Merge tag 'gcc-plugins-v4.9-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git...
[cascardo/linux.git] / arch / mips / kernel / pm-cps.c
1 /*
2  * Copyright (C) 2014 Imagination Technologies
3  * Author: Paul Burton <paul.burton@imgtec.com>
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms of the GNU General Public License as published by the
7  * Free Software Foundation;  either version 2 of the  License, or (at your
8  * option) any later version.
9  */
10
11 #include <linux/cpuhotplug.h>
12 #include <linux/init.h>
13 #include <linux/percpu.h>
14 #include <linux/slab.h>
15
16 #include <asm/asm-offsets.h>
17 #include <asm/cacheflush.h>
18 #include <asm/cacheops.h>
19 #include <asm/idle.h>
20 #include <asm/mips-cm.h>
21 #include <asm/mips-cpc.h>
22 #include <asm/mipsmtregs.h>
23 #include <asm/pm.h>
24 #include <asm/pm-cps.h>
25 #include <asm/smp-cps.h>
26 #include <asm/uasm.h>
27
28 /*
29  * cps_nc_entry_fn - type of a generated non-coherent state entry function
30  * @online: the count of online coupled VPEs
31  * @nc_ready_count: pointer to a non-coherent mapping of the core ready_count
32  *
33  * The code entering & exiting non-coherent states is generated at runtime
34  * using uasm, in order to ensure that the compiler cannot insert a stray
35  * memory access at an unfortunate time and to allow the generation of optimal
36  * core-specific code particularly for cache routines. If coupled_coherence
37  * is non-zero and this is the entry function for the CPS_PM_NC_WAIT state,
38  * returns the number of VPEs that were in the wait state at the point this
39  * VPE left it. Returns garbage if coupled_coherence is zero or this is not
40  * the entry function for CPS_PM_NC_WAIT.
41  */
42 typedef unsigned (*cps_nc_entry_fn)(unsigned online, u32 *nc_ready_count);
43
44 /*
45  * The entry point of the generated non-coherent idle state entry/exit
46  * functions. Actually per-core rather than per-CPU.
47  */
48 static DEFINE_PER_CPU_READ_MOSTLY(cps_nc_entry_fn[CPS_PM_STATE_COUNT],
49                                   nc_asm_enter);
50
51 /* Bitmap indicating which states are supported by the system */
52 DECLARE_BITMAP(state_support, CPS_PM_STATE_COUNT);
53
54 /*
55  * Indicates the number of coupled VPEs ready to operate in a non-coherent
56  * state. Actually per-core rather than per-CPU.
57  */
58 static DEFINE_PER_CPU_ALIGNED(u32*, ready_count);
59 static DEFINE_PER_CPU_ALIGNED(void*, ready_count_alloc);
60
61 /* Indicates online CPUs coupled with the current CPU */
62 static DEFINE_PER_CPU_ALIGNED(cpumask_t, online_coupled);
63
64 /*
65  * Used to synchronize entry to deep idle states. Actually per-core rather
66  * than per-CPU.
67  */
68 static DEFINE_PER_CPU_ALIGNED(atomic_t, pm_barrier);
69
70 /* Saved CPU state across the CPS_PM_POWER_GATED state */
71 DEFINE_PER_CPU_ALIGNED(struct mips_static_suspend_state, cps_cpu_state);
72
73 /* A somewhat arbitrary number of labels & relocs for uasm */
74 static struct uasm_label labels[32];
75 static struct uasm_reloc relocs[32];
76
77 enum mips_reg {
78         zero, at, v0, v1, a0, a1, a2, a3,
79         t0, t1, t2, t3, t4, t5, t6, t7,
80         s0, s1, s2, s3, s4, s5, s6, s7,
81         t8, t9, k0, k1, gp, sp, fp, ra,
82 };
83
84 bool cps_pm_support_state(enum cps_pm_state state)
85 {
86         return test_bit(state, state_support);
87 }
88
89 static void coupled_barrier(atomic_t *a, unsigned online)
90 {
91         /*
92          * This function is effectively the same as
93          * cpuidle_coupled_parallel_barrier, which can't be used here since
94          * there's no cpuidle device.
