Linux Audio

Check our new training course

Loading...
v6.8
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * TLB Management (flush/create/diagnostics) for MMUv3 and MMUv4
  4 *
  5 * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
  6 *
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  7 */
  8
  9#include <linux/module.h>
 10#include <linux/bug.h>
 11#include <linux/mm_types.h>
 12
 13#include <asm/arcregs.h>
 14#include <asm/setup.h>
 15#include <asm/mmu_context.h>
 16#include <asm/mmu.h>
 17
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 18/* A copy of the ASID from the PID reg is kept in asid_cache */
 19DEFINE_PER_CPU(unsigned int, asid_cache) = MM_CTXT_FIRST_CYCLE;
 20
 21static struct cpuinfo_arc_mmu {
 22	unsigned int ver, pg_sz_k, s_pg_sz_m, pae, sets, ways;
 23} mmuinfo;
 24
 25/*
 26 * Utility Routine to erase a J-TLB entry
 27 * Caller needs to setup Index Reg (manually or via getIndex)
 28 */
 29static inline void __tlb_entry_erase(void)
 30{
 31	write_aux_reg(ARC_REG_TLBPD1, 0);
 32
 33	if (is_pae40_enabled())
 34		write_aux_reg(ARC_REG_TLBPD1HI, 0);
 35
 36	write_aux_reg(ARC_REG_TLBPD0, 0);
 37	write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
 38}
 39
 40static void utlb_invalidate(void)
 41{
 42	write_aux_reg(ARC_REG_TLBCOMMAND, TLBIVUTLB);
 43}
 44
 45#ifdef CONFIG_ARC_MMU_V3
 46
 47static inline unsigned int tlb_entry_lkup(unsigned long vaddr_n_asid)
 48{
 49	unsigned int idx;
 50
 51	write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid);
 52
 53	write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe);
 54	idx = read_aux_reg(ARC_REG_TLBINDEX);
 55
 56	return idx;
 57}
 58
 59static void tlb_entry_erase(unsigned int vaddr_n_asid)
 60{
 61	unsigned int idx;
 62
 63	/* Locate the TLB entry for this vaddr + ASID */
 64	idx = tlb_entry_lkup(vaddr_n_asid);
 65
 66	/* No error means entry found, zero it out */
 67	if (likely(!(idx & TLB_LKUP_ERR))) {
 68		__tlb_entry_erase();
 69	} else {
 70		/* Duplicate entry error */
 71		WARN(idx == TLB_DUP_ERR, "Probe returned Dup PD for %x\n",
 72					   vaddr_n_asid);
 73	}
 74}
 75
 76static void tlb_entry_insert(unsigned int pd0, phys_addr_t pd1)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 77{
 78	unsigned int idx;
 79
 80	/*
 81	 * First verify if entry for this vaddr+ASID already exists
 82	 * This also sets up PD0 (vaddr, ASID..) for final commit
 83	 */
 84	idx = tlb_entry_lkup(pd0);
 85
 86	/*
 87	 * If Not already present get a free slot from MMU.
 88	 * Otherwise, Probe would have located the entry and set INDEX Reg
 89	 * with existing location. This will cause Write CMD to over-write
 90	 * existing entry with new PD0 and PD1
 91	 */
 92	if (likely(idx & TLB_LKUP_ERR))
 93		write_aux_reg(ARC_REG_TLBCOMMAND, TLBGetIndex);
 94
 95	/* setup the other half of TLB entry (pfn, rwx..) */
 96	write_aux_reg(ARC_REG_TLBPD1, pd1);
 97
 98	/*
 99	 * Commit the Entry to MMU
100	 * It doesn't sound safe to use the TLBWriteNI cmd here
101	 * which doesn't flush uTLBs. I'd rather be safe than sorry.
102	 */
103	write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
104}
105
106#else	/* MMUv4 */
 
 
 
 
 
107
108static void tlb_entry_erase(unsigned int vaddr_n_asid)
109{
110	write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid | _PAGE_PRESENT);
111	write_aux_reg(ARC_REG_TLBCOMMAND, TLBDeleteEntry);
112}
113
114static void tlb_entry_insert(unsigned int pd0, phys_addr_t pd1)
115{
116	write_aux_reg(ARC_REG_TLBPD0, pd0);
 
117
118	if (!is_pae40_enabled()) {
119		write_aux_reg(ARC_REG_TLBPD1, pd1);
120	} else {
121		write_aux_reg(ARC_REG_TLBPD1, pd1 & 0xFFFFFFFF);
122		write_aux_reg(ARC_REG_TLBPD1HI, (u64)pd1 >> 32);
123	}
124
125	write_aux_reg(ARC_REG_TLBCOMMAND, TLBInsertEntry);
126}
127
128#endif
129
130/*
131 * Un-conditionally (without lookup) erase the entire MMU contents
132 */
133
134noinline void local_flush_tlb_all(void)
135{
136	struct cpuinfo_arc_mmu *mmu = &mmuinfo;
137	unsigned long flags;
138	unsigned int entry;
139	int num_tlb = mmu->sets * mmu->ways;
140
141	local_irq_save(flags);
142
143	/* Load PD0 and PD1 with template for a Blank Entry */
144	write_aux_reg(ARC_REG_TLBPD1, 0);
145
146	if (is_pae40_enabled())
147		write_aux_reg(ARC_REG_TLBPD1HI, 0);
148
149	write_aux_reg(ARC_REG_TLBPD0, 0);
150
151	for (entry = 0; entry < num_tlb; entry++) {
152		/* write this entry to the TLB */
153		write_aux_reg(ARC_REG_TLBINDEX, entry);
154		write_aux_reg(ARC_REG_TLBCOMMAND, TLBWriteNI);
155	}
156
157	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
158		const int stlb_idx = 0x800;
159
160		/* Blank sTLB entry */
161		write_aux_reg(ARC_REG_TLBPD0, _PAGE_HW_SZ);
162
163		for (entry = stlb_idx; entry < stlb_idx + 16; entry++) {
164			write_aux_reg(ARC_REG_TLBINDEX, entry);
165			write_aux_reg(ARC_REG_TLBCOMMAND, TLBWriteNI);
166		}
167	}
168
169	utlb_invalidate();
170
171	local_irq_restore(flags);
172}
173
174/*
175 * Flush the entire MM for userland. The fastest way is to move to Next ASID
176 */
177noinline void local_flush_tlb_mm(struct mm_struct *mm)
178{
179	/*
180	 * Small optimisation courtesy IA64
181	 * flush_mm called during fork,exit,munmap etc, multiple times as well.
182	 * Only for fork( ) do we need to move parent to a new MMU ctxt,
183	 * all other cases are NOPs, hence this check.
184	 */
185	if (atomic_read(&mm->mm_users) == 0)
186		return;
187
188	/*
189	 * - Move to a new ASID, but only if the mm is still wired in
190	 *   (Android Binder ended up calling this for vma->mm != tsk->mm,
191	 *    causing h/w - s/w ASID to get out of sync)
192	 * - Also get_new_mmu_context() new implementation allocates a new
193	 *   ASID only if it is not allocated already - so unallocate first
194	 */
195	destroy_context(mm);
196	if (current->mm == mm)
197		get_new_mmu_context(mm);
198}
199
200/*
201 * Flush a Range of TLB entries for userland.
202 * @start is inclusive, while @end is exclusive
203 * Difference between this and Kernel Range Flush is
204 *  -Here the fastest way (if range is too large) is to move to next ASID
205 *      without doing any explicit Shootdown
206 *  -In case of kernel Flush, entry has to be shot down explicitly
207 */
208void local_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
209			   unsigned long end)
210{
211	const unsigned int cpu = smp_processor_id();
212	unsigned long flags;
213
214	/* If range @start to @end is more than 32 TLB entries deep,
215	 * its better to move to a new ASID rather than searching for
216	 * individual entries and then shooting them down
217	 *
218	 * The calc above is rough, doesn't account for unaligned parts,
219	 * since this is heuristics based anyways
220	 */
221	if (unlikely((end - start) >= PAGE_SIZE * 32)) {
222		local_flush_tlb_mm(vma->vm_mm);
223		return;
224	}
225
226	/*
227	 * @start moved to page start: this alone suffices for checking
228	 * loop end condition below, w/o need for aligning @end to end
229	 * e.g. 2000 to 4001 will anyhow loop twice
230	 */
231	start &= PAGE_MASK;
232
233	local_irq_save(flags);
234
235	if (asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID) {
236		while (start < end) {
237			tlb_entry_erase(start | hw_pid(vma->vm_mm, cpu));
238			start += PAGE_SIZE;
239		}
240	}
241
 
