3 * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
4 * Rewritten by Cort Dougan (cort@cs.nmt.edu) for PReP
5 * Copyright (C) 1996 Cort Dougan <cort@cs.nmt.edu>
6 * Adapted for Power Macintosh by Paul Mackerras.
7 * Low-level exception handlers and MMU support
8 * rewritten by Paul Mackerras.
9 * Copyright (C) 1996 Paul Mackerras.
11 * This file contains low-level assembler routines for managing
12 * the PowerPC MMU hash table. (PPC 8xx processors don't use a
13 * hash table, so this file is not used on them.)
15 * This program is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU General Public License
17 * as published by the Free Software Foundation; either version
18 * 2 of the License, or (at your option) any later version.
24 #include <asm/pgtable.h>
25 #include <asm/cputable.h>
26 #include <asm/ppc_asm.h>
27 #include <asm/thread_info.h>
28 #include <asm/asm-offsets.h>
29 #include <asm/export.h>
37 EXPORT_SYMBOL(mmu_hash_lock)
38 #endif /* CONFIG_SMP */
41 * Load a PTE into the hash table, if possible.
42 * The address is in r4, and r3 contains an access flag:
43 * _PAGE_RW (0x400) if a write.
44 * r9 contains the SRR1 value, from which we use the MSR_PR bit.
45 * SPRG_THREAD contains the physical address of the current task's thread.
47 * Returns to the caller if the access is illegal or there is no
48 * mapping for the address. Otherwise it places an appropriate PTE
49 * in the hash table and returns from the exception.
50 * Uses r0, r3 - r8, r10, ctr, lr.
54 tophys(r7,0) /* gets -KERNELBASE into r7 */
56 addis r8,r7,mmu_hash_lock@h
57 ori r8,r8,mmu_hash_lock@l
70 /* Get PTE (linux-style) and check access */
71 lis r0,KERNELBASE@h /* check if kernel address */
73 mfspr r8,SPRN_SPRG_THREAD /* current task's THREAD (phys) */
74 ori r3,r3,_PAGE_USER|_PAGE_PRESENT /* test low addresses as user */
75 lwz r5,PGDIR(r8) /* virt page-table root */
76 blt+ 112f /* assume user more likely */
77 lis r5,swapper_pg_dir@ha /* if kernel address, use */
78 addi r5,r5,swapper_pg_dir@l /* kernel page table */
79 rlwimi r3,r9,32-12,29,29 /* MSR_PR -> _PAGE_USER */
80 112: add r5,r5,r7 /* convert to phys addr */
81 #ifndef CONFIG_PTE_64BIT
82 rlwimi r5,r4,12,20,29 /* insert top 10 bits of address */
83 lwz r8,0(r5) /* get pmd entry */
84 rlwinm. r8,r8,0,0,19 /* extract address of pte page */
86 rlwinm r8,r4,13,19,29 /* Compute pgdir/pmd offset */
87 lwzx r8,r8,r5 /* Get L1 entry */
88 rlwinm. r8,r8,0,0,20 /* extract pt base address */
91 beq- hash_page_out /* return if no mapping */
93 /* XXX it seems like the 601 will give a machine fault on the
94 rfi if its alignment is wrong (bottom 4 bits of address are
95 8 or 0xc) and we have had a not-taken conditional branch
96 to the address following the rfi. */
99 #ifndef CONFIG_PTE_64BIT
100 rlwimi r8,r4,22,20,29 /* insert next 10 bits of address */
102 rlwimi r8,r4,23,20,28 /* compute pte address */
104 rlwinm r0,r3,32-3,24,24 /* _PAGE_RW access -> _PAGE_DIRTY */
105 ori r0,r0,_PAGE_ACCESSED|_PAGE_HASHPTE
108 * Update the linux PTE atomically. We do the lwarx up-front
109 * because almost always, there won't be a permission violation
110 * and there won't already be an HPTE, and thus we will have
111 * to update the PTE to set _PAGE_HASHPTE. -- paulus.
113 * If PTE_64BIT is set, the low word is the flags word; use that
114 * word for locking since it contains all the interesting bits.
