linux/arch/powerpc/mm/hash_low_32.S
Benjamin Herrenschmidt ee4f2ea486 [POWERPC] Fix 32-bit mm operations when not using BATs
On hash table based 32 bits powerpc's, the hash management code runs with
a big spinlock. It's thus important that it never causes itself a hash
fault. That code is generally safe (it does memory accesses in real mode
among other things) with the exception of the actual access to the code
itself. That is, the kernel text needs to be accessible without taking
a hash miss exceptions.

This is currently guaranteed by having a BAT register mapping part of the
linear mapping permanently, which includes the kernel text. But this is
not true if using the "nobats" kernel command line option (which can be
useful for debugging) and will not be true when using DEBUG_PAGEALLOC
implemented in a subsequent patch.

This patch fixes this by pre-faulting in the hash table pages that hit
the kernel text, and making sure we never evict such a page under hash
pressure.

Signed-off-by: Benjamin Herrenchmidt <benh@kernel.crashing.org>

 arch/powerpc/mm/hash_low_32.S |   22 ++++++++++++++++++++--
 arch/powerpc/mm/mem.c         |    3 ---
 arch/powerpc/mm/mmu_decl.h    |    4 ++++
 arch/powerpc/mm/pgtable_32.c  |   11 +++++++----
 4 files changed, 31 insertions(+), 9 deletions(-)
Signed-off-by: Paul Mackerras <paulus@samba.org>
2007-04-13 04:09:39 +10:00

600 lines
17 KiB
ArmAsm

/*
* $Id: hashtable.S,v 1.6 1999/10/08 01:56:15 paulus Exp $
*
* PowerPC version
* Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
* Rewritten by Cort Dougan (cort@cs.nmt.edu) for PReP
* Copyright (C) 1996 Cort Dougan <cort@cs.nmt.edu>
* Adapted for Power Macintosh by Paul Mackerras.
* Low-level exception handlers and MMU support
* rewritten by Paul Mackerras.
* Copyright (C) 1996 Paul Mackerras.
*
* This file contains low-level assembler routines for managing
* the PowerPC MMU hash table. (PPC 8xx processors don't use a
* hash table, so this file is not used on them.)
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version
* 2 of the License, or (at your option) any later version.
*
*/
#include <asm/reg.h>
#include <asm/page.h>
#include <asm/pgtable.h>
#include <asm/cputable.h>
#include <asm/ppc_asm.h>
#include <asm/thread_info.h>
#include <asm/asm-offsets.h>
#ifdef CONFIG_SMP
.comm mmu_hash_lock,4
#endif /* CONFIG_SMP */
/*
* Sync CPUs with hash_page taking & releasing the hash
* table lock
*/
#ifdef CONFIG_SMP
.text
_GLOBAL(hash_page_sync)
lis r8,mmu_hash_lock@h
ori r8,r8,mmu_hash_lock@l
lis r0,0x0fff
b 10f
11: lwz r6,0(r8)
cmpwi 0,r6,0
bne 11b
10: lwarx r6,0,r8
cmpwi 0,r6,0
bne- 11b
stwcx. r0,0,r8
bne- 10b
isync
eieio
li r0,0
stw r0,0(r8)
blr
#endif
/*
* Load a PTE into the hash table, if possible.
* The address is in r4, and r3 contains an access flag:
* _PAGE_RW (0x400) if a write.
* r9 contains the SRR1 value, from which we use the MSR_PR bit.
* SPRG3 contains the physical address of the current task's thread.
*
* Returns to the caller if the access is illegal or there is no
* mapping for the address. Otherwise it places an appropriate PTE
* in the hash table and returns from the exception.
* Uses r0, r3 - r8, ctr, lr.
