kernel-fxtec-pro1x/arch/sparc/mm/ultra.S

879 lines
20 KiB
ArmAsm
Raw Normal View History

/*
* ultra.S: Don't expand these all over the place...
*
* Copyright (C) 1997, 2000, 2008 David S. Miller (davem@davemloft.net)
*/
#include <asm/asi.h>
#include <asm/pgtable.h>
#include <asm/page.h>
#include <asm/spitfire.h>
#include <asm/mmu_context.h>
#include <asm/mmu.h>
#include <asm/pil.h>
#include <asm/head.h>
#include <asm/thread_info.h>
#include <asm/cacheflush.h>
#include <asm/hypervisor.h>
#include <asm/cpudata.h>
/* Basically, most of the Spitfire vs. Cheetah madness
* has to do with the fact that Cheetah does not support
* IMMU flushes out of the secondary context. Someone needs
* to throw a south lake birthday party for the folks
* in Microelectronics who refused to fix this shit.
*/
/* This file is meant to be read efficiently by the CPU, not humans.
* Staraj sie tego nikomu nie pierdolnac...
*/
.text
.align 32
.globl __flush_tlb_mm
__flush_tlb_mm: /* 18 insns */
/* %o0=(ctx & TAG_CONTEXT_BITS), %o1=SECONDARY_CONTEXT */
ldxa [%o1] ASI_DMMU, %g2
cmp %g2, %o0
bne,pn %icc, __spitfire_flush_tlb_mm_slow
mov 0x50, %g3
stxa %g0, [%g3] ASI_DMMU_DEMAP
stxa %g0, [%g3] ASI_IMMU_DEMAP
sethi %hi(KERNBASE), %g3
flush %g3
retl
nop
nop
nop
nop
nop
nop
nop
nop
nop
nop
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
.align 32
.globl __flush_tlb_page
__flush_tlb_page: /* 22 insns */
/* %o0 = context, %o1 = vaddr */
rdpr %pstate, %g7
andn %g7, PSTATE_IE, %g2
wrpr %g2, %pstate
mov SECONDARY_CONTEXT, %o4
ldxa [%o4] ASI_DMMU, %g2
stxa %o0, [%o4] ASI_DMMU
andcc %o1, 1, %g0
andn %o1, 1, %o3
be,pn %icc, 1f
or %o3, 0x10, %o3
stxa %g0, [%o3] ASI_IMMU_DEMAP
1: stxa %g0, [%o3] ASI_DMMU_DEMAP
membar #Sync
stxa %g2, [%o4] ASI_DMMU
sethi %hi(KERNBASE), %o4
flush %o4
retl
wrpr %g7, 0x0, %pstate
nop
nop
nop
nop
.align 32
.globl __flush_tlb_pending
__flush_tlb_pending: /* 26 insns */
/* %o0 = context, %o1 = nr, %o2 = vaddrs[] */
rdpr %pstate, %g7
sllx %o1, 3, %o1
andn %g7, PSTATE_IE, %g2
wrpr %g2, %pstate
mov SECONDARY_CONTEXT, %o4
ldxa [%o4] ASI_DMMU, %g2
stxa %o0, [%o4] ASI_DMMU
1: sub %o1, (1 << 3), %o1
ldx [%o2 + %o1], %o3
andcc %o3, 1, %g0
andn %o3, 1, %o3
be,pn %icc, 2f
or %o3, 0x10, %o3
stxa %g0, [%o3] ASI_IMMU_DEMAP
2: stxa %g0, [%o3] ASI_DMMU_DEMAP
membar #Sync
brnz,pt %o1, 1b
nop
stxa %g2, [%o4] ASI_DMMU
sethi %hi(KERNBASE), %o4
flush %o4
retl
wrpr %g7, 0x0, %pstate
nop
nop
nop
nop
.align 32
.globl __flush_tlb_kernel_range
__flush_tlb_kernel_range: /* 16 insns */
/* %o0=start, %o1=end */
cmp %o0, %o1
be,pn %xcc, 2f
sethi %hi(PAGE_SIZE), %o4
sub %o1, %o0, %o3
sub %o3, %o4, %o3
or %o0, 0x20, %o0 ! Nucleus
1: stxa %g0, [%o0 + %o3] ASI_DMMU_DEMAP
stxa %g0, [%o0 + %o3] ASI_IMMU_DEMAP
membar #Sync
brnz,pt %o3, 1b
sub %o3, %o4, %o3
2: sethi %hi(KERNBASE), %o3
flush %o3
retl
nop
nop
__spitfire_flush_tlb_mm_slow:
rdpr %pstate, %g1
wrpr %g1, PSTATE_IE, %pstate
stxa %o0, [%o1] ASI_DMMU
stxa %g0, [%g3] ASI_DMMU_DEMAP
stxa %g0, [%g3] ASI_IMMU_DEMAP
flush %g6
stxa %g2, [%o1] ASI_DMMU
sethi %hi(KERNBASE), %o1
flush %o1
retl
wrpr %g1, 0, %pstate
/*
* The following code flushes one page_size worth.