95          */
96
97         if (!coupled_coherence)
98                 return;
99
100         smp_mb__before_atomic();
101         atomic_inc(a);
102
103         while (atomic_read(a) < online)
104                 cpu_relax();
105
106         if (atomic_inc_return(a) == online * 2) {
107                 atomic_set(a, 0);
108                 return;
109         }
110
111         while (atomic_read(a) > online)
112                 cpu_relax();
113 }
114
115 int cps_pm_enter_state(enum cps_pm_state state)
116 {
117         unsigned cpu = smp_processor_id();
118         unsigned core = current_cpu_data.core;
119         unsigned online, left;
120         cpumask_t *coupled_mask = this_cpu_ptr(&online_coupled);
121         u32 *core_ready_count, *nc_core_ready_count;
122         void *nc_addr;
123         cps_nc_entry_fn entry;
124         struct core_boot_config *core_cfg;
125         struct vpe_boot_config *vpe_cfg;
126
127         /* Check that there is an entry function for this state */
128         entry = per_cpu(nc_asm_enter, core)[state];
129         if (!entry)
130                 return -EINVAL;
131
132         /* Calculate which coupled CPUs (VPEs) are online */
133 #if defined(CONFIG_MIPS_MT) || defined(CONFIG_CPU_MIPSR6)
134         if (cpu_online(cpu)) {
135                 cpumask_and(coupled_mask, cpu_online_mask,
136                             &cpu_sibling_map[cpu]);
137                 online = cpumask_weight(coupled_mask);
138                 cpumask_clear_cpu(cpu, coupled_mask);
139         } else
140 #endif
141         {
142                 cpumask_clear(coupled_mask);
143                 online = 1;
144         }
145
146         /* Setup the VPE to run mips_cps_pm_restore when started again */
147         if (IS_ENABLED(CONFIG_CPU_PM) && state == CPS_PM_POWER_GATED) {
148                 /* Power gating relies upon CPS SMP */
149                 if (!mips_cps_smp_in_use())
150                         return -EINVAL;
151
152                 core_cfg = &mips_cps_core_bootcfg[core];
153                 vpe_cfg = &core_cfg->vpe_config[cpu_vpe_id(&current_cpu_data)];
154                 vpe_cfg->pc = (unsigned long)mips_cps_pm_restore;
155                 vpe_cfg->gp = (unsigned long)current_thread_info();
156                 vpe_cfg->sp = 0;
157         }
158
159         /* Indicate that this CPU might not be coherent */
160         cpumask_clear_cpu(cpu, &cpu_coherent_mask);
161         smp_mb__after_atomic();
162
163         /* Create a non-coherent mapping of the core ready_count */
164         core_ready_count = per_cpu(ready_count, core);
165         nc_addr = kmap_noncoherent(virt_to_page(core_ready_count),
166                                    (unsigned long)core_ready_count);
167         nc_addr += ((unsigned long)core_ready_count & ~PAGE_MASK);
168         nc_core_ready_count = nc_addr;
169
170         /* Ensure ready_count is zero-initialised before the assembly runs */
171         ACCESS_ONCE(*nc_core_ready_count) = 0;
172         coupled_barrier(&per_cpu(pm_barrier, core), online);
173
174         /* Run the generated entry code */
175         left = entry(online, nc_core_ready_count);
176
177         /* Remove the non-coherent mapping of ready_count */
178         kunmap_noncoherent();
179
180         /* Indicate that this CPU is definitely coherent */
181         cpumask_set_cpu(cpu, &cpu_coherent_mask);
182
183         /*
184          * If this VPE is the first to leave the non-coherent wait state then
185          * it needs to wake up any coupled VPEs still running their wait
186          * instruction so that they return to cpuidle, which can then complete
187          * coordination between the coupled VPEs & provide the governor with
188          * a chance to reflect on the length of time the VPEs were in the
189          * idle state.
