Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / 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/init.h>
12 #include <linux/percpu.h>
13 #include <linux/slab.h>
14
15 #include <asm/asm-offsets.h>
16 #include <asm/cacheflush.h>
17 #include <asm/cacheops.h>
18 #include <asm/idle.h>
19 #include <asm/mips-cm.h>
20 #include <asm/mips-cpc.h>
21 #include <asm/mipsmtregs.h>
22 #include <asm/pm.h>
23 #include <asm/pm-cps.h>
24 #include <asm/smp-cps.h>
25 #include <asm/uasm.h>
26
27 /*
28  * cps_nc_entry_fn - type of a generated non-coherent state entry function
29  * @online: the count of online coupled VPEs
30  * @nc_ready_count: pointer to a non-coherent mapping of the core ready_count
31  *
32  * The code entering & exiting non-coherent states is generated at runtime
33  * using uasm, in order to ensure that the compiler cannot insert a stray
34  * memory access at an unfortunate time and to allow the generation of optimal
35  * core-specific code particularly for cache routines. If coupled_coherence
36  * is non-zero and this is the entry function for the CPS_PM_NC_WAIT state,
37  * returns the number of VPEs that were in the wait state at the point this
38  * VPE left it. Returns garbage if coupled_coherence is zero or this is not
39  * the entry function for CPS_PM_NC_WAIT.
40  */
41 typedef unsigned (*cps_nc_entry_fn)(unsigned online, u32 *nc_ready_count);
42
43 /*
44  * The entry point of the generated non-coherent idle state entry/exit
45  * functions. Actually per-core rather than per-CPU.
46  */
47 static DEFINE_PER_CPU_READ_MOSTLY(cps_nc_entry_fn[CPS_PM_STATE_COUNT],
48                                   nc_asm_enter);
49
50 /* Bitmap indicating which states are supported by the system */
51 DECLARE_BITMAP(state_support, CPS_PM_STATE_COUNT);
52
53 /*
54  * Indicates the number of coupled VPEs ready to operate in a non-coherent
55  * state. Actually per-core rather than per-CPU.
56  */
57 static DEFINE_PER_CPU_ALIGNED(u32*, ready_count);
58 static DEFINE_PER_CPU_ALIGNED(void*, ready_count_alloc);
59
60 /* Indicates online CPUs coupled with the current CPU */
61 static DEFINE_PER_CPU_ALIGNED(cpumask_t, online_coupled);
62
63 /*
64  * Used to synchronize entry to deep idle states. Actually per-core rather
65  * than per-CPU.
66  */
67 static DEFINE_PER_CPU_ALIGNED(atomic_t, pm_barrier);
68
69 /* Saved CPU state across the CPS_PM_POWER_GATED state */
70 DEFINE_PER_CPU_ALIGNED(struct mips_static_suspend_state, cps_cpu_state);
71
72 /* A somewhat arbitrary number of labels & relocs for uasm */
73 static struct uasm_label labels[32] __initdata;
74 static struct uasm_reloc relocs[32] __initdata;
75
76 /* CPU dependant sync types */
77 static unsigned stype_intervention;
78 static unsigned stype_memory;
79 static unsigned stype_ordering;
80
81 enum mips_reg {
82         zero, at, v0, v1, a0, a1, a2, a3,
83         t0, t1, t2, t3, t4, t5, t6, t7,
84         s0, s1, s2, s3, s4, s5, s6, s7,
85         t8, t9, k0, k1, gp, sp, fp, ra,
86 };
87
88 bool cps_pm_support_state(enum cps_pm_state state)
89 {
90         return test_bit(state, state_support);
91 }
92
93 static void coupled_barrier(atomic_t *a, unsigned online)
94 {
95         /*
96          * This function is effectively the same as
97          * cpuidle_coupled_parallel_barrier, which can't be used here since
98          * there's no cpuidle device.