*/
.section .kprobes.text, "ax"
.align 32
.globl __flush_icache_page
__flush_icache_page: /* %o0 = phys_page */
srlx %o0, PAGE_SHIFT, %o0
sethi %hi(PAGE_OFFSET), %g1
sllx %o0, PAGE_SHIFT, %o0
sethi %hi(PAGE_SIZE), %g2
ldx [%g1 + %lo(PAGE_OFFSET)], %g1
add %o0, %g1, %o0
1: subcc %g2, 32, %g2
bne,pt %icc, 1b
flush %o0 + %g2
retl
nop
#ifdef DCACHE_ALIASING_POSSIBLE
#if (PAGE_SHIFT != 13)
#error only page shift of 13 is supported by dcache flush
#endif
#define DTAG_MASK 0x3
/* This routine is Spitfire specific so the hardcoded
* D-cache size and line-size are OK.
*/
.align 64
.globl __flush_dcache_page
__flush_dcache_page: /* %o0=kaddr, %o1=flush_icache */
sethi %hi(PAGE_OFFSET), %g1
ldx [%g1 + %lo(PAGE_OFFSET)], %g1
sub %o0, %g1, %o0 ! physical address
srlx %o0, 11, %o0 ! make D-cache TAG
sethi %hi(1 << 14), %o2 ! D-cache size
sub %o2, (1 << 5), %o2 ! D-cache line size
1: ldxa [%o2] ASI_DCACHE_TAG, %o3 ! load D-cache TAG
andcc %o3, DTAG_MASK, %g0 ! Valid?
be,pn %xcc, 2f ! Nope, branch
andn %o3, DTAG_MASK, %o3 ! Clear valid bits
cmp %o3, %o0 ! TAG match?
bne,pt %xcc, 2f ! Nope, branch
nop
stxa %g0, [%o2] ASI_DCACHE_TAG ! Invalidate TAG
membar #Sync
2: brnz,pt %o2, 1b
sub %o2, (1 << 5), %o2 ! D-cache line size
/* The I-cache does not snoop local stores so we
* better flush that too when necessary.
*/
brnz,pt %o1, __flush_icache_page
sllx %o0, 11, %o0
retl
nop
#endif /* DCACHE_ALIASING_POSSIBLE */
.previous
/* Cheetah specific versions, patched at boot time. */
__cheetah_flush_tlb_mm: /* 19 insns */
rdpr %pstate, %g7
andn %g7, PSTATE_IE, %g2
wrpr %g2, 0x0, %pstate
wrpr %g0, 1, %tl
mov PRIMARY_CONTEXT, %o2
mov 0x40, %g3
ldxa [%o2] ASI_DMMU, %g2
srlx %g2, CTX_PGSZ1_NUC_SHIFT, %o1
sllx %o1, CTX_PGSZ1_NUC_SHIFT, %o1
or %o0, %o1, %o0 /* Preserve nucleus page size fields */
stxa %o0, [%o2] ASI_DMMU
stxa %g0, [%g3] ASI_DMMU_DEMAP
stxa %g0, [%g3] ASI_IMMU_DEMAP
stxa %g2, [%o2] ASI_DMMU
sethi %hi(KERNBASE), %o2
flush %o2
wrpr %g0, 0, %tl
retl
wrpr %g7, 0x0, %pstate
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
__cheetah_flush_tlb_page: /* 22 insns */
/* %o0 = context, %o1 = vaddr */
rdpr %pstate, %g7
andn %g7, PSTATE_IE, %g2
wrpr %g2, 0x0, %pstate
wrpr %g0, 1, %tl
mov PRIMARY_CONTEXT, %o4
ldxa [%o4] ASI_DMMU, %g2
srlx %g2, CTX_PGSZ1_NUC_SHIFT, %o3
sllx %o3, CTX_PGSZ1_NUC_SHIFT, %o3
or %o0, %o3, %o0 /* Preserve nucleus page size fields */
stxa %o0, [%o4] ASI_DMMU
andcc %o1, 1, %g0
be,pn %icc, 1f
andn %o1, 1, %o3
stxa %g0, [%o3] ASI_IMMU_DEMAP
1: stxa %g0, [%o3] ASI_DMMU_DEMAP
membar #Sync
stxa %g2, [%o4] ASI_DMMU
sethi %hi(KERNBASE), %o4
flush %o4
wrpr %g0, 0, %tl
retl
wrpr %g7, 0x0, %pstate
__cheetah_flush_tlb_pending: /* 27 insns */
/* %o0 = context, %o1 = nr, %o2 = vaddrs[] */
rdpr %pstate, %g7
sllx %o1, 3, %o1
andn %g7, PSTATE_IE, %g2
wrpr %g2, 0x0, %pstate
wrpr %g0, 1, %tl
mov PRIMARY_CONTEXT, %o4
ldxa [%o4] ASI_DMMU, %g2