 
242	local_irq_restore(flags);
243}
244
245/* Flush the kernel TLB entries - vmalloc/modules (Global from MMU perspective)
246 *  @start, @end interpreted as kvaddr
247 * Interestingly, shared TLB entries can also be flushed using just
248 * @start,@end alone (interpreted as user vaddr), although technically SASID
249 * is also needed. However our smart TLbProbe lookup takes care of that.
250 */
251void local_flush_tlb_kernel_range(unsigned long start, unsigned long end)
252{
253	unsigned long flags;
254
255	/* exactly same as above, except for TLB entry not taking ASID */
256
257	if (unlikely((end - start) >= PAGE_SIZE * 32)) {
258		local_flush_tlb_all();
259		return;
260	}
261
262	start &= PAGE_MASK;
263
264	local_irq_save(flags);
265	while (start < end) {
266		tlb_entry_erase(start);
267		start += PAGE_SIZE;
268	}
269
 
 
270	local_irq_restore(flags);
271}
272
273/*
274 * Delete TLB entry in MMU for a given page (??? address)
275 * NOTE One TLB entry contains translation for single PAGE
276 */
277
278void local_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
279{
280	const unsigned int cpu = smp_processor_id();
281	unsigned long flags;
282
283	/* Note that it is critical that interrupts are DISABLED between
284	 * checking the ASID and using it flush the TLB entry
285	 */
286	local_irq_save(flags);
287
288	if (asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID) {
289		tlb_entry_erase((page & PAGE_MASK) | hw_pid(vma->vm_mm, cpu));
 
290	}
291
292	local_irq_restore(flags);
293}
294
295#ifdef CONFIG_SMP
296
297struct tlb_args {
298	struct vm_area_struct *ta_vma;
299	unsigned long ta_start;
300	unsigned long ta_end;
301};
302
303static inline void ipi_flush_tlb_page(void *arg)
304{
305	struct tlb_args *ta = arg;
306
307	local_flush_tlb_page(ta->ta_vma, ta->ta_start);
308}
309
310static inline void ipi_flush_tlb_range(void *arg)
311{
312	struct tlb_args *ta = arg;
313
314	local_flush_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
315}
316
317#ifdef CONFIG_TRANSPARENT_HUGEPAGE
318static inline void ipi_flush_pmd_tlb_range(void *arg)
319{
320	struct tlb_args *ta = arg;
321
322	local_flush_pmd_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
323}
324#endif
325
326static inline void ipi_flush_tlb_kernel_range(void *arg)
327{
328	struct tlb_args *ta = (struct tlb_args *)arg;
329
330	local_flush_tlb_kernel_range(ta->ta_start, ta->ta_end);
331}
332
333void flush_tlb_all(void)
334{
335	on_each_cpu((smp_call_func_t)local_flush_tlb_all, NULL, 1);
336}
337
338void flush_tlb_mm(struct mm_struct *mm)
339{
340	on_each_cpu_mask(mm_cpumask(mm), (smp_call_func_t)local_flush_tlb_mm,
341			 mm, 1);
342}
343
344void flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr)
345{
346	struct tlb_args ta = {
347		.ta_vma = vma,
348		.ta_start = uaddr
349	};
350
351	on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_page, &ta, 1);
352}
353
354void flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
355		     unsigned long end)
356{
357	struct tlb_args ta = {
358		.ta_vma = vma,
359		.ta_start = start,
360		.ta_end = end
361	};
362
363	on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_range, &ta, 1);
364}
365
366#ifdef CONFIG_TRANSPARENT_HUGEPAGE
367void flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
368			 unsigned long end)
369{
370	struct tlb_args ta = {
371		.ta_vma = vma,
372		.ta_start = start,
373		.ta_end = end
374	};
375
376	on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_pmd_tlb_range, &ta, 1);
377}
378#endif
379
380void flush_tlb_kernel_range(unsigned long start, unsigned long end)
381{
382	struct tlb_args ta = {
383		.ta_start = start,
384		.ta_end = end
385	};
386
387	on_each_cpu(ipi_flush_tlb_kernel_range, &ta, 1);
388}
389#endif
390
391/*
392 * Routine to create a TLB entry
393 */
394static void create_tlb(struct vm_area_struct *vma, unsigned long vaddr, pte_t *ptep)
395{
396	unsigned long flags;
397	unsigned int asid_or_sasid, rwx;
398	unsigned long pd0;
399	phys_addr_t pd1;
400
401	/*
402	 * create_tlb() assumes that current->mm == vma->mm, since
403	 * -it ASID for TLB entry is fetched from MMU ASID reg (valid for curr)
404	 * -completes the lazy write to SASID reg (again valid for curr tsk)
405	 *
406	 * Removing the assumption involves
407	 * -Using vma->mm->context{ASID,SASID}, as opposed to MMU reg.
 
408	 * -More importantly it makes this handler inconsistent with fast-path
409	 *  TLB Refill handler which always deals with "current"
410	 *
411	 * Lets see the use cases when current->mm != vma->mm and we land here
412	 *  1. execve->copy_strings()->__get_user_pages->handle_mm_fault
413	 *     Here VM wants to pre-install a TLB entry for user stack while
414	 *     current->mm still points to pre-execve mm (hence the condition).
415	 *     However the stack vaddr is soon relocated (randomization) and
416	 *     move_page_tables() tries to undo that TLB entry.
417	 *     Thus not creating TLB entry is not any worse.
418	 *
419	 *  2. ptrace(POKETEXT) causes a CoW - debugger(current) inserting a
420	 *     breakpoint in debugged task. Not creating a TLB now is not
421	 *     performance critical.
422	 *
423	 * Both the cases above are not good enough for code churn.
424	 */
425	if (current->active_mm != vma->vm_mm)
426		return;
427
428	local_irq_save(flags);
429
 
 
430	vaddr &= PAGE_MASK;
431
432	/* update this PTE credentials */
433	pte_val(*ptep) |= (_PAGE_PRESENT | _PAGE_ACCESSED);
434
435	/* Create HW TLB(PD0,PD1) from PTE  */
436
437	/* ASID for this task */
438	asid_or_sasid = read_aux_reg(ARC_REG_PID) & 0xff;
439
440	pd0 = vaddr | asid_or_sasid | (pte_val(*ptep) & PTE_BITS_IN_PD0);
441
442	/*
443	 * ARC MMU provides fully orthogonal access bits for K/U mode,
444	 * however Linux only saves 1 set to save PTE real-estate
445	 * Here we convert 3 PTE bits into 6 MMU bits:
446	 * -Kernel only entries have Kr Kw Kx 0 0 0
447	 * -User entries have mirrored K and U bits
448	 */
449	rwx = pte_val(*ptep) & PTE_BITS_RWX;
450
451	if (pte_val(*ptep) & _PAGE_GLOBAL)
452		rwx <<= 3;		/* r w x => Kr Kw Kx 0 0 0 */
453	else
454		rwx |= (rwx << 3);	/* r w x => Kr Kw Kx Ur Uw Ux */
455
456	pd1 = rwx | (pte_val(*ptep) & PTE_BITS_NON_RWX_IN_PD1);
457
458	tlb_entry_insert(pd0, pd1);
459
460	local_irq_restore(flags);
461}
462
463/*
464 * Called at the end of pagefault, for a userspace mapped page
465 *  -pre-install the corresponding TLB entry into MMU
466 *  -Finalize the delayed D-cache flush of kernel mapping of page due to
467 *  	flush_dcache_page(), copy_user_page()
468 *
469 * Note that flush (when done) involves both WBACK - so physical page is
470 * in sync as well as INV - so any non-congruent aliases don't remain
471 */
472void update_mmu_cache_range(struct vm_fault *vmf, struct vm_area_struct *vma,
473		unsigned long vaddr_unaligned, pte_t *ptep, unsigned int nr)
474{
475	unsigned long vaddr = vaddr_unaligned & PAGE_MASK;
476	phys_addr_t paddr = pte_val(*ptep) & PAGE_MASK_PHYS;
477	struct page *page = pfn_to_page(pte_pfn(*ptep));
478
479	create_tlb(vma, vaddr, ptep);
480
481	if (page == ZERO_PAGE(0))
482		return;
 