116 #if (PTE_FLAGS_OFFSET != 0)
117 addi r8,r8,PTE_FLAGS_OFFSET
120 lwarx r6,0,r8 /* get linux-style pte, flag word */
121 andc. r5,r3,r6 /* check access & ~permission */
123 bne- hash_page_out /* return if access not permitted */
127 or r5,r0,r6 /* set accessed/dirty bits */
128 #ifdef CONFIG_PTE_64BIT
130 subf r10,r6,r8 /* create false data dependency */
131 subi r10,r10,PTE_FLAGS_OFFSET
132 lwzx r10,r6,r10 /* Get upper PTE word */
134 lwz r10,-PTE_FLAGS_OFFSET(r8)
135 #endif /* CONFIG_SMP */
136 #endif /* CONFIG_PTE_64BIT */
137 stwcx. r5,0,r8 /* attempt to update PTE */
138 bne- retry /* retry if someone got there first */
140 mfsrin r3,r4 /* get segment reg for segment */
143 bl create_hpte /* add the hash table entry */
147 addis r8,r7,mmu_hash_lock@ha
149 stw r0,mmu_hash_lock@l(r8)
152 /* Return from the exception */
158 b fast_exception_return
163 addis r8,r7,mmu_hash_lock@ha
165 stw r0,mmu_hash_lock@l(r8)
167 #endif /* CONFIG_SMP */
170 * Add an entry for a particular page to the hash table.
172 * add_hash_page(unsigned context, unsigned long va, unsigned long pmdval)
174 * We assume any necessary modifications to the pte (e.g. setting
175 * the accessed bit) have already been done and that there is actually
176 * a hash table in use (i.e. we're not on a 603).
178 _GLOBAL(add_hash_page)
182 /* Convert context and va to VSID */
183 mulli r3,r3,897*16 /* multiply context by context skew */
184 rlwinm r0,r4,4,28,31 /* get ESID (top 4 bits of va) */
185 mulli r0,r0,0x111 /* multiply by ESID skew */
186 add r3,r3,r0 /* note create_hpte trims to 24 bits */
189 CURRENT_THREAD_INFO(r8, r1) /* use cpu number to make tag */
190 lwz r8,TI_CPU(r8) /* to go in mmu_hash_lock */
192 #endif /* CONFIG_SMP */
195 * We disable interrupts here, even on UP, because we don't
196 * want to race with hash_page, and because we want the
197 * _PAGE_HASHPTE bit to be a reliable indication of whether
198 * the HPTE exists (or at least whether one did once).
199 * We also turn off the MMU for data accesses so that we
200 * we can't take a hash table miss (assuming the code is
201 * covered by a BAT). -- paulus
205 rlwinm r0,r9,0,17,15 /* clear bit 16 (MSR_EE) */
206 rlwinm r0,r0,0,28,26 /* clear MSR_DR */
214 addis r6,r7,mmu_hash_lock@ha
215 addi r6,r6,mmu_hash_lock@l
216 10: lwarx r0,0,r6 /* take the mmu_hash_lock */
229 * Fetch the linux pte and test and set _PAGE_HASHPTE atomically.
230 * If _PAGE_HASHPTE was already set, we don't replace the existing
231 * HPTE, so we just unlock and return.
234 #ifndef CONFIG_PTE_64BIT
235 rlwimi r8,r4,22,20,29
237 rlwimi r8,r4,23,20,28
238 addi r8,r8,PTE_FLAGS_OFFSET
241 andi. r0,r6,_PAGE_HASHPTE
242 bne 9f /* if HASHPTE already set, done */
243 #ifdef CONFIG_PTE_64BIT
245 subf r10,r6,r8 /* create false data dependency */
246 subi r10,r10,PTE_FLAGS_OFFSET
247 lwzx r10,r6,r10 /* Get upper PTE word */
249 lwz r10,-PTE_FLAGS_OFFSET(r8)
250 #endif /* CONFIG_SMP */
251 #endif /* CONFIG_PTE_64BIT */
252 ori r5,r6,_PAGE_HASHPTE
260 addis r6,r7,mmu_hash_lock@ha
261 addi r6,r6,mmu_hash_lock@l
264 stw r0,0(r6) /* clear mmu_hash_lock */
267 /* reenable interrupts and DR */
277 * This routine adds a hardware PTE to the hash table.
278 * It is designed to be called with the MMU either on or off.
279 * r3 contains the VSID, r4 contains the virtual address,
280 * r5 contains the linux PTE, r6 contains the old value of the
281 * linux PTE (before setting _PAGE_HASHPTE) and r7 contains the
282 * offset to be added to addresses (0 if the MMU is on,
283 * -KERNELBASE if it is off). r10 contains the upper half of
284 * the PTE if CONFIG_PTE_64BIT.