srlx %g2, CTX_PGSZ1_NUC_SHIFT, %o3
sllx %o3, CTX_PGSZ1_NUC_SHIFT, %o3
or %o0, %o3, %o0 /* Preserve nucleus page size fields */
stxa %o0, [%o4] ASI_DMMU
1: sub %o1, (1 << 3), %o1
ldx [%o2 + %o1], %o3
andcc %o3, 1, %g0
be,pn %icc, 2f
andn %o3, 1, %o3
stxa %g0, [%o3] ASI_IMMU_DEMAP
2: stxa %g0, [%o3] ASI_DMMU_DEMAP
membar #Sync
brnz,pt %o1, 1b
nop
stxa %g2, [%o4] ASI_DMMU
sethi %hi(KERNBASE), %o4
flush %o4
wrpr %g0, 0, %tl
retl
wrpr %g7, 0x0, %pstate
#ifdef DCACHE_ALIASING_POSSIBLE
__cheetah_flush_dcache_page: /* 11 insns */
sethi %hi(PAGE_OFFSET), %g1
ldx [%g1 + %lo(PAGE_OFFSET)], %g1
sub %o0, %g1, %o0
sethi %hi(PAGE_SIZE), %o4
1: subcc %o4, (1 << 5), %o4
stxa %g0, [%o0 + %o4] ASI_DCACHE_INVALIDATE
membar #Sync
bne,pt %icc, 1b
nop
retl /* I-cache flush never needed on Cheetah, see callers. */
nop
#endif /* DCACHE_ALIASING_POSSIBLE */
/* Hypervisor specific versions, patched at boot time. */
__hypervisor_tlb_tl0_error:
save %sp, -192, %sp
mov %i0, %o0
call hypervisor_tlbop_error
mov %i1, %o1
ret
restore
__hypervisor_flush_tlb_mm: /* 10 insns */
mov %o0, %o2 /* ARG2: mmu context */
mov 0, %o0 /* ARG0: CPU lists unimplemented */
mov 0, %o1 /* ARG1: CPU lists unimplemented */
mov HV_MMU_ALL, %o3 /* ARG3: flags */
mov HV_FAST_MMU_DEMAP_CTX, %o5
ta HV_FAST_TRAP
brnz,pn %o0, __hypervisor_tlb_tl0_error
mov HV_FAST_MMU_DEMAP_CTX, %o1
retl
nop
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
__hypervisor_flush_tlb_page: /* 11 insns */
/* %o0 = context, %o1 = vaddr */
mov %o0, %g2
mov %o1, %o0 /* ARG0: vaddr + IMMU-bit */
mov %g2, %o1 /* ARG1: mmu context */
mov HV_MMU_ALL, %o2 /* ARG2: flags */
srlx %o0, PAGE_SHIFT, %o0
sllx %o0, PAGE_SHIFT, %o0
ta HV_MMU_UNMAP_ADDR_TRAP
brnz,pn %o0, __hypervisor_tlb_tl0_error
mov HV_MMU_UNMAP_ADDR_TRAP, %o1
retl
nop
__hypervisor_flush_tlb_pending: /* 16 insns */
/* %o0 = context, %o1 = nr, %o2 = vaddrs[] */
sllx %o1, 3, %g1
mov %o2, %g2
mov %o0, %g3
1: sub %g1, (1 << 3), %g1
ldx [%g2 + %g1], %o0 /* ARG0: vaddr + IMMU-bit */
mov %g3, %o1 /* ARG1: mmu context */
mov HV_MMU_ALL, %o2 /* ARG2: flags */
srlx %o0, PAGE_SHIFT, %o0
sllx %o0, PAGE_SHIFT, %o0
ta HV_MMU_UNMAP_ADDR_TRAP
brnz,pn %o0, __hypervisor_tlb_tl0_error
mov HV_MMU_UNMAP_ADDR_TRAP, %o1
brnz,pt %g1, 1b
nop
retl
nop
__hypervisor_flush_tlb_kernel_range: /* 16 insns */
/* %o0=start, %o1=end */
cmp %o0, %o1
be,pn %xcc, 2f
sethi %hi(PAGE_SIZE), %g3
mov %o0, %g1
sub %o1, %g1, %g2
sub %g2, %g3, %g2
1: add %g1, %g2, %o0 /* ARG0: virtual address */
mov 0, %o1 /* ARG1: mmu context */
mov HV_MMU_ALL, %o2 /* ARG2: flags */
ta HV_MMU_UNMAP_ADDR_TRAP
brnz,pn %o0, __hypervisor_tlb_tl0_error
mov HV_MMU_UNMAP_ADDR_TRAP, %o1
brnz,pt %g2, 1b
sub %g2, %g3, %g2
2: retl
nop
#ifdef DCACHE_ALIASING_POSSIBLE
/* XXX Niagara and friends have an 8K cache, so no aliasing is
* XXX possible, but nothing explicit in the Hypervisor API
* XXX guarantees this.
