1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * PowerPC64 SLB support.
4 *
5 * Copyright (C) 2004 David Gibson <dwg@au.ibm.com>, IBM
6 * Based on earlier code written by:
7 * Dave Engebretsen and Mike Corrigan {engebret|mikejc}@us.ibm.com
8 * Copyright (c) 2001 Dave Engebretsen
9 * Copyright (C) 2002 Anton Blanchard <anton@au.ibm.com>, IBM
10 */
11
12 #include <asm/asm-prototypes.h>
13 #include <asm/interrupt.h>
14 #include <asm/mmu.h>
15 #include <asm/mmu_context.h>
16 #include <asm/paca.h>
17 #include <asm/ppc-opcode.h>
18 #include <asm/cputable.h>
19 #include <asm/cacheflush.h>
20 #include <asm/smp.h>
21 #include <linux/compiler.h>
22 #include <linux/context_tracking.h>
23 #include <linux/mm_types.h>
24 #include <linux/pgtable.h>
25
26 #include <asm/udbg.h>
27 #include <asm/code-patching.h>
28
29 #include "internal.h"
30
31
32 static long slb_allocate_user(struct mm_struct *mm, unsigned long ea);
33
34 bool stress_slb_enabled __initdata;
35
parse_stress_slb(char * p)36 static int __init parse_stress_slb(char *p)
37 {
38 stress_slb_enabled = true;
39 return 0;
40 }
41 early_param("stress_slb", parse_stress_slb);
42
43 __ro_after_init DEFINE_STATIC_KEY_FALSE(stress_slb_key);
44
assert_slb_presence(bool present,unsigned long ea)45 static void assert_slb_presence(bool present, unsigned long ea)
46 {
47 #ifdef CONFIG_DEBUG_VM
48 unsigned long tmp;
49
50 WARN_ON_ONCE(mfmsr() & MSR_EE);
51
52 if (!cpu_has_feature(CPU_FTR_ARCH_206))
53 return;
54
55 /*
56 * slbfee. requires bit 24 (PPC bit 39) be clear in RB. Hardware
57 * ignores all other bits from 0-27, so just clear them all.
58 */
59 ea &= ~((1UL << SID_SHIFT) - 1);
60 asm volatile(__PPC_SLBFEE_DOT(%0, %1) : "=r"(tmp) : "r"(ea) : "cr0");
61
62 WARN_ON(present == (tmp == 0));
63 #endif
64 }
65
slb_shadow_update(unsigned long ea,int ssize,unsigned long flags,enum slb_index index)66 static inline void slb_shadow_update(unsigned long ea, int ssize,
67 unsigned long flags,
68 enum slb_index index)
69 {
70 struct slb_shadow *p = get_slb_shadow();
71
72 /*
73 * Clear the ESID first so the entry is not valid while we are
74 * updating it. No write barriers are needed here, provided
75 * we only update the current CPU's SLB shadow buffer.
76 */
77 WRITE_ONCE(p->save_area[index].esid, 0);
78 WRITE_ONCE(p->save_area[index].vsid, cpu_to_be64(mk_vsid_data(ea, ssize, flags)));
79 WRITE_ONCE(p->save_area[index].esid, cpu_to_be64(mk_esid_data(ea, ssize, index)));
80 }
81
slb_shadow_clear(enum slb_index index)82 static inline void slb_shadow_clear(enum slb_index index)
83 {
84 WRITE_ONCE(get_slb_shadow()->save_area[index].esid, cpu_to_be64(index));
85 }
86
create_shadowed_slbe(unsigned long ea,int ssize,unsigned long flags,enum slb_index index)87 static inline void create_shadowed_slbe(unsigned long ea, int ssize,
88 unsigned long flags,
89 enum slb_index index)
90 {
91 /*
92 * Updating the shadow buffer before writing the SLB ensures
93 * we don't get a stale entry here if we get preempted by PHYP
94 * between these two statements.
95 */
96 slb_shadow_update(ea, ssize, flags, index);
97
98 assert_slb_presence(false, ea);
99 asm volatile("slbmte %0,%1" :
100 : "r" (mk_vsid_data(ea, ssize, flags)),
101 "r" (mk_esid_data(ea, ssize, index))
102 : "memory" );
103 }
104
105 /*
106 * Insert bolted entries into SLB (which may not be empty, so don't clear
107 * slb_cache_ptr).
