Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * TLB Management (flush/create/diagnostics) for ARC700
4 *
5 * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
6 *
7 * vineetg: Aug 2011
8 * -Reintroduce duplicate PD fixup - some customer chips still have the issue
9 *
10 * vineetg: May 2011
11 * -No need to flush_cache_page( ) for each call to update_mmu_cache()
12 * some of the LMBench tests improved amazingly
13 * = page-fault thrice as fast (75 usec to 28 usec)
14 * = mmap twice as fast (9.6 msec to 4.6 msec),
15 * = fork (5.3 msec to 3.7 msec)
16 *
17 * vineetg: April 2011 :
18 * -MMU v3: PD{0,1} bits layout changed: They don't overlap anymore,
19 * helps avoid a shift when preparing PD0 from PTE
20 *
21 * vineetg: April 2011 : Preparing for MMU V3
22 * -MMU v2/v3 BCRs decoded differently
23 * -Remove TLB_SIZE hardcoding as it's variable now: 256 or 512
24 * -tlb_entry_erase( ) can be void
25 * -local_flush_tlb_range( ):
26 * = need not "ceil" @end
27 * = walks MMU only if range spans < 32 entries, as opposed to 256
28 *
29 * Vineetg: Sept 10th 2008
30 * -Changes related to MMU v2 (Rel 4.8)
31 *
32 * Vineetg: Aug 29th 2008
33 * -In TLB Flush operations (Metal Fix MMU) there is a explict command to
34 * flush Micro-TLBS. If TLB Index Reg is invalid prior to TLBIVUTLB cmd,
35 * it fails. Thus need to load it with ANY valid value before invoking
36 * TLBIVUTLB cmd
37 *
38 * Vineetg: Aug 21th 2008:
39 * -Reduced the duration of IRQ lockouts in TLB Flush routines
40 * -Multiple copies of TLB erase code seperated into a "single" function
41 * -In TLB Flush routines, interrupt disabling moved UP to retrieve ASID
42 * in interrupt-safe region.
43 *
44 * Vineetg: April 23rd Bug #93131
45 * Problem: tlb_flush_kernel_range() doesn't do anything if the range to
46 * flush is more than the size of TLB itself.
47 *
48 * Rahul Trivedi : Codito Technologies 2004
49 */
50
51#include <linux/module.h>
52#include <linux/bug.h>
53#include <linux/mm_types.h>
54
55#include <asm/arcregs.h>
56#include <asm/setup.h>
57#include <asm/mmu_context.h>
58#include <asm/mmu.h>
59
60/* Need for ARC MMU v2
61 *
62 * ARC700 MMU-v1 had a Joint-TLB for Code and Data and is 2 way set-assoc.
63 * For a memcpy operation with 3 players (src/dst/code) such that all 3 pages
64 * map into same set, there would be contention for the 2 ways causing severe
65 * Thrashing.
66 *
67 * Although J-TLB is 2 way set assoc, ARC700 caches J-TLB into uTLBS which has
68 * much higher associativity. u-D-TLB is 8 ways, u-I-TLB is 4 ways.
69 * Given this, the thrasing problem should never happen because once the 3
70 * J-TLB entries are created (even though 3rd will knock out one of the prev
71 * two), the u-D-TLB and u-I-TLB will have what is required to accomplish memcpy
72 *
73 * Yet we still see the Thrashing because a J-TLB Write cause flush of u-TLBs.
74 * This is a simple design for keeping them in sync. So what do we do?
75 * The solution which James came up was pretty neat. It utilised the assoc
76 * of uTLBs by not invalidating always but only when absolutely necessary.
77 *
78 * - Existing TLB commands work as before
79 * - New command (TLBWriteNI) for TLB write without clearing uTLBs
80 * - New command (TLBIVUTLB) to invalidate uTLBs.
81 *
82 * The uTLBs need only be invalidated when pages are being removed from the
83 * OS page table. If a 'victim' TLB entry is being overwritten in the main TLB
84 * as a result of a miss, the removed entry is still allowed to exist in the
85 * uTLBs as it is still valid and present in the OS page table. This allows the
86 * full associativity of the uTLBs to hide the limited associativity of the main
87 * TLB.
88 *
89 * During a miss handler, the new "TLBWriteNI" command is used to load
90 * entries without clearing the uTLBs.
91 *
92 * When the OS page table is updated, TLB entries that may be associated with a
93 * removed page are removed (flushed) from the TLB using TLBWrite. In this
94 * circumstance, the uTLBs must also be cleared. This is done by using the
95 * existing TLBWrite command. An explicit IVUTLB is also required for those
96 * corner cases when TLBWrite was not executed at all because the corresp
97 * J-TLB entry got evicted/replaced.
98 */
99
100
101/* A copy of the ASID from the PID reg is kept in asid_cache */
102DEFINE_PER_CPU(unsigned int, asid_cache) = MM_CTXT_FIRST_CYCLE;
103
104static int __read_mostly pae_exists;
105
106/*
107 * Utility Routine to erase a J-TLB entry
108 * Caller needs to setup Index Reg (manually or via getIndex)
109 */
110static inline void __tlb_entry_erase(void)
111{
112 write_aux_reg(ARC_REG_TLBPD1, 0);
113
114 if (is_pae40_enabled())
115 write_aux_reg(ARC_REG_TLBPD1HI, 0);
116
117 write_aux_reg(ARC_REG_TLBPD0, 0);
118 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
119}
120
121#if (CONFIG_ARC_MMU_VER < 4)
122
123static inline unsigned int tlb_entry_lkup(unsigned long vaddr_n_asid)
124{
125 unsigned int idx;
126
127 write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid);
128
129 write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe);
130 idx = read_aux_reg(ARC_REG_TLBINDEX);
131
132 return idx;
133}
134
135static void tlb_entry_erase(unsigned int vaddr_n_asid)
136{
137 unsigned int idx;
138
139 /* Locate the TLB entry for this vaddr + ASID */
140 idx = tlb_entry_lkup(vaddr_n_asid);
141
142 /* No error means entry found, zero it out */
143 if (likely(!(idx & TLB_LKUP_ERR))) {
144 __tlb_entry_erase();
145 } else {
146 /* Duplicate entry error */
147 WARN(idx == TLB_DUP_ERR, "Probe returned Dup PD for %x\n",
148 vaddr_n_asid);
149 }
150}
151
152/****************************************************************************
153 * ARC700 MMU caches recently used J-TLB entries (RAM) as uTLBs (FLOPs)
154 *
155 * New IVUTLB cmd in MMU v2 explictly invalidates the uTLB
156 *
157 * utlb_invalidate ( )
158 * -For v2 MMU calls Flush uTLB Cmd
159 * -For v1 MMU does nothing (except for Metal Fix v1 MMU)
160 * This is because in v1 TLBWrite itself invalidate uTLBs
161 ***************************************************************************/
162
163static void utlb_invalidate(void)
164{
165#if (CONFIG_ARC_MMU_VER >= 2)
166
167#if (CONFIG_ARC_MMU_VER == 2)
168 /* MMU v2 introduced the uTLB Flush command.
169 * There was however an obscure hardware bug, where uTLB flush would
170 * fail when a prior probe for J-TLB (both totally unrelated) would
171 * return lkup err - because the entry didn't exist in MMU.
172 * The Workround was to set Index reg with some valid value, prior to
173 * flush. This was fixed in MMU v3 hence not needed any more
174 */
175 unsigned int idx;
176
177 /* make sure INDEX Reg is valid */
178 idx = read_aux_reg(ARC_REG_TLBINDEX);
179
180 /* If not write some dummy val */
181 if (unlikely(idx & TLB_LKUP_ERR))
182 write_aux_reg(ARC_REG_TLBINDEX, 0xa);
183#endif
184
185 write_aux_reg(ARC_REG_TLBCOMMAND, TLBIVUTLB);
186#endif
187
188}
189
190static void tlb_entry_insert(unsigned int pd0, pte_t pd1)
191{
192 unsigned int idx;
193
194 /*
195 * First verify if entry for this vaddr+ASID already exists
196 * This also sets up PD0 (vaddr, ASID..) for final commit
197 */
198 idx = tlb_entry_lkup(pd0);
199
200 /*
201 * If Not already present get a free slot from MMU.
202 * Otherwise, Probe would have located the entry and set INDEX Reg
203 * with existing location. This will cause Write CMD to over-write
204 * existing entry with new PD0 and PD1
205 */
206 if (likely(idx & TLB_LKUP_ERR))
207 write_aux_reg(ARC_REG_TLBCOMMAND, TLBGetIndex);
208
209 /* setup the other half of TLB entry (pfn, rwx..) */
210 write_aux_reg(ARC_REG_TLBPD1, pd1);
211
212 /*
213 * Commit the Entry to MMU
214 * It doesn't sound safe to use the TLBWriteNI cmd here
215 * which doesn't flush uTLBs. I'd rather be safe than sorry.
