2 * $Id: hashtable.S,v 1.6 1999/10/08 01:56:15 paulus Exp $
5 * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
6 * Rewritten by Cort Dougan (cort@cs.nmt.edu) for PReP
7 * Copyright (C) 1996 Cort Dougan <cort@cs.nmt.edu>
8 * Adapted for Power Macintosh by Paul Mackerras.
9 * Low-level exception handlers and MMU support
10 * rewritten by Paul Mackerras.
11 * Copyright (C) 1996 Paul Mackerras.
13 * This file contains low-level assembler routines for managing
14 * the PowerPC MMU hash table. (PPC 8xx processors don't use a
15 * hash table, so this file is not used on them.)
17 * This program is free software; you can redistribute it and/or
18 * modify it under the terms of the GNU General Public License
19 * as published by the Free Software Foundation; either version
20 * 2 of the License, or (at your option) any later version.
26 #include <asm/pgtable.h>
27 #include <asm/cputable.h>
28 #include <asm/ppc_asm.h>
29 #include <asm/thread_info.h>
30 #include <asm/asm-offsets.h>
34 #endif /* CONFIG_SMP */
37 * Sync CPUs with hash_page taking & releasing the hash
42 _GLOBAL(hash_page_sync)
43 lis r8,mmu_hash_lock@h
44 ori r8,r8,mmu_hash_lock@l
63 * Load a PTE into the hash table, if possible.
64 * The address is in r4, and r3 contains an access flag:
65 * _PAGE_RW (0x400) if a write.
66 * r9 contains the SRR1 value, from which we use the MSR_PR bit.
67 * SPRG3 contains the physical address of the current task's thread.
69 * Returns to the caller if the access is illegal or there is no
70 * mapping for the address. Otherwise it places an appropriate PTE
71 * in the hash table and returns from the exception.
72 * Uses r0, r3 - r8, ctr, lr.
76 tophys(r7,0) /* gets -KERNELBASE into r7 */
78 addis r8,r7,mmu_hash_lock@h
79 ori r8,r8,mmu_hash_lock@l
92 /* Get PTE (linux-style) and check access */
93 lis r0,KERNELBASE@h /* check if kernel address */
95 mfspr r8,SPRN_SPRG3 /* current task's THREAD (phys) */
96 ori r3,r3,_PAGE_USER|_PAGE_PRESENT /* test low addresses as user */
97 lwz r5,PGDIR(r8) /* virt page-table root */
98 blt+ 112f /* assume user more likely */
99 lis r5,swapper_pg_dir@ha /* if kernel address, use */
100 addi r5,r5,swapper_pg_dir@l /* kernel page table */
101 rlwimi r3,r9,32-12,29,29 /* MSR_PR -> _PAGE_USER */
102 112: add r5,r5,r7 /* convert to phys addr */
103 rlwimi r5,r4,12,20,29 /* insert top 10 bits of address */
104 lwz r8,0(r5) /* get pmd entry */
105 rlwinm. r8,r8,0,0,19 /* extract address of pte page */
107 beq- hash_page_out /* return if no mapping */
109 /* XXX it seems like the 601 will give a machine fault on the
110 rfi if its alignment is wrong (bottom 4 bits of address are
111 8 or 0xc) and we have had a not-taken conditional branch
112 to the address following the rfi. */
115 rlwimi r8,r4,22,20,29 /* insert next 10 bits of address */
116 rlwinm r0,r3,32-3,24,24 /* _PAGE_RW access -> _PAGE_DIRTY */
117 ori r0,r0,_PAGE_ACCESSED|_PAGE_HASHPTE
120 * Update the linux PTE atomically. We do the lwarx up-front
121 * because almost always, there won't be a permission violation
122 * and there won't already be an HPTE, and thus we will have
123 * to update the PTE to set _PAGE_HASHPTE. -- paulus.
126 lwarx r6,0,r8 /* get linux-style pte */
127 andc. r5,r3,r6 /* check access & ~permission */
129 bne- hash_page_out /* return if access not permitted */
133 or r5,r0,r6 /* set accessed/dirty bits */
134 stwcx. r5,0,r8 /* attempt to update PTE */
135 bne- retry /* retry if someone got there first */
137 mfsrin r3,r4 /* get segment reg for segment */
140 bl create_hpte /* add the hash table entry */
144 addis r8,r7,mmu_hash_lock@ha
146 stw r0,mmu_hash_lock@l(r8)
149 /* Return from the exception */
155 b fast_exception_return
160 addis r8,r7,mmu_hash_lock@ha
162 stw r0,mmu_hash_lock@l(r8)
164 #endif /* CONFIG_SMP */
167 * Add an entry for a particular page to the hash table.