108 */
__slb_restore_bolted_realmode(void)109 void __slb_restore_bolted_realmode(void)
110 {
111 struct slb_shadow *p = get_slb_shadow();
112 enum slb_index index;
113
114 /* No isync needed because realmode. */
115 for (index = 0; index < SLB_NUM_BOLTED; index++) {
116 asm volatile("slbmte %0,%1" :
117 : "r" (be64_to_cpu(p->save_area[index].vsid)),
118 "r" (be64_to_cpu(p->save_area[index].esid)));
119 }
120
121 assert_slb_presence(true, local_paca->kstack);
122 }
123
124 /*
125 * Insert the bolted entries into an empty SLB.
126 */
slb_restore_bolted_realmode(void)127 void slb_restore_bolted_realmode(void)
128 {
129 __slb_restore_bolted_realmode();
130 get_paca()->slb_cache_ptr = 0;
131
132 get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1;
133 get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap;
134 }
135
136 /*
137 * This flushes all SLB entries including 0, so it must be realmode.
138 */
slb_flush_all_realmode(void)139 void slb_flush_all_realmode(void)
140 {
141 asm volatile("slbmte %0,%0; slbia" : : "r" (0));
142 }
143
__slb_flush_and_restore_bolted(bool preserve_kernel_lookaside)144 static __always_inline void __slb_flush_and_restore_bolted(bool preserve_kernel_lookaside)
145 {
146 struct slb_shadow *p = get_slb_shadow();
147 unsigned long ksp_esid_data, ksp_vsid_data;
148 u32 ih;
149
150 /*
151 * SLBIA IH=1 on ISA v2.05 and newer processors may preserve lookaside
152 * information created with Class=0 entries, which we use for kernel
153 * SLB entries (the SLB entries themselves are still invalidated).
154 *
155 * Older processors will ignore this optimisation. Over-invalidation
156 * is fine because we never rely on lookaside information existing.
157 */
158 if (preserve_kernel_lookaside)
159 ih = 1;
160 else
161 ih = 0;
162
163 ksp_esid_data = be64_to_cpu(p->save_area[KSTACK_INDEX].esid);
164 ksp_vsid_data = be64_to_cpu(p->save_area[KSTACK_INDEX].vsid);
165
166 asm volatile(PPC_SLBIA(%0)" \n"
167 "slbmte %1, %2 \n"
168 :: "i" (ih),
169 "r" (ksp_vsid_data),
170 "r" (ksp_esid_data)
171 : "memory");
172 }
173
174 /*
175 * This flushes non-bolted entries, it can be run in virtual mode. Must
176 * be called with interrupts disabled.
177 */
slb_flush_and_restore_bolted(void)178 void slb_flush_and_restore_bolted(void)
179 {
180 BUILD_BUG_ON(SLB_NUM_BOLTED != 2);
181
182 WARN_ON(!irqs_disabled());
183
184 /*
185 * We can't take a PMU exception in the following code, so hard
186 * disable interrupts.