216 */
217 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
218}
219
220#else /* CONFIG_ARC_MMU_VER >= 4) */
221
222static void utlb_invalidate(void)
223{
224 /* No need since uTLB is always in sync with JTLB */
225}
226
227static void tlb_entry_erase(unsigned int vaddr_n_asid)
228{
229 write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid | _PAGE_PRESENT);
230 write_aux_reg(ARC_REG_TLBCOMMAND, TLBDeleteEntry);
231}
232
233static void tlb_entry_insert(unsigned int pd0, pte_t pd1)
234{
235 write_aux_reg(ARC_REG_TLBPD0, pd0);
236 write_aux_reg(ARC_REG_TLBPD1, pd1);
237
238 if (is_pae40_enabled())
239 write_aux_reg(ARC_REG_TLBPD1HI, (u64)pd1 >> 32);
240
241 write_aux_reg(ARC_REG_TLBCOMMAND, TLBInsertEntry);
242}
243
244#endif
245
246/*
247 * Un-conditionally (without lookup) erase the entire MMU contents
248 */
249
250noinline void local_flush_tlb_all(void)
251{
252 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
253 unsigned long flags;
254 unsigned int entry;
255 int num_tlb = mmu->sets * mmu->ways;
256
257 local_irq_save(flags);
258
259 /* Load PD0 and PD1 with template for a Blank Entry */
260 write_aux_reg(ARC_REG_TLBPD1, 0);
261
262 if (is_pae40_enabled())
263 write_aux_reg(ARC_REG_TLBPD1HI, 0);
264
265 write_aux_reg(ARC_REG_TLBPD0, 0);
266
267 for (entry = 0; entry < num_tlb; entry++) {
268 /* write this entry to the TLB */
269 write_aux_reg(ARC_REG_TLBINDEX, entry);
270 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
271 }
272
273 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
274 const int stlb_idx = 0x800;
275
276 /* Blank sTLB entry */
277 write_aux_reg(ARC_REG_TLBPD0, _PAGE_HW_SZ);
278
279 for (entry = stlb_idx; entry < stlb_idx + 16; entry++) {
280 write_aux_reg(ARC_REG_TLBINDEX, entry);
281 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
282 }
283 }
284
285 utlb_invalidate();
286
287 local_irq_restore(flags);
288}
289
290/*
291 * Flush the entrie MM for userland. The fastest way is to move to Next ASID
292 */
293noinline void local_flush_tlb_mm(struct mm_struct *mm)
294{
295 /*
296 * Small optimisation courtesy IA64
297 * flush_mm called during fork,exit,munmap etc, multiple times as well.
298 * Only for fork( ) do we need to move parent to a new MMU ctxt,
299 * all other cases are NOPs, hence this check.
300 */
301 if (atomic_read(&mm->mm_users) == 0)
302 return;
303
304 /*
305 * - Move to a new ASID, but only if the mm is still wired in
306 * (Android Binder ended up calling this for vma->mm != tsk->mm,
307 * causing h/w - s/w ASID to get out of sync)
308 * - Also get_new_mmu_context() new implementation allocates a new
309 * ASID only if it is not allocated already - so unallocate first
310 */
311 destroy_context(mm);
312 if (current->mm == mm)
313 get_new_mmu_context(mm);
314}
315
316/*
317 * Flush a Range of TLB entries for userland.
318 * @start is inclusive, while @end is exclusive
319 * Difference between this and Kernel Range Flush is
320 * -Here the fastest way (if range is too large) is to move to next ASID
321 * without doing any explicit Shootdown
322 * -In case of kernel Flush, entry has to be shot down explictly
323 */
324void local_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
325 unsigned long end)
326{
327 const unsigned int cpu = smp_processor_id();
328 unsigned long flags;
329
330 /* If range @start to @end is more than 32 TLB entries deep,
331 * its better to move to a new ASID rather than searching for
332 * individual entries and then shooting them down
333 *
334 * The calc above is rough, doesn't account for unaligned parts,
335 * since this is heuristics based anyways
336 */
337 if (unlikely((end - start) >= PAGE_SIZE * 32)) {
338 local_flush_tlb_mm(vma->vm_mm);
339 return;
340 }
341
342 /*
343 * @start moved to page start: this alone suffices for checking
344 * loop end condition below, w/o need for aligning @end to end
345 * e.g. 2000 to 4001 will anyhow loop twice
346 */
347 start &= PAGE_MASK;
348
349 local_irq_save(flags);
350
351 if (asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID) {
352 while (start < end) {
353 tlb_entry_erase(start | hw_pid(vma->vm_mm, cpu));
354 start += PAGE_SIZE;
355 }
356 }
357
358 utlb_invalidate();
359
360 local_irq_restore(flags);
361}
362
363/* Flush the kernel TLB entries - vmalloc/modules (Global from MMU perspective)
364 * @start, @end interpreted as kvaddr
365 * Interestingly, shared TLB entries can also be flushed using just
366 * @start,@end alone (interpreted as user vaddr), although technically SASID
367 * is also needed. However our smart TLbProbe lookup takes care of that.
368 */
369void local_flush_tlb_kernel_range(unsigned long start, unsigned long end)
370{
371 unsigned long flags;
372
373 /* exactly same as above, except for TLB entry not taking ASID */
374
375 if (unlikely((end - start) >= PAGE_SIZE * 32)) {
376 local_flush_tlb_all();
377 return;
378 }
379
380 start &= PAGE_MASK;
381
382 local_irq_save(flags);
383 while (start < end) {
384 tlb_entry_erase(start);
385 start += PAGE_SIZE;
386 }
387
388 utlb_invalidate();
389
390 local_irq_restore(flags);
391}
392
393/*
394 * Delete TLB entry in MMU for a given page (??? address)
395 * NOTE One TLB entry contains translation for single PAGE
396 */
397
398void local_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
399{
400 const unsigned int cpu = smp_processor_id();
401 unsigned long flags;
402
403 /* Note that it is critical that interrupts are DISABLED between
404 * checking the ASID and using it flush the TLB entry
405 */
406 local_irq_save(flags);
407
408 if (asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID) {
409 tlb_entry_erase((page & PAGE_MASK) | hw_pid(vma->vm_mm, cpu));
410 utlb_invalidate();
411 }
412
413 local_irq_restore(flags);
414}
415
416#ifdef CONFIG_SMP
417
418struct tlb_args {
419 struct vm_area_struct *ta_vma;
420 unsigned long ta_start;
421 unsigned long ta_end;
422};
423
424static inline void ipi_flush_tlb_page(void *arg)
425{
426 struct tlb_args *ta = arg;
427
428 local_flush_tlb_page(ta->ta_vma, ta->ta_start);
429}
430
431static inline void ipi_flush_tlb_range(void *arg)
432{
433 struct tlb_args *ta = arg;
434
435 local_flush_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
436}
437
438#ifdef CONFIG_TRANSPARENT_HUGEPAGE
439static inline void ipi_flush_pmd_tlb_range(void *arg)
440{
441 struct tlb_args *ta = arg;
442
443 local_flush_pmd_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
444}
445#endif
446
447static inline void ipi_flush_tlb_kernel_range(void *arg)
448{
449 struct tlb_args *ta = (struct tlb_args *)arg;
450
451 local_flush_tlb_kernel_range(ta->ta_start, ta->ta_end);
452}
453
454void flush_tlb_all(void)
455{
456 on_each_cpu((smp_call_func_t)local_flush_tlb_all, NULL, 1);
457}
458
459void flush_tlb_mm(struct mm_struct *mm)
460{
461 on_each_cpu_mask(mm_cpumask(mm), (smp_call_func_t)local_flush_tlb_mm,
462 mm, 1);
463}
464
465void flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr)
466{
467 struct tlb_args ta = {
468 .ta_vma = vma,
469 .ta_start = uaddr
470 };
471
472 on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_page, &ta, 1);
473}
474
475void flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
476 unsigned long end)
477{
478 struct tlb_args ta = {
479 .ta_vma = vma,
480 .ta_start = start,
481 .ta_end = end
482 };
483
484 on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_range, &ta, 1);
485}
486
487#ifdef CONFIG_TRANSPARENT_HUGEPAGE
488void flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
489 unsigned long end)
490{
491 struct tlb_args ta = {
492 .ta_vma = vma,
493 .ta_start = start,
494 .ta_end = end
495 };
496
497 on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_pmd_tlb_range, &ta, 1);
498}
499#endif
500
501void flush_tlb_kernel_range(unsigned long start, unsigned long end)
502{
503 struct tlb_args ta = {
504 .ta_start = start,
505 .ta_end = end
506 };
507
508 on_each_cpu(ipi_flush_tlb_kernel_range, &ta, 1);
509}
510#endif
511
512/*
513 * Routine to create a TLB entry
514 */
515void create_tlb(struct vm_area_struct *vma, unsigned long vaddr, pte_t *ptep)
516{
517 unsigned long flags;
518 unsigned int asid_or_sasid, rwx;
519 unsigned long pd0;
520 pte_t pd1;
521
522 /*
523 * create_tlb() assumes that current->mm == vma->mm, since
524 * -it ASID for TLB entry is fetched from MMU ASID reg (valid for curr)
525 * -completes the lazy write to SASID reg (again valid for curr tsk)
526 *
527 * Removing the assumption involves
528 * -Using vma->mm->context{ASID,SASID}, as opposed to MMU reg.
529 * -Fix the TLB paranoid debug code to not trigger false negatives.
530 * -More importantly it makes this handler inconsistent with fast-path
531 * TLB Refill handler which always deals with "current"
532 *
533 * Lets see the use cases when current->mm != vma->mm and we land here
534 * 1. execve->copy_strings()->__get_user_pages->handle_mm_fault
535 * Here VM wants to pre-install a TLB entry for user stack while
536 * current->mm still points to pre-execve mm (hence the condition).