*/
__hypervisor_flush_dcache_page: /* 2 insns */
retl
nop
#endif
tlb_patch_one:
1: lduw [%o1], %g1
stw %g1, [%o0]
flush %o0
subcc %o2, 1, %o2
add %o1, 4, %o1
bne,pt %icc, 1b
add %o0, 4, %o0
retl
nop
.globl cheetah_patch_cachetlbops
cheetah_patch_cachetlbops:
save %sp, -128, %sp
sethi %hi(__flush_tlb_mm), %o0
or %o0, %lo(__flush_tlb_mm), %o0
sethi %hi(__cheetah_flush_tlb_mm), %o1
or %o1, %lo(__cheetah_flush_tlb_mm), %o1
call tlb_patch_one
mov 19, %o2
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
sethi %hi(__flush_tlb_page), %o0
or %o0, %lo(__flush_tlb_page), %o0
sethi %hi(__cheetah_flush_tlb_page), %o1
or %o1, %lo(__cheetah_flush_tlb_page), %o1
call tlb_patch_one
mov 22, %o2
sethi %hi(__flush_tlb_pending), %o0
or %o0, %lo(__flush_tlb_pending), %o0
sethi %hi(__cheetah_flush_tlb_pending), %o1
or %o1, %lo(__cheetah_flush_tlb_pending), %o1
call tlb_patch_one
mov 27, %o2
#ifdef DCACHE_ALIASING_POSSIBLE
sethi %hi(__flush_dcache_page), %o0
or %o0, %lo(__flush_dcache_page), %o0
sethi %hi(__cheetah_flush_dcache_page), %o1
or %o1, %lo(__cheetah_flush_dcache_page), %o1
call tlb_patch_one
mov 11, %o2
#endif /* DCACHE_ALIASING_POSSIBLE */
ret
restore
#ifdef CONFIG_SMP
/* These are all called by the slaves of a cross call, at
* trap level 1, with interrupts fully disabled.
*
* Register usage:
* %g5 mm->context (all tlb flushes)
* %g1 address arg 1 (tlb page and range flushes)
* %g7 address arg 2 (tlb range flush only)
*
[SPARC64]: Elminate all usage of hard-coded trap globals. UltraSPARC has special sets of global registers which are switched to for certain trap types. There is one set for MMU related traps, one set of Interrupt Vector processing, and another set (called the Alternate globals) for all other trap types. For what seems like forever we've hard coded the values in some of these trap registers. Some examples include: 1) Interrupt Vector global %g6 holds current processors interrupt work struct where received interrupts are managed for IRQ handler dispatch. 2) MMU global %g7 holds the base of the page tables of the currently active address space. 3) Alternate global %g6 held the current_thread_info() value. Such hardcoding has resulted in some serious issues in many areas. There are some code sequences where having another register available would help clean up the implementation. Taking traps such as cross-calls from the OBP firmware requires some trick code sequences wherein we have to save away and restore all of the special sets of global registers when we enter/exit OBP. We were also using the IMMU TSB register on SMP to hold the per-cpu area base address, which doesn't work any longer now that we actually use the TSB facility of the cpu. The implementation is pretty straight forward. One tricky bit is getting the current processor ID as that is different on different cpu variants. We use a stub with a fancy calling convention which we patch at boot time. The calling convention is that the stub is branched to and the (PC - 4) to return to is in register %g1. The cpu number is left in %g6. This stub can be invoked by using the __GET_CPUID macro. We use an array of per-cpu trap state to store the current thread and physical address of the