187 */
188 hard_irq_disable();
189
190 isync();
191 __slb_flush_and_restore_bolted(false);
192 isync();
193
194 assert_slb_presence(true, get_paca()->kstack);
195
196 get_paca()->slb_cache_ptr = 0;
197
198 get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1;
199 get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap;
200 }
201
slb_save_contents(struct slb_entry * slb_ptr)202 void slb_save_contents(struct slb_entry *slb_ptr)
203 {
204 int i;
205 unsigned long e, v;
206
207 /* Save slb_cache_ptr value. */
208 get_paca()->slb_save_cache_ptr = get_paca()->slb_cache_ptr;
209
210 if (!slb_ptr)
211 return;
212
213 for (i = 0; i < mmu_slb_size; i++) {
214 asm volatile("slbmfee %0,%1" : "=r" (e) : "r" (i));
215 asm volatile("slbmfev %0,%1" : "=r" (v) : "r" (i));
216 slb_ptr->esid = e;
217 slb_ptr->vsid = v;
218 slb_ptr++;
219 }
220 }
221
slb_dump_contents(struct slb_entry * slb_ptr)222 void slb_dump_contents(struct slb_entry *slb_ptr)
223 {
224 int i, n;
225 unsigned long e, v;
226 unsigned long llp;
227
228 if (!slb_ptr)
229 return;
230
231 pr_err("SLB contents of cpu 0x%x\n", smp_processor_id());
232
233 for (i = 0; i < mmu_slb_size; i++) {
234 e = slb_ptr->esid;
235 v = slb_ptr->vsid;
236 slb_ptr++;
237
238 if (!e && !v)
239 continue;
240
241 pr_err("%02d %016lx %016lx %s\n", i, e, v,
242 (e & SLB_ESID_V) ? "VALID" : "NOT VALID");
243
244 if (!(e & SLB_ESID_V))
245 continue;
246
247 llp = v & SLB_VSID_LLP;
248 if (v & SLB_VSID_B_1T) {
249 pr_err(" 1T ESID=%9lx VSID=%13lx LLP:%3lx\n",
250 GET_ESID_1T(e),
251 (v & ~SLB_VSID_B) >> SLB_VSID_SHIFT_1T, llp);
252 } else {
253 pr_err(" 256M ESID=%9lx VSID=%13lx LLP:%3lx\n",
254 GET_ESID(e),
255 (v & ~SLB_VSID_B) >> SLB_VSID_SHIFT, llp);
256 }
257 }
258
259 if (!early_cpu_has_feature(CPU_FTR_ARCH_300)) {
260 /* RR is not so useful as it's often not used for allocation */
261 pr_err("SLB RR allocator index %d\n", get_paca()->stab_rr);
262
263 /* Dump slb cache entires as well. */
264 pr_err("SLB cache ptr value = %d\n", get_paca()->slb_save_cache_ptr);
265 pr_err("Valid SLB cache entries:\n");
266 n = min_t(int, get_paca()->slb_save_cache_ptr, SLB_CACHE_ENTRIES);
267 for (i = 0; i < n; i++)
268 pr_err("%02d EA[0-35]=%9x\n", i, get_paca()->slb_cache[i]);
269 pr_err("Rest of SLB cache entries:\n");
270 for (i = n; i < SLB_CACHE_ENTRIES; i++)
271 pr_err("%02d EA[0-35]=%9x\n", i, get_paca()->slb_cache[i]);
272 }
273 }
274
slb_vmalloc_update(void)275 void slb_vmalloc_update(void)
276 {
277 /*
278 * vmalloc is not bolted, so just have to flush non-bolted.
279 */
280 slb_flush_and_restore_bolted();
281 }
282
preload_hit(struct thread_info * ti,unsigned long esid)283 static bool preload_hit(struct thread_info *ti, unsigned long esid)
284 {
285 unsigned char i;
286
287 for (i = 0; i < ti->slb_preload_nr; i++) {
288 unsigned char idx;
289
290 idx = (ti->slb_preload_tail + i) % SLB_PRELOAD_NR;
291 if (esid == ti->slb_preload_esid[idx])
292 return true;
293 }
294 return false;
295 }
296
preload_add(struct thread_info * ti,unsigned long ea)297 static bool preload_add(struct thread_info *ti, unsigned long ea)
298 {
299 unsigned char idx;
300 unsigned long esid;
301
302 if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
303 /* EAs are stored >> 28 so 256MB segments don't need clearing */
304 if (ea & ESID_MASK_1T)
305 ea &= ESID_MASK_1T;
306 }
307
308 esid = ea >> SID_SHIFT;
309
310 if (preload_hit(ti, esid))
311 return false;
312
313 idx = (ti->slb_preload_tail + ti->slb_preload_nr) % SLB_PRELOAD_NR;
314 ti->slb_preload_esid[idx] = esid;
315 if (ti->slb_preload_nr == SLB_PRELOAD_NR)
316 ti->slb_preload_tail = (ti->slb_preload_tail + 1) % SLB_PRELOAD_NR;
317 else
318 ti->slb_preload_nr++;
319
320 return true;
321 }
322
preload_age(struct thread_info * ti)323 static void preload_age(struct thread_info *ti)
324 {
325 if (!ti->slb_preload_nr)
326 return;
327 ti->slb_preload_nr--;
328 ti->slb_preload_tail = (ti->slb_preload_tail + 1) % SLB_PRELOAD_NR;
329 }
330
slb_setup_new_exec(void)331 void slb_setup_new_exec(void)
332 {
333 struct thread_info *ti = current_thread_info();
334 struct mm_struct *mm = current->mm;
335 unsigned long exec = 0x10000000;
336
337 WARN_ON(irqs_disabled());
338
339 /*
340 * preload cache can only be used to determine whether a SLB
341 * entry exists if it does not start to overflow.
