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>
36 #endif /* CONFIG_SMP */
39 * Sync CPUs with hash_page taking & releasing the hash
44 _GLOBAL(hash_page_sync)
46 rlwinm r0,r10,0,17,15 /* clear bit 16 (MSR_EE) */
48 lis r8,mmu_hash_lock@h
49 ori r8,r8,mmu_hash_lock@l
66 #endif /* CONFIG_SMP */
69 * Load a PTE into the hash table, if possible.
70 * The address is in r4, and r3 contains an access flag:
71 * _PAGE_RW (0x400) if a write.
72 * r9 contains the SRR1 value, from which we use the MSR_PR bit.
73 * SPRG3 contains the physical address of the current task's thread.
75 * Returns to the caller if the access is illegal or there is no
76 * mapping for the address. Otherwise it places an appropriate PTE
77 * in the hash table and returns from the exception.
78 * Uses r0, r3 - r8, ctr, lr.
82 tophys(r7,0) /* gets -KERNELBASE into r7 */
84 addis r8,r7,mmu_hash_lock@h
85 ori r8,r8,mmu_hash_lock@l
98 /* Get PTE (linux-style) and check access */
99 lis r0,KERNELBASE@h /* check if kernel address */
101 mfspr r8,SPRN_SPRG3 /* current task's THREAD (phys) */
102 ori r3,r3,_PAGE_USER|_PAGE_PRESENT /* test low addresses as user */
103 lwz r5,PGDIR(r8) /* virt page-table root */
104 blt+ 112f /* assume user more likely */
105 lis r5,swapper_pg_dir@ha /* if kernel address, use */
106 addi r5,r5,swapper_pg_dir@l /* kernel page table */
107 rlwimi r3,r9,32-12,29,29 /* MSR_PR -> _PAGE_USER */
108 112: add r5,r5,r7 /* convert to phys addr */
109 rlwimi r5,r4,12,20,29 /* insert top 10 bits of address */
110 lwz r8,0(r5) /* get pmd entry */
111 rlwinm. r8,r8,0,0,19 /* extract address of pte page */
113 beq- hash_page_out /* return if no mapping */
115 /* XXX it seems like the 601 will give a machine fault on the
116 rfi if its alignment is wrong (bottom 4 bits of address are
117 8 or 0xc) and we have had a not-taken conditional branch
118 to the address following the rfi. */
121 rlwimi r8,r4,22,20,29 /* insert next 10 bits of address */
122 rlwinm r0,r3,32-3,24,24 /* _PAGE_RW access -> _PAGE_DIRTY */
123 ori r0,r0,_PAGE_ACCESSED|_PAGE_HASHPTE
126 * Update the linux PTE atomically. We do the lwarx up-front
127 * because almost always, there won't be a permission violation
128 * and there won't already be an HPTE, and thus we will have
129 * to update the PTE to set _PAGE_HASHPTE. -- paulus.
132 lwarx r6,0,r8 /* get linux-style pte */
133 andc. r5,r3,r6 /* check access & ~permission */
135 bne- hash_page_out /* return if access not permitted */
139 or r5,r0,r6 /* set accessed/dirty bits */
140 stwcx. r5,0,r8 /* attempt to update PTE */
141 bne- retry /* retry if someone got there first */
143 mfsrin r3,r4 /* get segment reg for segment */
146 bl create_hpte /* add the hash table entry */
150 addis r8,r7,mmu_hash_lock@ha
152 stw r0,mmu_hash_lock@l(r8)
155 /* Return from the exception */
161 b fast_exception_return
166 addis r8,r7,mmu_hash_lock@ha
168 stw r0,mmu_hash_lock@l(r8)
170 #endif /* CONFIG_SMP */
173 * Add an entry for a particular page to the hash table.
175 * add_hash_page(unsigned context, unsigned long va, unsigned long pmdval)
177 * We assume any necessary modifications to the pte (e.g. setting
178 * the accessed bit) have already been done and that there is actually
179 * a hash table in use (i.e. we're not on a 603).
181 _GLOBAL(add_hash_page)
185 /* Convert context and va to VSID */
186 mulli r3,r3,897*16 /* multiply context by context skew */
187 rlwinm r0,r4,4,28,31 /* get ESID (top 4 bits of va) */
188 mulli r0,r0,0x111 /* multiply by ESID skew */
189 add r3,r3,r0 /* note create_hpte trims to 24 bits */
192 rlwinm r8,r1,0,0,(31-THREAD_SHIFT) /* use cpu number to make tag */
193 lwz r8,TI_CPU(r8) /* to go in mmu_hash_lock */
195 #endif /* CONFIG_SMP */
198 * We disable interrupts here, even on UP, because we don't
199 * want to race with hash_page, and because we want the
200 * _PAGE_HASHPTE bit to be a reliable indication of whether
201 * the HPTE exists (or at least whether one did once).