*/
.text
_GLOBAL(hash_page)
tophys(r7,0) /* gets -KERNELBASE into r7 */
#ifdef CONFIG_SMP
addis r8,r7,mmu_hash_lock@h
ori r8,r8,mmu_hash_lock@l
lis r0,0x0fff
b 10f
11: lwz r6,0(r8)
cmpwi 0,r6,0
bne 11b
10: lwarx r6,0,r8
cmpwi 0,r6,0
bne- 11b
stwcx. r0,0,r8
bne- 10b
isync
#endif
/* Get PTE (linux-style) and check access */
lis r0,KERNELBASE@h /* check if kernel address */
cmplw 0,r4,r0
mfspr r8,SPRN_SPRG3 /* current task's THREAD (phys) */
ori r3,r3,_PAGE_USER|_PAGE_PRESENT /* test low addresses as user */
lwz r5,PGDIR(r8) /* virt page-table root */
blt+ 112f /* assume user more likely */
lis r5,swapper_pg_dir@ha /* if kernel address, use */
addi r5,r5,swapper_pg_dir@l /* kernel page table */
rlwimi r3,r9,32-12,29,29 /* MSR_PR -> _PAGE_USER */
112: add r5,r5,r7 /* convert to phys addr */
rlwimi r5,r4,12,20,29 /* insert top 10 bits of address */
lwz r8,0(r5) /* get pmd entry */
rlwinm. r8,r8,0,0,19 /* extract address of pte page */
#ifdef CONFIG_SMP
beq- hash_page_out /* return if no mapping */
#else
/* XXX it seems like the 601 will give a machine fault on the
rfi if its alignment is wrong (bottom 4 bits of address are
8 or 0xc) and we have had a not-taken conditional branch
to the address following the rfi. */
beqlr-
#endif
rlwimi r8,r4,22,20,29 /* insert next 10 bits of address */
rlwinm r0,r3,32-3,24,24 /* _PAGE_RW access -> _PAGE_DIRTY */
ori r0,r0,_PAGE_ACCESSED|_PAGE_HASHPTE
/*
* Update the linux PTE atomically. We do the lwarx up-front
* because almost always, there won't be a permission violation
* and there won't already be an HPTE, and thus we will have
* to update the PTE to set _PAGE_HASHPTE. -- paulus.
*/
retry:
lwarx r6,0,r8 /* get linux-style pte */
andc. r5,r3,r6 /* check access & ~permission */
#ifdef CONFIG_SMP
bne- hash_page_out /* return if access not permitted */
#else
bnelr-
#endif
or r5,r0,r6 /* set accessed/dirty bits */
stwcx. r5,0,r8 /* attempt to update PTE */
bne- retry /* retry if someone got there first */
mfsrin r3,r4 /* get segment reg for segment */
mfctr r0
stw r0,_CTR(r11)
bl create_hpte /* add the hash table entry */
#ifdef CONFIG_SMP
eieio
addis r8,r7,mmu_hash_lock@ha
li r0,0
stw r0,mmu_hash_lock@l(r8)
#endif
/* Return from the exception */
lwz r5,_CTR(r11)
mtctr r5
lwz r0,GPR0(r11)
lwz r7,GPR7(r11)
lwz r8,GPR8(r11)
b fast_exception_return
#ifdef CONFIG_SMP
hash_page_out:
eieio
addis r8,r7,mmu_hash_lock@ha
li r0,0
stw r0,mmu_hash_lock@l(r8)
blr
#endif /* CONFIG_SMP */
/*
* Add an entry for a particular page to the hash table.
*
* add_hash_page(unsigned context, unsigned long va, unsigned long pmdval)
*
* We assume any necessary modifications to the pte (e.g. setting
* the accessed bit) have already been done and that there is actually
* a hash table in use (i.e. we're not on a 603).