342 */
343 if (ti->slb_preload_nr + 2 > SLB_PRELOAD_NR)
344 return;
345
346 hard_irq_disable();
347
348 /*
349 * We have no good place to clear the slb preload cache on exec,
350 * flush_thread is about the earliest arch hook but that happens
351 * after we switch to the mm and have aleady preloaded the SLBEs.
352 *
353 * For the most part that's probably okay to use entries from the
354 * previous exec, they will age out if unused. It may turn out to
355 * be an advantage to clear the cache before switching to it,
356 * however.
357 */
358
359 /*
360 * preload some userspace segments into the SLB.
361 * Almost all 32 and 64bit PowerPC executables are linked at
362 * 0x10000000 so it makes sense to preload this segment.
363 */
364 if (!is_kernel_addr(exec)) {
365 if (preload_add(ti, exec))
366 slb_allocate_user(mm, exec);
367 }
368
369 /* Libraries and mmaps. */
370 if (!is_kernel_addr(mm->mmap_base)) {
371 if (preload_add(ti, mm->mmap_base))
372 slb_allocate_user(mm, mm->mmap_base);
373 }
374
375 /* see switch_slb */
376 asm volatile("isync" : : : "memory");
377
378 local_irq_enable();
379 }
380
preload_new_slb_context(unsigned long start,unsigned long sp)381 void preload_new_slb_context(unsigned long start, unsigned long sp)
382 {
383 struct thread_info *ti = current_thread_info();
384 struct mm_struct *mm = current->mm;
385 unsigned long heap = mm->start_brk;
386
387 WARN_ON(irqs_disabled());
388
389 /* see above */
390 if (ti->slb_preload_nr + 3 > SLB_PRELOAD_NR)
391 return;
392
393 hard_irq_disable();
394
395 /* Userspace entry address. */
396 if (!is_kernel_addr(start)) {
397 if (preload_add(ti, start))
398 slb_allocate_user(mm, start);
399 }
400
401 /* Top of stack, grows down. */
402 if (!is_kernel_addr(sp)) {
403 if (preload_add(ti, sp))
404 slb_allocate_user(mm, sp);
405 }
406
407 /* Bottom of heap, grows up. */
408 if (heap && !is_kernel_addr(heap)) {
409 if (preload_add(ti, heap))
410 slb_allocate_user(mm, heap);
411 }
412
413 /* see switch_slb */
414 asm volatile("isync" : : : "memory");
415
416 local_irq_enable();
417 }
418
slb_cache_slbie_kernel(unsigned int index)419 static void slb_cache_slbie_kernel(unsigned int index)
420 {
421 unsigned long slbie_data = get_paca()->slb_cache[index];
422 unsigned long ksp = get_paca()->kstack;
423
424 slbie_data <<= SID_SHIFT;
425 slbie_data |= 0xc000000000000000ULL;
426 if ((ksp & slb_esid_mask(mmu_kernel_ssize)) == slbie_data)
427 return;
428 slbie_data |= mmu_kernel_ssize << SLBIE_SSIZE_SHIFT;
429
430 asm volatile("slbie %0" : : "r" (slbie_data));
431 }
432
slb_cache_slbie_user(unsigned int index)433 static void slb_cache_slbie_user(unsigned int index)
434 {
435 unsigned long slbie_data = get_paca()->slb_cache[index];
436
437 slbie_data <<= SID_SHIFT;
438 slbie_data |= user_segment_size(slbie_data) << SLBIE_SSIZE_SHIFT;
439 slbie_data |= SLBIE_C; /* user slbs have C=1 */
440
441 asm volatile("slbie %0" : : "r" (slbie_data));
442 }
443
444 /* Flush all user entries from the segment table of the current processor. */
switch_slb(struct task_struct * tsk,struct mm_struct * mm)445 void switch_slb(struct task_struct *tsk, struct mm_struct *mm)
446 {
447 struct thread_info *ti = task_thread_info(tsk);
448 unsigned char i;
449
450 /*
451 * We need interrupts hard-disabled here, not just soft-disabled,
452 * so that a PMU interrupt can't occur, which might try to access
453 * user memory (to get a stack trace) and possible cause an SLB miss
454 * which would update the slb_cache/slb_cache_ptr fields in the PACA.