99          */
100
101         if (!coupled_coherence)
102                 return;
103
104         smp_mb__before_atomic();
105         atomic_inc(a);
106
107         while (atomic_read(a) < online)
108                 cpu_relax();
109
110         if (atomic_inc_return(a) == online * 2) {
111                 atomic_set(a, 0);
112                 return;
113         }
114
115         while (atomic_read(a) > online)
116                 cpu_relax();
117 }
118
119 int cps_pm_enter_state(enum cps_pm_state state)
120 {
121         unsigned cpu = smp_processor_id();
122         unsigned core = current_cpu_data.core;
123         unsigned online, left;
124         cpumask_t *coupled_mask = this_cpu_ptr(&online_coupled);
125         u32 *core_ready_count, *nc_core_ready_count;
126         void *nc_addr;
127         cps_nc_entry_fn entry;
128         struct core_boot_config *core_cfg;
129         struct vpe_boot_config *vpe_cfg;
130
131         /* Check that there is an entry function for this state */
132         entry = per_cpu(nc_asm_enter, core)[state];
133         if (!entry)
134                 return -EINVAL;
135
136         /* Calculate which coupled CPUs (VPEs) are online */
137 #ifdef CONFIG_MIPS_MT
138         if (cpu_online(cpu)) {
139                 cpumask_and(coupled_mask, cpu_online_mask,
140                             &cpu_sibling_map[cpu]);
141                 online = cpumask_weight(coupled_mask);
142                 cpumask_clear_cpu(cpu, coupled_mask);
143         } else
144 #endif
145         {
146                 cpumask_clear(coupled_mask);
147                 online = 1;
148         }
149
150         /* Setup the VPE to run mips_cps_pm_restore when started again */
151         if (config_enabled(CONFIG_CPU_PM) && state == CPS_PM_POWER_GATED) {
152                 /* Power gating relies upon CPS SMP */
153                 if (!mips_cps_smp_in_use())
154                         return -EINVAL;
155
156                 core_cfg = &mips_cps_core_bootcfg[core];
157                 vpe_cfg = &core_cfg->vpe_config[cpu_vpe_id(&current_cpu_data)];
158                 vpe_cfg->pc = (unsigned long)mips_cps_pm_restore;
159                 vpe_cfg->gp = (unsigned long)current_thread_info();
160                 vpe_cfg->sp = 0;
161         }
162
163         /* Indicate that this CPU might not be coherent */
164         cpumask_clear_cpu(cpu, &cpu_coherent_mask);
165         smp_mb__after_atomic();
166
167         /* Create a non-coherent mapping of the core ready_count */
168         core_ready_count = per_cpu(ready_count, core);
169         nc_addr = kmap_noncoherent(virt_to_page(core_ready_count),
170                                    (unsigned long)core_ready_count);
171         nc_addr += ((unsigned long)core_ready_count & ~PAGE_MASK);
172         nc_core_ready_count = nc_addr;
173
174         /* Ensure ready_count is zero-initialised before the assembly runs */
175         ACCESS_ONCE(*nc_core_ready_count) = 0;
176         coupled_barrier(&per_cpu(pm_barrier, core), online);
177
178         /* Run the generated entry code */
179         left = entry(online, nc_core_ready_count);
180
181         /* Remove the non-coherent mapping of ready_count */
182         kunmap_noncoherent();
183
184         /* Indicate that this CPU is definitely coherent */
185         cpumask_set_cpu(cpu, &cpu_coherent_mask);
186
187         /*
188          * If this VPE is the first to leave the non-coherent wait state then
189          * it needs to wake up any coupled VPEs still running their wait
190          * instruction so that they return to cpuidle, which can then complete
191          * coordination between the coupled VPEs & provide the governor with
192          * a chance to reflect on the length of time the VPEs were in the
193          * idle state.