190          */
191         if (coupled_coherence && (state == CPS_PM_NC_WAIT) && (left == online))
192                 arch_send_call_function_ipi_mask(coupled_mask);
193
194         return 0;
195 }
196
197 static void cps_gen_cache_routine(u32 **pp, struct uasm_label **pl,
198                                   struct uasm_reloc **pr,
199                                   const struct cache_desc *cache,
200                                   unsigned op, int lbl)
201 {
202         unsigned cache_size = cache->ways << cache->waybit;
203         unsigned i;
204         const unsigned unroll_lines = 32;
205
206         /* If the cache isn't present this function has it easy */
207         if (cache->flags & MIPS_CACHE_NOT_PRESENT)
208                 return;
209
210         /* Load base address */
211         UASM_i_LA(pp, t0, (long)CKSEG0);
212
213         /* Calculate end address */
214         if (cache_size < 0x8000)
215                 uasm_i_addiu(pp, t1, t0, cache_size);
216         else
217                 UASM_i_LA(pp, t1, (long)(CKSEG0 + cache_size));
218
219         /* Start of cache op loop */
220         uasm_build_label(pl, *pp, lbl);
221
222         /* Generate the cache ops */
223         for (i = 0; i < unroll_lines; i++) {
224                 if (cpu_has_mips_r6) {
225                         uasm_i_cache(pp, op, 0, t0);
226                         uasm_i_addiu(pp, t0, t0, cache->linesz);
227                 } else {
228                         uasm_i_cache(pp, op, i * cache->linesz, t0);
229                 }
230         }
231
232         if (!cpu_has_mips_r6)
233                 /* Update the base address */
234                 uasm_i_addiu(pp, t0, t0, unroll_lines * cache->linesz);
235
236         /* Loop if we haven't reached the end address yet */
237         uasm_il_bne(pp, pr, t0, t1, lbl);
238         uasm_i_nop(pp);
239 }
240
241 static int cps_gen_flush_fsb(u32 **pp, struct uasm_label **pl,
242                              struct uasm_reloc **pr,
243                              const struct cpuinfo_mips *cpu_info,
244                              int lbl)
245 {
246         unsigned i, fsb_size = 8;
247         unsigned num_loads = (fsb_size * 3) / 2;
248         unsigned line_stride = 2;
249         unsigned line_size = cpu_info->dcache.linesz;
250         unsigned perf_counter, perf_event;
251         unsigned revision = cpu_info->processor_id & PRID_REV_MASK;
252
253         /*
254          * Determine whether this CPU requires an FSB flush, and if so which
255          * performance counter/event reflect stalls due to a full FSB.
256          */
257         switch (__get_cpu_type(cpu_info->cputype)) {
258         case CPU_INTERAPTIV:
259                 perf_counter = 1;
260                 perf_event = 51;
261                 break;
262
263         case CPU_PROAPTIV:
264                 /* Newer proAptiv cores don't require this workaround */
265                 if (revision >= PRID_REV_ENCODE_332(1, 1, 0))
266                         return 0;
267
268                 /* On older ones it's unavailable */
269                 return -1;
270
271         default:
272                 /* Assume that the CPU does not need this workaround */
273                 return 0;
274         }
275
276         /*
277          * Ensure that the fill/store buffer (FSB) is not holding the results
278          * of a prefetch, since if it is then the CPC sequencer may become
279          * stuck in the D3 (ClrBus) state whilst entering a low power state.
280          */
281
282         /* Preserve perf counter setup */
283         uasm_i_mfc0(pp, t2, 25, (perf_counter * 2) + 0); /* PerfCtlN */
284         uasm_i_mfc0(pp, t3, 25, (perf_counter * 2) + 1); /* PerfCntN */
285
286         /* Setup perf counter to count FSB full pipeline stalls */
287         uasm_i_addiu(pp, t0, zero, (perf_event << 5) | 0xf);
288         uasm_i_mtc0(pp, t0, 25, (perf_counter * 2) + 0); /* PerfCtlN */
289         uasm_i_ehb(pp);
290         uasm_i_mtc0(pp, zero, 25, (perf_counter * 2) + 1); /* PerfCntN */
291         uasm_i_ehb(pp);
292
293         /* Base address for loads */
294         UASM_i_LA(pp, t0, (long)CKSEG0);
295
296         /* Start of clear loop */
297         uasm_build_label(pl, *pp, lbl);
298
299         /* Perform some loads to fill the FSB */
300         for (i = 0; i < num_loads; i++)
301                 uasm_i_lw(pp, zero, i * line_size * line_stride, t0);
302
303         /*
304          * Invalidate the new D-cache entries so that the cache will need
305          * refilling (via the FSB) if the loop is executed again.