483
484	/*
485	 * For executable pages, since icache doesn't snoop dcache, any
486	 * dirty K-mapping of a code page needs to be wback+inv so that
487	 * icache fetch by userspace sees code correctly.
488	 */
489	if (vma->vm_flags & VM_EXEC) {
490		struct folio *folio = page_folio(page);
491		int dirty = !test_and_set_bit(PG_dc_clean, &folio->flags);
 
 
 
 
 
492		if (dirty) {
493			unsigned long offset = offset_in_folio(folio, paddr);
494			nr = folio_nr_pages(folio);
495			paddr -= offset;
496			vaddr -= offset;
497			/* wback + inv dcache lines (K-mapping) */
498			__flush_dcache_pages(paddr, paddr, nr);
499
500			/* invalidate any existing icache lines (U-mapping) */
501			if (vma->vm_flags & VM_EXEC)
502				__inv_icache_pages(paddr, vaddr, nr);
503		}
504	}
505}
506
507#ifdef CONFIG_TRANSPARENT_HUGEPAGE
508
509/*
510 * MMUv4 in HS38x cores supports Super Pages which are basis for Linux THP
511 * support.
512 *
513 * Normal and Super pages can co-exist (ofcourse not overlap) in TLB with a
514 * new bit "SZ" in TLB page descriptor to distinguish between them.
515 * Super Page size is configurable in hardware (4K to 16M), but fixed once
516 * RTL builds.
517 *
518 * The exact THP size a Linux configuration will support is a function of:
519 *  - MMU page size (typical 8K, RTL fixed)
520 *  - software page walker address split between PGD:PTE:PFN (typical
521 *    11:8:13, but can be changed with 1 line)
522 * So for above default, THP size supported is 8K * (2^8) = 2M
523 *
524 * Default Page Walker is 2 levels, PGD:PTE:PFN, which in THP regime
525 * reduces to 1 level (as PTE is folded into PGD and canonically referred
526 * to as PMD).
527 * Thus THP PMD accessors are implemented in terms of PTE (just like sparc)
528 */
529
530void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
531				 pmd_t *pmd)
532{
533	pte_t pte = __pte(pmd_val(*pmd));
534	update_mmu_cache_range(NULL, vma, addr, &pte, HPAGE_PMD_NR);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
535}
536
537void local_flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
538			       unsigned long end)
539{
540	unsigned int cpu;
541	unsigned long flags;
542
543	local_irq_save(flags);
544
545	cpu = smp_processor_id();
546
547	if (likely(asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID)) {
548		unsigned int asid = hw_pid(vma->vm_mm, cpu);
549
550		/* No need to loop here: this will always be for 1 Huge Page */
551		tlb_entry_erase(start | _PAGE_HW_SZ | asid);
552	}
553
554	local_irq_restore(flags);
555}
556
557#endif
558
559/* Read the Cache Build Configuration Registers, Decode them and save into
560 * the cpuinfo structure for later use.
561 * No Validation is done here, simply read/convert the BCRs
562 */
563int arc_mmu_mumbojumbo(int c, char *buf, int len)
564{
565	struct cpuinfo_arc_mmu *mmu = &mmuinfo;
566	unsigned int bcr, u_dtlb, u_itlb, sasid;
567	struct bcr_mmu_3 *mmu3;
568	struct bcr_mmu_4 *mmu4;
569	char super_pg[64] = "";
570	int n = 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
571
572	bcr = read_aux_reg(ARC_REG_MMU_BCR);
573	mmu->ver = (bcr >> 24);
574
575	if (is_isa_arcompact() && mmu->ver == 3) {
576		mmu3 = (struct bcr_mmu_3 *)&bcr;
 
 
 
 
 
 
 
577		mmu->pg_sz_k = 1 << (mmu3->pg_sz - 1);
578		mmu->sets = 1 << mmu3->sets;
579		mmu->ways = 1 << mmu3->ways;
580		u_dtlb = mmu3->u_dtlb;
581		u_itlb = mmu3->u_itlb;
582		sasid = mmu3->sasid;
583	} else {
584		mmu4 = (struct bcr_mmu_4 *)&bcr;
585		mmu->pg_sz_k = 1 << (mmu4->sz0 - 1);
586		mmu->s_pg_sz_m = 1 << (mmu4->sz1 - 11);
587		mmu->sets = 64 << mmu4->n_entry;
588		mmu->ways = mmu4->n_ways * 2;
589		u_dtlb = mmu4->u_dtlb * 4;
590		u_itlb = mmu4->u_itlb * 4;
591		sasid = mmu4->sasid;
592		mmu->pae = mmu4->pae;
593	}
 
594
595	if (mmu->s_pg_sz_m)
596		scnprintf(super_pg, 64, "/%dM%s",
597			  mmu->s_pg_sz_m,
598			  IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) ? " (THP enabled)":"");
 
599
600	n += scnprintf(buf + n, len - n,
601		      "MMU [v%x]\t: %dk%s, swalk %d lvl, JTLB %dx%d, uDTLB %d, uITLB %d%s%s%s\n",
602		       mmu->ver, mmu->pg_sz_k, super_pg, CONFIG_PGTABLE_LEVELS,
603		       mmu->sets, mmu->ways,
604		       u_dtlb, u_itlb,
605		       IS_AVAIL1(sasid, ", SASID"),
606		       IS_AVAIL2(mmu->pae, ", PAE40 ", CONFIG_ARC_HAS_PAE40));
607
608	return n;
609}
 
 
 
 
610
611int pae40_exist_but_not_enab(void)
612{
613	return mmuinfo.pae && !is_pae40_enabled();
614}
615
616void arc_mmu_init(void)
617{
618	struct cpuinfo_arc_mmu *mmu = &mmuinfo;
619	int compat = 0;
620
621	/*
622	 * Can't be done in processor.h due to header include dependencies
623	 */
624	BUILD_BUG_ON(!IS_ALIGNED((CONFIG_ARC_KVADDR_SIZE << 20), PMD_SIZE));
625
626	/*
627	 * stack top size sanity check,
628	 * Can't be done in processor.h due to header include dependencies
629	 */
630	BUILD_BUG_ON(!IS_ALIGNED(STACK_TOP, PMD_SIZE));
631
632	/*
633	 * Ensure that MMU features assumed by kernel exist in hardware.
634	 *  - For older ARC700 cpus, only v3 supported
635	 *  - For HS cpus, v4 was baseline and v5 is backwards compatible
636	 *    (will run older software).
637	 */
638	if (is_isa_arcompact() && mmu->ver == 3)
639		compat = 1;
640	else if (is_isa_arcv2() && mmu->ver >= 4)
641		compat = 1;
642
643	if (!compat)
644		panic("MMU ver %d doesn't match kernel built for\n", mmu->ver);
 
 
 
 
 
 
 
 
 