194          */
195         if (coupled_coherence && (state == CPS_PM_NC_WAIT) && (left == online))
196                 arch_send_call_function_ipi_mask(coupled_mask);
197
198         return 0;
199 }
200
201 static void __init cps_gen_cache_routine(u32 **pp, struct uasm_label **pl,
202                                          struct uasm_reloc **pr,
203                                          const struct cache_desc *cache,
204                                          unsigned op, int lbl)
205 {
206         unsigned cache_size = cache->ways << cache->waybit;
207         unsigned i;
208         const unsigned unroll_lines = 32;
209
210         /* If the cache isn't present this function has it easy */
211         if (cache->flags & MIPS_CACHE_NOT_PRESENT)
212                 return;
213
214         /* Load base address */
215         UASM_i_LA(pp, t0, (long)CKSEG0);
216
217         /* Calculate end address */
218         if (cache_size < 0x8000)
219                 uasm_i_addiu(pp, t1, t0, cache_size);
220         else
221                 UASM_i_LA(pp, t1, (long)(CKSEG0 + cache_size));
222
223         /* Start of cache op loop */
224         uasm_build_label(pl, *pp, lbl);
225
226         /* Generate the cache ops */
227         for (i = 0; i < unroll_lines; i++)
228                 uasm_i_cache(pp, op, i * cache->linesz, t0);
229
230         /* Update the base address */
231         uasm_i_addiu(pp, t0, t0, unroll_lines * cache->linesz);
232
233         /* Loop if we haven't reached the end address yet */
234         uasm_il_bne(pp, pr, t0, t1, lbl);
235         uasm_i_nop(pp);
236 }
237
238 static int __init cps_gen_flush_fsb(u32 **pp, struct uasm_label **pl,
239                                     struct uasm_reloc **pr,
240                                     const struct cpuinfo_mips *cpu_info,
241                                     int lbl)
242 {
243         unsigned i, fsb_size = 8;
244         unsigned num_loads = (fsb_size * 3) / 2;
245         unsigned line_stride = 2;
246         unsigned line_size = cpu_info->dcache.linesz;
247         unsigned perf_counter, perf_event;
248         unsigned revision = cpu_info->processor_id & PRID_REV_MASK;
249
250         /*
251          * Determine whether this CPU requires an FSB flush, and if so which
252          * performance counter/event reflect stalls due to a full FSB.
253          */
254         switch (__get_cpu_type(cpu_info->cputype)) {
255         case CPU_INTERAPTIV:
256                 perf_counter = 1;
257                 perf_event = 51;
258                 break;
259
260         case CPU_PROAPTIV:
261                 /* Newer proAptiv cores don't require this workaround */
262                 if (revision >= PRID_REV_ENCODE_332(1, 1, 0))
263                         return 0;
264
265                 /* On older ones it's unavailable */
266                 return -1;
267
268         /* CPUs which do not require the workaround */
269         case CPU_P5600:
270                 return 0;
271
272         default:
273                 WARN_ONCE(1, "pm-cps: FSB flush unsupported for this CPU\n");
274                 return -1;
275         }
276
277         /*
278          * Ensure that the fill/store buffer (FSB) is not holding the results
279          * of a prefetch, since if it is then the CPC sequencer may become
280          * stuck in the D3 (ClrBus) state whilst entering a low power state.
281          */
282
283         /* Preserve perf counter setup */
284         uasm_i_mfc0(pp, t2, 25, (perf_counter * 2) + 0); /* PerfCtlN */
285         uasm_i_mfc0(pp, t3, 25, (perf_counter * 2) + 1); /* PerfCntN */
286
287         /* Setup perf counter to count FSB full pipeline stalls */
288         uasm_i_addiu(pp, t0, zero, (perf_event << 5) | 0xf);
289         uasm_i_mtc0(pp, t0, 25, (perf_counter * 2) + 0); /* PerfCtlN */
290         uasm_i_ehb(pp);
291         uasm_i_mtc0(pp, zero, 25, (perf_counter * 2) + 1); /* PerfCntN */
292         uasm_i_ehb(pp);
293
294         /* Base address for loads */
295         UASM_i_LA(pp, t0, (long)CKSEG0);
296
297         /* Start of clear loop */
298         uasm_build_label(pl, *pp, lbl);
299
300         /* Perform some loads to fill the FSB */
301         for (i = 0; i < num_loads; i++)
302                 uasm_i_lw(pp, zero, i * line_size * line_stride, t0);
303
304         /*
305          * Invalidate the new D-cache entries so that the cache will need
306          * refilling (via the FSB) if the loop is executed again.