*/
_GLOBAL(add_hash_page)
mflr r0
stw r0,4(r1)
/* Convert context and va to VSID */
mulli r3,r3,897*16 /* multiply context by context skew */
rlwinm r0,r4,4,28,31 /* get ESID (top 4 bits of va) */
mulli r0,r0,0x111 /* multiply by ESID skew */
add r3,r3,r0 /* note create_hpte trims to 24 bits */
#ifdef CONFIG_SMP
rlwinm r8,r1,0,0,18 /* use cpu number to make tag */
lwz r8,TI_CPU(r8) /* to go in mmu_hash_lock */
oris r8,r8,12
#endif /* CONFIG_SMP */
/*
* We disable interrupts here, even on UP, because we don't
* want to race with hash_page, and because we want the
* _PAGE_HASHPTE bit to be a reliable indication of whether
* the HPTE exists (or at least whether one did once).
* We also turn off the MMU for data accesses so that we
* we can't take a hash table miss (assuming the code is
* covered by a BAT). -- paulus
*/
mfmsr r10
SYNC
rlwinm r0,r10,0,17,15 /* clear bit 16 (MSR_EE) */
rlwinm r0,r0,0,28,26 /* clear MSR_DR */
mtmsr r0
SYNC_601
isync
tophys(r7,0)
#ifdef CONFIG_SMP
addis r9,r7,mmu_hash_lock@ha
addi r9,r9,mmu_hash_lock@l
10: lwarx r0,0,r9 /* take the mmu_hash_lock */
cmpi 0,r0,0
bne- 11f
stwcx. r8,0,r9
beq+ 12f
11: lwz r0,0(r9)
cmpi 0,r0,0
beq 10b
b 11b
12: isync
#endif
/*
* Fetch the linux pte and test and set _PAGE_HASHPTE atomically.
* If _PAGE_HASHPTE was already set, we don't replace the existing
* HPTE, so we just unlock and return.
*/
mr r8,r5
rlwimi r8,r4,22,20,29
1: lwarx r6,0,r8
andi. r0,r6,_PAGE_HASHPTE
bne 9f /* if HASHPTE already set, done */
ori r5,r6,_PAGE_HASHPTE
stwcx. r5,0,r8
bne- 1b
bl create_hpte
9:
#ifdef CONFIG_SMP
eieio
li r0,0
stw r0,0(r9) /* clear mmu_hash_lock */
#endif
/* reenable interrupts and DR */
mtmsr r10
SYNC_601
isync
lwz r0,4(r1)
mtlr r0
blr
/*
* This routine adds a hardware PTE to the hash table.
* It is designed to be called with the MMU either on or off.
* r3 contains the VSID, r4 contains the virtual address,
* r5 contains the linux PTE, r6 contains the old value of the
* linux PTE (before setting _PAGE_HASHPTE) and r7 contains the
* offset to be added to addresses (0 if the MMU is on,
* -KERNELBASE if it is off).
* On SMP, the caller should have the mmu_hash_lock held.
* We assume that the caller has (or will) set the _PAGE_HASHPTE
* bit in the linux PTE in memory. The value passed in r6 should
* be the old linux PTE value; if it doesn't have _PAGE_HASHPTE set
* this routine will skip the search for an existing HPTE.
* This procedure modifies r0, r3 - r6, r8, cr0.
* -- paulus.
*
* For speed, 4 of the instructions get patched once the size and
* physical address of the hash table are known. These definitions
* of Hash_base and Hash_bits below are just an example.