537 * However the stack vaddr is soon relocated (randomization) and
538 * move_page_tables() tries to undo that TLB entry.
539 * Thus not creating TLB entry is not any worse.
540 *
541 * 2. ptrace(POKETEXT) causes a CoW - debugger(current) inserting a
542 * breakpoint in debugged task. Not creating a TLB now is not
543 * performance critical.
544 *
545 * Both the cases above are not good enough for code churn.
546 */
547 if (current->active_mm != vma->vm_mm)
548 return;
549
550 local_irq_save(flags);
551
552 tlb_paranoid_check(asid_mm(vma->vm_mm, smp_processor_id()), vaddr);
553
554 vaddr &= PAGE_MASK;
555
556 /* update this PTE credentials */
557 pte_val(*ptep) |= (_PAGE_PRESENT | _PAGE_ACCESSED);
558
559 /* Create HW TLB(PD0,PD1) from PTE */
560
561 /* ASID for this task */
562 asid_or_sasid = read_aux_reg(ARC_REG_PID) & 0xff;
563
564 pd0 = vaddr | asid_or_sasid | (pte_val(*ptep) & PTE_BITS_IN_PD0);
565
566 /*
567 * ARC MMU provides fully orthogonal access bits for K/U mode,
568 * however Linux only saves 1 set to save PTE real-estate
569 * Here we convert 3 PTE bits into 6 MMU bits:
570 * -Kernel only entries have Kr Kw Kx 0 0 0
571 * -User entries have mirrored K and U bits
572 */
573 rwx = pte_val(*ptep) & PTE_BITS_RWX;
574
575 if (pte_val(*ptep) & _PAGE_GLOBAL)
576 rwx <<= 3; /* r w x => Kr Kw Kx 0 0 0 */
577 else
578 rwx |= (rwx << 3); /* r w x => Kr Kw Kx Ur Uw Ux */
579
580 pd1 = rwx | (pte_val(*ptep) & PTE_BITS_NON_RWX_IN_PD1);
581
582 tlb_entry_insert(pd0, pd1);
583
584 local_irq_restore(flags);
585}
586
587/*
588 * Called at the end of pagefault, for a userspace mapped page
589 * -pre-install the corresponding TLB entry into MMU
590 * -Finalize the delayed D-cache flush of kernel mapping of page due to
591 * flush_dcache_page(), copy_user_page()
592 *
593 * Note that flush (when done) involves both WBACK - so physical page is
594 * in sync as well as INV - so any non-congruent aliases don't remain
595 */
596void update_mmu_cache(struct vm_area_struct *vma, unsigned long vaddr_unaligned,
597 pte_t *ptep)
598{
599 unsigned long vaddr = vaddr_unaligned & PAGE_MASK;
600 phys_addr_t paddr = pte_val(*ptep) & PAGE_MASK;
601 struct page *page = pfn_to_page(pte_pfn(*ptep));
602
603 create_tlb(vma, vaddr, ptep);
604
605 if (page == ZERO_PAGE(0)) {
606 return;
607 }
608
609 /*
610 * Exec page : Independent of aliasing/page-color considerations,
611 * since icache doesn't snoop dcache on ARC, any dirty
612 * K-mapping of a code page needs to be wback+inv so that
613 * icache fetch by userspace sees code correctly.
614 * !EXEC page: If K-mapping is NOT congruent to U-mapping, flush it
615 * so userspace sees the right data.
616 * (Avoids the flush for Non-exec + congruent mapping case)
617 */
618 if ((vma->vm_flags & VM_EXEC) ||
619 addr_not_cache_congruent(paddr, vaddr)) {
620
621 int dirty = !test_and_set_bit(PG_dc_clean, &page->flags);
622 if (dirty) {
623 /* wback + inv dcache lines (K-mapping) */
624 __flush_dcache_page(paddr, paddr);
625
626 /* invalidate any existing icache lines (U-mapping) */
627 if (vma->vm_flags & VM_EXEC)
628 __inv_icache_page(paddr, vaddr);
629 }
630 }
631}
632
633#ifdef CONFIG_TRANSPARENT_HUGEPAGE
634
635/*
636 * MMUv4 in HS38x cores supports Super Pages which are basis for Linux THP
637 * support.
638 *
639 * Normal and Super pages can co-exist (ofcourse not overlap) in TLB with a
640 * new bit "SZ" in TLB page descriptor to distinguish between them.
641 * Super Page size is configurable in hardware (4K to 16M), but fixed once
642 * RTL builds.
643 *
644 * The exact THP size a Linx configuration will support is a function of:
645 * - MMU page size (typical 8K, RTL fixed)
646 * - software page walker address split between PGD:PTE:PFN (typical
647 * 11:8:13, but can be changed with 1 line)
648 * So for above default, THP size supported is 8K * (2^8) = 2M
649 *
650 * Default Page Walker is 2 levels, PGD:PTE:PFN, which in THP regime
651 * reduces to 1 level (as PTE is folded into PGD and canonically referred
652 * to as PMD).
653 * Thus THP PMD accessors are implemented in terms of PTE (just like sparc)
654 */
655
656void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
657 pmd_t *pmd)
658{
659 pte_t pte = __pte(pmd_val(*pmd));
660 update_mmu_cache(vma, addr, &pte);
661}
662
663void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
664 pgtable_t pgtable)
665{
666 struct list_head *lh = (struct list_head *) pgtable;
667
668 assert_spin_locked(&mm->page_table_lock);
669
670 /* FIFO */
671 if (!pmd_huge_pte(mm, pmdp))
672 INIT_LIST_HEAD(lh);
673 else
674 list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
675 pmd_huge_pte(mm, pmdp) = pgtable;
676}
677
678pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
679{
680 struct list_head *lh;
681 pgtable_t pgtable;
682
683 assert_spin_locked(&mm->page_table_lock);
684
685 pgtable = pmd_huge_pte(mm, pmdp);
686 lh = (struct list_head *) pgtable;
687 if (list_empty(lh))
688 pmd_huge_pte(mm, pmdp) = NULL;
689 else {
690 pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
691 list_del(lh);
692 }
693
694 pte_val(pgtable[0]) = 0;
695 pte_val(pgtable[1]) = 0;
696
697 return pgtable;
698}
699
700void local_flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
701 unsigned long end)
702{
703 unsigned int cpu;
704 unsigned long flags;
705
706 local_irq_save(flags);
707
708 cpu = smp_processor_id();
709
710 if (likely(asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID)) {
711 unsigned int asid = hw_pid(vma->vm_mm, cpu);
712
713 /* No need to loop here: this will always be for 1 Huge Page */
714 tlb_entry_erase(start | _PAGE_HW_SZ | asid);
715 }
716
717 local_irq_restore(flags);
718}
719
720#endif
721
722/* Read the Cache Build Confuration Registers, Decode them and save into
723 * the cpuinfo structure for later use.
