852 void __iomem *base;
1853 unsigned long phys;
1854 enum cci_ace_port_type type;
1855 struct device_node *dn;
1856};
1857
1858static struct cci_ace_port *ports;
1859static unsigned int nb_cci_ports;
1860
1861struct cpu_port {
1862 u64 mpidr;
1863 u32 port;
1864};
1865
1866/*
1867 * Use the port MSB as valid flag, shift can be made dynamic
1868 * by computing number of bits required for port indexes.
1869 * Code disabling CCI cpu ports runs with D-cache invalidated
1870 * and SCTLR bit clear so data accesses must be kept to a minimum
1871 * to improve performance; for now shift is left static to
1872 * avoid one more data access while disabling the CCI port.
1873 */
1874#define PORT_VALID_SHIFT 31
1875#define PORT_VALID (0x1 << PORT_VALID_SHIFT)
1876
1877static inline void init_cpu_port(struct cpu_port *port, u32 index, u64 mpidr)
1878{
1879 port->port = PORT_VALID | index;
1880 port->mpidr = mpidr;
1881}
1882
1883static inline bool cpu_port_is_valid(struct cpu_port *port)
1884{
1885 return !!(port->port & PORT_VALID);
1886}
1887
1888static inline bool cpu_port_match(struct cpu_port *port, u64 mpidr)
1889{
1890 return port->mpidr == (mpidr & MPIDR_HWID_BITMASK);
1891}
1892
1893static struct cpu_port cpu_port[NR_CPUS];
1894
1895/**
1896 * __cci_ace_get_port - Function to retrieve the port index connected to
1897 * a cpu or device.
1898 *
1899 * @dn: device node of the device to look-up
1900 * @type: port type
1901 *
1902 * Return value:
1903 * - CCI port index if success
1904 * - -ENODEV if failure
1905 */
1906static int __cci_ace_get_port(struct device_node *dn, int type)
1907{
1908 int i;
1909 bool ace_match;
1910 struct device_node *cci_portn;
1911
1912 cci_portn = of_parse_phandle(dn, "cci-control-port", 0);
1913 for (i = 0; i < nb_cci_ports; i++) {
1914 ace_match = ports[i].type == type;
1915 if (ace_match && cci_portn == ports[i].dn)
1916 return i;
1917 }
1918 return -ENODEV;
1919}
1920
1921int cci_ace_get_port(struct device_node *dn)
1922{
1923 return __cci_ace_get_port(dn, ACE_LITE_PORT);
1924}
1925EXPORT_SYMBOL_GPL(cci_ace_get_port);
1926
1927static void cci_ace_init_ports(void)
1928{
1929 int port, cpu;
1930 struct device_node *cpun;
1931
1932 /*
1933 * Port index look-up speeds up the function disabling ports by CPU,
1934 * since the logical to port index mapping is done once and does
1935 * not change after system boot.
1936 * The stashed index array is initialized for all possible CPUs
1937 * at probe time.
1938 */
1939 for_each_possible_cpu(cpu) {
1940 /* too early to use cpu->of_node */
1941 cpun = of_get_cpu_node(cpu, NULL);
1942
1943 if (WARN(!cpun, "Missing cpu device node\n"))
1944 continue;
1945
1946 port = __cci_ace_get_port(cpun, ACE_PORT);
1947 if (port < 0)
1948 continue;
1949
1950 init_cpu_port(&cpu_port[cpu], port, cpu_logical_map(cpu));
1951 }
1952
1953 for_each_possible_cpu(cpu) {
1954 WARN(!cpu_port_is_valid(&cpu_port[cpu]),
1955 "CPU %u does not have an associated CCI port\n",
1956 cpu);
1957 }
1958}
1959/*
1960 * Functions to enable/disable a CCI interconnect slave port
1961 *
1962 * They are called by low-level power management code to disable slave
1963 * interfaces snoops and DVM broadcast.
1964 * Since they may execute with cache data allocation disabled and
1965 * after the caches have been cleaned and invalidated the functions provide
1966 * no explicit locking since they may run with D-cache disabled, so normal
1967 * cacheable kernel locks based on ldrex/strex may not work.
1968 * Locking has to be provided by BSP implementations to ensure proper
1969 * operations.
1970 */
1971
1972/**
1973 * cci_port_control() - function to control a CCI port
1974 *
1975 * @port: index of the port to setup
1976 * @enable: if true enables the port, if false disables it
1977 */
1978static void notrace cci_port_control(unsigned int port, bool enable)
1979{
1980 void __iomem *base = ports[port].base;
1981
1982 writel_relaxed(enable ? CCI_ENABLE_REQ : 0, base + CCI_PORT_CTRL);
1983 /*
1984 * This function is called from power down procedures
1985 * and must not execute any instruction that might
1986 * cause the processor to be put in a quiescent state
1987 * (eg wfi). Hence, cpu_relax() can not be added to this
1988 * read loop to optimize power, since it might hide possibly
1989 * disruptive operations.
