• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * TLB Management (flush/create/diagnostics) for ARC700
3  *
4  * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License version 2 as
8  * published by the Free Software Foundation.
9  *
10  * vineetg: Aug 2011
11  *  -Reintroduce duplicate PD fixup - some customer chips still have the issue
12  *
13  * vineetg: May 2011
14  *  -No need to flush_cache_page( ) for each call to update_mmu_cache()
15  *   some of the LMBench tests improved amazingly
16  *      = page-fault thrice as fast (75 usec to 28 usec)
17  *      = mmap twice as fast (9.6 msec to 4.6 msec),
18  *      = fork (5.3 msec to 3.7 msec)
19  *
20  * vineetg: April 2011 :
21  *  -MMU v3: PD{0,1} bits layout changed: They don't overlap anymore,
22  *      helps avoid a shift when preparing PD0 from PTE
23  *
24  * vineetg: April 2011 : Preparing for MMU V3
25  *  -MMU v2/v3 BCRs decoded differently
26  *  -Remove TLB_SIZE hardcoding as it's variable now: 256 or 512
27  *  -tlb_entry_erase( ) can be void
28  *  -local_flush_tlb_range( ):
29  *      = need not "ceil" @end
30  *      = walks MMU only if range spans < 32 entries, as opposed to 256
31  *
32  * Vineetg: Sept 10th 2008
33  *  -Changes related to MMU v2 (Rel 4.8)
34  *
35  * Vineetg: Aug 29th 2008
36  *  -In TLB Flush operations (Metal Fix MMU) there is a explict command to
37  *    flush Micro-TLBS. If TLB Index Reg is invalid prior to TLBIVUTLB cmd,
38  *    it fails. Thus need to load it with ANY valid value before invoking
39  *    TLBIVUTLB cmd
40  *
41  * Vineetg: Aug 21th 2008:
42  *  -Reduced the duration of IRQ lockouts in TLB Flush routines
43  *  -Multiple copies of TLB erase code seperated into a "single" function
44  *  -In TLB Flush routines, interrupt disabling moved UP to retrieve ASID
45  *       in interrupt-safe region.
46  *
47  * Vineetg: April 23rd Bug #93131
48  *    Problem: tlb_flush_kernel_range() doesnt do anything if the range to
49  *              flush is more than the size of TLB itself.
50  *
51  * Rahul Trivedi : Codito Technologies 2004
52  */
53 
54 #include <linux/module.h>
55 #include <asm/arcregs.h>
56 #include <asm/setup.h>
57 #include <asm/mmu_context.h>
58 #include <asm/tlb.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 /* A copy of the ASID from the PID reg is kept in asid_cache */
101 int asid_cache = FIRST_ASID;
102 
103 /* ASID to mm struct mapping. We have one extra entry corresponding to
104  * NO_ASID to save us a compare when clearing the mm entry for old asid
105  * see get_new_mmu_context (asm-arc/mmu_context.h)
106  */
107 struct mm_struct *asid_mm_map[NUM_ASID + 1];
108 
109 /*
110  * Utility Routine to erase a J-TLB entry
111  * The procedure is to look it up in the MMU. If found, ERASE it by
112  *  issuing a TlbWrite CMD with PD0 = PD1 = 0
113  */
114 
__tlb_entry_erase(void)115 static void __tlb_entry_erase(void)
116 {
117 	write_aux_reg(ARC_REG_TLBPD1, 0);
118 	write_aux_reg(ARC_REG_TLBPD0, 0);
119 	write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
120 }
121 
tlb_entry_erase(unsigned int vaddr_n_asid)122 static void tlb_entry_erase(unsigned int vaddr_n_asid)
123 {
124 	unsigned int idx;
125 
126 	/* Locate the TLB entry for this vaddr + ASID */
127 	write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid);
128 	write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe);
129 	idx = read_aux_reg(ARC_REG_TLBINDEX);
130 