*/
Hash_base = 0xc0180000
Hash_bits = 12 /* e.g. 256kB hash table */
Hash_msk = (((1 << Hash_bits) - 1) * 64)
/* defines for the PTE format for 32-bit PPCs */
#define PTE_SIZE 8
#define PTEG_SIZE 64
#define LG_PTEG_SIZE 6
#define LDPTEu lwzu
#define LDPTE lwz
#define STPTE stw
#define CMPPTE cmpw
#define PTE_H 0x40
#define PTE_V 0x80000000
#define TST_V(r) rlwinm. r,r,0,0,0
#define SET_V(r) oris r,r,PTE_V@h
#define CLR_V(r,t) rlwinm r,r,0,1,31
#define HASH_LEFT 31-(LG_PTEG_SIZE+Hash_bits-1)
#define HASH_RIGHT 31-LG_PTEG_SIZE
_GLOBAL(create_hpte)
/* Convert linux-style PTE (r5) to low word of PPC-style PTE (r8) */
rlwinm r8,r5,32-10,31,31 /* _PAGE_RW -> PP lsb */
rlwinm r0,r5,32-7,31,31 /* _PAGE_DIRTY -> PP lsb */
and r8,r8,r0 /* writable if _RW & _DIRTY */
rlwimi r5,r5,32-1,30,30 /* _PAGE_USER -> PP msb */
rlwimi r5,r5,32-2,31,31 /* _PAGE_USER -> PP lsb */
ori r8,r8,0xe14 /* clear out reserved bits and M */
andc r8,r5,r8 /* PP = user? (rw&dirty? 2: 3): 0 */
BEGIN_FTR_SECTION
ori r8,r8,_PAGE_COHERENT /* set M (coherence required) */
END_FTR_SECTION_IFSET(CPU_FTR_NEED_COHERENT)
/* Construct the high word of the PPC-style PTE (r5) */
rlwinm r5,r3,7,1,24 /* put VSID in 0x7fffff80 bits */
rlwimi r5,r4,10,26,31 /* put in API (abbrev page index) */
SET_V(r5) /* set V (valid) bit */
/* Get the address of the primary PTE group in the hash table (r3) */
_GLOBAL(hash_page_patch_A)
addis r0,r7,Hash_base@h /* base address of hash table */
rlwimi r0,r3,LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* VSID -> hash */
rlwinm r3,r4,20+LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* PI -> hash */
xor r3,r3,r0 /* make primary hash */
li r0,8 /* PTEs/group */
/*
* Test the _PAGE_HASHPTE bit in the old linux PTE, and skip the search
* if it is clear, meaning that the HPTE isn't there already...
*/
andi. r6,r6,_PAGE_HASHPTE
beq+ 10f /* no PTE: go look for an empty slot */
tlbie r4
addis r4,r7,htab_hash_searches@ha
lwz r6,htab_hash_searches@l(r4)
addi r6,r6,1 /* count how many searches we do */
stw r6,htab_hash_searches@l(r4)
/* Search the primary PTEG for a PTE whose 1st (d)word matches r5 */
mtctr r0
addi r4,r3,-PTE_SIZE
1: LDPTEu r6,PTE_SIZE(r4) /* get next PTE */
CMPPTE 0,r6,r5
bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
beq+ found_slot
/* Search the secondary PTEG for a matching PTE */
ori r5,r5,PTE_H /* set H (secondary hash) bit */
_GLOBAL(hash_page_patch_B)
xoris r4,r3,Hash_msk>>16 /* compute secondary hash */
xori r4,r4,(-PTEG_SIZE & 0xffff)
addi r4,r4,-PTE_SIZE
mtctr r0
2: LDPTEu r6,PTE_SIZE(r4)
CMPPTE 0,r6,r5
bdnzf 2,2b
beq+ found_slot
xori r5,r5,PTE_H /* clear H bit again */
/* Search the primary PTEG for an empty slot */
10: mtctr r0
addi r4,r3,-PTE_SIZE /* search primary PTEG */
1: LDPTEu r6,PTE_SIZE(r4) /* get next PTE */
TST_V(r6) /* test valid bit */
bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
beq+ found_empty
/* update counter of times that the primary PTEG is full */
addis r4,r7,primary_pteg_full@ha
lwz r6,primary_pteg_full@l(r4)
addi r6,r6,1
stw r6,primary_pteg_full@l(r4)
/* Search the secondary PTEG for an empty slot */
ori r5,r5,PTE_H /* set H (secondary hash) bit */
_GLOBAL(hash_page_patch_C)
xoris r4,r3,Hash_msk>>16 /* compute secondary hash */
xori r4,r4,(-PTEG_SIZE & 0xffff)
addi r4,r4,-PTE_SIZE
mtctr r0
2: LDPTEu r6,PTE_SIZE(r4)
TST_V(r6)
bdnzf 2,2b
beq+ found_empty
xori r5,r5,PTE_H /* clear H bit again */
/*
* Choose an arbitrary slot in the primary PTEG to overwrite.