307          */
308         for (i = 0; i < num_loads; i++) {
309                 uasm_i_cache(pp, Hit_Invalidate_D,
310                              i * line_size * line_stride, t0);
311                 uasm_i_cache(pp, Hit_Writeback_Inv_SD,
312                              i * line_size * line_stride, t0);
313         }
314
315         /* Completion barrier */
316         uasm_i_sync(pp, stype_memory);
317         uasm_i_ehb(pp);
318
319         /* Check whether the pipeline stalled due to the FSB being full */
320         uasm_i_mfc0(pp, t1, 25, (perf_counter * 2) + 1); /* PerfCntN */
321
322         /* Loop if it didn't */
323         uasm_il_beqz(pp, pr, t1, lbl);
324         uasm_i_nop(pp);
325
326         /* Restore perf counter 1. The count may well now be wrong... */
327         uasm_i_mtc0(pp, t2, 25, (perf_counter * 2) + 0); /* PerfCtlN */
328         uasm_i_ehb(pp);
329         uasm_i_mtc0(pp, t3, 25, (perf_counter * 2) + 1); /* PerfCntN */
330         uasm_i_ehb(pp);
331
332         return 0;
333 }
334
335 static void __init cps_gen_set_top_bit(u32 **pp, struct uasm_label **pl,
336                                        struct uasm_reloc **pr,
337                                        unsigned r_addr, int lbl)
338 {
339         uasm_i_lui(pp, t0, uasm_rel_hi(0x80000000));
340         uasm_build_label(pl, *pp, lbl);
341         uasm_i_ll(pp, t1, 0, r_addr);
342         uasm_i_or(pp, t1, t1, t0);
343         uasm_i_sc(pp, t1, 0, r_addr);
344         uasm_il_beqz(pp, pr, t1, lbl);
345         uasm_i_nop(pp);
346 }
347
348 static void * __init cps_gen_entry_code(unsigned cpu, enum cps_pm_state state)
349 {
350         struct uasm_label *l = labels;
351         struct uasm_reloc *r = relocs;
352         u32 *buf, *p;
353         const unsigned r_online = a0;
354         const unsigned r_nc_count = a1;
355         const unsigned r_pcohctl = t7;
356         const unsigned max_instrs = 256;
357         unsigned cpc_cmd;
358         int err;
359         enum {
360                 lbl_incready = 1,
361                 lbl_poll_cont,
362                 lbl_secondary_hang,
363                 lbl_disable_coherence,
364                 lbl_flush_fsb,
365                 lbl_invicache,
366                 lbl_flushdcache,
367                 lbl_hang,
368                 lbl_set_cont,
369                 lbl_secondary_cont,
370                 lbl_decready,
371         };
372
373         /* Allocate a buffer to hold the generated code */
374         p = buf = kcalloc(max_instrs, sizeof(u32), GFP_KERNEL);
375         if (!buf)
376                 return NULL;
377
378         /* Clear labels & relocs ready for (re)use */
379         memset(labels, 0, sizeof(labels));
380         memset(relocs, 0, sizeof(relocs));
381
382         if (config_enabled(CONFIG_CPU_PM) && state == CPS_PM_POWER_GATED) {
383                 /* Power gating relies upon CPS SMP */
384                 if (!mips_cps_smp_in_use())
385                         goto out_err;
386
387                 /*
388                  * Save CPU state. Note the non-standard calling convention
389                  * with the return address placed in v0 to avoid clobbering
390                  * the ra register before it is saved.