202 * We also turn off the MMU for data accesses so that we
203 * we can't take a hash table miss (assuming the code is
204 * covered by a BAT). -- paulus
208 rlwinm r0,r10,0,17,15 /* clear bit 16 (MSR_EE) */
209 rlwinm r0,r0,0,28,26 /* clear MSR_DR */
217 addis r9,r7,mmu_hash_lock@ha
218 addi r9,r9,mmu_hash_lock@l
219 10: lwarx r0,0,r9 /* take the mmu_hash_lock */
232 * Fetch the linux pte and test and set _PAGE_HASHPTE atomically.
233 * If _PAGE_HASHPTE was already set, we don't replace the existing
234 * HPTE, so we just unlock and return.
237 rlwimi r8,r4,22,20,29
239 andi. r0,r6,_PAGE_HASHPTE
240 bne 9f /* if HASHPTE already set, done */
241 ori r5,r6,_PAGE_HASHPTE
251 stw r0,0(r9) /* clear mmu_hash_lock */
254 /* reenable interrupts and DR */
264 * This routine adds a hardware PTE to the hash table.
265 * It is designed to be called with the MMU either on or off.
266 * r3 contains the VSID, r4 contains the virtual address,
267 * r5 contains the linux PTE, r6 contains the old value of the
268 * linux PTE (before setting _PAGE_HASHPTE) and r7 contains the
269 * offset to be added to addresses (0 if the MMU is on,
270 * -KERNELBASE if it is off).
271 * On SMP, the caller should have the mmu_hash_lock held.
272 * We assume that the caller has (or will) set the _PAGE_HASHPTE
273 * bit in the linux PTE in memory. The value passed in r6 should
274 * be the old linux PTE value; if it doesn't have _PAGE_HASHPTE set
275 * this routine will skip the search for an existing HPTE.
276 * This procedure modifies r0, r3 - r6, r8, cr0.
279 * For speed, 4 of the instructions get patched once the size and
280 * physical address of the hash table are known. These definitions
281 * of Hash_base and Hash_bits below are just an example.
283 Hash_base = 0xc0180000
284 Hash_bits = 12 /* e.g. 256kB hash table */
285 Hash_msk = (((1 << Hash_bits) - 1) * 64)
287 /* defines for the PTE format for 32-bit PPCs */
290 #define LG_PTEG_SIZE 6
296 #define PTE_V 0x80000000
297 #define TST_V(r) rlwinm. r,r,0,0,0
298 #define SET_V(r) oris r,r,PTE_V@h
299 #define CLR_V(r,t) rlwinm r,r,0,1,31
301 #define HASH_LEFT 31-(LG_PTEG_SIZE+Hash_bits-1)
302 #define HASH_RIGHT 31-LG_PTEG_SIZE
305 /* Convert linux-style PTE (r5) to low word of PPC-style PTE (r8) */
306 rlwinm r8,r5,32-10,31,31 /* _PAGE_RW -> PP lsb */
307 rlwinm r0,r5,32-7,31,31 /* _PAGE_DIRTY -> PP lsb */
308 and r8,r8,r0 /* writable if _RW & _DIRTY */
309 rlwimi r5,r5,32-1,30,30 /* _PAGE_USER -> PP msb */
310 rlwimi r5,r5,32-2,31,31 /* _PAGE_USER -> PP lsb */
311 ori r8,r8,0xe14 /* clear out reserved bits and M */
312 andc r8,r5,r8 /* PP = user? (rw&dirty? 2: 3): 0 */
314 ori r8,r8,_PAGE_COHERENT /* set M (coherence required) */
315 END_FTR_SECTION_IFSET(CPU_FTR_NEED_COHERENT)
317 /* Construct the high word of the PPC-style PTE (r5) */
318 rlwinm r5,r3,7,1,24 /* put VSID in 0x7fffff80 bits */
319 rlwimi r5,r4,10,26,31 /* put in API (abbrev page index) */
320 SET_V(r5) /* set V (valid) bit */
322 /* Get the address of the primary PTE group in the hash table (r3) */
323 _GLOBAL(hash_page_patch_A)
324 addis r0,r7,Hash_base@h /* base address of hash table */
325 rlwimi r0,r3,LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* VSID -> hash */
326 rlwinm r3,r4,20+LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* PI -> hash */
327 xor r3,r3,r0 /* make primary hash */
328 li r0,8 /* PTEs/group */
331 * Test the _PAGE_HASHPTE bit in the old linux PTE, and skip the search
332 * if it is clear, meaning that the HPTE isn't there already...