 
645
646	if (mmu->pg_sz_k != TO_KB(PAGE_SIZE))
647		panic("MMU pg size != PAGE_SIZE (%luk)\n", TO_KB(PAGE_SIZE));
648
649	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
650	    mmu->s_pg_sz_m != TO_MB(HPAGE_PMD_SIZE))
651		panic("MMU Super pg size != Linux HPAGE_PMD_SIZE (%luM)\n",
652		      (unsigned long)TO_MB(HPAGE_PMD_SIZE));
653
654	if (IS_ENABLED(CONFIG_ARC_HAS_PAE40) && !mmu->pae)
655		panic("Hardware doesn't support PAE40\n");
656
657	/* Enable the MMU with ASID 0 */
658	mmu_setup_asid(NULL, 0);
659
660	/* cache the pgd pointer in MMU SCRATCH reg (ARCv2 only) */
661	mmu_setup_pgd(NULL, swapper_pg_dir);
662
663	if (pae40_exist_but_not_enab())
664		write_aux_reg(ARC_REG_TLBPD1HI, 0);
665}
666
667/*
668 * TLB Programmer's Model uses Linear Indexes: 0 to {255, 511} for 128 x {2,4}
669 * The mapping is Column-first.
670 *		---------------------	-----------
671 *		|way0|way1|way2|way3|	|way0|way1|
672 *		---------------------	-----------
673 * [set0]	|  0 |  1 |  2 |  3 |	|  0 |  1 |
674 * [set1]	|  4 |  5 |  6 |  7 |	|  2 |  3 |
675 *		~		    ~	~	  ~
676 * [set127]	| 508| 509| 510| 511|	| 254| 255|
677 *		---------------------	-----------
678 * For normal operations we don't(must not) care how above works since
679 * MMU cmd getIndex(vaddr) abstracts that out.
680 * However for walking WAYS of a SET, we need to know this
681 */
682#define SET_WAY_TO_IDX(mmu, set, way)  ((set) * mmu->ways + (way))
683
684/* Handling of Duplicate PD (TLB entry) in MMU.
685 * -Could be due to buggy customer tapeouts or obscure kernel bugs
686 * -MMU complaints not at the time of duplicate PD installation, but at the
687 *      time of lookup matching multiple ways.
688 * -Ideally these should never happen - but if they do - workaround by deleting
689 *      the duplicate one.
690 * -Knob to be verbose abt it.(TODO: hook them up to debugfs)
691 */
692volatile int dup_pd_silent; /* Be silent abt it or complain (default) */
693
694void do_tlb_overlap_fault(unsigned long cause, unsigned long address,
695			  struct pt_regs *regs)
696{
697	struct cpuinfo_arc_mmu *mmu = &mmuinfo;
 
698	unsigned long flags;
699	int set, n_ways = mmu->ways;
700
701	n_ways = min(n_ways, 4);
702	BUG_ON(mmu->ways > 4);
703
704	local_irq_save(flags);
705
 
 
 
706	/* loop thru all sets of TLB */
707	for (set = 0; set < mmu->sets; set++) {
708
709		int is_valid, way;
710		unsigned int pd0[4];
711
712		/* read out all the ways of current set */
713		for (way = 0, is_valid = 0; way < n_ways; way++) {
714			write_aux_reg(ARC_REG_TLBINDEX,
715					  SET_WAY_TO_IDX(mmu, set, way));
716			write_aux_reg(ARC_REG_TLBCOMMAND, TLBRead);
717			pd0[way] = read_aux_reg(ARC_REG_TLBPD0);
718			is_valid |= pd0[way] & _PAGE_PRESENT;
719			pd0[way] &= PAGE_MASK;
720		}
721
722		/* If all the WAYS in SET are empty, skip to next SET */
723		if (!is_valid)
724			continue;
725
726		/* Scan the set for duplicate ways: needs a nested loop */
727		for (way = 0; way < n_ways - 1; way++) {
728
729			int n;
730
731			if (!pd0[way])
732				continue;
733
734			for (n = way + 1; n < n_ways; n++) {
735				if (pd0[way] != pd0[n])
736					continue;
737
738				if (!dup_pd_silent)
739					pr_info("Dup TLB PD0 %08x @ set %d ways %d,%d\n",
740						pd0[way], set, way, n);
741
742				/*
743				 * clear entry @way and not @n.
744				 * This is critical to our optimised loop
745				 */
746				pd0[way] = 0;
747				write_aux_reg(ARC_REG_TLBINDEX,
748						SET_WAY_TO_IDX(mmu, set, way));
749				__tlb_entry_erase();
750			}
751		}
752	}
753
754	local_irq_restore(flags);
755}
v4.6
 
  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() doesn't 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 <linux/bug.h>
 
 
 56#include <asm/arcregs.h>
 57#include <asm/setup.h>
 58#include <asm/mmu_context.h>
 59#include <asm/mmu.h>
 60
 61/*			Need for ARC MMU v2
 62 *
 63 * ARC700 MMU-v1 had a Joint-TLB for Code and Data and is 2 way set-assoc.
 64 * For a memcpy operation with 3 players (src/dst/code) such that all 3 pages
 65 * map into same set, there would be contention for the 2 ways causing severe
 66 * Thrashing.
 67 *
 68 * Although J-TLB is 2 way set assoc, ARC700 caches J-TLB into uTLBS which has
 69 * much higher associativity. u-D-TLB is 8 ways, u-I-TLB is 4 ways.
 70 * Given this, the thrasing problem should never happen because once the 3
 71 * J-TLB entries are created (even though 3rd will knock out one of the prev
 72 * two), the u-D-TLB and u-I-TLB will have what is required to accomplish memcpy
 73 *
 74 * Yet we still see the Thrashing because a J-TLB Write cause flush of u-TLBs.
 75 * This is a simple design for keeping them in sync. So what do we do?
 76 * The solution which James came up was pretty neat. It utilised the assoc
 77 * of uTLBs by not invalidating always but only when absolutely necessary.
 78 *
 79 * - Existing TLB commands work as before
 80 * - New command (TLBWriteNI) for TLB write without clearing uTLBs
 81 * - New command (TLBIVUTLB) to invalidate uTLBs.
 82 *
 83 * The uTLBs need only be invalidated when pages are being removed from the
 84 * OS page table. If a 'victim' TLB entry is being overwritten in the main TLB
 85 * as a result of a miss, the removed entry is still allowed to exist in the
 86 * uTLBs as it is still valid and present in the OS page table. This allows the
 87 * full associativity of the uTLBs to hide the limited associativity of the main
 88 * TLB.
 89 *
 90 * During a miss handler, the new "TLBWriteNI" command is used to load
 91 * entries without clearing the uTLBs.
 92 *
 93 * When the OS page table is updated, TLB entries that may be associated with a
 94 * removed page are removed (flushed) from the TLB using TLBWrite. In this
 95 * circumstance, the uTLBs must also be cleared. This is done by using the
 96 * existing TLBWrite command. An explicit IVUTLB is also required for those
 97 * corner cases when TLBWrite was not executed at all because the corresp
 98 * J-TLB entry got evicted/replaced.
 99 */
100
101
102/* A copy of the ASID from the PID reg is kept in asid_cache */
103DEFINE_PER_CPU(unsigned int, asid_cache) = MM_CTXT_FIRST_CYCLE;
104
 
 
 
 
105/*
106 * Utility Routine to erase a J-TLB entry
107 * Caller needs to setup Index Reg (manually or via getIndex)
108 */
109static inline void __tlb_entry_erase(void)
110{
111	write_aux_reg(ARC_REG_TLBPD1, 0);
112
113	if (is_pae40_enabled())
114		write_aux_reg(ARC_REG_TLBPD1HI, 0);
115
116	write_aux_reg(ARC_REG_TLBPD0, 0);
117	write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
118}
119
120#if (CONFIG_ARC_MMU_VER < 4)
 
 
 
 
 