455 */
456 hard_irq_disable();
457 isync();
458 if (stress_slb()) {
459 __slb_flush_and_restore_bolted(false);
460 isync();
461 get_paca()->slb_cache_ptr = 0;
462 get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1;
463
464 } else if (cpu_has_feature(CPU_FTR_ARCH_300)) {
465 /*
466 * SLBIA IH=3 invalidates all Class=1 SLBEs and their
467 * associated lookaside structures, which matches what
468 * switch_slb wants. So ARCH_300 does not use the slb
469 * cache.
470 */
471 asm volatile(PPC_SLBIA(3));
472
473 } else {
474 unsigned long offset = get_paca()->slb_cache_ptr;
475
476 if (!mmu_has_feature(MMU_FTR_NO_SLBIE_B) &&
477 offset <= SLB_CACHE_ENTRIES) {
478 /*
479 * Could assert_slb_presence(true) here, but
480 * hypervisor or machine check could have come
481 * in and removed the entry at this point.
482 */
483
484 for (i = 0; i < offset; i++)
485 slb_cache_slbie_user(i);
486
487 /* Workaround POWER5 < DD2.1 issue */
488 if (!cpu_has_feature(CPU_FTR_ARCH_207S) && offset == 1)
489 slb_cache_slbie_user(0);
490
491 } else {
492 /* Flush but retain kernel lookaside information */
493 __slb_flush_and_restore_bolted(true);
494 isync();
495
496 get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1;
497 }
498
499 get_paca()->slb_cache_ptr = 0;
500 }
501 get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap;
502
503 copy_mm_to_paca(mm);
504
505 /*
506 * We gradually age out SLBs after a number of context switches to
507 * reduce reload overhead of unused entries (like we do with FP/VEC
508 * reload). Each time we wrap 256 switches, take an entry out of the
509 * SLB preload cache.
510 */
511 tsk->thread.load_slb++;
512 if (!tsk->thread.load_slb) {
513 unsigned long pc = KSTK_EIP(tsk);
514
515 preload_age(ti);
516 preload_add(ti, pc);
517 }
518
519 for (i = 0; i < ti->slb_preload_nr; i++) {
520 unsigned char idx;
521 unsigned long ea;
522
523 idx = (ti->slb_preload_tail + i) % SLB_PRELOAD_NR;
524 ea = (unsigned long)ti->slb_preload_esid[idx] << SID_SHIFT;
525
526 slb_allocate_user(mm, ea);
527 }
528
529 /*
530 * Synchronize slbmte preloads with possible subsequent user memory
531 * address accesses by the kernel (user mode won't happen until
532 * rfid, which is safe).
