diff options
Diffstat (limited to 'include/asm-sparc64/mmu_context.h')
| -rw-r--r-- | include/asm-sparc64/mmu_context.h | 162 |
1 files changed, 81 insertions, 81 deletions
diff --git a/include/asm-sparc64/mmu_context.h b/include/asm-sparc64/mmu_context.h index 57ee7b306189..e7974321d052 100644 --- a/include/asm-sparc64/mmu_context.h +++ b/include/asm-sparc64/mmu_context.h | |||
| @@ -19,96 +19,98 @@ extern unsigned long tlb_context_cache; | |||
| 19 | extern unsigned long mmu_context_bmap[]; | 19 | extern unsigned long mmu_context_bmap[]; |
| 20 | 20 | ||
| 21 | extern void get_new_mmu_context(struct mm_struct *mm); | 21 | extern void get_new_mmu_context(struct mm_struct *mm); |
| 22 | #ifdef CONFIG_SMP | ||
| 23 | extern void smp_new_mmu_context_version(void); | ||
| 24 | #else | ||
| 25 | #define smp_new_mmu_context_version() do { } while (0) | ||
| 26 | #endif | ||
| 27 | |||
| 28 | extern int init_new_context(struct task_struct *tsk, struct mm_struct *mm); | ||
| 29 | extern void destroy_context(struct mm_struct *mm); | ||
| 30 | |||
| 31 | extern void __tsb_context_switch(unsigned long pgd_pa, | ||
| 32 | unsigned long tsb_reg, | ||
| 33 | unsigned long tsb_vaddr, | ||
| 34 | unsigned long tsb_pte, | ||
| 35 | unsigned long tsb_descr_pa); | ||
| 36 | |||
| 37 | static inline void tsb_context_switch(struct mm_struct *mm) | ||
| 38 | { | ||
| 39 | __tsb_context_switch(__pa(mm->pgd), mm->context.tsb_reg_val, | ||
| 40 | mm->context.tsb_map_vaddr, | ||
| 41 | mm->context.tsb_map_pte, | ||
| 42 | __pa(&mm->context.tsb_descr)); | ||
| 43 | } | ||
| 22 | 44 | ||
| 23 | /* Initialize a new mmu context. This is invoked when a new | 45 | extern void tsb_grow(struct mm_struct *mm, unsigned long mm_rss); |
| 24 | * address space instance (unique or shared) is instantiated. | 46 | #ifdef CONFIG_SMP |
| 25 | * This just needs to set mm->context to an invalid context. | 47 | extern void smp_tsb_sync(struct mm_struct *mm); |
| 26 | */ | 48 | #else |
| 27 | #define init_new_context(__tsk, __mm) \ | 49 | #define smp_tsb_sync(__mm) do { } while (0) |
| 28 | (((__mm)->context.sparc64_ctx_val = 0UL), 0) | 50 | #endif |
| 29 | |||
| 30 | /* Destroy a dead context. This occurs when mmput drops the | ||
| 31 | * mm_users count to zero, the mmaps have been released, and | ||
| 32 | * all the page tables have been flushed. Our job is to destroy | ||
| 33 | * any remaining processor-specific state, and in the sparc64 | ||
| 34 | * case this just means freeing up the mmu context ID held by | ||
| 35 | * this task if valid. | ||
| 36 | */ | ||
| 37 | #define destroy_context(__mm) \ | ||
| 38 | do { spin_lock(&ctx_alloc_lock); \ | ||
| 39 | if (CTX_VALID((__mm)->context)) { \ | ||
| 40 | unsigned long nr = CTX_NRBITS((__mm)->context); \ | ||
| 41 | mmu_context_bmap[nr>>6] &= ~(1UL << (nr & 63)); \ | ||
| 42 | } \ | ||
| 43 | spin_unlock(&ctx_alloc_lock); \ | ||
| 44 | } while(0) | ||
| 45 | |||
| 46 | /* Reload the two core values used by TLB miss handler | ||