724 * No Validation is done here, simply read/convert the BCRs
725 */
726void read_decode_mmu_bcr(void)
727{
728 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
729 unsigned int tmp;
730 struct bcr_mmu_1_2 {
731#ifdef CONFIG_CPU_BIG_ENDIAN
732 unsigned int ver:8, ways:4, sets:4, u_itlb:8, u_dtlb:8;
733#else
734 unsigned int u_dtlb:8, u_itlb:8, sets:4, ways:4, ver:8;
735#endif
736 } *mmu2;
737
738 struct bcr_mmu_3 {
739#ifdef CONFIG_CPU_BIG_ENDIAN
740 unsigned int ver:8, ways:4, sets:4, res:3, sasid:1, pg_sz:4,
741 u_itlb:4, u_dtlb:4;
742#else
743 unsigned int u_dtlb:4, u_itlb:4, pg_sz:4, sasid:1, res:3, sets:4,
744 ways:4, ver:8;
745#endif
746 } *mmu3;
747
748 struct bcr_mmu_4 {
749#ifdef CONFIG_CPU_BIG_ENDIAN
750 unsigned int ver:8, sasid:1, sz1:4, sz0:4, res:2, pae:1,
751 n_ways:2, n_entry:2, n_super:2, u_itlb:3, u_dtlb:3;
752#else
753 /* DTLB ITLB JES JE JA */
754 unsigned int u_dtlb:3, u_itlb:3, n_super:2, n_entry:2, n_ways:2,
755 pae:1, res:2, sz0:4, sz1:4, sasid:1, ver:8;
756#endif
757 } *mmu4;
758
759 tmp = read_aux_reg(ARC_REG_MMU_BCR);
760 mmu->ver = (tmp >> 24);
761
762 if (is_isa_arcompact()) {
763 if (mmu->ver <= 2) {
764 mmu2 = (struct bcr_mmu_1_2 *)&tmp;
765 mmu->pg_sz_k = TO_KB(0x2000);
766 mmu->sets = 1 << mmu2->sets;
767 mmu->ways = 1 << mmu2->ways;
768 mmu->u_dtlb = mmu2->u_dtlb;
769 mmu->u_itlb = mmu2->u_itlb;
770 } else {
771 mmu3 = (struct bcr_mmu_3 *)&tmp;
772 mmu->pg_sz_k = 1 << (mmu3->pg_sz - 1);
773 mmu->sets = 1 << mmu3->sets;
774 mmu->ways = 1 << mmu3->ways;
775 mmu->u_dtlb = mmu3->u_dtlb;
776 mmu->u_itlb = mmu3->u_itlb;
777 mmu->sasid = mmu3->sasid;
778 }
779 } else {
780 mmu4 = (struct bcr_mmu_4 *)&tmp;
781 mmu->pg_sz_k = 1 << (mmu4->sz0 - 1);
782 mmu->s_pg_sz_m = 1 << (mmu4->sz1 - 11);
783 mmu->sets = 64 << mmu4->n_entry;
784 mmu->ways = mmu4->n_ways * 2;
785 mmu->u_dtlb = mmu4->u_dtlb * 4;
786 mmu->u_itlb = mmu4->u_itlb * 4;
787 mmu->sasid = mmu4->sasid;
788 pae_exists = mmu->pae = mmu4->pae;
789 }
790}
791
792char *arc_mmu_mumbojumbo(int cpu_id, char *buf, int len)
793{
794 int n = 0;
795 struct cpuinfo_arc_mmu *p_mmu = &cpuinfo_arc700[cpu_id].mmu;
796 char super_pg[64] = "";
797
798 if (p_mmu->s_pg_sz_m)
799 scnprintf(super_pg, 64, "%dM Super Page %s",
800 p_mmu->s_pg_sz_m,
801 IS_USED_CFG(CONFIG_TRANSPARENT_HUGEPAGE));
802
803 n += scnprintf(buf + n, len - n,
804 "MMU [v%x]\t: %dk PAGE, %sJTLB %d (%dx%d), uDTLB %d, uITLB %d%s%s\n",
805 p_mmu->ver, p_mmu->pg_sz_k, super_pg,
806 p_mmu->sets * p_mmu->ways, p_mmu->sets, p_mmu->ways,
807 p_mmu->u_dtlb, p_mmu->u_itlb,
808 IS_AVAIL2(p_mmu->pae, ", PAE40 ", CONFIG_ARC_HAS_PAE40));
809
810 return buf;
811}
812
813int pae40_exist_but_not_enab(void)
814{
815 return pae_exists && !is_pae40_enabled();
816}
817
818void arc_mmu_init(void)
819{
820 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
821 char str[256];
822 int compat = 0;
823
824 pr_info("%s", arc_mmu_mumbojumbo(0, str, sizeof(str)));
825
826 /*
827 * Can't be done in processor.h due to header include depenedencies
828 */
829 BUILD_BUG_ON(!IS_ALIGNED((CONFIG_ARC_KVADDR_SIZE << 20), PMD_SIZE));
830
831 /*
832 * stack top size sanity check,
833 * Can't be done in processor.h due to header include depenedencies
834 */
835 BUILD_BUG_ON(!IS_ALIGNED(STACK_TOP, PMD_SIZE));
836
837 /*
838 * Ensure that MMU features assumed by kernel exist in hardware.
839 * For older ARC700 cpus, it has to be exact match, since the MMU
840 * revisions were not backwards compatible (MMUv3 TLB layout changed
841 * so even if kernel for v2 didn't use any new cmds of v3, it would
842 * still not work.
843 * For HS cpus, MMUv4 was baseline and v5 is backwards compatible
844 * (will run older software).
845 */
846 if (is_isa_arcompact() && mmu->ver == CONFIG_ARC_MMU_VER)
847 compat = 1;
848 else if (is_isa_arcv2() && mmu->ver >= CONFIG_ARC_MMU_VER)
849 compat = 1;
850
851 if (!compat) {
852 panic("MMU ver %d doesn't match kernel built for %d...\n",
853 mmu->ver, CONFIG_ARC_MMU_VER);
854 }
855
856 if (mmu->pg_sz_k != TO_KB(PAGE_SIZE))
857 panic("MMU pg size != PAGE_SIZE (%luk)\n", TO_KB(PAGE_SIZE));
858
859 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
860 mmu->s_pg_sz_m != TO_MB(HPAGE_PMD_SIZE))
861 panic("MMU Super pg size != Linux HPAGE_PMD_SIZE (%luM)\n",
862 (unsigned long)TO_MB(HPAGE_PMD_SIZE));
863
864 if (IS_ENABLED(CONFIG_ARC_HAS_PAE40) && !mmu->pae)
865 panic("Hardware doesn't support PAE40\n");
866
867 /* Enable the MMU */
868 write_aux_reg(ARC_REG_PID, MMU_ENABLE);
869
870 /* In smp we use this reg for interrupt 1 scratch */
871#ifndef CONFIG_SMP
872 /* swapper_pg_dir is the pgd for the kernel, used by vmalloc */
873 write_aux_reg(ARC_REG_SCRATCH_DATA0, swapper_pg_dir);
874#endif
875
876 if (pae40_exist_but_not_enab())
877 write_aux_reg(ARC_REG_TLBPD1HI, 0);
878}
879
880/*
881 * TLB Programmer's Model uses Linear Indexes: 0 to {255, 511} for 128 x {2,4}
882 * The mapping is Column-first.
883 * --------------------- -----------
884 * |way0|way1|way2|way3| |way0|way1|
885 * --------------------- -----------
886 * [set0] | 0 | 1 | 2 | 3 | | 0 | 1 |
887 * [set1] | 4 | 5 | 6 | 7 | | 2 | 3 |
888 * ~ ~ ~ ~
889 * [set127] | 508| 509| 510| 511| | 254| 255|
890 * --------------------- -----------
891 * For normal operations we don't(must not) care how above works since
892 * MMU cmd getIndex(vaddr) abstracts that out.
893 * However for walking WAYS of a SET, we need to know this
894 */
895#define SET_WAY_TO_IDX(mmu, set, way) ((set) * mmu->ways + (way))
896
897/* Handling of Duplicate PD (TLB entry) in MMU.
898 * -Could be due to buggy customer tapeouts or obscure kernel bugs
899 * -MMU complaints not at the time of duplicate PD installation, but at the
900 * time of lookup matching multiple ways.
901 * -Ideally these should never happen - but if they do - workaround by deleting
902 * the duplicate one.
903 * -Knob to be verbose abt it.(TODO: hook them up to debugfs)
904 */
905volatile int dup_pd_silent; /* Be slient abt it or complain (default) */
906
907void do_tlb_overlap_fault(unsigned long cause, unsigned long address,
908 struct pt_regs *regs)
909{
910 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
911 unsigned long flags;
912 int set, n_ways = mmu->ways;
913
914 n_ways = min(n_ways, 4);
915 BUG_ON(mmu->ways > 4);
916
917 local_irq_save(flags);
918
919 /* loop thru all sets of TLB */
920 for (set = 0; set < mmu->sets; set++) {
921
922 int is_valid, way;
923 unsigned int pd0[4];
924
925 /* read out all the ways of current set */
926 for (way = 0, is_valid = 0; way < n_ways; way++) {
927 write_aux_reg(ARC_REG_TLBINDEX,
928 SET_WAY_TO_IDX(mmu, set, way));
929 write_aux_reg(ARC_REG_TLBCOMMAND, TLBRead);
930 pd0[way] = read_aux_reg(ARC_REG_TLBPD0);
931 is_valid |= pd0[way] & _PAGE_PRESENT;
932 pd0[way] &= PAGE_MASK;
933 }
934
935 /* If all the WAYS in SET are empty, skip to next SET */
936 if (!is_valid)
937 continue;
938
939 /* Scan the set for duplicate ways: needs a nested loop */
940 for (way = 0; way < n_ways - 1; way++) {
941
942 int n;
943
944 if (!pd0[way])
945 continue;
946
947 for (n = way + 1; n < n_ways; n++) {
948 if (pd0[way] != pd0[n])
949 continue;
950
951 if (!dup_pd_silent)
952 pr_info("Dup TLB PD0 %08x @ set %d ways %d,%d\n",
953 pd0[way], set, way, n);
954
955 /*
956 * clear entry @way and not @n.
