2 * TLB Management (flush/create/diagnostics) for ARC700
4 * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com)
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.
11 * -Reintroduce duplicate PD fixup - some customer chips still have the issue
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)
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
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
32 * Vineetg: Sept 10th 2008
33 * -Changes related to MMU v2 (Rel 4.8)
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
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.
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.
51 * Rahul Trivedi : Codito Technologies 2004
54 #include <linux/module.h>
55 #include <asm/arcregs.h>
56 #include <asm/setup.h>
57 #include <asm/mmu_context.h>
60 /* Need for ARC MMU v2
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
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
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.
78 * - Existing TLB commands work as before
79 * - New command (TLBWriteNI) for TLB write without clearing uTLBs
80 * - New command (TLBIVUTLB) to invalidate uTLBs.
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
89 * During a miss handler, the new "TLBWriteNI" command is used to load
90 * entries without clearing the uTLBs.
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.
101 /* A copy of the ASID from the PID reg is kept in asid_cache */
102 int asid_cache = FIRST_ASID;
104 /* ASID to mm struct mapping. We have one extra entry corresponding to
105 * NO_ASID to save us a compare when clearing the mm entry for old asid
106 * see get_new_mmu_context (asm-arc/mmu_context.h)
108 struct mm_struct *asid_mm_map[NUM_ASID + 1];
111 * Utility Routine to erase a J-TLB entry
112 * The procedure is to look it up in the MMU. If found, ERASE it by
113 * issuing a TlbWrite CMD with PD0 = PD1 = 0
116 static void __tlb_entry_erase(void)
118 write_aux_reg(ARC_REG_TLBPD1, 0);
119 write_aux_reg(ARC_REG_TLBPD0, 0);
120 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
123 static void tlb_entry_erase(unsigned int vaddr_n_asid)
127 /* Locate the TLB entry for this vaddr + ASID */
128 write_aux_reg(ARC_REG_TLBPD0, vaddr_n_asid);
129 write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe);
130 idx = read_aux_reg(ARC_REG_TLBINDEX);
132 /* No error means entry found, zero it out */
133 if (likely(!(idx & TLB_LKUP_ERR))) {
135 } else { /* Some sort of Error */
137 /* Duplicate entry error */
139 /* TODO we need to handle this case too */
140 pr_emerg("unhandled Duplicate flush for %x\n",
143 /* else entry not found so nothing to do */
147 /****************************************************************************
148 * ARC700 MMU caches recently used J-TLB entries (RAM) as uTLBs (FLOPs)
150 * New IVUTLB cmd in MMU v2 explictly invalidates the uTLB
152 * utlb_invalidate ( )
153 * -For v2 MMU calls Flush uTLB Cmd
154 * -For v1 MMU does nothing (except for Metal Fix v1 MMU)
155 * This is because in v1 TLBWrite itself invalidate uTLBs
156 ***************************************************************************/
158 static void utlb_invalidate(void)
160 #if (CONFIG_ARC_MMU_VER >= 2)
162 #if (CONFIG_ARC_MMU_VER < 3)
163 /* MMU v2 introduced the uTLB Flush command.
164 * There was however an obscure hardware bug, where uTLB flush would
165 * fail when a prior probe for J-TLB (both totally unrelated) would
166 * return lkup err - because the entry didnt exist in MMU.
167 * The Workround was to set Index reg with some valid value, prior to
168 * flush. This was fixed in MMU v3 hence not needed any more
172 /* make sure INDEX Reg is valid */
173 idx = read_aux_reg(ARC_REG_TLBINDEX);
175 /* If not write some dummy val */
176 if (unlikely(idx & TLB_LKUP_ERR))
177 write_aux_reg(ARC_REG_TLBINDEX, 0xa);
180 write_aux_reg(ARC_REG_TLBCOMMAND, TLBIVUTLB);
186 * Un-conditionally (without lookup) erase the entire MMU contents
189 noinline void local_flush_tlb_all(void)
193 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
195 local_irq_save(flags);
197 /* Load PD0 and PD1 with template for a Blank Entry */
198 write_aux_reg(ARC_REG_TLBPD1, 0);
199 write_aux_reg(ARC_REG_TLBPD0, 0);
201 for (entry = 0; entry < mmu->num_tlb; entry++) {
202 /* write this entry to the TLB */
203 write_aux_reg(ARC_REG_TLBINDEX, entry);
204 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
209 local_irq_restore(flags);
213 * Flush the entrie MM for userland. The fastest way is to move to Next ASID
215 noinline void local_flush_tlb_mm(struct mm_struct *mm)
218 * Small optimisation courtesy IA64
219 * flush_mm called during fork,exit,munmap etc, multiple times as well.