current address space's page tables. The TRAP_LOAD_THREAD_REG loads %g6 with the current thread from this table, it uses __GET_CPUID and also clobbers %g1. TRAP_LOAD_IRQ_WORK is used by the interrupt vector processing to load the current processor's IRQ software state into %g6. It also uses __GET_CPUID and clobbers %g1. Finally, TRAP_LOAD_PGD_PHYS loads the physical address base of the current address space's page tables into %g7, it clobbers %g1 and uses __GET_CPUID. Many refinements are possible, as well as some tuning, with this stuff in place. Signed-off-by: David S. Miller <davem@davemloft.net>
2006-02-27 00:24:22 -07:00
* %g6 scratch 1
* %g2 scratch 2
* %g3 scratch 3
* %g4 scratch 4
*/
.align 32
.globl xcall_flush_tlb_mm
xcall_flush_tlb_mm: /* 21 insns */
mov PRIMARY_CONTEXT, %g2
ldxa [%g2] ASI_DMMU, %g3
srlx %g3, CTX_PGSZ1_NUC_SHIFT, %g4
sllx %g4, CTX_PGSZ1_NUC_SHIFT, %g4
or %g5, %g4, %g5 /* Preserve nucleus page size fields */
stxa %g5, [%g2] ASI_DMMU
mov 0x40, %g4
stxa %g0, [%g4] ASI_DMMU_DEMAP
stxa %g0, [%g4] ASI_IMMU_DEMAP
stxa %g3, [%g2] ASI_DMMU
retry
nop
nop
nop
nop
nop
nop
nop
nop
nop
nop
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
.globl xcall_flush_tlb_page
xcall_flush_tlb_page: /* 17 insns */
/* %g5=context, %g1=vaddr */
mov PRIMARY_CONTEXT, %g4
ldxa [%g4] ASI_DMMU, %g2
srlx %g2, CTX_PGSZ1_NUC_SHIFT, %g4
sllx %g4, CTX_PGSZ1_NUC_SHIFT, %g4
or %g5, %g4, %g5
mov PRIMARY_CONTEXT, %g4
stxa %g5, [%g4] ASI_DMMU
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
andcc %g1, 0x1, %g0
be,pn %icc, 2f
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
andn %g1, 0x1, %g5
stxa %g0, [%g5] ASI_IMMU_DEMAP
2: stxa %g0, [%g5] ASI_DMMU_DEMAP
membar #Sync
stxa %g2, [%g4] ASI_DMMU
retry
nop
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
nop
.globl xcall_flush_tlb_kernel_range
xcall_flush_tlb_kernel_range: /* 25 insns */
sethi %hi(PAGE_SIZE - 1), %g2
or %g2, %lo(PAGE_SIZE - 1), %g2
andn %g1, %g2, %g1
andn %g7, %g2, %g7
sub %g7, %g1, %g3
add %g2, 1, %g2
sub %g3, %g2, %g3
or %g1, 0x20, %g1 ! Nucleus
1: stxa %g0, [%g1 + %g3] ASI_DMMU_DEMAP
stxa %g0, [%g1 + %g3] ASI_IMMU_DEMAP
membar #Sync
brnz,pt %g3, 1b
sub %g3, %g2, %g3
retry
nop
nop
nop
nop
nop
nop
nop
nop
nop
nop
nop
/* This runs in a very controlled environment, so we do
* not need to worry about BH races etc.
*/
.globl xcall_sync_tick
xcall_sync_tick:
661: rdpr %pstate, %g2
wrpr %g2, PSTATE_IG | PSTATE_AG, %pstate
.section .sun4v_2insn_patch, "ax"
.word 661b
nop
nop
.previous
rdpr %pil, %g2
wrpr %g0, PIL_NORMAL_MAX, %pil
sethi %hi(109f), %g7
b,pt %xcc, etrap_irq
109: or %g7, %lo(109b), %g7
#ifdef CONFIG_TRACE_IRQFLAGS
call trace_hardirqs_off
nop
#endif
call smp_synchronize_tick_client
nop
b rtrap_xcall
ldx [%sp + PTREGS_OFF + PT_V9_TSTATE], %l1
.globl xcall_fetch_glob_regs
xcall_fetch_glob_regs:
sethi %hi(global_cpu_snapshot), %g1
or %g1, %lo(global_cpu_snapshot), %g1
__GET_CPUID(%g2)
sllx %g2, 6, %g3
add %g1, %g3, %g1
rdpr %tstate, %g7
stx %g7, [%g1 + GR_SNAP_TSTATE]
rdpr %tpc, %g7
stx %g7, [%g1 + GR_SNAP_TPC]
rdpr %tnpc, %g7
stx %g7, [%g1 + GR_SNAP_TNPC]
stx %o7, [%g1 + GR_SNAP_O7]
stx %i7, [%g1 + GR_SNAP_I7]
/* Don't try this at home kids... */
rdpr %cwp, %g3
sub %g3, 1, %g7