131 	/* No error means entry found, zero it out */
132 	if (likely(!(idx & TLB_LKUP_ERR))) {
133 		__tlb_entry_erase();
134 	} else {		/* Some sort of Error */
135 
136 		/* Duplicate entry error */
137 		if (idx & 0x1) {
138 			/* TODO we need to handle this case too */
139 			pr_emerg("unhandled Duplicate flush for %x\n",
140 			       vaddr_n_asid);
141 		}
142 		/* else entry not found so nothing to do */
143 	}
144 }
145 
146 /****************************************************************************
147  * ARC700 MMU caches recently used J-TLB entries (RAM) as uTLBs (FLOPs)
148  *
149  * New IVUTLB cmd in MMU v2 explictly invalidates the uTLB
150  *
151  * utlb_invalidate ( )
152  *  -For v2 MMU calls Flush uTLB Cmd
153  *  -For v1 MMU does nothing (except for Metal Fix v1 MMU)
154  *      This is because in v1 TLBWrite itself invalidate uTLBs
155  ***************************************************************************/
156 
utlb_invalidate(void)157 static void utlb_invalidate(void)
158 {
159 #if (CONFIG_ARC_MMU_VER >= 2)
160 
161 #if (CONFIG_ARC_MMU_VER < 3)
162 	/* MMU v2 introduced the uTLB Flush command.
163 	 * There was however an obscure hardware bug, where uTLB flush would
164 	 * fail when a prior probe for J-TLB (both totally unrelated) would
165 	 * return lkup err - because the entry didnt exist in MMU.
166 	 * The Workround was to set Index reg with some valid value, prior to
167 	 * flush. This was fixed in MMU v3 hence not needed any more
168 	 */
169 	unsigned int idx;
170 
171 	/* make sure INDEX Reg is valid */
172 	idx = read_aux_reg(ARC_REG_TLBINDEX);
173 
174 	/* If not write some dummy val */
175 	if (unlikely(idx & TLB_LKUP_ERR))
176 		write_aux_reg(ARC_REG_TLBINDEX, 0xa);
177 #endif
178 
179 	write_aux_reg(ARC_REG_TLBCOMMAND, TLBIVUTLB);
180 #endif
181 
182 }
183 
184 /*
185  * Un-conditionally (without lookup) erase the entire MMU contents
186  */
187 
local_flush_tlb_all(void)188 noinline void local_flush_tlb_all(void)
189 {
190 	unsigned long flags;
191 	unsigned int entry;
192 	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
193 
194 	local_irq_save(flags);
195 
196 	/* Load PD0 and PD1 with template for a Blank Entry */
197 	write_aux_reg(ARC_REG_TLBPD1, 0);
198 	write_aux_reg(ARC_REG_TLBPD0, 0);
199 
200 	for (entry = 0; entry < mmu->num_tlb; entry++) {
201 		/* write this entry to the TLB */
202 		write_aux_reg(ARC_REG_TLBINDEX, entry);
203 		write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
204 	}
205 
206 	utlb_invalidate();
207 
208 	local_irq_restore(flags);
209 }
210 
211 /*
212  * Flush the entrie MM for userland. The fastest way is to move to Next ASID
213  */
local_flush_tlb_mm(struct mm_struct * mm)214 noinline void local_flush_tlb_mm(struct mm_struct *mm)
215 {
216 	/*
217 	 * Small optimisation courtesy IA64
218 	 * flush_mm called during fork,exit,munmap etc, multiple times as well.
219 	 * Only for fork( ) do we need to move parent to a new MMU ctxt,
220 	 * all other cases are NOPs, hence this check.
221 	 */
222 	if (atomic_read(&mm->mm_users) == 0)
223 		return;
224 
225 	/*
226 	 * Workaround for Android weirdism:
227 	 * A binder VMA could end up in a task such that vma->mm != tsk->mm
228 	 * old code would cause h/w - s/w ASID to get out of sync
229 	 */
230 	if (current->mm != mm)
231 		destroy_context(mm);
232 	else
233 		get_new_mmu_context(mm);
234 }
235 
236 /*
237  * Flush a Range of TLB entries for userland.