391                  */
392                 UASM_i_LA(&p, t0, (long)mips_cps_pm_save);
393                 uasm_i_jalr(&p, v0, t0);
394                 uasm_i_nop(&p);
395         }
396
397         /*
398          * Load addresses of required CM & CPC registers. This is done early
399          * because they're needed in both the enable & disable coherence steps
400          * but in the coupled case the enable step will only run on one VPE.
401          */
402         UASM_i_LA(&p, r_pcohctl, (long)addr_gcr_cl_coherence());
403
404         if (coupled_coherence) {
405                 /* Increment ready_count */
406                 uasm_i_sync(&p, stype_ordering);
407                 uasm_build_label(&l, p, lbl_incready);
408                 uasm_i_ll(&p, t1, 0, r_nc_count);
409                 uasm_i_addiu(&p, t2, t1, 1);
410                 uasm_i_sc(&p, t2, 0, r_nc_count);
411                 uasm_il_beqz(&p, &r, t2, lbl_incready);
412                 uasm_i_addiu(&p, t1, t1, 1);
413
414                 /* Ordering barrier */
415                 uasm_i_sync(&p, stype_ordering);
416
417                 /*
418                  * If this is the last VPE to become ready for non-coherence
419                  * then it should branch below.
420                  */
421                 uasm_il_beq(&p, &r, t1, r_online, lbl_disable_coherence);
422                 uasm_i_nop(&p);
423
424                 if (state < CPS_PM_POWER_GATED) {
425                         /*
426                          * Otherwise this is not the last VPE to become ready
427                          * for non-coherence. It needs to wait until coherence
428                          * has been disabled before proceeding, which it will do
429                          * by polling for the top bit of ready_count being set.
430                          */
431                         uasm_i_addiu(&p, t1, zero, -1);
432                         uasm_build_label(&l, p, lbl_poll_cont);
433                         uasm_i_lw(&p, t0, 0, r_nc_count);
434                         uasm_il_bltz(&p, &r, t0, lbl_secondary_cont);
435                         uasm_i_ehb(&p);
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                         uasm_i_addiu(&p, t0, zero, TCHALT_H);
445                         uasm_i_mtc0(&p, t0, 2, 4);
446                         uasm_build_label(&l, p, lbl_secondary_hang);
447                         uasm_il_b(&p, &r, lbl_secondary_hang);
448                         uasm_i_nop(&p);
449                 }
450         }
451
452         /*
453          * This is the point of no return - this VPE will now proceed to
454          * disable coherence. At this point we *must* be sure that no other
455          * VPE within the core will interfere with the L1 dcache.
456          */
457         uasm_build_label(&l, p, lbl_disable_coherence);
458
459         /* Invalidate the L1 icache */
460         cps_gen_cache_routine(&p, &l, &r, &cpu_data[cpu].icache,
461                               Index_Invalidate_I, lbl_invicache);
462
463         /* Writeback & invalidate the L1 dcache */
464         cps_gen_cache_routine(&p, &l, &r, &cpu_data[cpu].dcache,
465                               Index_Writeback_Inv_D, lbl_flushdcache);
466
467         /* Completion barrier */
468         uasm_i_sync(&p, stype_memory);
469         uasm_i_ehb(&p);
470
471         /*
472          * Disable all but self interventions. The load from COHCTL is defined
473          * by the interAptiv & proAptiv SUMs as ensuring that the operation
474          * resulting from the preceeding store is complete.