306          */
307         for (i = 0; i < num_loads; i++) {
308                 uasm_i_cache(pp, Hit_Invalidate_D,
309                              i * line_size * line_stride, t0);
310                 uasm_i_cache(pp, Hit_Writeback_Inv_SD,
311                              i * line_size * line_stride, t0);
312         }
313
314         /* Barrier ensuring previous cache invalidates are complete */
315         uasm_i_sync(pp, STYPE_SYNC);
316         uasm_i_ehb(pp);
317
318         /* Check whether the pipeline stalled due to the FSB being full */
319         uasm_i_mfc0(pp, t1, 25, (perf_counter * 2) + 1); /* PerfCntN */
320
321         /* Loop if it didn't */
322         uasm_il_beqz(pp, pr, t1, lbl);
323         uasm_i_nop(pp);
324
325         /* Restore perf counter 1. The count may well now be wrong... */
326         uasm_i_mtc0(pp, t2, 25, (perf_counter * 2) + 0); /* PerfCtlN */
327         uasm_i_ehb(pp);
328         uasm_i_mtc0(pp, t3, 25, (perf_counter * 2) + 1); /* PerfCntN */
329         uasm_i_ehb(pp);
330
331         return 0;
332 }
333
334 static void cps_gen_set_top_bit(u32 **pp, struct uasm_label **pl,
335                                 struct uasm_reloc **pr,
336                                 unsigned r_addr, int lbl)
337 {
338         uasm_i_lui(pp, t0, uasm_rel_hi(0x80000000));
339         uasm_build_label(pl, *pp, lbl);
340         uasm_i_ll(pp, t1, 0, r_addr);
341         uasm_i_or(pp, t1, t1, t0);
342         uasm_i_sc(pp, t1, 0, r_addr);
343         uasm_il_beqz(pp, pr, t1, lbl);
344         uasm_i_nop(pp);
345 }
346
347 static void *cps_gen_entry_code(unsigned cpu, enum cps_pm_state state)
348 {
349         struct uasm_label *l = labels;
350         struct uasm_reloc *r = relocs;
351         u32 *buf, *p;
352         const unsigned r_online = a0;
353         const unsigned r_nc_count = a1;
354         const unsigned r_pcohctl = t7;
355         const unsigned max_instrs = 256;
356         unsigned cpc_cmd;
357         int err;
358         enum {
359                 lbl_incready = 1,
360                 lbl_poll_cont,
361                 lbl_secondary_hang,
362                 lbl_disable_coherence,
363                 lbl_flush_fsb,
364                 lbl_invicache,
365                 lbl_flushdcache,
366                 lbl_hang,
367                 lbl_set_cont,
368                 lbl_secondary_cont,
369                 lbl_decready,
370         };
371
372         /* Allocate a buffer to hold the generated code */
373         p = buf = kcalloc(max_instrs, sizeof(u32), GFP_KERNEL);
374         if (!buf)
375                 return NULL;
376
377         /* Clear labels & relocs ready for (re)use */
378         memset(labels, 0, sizeof(labels));
379         memset(relocs, 0, sizeof(relocs));
380
381         if (IS_ENABLED(CONFIG_CPU_PM) && state == CPS_PM_POWER_GATED) {
382                 /* Power gating relies upon CPS SMP */
383                 if (!mips_cps_smp_in_use())
384                         goto out_err;
385
386                 /*
387                  * Save CPU state. Note the non-standard calling convention
388                  * with the return address placed in v0 to avoid clobbering
389                  * the ra register before it is saved.
390                  */
391                 UASM_i_LA(&p, t0, (long)mips_cps_pm_save);
392                 uasm_i_jalr(&p, v0, t0);
393                 uasm_i_nop(&p);
394         }
395
396         /*
397          * Load addresses of required CM & CPC registers. This is done early
398          * because they're needed in both the enable & disable coherence steps
399          * but in the coupled case the enable step will only run on one VPE.
400          */
401         UASM_i_LA(&p, r_pcohctl, (long)addr_gcr_cl_coherence());
402
403         if (coupled_coherence) {
404                 /* Increment ready_count */
405                 uasm_i_sync(&p, STYPE_SYNC_MB);
406                 uasm_build_label(&l, p, lbl_incready);
407                 uasm_i_ll(&p, t1, 0, r_nc_count);
408                 uasm_i_addiu(&p, t2, t1, 1);
409                 uasm_i_sc(&p, t2, 0, r_nc_count);
410                 uasm_il_beqz(&p, &r, t2, lbl_incready);
411                 uasm_i_addiu(&p, t1, t1, 1);
412
413                 /* Barrier ensuring all CPUs see the updated r_nc_count value */
414                 uasm_i_sync(&p, STYPE_SYNC_MB);
415
416                 /*
417                  * If this is the last VPE to become ready for non-coherence
418                  * then it should branch below.