220 * Only for fork( ) do we need to move parent to a new MMU ctxt,
221 * all other cases are NOPs, hence this check.
223 if (atomic_read(&mm->mm_users) == 0)
227 * Workaround for Android weirdism:
228 * A binder VMA could end up in a task such that vma->mm != tsk->mm
229 * old code would cause h/w - s/w ASID to get out of sync
231 if (current->mm != mm)
234 get_new_mmu_context(mm);
238 * Flush a Range of TLB entries for userland.
239 * @start is inclusive, while @end is exclusive
240 * Difference between this and Kernel Range Flush is
241 * -Here the fastest way (if range is too large) is to move to next ASID
242 * without doing any explicit Shootdown
243 * -In case of kernel Flush, entry has to be shot down explictly
245 void local_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
251 /* If range @start to @end is more than 32 TLB entries deep,
252 * its better to move to a new ASID rather than searching for
253 * individual entries and then shooting them down
255 * The calc above is rough, doesn't account for unaligned parts,
256 * since this is heuristics based anyways
258 if (unlikely((end - start) >= PAGE_SIZE * 32)) {
259 local_flush_tlb_mm(vma->vm_mm);
264 * @start moved to page start: this alone suffices for checking
265 * loop end condition below, w/o need for aligning @end to end
266 * e.g. 2000 to 4001 will anyhow loop twice
270 local_irq_save(flags);
271 asid = vma->vm_mm->context.asid;
273 if (asid != NO_ASID) {
274 while (start < end) {
275 tlb_entry_erase(start | (asid & 0xff));
282 local_irq_restore(flags);
285 /* Flush the kernel TLB entries - vmalloc/modules (Global from MMU perspective)
286 * @start, @end interpreted as kvaddr
287 * Interestingly, shared TLB entries can also be flushed using just
288 * @start,@end alone (interpreted as user vaddr), although technically SASID
289 * is also needed. However our smart TLbProbe lookup takes care of that.
291 void local_flush_tlb_kernel_range(unsigned long start, unsigned long end)
295 /* exactly same as above, except for TLB entry not taking ASID */
297 if (unlikely((end - start) >= PAGE_SIZE * 32)) {
298 local_flush_tlb_all();
304 local_irq_save(flags);
305 while (start < end) {
306 tlb_entry_erase(start);
312 local_irq_restore(flags);
316 * Delete TLB entry in MMU for a given page (??? address)
317 * NOTE One TLB entry contains translation for single PAGE
320 void local_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
324 /* Note that it is critical that interrupts are DISABLED between
325 * checking the ASID and using it flush the TLB entry
327 local_irq_save(flags);
329 if (vma->vm_mm->context.asid != NO_ASID) {
330 tlb_entry_erase((page & PAGE_MASK) |
331 (vma->vm_mm->context.asid & 0xff));
335 local_irq_restore(flags);
339 * Routine to create a TLB entry
341 void create_tlb(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
344 unsigned int idx, asid_or_sasid;
345 unsigned long pd0_flags;
348 * create_tlb() assumes that current->mm == vma->mm, since
349 * -it ASID for TLB entry is fetched from MMU ASID reg (valid for curr)
350 * -completes the lazy write to SASID reg (again valid for curr tsk)
352 * Removing the assumption involves
353 * -Using vma->mm->context{ASID,SASID}, as opposed to MMU reg.
354 * -Fix the TLB paranoid debug code to not trigger false negatives.
355 * -More importantly it makes this handler inconsistent with fast-path
356 * TLB Refill handler which always deals with "current"
358 * Lets see the use cases when current->mm != vma->mm and we land here
359 * 1. execve->copy_strings()->__get_user_pages->handle_mm_fault
360 * Here VM wants to pre-install a TLB entry for user stack while
361 * current->mm still points to pre-execve mm (hence the condition).