957 * This is critical to our optimised loop
958 */
959 pd0[way] = 0;
960 write_aux_reg(ARC_REG_TLBINDEX,
961 SET_WAY_TO_IDX(mmu, set, way));
962 __tlb_entry_erase();
963 }
964 }
965 }
966
967 local_irq_restore(flags);
968}
969
970/***********************************************************************
971 * Diagnostic Routines
972 * -Called from Low Level TLB Hanlders if things don;t look good
973 **********************************************************************/
974
975#ifdef CONFIG_ARC_DBG_TLB_PARANOIA
976
977/*
978 * Low Level ASM TLB handler calls this if it finds that HW and SW ASIDS
979 * don't match
980 */
981void print_asid_mismatch(int mm_asid, int mmu_asid, int is_fast_path)
982{
983 pr_emerg("ASID Mismatch in %s Path Handler: sw-pid=0x%x hw-pid=0x%x\n",
984 is_fast_path ? "Fast" : "Slow", mm_asid, mmu_asid);
985
986 __asm__ __volatile__("flag 1");
987}
988
989void tlb_paranoid_check(unsigned int mm_asid, unsigned long addr)
990{
991 unsigned int mmu_asid;
992
993 mmu_asid = read_aux_reg(ARC_REG_PID) & 0xff;
994
995 /*
996 * At the time of a TLB miss/installation
997 * - HW version needs to match SW version
998 * - SW needs to have a valid ASID
999 */
1000 if (addr < 0x70000000 &&
1001 ((mm_asid == MM_CTXT_NO_ASID) ||
1002 (mmu_asid != (mm_asid & MM_CTXT_ASID_MASK))))
1003 print_asid_mismatch(mm_asid, mmu_asid, 0);
1004}
1005#endif
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * TLB Management (flush/create/diagnostics) for ARC700
4 *
5 * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
6 *
7 * vineetg: Aug 2011
8 * -Reintroduce duplicate PD fixup - some customer chips still have the issue
9 *
10 * vineetg: May 2011
11 * -No need to flush_cache_page( ) for each call to update_mmu_cache()
12 * some of the LMBench tests improved amazingly
13 * = page-fault thrice as fast (75 usec to 28 usec)
14 * = mmap twice as fast (9.6 msec to 4.6 msec),
15 * = fork (5.3 msec to 3.7 msec)
16 *
17 * vineetg: April 2011 :
18 * -MMU v3: PD{0,1} bits layout changed: They don't overlap anymore,
19 * helps avoid a shift when preparing PD0 from PTE
20 *
21 * vineetg: April 2011 : Preparing for MMU V3
22 * -MMU v2/v3 BCRs decoded differently
23 * -Remove TLB_SIZE hardcoding as it's variable now: 256 or 512
24 * -tlb_entry_erase( ) can be void
25 * -local_flush_tlb_range( ):
26 * = need not "ceil" @end
27 * = walks MMU only if range spans < 32 entries, as opposed to 256
28 *
29 * Vineetg: Sept 10th 2008
30 * -Changes related to MMU v2 (Rel 4.8)
31 *
32 * Vineetg: Aug 29th 2008
33 * -In TLB Flush operations (Metal Fix MMU) there is a explicit command to
34 * flush Micro-TLBS. If TLB Index Reg is invalid prior to TLBIVUTLB cmd,
35 * it fails. Thus need to load it with ANY valid value before invoking
36 * TLBIVUTLB cmd
37 *
38 * Vineetg: Aug 21th 2008:
39 * -Reduced the duration of IRQ lockouts in TLB Flush routines
40 * -Multiple copies of TLB erase code separated into a "single" function
41 * -In TLB Flush routines, interrupt disabling moved UP to retrieve ASID
42 * in interrupt-safe region.
43 *
44 * Vineetg: April 23rd Bug #93131
45 * Problem: tlb_flush_kernel_range() doesn't do anything if the range to
46 * flush is more than the size of TLB itself.
47 *
48 * Rahul Trivedi : Codito Technologies 2004
49 */
50
51#include <linux/module.h>
52#include <linux/bug.h>
53#include <linux/mm_types.h>
54
55#include <asm/arcregs.h>
56#include <asm/setup.h>
57#include <asm/mmu_context.h>
58#include <asm/mmu.h>
59
60/* Need for ARC MMU v2
61 *
62 * ARC700 MMU-v1 had a Joint-TLB for Code and Data and is 2 way set-assoc.
63 * For a memcpy operation with 3 players (src/dst/code) such that all 3 pages
64 * map into same set, there would be contention for the 2 ways causing severe
65 * Thrashing.
66 *
67 * Although J-TLB is 2 way set assoc, ARC700 caches J-TLB into uTLBS which has
68 * much higher associativity. u-D-TLB is 8 ways, u-I-TLB is 4 ways.
69 * Given this, the thrashing problem should never happen because once the 3
70 * J-TLB entries are created (even though 3rd will knock out one of the prev
71 * two), the u-D-TLB and u-I-TLB will have what is required to accomplish memcpy
72 *
73 * Yet we still see the Thrashing because a J-TLB Write cause flush of u-TLBs.
74 * This is a simple design for keeping them in sync. So what do we do?
75 * The solution which James came up was pretty neat. It utilised the assoc
76 * of uTLBs by not invalidating always but only when absolutely necessary.
77 *
78 * - Existing TLB commands work as before
79 * - New command (TLBWriteNI) for TLB write without clearing uTLBs
80 * - New command (TLBIVUTLB) to invalidate uTLBs.
81 *
82 * The uTLBs need only be invalidated when pages are being removed from the
83 * OS page table. If a 'victim' TLB entry is being overwritten in the main TLB
84 * as a result of a miss, the removed entry is still allowed to exist in the
85 * uTLBs as it is still valid and present in the OS page table. This allows the
86 * full associativity of the uTLBs to hide the limited associativity of the main
87 * TLB.
88 *
89 * During a miss handler, the new "TLBWriteNI" command is used to load
90 * entries without clearing the uTLBs.
91 *
92 * When the OS page table is updated, TLB entries that may be associated with a
93 * removed page are removed (flushed) from the TLB using TLBWrite. In this
94 * circumstance, the uTLBs must also be cleared. This is done by using the
95 * existing TLBWrite command. An explicit IVUTLB is also required for those
96 * corner cases when TLBWrite was not executed at all because the corresp
97 * J-TLB entry got evicted/replaced.
98 */
99
100
101/* A copy of the ASID from the PID reg is kept in asid_cache */
102DEFINE_PER_CPU(unsigned int, asid_cache) = MM_CTXT_FIRST_CYCLE;
103
104static int __read_mostly pae_exists;
105
106/*
107 * Utility Routine to erase a J-TLB entry
108 * Caller needs to setup Index Reg (manually or via getIndex)
109 */
110static inline void __tlb_entry_erase(void)
111{
112 write_aux_reg(ARC_REG_TLBPD1, 0);
113
114 if (is_pae40_enabled())
115 write_aux_reg(ARC_REG_TLBPD1HI, 0);
116
117 write_aux_reg(ARC_REG_TLBPD0, 0);
118 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
119}
120
121static void utlb_invalidate(void)
122{
123#if (CONFIG_ARC_MMU_VER >= 2)
124
125#if (CONFIG_ARC_MMU_VER == 2)
126 /* MMU v2 introduced the uTLB Flush command.
127 * There was however an obscure hardware bug, where uTLB flush would
128 * fail when a prior probe for J-TLB (both totally unrelated) would
129 * return lkup err - because the entry didn't exist in MMU.
130 * The Workaround was to set Index reg with some valid value, prior to
131 * flush. This was fixed in MMU v3
132 */
133 unsigned int idx;
134
135 /* make sure INDEX Reg is valid */
136 idx = read_aux_reg(ARC_REG_TLBINDEX);
137
138 /* If not write some dummy val */
139 if (unlikely(idx & TLB_LKUP_ERR))
140 write_aux_reg(ARC_REG_TLBINDEX, 0xa);
141#endif
142
143 write_aux_reg(ARC_REG_TLBCOMMAND, TLBIVUTLB);
144#endif
145
146}
147
148#if (CONFIG_ARC_MMU_VER < 4)
149
150static inline unsigned int tlb_entry_lkup(unsigned long vaddr_n_asid)
151{
152 unsigned int idx;
153
154 write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid);
155
156 write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe);
157 idx = read_aux_reg(ARC_REG_TLBINDEX);
158
159 return idx;
160}
161
162static void tlb_entry_erase(unsigned int vaddr_n_asid)
163{
164 unsigned int idx;
165
166 /* Locate the TLB entry for this vaddr + ASID */
167 idx = tlb_entry_lkup(vaddr_n_asid);
168
169 /* No error means entry found, zero it out */
170 if (likely(!(idx & TLB_LKUP_ERR))) {
171 __tlb_entry_erase();
172 } else {
173 /* Duplicate entry error */
174 WARN(idx == TLB_DUP_ERR, "Probe returned Dup PD for %x\n",
175 vaddr_n_asid);
176 }
177}
178
179static void tlb_entry_insert(unsigned int pd0, pte_t pd1)
180{
181 unsigned int idx;
182
183 /*
184 * First verify if entry for this vaddr+ASID already exists
185 * This also sets up PD0 (vaddr, ASID..) for final commit
186 */
187 idx = tlb_entry_lkup(pd0);
188
189 /*
190 * If Not already present get a free slot from MMU.
191 * Otherwise, Probe would have located the entry and set INDEX Reg
192 * with existing location. This will cause Write CMD to over-write
193 * existing entry with new PD0 and PD1
194 */
195 if (likely(idx & TLB_LKUP_ERR))
196 write_aux_reg(ARC_REG_TLBCOMMAND, TLBGetIndex);
197
198 /* setup the other half of TLB entry (pfn, rwx..) */
199 write_aux_reg(ARC_REG_TLBPD1, pd1);
200
201 /*
202 * Commit the Entry to MMU
203 * It doesn't sound safe to use the TLBWriteNI cmd here
204 * which doesn't flush uTLBs. I'd rather be safe than sorry.