121
122static inline unsigned int tlb_entry_lkup(unsigned long vaddr_n_asid)
123{
124	unsigned int idx;
125
126	write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid);
127
128	write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe);
129	idx = read_aux_reg(ARC_REG_TLBINDEX);
130
131	return idx;
132}
133
134static void tlb_entry_erase(unsigned int vaddr_n_asid)
135{
136	unsigned int idx;
137
138	/* Locate the TLB entry for this vaddr + ASID */
139	idx = tlb_entry_lkup(vaddr_n_asid);
140
141	/* No error means entry found, zero it out */
142	if (likely(!(idx & TLB_LKUP_ERR))) {
143		__tlb_entry_erase();
144	} else {
145		/* Duplicate entry error */
146		WARN(idx == TLB_DUP_ERR, "Probe returned Dup PD for %x\n",
147					   vaddr_n_asid);
148	}
149}
150
151/****************************************************************************
152 * ARC700 MMU caches recently used J-TLB entries (RAM) as uTLBs (FLOPs)
153 *
154 * New IVUTLB cmd in MMU v2 explictly invalidates the uTLB
155 *
156 * utlb_invalidate ( )
157 *  -For v2 MMU calls Flush uTLB Cmd
158 *  -For v1 MMU does nothing (except for Metal Fix v1 MMU)
159 *      This is because in v1 TLBWrite itself invalidate uTLBs
160 ***************************************************************************/
161
162static void utlb_invalidate(void)
163{
164#if (CONFIG_ARC_MMU_VER >= 2)
165
166#if (CONFIG_ARC_MMU_VER == 2)
167	/* MMU v2 introduced the uTLB Flush command.
168	 * There was however an obscure hardware bug, where uTLB flush would
169	 * fail when a prior probe for J-TLB (both totally unrelated) would
170	 * return lkup err - because the entry didn't exist in MMU.
171	 * The Workround was to set Index reg with some valid value, prior to
172	 * flush. This was fixed in MMU v3 hence not needed any more
173	 */
174	unsigned int idx;
175
176	/* make sure INDEX Reg is valid */
177	idx = read_aux_reg(ARC_REG_TLBINDEX);
178
179	/* If not write some dummy val */
180	if (unlikely(idx & TLB_LKUP_ERR))
181		write_aux_reg(ARC_REG_TLBINDEX, 0xa);
182#endif
183
184	write_aux_reg(ARC_REG_TLBCOMMAND, TLBIVUTLB);
185#endif
186
187}
188
189static void tlb_entry_insert(unsigned int pd0, pte_t pd1)
190{
191	unsigned int idx;
192
193	/*
194	 * First verify if entry for this vaddr+ASID already exists
195	 * This also sets up PD0 (vaddr, ASID..) for final commit
196	 */
197	idx = tlb_entry_lkup(pd0);
198
199	/*
200	 * If Not already present get a free slot from MMU.
201	 * Otherwise, Probe would have located the entry and set INDEX Reg
202	 * with existing location. This will cause Write CMD to over-write
203	 * existing entry with new PD0 and PD1
204	 */
205	if (likely(idx & TLB_LKUP_ERR))
206		write_aux_reg(ARC_REG_TLBCOMMAND, TLBGetIndex);
207
208	/* setup the other half of TLB entry (pfn, rwx..) */
209	write_aux_reg(ARC_REG_TLBPD1, pd1);
210
211	/*
212	 * Commit the Entry to MMU
213	 * It doesn't sound safe to use the TLBWriteNI cmd here
214	 * which doesn't flush uTLBs. I'd rather be safe than sorry.
215	 */
216	write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
217}
218
219#else	/* CONFIG_ARC_MMU_VER >= 4) */
220
221static void utlb_invalidate(void)
222{
223	/* No need since uTLB is always in sync with JTLB */
224}
225
226static void tlb_entry_erase(unsigned int vaddr_n_asid)
227{
228	write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid | _PAGE_PRESENT);
229	write_aux_reg(ARC_REG_TLBCOMMAND, TLBDeleteEntry);
230}
231
232static void tlb_entry_insert(unsigned int pd0, pte_t pd1)
233{
234	write_aux_reg(ARC_REG_TLBPD0, pd0);
235	write_aux_reg(ARC_REG_TLBPD1, pd1);
236
237	if (is_pae40_enabled())
 
 
 
238		write_aux_reg(ARC_REG_TLBPD1HI, (u64)pd1 >> 32);
 