169 * add_hash_page(unsigned context, unsigned long va, unsigned long pmdval)
171 * We assume any necessary modifications to the pte (e.g. setting
172 * the accessed bit) have already been done and that there is actually
173 * a hash table in use (i.e. we're not on a 603).
175 _GLOBAL(add_hash_page)
179 /* Convert context and va to VSID */
180 mulli r3,r3,897*16 /* multiply context by context skew */
181 rlwinm r0,r4,4,28,31 /* get ESID (top 4 bits of va) */
182 mulli r0,r0,0x111 /* multiply by ESID skew */
183 add r3,r3,r0 /* note create_hpte trims to 24 bits */
186 rlwinm r8,r1,0,0,18 /* use cpu number to make tag */
187 lwz r8,TI_CPU(r8) /* to go in mmu_hash_lock */
189 #endif /* CONFIG_SMP */
192 * We disable interrupts here, even on UP, because we don't
193 * want to race with hash_page, and because we want the
194 * _PAGE_HASHPTE bit to be a reliable indication of whether
195 * the HPTE exists (or at least whether one did once).
196 * We also turn off the MMU for data accesses so that we
197 * we can't take a hash table miss (assuming the code is
198 * covered by a BAT). -- paulus
202 rlwinm r0,r10,0,17,15 /* clear bit 16 (MSR_EE) */
203 rlwinm r0,r0,0,28,26 /* clear MSR_DR */
211 addis r9,r7,mmu_hash_lock@ha
212 addi r9,r9,mmu_hash_lock@l
213 10: lwarx r0,0,r9 /* take the mmu_hash_lock */
226 * Fetch the linux pte and test and set _PAGE_HASHPTE atomically.
227 * If _PAGE_HASHPTE was already set, we don't replace the existing
228 * HPTE, so we just unlock and return.
231 rlwimi r8,r4,22,20,29
233 andi. r0,r6,_PAGE_HASHPTE
234 bne 9f /* if HASHPTE already set, done */
235 ori r5,r6,_PAGE_HASHPTE
245 stw r0,0(r9) /* clear mmu_hash_lock */
248 /* reenable interrupts and DR */
258 * This routine adds a hardware PTE to the hash table.
259 * It is designed to be called with the MMU either on or off.
260 * r3 contains the VSID, r4 contains the virtual address,
261 * r5 contains the linux PTE, r6 contains the old value of the
262 * linux PTE (before setting _PAGE_HASHPTE) and r7 contains the
263 * offset to be added to addresses (0 if the MMU is on,
264 * -KERNELBASE if it is off).
265 * On SMP, the caller should have the mmu_hash_lock held.
266 * We assume that the caller has (or will) set the _PAGE_HASHPTE
267 * bit in the linux PTE in memory. The value passed in r6 should
268 * be the old linux PTE value; if it doesn't have _PAGE_HASHPTE set
269 * this routine will skip the search for an existing HPTE.
270 * This procedure modifies r0, r3 - r6, r8, cr0.
273 * For speed, 4 of the instructions get patched once the size and
274 * physical address of the hash table are known. These definitions
275 * of Hash_base and Hash_bits below are just an example.
277 Hash_base = 0xc0180000
278 Hash_bits = 12 /* e.g. 256kB hash table */
279 Hash_msk = (((1 << Hash_bits) - 1) * 64)
281 /* defines for the PTE format for 32-bit PPCs */
284 #define LG_PTEG_SIZE 6
290 #define PTE_V 0x80000000
291 #define TST_V(r) rlwinm. r,r,0,0,0
292 #define SET_V(r) oris r,r,PTE_V@h
293 #define CLR_V(r,t) rlwinm r,r,0,1,31
295 #define HASH_LEFT 31-(LG_PTEG_SIZE+Hash_bits-1)
296 #define HASH_RIGHT 31-LG_PTEG_SIZE
299 /* Convert linux-style PTE (r5) to low word of PPC-style PTE (r8) */
300 rlwinm r8,r5,32-10,31,31 /* _PAGE_RW -> PP lsb */
301 rlwinm r0,r5,32-7,31,31 /* _PAGE_DIRTY -> PP lsb */
302 and r8,r8,r0 /* writable if _RW & _DIRTY */
303 rlwimi r5,r5,32-1,30,30 /* _PAGE_USER -> PP msb */
304 rlwimi r5,r5,32-2,31,31 /* _PAGE_USER -> PP lsb */
305 ori r8,r8,0xe14 /* clear out reserved bits and M */
306 andc r8,r5,r8 /* PP = user? (rw&dirty? 2: 3): 0 */
308 ori r8,r8,_PAGE_COHERENT /* set M (coherence required) */
309 END_FTR_SECTION_IFSET(CPU_FTR_NEED_COHERENT)
311 /* Construct the high word of the PPC-style PTE (r5) */
312 rlwinm r5,r3,7,1,24 /* put VSID in 0x7fffff80 bits */
313 rlwimi r5,r4,10,26,31 /* put in API (abbrev page index) */
314 SET_V(r5) /* set V (valid) bit */
316 /* Get the address of the primary PTE group in the hash table (r3) */
317 _GLOBAL(hash_page_patch_A)
318 addis r0,r7,Hash_base@h /* base address of hash table */
319 rlwimi r0,r3,LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* VSID -> hash */
320 rlwinm r3,r4,20+LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* PI -> hash */
321 xor r3,r3,r0 /* make primary hash */
322 li r0,8 /* PTEs/group */
325 * Test the _PAGE_HASHPTE bit in the old linux PTE, and skip the search
326 * if it is clear, meaning that the HPTE isn't there already...