wrpr %g7, %cwp
mov %i7, %g7
wrpr %g3, %cwp
stx %g7, [%g1 + GR_SNAP_RPC]
sethi %hi(trap_block), %g7
or %g7, %lo(trap_block), %g7
sllx %g2, TRAP_BLOCK_SZ_SHIFT, %g2
add %g7, %g2, %g7
ldx [%g7 + TRAP_PER_CPU_THREAD], %g3
stx %g3, [%g1 + GR_SNAP_THREAD]
retry
.globl xcall_fetch_glob_pmu
xcall_fetch_glob_pmu:
sethi %hi(global_cpu_snapshot), %g1
or %g1, %lo(global_cpu_snapshot), %g1
__GET_CPUID(%g2)
sllx %g2, 6, %g3
add %g1, %g3, %g1
rd %pic, %g7
stx %g7, [%g1 + (4 * 8)]
rd %pcr, %g7
stx %g7, [%g1 + (0 * 8)]
retry
.globl xcall_fetch_glob_pmu_n4
xcall_fetch_glob_pmu_n4:
sethi %hi(global_cpu_snapshot), %g1
or %g1, %lo(global_cpu_snapshot), %g1
__GET_CPUID(%g2)
sllx %g2, 6, %g3
add %g1, %g3, %g1
ldxa [%g0] ASI_PIC, %g7
stx %g7, [%g1 + (4 * 8)]
mov 0x08, %g3
ldxa [%g3] ASI_PIC, %g7
stx %g7, [%g1 + (5 * 8)]
mov 0x10, %g3
ldxa [%g3] ASI_PIC, %g7
stx %g7, [%g1 + (6 * 8)]
mov 0x18, %g3
ldxa [%g3] ASI_PIC, %g7
stx %g7, [%g1 + (7 * 8)]
mov %o0, %g2
mov %o1, %g3
mov %o5, %g7
mov HV_FAST_VT_GET_PERFREG, %o5
mov 3, %o0
ta HV_FAST_TRAP
stx %o1, [%g1 + (3 * 8)]
mov HV_FAST_VT_GET_PERFREG, %o5
mov 2, %o0
ta HV_FAST_TRAP
stx %o1, [%g1 + (2 * 8)]
mov HV_FAST_VT_GET_PERFREG, %o5
mov 1, %o0
ta HV_FAST_TRAP
stx %o1, [%g1 + (1 * 8)]
mov HV_FAST_VT_GET_PERFREG, %o5
mov 0, %o0
ta HV_FAST_TRAP
stx %o1, [%g1 + (0 * 8)]
mov %g2, %o0
mov %g3, %o1
mov %g7, %o5
retry
#ifdef DCACHE_ALIASING_POSSIBLE
.align 32
.globl xcall_flush_dcache_page_cheetah
xcall_flush_dcache_page_cheetah: /* %g1 == physical page address */
sethi %hi(PAGE_SIZE), %g3
1: subcc %g3, (1 << 5), %g3
stxa %g0, [%g1 + %g3] ASI_DCACHE_INVALIDATE
membar #Sync
bne,pt %icc, 1b
nop
retry
nop
#endif /* DCACHE_ALIASING_POSSIBLE */
.globl xcall_flush_dcache_page_spitfire
xcall_flush_dcache_page_spitfire: /* %g1 == physical page address
%g7 == kernel page virtual address
%g5 == (page->mapping != NULL) */
#ifdef DCACHE_ALIASING_POSSIBLE
srlx %g1, (13 - 2), %g1 ! Form tag comparitor
sethi %hi(L1DCACHE_SIZE), %g3 ! D$ size == 16K
sub %g3, (1 << 5), %g3 ! D$ linesize == 32
1: ldxa [%g3] ASI_DCACHE_TAG, %g2
andcc %g2, 0x3, %g0
be,pn %xcc, 2f
andn %g2, 0x3, %g2
cmp %g2, %g1
bne,pt %xcc, 2f
nop
stxa %g0, [%g3] ASI_DCACHE_TAG
membar #Sync
2: cmp %g3, 0
bne,pt %xcc, 1b
sub %g3, (1 << 5), %g3
brz,pn %g5, 2f
#endif /* DCACHE_ALIASING_POSSIBLE */
sethi %hi(PAGE_SIZE), %g3
1: flush %g7
subcc %g3, (1 << 5), %g3
bne,pt %icc, 1b
add %g7, (1 << 5), %g7
2: retry
nop
nop
/* %g5: error
* %g6: tlb op
*/
__hypervisor_tlb_xcall_error:
mov %g5, %g4
mov %g6, %g5
ba,pt %xcc, etrap
rd %pc, %g7
mov %l4, %o0
call hypervisor_tlbop_error_xcall
mov %l5, %o1
ba,a,pt %xcc, rtrap
.globl __hypervisor_xcall_flush_tlb_mm
__hypervisor_xcall_flush_tlb_mm: /* 21 insns */
/* %g5=ctx, g1,g2,g3,g4,g7=scratch, %g6=unusable */
mov %o0, %g2
mov %o1, %g3
mov %o2, %g4
mov %o3, %g1
mov %o5, %g7
clr %o0 /* ARG0: CPU lists unimplemented */
clr %o1 /* ARG1: CPU lists unimplemented */
mov %g5, %o2 /* ARG2: mmu context */
mov HV_MMU_ALL, %o3 /* ARG3: flags */
mov HV_FAST_MMU_DEMAP_CTX, %o5
ta HV_FAST_TRAP
mov HV_FAST_MMU_DEMAP_CTX, %g6
brnz,pn %o0, __hypervisor_tlb_xcall_error
mov %o0, %g5
mov %g2, %o0
mov %g3, %o1
mov %g4, %o2
mov %g1, %o3
mov %g7, %o5
membar #Sync
retry