| 47 | * processing on sparc64. They are: | ||
| 48 | * 1) The physical address of mm->pgd, when full page | ||
| 49 | * table walks are necessary, this is where the | ||
| 50 | * search begins. | ||
| 51 | * 2) A "PGD cache". For 32-bit tasks only pgd[0] is | ||
| 52 | * ever used since that maps the entire low 4GB | ||
| 53 | * completely. To speed up TLB miss processing we | ||
| 54 | * make this value available to the handlers. This | ||
| 55 | * decreases the amount of memory traffic incurred. | ||
| 56 | */ | ||
| 57 | #define reload_tlbmiss_state(__tsk, __mm) \ | ||
| 58 | do { \ | ||
| 59 | register unsigned long paddr asm("o5"); \ | ||
| 60 | register unsigned long pgd_cache asm("o4"); \ | ||
| 61 | paddr = __pa((__mm)->pgd); \ | ||
| 62 | pgd_cache = 0UL; \ | ||
| 63 | if (task_thread_info(__tsk)->flags & _TIF_32BIT) \ | ||
| 64 | pgd_cache = get_pgd_cache((__mm)->pgd); \ | ||
| 65 | __asm__ __volatile__("wrpr %%g0, 0x494, %%pstate\n\t" \ | ||
| 66 | "mov %3, %%g4\n\t" \ | ||
| 67 | "mov %0, %%g7\n\t" \ | ||
| 68 | "stxa %1, [%%g4] %2\n\t" \ | ||
| 69 | "membar #Sync\n\t" \ | ||
| 70 | "wrpr %%g0, 0x096, %%pstate" \ | ||
| 71 | : /* no outputs */ \ | ||
| 72 | : "r" (paddr), "r" (pgd_cache),\ | ||
| 73 | "i" (ASI_DMMU), "i" (TSB_REG)); \ | ||
| 74 | } while(0) | ||
| 75 | 51 | ||
| 76 | /* Set MMU context in the actual hardware. */ | 52 | /* Set MMU context in the actual hardware. */ |
| 77 | #define load_secondary_context(__mm) \ | 53 | #define load_secondary_context(__mm) \ |
| 78 | __asm__ __volatile__("stxa %0, [%1] %2\n\t" \ | 54 | __asm__ __volatile__( \ |
| 79 | "flush %%g6" \ | 55 | "\n661: stxa %0, [%1] %2\n" \ |
| 80 | : /* No outputs */ \ | 56 | " .section .sun4v_1insn_patch, \"ax\"\n" \ |
| 81 | : "r" (CTX_HWBITS((__mm)->context)), \ | 57 | " .word 661b\n" \ |
| 82 | "r" (SECONDARY_CONTEXT), "i" (ASI_DMMU)) | 58 | " stxa %0, [%1] %3\n" \ |
| 59 | " .previous\n" \ | ||
| 60 | " flush %%g6\n" \ | ||
| 61 | : /* No outputs */ \ | ||
| 62 | : "r" (CTX_HWBITS((__mm)->context)), \ | ||
| 63 | "r" (SECONDARY_CONTEXT), "i" (ASI_DMMU), "i" (ASI_MMU)) | ||
| 83 | 64 | ||
| 84 | extern void __flush_tlb_mm(unsigned long, unsigned long); | 65 | extern void __flush_tlb_mm(unsigned long, unsigned long); |
| 85 | 66 | ||
| 86 | /* Switch the current MM context. */ | 67 | /* Switch the current MM context. Interrupts are disabled. */ |
| 87 | static inline void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm, struct task_struct *tsk) | 68 | static inline void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm, struct task_struct *tsk) |
| 88 | { | 69 | { |
| 89 | unsigned long ctx_valid; | 70 | unsigned long ctx_valid, flags; |
| 90 | int cpu; | 71 | int cpu; |
| 91 | 72 | ||
| 92 | /* Note: page_table_lock is used here to serialize switch_mm | 73 | spin_lock_irqsave(&mm->context.lock, flags); |
| 93 | * and activate_mm, and their calls to get_new_mmu_context. | ||
| 94 | * This use of page_table_lock is unrelated to its other uses. | ||
| 95 | */ | ||