328 andi. r6,r6,_PAGE_HASHPTE
329 beq+ 10f /* no PTE: go look for an empty slot */
332 addis r4,r7,htab_hash_searches@ha
333 lwz r6,htab_hash_searches@l(r4)
334 addi r6,r6,1 /* count how many searches we do */
335 stw r6,htab_hash_searches@l(r4)
337 /* Search the primary PTEG for a PTE whose 1st (d)word matches r5 */
340 1: LDPTEu r6,PTE_SIZE(r4) /* get next PTE */
342 bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
345 /* Search the secondary PTEG for a matching PTE */
346 ori r5,r5,PTE_H /* set H (secondary hash) bit */
347 _GLOBAL(hash_page_patch_B)
348 xoris r4,r3,Hash_msk>>16 /* compute secondary hash */
349 xori r4,r4,(-PTEG_SIZE & 0xffff)
352 2: LDPTEu r6,PTE_SIZE(r4)
356 xori r5,r5,PTE_H /* clear H bit again */
358 /* Search the primary PTEG for an empty slot */
360 addi r4,r3,-PTE_SIZE /* search primary PTEG */
361 1: LDPTEu r6,PTE_SIZE(r4) /* get next PTE */
362 TST_V(r6) /* test valid bit */
363 bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
366 /* update counter of times that the primary PTEG is full */
367 addis r4,r7,primary_pteg_full@ha
368 lwz r6,primary_pteg_full@l(r4)
370 stw r6,primary_pteg_full@l(r4)
372 /* Search the secondary PTEG for an empty slot */
373 ori r5,r5,PTE_H /* set H (secondary hash) bit */
374 _GLOBAL(hash_page_patch_C)
375 xoris r4,r3,Hash_msk>>16 /* compute secondary hash */
376 xori r4,r4,(-PTEG_SIZE & 0xffff)
379 2: LDPTEu r6,PTE_SIZE(r4)
383 xori r5,r5,PTE_H /* clear H bit again */
386 * Choose an arbitrary slot in the primary PTEG to overwrite.
387 * Since both the primary and secondary PTEGs are full, and we
388 * have no information that the PTEs in the primary PTEG are
389 * more important or useful than those in the secondary PTEG,
390 * and we know there is a definite (although small) speed
391 * advantage to putting the PTE in the primary PTEG, we always
392 * put the PTE in the primary PTEG.
394 * In addition, we skip any slot that is mapping kernel text in
395 * order to avoid a deadlock when not using BAT mappings if
396 * trying to hash in the kernel hash code itself after it has
397 * already taken the hash table lock. This works in conjunction
398 * with pre-faulting of the kernel text.
400 * If the hash table bucket is full of kernel text entries, we'll
401 * lockup here but that shouldn't happen
404 1: addis r4,r7,next_slot@ha /* get next evict slot */
405 lwz r6,next_slot@l(r4)
406 addi r6,r6,PTE_SIZE /* search for candidate */
407 andi. r6,r6,7*PTE_SIZE
408 stw r6,next_slot@l(r4)
410 LDPTE r0,PTE_SIZE/2(r4) /* get PTE second word */
413 ori r6,r6,etext@l /* get etext */
415 cmpl cr0,r0,r6 /* compare and try again */
419 /* Store PTE in PTEG */
423 STPTE r8,PTE_SIZE/2(r4)
425 #else /* CONFIG_SMP */
427 * Between the tlbie above and updating the hash table entry below,
428 * another CPU could read the hash table entry and put it in its TLB.