239
240	write_aux_reg(ARC_REG_TLBCOMMAND, TLBInsertEntry);
241}
242
243#endif
244
245/*
246 * Un-conditionally (without lookup) erase the entire MMU contents
247 */
248
249noinline void local_flush_tlb_all(void)
250{
251	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
252	unsigned long flags;
253	unsigned int entry;
254	int num_tlb = mmu->sets * mmu->ways;
255
256	local_irq_save(flags);
257
258	/* Load PD0 and PD1 with template for a Blank Entry */
259	write_aux_reg(ARC_REG_TLBPD1, 0);
260
261	if (is_pae40_enabled())
262		write_aux_reg(ARC_REG_TLBPD1HI, 0);
263
264	write_aux_reg(ARC_REG_TLBPD0, 0);
265
266	for (entry = 0; entry < num_tlb; entry++) {
267		/* write this entry to the TLB */
268		write_aux_reg(ARC_REG_TLBINDEX, entry);
269		write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
270	}
271
272	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
273		const int stlb_idx = 0x800;
274
275		/* Blank sTLB entry */
276		write_aux_reg(ARC_REG_TLBPD0, _PAGE_HW_SZ);
277
278		for (entry = stlb_idx; entry < stlb_idx + 16; entry++) {
279			write_aux_reg(ARC_REG_TLBINDEX, entry);
280			write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
281		}
282	}
283
284	utlb_invalidate();
285
286	local_irq_restore(flags);
287}
288
289/*
290 * Flush the entrie MM for userland. The fastest way is to move to Next ASID
291 */
292noinline void local_flush_tlb_mm(struct mm_struct *mm)
293{
294	/*
295	 * Small optimisation courtesy IA64
296	 * flush_mm called during fork,exit,munmap etc, multiple times as well.
297	 * Only for fork( ) do we need to move parent to a new MMU ctxt,
298	 * all other cases are NOPs, hence this check.
299	 */
300	if (atomic_read(&mm->mm_users) == 0)
301		return;
302
303	/*
304	 * - Move to a new ASID, but only if the mm is still wired in
305	 *   (Android Binder ended up calling this for vma->mm != tsk->mm,
306	 *    causing h/w - s/w ASID to get out of sync)
307	 * - Also get_new_mmu_context() new implementation allocates a new
308	 *   ASID only if it is not allocated already - so unallocate first
309	 */
310	destroy_context(mm);
311	if (current->mm == mm)
312		get_new_mmu_context(mm);
313}
314
315/*
316 * Flush a Range of TLB entries for userland.
317 * @start is inclusive, while @end is exclusive
318 * Difference between this and Kernel Range Flush is
319 *  -Here the fastest way (if range is too large) is to move to next ASID
320 *      without doing any explicit Shootdown
321 *  -In case of kernel Flush, entry has to be shot down explictly
322 */
323void local_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
324			   unsigned long end)
325{
326	const unsigned int cpu = smp_processor_id();
327	unsigned long flags;
328
329	/* If range @start to @end is more than 32 TLB entries deep,
330	 * its better to move to a new ASID rather than searching for
331	 * individual entries and then shooting them down
332	 *
333	 * The calc above is rough, doesn't account for unaligned parts,
334	 * since this is heuristics based anyways
335	 */
336	if (unlikely((end - start) >= PAGE_SIZE * 32)) {
337		local_flush_tlb_mm(vma->vm_mm);
338		return;
339	}
340
341	/*
342	 * @start moved to page start: this alone suffices for checking
343	 * loop end condition below, w/o need for aligning @end to end
344	 * e.g. 2000 to 4001 will anyhow loop twice
345	 */
346	start &= PAGE_MASK;
347
348	local_irq_save(flags);
349
350	if (asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID) {
351		while (start < end) {
352			tlb_entry_erase(start | hw_pid(vma->vm_mm, cpu));
353			start += PAGE_SIZE;
354		}
355	}
356
357	utlb_invalidate();
358
359	local_irq_restore(flags);
360}
361
362/* Flush the kernel TLB entries - vmalloc/modules (Global from MMU perspective)
363 *  @start, @end interpreted as kvaddr
364 * Interestingly, shared TLB entries can also be flushed using just
365 * @start,@end alone (interpreted as user vaddr), although technically SASID
366 * is also needed. However our smart TLbProbe lookup takes care of that.
367 */
368void local_flush_tlb_kernel_range(unsigned long start, unsigned long end)
369{
370	unsigned long flags;
371
372	/* exactly same as above, except for TLB entry not taking ASID */
373
374	if (unlikely((end - start) >= PAGE_SIZE * 32)) {
375		local_flush_tlb_all();
376		return;
377	}
378
379	start &= PAGE_MASK;
380
381	local_irq_save(flags);
382	while (start < end) {
383		tlb_entry_erase(start);
384		start += PAGE_SIZE;
385	}
386
387	utlb_invalidate();
388
389	local_irq_restore(flags);
390}
391
392/*
393 * Delete TLB entry in MMU for a given page (??? address)
394 * NOTE One TLB entry contains translation for single PAGE
395 */
396
397void local_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
398{
399	const unsigned int cpu = smp_processor_id();
400	unsigned long flags;
401
402	/* Note that it is critical that interrupts are DISABLED between
403	 * checking the ASID and using it flush the TLB entry
404	 */
405	local_irq_save(flags);
406
407	if (asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID) {
408		tlb_entry_erase((page & PAGE_MASK) | hw_pid(vma->vm_mm, cpu));
409		utlb_invalidate();
410	}
411
412	local_irq_restore(flags);
413}
414
415#ifdef CONFIG_SMP
416
417struct tlb_args {
418	struct vm_area_struct *ta_vma;
419	unsigned long ta_start;
420	unsigned long ta_end;
421};
422
423static inline void ipi_flush_tlb_page(void *arg)
424{
425	struct tlb_args *ta = arg;
426
427	local_flush_tlb_page(ta->ta_vma, ta->ta_start);
428}
429
430static inline void ipi_flush_tlb_range(void *arg)
431{
432	struct tlb_args *ta = arg;
433
434	local_flush_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
435}
436
437#ifdef CONFIG_TRANSPARENT_HUGEPAGE
438static inline void ipi_flush_pmd_tlb_range(void *arg)
439{
440	struct tlb_args *ta = arg;
441
442	local_flush_pmd_tlb_range(ta->ta_vma, ta->ta_start, ta->ta_end);
443}
444#endif
445
446static inline void ipi_flush_tlb_kernel_range(void *arg)
447{
448	struct tlb_args *ta = (struct tlb_args *)arg;
449
450	local_flush_tlb_kernel_range(ta->ta_start, ta->ta_end);
451}
452
453void flush_tlb_all(void)
454{
455	on_each_cpu((smp_call_func_t)local_flush_tlb_all, NULL, 1);
456}
457
458void flush_tlb_mm(struct mm_struct *mm)
459{
460	on_each_cpu_mask(mm_cpumask(mm), (smp_call_func_t)local_flush_tlb_mm,
461			 mm, 1);
462}
463
464void flush_tlb_page(struct vm_area_struct *vma, unsigned long uaddr)
465{
466	struct tlb_args ta = {
467		.ta_vma = vma,
468		.ta_start = uaddr
469	};
470
471	on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_page, &ta, 1);
472}
473
474void flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
475		     unsigned long end)
476{
477	struct tlb_args ta = {
478		.ta_vma = vma,
479		.ta_start = start,
480		.ta_end = end
481	};
482
483	on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_tlb_range, &ta, 1);
484}
485
486#ifdef CONFIG_TRANSPARENT_HUGEPAGE
487void flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
488			 unsigned long end)
489{
490	struct tlb_args ta = {
491		.ta_vma = vma,
492		.ta_start = start,
493		.ta_end = end
494	};
495
496	on_each_cpu_mask(mm_cpumask(vma->vm_mm), ipi_flush_pmd_tlb_range, &ta, 1);
497}
498#endif
499
500void flush_tlb_kernel_range(unsigned long start, unsigned long end)
501{
502	struct tlb_args ta = {
503		.ta_start = start,
504		.ta_end = end
505	};
506
507	on_each_cpu(ipi_flush_tlb_kernel_range, &ta, 1);
508}
509#endif
510
511/*
512 * Routine to create a TLB entry
513 */
514void create_tlb(struct vm_area_struct *vma, unsigned long vaddr, pte_t *ptep)
515{
516	unsigned long flags;
517	unsigned int asid_or_sasid, rwx;
518	unsigned long pd0;
519	pte_t pd1;
520
521	/*
522	 * create_tlb() assumes that current->mm == vma->mm, since
523	 * -it ASID for TLB entry is fetched from MMU ASID reg (valid for curr)
524	 * -completes the lazy write to SASID reg (again valid for curr tsk)
525	 *
526	 * Removing the assumption involves
527	 * -Using vma->mm->context{ASID,SASID}, as opposed to MMU reg.
528	 * -Fix the TLB paranoid debug code to not trigger false negatives.
529	 * -More importantly it makes this handler inconsistent with fast-path
530	 *  TLB Refill handler which always deals with "current"
531	 *
532	 * Lets see the use cases when current->mm != vma->mm and we land here
533	 *  1. execve->copy_strings()->__get_user_pages->handle_mm_fault
534	 *     Here VM wants to pre-install a TLB entry for user stack while
535	 *     current->mm still points to pre-execve mm (hence the condition).
536	 *     However the stack vaddr is soon relocated (randomization) and
537	 *     move_page_tables() tries to undo that TLB entry.
538	 *     Thus not creating TLB entry is not any worse.
539	 *
540	 *  2. ptrace(POKETEXT) causes a CoW - debugger(current) inserting a
541	 *     breakpoint in debugged task. Not creating a TLB now is not
542	 *     performance critical.
543	 *
544	 * Both the cases above are not good enough for code churn.
545	 */
546	if (current->active_mm != vma->vm_mm)
547		return;
548
549	local_irq_save(flags);
550
551	tlb_paranoid_check(asid_mm(vma->vm_mm, smp_processor_id()), vaddr);
552
553	vaddr &= PAGE_MASK;
554
555	/* update this PTE credentials */
556	pte_val(*ptep) |= (_PAGE_PRESENT | _PAGE_ACCESSED);
557
558	/* Create HW TLB(PD0,PD1) from PTE  */
559
560	/* ASID for this task */
561	asid_or_sasid = read_aux_reg(ARC_REG_PID) & 0xff;
562
563	pd0 = vaddr | asid_or_sasid | (pte_val(*ptep) & PTE_BITS_IN_PD0);
564
565	/*
566	 * ARC MMU provides fully orthogonal access bits for K/U mode,
567	 * however Linux only saves 1 set to save PTE real-estate
568	 * Here we convert 3 PTE bits into 6 MMU bits:
569	 * -Kernel only entries have Kr Kw Kx 0 0 0
570	 * -User entries have mirrored K and U bits
571	 */
572	rwx = pte_val(*ptep) & PTE_BITS_RWX;
573
574	if (pte_val(*ptep) & _PAGE_GLOBAL)
575		rwx <<= 3;		/* r w x => Kr Kw Kx 0 0 0 */
576	else
577		rwx |= (rwx << 3);	/* r w x => Kr Kw Kx Ur Uw Ux */
578
579	pd1 = rwx | (pte_val(*ptep) & PTE_BITS_NON_RWX_IN_PD1);
580
581	tlb_entry_insert(pd0, pd1);
582
583	local_irq_restore(flags);
584}
585
586/*
587 * Called at the end of pagefault, for a userspace mapped page
588 *  -pre-install the corresponding TLB entry into MMU
589 *  -Finalize the delayed D-cache flush of kernel mapping of page due to
590 *  	flush_dcache_page(), copy_user_page()
591 *
592 * Note that flush (when done) involves both WBACK - so physical page is
593 * in sync as well as INV - so any non-congruent aliases don't remain
594 */
595void update_mmu_cache(struct vm_area_struct *vma, unsigned long vaddr_unaligned,
596		      pte_t *ptep)
597{
598	unsigned long vaddr = vaddr_unaligned & PAGE_MASK;
599	phys_addr_t paddr = pte_val(*ptep) & PAGE_MASK;
600	struct page *page = pfn_to_page(pte_pfn(*ptep));
601
602	create_tlb(vma, vaddr, ptep);
603
604	if (page == ZERO_PAGE(0)) {
605		return;
606	}
607
608	/*
609	 * Exec page : Independent of aliasing/page-color considerations,
610	 *	       since icache doesn't snoop dcache on ARC, any dirty
611	 *	       K-mapping of a code page needs to be wback+inv so that
612	 *	       icache fetch by userspace sees code correctly.
613	 * !EXEC page: If K-mapping is NOT congruent to U-mapping, flush it
614	 *	       so userspace sees the right data.
615	 *  (Avoids the flush for Non-exec + congruent mapping case)
616	 */
617	if ((vma->vm_flags & VM_EXEC) ||
618	     addr_not_cache_congruent(paddr, vaddr)) {
619
620		int dirty = !test_and_set_bit(PG_dc_clean, &page->flags);
621		if (dirty) {
 