1990 */
1991 while (readl_relaxed(cci_ctrl_base + CCI_CTRL_STATUS) & 0x1)
1992 ;
1993}
1994
1995/**
1996 * cci_disable_port_by_cpu() - function to disable a CCI port by CPU
1997 * reference
1998 *
1999 * @mpidr: mpidr of the CPU whose CCI port should be disabled
2000 *
2001 * Disabling a CCI port for a CPU implies disabling the CCI port
2002 * controlling that CPU cluster. Code disabling CPU CCI ports
2003 * must make sure that the CPU running the code is the last active CPU
2004 * in the cluster ie all other CPUs are quiescent in a low power state.
2005 *
2006 * Return:
2007 * 0 on success
2008 * -ENODEV on port look-up failure
2009 */
2010int notrace cci_disable_port_by_cpu(u64 mpidr)
2011{
2012 int cpu;
2013 bool is_valid;
2014 for (cpu = 0; cpu < nr_cpu_ids; cpu++) {
2015 is_valid = cpu_port_is_valid(&cpu_port[cpu]);
2016 if (is_valid && cpu_port_match(&cpu_port[cpu], mpidr)) {
2017 cci_port_control(cpu_port[cpu].port, false);
2018 return 0;
2019 }
2020 }
2021 return -ENODEV;
2022}
2023EXPORT_SYMBOL_GPL(cci_disable_port_by_cpu);
2024
2025/**
2026 * cci_enable_port_for_self() - enable a CCI port for calling CPU
2027 *
2028 * Enabling a CCI port for the calling CPU implies enabling the CCI
2029 * port controlling that CPU's cluster. Caller must make sure that the
2030 * CPU running the code is the first active CPU in the cluster and all
2031 * other CPUs are quiescent in a low power state or waiting for this CPU
2032 * to complete the CCI initialization.
2033 *
2034 * Because this is called when the MMU is still off and with no stack,
2035 * the code must be position independent and ideally rely on callee
2036 * clobbered registers only. To achieve this we must code this function
2037 * entirely in assembler.
2038 *
2039 * On success this returns with the proper CCI port enabled. In case of
2040 * any failure this never returns as the inability to enable the CCI is
2041 * fatal and there is no possible recovery at this stage.
2042 */
2043asmlinkage void __naked cci_enable_port_for_self(void)
2044{
2045 asm volatile ("\n"
2046" .arch armv7-a\n"
2047" mrc p15, 0, r0, c0, c0, 5 @ get MPIDR value \n"
2048" and r0, r0, #"__stringify(MPIDR_HWID_BITMASK)" \n"
2049" adr r1, 5f \n"
2050" ldr r2, [r1] \n"
2051" add r1, r1, r2 @ &cpu_port \n"
2052" add ip, r1, %[sizeof_cpu_port] \n"
2053
2054 /* Loop over the cpu_port array looking for a matching MPIDR */
2055"1: ldr r2, [r1, %[offsetof_cpu_port_mpidr_lsb]] \n"
2056" cmp r2, r0 @ compare MPIDR \n"
2057" bne 2f \n"
2058
2059 /* Found a match, now test port validity */
2060" ldr r3, [r1, %[offsetof_cpu_port_port]] \n"
2061" tst r3, #"__stringify(PORT_VALID)" \n"
2062" bne 3f \n"
2063
2064 /* no match, loop with the next cpu_port entry */
2065"2: add r1, r1, %[sizeof_struct_cpu_port] \n"
2066" cmp r1, ip @ done? \n"
2067" blo 1b \n"
2068
2069 /* CCI port not found -- cheaply try to stall this CPU */
2070"cci_port_not_found: \n"
2071" wfi \n"
2072" wfe \n"
2073" b cci_port_not_found \n"
2074
2075 /* Use matched port index to look up the corresponding ports entry */
2076"3: bic r3, r3, #"__stringify(PORT_VALID)" \n"