533 */
534 isync();
535 }
536
slb_set_size(u16 size)537 void slb_set_size(u16 size)
538 {
539 mmu_slb_size = size;
540 }
541
slb_initialize(void)542 void slb_initialize(void)
543 {
544 unsigned long linear_llp, vmalloc_llp, io_llp;
545 unsigned long lflags;
546 static int slb_encoding_inited;
547 #ifdef CONFIG_SPARSEMEM_VMEMMAP
548 unsigned long vmemmap_llp;
549 #endif
550
551 /* Prepare our SLB miss handler based on our page size */
552 linear_llp = mmu_psize_defs[mmu_linear_psize].sllp;
553 io_llp = mmu_psize_defs[mmu_io_psize].sllp;
554 vmalloc_llp = mmu_psize_defs[mmu_vmalloc_psize].sllp;
555 get_paca()->vmalloc_sllp = SLB_VSID_KERNEL | vmalloc_llp;
556 #ifdef CONFIG_SPARSEMEM_VMEMMAP
557 vmemmap_llp = mmu_psize_defs[mmu_vmemmap_psize].sllp;
558 #endif
559 if (!slb_encoding_inited) {
560 slb_encoding_inited = 1;
561 pr_devel("SLB: linear LLP = %04lx\n", linear_llp);
562 pr_devel("SLB: io LLP = %04lx\n", io_llp);
563 #ifdef CONFIG_SPARSEMEM_VMEMMAP
564 pr_devel("SLB: vmemmap LLP = %04lx\n", vmemmap_llp);
565 #endif
566 }
567
568 get_paca()->stab_rr = SLB_NUM_BOLTED - 1;
569 get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1;
570 get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap;
571
572 lflags = SLB_VSID_KERNEL | linear_llp;
573
574 /* Invalidate the entire SLB (even entry 0) & all the ERATS */
575 asm volatile("isync":::"memory");
576 asm volatile("slbmte %0,%0"::"r" (0) : "memory");
577 asm volatile("isync; slbia; isync":::"memory");
578 create_shadowed_slbe(PAGE_OFFSET, mmu_kernel_ssize, lflags, LINEAR_INDEX);
579
580 /*
581 * For the boot cpu, we're running on the stack in init_thread_union,
582 * which is in the first segment of the linear mapping, and also
583 * get_paca()->kstack hasn't been initialized yet.
584 * For secondary cpus, we need to bolt the kernel stack entry now.
585 */
586 slb_shadow_clear(KSTACK_INDEX);
587 if (raw_smp_processor_id() != boot_cpuid &&
588 (get_paca()->kstack & slb_esid_mask(mmu_kernel_ssize)) > PAGE_OFFSET)
589 create_shadowed_slbe(get_paca()->kstack,
590 mmu_kernel_ssize, lflags, KSTACK_INDEX);
591
592 asm volatile("isync":::"memory");
593 }
594
slb_cache_update(unsigned long esid_data)595 static void slb_cache_update(unsigned long esid_data)
596 {
597 int slb_cache_index;
598
599 if (cpu_has_feature(CPU_FTR_ARCH_300))
600 return; /* ISAv3.0B and later does not use slb_cache */
601
602 if (stress_slb())
603 return;
604
605 /*
606 * Now update slb cache entries
607 */
608 slb_cache_index = local_paca->slb_cache_ptr;
609 if (slb_cache_index < SLB_CACHE_ENTRIES) {
610 /*
611 * We have space in slb cache for optimized switch_slb().
612 * Top 36 bits from esid_data as per ISA
613 */
614 local_paca->slb_cache[slb_cache_index++] = esid_data >> SID_SHIFT;
615 local_paca->slb_cache_ptr++;
616 } else {
617 /*
618 * Our cache is full and the current cache content strictly
619 * doesn't indicate the active SLB conents. Bump the ptr
620 * so that switch_slb() will ignore the cache.
621 */
622 local_paca->slb_cache_ptr = SLB_CACHE_ENTRIES + 1;
623 }
624 }
625
alloc_slb_index(bool kernel)626 static enum slb_index alloc_slb_index(bool kernel)
627 {
628 enum slb_index index;
629
630 /*
631 * The allocation bitmaps can become out of synch with the SLB
632 * when the _switch code does slbie when bolting a new stack
633 * segment and it must not be anywhere else in the SLB. This leaves
634 * a kernel allocated entry that is unused in the SLB. With very
635 * large systems or small segment sizes, the bitmaps could slowly
636 * fill with these entries. They will eventually be cleared out
637 * by the round robin allocator in that case, so it's probably not
638 * worth accounting for.
639 */
640
641 /*
642 * SLBs beyond 32 entries are allocated with stab_rr only
643 * POWER7/8/9 have 32 SLB entries, this could be expanded if a
644 * future CPU has more.