238  * @start is inclusive, while @end is exclusive
239  * Difference between this and Kernel Range Flush is
240  *  -Here the fastest way (if range is too large) is to move to next ASID
241  *      without doing any explicit Shootdown
242  *  -In case of kernel Flush, entry has to be shot down explictly
243  */
local_flush_tlb_range(struct vm_area_struct * vma,unsigned long start,unsigned long end)244 void local_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
245 			   unsigned long end)
246 {
247 	unsigned long flags;
248 	unsigned int asid;
249 
250 	/* If range @start to @end is more than 32 TLB entries deep,
251 	 * its better to move to a new ASID rather than searching for
252 	 * individual entries and then shooting them down
253 	 *
254 	 * The calc above is rough, doesn't account for unaligned parts,
255 	 * since this is heuristics based anyways
256 	 */
257 	if (unlikely((end - start) >= PAGE_SIZE * 32)) {
258 		local_flush_tlb_mm(vma->vm_mm);
259 		return;
260 	}
261 
262 	/*
263 	 * @start moved to page start: this alone suffices for checking
264 	 * loop end condition below, w/o need for aligning @end to end
265 	 * e.g. 2000 to 4001 will anyhow loop twice
266 	 */
267 	start &= PAGE_MASK;
268 
269 	local_irq_save(flags);
270 	asid = vma->vm_mm->context.asid;
271 
272 	if (asid != NO_ASID) {
273 		while (start < end) {
274 			tlb_entry_erase(start | (asid & 0xff));
275 			start += PAGE_SIZE;
276 		}
277 	}
278 
279 	utlb_invalidate();
280 
281 	local_irq_restore(flags);
282 }
283 
284 /* Flush the kernel TLB entries - vmalloc/modules (Global from MMU perspective)
285  *  @start, @end interpreted as kvaddr
286  * Interestingly, shared TLB entries can also be flushed using just
287  * @start,@end alone (interpreted as user vaddr), although technically SASID
288  * is also needed. However our smart TLbProbe lookup takes care of that.
289  */
local_flush_tlb_kernel_range(unsigned long start,unsigned long end)290 void local_flush_tlb_kernel_range(unsigned long start, unsigned long end)
291 {
292 	unsigned long flags;
293 
294 	/* exactly same as above, except for TLB entry not taking ASID */
295 
296 	if (unlikely((end - start) >= PAGE_SIZE * 32)) {
297 		local_flush_tlb_all();
298 		return;
299 	}
300 
301 	start &= PAGE_MASK;
302 
303 	local_irq_save(flags);
304 	while (start < end) {
305 		tlb_entry_erase(start);
306 		start += PAGE_SIZE;
307 	}
308 
309 	utlb_invalidate();
310 
311 	local_irq_restore(flags);
312 }
313 
314 /*
315  * Delete TLB entry in MMU for a given page (??? address)
316  * NOTE One TLB entry contains translation for single PAGE
317  */
318 
local_flush_tlb_page(struct vm_area_struct * vma,unsigned long page)319 void local_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
320 {
321 	unsigned long flags;
322 
323 	/* Note that it is critical that interrupts are DISABLED between
324 	 * checking the ASID and using it flush the TLB entry
325 	 */
326 	local_irq_save(flags);
327 
328 	if (vma->vm_mm->context.asid != NO_ASID) {
329 		tlb_entry_erase((page & PAGE_MASK) |
330 				(vma->vm_mm->context.asid & 0xff));
331 		utlb_invalidate();
332 	}
333 
334 	local_irq_restore(flags);
335 }
336 
337 /*
338  * Routine to create a TLB entry
339  */
create_tlb(struct vm_area_struct * vma,unsigned long address,pte_t * ptep)340 void create_tlb(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
341 {
342 	unsigned long flags;
343 	unsigned int idx, asid_or_sasid;
344 	unsigned long pd0_flags;
345 
346 	/*
347 	 * create_tlb() assumes that current->mm == vma->mm, since
348 	 * -it ASID for TLB entry is fetched from MMU ASID reg (valid for curr)
349 	 * -completes the lazy write to SASID reg (again valid for curr tsk)
350 	 *
351 	 * Removing the assumption involves
352 	 * -Using vma->mm->context{ASID,SASID}, as opposed to MMU reg.