419                  */
420                 uasm_il_beq(&p, &r, t1, r_online, lbl_disable_coherence);
421                 uasm_i_nop(&p);
422
423                 if (state < CPS_PM_POWER_GATED) {
424                         /*
425                          * Otherwise this is not the last VPE to become ready
426                          * for non-coherence. It needs to wait until coherence
427                          * has been disabled before proceeding, which it will do
428                          * by polling for the top bit of ready_count being set.
429                          */
430                         uasm_i_addiu(&p, t1, zero, -1);
431                         uasm_build_label(&l, p, lbl_poll_cont);
432                         uasm_i_lw(&p, t0, 0, r_nc_count);
433                         uasm_il_bltz(&p, &r, t0, lbl_secondary_cont);
434                         uasm_i_ehb(&p);
435                         if (cpu_has_mipsmt)
436                                 uasm_i_yield(&p, zero, t1);
437                         uasm_il_b(&p, &r, lbl_poll_cont);
438                         uasm_i_nop(&p);
439                 } else {
440                         /*
441                          * The core will lose power & this VPE will not continue
442                          * so it can simply halt here.
443                          */
444                         if (cpu_has_mipsmt) {
445                                 /* Halt the VPE via C0 tchalt register */
446                                 uasm_i_addiu(&p, t0, zero, TCHALT_H);
447                                 uasm_i_mtc0(&p, t0, 2, 4);
448                         } else if (cpu_has_vp) {
449                                 /* Halt the VP via the CPC VP_STOP register */
450                                 unsigned int vpe_id;
451
452                                 vpe_id = cpu_vpe_id(&cpu_data[cpu]);
453                                 uasm_i_addiu(&p, t0, zero, 1 << vpe_id);
454                                 UASM_i_LA(&p, t1, (long)addr_cpc_cl_vp_stop());
455                                 uasm_i_sw(&p, t0, 0, t1);
456                         } else {
457                                 BUG();
458                         }
459                         uasm_build_label(&l, p, lbl_secondary_hang);
460                         uasm_il_b(&p, &r, lbl_secondary_hang);
461                         uasm_i_nop(&p);
462                 }
463         }
464
465         /*
466          * This is the point of no return - this VPE will now proceed to
467          * disable coherence. At this point we *must* be sure that no other
468          * VPE within the core will interfere with the L1 dcache.
469          */
470         uasm_build_label(&l, p, lbl_disable_coherence);
471
472         /* Invalidate the L1 icache */
473         cps_gen_cache_routine(&p, &l, &r, &cpu_data[cpu].icache,
474                               Index_Invalidate_I, lbl_invicache);
475
476         /* Writeback & invalidate the L1 dcache */
477         cps_gen_cache_routine(&p, &l, &r, &cpu_data[cpu].dcache,
478                               Index_Writeback_Inv_D, lbl_flushdcache);
479
480         /* Barrier ensuring previous cache invalidates are complete */
481         uasm_i_sync(&p, STYPE_SYNC);
482         uasm_i_ehb(&p);
483
484         if (mips_cm_revision() < CM_REV_CM3) {
485                 /*
486                 * Disable all but self interventions. The load from COHCTL is
487                 * defined by the interAptiv & proAptiv SUMs as ensuring that the
488                 *  operation resulting from the preceding store is complete.