334 andi. r6,r6,_PAGE_HASHPTE
335 beq+ 10f /* no PTE: go look for an empty slot */
338 addis r4,r7,htab_hash_searches@ha
339 lwz r6,htab_hash_searches@l(r4)
340 addi r6,r6,1 /* count how many searches we do */
341 stw r6,htab_hash_searches@l(r4)
343 /* Search the primary PTEG for a PTE whose 1st (d)word matches r5 */
346 1: LDPTEu r6,PTE_SIZE(r4) /* get next PTE */
348 bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
351 /* Search the secondary PTEG for a matching PTE */
352 ori r5,r5,PTE_H /* set H (secondary hash) bit */
353 _GLOBAL(hash_page_patch_B)
354 xoris r4,r3,Hash_msk>>16 /* compute secondary hash */
355 xori r4,r4,(-PTEG_SIZE & 0xffff)
358 2: LDPTEu r6,PTE_SIZE(r4)
362 xori r5,r5,PTE_H /* clear H bit again */
364 /* Search the primary PTEG for an empty slot */
366 addi r4,r3,-PTE_SIZE /* search primary PTEG */
367 1: LDPTEu r6,PTE_SIZE(r4) /* get next PTE */
368 TST_V(r6) /* test valid bit */
369 bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
372 /* update counter of times that the primary PTEG is full */
373 addis r4,r7,primary_pteg_full@ha
374 lwz r6,primary_pteg_full@l(r4)
376 stw r6,primary_pteg_full@l(r4)
378 /* Search the secondary PTEG for an empty slot */
379 ori r5,r5,PTE_H /* set H (secondary hash) bit */
380 _GLOBAL(hash_page_patch_C)
381 xoris r4,r3,Hash_msk>>16 /* compute secondary hash */
382 xori r4,r4,(-PTEG_SIZE & 0xffff)
385 2: LDPTEu r6,PTE_SIZE(r4)
389 xori r5,r5,PTE_H /* clear H bit again */
392 * Choose an arbitrary slot in the primary PTEG to overwrite.
393 * Since both the primary and secondary PTEGs are full, and we
394 * have no information that the PTEs in the primary PTEG are
395 * more important or useful than those in the secondary PTEG,
396 * and we know there is a definite (although small) speed
397 * advantage to putting the PTE in the primary PTEG, we always
398 * put the PTE in the primary PTEG.
400 * In addition, we skip any slot that is mapping kernel text in
401 * order to avoid a deadlock when not using BAT mappings if
402 * trying to hash in the kernel hash code itself after it has
403 * already taken the hash table lock. This works in conjunction
404 * with pre-faulting of the kernel text.
406 * If the hash table bucket is full of kernel text entries, we'll
407 * lockup here but that shouldn't happen
410 1: addis r4,r7,next_slot@ha /* get next evict slot */
411 lwz r6,next_slot@l(r4)
412 addi r6,r6,PTE_SIZE /* search for candidate */
413 andi. r6,r6,7*PTE_SIZE
414 stw r6,next_slot@l(r4)
416 LDPTE r0,PTE_SIZE/2(r4) /* get PTE second word */
419 ori r6,r6,etext@l /* get etext */
421 cmpl cr0,r0,r6 /* compare and try again */
425 /* Store PTE in PTEG */
429 STPTE r8,PTE_SIZE/2(r4)
431 #else /* CONFIG_SMP */
433 * Between the tlbie above and updating the hash table entry below,
434 * another CPU could read the hash table entry and put it in its TLB.
436 * 1. using an empty slot
437 * 2. updating an earlier entry to change permissions (i.e. enable write)
438 * 3. taking over the PTE for an unrelated address
440 * In each case it doesn't really matter if the other CPUs have the old
441 * PTE in their TLB. So we don't need to bother with another tlbie here,
442 * which is convenient as we've overwritten the register that had the
443 * address. :-) The tlbie above is mainly to make sure that this CPU comes
444 * and gets the new PTE from the hash table.