353 	 * -Fix the TLB paranoid debug code to not trigger false negatives.
354 	 * -More importantly it makes this handler inconsistent with fast-path
355 	 *  TLB Refill handler which always deals with "current"
356 	 *
357 	 * Lets see the use cases when current->mm != vma->mm and we land here
358 	 *  1. execve->copy_strings()->__get_user_pages->handle_mm_fault
359 	 *     Here VM wants to pre-install a TLB entry for user stack while
360 	 *     current->mm still points to pre-execve mm (hence the condition).
361 	 *     However the stack vaddr is soon relocated (randomization) and
362 	 *     move_page_tables() tries to undo that TLB entry.
363 	 *     Thus not creating TLB entry is not any worse.
364 	 *
365 	 *  2. ptrace(POKETEXT) causes a CoW - debugger(current) inserting a
366 	 *     breakpoint in debugged task. Not creating a TLB now is not
367 	 *     performance critical.
368 	 *
369 	 * Both the cases above are not good enough for code churn.
370 	 */
371 	if (current->active_mm != vma->vm_mm)
372 		return;
373 
374 	local_irq_save(flags);
375 
376 	tlb_paranoid_check(vma->vm_mm->context.asid, address);
377 
378 	address &= PAGE_MASK;
379 
380 	/* update this PTE credentials */
381 	pte_val(*ptep) |= (_PAGE_PRESENT | _PAGE_ACCESSED);
382 
383 	/* Create HW TLB entry Flags (in PD0) from PTE Flags */
384 #if (CONFIG_ARC_MMU_VER <= 2)
385 	pd0_flags = ((pte_val(*ptep) & PTE_BITS_IN_PD0) >> 1);
386 #else
387 	pd0_flags = ((pte_val(*ptep) & PTE_BITS_IN_PD0));
388 #endif
389 
390 	/* ASID for this task */
391 	asid_or_sasid = read_aux_reg(ARC_REG_PID) & 0xff;
392 
393 	write_aux_reg(ARC_REG_TLBPD0, address | pd0_flags | asid_or_sasid);
394 
395 	/* Load remaining info in PD1 (Page Frame Addr and Kx/Kw/Kr Flags) */
396 	write_aux_reg(ARC_REG_TLBPD1, (pte_val(*ptep) & PTE_BITS_IN_PD1));
397 
398 	/* First verify if entry for this vaddr+ASID already exists */
399 	write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe);
400 	idx = read_aux_reg(ARC_REG_TLBINDEX);
401 
402 	/*
403 	 * If Not already present get a free slot from MMU.
404 	 * Otherwise, Probe would have located the entry and set INDEX Reg
405 	 * with existing location. This will cause Write CMD to over-write
406 	 * existing entry with new PD0 and PD1
407 	 */
408 	if (likely(idx & TLB_LKUP_ERR))
409 		write_aux_reg(ARC_REG_TLBCOMMAND, TLBGetIndex);
410 
411 	/*
412 	 * Commit the Entry to MMU
413 	 * It doesnt sound safe to use the TLBWriteNI cmd here
414 	 * which doesn't flush uTLBs. I'd rather be safe than sorry.