205 */
206 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
207}
208
209#else /* CONFIG_ARC_MMU_VER >= 4) */
210
211static void tlb_entry_erase(unsigned int vaddr_n_asid)
212{
213 write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid | _PAGE_PRESENT);
214 write_aux_reg(ARC_REG_TLBCOMMAND, TLBDeleteEntry);
215}
216
217static void tlb_entry_insert(unsigned int pd0, pte_t pd1)
218{
219 write_aux_reg(ARC_REG_TLBPD0, pd0);
220 write_aux_reg(ARC_REG_TLBPD1, pd1);
221
222 if (is_pae40_enabled())
223 write_aux_reg(ARC_REG_TLBPD1HI, (u64)pd1 >> 32);
224
225 write_aux_reg(ARC_REG_TLBCOMMAND, TLBInsertEntry);
226}
227
228#endif
229
230/*
231 * Un-conditionally (without lookup) erase the entire MMU contents
232 */
233
234noinline void local_flush_tlb_all(void)
235{
236 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
237 unsigned long flags;
238 unsigned int entry;
239 int num_tlb = mmu->sets * mmu->ways;
240
241 local_irq_save(flags);
242
243 /* Load PD0 and PD1 with template for a Blank Entry */
244 write_aux_reg(ARC_REG_TLBPD1, 0);
245
246 if (is_pae40_enabled())
247 write_aux_reg(ARC_REG_TLBPD1HI, 0);
248
249 write_aux_reg(ARC_REG_TLBPD0, 0);
250
251 for (entry = 0; entry < num_tlb; entry++) {
252 /* write this entry to the TLB */
253 write_aux_reg(ARC_REG_TLBINDEX, entry);
254 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWriteNI);
255 }
256
257 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
258 const int stlb_idx = 0x800;
259
260 /* Blank sTLB entry */
261 write_aux_reg(ARC_REG_TLBPD0, _PAGE_HW_SZ);
262
263 for (entry = stlb_idx; entry < stlb_idx + 16; entry++) {
264 write_aux_reg(ARC_REG_TLBINDEX, entry);
265 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWriteNI);
266 }
267 }
268
269 utlb_invalidate();
270
271 local_irq_restore(flags);
272}
273
274/*
275 * Flush the entire MM for userland. The fastest way is to move to Next ASID
276 */
277noinline void local_flush_tlb_mm(struct mm_struct *mm)
278{
279 /*
280 * Small optimisation courtesy IA64
281 * flush_mm called during fork,exit,munmap etc, multiple times as well.
282 * Only for fork( ) do we need to move parent to a new MMU ctxt,
283 * all other cases are NOPs, hence this check.
284 */
285 if (atomic_read(&mm->mm_users) == 0)
286 return;
287
288 /*
289 * - Move to a new ASID, but only if the mm is still wired in
290 * (Android Binder ended up calling this for vma->mm != tsk->mm,
291 * causing h/w - s/w ASID to get out of sync)
292 * - Also get_new_mmu_context() new implementation allocates a new
293 * ASID only if it is not allocated already - so unallocate first
294 */
295 destroy_context(mm);
296 if (current->mm == mm)
297 get_new_mmu_context(mm);
298}
299
300/*
301 * Flush a Range of TLB entries for userland.
302 * @start is inclusive, while @end is exclusive
303 * Difference between this and Kernel Range Flush is
304 * -Here the fastest way (if range is too large) is to move to next ASID
305 * without doing any explicit Shootdown
306 * -In case of kernel Flush, entry has to be shot down explicitly
307 */
308void local_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
309 unsigned long end)
310{
311 const unsigned int cpu = smp_processor_id();
312 unsigned long flags;
313
314 /* If range @start to @end is more than 32 TLB entries deep,
315 * its better to move to a new ASID rather than searching for
316 * individual entries and then shooting them down
317 *
318 * The calc above is rough, doesn't account for unaligned parts,
319 * since this is heuristics based anyways
320 */
321 if (unlikely((end - start) >= PAGE_SIZE * 32)) {
322 local_flush_tlb_mm(vma->vm_mm);
323 return;
324 }
325
326 /*
327 * @start moved to page start: this alone suffices for checking
328 * loop end condition below, w/o need for aligning @end to end
329 * e.g. 2000 to 4001 will anyhow loop twice
330 */
331 start &= PAGE_MASK;
332
333 local_irq_save(flags);
334
335 if (asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID) {
336 while (start < end) {
337 tlb_entry_erase(start | hw_pid(vma->vm_mm, cpu));
338 start += PAGE_SIZE;
339 }
340 }
341
342 local_irq_restore(flags);
343}
344
345/* Flush the kernel TLB entries - vmalloc/modules (Global from MMU perspective)
346 * @start, @end interpreted as kvaddr
347 * Interestingly, shared TLB entries can also be flushed using just
348 * @start,@end alone (interpreted as user vaddr), although technically SASID
349 * is also needed. However our smart TLbProbe lookup takes care of that.
350 */
351void local_flush_tlb_kernel_range(unsigned long start, unsigned long end)
352{
353 unsigned long flags;
354
355 /* exactly same as above, except for TLB entry not taking ASID */
356
357 if (unlikely((end - start) >= PAGE_SIZE * 32)) {
358 local_flush_tlb_all();
359 return;
360 }
361
362 start &= PAGE_MASK;
363
364 local_irq_save(flags);
365 while (start < end) {
366 tlb_entry_erase(start);
367 start += PAGE_SIZE;
368 }
369
370 local_irq_restore(flags);
371}
372
373/*
374 * Delete TLB entry in MMU for a given page (??? address)
375 * NOTE One TLB entry contains translation for single PAGE
376 */
377
378void local_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
379{
380 const unsigned int cpu = smp_processor_id();
381 unsigned long flags;
382
383 /* Note that it is critical that interrupts are DISABLED between
384 * checking the ASID and using it flush the TLB entry
385 */
386 local_irq_save(flags);
387
388 if (asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID) {
389 tlb_entry_erase((page & PAGE_MASK) | hw_pid(vma->vm_mm, cpu));
390 }
391
392 local_irq_restore(flags);
393}
394
395#ifdef CONFIG_SMP
396
397struct tlb_args {
398 struct vm_area_struct *ta_vma;
399 unsigned long ta_start;
400 unsigned long ta_end;
401};
402
403static inline void ipi_flush_tlb_page(void *arg)
404{
405 struct tlb_args *ta = arg;
406
407 local_flush_tlb_page(ta->ta_vma, ta->ta_start);
408}
409
410static inline void ipi_flush_tlb_range(void *arg)
411{
412 struct tlb_args *ta = arg;
413
414 local_flush_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
415}
416
417#ifdef CONFIG_TRANSPARENT_HUGEPAGE
418static inline void ipi_flush_pmd_tlb_range(void *arg)
419{
420 struct tlb_args *ta = arg;
421
422 local_flush_pmd_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
423}
424#endif
425
426static inline void ipi_flush_tlb_kernel_range(void *arg)
427{
428 struct tlb_args *ta = (struct tlb_args *)arg;
429
430 local_flush_tlb_kernel_range(ta->ta_start, ta->ta_end);
431}
432
433void flush_tlb_all(void)
434{
435 on_each_cpu((smp_call_func_t)local_flush_tlb_all, NULL, 1);
436}
437
438void flush_tlb_mm(struct mm_struct *mm)
439{
440 on_each_cpu_mask(mm_cpumask(mm), (smp_call_func_t)local_flush_tlb_mm,
441 mm, 1);
442}
443
444void flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr)
445{
446 struct tlb_args ta = {
447 .ta_vma = vma,
448 .ta_start = uaddr
449 };
450
451 on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_page, &ta, 1);
452}
453
454void flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
455 unsigned long end)
456{
457 struct tlb_args ta = {
458 .ta_vma = vma,
459 .ta_start = start,
460 .ta_end = end
461 };
462
463 on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_range, &ta, 1);
464}
465
466#ifdef CONFIG_TRANSPARENT_HUGEPAGE
467void flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
468 unsigned long end)
469{
470 struct tlb_args ta = {
471 .ta_vma = vma,
472 .ta_start = start,
473 .ta_end = end
474 };
475
476 on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_pmd_tlb_range, &ta, 1);
477}
478#endif
479
480void flush_tlb_kernel_range(unsigned long start, unsigned long end)
481{
482 struct tlb_args ta = {
483 .ta_start = start,
484 .ta_end = end
485 };
486
487 on_each_cpu(ipi_flush_tlb_kernel_range, &ta, 1);
488}
489#endif
490
491/*
492 * Routine to create a TLB entry
493 */
494void create_tlb(struct vm_area_struct *vma, unsigned long vaddr, pte_t *ptep)
495{
496 unsigned long flags;
497 unsigned int asid_or_sasid, rwx;
498 unsigned long pd0;
499 pte_t pd1;
500
501 /*
502 * create_tlb() assumes that current->mm == vma->mm, since
503 * -it ASID for TLB entry is fetched from MMU ASID reg (valid for curr)
504 * -completes the lazy write to SASID reg (again valid for curr tsk)
505 *
506 * Removing the assumption involves
507 * -Using vma->mm->context{ASID,SASID}, as opposed to MMU reg.
508 * -Fix the TLB paranoid debug code to not trigger false negatives.
509 * -More importantly it makes this handler inconsistent with fast-path
510 * TLB Refill handler which always deals with "current"
511 *
512 * Lets see the use cases when current->mm != vma->mm and we land here
513 * 1. execve->copy_strings()->__get_user_pages->handle_mm_fault
514 * Here VM wants to pre-install a TLB entry for user stack while
515 * current->mm still points to pre-execve mm (hence the condition).
516 * However the stack vaddr is soon relocated (randomization) and
517 * move_page_tables() tries to undo that TLB entry.
518 * Thus not creating TLB entry is not any worse.