362 * However the stack vaddr is soon relocated (randomization) and
363 * move_page_tables() tries to undo that TLB entry.
364 * Thus not creating TLB entry is not any worse.
366 * 2. ptrace(POKETEXT) causes a CoW - debugger(current) inserting a
367 * breakpoint in debugged task. Not creating a TLB now is not
368 * performance critical.
370 * Both the cases above are not good enough for code churn.
372 if (current->active_mm != vma->vm_mm)
375 local_irq_save(flags);
377 tlb_paranoid_check(vma->vm_mm->context.asid, address);
379 address &= PAGE_MASK;
381 /* update this PTE credentials */
382 pte_val(*ptep) |= (_PAGE_PRESENT | _PAGE_ACCESSED);
384 /* Create HW TLB entry Flags (in PD0) from PTE Flags */
385 #if (CONFIG_ARC_MMU_VER <= 2)
386 pd0_flags = ((pte_val(*ptep) & PTE_BITS_IN_PD0) >> 1);
388 pd0_flags = ((pte_val(*ptep) & PTE_BITS_IN_PD0));
391 /* ASID for this task */
392 asid_or_sasid = read_aux_reg(ARC_REG_PID) & 0xff;
394 write_aux_reg(ARC_REG_TLBPD0, address | pd0_flags | asid_or_sasid);
396 /* Load remaining info in PD1 (Page Frame Addr and Kx/Kw/Kr Flags) */
397 write_aux_reg(ARC_REG_TLBPD1, (pte_val(*ptep) & PTE_BITS_IN_PD1));
399 /* First verify if entry for this vaddr+ASID already exists */
400 write_aux_reg(ARC_REG_TLBCOMMAND, TLBProbe);
401 idx = read_aux_reg(ARC_REG_TLBINDEX);
404 * If Not already present get a free slot from MMU.
405 * Otherwise, Probe would have located the entry and set INDEX Reg
406 * with existing location. This will cause Write CMD to over-write
407 * existing entry with new PD0 and PD1
409 if (likely(idx & TLB_LKUP_ERR))
410 write_aux_reg(ARC_REG_TLBCOMMAND, TLBGetIndex);
413 * Commit the Entry to MMU
414 * It doesnt sound safe to use the TLBWriteNI cmd here
415 * which doesn't flush uTLBs. I'd rather be safe than sorry.
417 write_aux_reg(ARC_REG_TLBCOMMAND, TLBWrite);
419 local_irq_restore(flags);
423 * Called at the end of pagefault, for a userspace mapped page
424 * -pre-install the corresponding TLB entry into MMU
425 * -Finalize the delayed D-cache flush of kernel mapping of page due to
426 * flush_dcache_page(), copy_user_page()
428 * Note that flush (when done) involves both WBACK - so physical page is
429 * in sync as well as INV - so any non-congruent aliases don't remain
431 void update_mmu_cache(struct vm_area_struct *vma, unsigned long vaddr_unaligned,
434 unsigned long vaddr = vaddr_unaligned & PAGE_MASK;
435 unsigned long paddr = pte_val(*ptep) & PAGE_MASK;
436 struct page *page = pfn_to_page(pte_pfn(*ptep));
438 create_tlb(vma, vaddr, ptep);
440 if (page == ZERO_PAGE(0)) {
445 * Exec page : Independent of aliasing/page-color considerations,
446 * since icache doesn't snoop dcache on ARC, any dirty
447 * K-mapping of a code page needs to be wback+inv so that
448 * icache fetch by userspace sees code correctly.
449 * !EXEC page: If K-mapping is NOT congruent to U-mapping, flush it
450 * so userspace sees the right data.
451 * (Avoids the flush for Non-exec + congruent mapping case)
453 if ((vma->vm_flags & VM_EXEC) ||
454 addr_not_cache_congruent(paddr, vaddr)) {
456 int dirty = !test_and_set_bit(PG_dc_clean, &page->flags);
458 /* wback + inv dcache lines */
459 __flush_dcache_page(paddr, paddr);
461 /* invalidate any existing icache lines */
462 if (vma->vm_flags & VM_EXEC)
463 __inv_icache_page(paddr, vaddr);
468 /* Read the Cache Build Confuration Registers, Decode them and save into
469 * the cpuinfo structure for later use.