415 	 */
416 	write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
417 
418 	local_irq_restore(flags);
419 }
420 
421 /*
422  * Called at the end of pagefault, for a userspace mapped page
423  *  -pre-install the corresponding TLB entry into MMU
424  *  -Finalize the delayed D-cache flush of kernel mapping of page due to
425  *  	flush_dcache_page(), copy_user_page()
426  *
427  * Note that flush (when done) involves both WBACK - so physical page is
428  * in sync as well as INV - so any non-congruent aliases don't remain
429  */
update_mmu_cache(struct vm_area_struct * vma,unsigned long vaddr_unaligned,pte_t * ptep)430 void update_mmu_cache(struct vm_area_struct *vma, unsigned long vaddr_unaligned,
431 		      pte_t *ptep)
432 {
433 	unsigned long vaddr = vaddr_unaligned & PAGE_MASK;
434 	unsigned long paddr = pte_val(*ptep) & PAGE_MASK;
435 
436 	create_tlb(vma, vaddr, ptep);
437 
438 	/*
439 	 * Exec page : Independent of aliasing/page-color considerations,
440 	 *	       since icache doesn't snoop dcache on ARC, any dirty
441 	 *	       K-mapping of a code page needs to be wback+inv so that
442 	 *	       icache fetch by userspace sees code correctly.
443 	 * !EXEC page: If K-mapping is NOT congruent to U-mapping, flush it
444 	 *	       so userspace sees the right data.
445 	 *  (Avoids the flush for Non-exec + congruent mapping case)
446 	 */
447 	if ((vma->vm_flags & VM_EXEC) ||
448 	     addr_not_cache_congruent(paddr, vaddr)) {
449 		struct page *page = pfn_to_page(pte_pfn(*ptep));
450 
451 		int dirty = test_and_clear_bit(PG_arch_1, &page->flags);
452 		if (dirty) {
453 			/* wback + inv dcache lines */
454 			__flush_dcache_page(paddr, paddr);
455 
456 			/* invalidate any existing icache lines */
457 			if (vma->vm_flags & VM_EXEC)
458 				__inv_icache_page(paddr, vaddr);
459 		}
460 	}
461 }
462 
463 /* Read the Cache Build Confuration Registers, Decode them and save into
464  * the cpuinfo structure for later use.
465  * No Validation is done here, simply read/convert the BCRs
466  */
read_decode_mmu_bcr(void)467 void __cpuinit read_decode_mmu_bcr(void)
468 {
469 	unsigned int tmp;
470 	struct bcr_mmu_1_2 *mmu2;	/* encoded MMU2 attr */
471 	struct bcr_mmu_3 *mmu3;		/* encoded MMU3 attr */
472 	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
473 
474 	tmp = read_aux_reg(ARC_REG_MMU_BCR);
475 	mmu->ver = (tmp >> 24);
476 
477 	if (mmu->ver <= 2) {
478 		mmu2 = (struct bcr_mmu_1_2 *)&tmp;
479 		mmu->pg_sz = PAGE_SIZE;
480 		mmu->sets = 1 << mmu2->sets;
481 		mmu->ways = 1 << mmu2->ways;
482 		mmu->u_dtlb = mmu2->u_dtlb;
483 		mmu->u_itlb = mmu2->u_itlb;
484 	} else {
485 		mmu3 = (struct bcr_mmu_3 *)&tmp;
486 		mmu->pg_sz = 512 << mmu3->pg_sz;
487 		mmu->sets = 1 << mmu3->sets;
488 		mmu->ways = 1 << mmu3->ways;
489 		mmu->u_dtlb = mmu3->u_dtlb;
490 		mmu->u_itlb = mmu3->u_itlb;
491 	}
492 
493 	mmu->num_tlb = mmu->sets * mmu->ways;
494 }
495 
arc_mmu_mumbojumbo(int cpu_id,char * buf,int len)496 char *arc_mmu_mumbojumbo(int cpu_id, char *buf, int len)
497 {
498 	int n = 0;
499 	struct cpuinfo_arc_mmu *p_mmu = &cpuinfo_arc700[cpu_id].mmu;
500 
501 	n += scnprintf(buf + n, len - n, "ARC700 MMU [v%x]\t: %dk PAGE, ",
502 		       p_mmu->ver, TO_KB(p_mmu->pg_sz));
503 
504 	n += scnprintf(buf + n, len - n,
505 		       "J-TLB %d (%dx%d), uDTLB %d, uITLB %d, %s\n",
506 		       p_mmu->num_tlb, p_mmu->sets, p_mmu->ways,
507 		       p_mmu->u_dtlb, p_mmu->u_itlb,
508 		       __CONFIG_ARC_MMU_SASID_VAL ? "SASID" : "");
509 
510 	return buf;
511 }
512 
arc_mmu_init(void)513 void __cpuinit arc_mmu_init(void)
514 {
515 	char str[256];
516 	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
517 
518 	printk(arc_mmu_mumbojumbo(0, str, sizeof(str)));
519 
520 	/* For efficiency sake, kernel is compile time built for a MMU ver
521 	 * This must match the hardware it is running on.