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
.globl __hypervisor_xcall_flush_tlb_page
__hypervisor_xcall_flush_tlb_page: /* 17 insns */
/* %g5=ctx, %g1=vaddr */
mov %o0, %g2
mov %o1, %g3
mov %o2, %g4
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
mov %g1, %o0 /* ARG0: virtual address */
mov %g5, %o1 /* ARG1: mmu context */
mov HV_MMU_ALL, %o2 /* ARG2: flags */
srlx %o0, PAGE_SHIFT, %o0
sllx %o0, PAGE_SHIFT, %o0
ta HV_MMU_UNMAP_ADDR_TRAP
mov HV_MMU_UNMAP_ADDR_TRAP, %g6
brnz,a,pn %o0, __hypervisor_tlb_xcall_error
mov %o0, %g5
mov %g2, %o0
mov %g3, %o1
mov %g4, %o2
membar #Sync
retry
.globl __hypervisor_xcall_flush_tlb_kernel_range
__hypervisor_xcall_flush_tlb_kernel_range: /* 25 insns */
/* %g1=start, %g7=end, g2,g3,g4,g5,g6=scratch */
sethi %hi(PAGE_SIZE - 1), %g2
or %g2, %lo(PAGE_SIZE - 1), %g2
andn %g1, %g2, %g1
andn %g7, %g2, %g7
sub %g7, %g1, %g3
add %g2, 1, %g2
sub %g3, %g2, %g3
mov %o0, %g2
mov %o1, %g4
mov %o2, %g7
1: add %g1, %g3, %o0 /* ARG0: virtual address */
mov 0, %o1 /* ARG1: mmu context */
mov HV_MMU_ALL, %o2 /* ARG2: flags */
ta HV_MMU_UNMAP_ADDR_TRAP
mov HV_MMU_UNMAP_ADDR_TRAP, %g6
brnz,pn %o0, __hypervisor_tlb_xcall_error
mov %o0, %g5
sethi %hi(PAGE_SIZE), %o2
brnz,pt %g3, 1b
sub %g3, %o2, %g3
mov %g2, %o0
mov %g4, %o1
mov %g7, %o2
membar #Sync
retry
/* These just get rescheduled to PIL vectors. */
.globl xcall_call_function
xcall_call_function:
wr %g0, (1 << PIL_SMP_CALL_FUNC), %set_softint
retry
.globl xcall_call_function_single
xcall_call_function_single:
wr %g0, (1 << PIL_SMP_CALL_FUNC_SNGL), %set_softint
retry
.globl xcall_receive_signal
xcall_receive_signal:
wr %g0, (1 << PIL_SMP_RECEIVE_SIGNAL), %set_softint
retry
.globl xcall_capture
xcall_capture:
wr %g0, (1 << PIL_SMP_CAPTURE), %set_softint
retry
.globl xcall_new_mmu_context_version
xcall_new_mmu_context_version:
wr %g0, (1 << PIL_SMP_CTX_NEW_VERSION), %set_softint
retry
#ifdef CONFIG_KGDB
.globl xcall_kgdb_capture
xcall_kgdb_capture:
wr %g0, (1 << PIL_KGDB_CAPTURE), %set_softint
retry
#endif
#endif /* CONFIG_SMP */
.globl hypervisor_patch_cachetlbops
hypervisor_patch_cachetlbops:
save %sp, -128, %sp
sethi %hi(__flush_tlb_mm), %o0
or %o0, %lo(__flush_tlb_mm), %o0
sethi %hi(__hypervisor_flush_tlb_mm), %o1
or %o1, %lo(__hypervisor_flush_tlb_mm), %o1
call tlb_patch_one
mov 10, %o2
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
sethi %hi(__flush_tlb_page), %o0
or %o0, %lo(__flush_tlb_page), %o0
sethi %hi(__hypervisor_flush_tlb_page), %o1
or %o1, %lo(__hypervisor_flush_tlb_page), %o1
call tlb_patch_one
mov 11, %o2
sethi %hi(__flush_tlb_pending), %o0
or %o0, %lo(__flush_tlb_pending), %o0
sethi %hi(__hypervisor_flush_tlb_pending), %o1
or %o1, %lo(__hypervisor_flush_tlb_pending), %o1
call tlb_patch_one
mov 16, %o2
sethi %hi(__flush_tlb_kernel_range), %o0
or %o0, %lo(__flush_tlb_kernel_range), %o0
sethi %hi(__hypervisor_flush_tlb_kernel_range), %o1
or %o1, %lo(__hypervisor_flush_tlb_kernel_range), %o1
call tlb_patch_one
mov 16, %o2
#ifdef DCACHE_ALIASING_POSSIBLE
sethi %hi(__flush_dcache_page), %o0
or %o0, %lo(__flush_dcache_page), %o0
sethi %hi(__hypervisor_flush_dcache_page), %o1
or %o1, %lo(__hypervisor_flush_dcache_page), %o1
call tlb_patch_one
mov 2, %o2