* Since both the primary and secondary PTEGs are full, and we
* have no information that the PTEs in the primary PTEG are
* more important or useful than those in the secondary PTEG,
* and we know there is a definite (although small) speed
* advantage to putting the PTE in the primary PTEG, we always
* put the PTE in the primary PTEG.
*
* In addition, we skip any slot that is mapping kernel text in
* order to avoid a deadlock when not using BAT mappings if
* trying to hash in the kernel hash code itself after it has
* already taken the hash table lock. This works in conjunction
* with pre-faulting of the kernel text.
*
* If the hash table bucket is full of kernel text entries, we'll
* lockup here but that shouldn't happen
*/
1: addis r4,r7,next_slot@ha /* get next evict slot */
lwz r6,next_slot@l(r4)
addi r6,r6,PTE_SIZE /* search for candidate */
andi. r6,r6,7*PTE_SIZE
stw r6,next_slot@l(r4)
add r4,r3,r6
LDPTE r0,PTE_SIZE/2(r4) /* get PTE second word */
clrrwi r0,r0,12
lis r6,etext@h
ori r6,r6,etext@l /* get etext */
tophys(r6,r6)
cmpl cr0,r0,r6 /* compare and try again */
blt 1b
#ifndef CONFIG_SMP
/* Store PTE in PTEG */
found_empty:
STPTE r5,0(r4)
found_slot:
STPTE r8,PTE_SIZE/2(r4)
#else /* CONFIG_SMP */
/*
* Between the tlbie above and updating the hash table entry below,
* another CPU could read the hash table entry and put it in its TLB.
* There are 3 cases:
* 1. using an empty slot
* 2. updating an earlier entry to change permissions (i.e. enable write)
* 3. taking over the PTE for an unrelated address
*
* In each case it doesn't really matter if the other CPUs have the old
* PTE in their TLB. So we don't need to bother with another tlbie here,
* which is convenient as we've overwritten the register that had the
* address. :-) The tlbie above is mainly to make sure that this CPU comes
* and gets the new PTE from the hash table.
*
* We do however have to make sure that the PTE is never in an invalid
* state with the V bit set.
*/
found_empty:
found_slot:
CLR_V(r5,r0) /* clear V (valid) bit in PTE */
STPTE r5,0(r4)
sync
TLBSYNC
STPTE r8,PTE_SIZE/2(r4) /* put in correct RPN, WIMG, PP bits */
sync
SET_V(r5)
STPTE r5,0(r4) /* finally set V bit in PTE */
#endif /* CONFIG_SMP */
sync /* make sure pte updates get to memory */
blr
.comm next_slot,4
.comm primary_pteg_full,4
.comm htab_hash_searches,4
/*
* Flush the entry for a particular page from the hash table.
*
* flush_hash_pages(unsigned context, unsigned long va, unsigned long pmdval,
* int count)
*
* We assume that there is a hash table in use (Hash != 0).
*/
_GLOBAL(flush_hash_pages)
tophys(r7,0)
/*
* We disable interrupts here, even on UP, because we want
* the _PAGE_HASHPTE bit to be a reliable indication of
* whether the HPTE exists (or at least whether one did once).