519 *
520 * 2. ptrace(POKETEXT) causes a CoW - debugger(current) inserting a
521 * breakpoint in debugged task. Not creating a TLB now is not
522 * performance critical.
523 *
524 * Both the cases above are not good enough for code churn.
525 */
526 if (current->active_mm != vma->vm_mm)
527 return;
528
529 local_irq_save(flags);
530
531 tlb_paranoid_check(asid_mm(vma->vm_mm, smp_processor_id()), vaddr);
532
533 vaddr &= PAGE_MASK;
534
535 /* update this PTE credentials */
536 pte_val(*ptep) |= (_PAGE_PRESENT | _PAGE_ACCESSED);
537
538 /* Create HW TLB(PD0,PD1) from PTE */
539
540 /* ASID for this task */
541 asid_or_sasid = read_aux_reg(ARC_REG_PID) & 0xff;
542
543 pd0 = vaddr | asid_or_sasid | (pte_val(*ptep) & PTE_BITS_IN_PD0);
544
545 /*
546 * ARC MMU provides fully orthogonal access bits for K/U mode,
547 * however Linux only saves 1 set to save PTE real-estate
548 * Here we convert 3 PTE bits into 6 MMU bits:
549 * -Kernel only entries have Kr Kw Kx 0 0 0
550 * -User entries have mirrored K and U bits
551 */
552 rwx = pte_val(*ptep) & PTE_BITS_RWX;
553
554 if (pte_val(*ptep) & _PAGE_GLOBAL)
555 rwx <<= 3; /* r w x => Kr Kw Kx 0 0 0 */
556 else
557 rwx |= (rwx << 3); /* r w x => Kr Kw Kx Ur Uw Ux */
558
559 pd1 = rwx | (pte_val(*ptep) & PTE_BITS_NON_RWX_IN_PD1);
560
561 tlb_entry_insert(pd0, pd1);
562
563 local_irq_restore(flags);
564}
565
566/*
567 * Called at the end of pagefault, for a userspace mapped page
568 * -pre-install the corresponding TLB entry into MMU
569 * -Finalize the delayed D-cache flush of kernel mapping of page due to
570 * flush_dcache_page(), copy_user_page()
571 *
572 * Note that flush (when done) involves both WBACK - so physical page is
573 * in sync as well as INV - so any non-congruent aliases don't remain
574 */
575void update_mmu_cache(struct vm_area_struct *vma, unsigned long vaddr_unaligned,
576 pte_t *ptep)
577{
578 unsigned long vaddr = vaddr_unaligned & PAGE_MASK;
579 phys_addr_t paddr = pte_val(*ptep) & PAGE_MASK_PHYS;
580 struct page *page = pfn_to_page(pte_pfn(*ptep));
581
582 create_tlb(vma, vaddr, ptep);
583
584 if (page == ZERO_PAGE(0)) {
585 return;
586 }
587
588 /*
589 * Exec page : Independent of aliasing/page-color considerations,
590 * since icache doesn't snoop dcache on ARC, any dirty
591 * K-mapping of a code page needs to be wback+inv so that
592 * icache fetch by userspace sees code correctly.
593 * !EXEC page: If K-mapping is NOT congruent to U-mapping, flush it
594 * so userspace sees the right data.
595 * (Avoids the flush for Non-exec + congruent mapping case)
596 */
597 if ((vma->vm_flags & VM_EXEC) ||
598 addr_not_cache_congruent(paddr, vaddr)) {
599
600 int dirty = !test_and_set_bit(PG_dc_clean, &page->flags);
601 if (dirty) {
602 /* wback + inv dcache lines (K-mapping) */
603 __flush_dcache_page(paddr, paddr);
604
605 /* invalidate any existing icache lines (U-mapping) */
606 if (vma->vm_flags & VM_EXEC)
607 __inv_icache_page(paddr, vaddr);
608 }
609 }
610}
611
612#ifdef CONFIG_TRANSPARENT_HUGEPAGE
613
614/*
615 * MMUv4 in HS38x cores supports Super Pages which are basis for Linux THP
616 * support.
617 *
618 * Normal and Super pages can co-exist (ofcourse not overlap) in TLB with a
619 * new bit "SZ" in TLB page descriptor to distinguish between them.
620 * Super Page size is configurable in hardware (4K to 16M), but fixed once
621 * RTL builds.
622 *
623 * The exact THP size a Linux configuration will support is a function of:
624 * - MMU page size (typical 8K, RTL fixed)
625 * - software page walker address split between PGD:PTE:PFN (typical
626 * 11:8:13, but can be changed with 1 line)
627 * So for above default, THP size supported is 8K * (2^8) = 2M
628 *
629 * Default Page Walker is 2 levels, PGD:PTE:PFN, which in THP regime
630 * reduces to 1 level (as PTE is folded into PGD and canonically referred
631 * to as PMD).
632 * Thus THP PMD accessors are implemented in terms of PTE (just like sparc)
633 */
634
635void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
636 pmd_t *pmd)
637{
638 pte_t pte = __pte(pmd_val(*pmd));
639 update_mmu_cache(vma, addr, &pte);
640}
641
642void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
643 pgtable_t pgtable)
644{
645 struct list_head *lh = (struct list_head *) pgtable;
646
647 assert_spin_locked(&mm->page_table_lock);
648
649 /* FIFO */
650 if (!pmd_huge_pte(mm, pmdp))
651 INIT_LIST_HEAD(lh);
652 else
653 list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
654 pmd_huge_pte(mm, pmdp) = pgtable;
655}
656
657pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
658{
659 struct list_head *lh;
660 pgtable_t pgtable;
661
662 assert_spin_locked(&mm->page_table_lock);
663
664 pgtable = pmd_huge_pte(mm, pmdp);
665 lh = (struct list_head *) pgtable;
666 if (list_empty(lh))
667 pmd_huge_pte(mm, pmdp) = NULL;
668 else {
669 pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
670 list_del(lh);
671 }
672
673 pte_val(pgtable[0]) = 0;
674 pte_val(pgtable[1]) = 0;
675
676 return pgtable;
677}
678
679void local_flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
680 unsigned long end)
681{
682 unsigned int cpu;
683 unsigned long flags;
684
685 local_irq_save(flags);
686
687 cpu = smp_processor_id();
688
689 if (likely(asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID)) {
690 unsigned int asid = hw_pid(vma->vm_mm, cpu);
691
692 /* No need to loop here: this will always be for 1 Huge Page */
693 tlb_entry_erase(start | _PAGE_HW_SZ | asid);
694 }
695
696 local_irq_restore(flags);
697}
698
699#endif
700
701/* Read the Cache Build Configuration Registers, Decode them and save into
702 * the cpuinfo structure for later use.
703 * No Validation is done here, simply read/convert the BCRs
704 */
705void read_decode_mmu_bcr(void)
706{
707 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
708 unsigned int tmp;
709 struct bcr_mmu_1_2 {
710#ifdef CONFIG_CPU_BIG_ENDIAN
711 unsigned int ver:8, ways:4, sets:4, u_itlb:8, u_dtlb:8;
712#else
713 unsigned int u_dtlb:8, u_itlb:8, sets:4, ways:4, ver:8;
714#endif
715 } *mmu2;
716
717 struct bcr_mmu_3 {
718#ifdef CONFIG_CPU_BIG_ENDIAN
719 unsigned int ver:8, ways:4, sets:4, res:3, sasid:1, pg_sz:4,
720 u_itlb:4, u_dtlb:4;
721#else
722 unsigned int u_dtlb:4, u_itlb:4, pg_sz:4, sasid:1, res:3, sets:4,
723 ways:4, ver:8;
724#endif
725 } *mmu3;
726
727 struct bcr_mmu_4 {
728#ifdef CONFIG_CPU_BIG_ENDIAN
729 unsigned int ver:8, sasid:1, sz1:4, sz0:4, res:2, pae:1,
730 n_ways:2, n_entry:2, n_super:2, u_itlb:3, u_dtlb:3;
731#else
732 /* DTLB ITLB JES JE JA */
733 unsigned int u_dtlb:3, u_itlb:3, n_super:2, n_entry:2, n_ways:2,
734 pae:1, res:2, sz0:4, sz1:4, sasid:1, ver:8;
735#endif
736 } *mmu4;
737
738 tmp = read_aux_reg(ARC_REG_MMU_BCR);
739 mmu->ver = (tmp >> 24);
740
741 if (is_isa_arcompact()) {
742 if (mmu->ver <= 2) {
743 mmu2 = (struct bcr_mmu_1_2 *)&tmp;
744 mmu->pg_sz_k = TO_KB(0x2000);
745 mmu->sets = 1 << mmu2->sets;
746 mmu->ways = 1 << mmu2->ways;
747 mmu->u_dtlb = mmu2->u_dtlb;
748 mmu->u_itlb = mmu2->u_itlb;
749 } else {
750 mmu3 = (struct bcr_mmu_3 *)&tmp;
751 mmu->pg_sz_k = 1 << (mmu3->pg_sz - 1);
752 mmu->sets = 1 << mmu3->sets;
753 mmu->ways = 1 << mmu3->ways;
754 mmu->u_dtlb = mmu3->u_dtlb;
755 mmu->u_itlb = mmu3->u_itlb;
756 mmu->sasid = mmu3->sasid;
757 }
758 } else {
759 mmu4 = (struct bcr_mmu_4 *)&tmp;
760 mmu->pg_sz_k = 1 << (mmu4->sz0 - 1);
761 mmu->s_pg_sz_m = 1 << (mmu4->sz1 - 11);
762 mmu->sets = 64 << mmu4->n_entry;
763 mmu->ways = mmu4->n_ways * 2;
764 mmu->u_dtlb = mmu4->u_dtlb * 4;
765 mmu->u_itlb = mmu4->u_itlb * 4;
766 mmu->sasid = mmu4->sasid;
767 pae_exists = mmu->pae = mmu4->pae;
768 }
769}
770
771char *arc_mmu_mumbojumbo(int cpu_id, char *buf, int len)
772{
773 int n = 0;
774 struct cpuinfo_arc_mmu *p_mmu = &cpuinfo_arc700[cpu_id].mmu;
775 char super_pg[64] = "";
776
777 if (p_mmu->s_pg_sz_m)
778 scnprintf(super_pg, 64, "%dM Super Page %s",
779 p_mmu->s_pg_sz_m,
780 IS_USED_CFG(CONFIG_TRANSPARENT_HUGEPAGE));
781
782 n += scnprintf(buf + n, len - n,
783 "MMU [v%x]\t: %dk PAGE, %sJTLB %d (%dx%d), uDTLB %d, uITLB %d%s%s\n",
784 p_mmu->ver, p_mmu->pg_sz_k, super_pg,
785 p_mmu->sets * p_mmu->ways, p_mmu->sets, p_mmu->ways,
786 p_mmu->u_dtlb, p_mmu->u_itlb,
787 IS_AVAIL2(p_mmu->pae, ", PAE40 ", CONFIG_ARC_HAS_PAE40));
788
789 return buf;
790}
791
792int pae40_exist_but_not_enab(void)
793{
794 return pae_exists && !is_pae40_enabled();
795}
796
797void arc_mmu_init(void)
798{
799 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
800 char str[256];
801 int compat = 0;
802
803 pr_info("%s", arc_mmu_mumbojumbo(0, str, sizeof(str)));
804
805 /*
806 * Can't be done in processor.h due to header include dependencies
807 */
808 BUILD_BUG_ON(!IS_ALIGNED((CONFIG_ARC_KVADDR_SIZE << 20), PMD_SIZE));
809
810 /*
811 * stack top size sanity check,
812 * Can't be done in processor.h due to header include dependencies
813 */
814 BUILD_BUG_ON(!IS_ALIGNED(STACK_TOP, PMD_SIZE));
815
816 /*
817 * Ensure that MMU features assumed by kernel exist in hardware.