475          */
476         uasm_i_addiu(&p, t0, zero, 1 << cpu_data[cpu].core);
477         uasm_i_sw(&p, t0, 0, r_pcohctl);
478         uasm_i_lw(&p, t0, 0, r_pcohctl);
479
480         /* Sync to ensure previous interventions are complete */
481         uasm_i_sync(&p, stype_intervention);
482         uasm_i_ehb(&p);
483
484         /* Disable coherence */
485         uasm_i_sw(&p, zero, 0, r_pcohctl);
486         uasm_i_lw(&p, t0, 0, r_pcohctl);
487
488         if (state >= CPS_PM_CLOCK_GATED) {
489                 err = cps_gen_flush_fsb(&p, &l, &r, &cpu_data[cpu],
490                                         lbl_flush_fsb);
491                 if (err)
492                         goto out_err;
493
494                 /* Determine the CPC command to issue */
495                 switch (state) {
496                 case CPS_PM_CLOCK_GATED:
497                         cpc_cmd = CPC_Cx_CMD_CLOCKOFF;
498                         break;
499                 case CPS_PM_POWER_GATED:
500                         cpc_cmd = CPC_Cx_CMD_PWRDOWN;
501                         break;
502                 default:
503                         BUG();
504                         goto out_err;
505                 }
506
507                 /* Issue the CPC command */
508                 UASM_i_LA(&p, t0, (long)addr_cpc_cl_cmd());
509                 uasm_i_addiu(&p, t1, zero, cpc_cmd);
510                 uasm_i_sw(&p, t1, 0, t0);
511
512                 if (state == CPS_PM_POWER_GATED) {
513                         /* If anything goes wrong just hang */
514                         uasm_build_label(&l, p, lbl_hang);
515                         uasm_il_b(&p, &r, lbl_hang);
516                         uasm_i_nop(&p);
517
518                         /*
519                          * There's no point generating more code, the core is
520                          * powered down & if powered back up will run from the
521                          * reset vector not from here.
522                          */
523                         goto gen_done;
524                 }
525
526                 /* Completion barrier */
527                 uasm_i_sync(&p, stype_memory);
528                 uasm_i_ehb(&p);
529         }
530
531         if (state == CPS_PM_NC_WAIT) {
532                 /*
533                  * At this point it is safe for all VPEs to proceed with
534                  * execution. This VPE will set the top bit of ready_count
535                  * to indicate to the other VPEs that they may continue.
536                  */
537                 if (coupled_coherence)
538                         cps_gen_set_top_bit(&p, &l, &r, r_nc_count,
539                                             lbl_set_cont);
540
541                 /*
542                  * VPEs which did not disable coherence will continue
543                  * executing, after coherence has been disabled, from this
544                  * point.
545                  */
546                 uasm_build_label(&l, p, lbl_secondary_cont);
547
548                 /* Now perform our wait */
549                 uasm_i_wait(&p, 0);
550         }
551
552         /*
553          * Re-enable coherence. Note that for CPS_PM_NC_WAIT all coupled VPEs
554          * will run this. The first will actually re-enable coherence & the
555          * rest will just be performing a rather unusual nop.
556          */
557         uasm_i_addiu(&p, t0, zero, CM_GCR_Cx_COHERENCE_COHDOMAINEN_MSK);
558         uasm_i_sw(&p, t0, 0, r_pcohctl);
559         uasm_i_lw(&p, t0, 0, r_pcohctl);
560
561         /* Completion barrier */
562         uasm_i_sync(&p, stype_memory);
563         uasm_i_ehb(&p);
564
565         if (coupled_coherence && (state == CPS_PM_NC_WAIT)) {
566                 /* Decrement ready_count */
567                 uasm_build_label(&l, p, lbl_decready);
568                 uasm_i_sync(&p, stype_ordering);
569                 uasm_i_ll(&p, t1, 0, r_nc_count);
570                 uasm_i_addiu(&p, t2, t1, -1);
571                 uasm_i_sc(&p, t2, 0, r_nc_count);
572                 uasm_il_beqz(&p, &r, t2, lbl_decready);
573                 uasm_i_andi(&p, v0, t1, (1 << fls(smp_num_siblings)) - 1);
574
575                 /* Ordering barrier */
576                 uasm_i_sync(&p, stype_ordering);
577         }
578
579         if (coupled_coherence && (state == CPS_PM_CLOCK_GATED)) {
580                 /*
581                  * At this point it is safe for all VPEs to proceed with
582                  * execution. This VPE will set the top bit of ready_count
583                  * to indicate to the other VPEs that they may continue.