* We also turn off the MMU for data accesses so that we
* we can't take a hash table miss (assuming the code is
* covered by a BAT). -- paulus
*/
mfmsr r10
SYNC
rlwinm r0,r10,0,17,15 /* clear bit 16 (MSR_EE) */
rlwinm r0,r0,0,28,26 /* clear MSR_DR */
mtmsr r0
SYNC_601
isync
/* First find a PTE in the range that has _PAGE_HASHPTE set */
rlwimi r5,r4,22,20,29
1: lwz r0,0(r5)
cmpwi cr1,r6,1
andi. r0,r0,_PAGE_HASHPTE
bne 2f
ble cr1,19f
addi r4,r4,0x1000
addi r5,r5,4
addi r6,r6,-1
b 1b
/* Convert context and va to VSID */
2: mulli r3,r3,897*16 /* multiply context by context skew */
rlwinm r0,r4,4,28,31 /* get ESID (top 4 bits of va) */
mulli r0,r0,0x111 /* multiply by ESID skew */
add r3,r3,r0 /* note code below trims to 24 bits */
/* Construct the high word of the PPC-style PTE (r11) */
rlwinm r11,r3,7,1,24 /* put VSID in 0x7fffff80 bits */
rlwimi r11,r4,10,26,31 /* put in API (abbrev page index) */
SET_V(r11) /* set V (valid) bit */
#ifdef CONFIG_SMP
addis r9,r7,mmu_hash_lock@ha
addi r9,r9,mmu_hash_lock@l
rlwinm r8,r1,0,0,18
add r8,r8,r7
lwz r8,TI_CPU(r8)
oris r8,r8,9
10: lwarx r0,0,r9
cmpi 0,r0,0
bne- 11f
stwcx. r8,0,r9
beq+ 12f
11: lwz r0,0(r9)
cmpi 0,r0,0
beq 10b
b 11b
12: isync
#endif
/*
* Check the _PAGE_HASHPTE bit in the linux PTE. If it is
* already clear, we're done (for this pte). If not,
* clear it (atomically) and proceed. -- paulus.
*/
33: lwarx r8,0,r5 /* fetch the pte */
andi. r0,r8,_PAGE_HASHPTE
beq 8f /* done if HASHPTE is already clear */
rlwinm r8,r8,0,31,29 /* clear HASHPTE bit */
stwcx. r8,0,r5 /* update the pte */
bne- 33b
/* Get the address of the primary PTE group in the hash table (r3) */
_GLOBAL(flush_hash_patch_A)
addis r8,r7,Hash_base@h /* base address of hash table */
rlwimi r8,r3,LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* VSID -> hash */
rlwinm r0,r4,20+LG_PTEG_SIZE,HASH_LEFT,HASH_RIGHT /* PI -> hash */
xor r8,r0,r8 /* make primary hash */
/* Search the primary PTEG for a PTE whose 1st (d)word matches r5 */
li r0,8 /* PTEs/group */
mtctr r0
addi r12,r8,-PTE_SIZE
1: LDPTEu r0,PTE_SIZE(r12) /* get next PTE */
CMPPTE 0,r0,r11
bdnzf 2,1b /* loop while ctr != 0 && !cr0.eq */
beq+ 3f
/* Search the secondary PTEG for a matching PTE */
ori r11,r11,PTE_H /* set H (secondary hash) bit */
li r0,8 /* PTEs/group */
_GLOBAL(flush_hash_patch_B)
xoris r12,r8,Hash_msk>>16 /* compute secondary hash */
xori r12,r12,(-PTEG_SIZE & 0xffff)
addi r12,r12,-PTE_SIZE
mtctr r0
2: LDPTEu r0,PTE_SIZE(r12)
CMPPTE 0,r0,r11
bdnzf 2,2b
xori r11,r11,PTE_H /* clear H again */
bne- 4f /* should rarely fail to find it */
3: li r0,0
STPTE r0,0(r12) /* invalidate entry */
4: sync
tlbie r4 /* in hw tlb too */
sync
8: ble cr1,9f /* if all ptes checked */
81: addi r6,r6,-1
addi r5,r5,4 /* advance to next pte */
addi r4,r4,0x1000
lwz r0,0(r5) /* check next pte */
cmpwi cr1,r6,1
andi. r0,r0,_PAGE_HASHPTE
bne 33b
bgt cr1,81b
9:
#ifdef CONFIG_SMP
TLBSYNC
li r0,0
stw r0,0(r9) /* clear mmu_hash_lock */
#endif
19: mtmsr r10
SYNC_601
isync
blr