470 * No Validation is done here, simply read/convert the BCRs
472 void read_decode_mmu_bcr(void)
474 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
477 #ifdef CONFIG_CPU_BIG_ENDIAN
478 unsigned int ver:8, ways:4, sets:4, u_itlb:8, u_dtlb:8;
480 unsigned int u_dtlb:8, u_itlb:8, sets:4, ways:4, ver:8;
485 #ifdef CONFIG_CPU_BIG_ENDIAN
486 unsigned int ver:8, ways:4, sets:4, osm:1, reserv:3, pg_sz:4,
489 unsigned int u_dtlb:4, u_itlb:4, pg_sz:4, reserv:3, osm:1, sets:4,
494 tmp = read_aux_reg(ARC_REG_MMU_BCR);
495 mmu->ver = (tmp >> 24);
498 mmu2 = (struct bcr_mmu_1_2 *)&tmp;
499 mmu->pg_sz = PAGE_SIZE;
500 mmu->sets = 1 << mmu2->sets;
501 mmu->ways = 1 << mmu2->ways;
502 mmu->u_dtlb = mmu2->u_dtlb;
503 mmu->u_itlb = mmu2->u_itlb;
505 mmu3 = (struct bcr_mmu_3 *)&tmp;
506 mmu->pg_sz = 512 << mmu3->pg_sz;
507 mmu->sets = 1 << mmu3->sets;
508 mmu->ways = 1 << mmu3->ways;
509 mmu->u_dtlb = mmu3->u_dtlb;
510 mmu->u_itlb = mmu3->u_itlb;
513 mmu->num_tlb = mmu->sets * mmu->ways;
516 char *arc_mmu_mumbojumbo(int cpu_id, char *buf, int len)
519 struct cpuinfo_arc_mmu *p_mmu = &cpuinfo_arc700[cpu_id].mmu;
521 n += scnprintf(buf + n, len - n, "ARC700 MMU [v%x]\t: %dk PAGE, ",
522 p_mmu->ver, TO_KB(p_mmu->pg_sz));
524 n += scnprintf(buf + n, len - n,
525 "J-TLB %d (%dx%d), uDTLB %d, uITLB %d, %s\n",
526 p_mmu->num_tlb, p_mmu->sets, p_mmu->ways,
527 p_mmu->u_dtlb, p_mmu->u_itlb,
528 IS_ENABLED(CONFIG_ARC_MMU_SASID) ? "SASID" : "");
533 void arc_mmu_init(void)
536 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
538 printk(arc_mmu_mumbojumbo(0, str, sizeof(str)));
540 /* For efficiency sake, kernel is compile time built for a MMU ver
541 * This must match the hardware it is running on.
542 * Linux built for MMU V2, if run on MMU V1 will break down because V1
543 * hardware doesn't understand cmds such as WriteNI, or IVUTLB
544 * On the other hand, Linux built for V1 if run on MMU V2 will do
545 * un-needed workarounds to prevent memcpy thrashing.
546 * Similarly MMU V3 has new features which won't work on older MMU
548 if (mmu->ver != CONFIG_ARC_MMU_VER) {
549 panic("MMU ver %d doesn't match kernel built for %d...\n",
550 mmu->ver, CONFIG_ARC_MMU_VER);
553 if (mmu->pg_sz != PAGE_SIZE)
554 panic("MMU pg size != PAGE_SIZE (%luk)\n", TO_KB(PAGE_SIZE));
557 * ASID mgmt data structures are compile time init
558 * asid_cache = FIRST_ASID and asid_mm_map[] all zeroes
561 local_flush_tlb_all();
564 write_aux_reg(ARC_REG_PID, MMU_ENABLE);
566 /* In smp we use this reg for interrupt 1 scratch */
568 /* swapper_pg_dir is the pgd for the kernel, used by vmalloc */
569 write_aux_reg(ARC_REG_SCRATCH_DATA0, swapper_pg_dir);
574 * TLB Programmer's Model uses Linear Indexes: 0 to {255, 511} for 128 x {2,4}
575 * The mapping is Column-first.