522 	 * Linux built for MMU V2, if run on MMU V1 will break down because V1
523 	 *  hardware doesn't understand cmds such as WriteNI, or IVUTLB
524 	 * On the other hand, Linux built for V1 if run on MMU V2 will do
525 	 *   un-needed workarounds to prevent memcpy thrashing.
526 	 * Similarly MMU V3 has new features which won't work on older MMU
527 	 */
528 	if (mmu->ver != CONFIG_ARC_MMU_VER) {
529 		panic("MMU ver %d doesn't match kernel built for %d...\n",
530 		      mmu->ver, CONFIG_ARC_MMU_VER);
531 	}
532 
533 	if (mmu->pg_sz != PAGE_SIZE)
534 		panic("MMU pg size != PAGE_SIZE (%luk)\n", TO_KB(PAGE_SIZE));
535 
536 	/*
537 	 * ASID mgmt data structures are compile time init
538 	 *  asid_cache = FIRST_ASID and asid_mm_map[] all zeroes
539 	 */
540 
541 	local_flush_tlb_all();
542 
543 	/* Enable the MMU */
544 	write_aux_reg(ARC_REG_PID, MMU_ENABLE);
545 
546 	/* In smp we use this reg for interrupt 1 scratch */
547 #ifndef CONFIG_SMP
548 	/* swapper_pg_dir is the pgd for the kernel, used by vmalloc */
549 	write_aux_reg(ARC_REG_SCRATCH_DATA0, swapper_pg_dir);
550 #endif
551 }
552 
553 /*
554  * TLB Programmer's Model uses Linear Indexes: 0 to {255, 511} for 128 x {2,4}
555  * The mapping is Column-first.
556  *		---------------------	-----------
557  *		|way0|way1|way2|way3|	|way0|way1|
558  *		---------------------	-----------
559  * [set0]	|  0 |  1 |  2 |  3 |	|  0 |  1 |
560  * [set1]	|  4 |  5 |  6 |  7 |	|  2 |  3 |
561  *		~		    ~	~	  ~
562  * [set127]	| 508| 509| 510| 511|	| 254| 255|
563  *		---------------------	-----------
564  * For normal operations we don't(must not) care how above works since
565  * MMU cmd getIndex(vaddr) abstracts that out.
566  * However for walking WAYS of a SET, we need to know this
567  */
568 #define SET_WAY_TO_IDX(mmu, set, way)  ((set) * mmu->ways + (way))
569 
570 /* Handling of Duplicate PD (TLB entry) in MMU.
571  * -Could be due to buggy customer tapeouts or obscure kernel bugs
572  * -MMU complaints not at the time of duplicate PD installation, but at the
573  *      time of lookup matching multiple ways.
574  * -Ideally these should never happen - but if they do - workaround by deleting
575  *      the duplicate one.