285 * On SMP, the caller should have the mmu_hash_lock held.
286 * We assume that the caller has (or will) set the _PAGE_HASHPTE
287 * bit in the linux PTE in memory. The value passed in r6 should
288 * be the old linux PTE value; if it doesn't have _PAGE_HASHPTE set
289 * this routine will skip the search for an existing HPTE.
290 * This procedure modifies r0, r3 - r6, r8, cr0.
293 * For speed, 4 of the instructions get patched once the size and
294 * physical address of the hash table are known. These definitions
295 * of Hash_base and Hash_bits below are just an example.
297 Hash_base = 0xc0180000
298 Hash_bits = 12 /* e.g. 256kB hash table */
299 Hash_msk = (((1 << Hash_bits) - 1) * 64)
301 /* defines for the PTE format for 32-bit PPCs */
304 #define LG_PTEG_SIZE 6
310 #define PTE_V 0x80000000
311 #define TST_V(r) rlwinm. r,r,0,0,0
312 #define SET_V(r) oris r,r,PTE_V@h
313 #define CLR_V(r,t) rlwinm r,r,0,1,31
315 #define HASH_LEFT 31-(LG_PTEG_SIZE+Hash_bits-1)
316 #define HASH_RIGHT 31-LG_PTEG_SIZE
319 /* Convert linux-style PTE (r5) to low word of PPC-style PTE (r8) */
320 rlwinm r8,r5,32-10,31,31 /* _PAGE_RW -> PP lsb */
321 rlwinm r0,r5,32-7,31,31 /* _PAGE_DIRTY -> PP lsb */
322 and r8,r8,r0 /* writable if _RW & _DIRTY */
323 rlwimi r5,r5,32-1,30,30 /* _PAGE_USER -> PP msb */
324 rlwimi r5,r5,32-2,31,31 /* _PAGE_USER -> PP lsb */
325 ori r8,r8,0xe04 /* clear out reserved bits */
326 andc r8,r5,r8 /* PP = user? (rw&dirty? 2: 3): 0 */
328 rlwinm r8,r8,0,~_PAGE_COHERENT /* clear M (coherence not required) */
329 END_FTR_SECTION_IFCLR(CPU_FTR_NEED_COHERENT)
330 #ifdef CONFIG_PTE_64BIT
331 /* Put the XPN bits into the PTE */
332 rlwimi r8,r10,8,20,22
333 rlwimi r8,r10,2,29,29
336 /* Construct the high word of the PPC-style PTE (r5) */
337 rlwinm r5,r3,7,1,24 /* put VSID in 0x7fffff80 bits */
338 rlwimi r5,r4,10,26,31 /* put in API (abbrev page index) */
339 SET_V(r5) /* set V (valid) bit */
341 /* Get the address of the primary PTE group in the hash table (r3) */
342 _GLOBAL(hash_page_patch_A)
343 addis r0,r7,Hash_base@h /* base address of hash table */
344 rlwimi r0,r3,LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* VSID -> hash */
345 rlwinm r3,r4,20+LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* PI -> hash */
346 xor r3,r3,r0 /* make primary hash */
347 li r0,8 /* PTEs/group */
350 * Test the _PAGE_HASHPTE bit in the old linux PTE, and skip the search
351 * if it is clear, meaning that the HPTE isn't there already...
353 andi. r6,r6,_PAGE_HASHPTE
354 beq+ 10f /* no PTE: go look for an empty slot */
357 addis r4,r7,htab_hash_searches@ha
358 lwz r6,htab_hash_searches@l(r4)
359 addi r6,r6,1 /* count how many searches we do */
360 stw r6,htab_hash_searches@l(r4)
362 /* Search the primary PTEG for a PTE whose 1st (d)word matches r5 */
364 addi r4,r3,-HPTE_SIZE
365 1: LDPTEu r6,HPTE_SIZE(r4) /* get next PTE */
367 bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
370 /* Search the secondary PTEG for a matching PTE */
371 ori r5,r5,PTE_H /* set H (secondary hash) bit */
372 _GLOBAL(hash_page_patch_B)
373 xoris r4,r3,Hash_msk>>16 /* compute secondary hash */
374 xori r4,r4,(-PTEG_SIZE & 0xffff)
375 addi r4,r4,-HPTE_SIZE
377 2: LDPTEu r6,HPTE_SIZE(r4)
381 xori r5,r5,PTE_H /* clear H bit again */
383 /* Search the primary PTEG for an empty slot */
385 addi r4,r3,-HPTE_SIZE /* search primary PTEG */
386 1: LDPTEu r6,HPTE_SIZE(r4) /* get next PTE */
387 TST_V(r6) /* test valid bit */
388 bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
391 /* update counter of times that the primary PTEG is full */
392 addis r4,r7,primary_pteg_full@ha
393 lwz r6,primary_pteg_full@l(r4)
395 stw r6,primary_pteg_full@l(r4)
397 /* Search the secondary PTEG for an empty slot */
398 ori r5,r5,PTE_H /* set H (secondary hash) bit */
399 _GLOBAL(hash_page_patch_C)
400 xoris r4,r3,Hash_msk>>16 /* compute secondary hash */
401 xori r4,r4,(-PTEG_SIZE & 0xffff)
402 addi r4,r4,-HPTE_SIZE
404 2: LDPTEu r6,HPTE_SIZE(r4)
408 xori r5,r5,PTE_H /* clear H bit again */
411 * Choose an arbitrary slot in the primary PTEG to overwrite.