 
 
 
622			/* wback + inv dcache lines (K-mapping) */
623			__flush_dcache_page(paddr, paddr);
624
625			/* invalidate any existing icache lines (U-mapping) */
626			if (vma->vm_flags & VM_EXEC)
627				__inv_icache_page(paddr, vaddr);
628		}
629	}
630}
631
632#ifdef CONFIG_TRANSPARENT_HUGEPAGE
633
634/*
635 * MMUv4 in HS38x cores supports Super Pages which are basis for Linux THP
636 * support.
637 *
638 * Normal and Super pages can co-exist (ofcourse not overlap) in TLB with a
639 * new bit "SZ" in TLB page descriptor to distinguish between them.
640 * Super Page size is configurable in hardware (4K to 16M), but fixed once
641 * RTL builds.
642 *
643 * The exact THP size a Linx configuration will support is a function of:
644 *  - MMU page size (typical 8K, RTL fixed)
645 *  - software page walker address split between PGD:PTE:PFN (typical
646 *    11:8:13, but can be changed with 1 line)
647 * So for above default, THP size supported is 8K * (2^8) = 2M
648 *
649 * Default Page Walker is 2 levels, PGD:PTE:PFN, which in THP regime
650 * reduces to 1 level (as PTE is folded into PGD and canonically referred
651 * to as PMD).
652 * Thus THP PMD accessors are implemented in terms of PTE (just like sparc)
653 */
654
655void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
656				 pmd_t *pmd)
657{
658	pte_t pte = __pte(pmd_val(*pmd));
659	update_mmu_cache(vma, addr, &pte);
660}
661
662void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
663				pgtable_t pgtable)
664{
665	struct list_head *lh = (struct list_head *) pgtable;
666
667	assert_spin_locked(&mm->page_table_lock);
668
669	/* FIFO */
670	if (!pmd_huge_pte(mm, pmdp))
671		INIT_LIST_HEAD(lh);
672	else
673		list_add(lh, (struct list_head *) pmd_huge_pte(mm, pmdp));
674	pmd_huge_pte(mm, pmdp) = pgtable;
675}
676
677pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
678{
679	struct list_head *lh;
680	pgtable_t pgtable;
681
682	assert_spin_locked(&mm->page_table_lock);
683
684	pgtable = pmd_huge_pte(mm, pmdp);
685	lh = (struct list_head *) pgtable;
686	if (list_empty(lh))
687		pmd_huge_pte(mm, pmdp) = NULL;
688	else {
689		pmd_huge_pte(mm, pmdp) = (pgtable_t) lh->next;
690		list_del(lh);
691	}
692
693	pte_val(pgtable[0]) = 0;
694	pte_val(pgtable[1]) = 0;
695
696	return pgtable;
697}
698
699void local_flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
700			       unsigned long end)
701{
702	unsigned int cpu;
703	unsigned long flags;
704
705	local_irq_save(flags);
706
707	cpu = smp_processor_id();
708
709	if (likely(asid_mm(vma->vm_mm, cpu) != MM_CTXT_NO_ASID)) {
710		unsigned int asid = hw_pid(vma->vm_mm, cpu);
711
712		/* No need to loop here: this will always be for 1 Huge Page */
713		tlb_entry_erase(start | _PAGE_HW_SZ | asid);
714	}
715
716	local_irq_restore(flags);
717}
718
719#endif
720
721/* Read the Cache Build Confuration Registers, Decode them and save into
722 * the cpuinfo structure for later use.
723 * No Validation is done here, simply read/convert the BCRs
724 */
725void read_decode_mmu_bcr(void)
726{
727	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
728	unsigned int tmp;
729	struct bcr_mmu_1_2 {
730#ifdef CONFIG_CPU_BIG_ENDIAN
731		unsigned int ver:8, ways:4, sets:4, u_itlb:8, u_dtlb:8;
732#else
733		unsigned int u_dtlb:8, u_itlb:8, sets:4, ways:4, ver:8;
734#endif
735	} *mmu2;
736
737	struct bcr_mmu_3 {
738#ifdef CONFIG_CPU_BIG_ENDIAN
739	unsigned int ver:8, ways:4, sets:4, res:3, sasid:1, pg_sz:4,
740		     u_itlb:4, u_dtlb:4;
741#else
742	unsigned int u_dtlb:4, u_itlb:4, pg_sz:4, sasid:1, res:3, sets:4,
743		     ways:4, ver:8;
744#endif
745	} *mmu3;
746
747	struct bcr_mmu_4 {
748#ifdef CONFIG_CPU_BIG_ENDIAN
749	unsigned int ver:8, sasid:1, sz1:4, sz0:4, res:2, pae:1,
750		     n_ways:2, n_entry:2, n_super:2, u_itlb:3, u_dtlb:3;
751#else
752	/*           DTLB      ITLB      JES        JE         JA      */
753	unsigned int u_dtlb:3, u_itlb:3, n_super:2, n_entry:2, n_ways:2,
754		     pae:1, res:2, sz0:4, sz1:4, sasid:1, ver:8;
755#endif
756	} *mmu4;
757
758	tmp = read_aux_reg(ARC_REG_MMU_BCR);
759	mmu->ver = (tmp >> 24);
760
761	if (mmu->ver <= 2) {
762		mmu2 = (struct bcr_mmu_1_2 *)&tmp;
763		mmu->pg_sz_k = TO_KB(0x2000);
764		mmu->sets = 1 << mmu2->sets;
765		mmu->ways = 1 << mmu2->ways;
766		mmu->u_dtlb = mmu2->u_dtlb;
767		mmu->u_itlb = mmu2->u_itlb;
768	} else if (mmu->ver == 3) {
769		mmu3 = (struct bcr_mmu_3 *)&tmp;
770		mmu->pg_sz_k = 1 << (mmu3->pg_sz - 1);
771		mmu->sets = 1 << mmu3->sets;
772		mmu->ways = 1 << mmu3->ways;
773		mmu->u_dtlb = mmu3->u_dtlb;
774		mmu->u_itlb = mmu3->u_itlb;
775		mmu->sasid = mmu3->sasid;
776	} else {
777		mmu4 = (struct bcr_mmu_4 *)&tmp;
778		mmu->pg_sz_k = 1 << (mmu4->sz0 - 1);
779		mmu->s_pg_sz_m = 1 << (mmu4->sz1 - 11);
780		mmu->sets = 64 << mmu4->n_entry;
781		mmu->ways = mmu4->n_ways * 2;
782		mmu->u_dtlb = mmu4->u_dtlb * 4;
783		mmu->u_itlb = mmu4->u_itlb * 4;
784		mmu->sasid = mmu4->sasid;
785		mmu->pae = mmu4->pae;
786	}
787}
788
789char *arc_mmu_mumbojumbo(int cpu_id, char *buf, int len)
790{
791	int n = 0;
792	struct cpuinfo_arc_mmu *p_mmu = &cpuinfo_arc700[cpu_id].mmu;
793	char super_pg[64] = "";
794
795	if (p_mmu->s_pg_sz_m)
796		scnprintf(super_pg, 64, "%dM Super Page%s, ",
797			  p_mmu->s_pg_sz_m,
798			  IS_USED_CFG(CONFIG_TRANSPARENT_HUGEPAGE));
 