2077" adr r0, 6f \n"
2078" ldmia r0, {r1, r2} \n"
2079" sub r1, r1, r0 @ virt - phys \n"
2080" ldr r0, [r0, r2] @ *(&ports) \n"
2081" mov r2, %[sizeof_struct_ace_port] \n"
2082" mla r0, r2, r3, r0 @ &ports[index] \n"
2083" sub r0, r0, r1 @ virt_to_phys() \n"
2084
2085 /* Enable the CCI port */
2086" ldr r0, [r0, %[offsetof_port_phys]] \n"
2087" mov r3, %[cci_enable_req]\n"
2088" str r3, [r0, #"__stringify(CCI_PORT_CTRL)"] \n"
2089
2090 /* poll the status reg for completion */
2091" adr r1, 7f \n"
2092" ldr r0, [r1] \n"
2093" ldr r0, [r0, r1] @ cci_ctrl_base \n"
2094"4: ldr r1, [r0, #"__stringify(CCI_CTRL_STATUS)"] \n"
2095" tst r1, %[cci_control_status_bits] \n"
2096" bne 4b \n"
2097
2098" mov r0, #0 \n"
2099" bx lr \n"
2100
2101" .align 2 \n"
2102"5: .word cpu_port - . \n"
2103"6: .word . \n"
2104" .word ports - 6b \n"
2105"7: .word cci_ctrl_phys - . \n"
2106 : :
2107 [sizeof_cpu_port] "i" (sizeof(cpu_port)),
2108 [cci_enable_req] "i" cpu_to_le32(CCI_ENABLE_REQ),
2109 [cci_control_status_bits] "i" cpu_to_le32(1),
2110#ifndef __ARMEB__
2111 [offsetof_cpu_port_mpidr_lsb] "i" (offsetof(struct cpu_port, mpidr)),
2112#else
2113 [offsetof_cpu_port_mpidr_lsb] "i" (offsetof(struct cpu_port, mpidr)+4),
2114#endif
2115 [offsetof_cpu_port_port] "i" (offsetof(struct cpu_port, port)),
2116 [sizeof_struct_cpu_port] "i" (sizeof(struct cpu_port)),
2117 [sizeof_struct_ace_port] "i" (sizeof(struct cci_ace_port)),
2118 [offsetof_port_phys] "i" (offsetof(struct cci_ace_port, phys)) );
2119
2120 unreachable();
2121}
2122
2123/**
2124 * __cci_control_port_by_device() - function to control a CCI port by device
2125 * reference
2126 *
2127 * @dn: device node pointer of the device whose CCI port should be
2128 * controlled
2129 * @enable: if true enables the port, if false disables it
2130 *
2131 * Return:
2132 * 0 on success
2133 * -ENODEV on port look-up failure
2134 */
2135int notrace __cci_control_port_by_device(struct device_node *dn, bool enable)
2136{
2137 int port;
2138
2139 if (!dn)
2140 return -ENODEV;
2141
2142 port = __cci_ace_get_port(dn, ACE_LITE_PORT);
2143 if (WARN_ONCE(port < 0, "node %s ACE lite port look-up failure\n",
2144 dn->full_name))
2145 return -ENODEV;
2146 cci_port_control(port, enable);
2147 return 0;
2148}
2149EXPORT_SYMBOL_GPL(__cci_control_port_by_device);
2150
2151/**
2152 * __cci_control_port_by_index() - function to control a CCI port by port index
2153 *
2154 * @port: port index previously retrieved with cci_ace_get_port()
2155 * @enable: if true enables the port, if false disables it
2156 *
2157 * Return:
2158 * 0 on success
2159 * -ENODEV on port index out of range
2160 * -EPERM if operation carried out on an ACE PORT
2161 */
2162int notrace __cci_control_port_by_index(u32 port, bool enable)
2163{
2164 if (port >= nb_cci_ports || ports[port].type == ACE_INVALID_PORT)
2165 return -ENODEV;
2166 /*
2167 * CCI control for ports connected to CPUS is extremely fragile
2168 * and must be made to go through a specific and controlled
2169 * interface (ie cci_disable_port_by_cpu(); control by general purpose
2170 * indexing is therefore disabled for ACE ports.