#endif /* DCACHE_ALIASING_POSSIBLE */
#ifdef CONFIG_SMP
sethi %hi(xcall_flush_tlb_mm), %o0
or %o0, %lo(xcall_flush_tlb_mm), %o0
sethi %hi(__hypervisor_xcall_flush_tlb_mm), %o1
or %o1, %lo(__hypervisor_xcall_flush_tlb_mm), %o1
call tlb_patch_one
mov 21, %o2
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
sethi %hi(xcall_flush_tlb_page), %o0
or %o0, %lo(xcall_flush_tlb_page), %o0
sethi %hi(__hypervisor_xcall_flush_tlb_page), %o1
or %o1, %lo(__hypervisor_xcall_flush_tlb_page), %o1
call tlb_patch_one
sparc64: Fix race in TLB batch processing. As reported by Dave Kleikamp, when we emit cross calls to do batched TLB flush processing we have a race because we do not synchronize on the sibling cpus completing the cross call. So meanwhile the TLB batch can be reset (tb->tlb_nr set to zero, etc.) and either flushes are missed or flushes will flush the wrong addresses. Fix this by using generic infrastructure to synchonize on the completion of the cross call. This first required getting the flush_tlb_pending() call out from switch_to() which operates with locks held and interrupts disabled. The problem is that smp_call_function_many() cannot be invoked with IRQs disabled and this is explicitly checked for with WARN_ON_ONCE(). We get the batch processing outside of locked IRQ disabled sections by using some ideas from the powerpc port. Namely, we only batch inside of arch_{enter,leave}_lazy_mmu_mode() calls. If we're not in such a region, we flush TLBs synchronously. 1) Get rid of xcall_flush_tlb_pending and per-cpu type implementations. 2) Do TLB batch cross calls instead via: smp_call_function_many() tlb_pending_func() __flush_tlb_pending() 3) Batch only in lazy mmu sequences: a) Add 'active' member to struct tlb_batch b) Define __HAVE_ARCH_ENTER_LAZY_MMU_MODE c) Set 'active' in arch_enter_lazy_mmu_mode() d) Run batch and clear 'active' in arch_leave_lazy_mmu_mode() e) Check 'active' in tlb_batch_add_one() and do a synchronous flush if it's clear. 4) Add infrastructure for synchronous TLB page flushes. a) Implement __flush_tlb_page and per-cpu variants, patch as needed. b) Likewise for xcall_flush_tlb_page. c) Implement smp_flush_tlb_page() to invoke the cross-call. d) Wire up global_flush_tlb_page() to the right routine based upon CONFIG_SMP 5) It turns out that singleton batches are very common, 2 out of every 3 batch flushes have only a single entry in them. The batch flush waiting is very expensive, both because of the poll on sibling cpu completeion, as well as because passing the tlb batch pointer to the sibling cpus invokes a shared memory dereference. Therefore, in flush_tlb_pending(), if there is only one entry in the batch perform a completely asynchronous global_flush_tlb_page() instead. Reported-by: Dave Kleikamp <dave.kleikamp@oracle.com> Signed-off-by: David S. Miller <davem@davemloft.net> Acked-by: Dave Kleikamp <dave.kleikamp@oracle.com>
2013-04-19 15:26:26 -06:00
mov 17, %o2
sethi %hi(xcall_flush_tlb_kernel_range), %o0
or %o0, %lo(xcall_flush_tlb_kernel_range), %o0
sethi %hi(__hypervisor_xcall_flush_tlb_kernel_range), %o1
or %o1, %lo(__hypervisor_xcall_flush_tlb_kernel_range), %o1
call tlb_patch_one
mov 25, %o2
#endif /* CONFIG_SMP */
ret
restore