584                  */
585                 cps_gen_set_top_bit(&p, &l, &r, r_nc_count, lbl_set_cont);
586
587                 /*
588                  * This core will be reliant upon another core sending a
589                  * power-up command to the CPC in order to resume operation.
590                  * Thus an arbitrary VPE can't trigger the core leaving the
591                  * idle state and the one that disables coherence might as well
592                  * be the one to re-enable it. The rest will continue from here
593                  * after that has been done.
594                  */
595                 uasm_build_label(&l, p, lbl_secondary_cont);
596
597                 /* Ordering barrier */
598                 uasm_i_sync(&p, stype_ordering);
599         }
600
601         /* The core is coherent, time to return to C code */
602         uasm_i_jr(&p, ra);
603         uasm_i_nop(&p);
604
605 gen_done:
606         /* Ensure the code didn't exceed the resources allocated for it */
607         BUG_ON((p - buf) > max_instrs);
608         BUG_ON((l - labels) > ARRAY_SIZE(labels));
609         BUG_ON((r - relocs) > ARRAY_SIZE(relocs));
610
611         /* Patch branch offsets */
612         uasm_resolve_relocs(relocs, labels);
613
614         /* Flush the icache */
615         local_flush_icache_range((unsigned long)buf, (unsigned long)p);
616
617         return buf;
618 out_err:
619         kfree(buf);
620         return NULL;
621 }
622
623 static int __init cps_gen_core_entries(unsigned cpu)
624 {
625         enum cps_pm_state state;
626         unsigned core = cpu_data[cpu].core;
627         unsigned dlinesz = cpu_data[cpu].dcache.linesz;
628         void *entry_fn, *core_rc;
629
630         for (state = CPS_PM_NC_WAIT; state < CPS_PM_STATE_COUNT; state++) {
631                 if (per_cpu(nc_asm_enter, core)[state])
632                         continue;
633                 if (!test_bit(state, state_support))
634                         continue;
635
636                 entry_fn = cps_gen_entry_code(cpu, state);
637                 if (!entry_fn) {
638                         pr_err("Failed to generate core %u state %u entry\n",
639                                core, state);
640                         clear_bit(state, state_support);
641                 }
642
643                 per_cpu(nc_asm_enter, core)[state] = entry_fn;
644         }
645
646         if (!per_cpu(ready_count, core)) {
647                 core_rc = kmalloc(dlinesz * 2, GFP_KERNEL);
648                 if (!core_rc) {
649                         pr_err("Failed allocate core %u ready_count\n", core);
650                         return -ENOMEM;
651                 }
652                 per_cpu(ready_count_alloc, core) = core_rc;
653
654                 /* Ensure ready_count is aligned to a cacheline boundary */
655                 core_rc += dlinesz - 1;
656                 core_rc = (void *)((unsigned long)core_rc & ~(dlinesz - 1));
657                 per_cpu(ready_count, core) = core_rc;
658         }
659
660         return 0;
661 }
662
663 static int __init cps_pm_init(void)
664 {
665         unsigned cpu;
666         int err;
667
668         /* Detect appropriate sync types for the system */
669         switch (current_cpu_data.cputype) {
670         case CPU_INTERAPTIV:
671         case CPU_PROAPTIV:
672         case CPU_M5150:
673         case CPU_P5600:
674                 stype_intervention = 0x2;
675                 stype_memory = 0x3;
676                 stype_ordering = 0x10;
677                 break;
678
679         default:
680                 pr_warn("Power management is using heavyweight sync 0\n");
681         }
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                 goto out;
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         for_each_present_cpu(cpu) {
717                 err = cps_gen_core_entries(cpu);
718                 if (err)
719                         return err;
720         }
721 out:
722         return 0;
723 }
724 arch_initcall(cps_pm_init);