412 * Since both the primary and secondary PTEGs are full, and we
413 * have no information that the PTEs in the primary PTEG are
414 * more important or useful than those in the secondary PTEG,
415 * and we know there is a definite (although small) speed
416 * advantage to putting the PTE in the primary PTEG, we always
417 * put the PTE in the primary PTEG.
419 * In addition, we skip any slot that is mapping kernel text in
420 * order to avoid a deadlock when not using BAT mappings if
421 * trying to hash in the kernel hash code itself after it has
422 * already taken the hash table lock. This works in conjunction
423 * with pre-faulting of the kernel text.
425 * If the hash table bucket is full of kernel text entries, we'll
426 * lockup here but that shouldn't happen
429 1: addis r4,r7,next_slot@ha /* get next evict slot */
430 lwz r6,next_slot@l(r4)
431 addi r6,r6,HPTE_SIZE /* search for candidate */
432 andi. r6,r6,7*HPTE_SIZE
433 stw r6,next_slot@l(r4)
435 LDPTE r0,HPTE_SIZE/2(r4) /* get PTE second word */
438 ori r6,r6,etext@l /* get etext */
440 cmpl cr0,r0,r6 /* compare and try again */
444 /* Store PTE in PTEG */
448 STPTE r8,HPTE_SIZE/2(r4)
450 #else /* CONFIG_SMP */
452 * Between the tlbie above and updating the hash table entry below,
453 * another CPU could read the hash table entry and put it in its TLB.
455 * 1. using an empty slot
456 * 2. updating an earlier entry to change permissions (i.e. enable write)
457 * 3. taking over the PTE for an unrelated address
459 * In each case it doesn't really matter if the other CPUs have the old
460 * PTE in their TLB. So we don't need to bother with another tlbie here,
461 * which is convenient as we've overwritten the register that had the
462 * address. :-) The tlbie above is mainly to make sure that this CPU comes
463 * and gets the new PTE from the hash table.
465 * We do however have to make sure that the PTE is never in an invalid
466 * state with the V bit set.
470 CLR_V(r5,r0) /* clear V (valid) bit in PTE */
474 STPTE r8,HPTE_SIZE/2(r4) /* put in correct RPN, WIMG, PP bits */
477 STPTE r5,0(r4) /* finally set V bit in PTE */
478 #endif /* CONFIG_SMP */
480 sync /* make sure pte updates get to memory */
494 * Flush the entry for a particular page from the hash table.
496 * flush_hash_pages(unsigned context, unsigned long va, unsigned long pmdval,
499 * We assume that there is a hash table in use (Hash != 0).
501 _GLOBAL(flush_hash_pages)
505 * We disable interrupts here, even on UP, because we want
506 * the _PAGE_HASHPTE bit to be a reliable indication of
507 * whether the HPTE exists (or at least whether one did once).