489                 */
490                 uasm_i_addiu(&p, t0, zero, 1 << cpu_data[cpu].core);
491                 uasm_i_sw(&p, t0, 0, r_pcohctl);
492                 uasm_i_lw(&p, t0, 0, r_pcohctl);
493
494                 /* Barrier to ensure write to coherence control is complete */
495                 uasm_i_sync(&p, STYPE_SYNC);
496                 uasm_i_ehb(&p);
497         }
498
499         /* Disable coherence */
500         uasm_i_sw(&p, zero, 0, r_pcohctl);
501         uasm_i_lw(&p, t0, 0, r_pcohctl);
502
503         if (state >= CPS_PM_CLOCK_GATED) {
504                 err = cps_gen_flush_fsb(&p, &l, &r, &cpu_data[cpu],
505                                         lbl_flush_fsb);
506                 if (err)
507                         goto out_err;
508
509                 /* Determine the CPC command to issue */
510                 switch (state) {
511                 case CPS_PM_CLOCK_GATED:
512                         cpc_cmd = CPC_Cx_CMD_CLOCKOFF;
513                         break;
514                 case CPS_PM_POWER_GATED:
515                         cpc_cmd = CPC_Cx_CMD_PWRDOWN;
516                         break;
517                 default:
518                         BUG();
519                         goto out_err;
520                 }
521
522                 /* Issue the CPC command */
523                 UASM_i_LA(&p, t0, (long)addr_cpc_cl_cmd());
524                 uasm_i_addiu(&p, t1, zero, cpc_cmd);
525                 uasm_i_sw(&p, t1, 0, t0);
526
527                 if (state == CPS_PM_POWER_GATED) {
528                         /* If anything goes wrong just hang */
529                         uasm_build_label(&l, p, lbl_hang);
530                         uasm_il_b(&p, &r, lbl_hang);
531                         uasm_i_nop(&p);
532
533                         /*
534                          * There's no point generating more code, the core is
535                          * powered down & if powered back up will run from the
536                          * reset vector not from here.
537                          */
538                         goto gen_done;
539                 }
540
541                 /* Barrier to ensure write to CPC command is complete */
542                 uasm_i_sync(&p, STYPE_SYNC);
543                 uasm_i_ehb(&p);
544         }
545
546         if (state == CPS_PM_NC_WAIT) {
547                 /*
548                  * At this point it is safe for all VPEs to proceed with
549                  * execution. This VPE will set the top bit of ready_count
550                  * to indicate to the other VPEs that they may continue.
551                  */
552                 if (coupled_coherence)
553                         cps_gen_set_top_bit(&p, &l, &r, r_nc_count,
554                                             lbl_set_cont);
555
556                 /*
557                  * VPEs which did not disable coherence will continue
558                  * executing, after coherence has been disabled, from this
559                  * point.
560                  */
561                 uasm_build_label(&l, p, lbl_secondary_cont);
562
563                 /* Now perform our wait */
564                 uasm_i_wait(&p, 0);
565         }
566
567         /*
568          * Re-enable coherence. Note that for CPS_PM_NC_WAIT all coupled VPEs
569          * will run this. The first will actually re-enable coherence & the
570          * rest will just be performing a rather unusual nop.
571          */
572         uasm_i_addiu(&p, t0, zero, mips_cm_revision() < CM_REV_CM3
573                                 ? CM_GCR_Cx_COHERENCE_COHDOMAINEN_MSK
574                                 : CM3_GCR_Cx_COHERENCE_COHEN_MSK);
575
576         uasm_i_sw(&p, t0, 0, r_pcohctl);
577         uasm_i_lw(&p, t0, 0, r_pcohctl);
578
579         /* Barrier to ensure write to coherence control is complete */
580         uasm_i_sync(&p, STYPE_SYNC);
581         uasm_i_ehb(&p);
582
583         if (coupled_coherence && (state == CPS_PM_NC_WAIT)) {
584                 /* Decrement ready_count */
585                 uasm_build_label(&l, p, lbl_decready);
586                 uasm_i_sync(&p, STYPE_SYNC_MB);
587                 uasm_i_ll(&p, t1, 0, r_nc_count);
588                 uasm_i_addiu(&p, t2, t1, -1);
589                 uasm_i_sc(&p, t2, 0, r_nc_count);
590                 uasm_il_beqz(&p, &r, t2, lbl_decready);
591                 uasm_i_andi(&p, v0, t1, (1 << fls(smp_num_siblings)) - 1);
592
593                 /* Barrier ensuring all CPUs see the updated r_nc_count value */
594                 uasm_i_sync(&p, STYPE_SYNC_MB);
595         }
596
597         if (coupled_coherence && (state == CPS_PM_CLOCK_GATED)) {
598                 /*
599                  * At this point it is safe for all VPEs to proceed with
600                  * execution. This VPE will set the top bit of ready_count
601                  * to indicate to the other VPEs that they may continue.
602                  */
603                 cps_gen_set_top_bit(&p, &l, &r, r_nc_count, lbl_set_cont);
604
605                 /*
606                  * This core will be reliant upon another core sending a
607                  * power-up command to the CPC in order to resume operation.