645 */
646 if (local_paca->slb_used_bitmap != U32_MAX) {
647 index = ffz(local_paca->slb_used_bitmap);
648 local_paca->slb_used_bitmap |= 1U << index;
649 if (kernel)
650 local_paca->slb_kern_bitmap |= 1U << index;
651 } else {
652 /* round-robin replacement of slb starting at SLB_NUM_BOLTED. */
653 index = local_paca->stab_rr;
654 if (index < (mmu_slb_size - 1))
655 index++;
656 else
657 index = SLB_NUM_BOLTED;
658 local_paca->stab_rr = index;
659 if (index < 32) {
660 if (kernel)
661 local_paca->slb_kern_bitmap |= 1U << index;
662 else
663 local_paca->slb_kern_bitmap &= ~(1U << index);
664 }
665 }
666 BUG_ON(index < SLB_NUM_BOLTED);
667
668 return index;
669 }
670
slb_insert_entry(unsigned long ea,unsigned long context,unsigned long flags,int ssize,bool kernel)671 static long slb_insert_entry(unsigned long ea, unsigned long context,
672 unsigned long flags, int ssize, bool kernel)
673 {
674 unsigned long vsid;
675 unsigned long vsid_data, esid_data;
676 enum slb_index index;
677
678 vsid = get_vsid(context, ea, ssize);
679 if (!vsid)
680 return -EFAULT;
681
682 /*
683 * There must not be a kernel SLB fault in alloc_slb_index or before
684 * slbmte here or the allocation bitmaps could get out of whack with
685 * the SLB.
686 *
687 * User SLB faults or preloads take this path which might get inlined
688 * into the caller, so add compiler barriers here to ensure unsafe
689 * memory accesses do not come between.
690 */
691 barrier();
692
693 index = alloc_slb_index(kernel);
694
695 vsid_data = __mk_vsid_data(vsid, ssize, flags);
696 esid_data = mk_esid_data(ea, ssize, index);
697
698 /*
699 * No need for an isync before or after this slbmte. The exception
700 * we enter with and the rfid we exit with are context synchronizing.
701 * User preloads should add isync afterwards in case the kernel
702 * accesses user memory before it returns to userspace with rfid.
703 */
704 assert_slb_presence(false, ea);
705 if (stress_slb()) {
706 int slb_cache_index = local_paca->slb_cache_ptr;
707
708 /*
709 * stress_slb() does not use slb cache, repurpose as a
710 * cache of inserted (non-bolted) kernel SLB entries. All
711 * non-bolted kernel entries are flushed on any user fault,
712 * or if there are already 3 non-boled kernel entries.
713 */
714 BUILD_BUG_ON(SLB_CACHE_ENTRIES < 3);
715 if (!kernel || slb_cache_index == 3) {
716 int i;
717
718 for (i = 0; i < slb_cache_index; i++)
719 slb_cache_slbie_kernel(i);
720 slb_cache_index = 0;
721 }
722
723 if (kernel)
724 local_paca->slb_cache[slb_cache_index++] = esid_data >> SID_SHIFT;
725 local_paca->slb_cache_ptr = slb_cache_index;
726 }
727 asm volatile("slbmte %0, %1" : : "r" (vsid_data), "r" (esid_data));
728
729 barrier();
730
731 if (!kernel)
732 slb_cache_update(esid_data);
733
734 return 0;
735 }
736
slb_allocate_kernel(unsigned long ea,unsigned long id)737 static long slb_allocate_kernel(unsigned long ea, unsigned long id)
738 {
739 unsigned long context;
740 unsigned long flags;
741 int ssize;
742
743 if (id == LINEAR_MAP_REGION_ID) {
744
745 /* We only support upto H_MAX_PHYSMEM_BITS */
746 if ((ea & EA_MASK) > (1UL << H_MAX_PHYSMEM_BITS))
747 return -EFAULT;
748
749 flags = SLB_VSID_KERNEL | mmu_psize_defs[mmu_linear_psize].sllp;
750
751 #ifdef CONFIG_SPARSEMEM_VMEMMAP
752 } else if (id == VMEMMAP_REGION_ID) {
753
754 if (ea >= H_VMEMMAP_END)
755 return -EFAULT;
756
757 flags = SLB_VSID_KERNEL | mmu_psize_defs[mmu_vmemmap_psize].sllp;
758 #endif
759 } else if (id == VMALLOC_REGION_ID) {
760
761 if (ea >= H_VMALLOC_END)
762 return -EFAULT;
763
764 flags = local_paca->vmalloc_sllp;
765
766 } else if (id == IO_REGION_ID) {
767
768 if (ea >= H_KERN_IO_END)
769 return -EFAULT;
770
771 flags = SLB_VSID_KERNEL | mmu_psize_defs[mmu_io_psize].sllp;
772
773 } else {
774 return -EFAULT;
775 }
776
777 ssize = MMU_SEGSIZE_1T;
778 if (!mmu_has_feature(MMU_FTR_1T_SEGMENT))
779 ssize = MMU_SEGSIZE_256M;
780
781 context = get_kernel_context(ea);
782
783 return slb_insert_entry(ea, context, flags, ssize, true);
784 }
785
slb_allocate_user(struct mm_struct * mm,unsigned long ea)786 static long slb_allocate_user(struct mm_struct *mm, unsigned long ea)
787 {
788 unsigned long context;
789 unsigned long flags;
790 int bpsize;
791 int ssize;
792
793 /*
794 * consider this as bad access if we take a SLB miss
795 * on an address above addr limit.