2171 */
2172 if (ports[port].type == ACE_PORT)
2173 return -EPERM;
2174
2175 cci_port_control(port, enable);
2176 return 0;
2177}
2178EXPORT_SYMBOL_GPL(__cci_control_port_by_index);
2179
2180static const struct of_device_id arm_cci_ctrl_if_matches[] = {
2181 {.compatible = "arm,cci-400-ctrl-if", },
2182 {},
2183};
2184
2185static int cci_probe_ports(struct device_node *np)
2186{
2187 struct cci_nb_ports const *cci_config;
2188 int ret, i, nb_ace = 0, nb_ace_lite = 0;
2189 struct device_node *cp;
2190 struct resource res;
2191 const char *match_str;
2192 bool is_ace;
2193
2194
2195 cci_config = of_match_node(arm_cci_matches, np)->data;
2196 if (!cci_config)
2197 return -ENODEV;
2198
2199 nb_cci_ports = cci_config->nb_ace + cci_config->nb_ace_lite;
2200
2201 ports = kcalloc(nb_cci_ports, sizeof(*ports), GFP_KERNEL);
2202 if (!ports)
2203 return -ENOMEM;
2204
2205 for_each_child_of_node(np, cp) {
2206 if (!of_match_node(arm_cci_ctrl_if_matches, cp))
2207 continue;
2208
2209 i = nb_ace + nb_ace_lite;
2210
2211 if (i >= nb_cci_ports)
2212 break;
2213
2214 if (of_property_read_string(cp, "interface-type",
2215 &match_str)) {
2216 WARN(1, "node %s missing interface-type property\n",
2217 cp->full_name);
2218 continue;
2219 }
2220 is_ace = strcmp(match_str, "ace") == 0;
2221 if (!is_ace && strcmp(match_str, "ace-lite")) {
2222 WARN(1, "node %s containing invalid interface-type property, skipping it\n",
2223 cp->full_name);
2224 continue;
2225 }
2226
2227 ret = of_address_to_resource(cp, 0, &res);
2228 if (!ret) {
2229 ports[i].base = ioremap(res.start, resource_size(&res));
2230 ports[i].phys = res.start;
2231 }
2232 if (ret || !ports[i].base) {
2233 WARN(1, "unable to ioremap CCI port %d\n", i);
2234 continue;
2235 }
2236
2237 if (is_ace) {
2238 if (WARN_ON(nb_ace >= cci_config->nb_ace))
2239 continue;
2240 ports[i].type = ACE_PORT;
2241 ++nb_ace;
2242 } else {
2243 if (WARN_ON(nb_ace_lite >= cci_config->nb_ace_lite))
2244 continue;
2245 ports[i].type = ACE_LITE_PORT;
2246 ++nb_ace_lite;
2247 }
2248 ports[i].dn = cp;
2249 }
2250
2251 /* initialize a stashed array of ACE ports to speed-up look-up */
2252 cci_ace_init_ports();
2253
2254 /*
2255 * Multi-cluster systems may need this data when non-coherent, during
2256 * cluster power-up/power-down. Make sure it reaches main memory.
2257 */
2258 sync_cache_w(&cci_ctrl_base);
2259 sync_cache_w(&cci_ctrl_phys);
2260 sync_cache_w(&ports);
2261 sync_cache_w(&cpu_port);
2262 __sync_cache_range_w(ports, sizeof(*ports) * nb_cci_ports);
2263 pr_info("ARM CCI driver probed\n");
2264
2265 return 0;
2266}
2267#else /* !CONFIG_ARM_CCI400_PORT_CTRL */
2268static inline int cci_probe_ports(struct device_node *np)
2269{
2270 return 0;
2271}
2272#endif /* CONFIG_ARM_CCI400_PORT_CTRL */
2273
2274static int cci_probe(void)
2275{
2276 int ret;
2277 struct device_node *np;
2278 struct resource res;
2279
2280 np = of_find_matching_node(NULL, arm_cci_matches);
2281 if(!np || !of_device_is_available(np))
2282 return -ENODEV;
2283
2284 ret = of_address_to_resource(np, 0, &res);
2285 if (!ret) {
2286 cci_ctrl_base = ioremap(res.start, resource_size(&res));
2287 cci_ctrl_phys = res.start;
2288 }
2289 if (ret || !cci_ctrl_base) {
2290 WARN(1, "unable to ioremap CCI ctrl\n");
2291 return -ENXIO;
2292 }
2293
2294 return cci_probe_ports(np);
2295}
2296
2297static int cci_init_status = -EAGAIN;
2298static DEFINE_MUTEX(cci_probing);
2299
2300static int cci_init(void)
2301{
2302 if (cci_init_status != -EAGAIN)
2303 return cci_init_status;
2304
2305 mutex_lock(&cci_probing);
2306 if (cci_init_status == -EAGAIN)
2307 cci_init_status = cci_probe();
2308 mutex_unlock(&cci_probing);
2309 return cci_init_status;
2310}
2311
2312/*
2313 * To sort out early init calls ordering a helper function is provided to
2314 * check if the CCI driver has beed initialized. Function check if the driver
2315 * has been initialized, if not it calls the init function that probes
2316 * the driver and updates the return value.
2317 */
2318bool cci_probed(void)
2319{
2320 return cci_init() == 0;
2321}
2322EXPORT_SYMBOL_GPL(cci_probed);
2323
2324early_initcall(cci_init);
2325core_initcall(cci_platform_init);
2326MODULE_LICENSE("GPL");
2327MODULE_DESCRIPTION("ARM CCI support");