608                  * Thus an arbitrary VPE can't trigger the core leaving the
609                  * idle state and the one that disables coherence might as well
610                  * be the one to re-enable it. The rest will continue from here
611                  * after that has been done.
612                  */
613                 uasm_build_label(&l, p, lbl_secondary_cont);
614
615                 /* Barrier ensuring all CPUs see the updated r_nc_count value */
616                 uasm_i_sync(&p, STYPE_SYNC_MB);
617         }
618
619         /* The core is coherent, time to return to C code */
620         uasm_i_jr(&p, ra);
621         uasm_i_nop(&p);
622
623 gen_done:
624         /* Ensure the code didn't exceed the resources allocated for it */
625         BUG_ON((p - buf) > max_instrs);
626         BUG_ON((l - labels) > ARRAY_SIZE(labels));
627         BUG_ON((r - relocs) > ARRAY_SIZE(relocs));
628
629         /* Patch branch offsets */
630         uasm_resolve_relocs(relocs, labels);
631
632         /* Flush the icache */
633         local_flush_icache_range((unsigned long)buf, (unsigned long)p);
634
635         return buf;
636 out_err:
637         kfree(buf);
638         return NULL;
639 }
640
641 static int cps_pm_online_cpu(unsigned int cpu)
642 {
643         enum cps_pm_state state;
644         unsigned core = cpu_data[cpu].core;
645         unsigned dlinesz = cpu_data[cpu].dcache.linesz;
646         void *entry_fn, *core_rc;
647
648         for (state = CPS_PM_NC_WAIT; state < CPS_PM_STATE_COUNT; state++) {
649                 if (per_cpu(nc_asm_enter, core)[state])
650                         continue;
651                 if (!test_bit(state, state_support))
652                         continue;
653
654                 entry_fn = cps_gen_entry_code(cpu, state);
655                 if (!entry_fn) {
656                         pr_err("Failed to generate core %u state %u entry\n",
657                                core, state);
658                         clear_bit(state, state_support);
659                 }
660
661                 per_cpu(nc_asm_enter, core)[state] = entry_fn;
662         }
663
664         if (!per_cpu(ready_count, core)) {
665                 core_rc = kmalloc(dlinesz * 2, GFP_KERNEL);
666                 if (!core_rc) {
667                         pr_err("Failed allocate core %u ready_count\n", core);
668                         return -ENOMEM;
669                 }
670                 per_cpu(ready_count_alloc, core) = core_rc;
671
672                 /* Ensure ready_count is aligned to a cacheline boundary */
673                 core_rc += dlinesz - 1;
674                 core_rc = (void *)((unsigned long)core_rc & ~(dlinesz - 1));
675                 per_cpu(ready_count, core) = core_rc;
676         }
677
678         return 0;
679 }
680
681 static int __init cps_pm_init(void)
682 {
683         /* A CM is required for all non-coherent states */
684         if (!mips_cm_present()) {
685                 pr_warn("pm-cps: no CM, non-coherent states unavailable\n");
686                 return 0;
687         }
688
689         /*
690          * If interrupts were enabled whilst running a wait instruction on a
691          * non-coherent core then the VPE may end up processing interrupts
692          * whilst non-coherent. That would be bad.
693          */
694         if (cpu_wait == r4k_wait_irqoff)
695                 set_bit(CPS_PM_NC_WAIT, state_support);
696         else
697                 pr_warn("pm-cps: non-coherent wait unavailable\n");
698
699         /* Detect whether a CPC is present */
700         if (mips_cpc_present()) {
701                 /* Detect whether clock gating is implemented */
702                 if (read_cpc_cl_stat_conf() & CPC_Cx_STAT_CONF_CLKGAT_IMPL_MSK)
703                         set_bit(CPS_PM_CLOCK_GATED, state_support);
704                 else
705                         pr_warn("pm-cps: CPC does not support clock gating\n");
706
707                 /* Power gating is available with CPS SMP & any CPC */
708                 if (mips_cps_smp_in_use())
709                         set_bit(CPS_PM_POWER_GATED, state_support);
710                 else
711                         pr_warn("pm-cps: CPS SMP not in use, power gating unavailable\n");
712         } else {
713                 pr_warn("pm-cps: no CPC, clock & power gating unavailable\n");
714         }
715
716         return cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "AP_PM_CPS_CPU_ONLINE",
717                                  cps_pm_online_cpu, NULL);
718 }
719 arch_initcall(cps_pm_init);