796 */
797 if (ea >= mm_ctx_slb_addr_limit(&mm->context))
798 return -EFAULT;
799
800 context = get_user_context(&mm->context, ea);
801 if (!context)
802 return -EFAULT;
803
804 if (unlikely(ea >= H_PGTABLE_RANGE)) {
805 WARN_ON(1);
806 return -EFAULT;
807 }
808
809 ssize = user_segment_size(ea);
810
811 bpsize = get_slice_psize(mm, ea);
812 flags = SLB_VSID_USER | mmu_psize_defs[bpsize].sllp;
813
814 return slb_insert_entry(ea, context, flags, ssize, false);
815 }
816
DEFINE_INTERRUPT_HANDLER_RAW(do_slb_fault)817 DEFINE_INTERRUPT_HANDLER_RAW(do_slb_fault)
818 {
819 unsigned long ea = regs->dar;
820 unsigned long id = get_region_id(ea);
821
822 /* IRQs are not reconciled here, so can't check irqs_disabled */
823 VM_WARN_ON(mfmsr() & MSR_EE);
824
825 if (unlikely(!(regs->msr & MSR_RI)))
826 return -EINVAL;
827
828 /*
829 * SLB kernel faults must be very careful not to touch anything that is
830 * not bolted. E.g., PACA and global variables are okay, mm->context
831 * stuff is not. SLB user faults may access all of memory (and induce
832 * one recursive SLB kernel fault), so the kernel fault must not
833 * trample on the user fault state at those points.
834 */
835
836 /*
837 * This is a raw interrupt handler, for performance, so that
838 * fast_interrupt_return can be used. The handler must not touch local
839 * irq state, or schedule. We could test for usermode and upgrade to a
840 * normal process context (synchronous) interrupt for those, which
841 * would make them first-class kernel code and able to be traced and
842 * instrumented, although performance would suffer a bit, it would
843 * probably be a good tradeoff.
844 */
845 if (id >= LINEAR_MAP_REGION_ID) {
846 long err;
847 #ifdef CONFIG_DEBUG_VM
848 /* Catch recursive kernel SLB faults. */
849 BUG_ON(local_paca->in_kernel_slb_handler);
850 local_paca->in_kernel_slb_handler = 1;
851 #endif
852 err = slb_allocate_kernel(ea, id);
853 #ifdef CONFIG_DEBUG_VM
854 local_paca->in_kernel_slb_handler = 0;
855 #endif
856 return err;
857 } else {
858 struct mm_struct *mm = current->mm;
859 long err;
860
861 if (unlikely(!mm))
862 return -EFAULT;
863
864 err = slb_allocate_user(mm, ea);
865 if (!err)
866 preload_add(current_thread_info(), ea);
867
868 return err;
869 }
870 }
871
DEFINE_INTERRUPT_HANDLER(do_bad_slb_fault)872 DEFINE_INTERRUPT_HANDLER(do_bad_slb_fault)
873 {
874 int err = regs->result;
875
876 if (err == -EFAULT) {
877 if (user_mode(regs))
878 _exception(SIGSEGV, regs, SEGV_BNDERR, regs->dar);
879 else
880 bad_page_fault(regs, SIGSEGV);
881 } else if (err == -EINVAL) {
882 unrecoverable_exception(regs);
883 } else {
884 BUG();
885 }
886 }
887