508 * We also turn off the MMU for data accesses so that we
509 * we can't take a hash table miss (assuming the code is
510 * covered by a BAT). -- paulus
514 rlwinm r0,r10,0,17,15 /* clear bit 16 (MSR_EE) */
515 rlwinm r0,r0,0,28,26 /* clear MSR_DR */
520 /* First find a PTE in the range that has _PAGE_HASHPTE set */
521 #ifndef CONFIG_PTE_64BIT
522 rlwimi r5,r4,22,20,29
524 rlwimi r5,r4,23,20,28
526 1: lwz r0,PTE_FLAGS_OFFSET(r5)
528 andi. r0,r0,_PAGE_HASHPTE
536 /* Convert context and va to VSID */
537 2: mulli r3,r3,897*16 /* multiply context by context skew */
538 rlwinm r0,r4,4,28,31 /* get ESID (top 4 bits of va) */
539 mulli r0,r0,0x111 /* multiply by ESID skew */
540 add r3,r3,r0 /* note code below trims to 24 bits */
542 /* Construct the high word of the PPC-style PTE (r11) */
543 rlwinm r11,r3,7,1,24 /* put VSID in 0x7fffff80 bits */
544 rlwimi r11,r4,10,26,31 /* put in API (abbrev page index) */
545 SET_V(r11) /* set V (valid) bit */
548 addis r9,r7,mmu_hash_lock@ha
549 addi r9,r9,mmu_hash_lock@l
550 CURRENT_THREAD_INFO(r8, r1)
567 * Check the _PAGE_HASHPTE bit in the linux PTE. If it is
568 * already clear, we're done (for this pte). If not,
569 * clear it (atomically) and proceed. -- paulus.
571 #if (PTE_FLAGS_OFFSET != 0)
572 addi r5,r5,PTE_FLAGS_OFFSET
574 33: lwarx r8,0,r5 /* fetch the pte flags word */
575 andi. r0,r8,_PAGE_HASHPTE
576 beq 8f /* done if HASHPTE is already clear */
577 rlwinm r8,r8,0,31,29 /* clear HASHPTE bit */
578 stwcx. r8,0,r5 /* update the pte */
580 EXPORT_SYMBOL(flush_hash_pages)
582 /* Get the address of the primary PTE group in the hash table (r3) */
583 _GLOBAL(flush_hash_patch_A)
584 addis r8,r7,Hash_base@h /* base address of hash table */
585 rlwimi r8,r3,LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* VSID -> hash */
586 rlwinm r0,r4,20+LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* PI -> hash */
587 xor r8,r0,r8 /* make primary hash */
589 /* Search the primary PTEG for a PTE whose 1st (d)word matches r5 */
590 li r0,8 /* PTEs/group */
592 addi r12,r8,-HPTE_SIZE
593 1: LDPTEu r0,HPTE_SIZE(r12) /* get next PTE */
595 bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
598 /* Search the secondary PTEG for a matching PTE */
599 ori r11,r11,PTE_H /* set H (secondary hash) bit */
600 li r0,8 /* PTEs/group */
601 _GLOBAL(flush_hash_patch_B)
602 xoris r12,r8,Hash_msk>>16 /* compute secondary hash */
603 xori r12,r12,(-PTEG_SIZE & 0xffff)
604 addi r12,r12,-HPTE_SIZE
606 2: LDPTEu r0,HPTE_SIZE(r12)
609 xori r11,r11,PTE_H /* clear H again */
610 bne- 4f /* should rarely fail to find it */
613 STPTE r0,0(r12) /* invalidate entry */
615 tlbie r4 /* in hw tlb too */
618 8: ble cr1,9f /* if all ptes checked */
622 lwz r0,0(r5) /* check next pte */
624 andi. r0,r0,_PAGE_HASHPTE
632 stw r0,0(r9) /* clear mmu_hash_lock */
641 * Flush an entry from the TLB
645 CURRENT_THREAD_INFO(r8, r1)
650 rlwinm r0,r10,0,17,15 /* clear bit 16 (MSR_EE) */
651 rlwinm r0,r0,0,28,26 /* clear DR */
655 lis r9,mmu_hash_lock@h
656 ori r9,r9,mmu_hash_lock@l
668 stw r0,0(r9) /* clear mmu_hash_lock */
672 #else /* CONFIG_SMP */
675 #endif /* CONFIG_SMP */
679 * Flush the entire TLB. 603/603e only
682 #if defined(CONFIG_SMP)
683 CURRENT_THREAD_INFO(r8, r1)
688 rlwinm r0,r10,0,17,15 /* clear bit 16 (MSR_EE) */
689 rlwinm r0,r0,0,28,26 /* clear DR */
693 lis r9,mmu_hash_lock@h
694 ori r9,r9,mmu_hash_lock@l
706 stw r0,0(r9) /* clear mmu_hash_lock */
710 #else /* CONFIG_SMP */
714 #endif /* CONFIG_SMP */