576 * --------------------- -----------
577 * |way0|way1|way2|way3| |way0|way1|
578 * --------------------- -----------
579 * [set0] | 0 | 1 | 2 | 3 | | 0 | 1 |
580 * [set1] | 4 | 5 | 6 | 7 | | 2 | 3 |
582 * [set127] | 508| 509| 510| 511| | 254| 255|
583 * --------------------- -----------
584 * For normal operations we don't(must not) care how above works since
585 * MMU cmd getIndex(vaddr) abstracts that out.
586 * However for walking WAYS of a SET, we need to know this
588 #define SET_WAY_TO_IDX(mmu, set, way) ((set) * mmu->ways + (way))
590 /* Handling of Duplicate PD (TLB entry) in MMU.
591 * -Could be due to buggy customer tapeouts or obscure kernel bugs
592 * -MMU complaints not at the time of duplicate PD installation, but at the
593 * time of lookup matching multiple ways.
594 * -Ideally these should never happen - but if they do - workaround by deleting
596 * -Knob to be verbose abt it.(TODO: hook them up to debugfs)
598 volatile int dup_pd_verbose = 1;/* Be slient abt it or complain (default) */
600 void do_tlb_overlap_fault(unsigned long cause, unsigned long address,
601 struct pt_regs *regs)
604 unsigned int pd0[4], pd1[4]; /* assume max 4 ways */
605 unsigned long flags, is_valid;
606 struct cpuinfo_arc_mmu *mmu = &cpuinfo_arc700[smp_processor_id()].mmu;
608 local_irq_save(flags);
610 /* re-enable the MMU */
611 write_aux_reg(ARC_REG_PID, MMU_ENABLE | read_aux_reg(ARC_REG_PID));
613 /* loop thru all sets of TLB */
614 for (set = 0; set < mmu->sets; set++) {
616 /* read out all the ways of current set */
617 for (way = 0, is_valid = 0; way < mmu->ways; way++) {
618 write_aux_reg(ARC_REG_TLBINDEX,
619 SET_WAY_TO_IDX(mmu, set, way));
620 write_aux_reg(ARC_REG_TLBCOMMAND, TLBRead);
621 pd0[way] = read_aux_reg(ARC_REG_TLBPD0);
622 pd1[way] = read_aux_reg(ARC_REG_TLBPD1);
623 is_valid |= pd0[way] & _PAGE_PRESENT;
626 /* If all the WAYS in SET are empty, skip to next SET */
630 /* Scan the set for duplicate ways: needs a nested loop */
631 for (way = 0; way < mmu->ways; way++) {
635 for (n = way + 1; n < mmu->ways; n++) {
636 if ((pd0[way] & PAGE_MASK) ==
637 (pd0[n] & PAGE_MASK)) {
639 if (dup_pd_verbose) {
640 pr_info("Duplicate PD's @"
643 pr_info("TLBPD0[%u]: %08x\n",
648 * clear entry @way and not @n. This is
649 * critical to our optimised loop
651 pd0[way] = pd1[way] = 0;
652 write_aux_reg(ARC_REG_TLBINDEX,
653 SET_WAY_TO_IDX(mmu, set, way));
660 local_irq_restore(flags);
663 /***********************************************************************
664 * Diagnostic Routines
665 * -Called from Low Level TLB Hanlders if things don;t look good
666 **********************************************************************/
668 #ifdef CONFIG_ARC_DBG_TLB_PARANOIA
671 * Low Level ASM TLB handler calls this if it finds that HW and SW ASIDS
674 void print_asid_mismatch(int is_fast_path)
677 pid_sw = current->active_mm->context.asid;
678 pid_hw = read_aux_reg(ARC_REG_PID) & 0xff;
680 pr_emerg("ASID Mismatch in %s Path Handler: sw-pid=0x%x hw-pid=0x%x\n",
681 is_fast_path ? "Fast" : "Slow", pid_sw, pid_hw);
683 __asm__ __volatile__("flag 1");
686 void tlb_paranoid_check(unsigned int pid_sw, unsigned long addr)
690 pid_hw = read_aux_reg(ARC_REG_PID) & 0xff;
692 if (addr < 0x70000000 && ((pid_hw != pid_sw) || (pid_sw == NO_ASID)))
693 print_asid_mismatch(0);