818 * For older ARC700 cpus, it has to be exact match, since the MMU
819 * revisions were not backwards compatible (MMUv3 TLB layout changed
820 * so even if kernel for v2 didn't use any new cmds of v3, it would
821 * still not work.
822 * For HS cpus, MMUv4 was baseline and v5 is backwards compatible
823 * (will run older software).
824 */
825 if (is_isa_arcompact() && mmu->ver == CONFIG_ARC_MMU_VER)
826 compat = 1;
827 else if (is_isa_arcv2() && mmu->ver >= CONFIG_ARC_MMU_VER)
828 compat = 1;
829
830 if (!compat) {
831 panic("MMU ver %d doesn't match kernel built for %d...\n",
832 mmu->ver, CONFIG_ARC_MMU_VER);
833 }
834
835 if (mmu->pg_sz_k != TO_KB(PAGE_SIZE))
836 panic("MMU pg size != PAGE_SIZE (%luk)\n", TO_KB(PAGE_SIZE));
837
838 if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
839 mmu->s_pg_sz_m != TO_MB(HPAGE_PMD_SIZE))
840 panic("MMU Super pg size != Linux HPAGE_PMD_SIZE (%luM)\n",
841 (unsigned long)TO_MB(HPAGE_PMD_SIZE));
842
843 if (IS_ENABLED(CONFIG_ARC_HAS_PAE40) && !mmu->pae)
844 panic("Hardware doesn't support PAE40\n");
845
846 /* Enable the MMU */
847 write_aux_reg(ARC_REG_PID, MMU_ENABLE);
848
849 /* In smp we use this reg for interrupt 1 scratch */
850#ifdef ARC_USE_SCRATCH_REG
851 /* swapper_pg_dir is the pgd for the kernel, used by vmalloc */
852 write_aux_reg(ARC_REG_SCRATCH_DATA0, swapper_pg_dir);
853#endif
854
855 if (pae40_exist_but_not_enab())
856 write_aux_reg(ARC_REG_TLBPD1HI, 0);
857}
858
859/*
860 * TLB Programmer's Model uses Linear Indexes: 0 to {255, 511} for 128 x {2,4}
861 * The mapping is Column-first.
862 * --------------------- -----------
863 * |way0|way1|way2|way3| |way0|way1|
864 * --------------------- -----------
865 * [set0] | 0 | 1 | 2 | 3 | | 0 | 1 |
866 * [set1] | 4 | 5 | 6 | 7 | | 2 | 3 |
867 * ~ ~ ~ ~
868 * [set127] | 508| 509| 510| 511| | 254| 255|
869 * --------------------- -----------
870 * For normal operations we don't(must not) care how above works since
871 * MMU cmd getIndex(vaddr) abstracts that out.
872 * However for walking WAYS of a SET, we need to know this
873 */
874#define SET_WAY_TO_IDX(mmu, set, way) ((set) * mmu->ways + (way))
875
876/* Handling of Duplicate PD (TLB entry) in MMU.
877 * -Could be due to buggy customer tapeouts or obscure kernel bugs
878 * -MMU complaints not at the time of duplicate PD installation, but at the
879 * time of lookup matching multiple ways.
880 * -Ideally these should never happen - but if they do - workaround by deleting
881 * the duplicate one.
882 * -Knob to be verbose abt it.(TODO: hook them up to debugfs)
883 */
884volatile int dup_pd_silent; /* Be silent abt it or complain (default) */
885
886void do_tlb_overlap_fault(unsigned long cause, unsigned long address,
887 struct pt_regs *regs)
888{
889 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
890 unsigned long flags;
891 int set, n_ways = mmu->ways;
892
893 n_ways = min(n_ways, 4);
894 BUG_ON(mmu->ways > 4);
895
896 local_irq_save(flags);
897
898 /* loop thru all sets of TLB */
899 for (set = 0; set < mmu->sets; set++) {
900
901 int is_valid, way;
902 unsigned int pd0[4];
903
904 /* read out all the ways of current set */
905 for (way = 0, is_valid = 0; way < n_ways; way++) {
906 write_aux_reg(ARC_REG_TLBINDEX,
907 SET_WAY_TO_IDX(mmu, set, way));
908 write_aux_reg(ARC_REG_TLBCOMMAND, TLBRead);
909 pd0[way] = read_aux_reg(ARC_REG_TLBPD0);
910 is_valid |= pd0[way] & _PAGE_PRESENT;
911 pd0[way] &= PAGE_MASK;
912 }
913
914 /* If all the WAYS in SET are empty, skip to next SET */
915 if (!is_valid)
916 continue;
917
918 /* Scan the set for duplicate ways: needs a nested loop */
919 for (way = 0; way < n_ways - 1; way++) {
920
921 int n;
922
923 if (!pd0[way])
924 continue;
925
926 for (n = way + 1; n < n_ways; n++) {
927 if (pd0[way] != pd0[n])
928 continue;
929
930 if (!dup_pd_silent)
931 pr_info("Dup TLB PD0 %08x @ set %d ways %d,%d\n",
932 pd0[way], set, way, n);
933
934 /*
935 * clear entry @way and not @n.
936 * This is critical to our optimised loop
937 */
938 pd0[way] = 0;
939 write_aux_reg(ARC_REG_TLBINDEX,
940 SET_WAY_TO_IDX(mmu, set, way));
941 __tlb_entry_erase();
942 }
943 }
944 }
945
946 local_irq_restore(flags);
947}
948
949/***********************************************************************
950 * Diagnostic Routines
951 * -Called from Low Level TLB Handlers if things don;t look good
952 **********************************************************************/
953
954#ifdef CONFIG_ARC_DBG_TLB_PARANOIA
955
956/*
957 * Low Level ASM TLB handler calls this if it finds that HW and SW ASIDS
958 * don't match
959 */
960void print_asid_mismatch(int mm_asid, int mmu_asid, int is_fast_path)
961{
962 pr_emerg("ASID Mismatch in %s Path Handler: sw-pid=0x%x hw-pid=0x%x\n",
963 is_fast_path ? "Fast" : "Slow", mm_asid, mmu_asid);
964
965 __asm__ __volatile__("flag 1");
966}
967
968void tlb_paranoid_check(unsigned int mm_asid, unsigned long addr)
969{
970 unsigned int mmu_asid;
971
972 mmu_asid = read_aux_reg(ARC_REG_PID) & 0xff;
973
974 /*
975 * At the time of a TLB miss/installation
976 * - HW version needs to match SW version
977 * - SW needs to have a valid ASID
978 */
979 if (addr < 0x70000000 &&
980 ((mm_asid == MM_CTXT_NO_ASID) ||
981 (mmu_asid != (mm_asid & MM_CTXT_ASID_MASK))))
982 print_asid_mismatch(mm_asid, mmu_asid, 0);
983}
984#endif