446 * We do however have to make sure that the PTE is never in an invalid
447 * state with the V bit set.
451 CLR_V(r5,r0) /* clear V (valid) bit in PTE */
455 STPTE r8,PTE_SIZE/2(r4) /* put in correct RPN, WIMG, PP bits */
458 STPTE r5,0(r4) /* finally set V bit in PTE */
459 #endif /* CONFIG_SMP */
461 sync /* make sure pte updates get to memory */
475 * Flush the entry for a particular page from the hash table.
477 * flush_hash_pages(unsigned context, unsigned long va, unsigned long pmdval,
480 * We assume that there is a hash table in use (Hash != 0).
482 _GLOBAL(flush_hash_pages)
486 * We disable interrupts here, even on UP, because we want
487 * the _PAGE_HASHPTE bit to be a reliable indication of
488 * whether the HPTE exists (or at least whether one did once).
489 * We also turn off the MMU for data accesses so that we
490 * we can't take a hash table miss (assuming the code is
491 * covered by a BAT). -- paulus
495 rlwinm r0,r10,0,17,15 /* clear bit 16 (MSR_EE) */
496 rlwinm r0,r0,0,28,26 /* clear MSR_DR */
501 /* First find a PTE in the range that has _PAGE_HASHPTE set */
502 rlwimi r5,r4,22,20,29
505 andi. r0,r0,_PAGE_HASHPTE
513 /* Convert context and va to VSID */
514 2: mulli r3,r3,897*16 /* multiply context by context skew */
515 rlwinm r0,r4,4,28,31 /* get ESID (top 4 bits of va) */
516 mulli r0,r0,0x111 /* multiply by ESID skew */
517 add r3,r3,r0 /* note code below trims to 24 bits */
519 /* Construct the high word of the PPC-style PTE (r11) */
520 rlwinm r11,r3,7,1,24 /* put VSID in 0x7fffff80 bits */
521 rlwimi r11,r4,10,26,31 /* put in API (abbrev page index) */
522 SET_V(r11) /* set V (valid) bit */
525 addis r9,r7,mmu_hash_lock@ha
526 addi r9,r9,mmu_hash_lock@l
527 rlwinm r8,r1,0,0,(31-THREAD_SHIFT)
544 * Check the _PAGE_HASHPTE bit in the linux PTE. If it is
545 * already clear, we're done (for this pte). If not,
546 * clear it (atomically) and proceed. -- paulus.
548 33: lwarx r8,0,r5 /* fetch the pte */
549 andi. r0,r8,_PAGE_HASHPTE
550 beq 8f /* done if HASHPTE is already clear */
551 rlwinm r8,r8,0,31,29 /* clear HASHPTE bit */
552 stwcx. r8,0,r5 /* update the pte */
555 /* Get the address of the primary PTE group in the hash table (r3) */
556 _GLOBAL(flush_hash_patch_A)
557 addis r8,r7,Hash_base@h /* base address of hash table */
558 rlwimi r8,r3,LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* VSID -> hash */
559 rlwinm r0,r4,20+LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* PI -> hash */
560 xor r8,r0,r8 /* make primary hash */
562 /* Search the primary PTEG for a PTE whose 1st (d)word matches r5 */
563 li r0,8 /* PTEs/group */
565 addi r12,r8,-PTE_SIZE
566 1: LDPTEu r0,PTE_SIZE(r12) /* get next PTE */
568 bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
571 /* Search the secondary PTEG for a matching PTE */
572 ori r11,r11,PTE_H /* set H (secondary hash) bit */
573 li r0,8 /* PTEs/group */
574 _GLOBAL(flush_hash_patch_B)
575 xoris r12,r8,Hash_msk>>16 /* compute secondary hash */
576 xori r12,r12,(-PTEG_SIZE & 0xffff)
577 addi r12,r12,-PTE_SIZE
579 2: LDPTEu r0,PTE_SIZE(r12)
582 xori r11,r11,PTE_H /* clear H again */
583 bne- 4f /* should rarely fail to find it */
586 STPTE r0,0(r12) /* invalidate entry */
588 tlbie r4 /* in hw tlb too */
591 8: ble cr1,9f /* if all ptes checked */
593 addi r5,r5,4 /* advance to next pte */
595 lwz r0,0(r5) /* check next pte */
597 andi. r0,r0,_PAGE_HASHPTE
605 stw r0,0(r9) /* clear mmu_hash_lock */