430 * 1. using an empty slot
431 * 2. updating an earlier entry to change permissions (i.e. enable write)
432 * 3. taking over the PTE for an unrelated address
434 * In each case it doesn't really matter if the other CPUs have the old
435 * PTE in their TLB. So we don't need to bother with another tlbie here,
436 * which is convenient as we've overwritten the register that had the
437 * address. :-) The tlbie above is mainly to make sure that this CPU comes
438 * and gets the new PTE from the hash table.
440 * We do however have to make sure that the PTE is never in an invalid
441 * state with the V bit set.
445 CLR_V(r5,r0) /* clear V (valid) bit in PTE */
449 STPTE r8,PTE_SIZE/2(r4) /* put in correct RPN, WIMG, PP bits */
452 STPTE r5,0(r4) /* finally set V bit in PTE */
453 #endif /* CONFIG_SMP */
455 sync /* make sure pte updates get to memory */
459 .comm primary_pteg_full,4
460 .comm htab_hash_searches,4
463 * Flush the entry for a particular page from the hash table.
465 * flush_hash_pages(unsigned context, unsigned long va, unsigned long pmdval,
468 * We assume that there is a hash table in use (Hash != 0).
470 _GLOBAL(flush_hash_pages)
474 * We disable interrupts here, even on UP, because we want
475 * the _PAGE_HASHPTE bit to be a reliable indication of
476 * whether the HPTE exists (or at least whether one did once).
477 * We also turn off the MMU for data accesses so that we
478 * we can't take a hash table miss (assuming the code is
479 * covered by a BAT). -- paulus
483 rlwinm r0,r10,0,17,15 /* clear bit 16 (MSR_EE) */
484 rlwinm r0,r0,0,28,26 /* clear MSR_DR */
489 /* First find a PTE in the range that has _PAGE_HASHPTE set */
490 rlwimi r5,r4,22,20,29
493 andi. r0,r0,_PAGE_HASHPTE
501 /* Convert context and va to VSID */
502 2: mulli r3,r3,897*16 /* multiply context by context skew */
503 rlwinm r0,r4,4,28,31 /* get ESID (top 4 bits of va) */
504 mulli r0,r0,0x111 /* multiply by ESID skew */
505 add r3,r3,r0 /* note code below trims to 24 bits */
507 /* Construct the high word of the PPC-style PTE (r11) */
508 rlwinm r11,r3,7,1,24 /* put VSID in 0x7fffff80 bits */
509 rlwimi r11,r4,10,26,31 /* put in API (abbrev page index) */
510 SET_V(r11) /* set V (valid) bit */
513 addis r9,r7,mmu_hash_lock@ha
514 addi r9,r9,mmu_hash_lock@l
532 * Check the _PAGE_HASHPTE bit in the linux PTE. If it is
533 * already clear, we're done (for this pte). If not,
534 * clear it (atomically) and proceed. -- paulus.
536 33: lwarx r8,0,r5 /* fetch the pte */
537 andi. r0,r8,_PAGE_HASHPTE
538 beq 8f /* done if HASHPTE is already clear */
539 rlwinm r8,r8,0,31,29 /* clear HASHPTE bit */
540 stwcx. r8,0,r5 /* update the pte */
543 /* Get the address of the primary PTE group in the hash table (r3) */
544 _GLOBAL(flush_hash_patch_A)
545 addis r8,r7,Hash_base@h /* base address of hash table */
546 rlwimi r8,r3,LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* VSID -> hash */
547 rlwinm r0,r4,20+LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* PI -> hash */
548 xor r8,r0,r8 /* make primary hash */
550 /* Search the primary PTEG for a PTE whose 1st (d)word matches r5 */
551 li r0,8 /* PTEs/group */
553 addi r12,r8,-PTE_SIZE
554 1: LDPTEu r0,PTE_SIZE(r12) /* get next PTE */
556 bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
559 /* Search the secondary PTEG for a matching PTE */
560 ori r11,r11,PTE_H /* set H (secondary hash) bit */
561 li r0,8 /* PTEs/group */
562 _GLOBAL(flush_hash_patch_B)
563 xoris r12,r8,Hash_msk>>16 /* compute secondary hash */
564 xori r12,r12,(-PTEG_SIZE & 0xffff)
565 addi r12,r12,-PTE_SIZE
567 2: LDPTEu r0,PTE_SIZE(r12)
570 xori r11,r11,PTE_H /* clear H again */
571 bne- 4f /* should rarely fail to find it */
574 STPTE r0,0(r12) /* invalidate entry */
576 tlbie r4 /* in hw tlb too */
579 8: ble cr1,9f /* if all ptes checked */
581 addi r5,r5,4 /* advance to next pte */
583 lwz r0,0(r5) /* check next pte */
585 andi. r0,r0,_PAGE_HASHPTE
593 stw r0,0(r9) /* clear mmu_hash_lock */