576  * -Knob to be verbose abt it.(TODO: hook them up to debugfs)
577  */
578 volatile int dup_pd_verbose = 1;/* Be slient abt it or complain (default) */
579 
do_tlb_overlap_fault(unsigned long cause,unsigned long address,struct pt_regs * regs)580 void do_tlb_overlap_fault(unsigned long cause, unsigned long address,
581 			  struct pt_regs *regs)
582 {
583 	int set, way, n;
584 	unsigned int pd0[4], pd1[4];	/* assume max 4 ways */
585 	unsigned long flags, is_valid;
586 	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
587 
588 	local_irq_save(flags);
589 
590 	/* re-enable the MMU */
591 	write_aux_reg(ARC_REG_PID, MMU_ENABLE | read_aux_reg(ARC_REG_PID));
592 
593 	/* loop thru all sets of TLB */
594 	for (set = 0; set < mmu->sets; set++) {
595 
596 		/* read out all the ways of current set */
597 		for (way = 0, is_valid = 0; way < mmu->ways; way++) {
598 			write_aux_reg(ARC_REG_TLBINDEX,
599 					  SET_WAY_TO_IDX(mmu, set, way));
600 			write_aux_reg(ARC_REG_TLBCOMMAND, TLBRead);
601 			pd0[way] = read_aux_reg(ARC_REG_TLBPD0);
602 			pd1[way] = read_aux_reg(ARC_REG_TLBPD1);
603 			is_valid |= pd0[way] & _PAGE_PRESENT;
604 		}
605 
606 		/* If all the WAYS in SET are empty, skip to next SET */
607 		if (!is_valid)
608 			continue;
609 
610 		/* Scan the set for duplicate ways: needs a nested loop */
611 		for (way = 0; way < mmu->ways; way++) {
612 			if (!pd0[way])
613 				continue;
614 
615 			for (n = way + 1; n < mmu->ways; n++) {
616 				if ((pd0[way] & PAGE_MASK) ==
617 				    (pd0[n] & PAGE_MASK)) {
618 
619 					if (dup_pd_verbose) {
620 						pr_info("Duplicate PD's @"
621 							"[%d:%d]/[%d:%d]\n",
622 						     set, way, set, n);
623 						pr_info("TLBPD0[%u]: %08x\n",
624 						     way, pd0[way]);
625 					}
626 
627 					/*
628 					 * clear entry @way and not @n. This is
629 					 * critical to our optimised loop
630 					 */
631 					pd0[way] = pd1[way] = 0;
632 					write_aux_reg(ARC_REG_TLBINDEX,
633 						SET_WAY_TO_IDX(mmu, set, way));
634 					__tlb_entry_erase();
635 				}
636 			}
637 		}
638 	}
639 
640 	local_irq_restore(flags);
641 }
642 
643 /***********************************************************************
644  * Diagnostic Routines
645  *  -Called from Low Level TLB Hanlders if things don;t look good
646  **********************************************************************/
647 
648 #ifdef CONFIG_ARC_DBG_TLB_PARANOIA
649 
650 /*
651  * Low Level ASM TLB handler calls this if it finds that HW and SW ASIDS
652  * don't match
653  */
print_asid_mismatch(int is_fast_path)654 void print_asid_mismatch(int is_fast_path)
655 {
656 	int pid_sw, pid_hw;
657 	pid_sw = current->active_mm->context.asid;
658 	pid_hw = read_aux_reg(ARC_REG_PID) & 0xff;
659 
660 	pr_emerg("ASID Mismatch in %s Path Handler: sw-pid=0x%x hw-pid=0x%x\n",
661 	       is_fast_path ? "Fast" : "Slow", pid_sw, pid_hw);
662 
663 	__asm__ __volatile__("flag 1");
664 }
665 
tlb_paranoid_check(unsigned int pid_sw,unsigned long addr)666 void tlb_paranoid_check(unsigned int pid_sw, unsigned long addr)
667 {
668 	unsigned int pid_hw;
669 
670 	pid_hw = read_aux_reg(ARC_REG_PID) & 0xff;
671 
672 	if (addr < 0x70000000 && ((pid_hw != pid_sw) || (pid_sw == NO_ASID)))
673 		print_asid_mismatch(0);
674 }
675 #endif
676