 
 
799
800	n += scnprintf(buf + n, len - n,
801		      "MMU [v%x]\t: %dk PAGE, %sJTLB %d (%dx%d), uDTLB %d, uITLB %d %s%s\n",
802		       p_mmu->ver, p_mmu->pg_sz_k, super_pg,
803		       p_mmu->sets * p_mmu->ways, p_mmu->sets, p_mmu->ways,
804		       p_mmu->u_dtlb, p_mmu->u_itlb,
805		       IS_AVAIL2(p_mmu->pae, "PAE40 ", CONFIG_ARC_HAS_PAE40));
806
807	return buf;
 
 
808}
809
810void arc_mmu_init(void)
811{
812	char str[256];
813	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
814
815	printk(arc_mmu_mumbojumbo(0, str, sizeof(str)));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
816
817	/* For efficiency sake, kernel is compile time built for a MMU ver
818	 * This must match the hardware it is running on.
819	 * Linux built for MMU V2, if run on MMU V1 will break down because V1
820	 *  hardware doesn't understand cmds such as WriteNI, or IVUTLB
821	 * On the other hand, Linux built for V1 if run on MMU V2 will do
822	 *   un-needed workarounds to prevent memcpy thrashing.
823	 * Similarly MMU V3 has new features which won't work on older MMU
824	 */
825	if (mmu->ver != CONFIG_ARC_MMU_VER) {
826		panic("MMU ver %d doesn't match kernel built for %d...\n",
827		      mmu->ver, CONFIG_ARC_MMU_VER);
828	}
829
830	if (mmu->pg_sz_k != TO_KB(PAGE_SIZE))
831		panic("MMU pg size != PAGE_SIZE (%luk)\n", TO_KB(PAGE_SIZE));
832
833	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
834	    mmu->s_pg_sz_m != TO_MB(HPAGE_PMD_SIZE))
835		panic("MMU Super pg size != Linux HPAGE_PMD_SIZE (%luM)\n",
836		      (unsigned long)TO_MB(HPAGE_PMD_SIZE));
837
838	if (IS_ENABLED(CONFIG_ARC_HAS_PAE40) && !mmu->pae)
839		panic("Hardware doesn't support PAE40\n");
840
841	/* Enable the MMU */
842	write_aux_reg(ARC_REG_PID, MMU_ENABLE);
843
844	/* In smp we use this reg for interrupt 1 scratch */
845#ifndef CONFIG_SMP
846	/* swapper_pg_dir is the pgd for the kernel, used by vmalloc */
847	write_aux_reg(ARC_REG_SCRATCH_DATA0, swapper_pg_dir);
848#endif
849}
850
851/*
852 * TLB Programmer's Model uses Linear Indexes: 0 to {255, 511} for 128 x {2,4}
853 * The mapping is Column-first.
854 *		---------------------	-----------
855 *		|way0|way1|way2|way3|	|way0|way1|
856 *		---------------------	-----------
857 * [set0]	|  0 |  1 |  2 |  3 |	|  0 |  1 |
858 * [set1]	|  4 |  5 |  6 |  7 |	|  2 |  3 |
859 *		~		    ~	~	  ~
860 * [set127]	| 508| 509| 510| 511|	| 254| 255|
861 *		---------------------	-----------
862 * For normal operations we don't(must not) care how above works since
863 * MMU cmd getIndex(vaddr) abstracts that out.
864 * However for walking WAYS of a SET, we need to know this
865 */
866#define SET_WAY_TO_IDX(mmu, set, way)  ((set) * mmu->ways + (way))
867
868/* Handling of Duplicate PD (TLB entry) in MMU.
869 * -Could be due to buggy customer tapeouts or obscure kernel bugs
870 * -MMU complaints not at the time of duplicate PD installation, but at the
871 *      time of lookup matching multiple ways.
872 * -Ideally these should never happen - but if they do - workaround by deleting
873 *      the duplicate one.
874 * -Knob to be verbose abt it.(TODO: hook them up to debugfs)
875 */
876volatile int dup_pd_silent; /* Be slient abt it or complain (default) */
877
878void do_tlb_overlap_fault(unsigned long cause, unsigned long address,
879			  struct pt_regs *regs)
880{
881	struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
882	unsigned int pd0[mmu->ways];
883	unsigned long flags;
884	int set;
 
 
 
885
886	local_irq_save(flags);
887
888	/* re-enable the MMU */
889	write_aux_reg(ARC_REG_PID, MMU_ENABLE | read_aux_reg(ARC_REG_PID));
890
891	/* loop thru all sets of TLB */
892	for (set = 0; set < mmu->sets; set++) {
893
894		int is_valid, way;
 
895
896		/* read out all the ways of current set */
897		for (way = 0, is_valid = 0; way < mmu->ways; way++) {
898			write_aux_reg(ARC_REG_TLBINDEX,
899					  SET_WAY_TO_IDX(mmu, set, way));
900			write_aux_reg(ARC_REG_TLBCOMMAND, TLBRead);
901			pd0[way] = read_aux_reg(ARC_REG_TLBPD0);
902			is_valid |= pd0[way] & _PAGE_PRESENT;
903			pd0[way] &= PAGE_MASK;
904		}
905
906		/* If all the WAYS in SET are empty, skip to next SET */
907		if (!is_valid)
908			continue;
909
910		/* Scan the set for duplicate ways: needs a nested loop */
911		for (way = 0; way < mmu->ways - 1; way++) {
912
913			int n;
914
915			if (!pd0[way])
916				continue;
917
918			for (n = way + 1; n < mmu->ways; n++) {
919				if (pd0[way] != pd0[n])
920					continue;
921
922				if (!dup_pd_silent)
923					pr_info("Dup TLB PD0 %08x @ set %d ways %d,%d\n",
924						pd0[way], set, way, n);
925
926				/*
927				 * clear entry @way and not @n.
928				 * This is critical to our optimised loop
929				 */
930				pd0[way] = 0;
931				write_aux_reg(ARC_REG_TLBINDEX,
932						SET_WAY_TO_IDX(mmu, set, way));
933				__tlb_entry_erase();
934			}
935		}
936	}
937
938	local_irq_restore(flags);
939}
940
941/***********************************************************************
942 * Diagnostic Routines
943 *  -Called from Low Level TLB Hanlders if things don;t look good
944 **********************************************************************/
945
946#ifdef CONFIG_ARC_DBG_TLB_PARANOIA
947
948/*
949 * Low Level ASM TLB handler calls this if it finds that HW and SW ASIDS
950 * don't match
951 */
952void print_asid_mismatch(int mm_asid, int mmu_asid, int is_fast_path)
953{
954	pr_emerg("ASID Mismatch in %s Path Handler: sw-pid=0x%x hw-pid=0x%x\n",
955	       is_fast_path ? "Fast" : "Slow", mm_asid, mmu_asid);
956
957	__asm__ __volatile__("flag 1");
958}
959
960void tlb_paranoid_check(unsigned int mm_asid, unsigned long addr)
961{
962	unsigned int mmu_asid;
963
964	mmu_asid = read_aux_reg(ARC_REG_PID) & 0xff;
965
966	/*
967	 * At the time of a TLB miss/installation
968	 *   - HW version needs to match SW version
969	 *   - SW needs to have a valid ASID
970	 */
971	if (addr < 0x70000000 &&
972	    ((mm_asid == MM_CTXT_NO_ASID) ||
973	      (mmu_asid != (mm_asid & MM_CTXT_ASID_MASK))))
974		print_asid_mismatch(mm_asid, mmu_asid, 0);
975}
976#endif