| 96 | spin_lock(&mm->page_table_lock); | ||
| 97 | ctx_valid = CTX_VALID(mm->context); | 74 | ctx_valid = CTX_VALID(mm->context); |
| 98 | if (!ctx_valid) | 75 | if (!ctx_valid) |
| 99 | get_new_mmu_context(mm); | 76 | get_new_mmu_context(mm); |
| 100 | spin_unlock(&mm->page_table_lock); | ||
| 101 | 77 | ||
| 102 | if (!ctx_valid || (old_mm != mm)) { | 78 | /* We have to be extremely careful here or else we will miss |
| 103 | load_secondary_context(mm); | 79 | * a TSB grow if we switch back and forth between a kernel |
| 104 | reload_tlbmiss_state(tsk, mm); | 80 | * thread and an address space which has it's TSB size increased |
| 105 | } | 81 | * on another processor. |
| 82 | * | ||
| 83 | * It is possible to play some games in order to optimize the | ||
| 84 | * switch, but the safest thing to do is to unconditionally | ||
| 85 | * perform the secondary context load and the TSB context switch. | ||
| 86 | * | ||
| 87 | * For reference the bad case is, for address space "A": | ||
| 88 | * | ||
| 89 | * CPU 0 CPU 1 | ||
| 90 | * run address space A | ||
| 91 | * set cpu0's bits in cpu_vm_mask | ||
| 92 | * switch to kernel thread, borrow | ||
| 93 | * address space A via entry_lazy_tlb | ||
| 94 | * run address space A | ||
| 95 | * set cpu1's bit in cpu_vm_mask | ||
| 96 | * flush_tlb_pending() | ||
| 97 | * reset cpu_vm_mask to just cpu1 | ||
| 98 | * TSB grow | ||
| 99 | * run address space A | ||
| 100 | * context was valid, so skip | ||
| 101 | * TSB context switch | ||
| 102 | * | ||
| 103 | * At that point cpu0 continues to use a stale TSB, the one from | ||
| 104 | * before the TSB grow performed on cpu1. cpu1 did not cross-call | ||
| 105 | * cpu0 to update it's TSB because at that point the cpu_vm_mask | ||
| 106 | * only had cpu1 set in it. | ||
| 107 | */ | ||
| 108 | load_secondary_context(mm); | ||
| 109 | tsb_context_switch(mm); | ||
| 106 | 110 | ||
| 107 | /* Even if (mm == old_mm) we _must_ check | 111 | /* Any time a processor runs a context on an address space |
| 108 | * the cpu_vm_mask. If we do not we could | 112 | * for the first time, we must flush that context out of the |
| 109 | * corrupt the TLB state because of how | 113 | * local TLB. |
| 110 | * smp_flush_tlb_{page,range,mm} on sparc64 | ||
| 111 | * and lazy tlb switches work. -DaveM | ||
| 112 | */ | 114 | */ |
| 113 | cpu = smp_processor_id(); | 115 | cpu = smp_processor_id(); |
| 114 | if (!ctx_valid || !cpu_isset(cpu, mm->cpu_vm_mask)) { | 116 | if (!ctx_valid || !cpu_isset(cpu, mm->cpu_vm_mask)) { |
| @@ -116,6 +118,7 @@ static inline void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm, str | |||
| 116 | __flush_tlb_mm(CTX_HWBITS(mm->context), | 118 | __flush_tlb_mm(CTX_HWBITS(mm->context), |
| 117 | SECONDARY_CONTEXT); | 119 | SECONDARY_CONTEXT); |
| 118 | } | 120 | } |
| 121 | spin_unlock_irqrestore(&mm->context.lock, flags); | ||
| 119 | } | 122 | } |
| 120 | 123 | ||
| 121 | #define deactivate_mm(tsk,mm) do { } while (0) | 124 | #define deactivate_mm(tsk,mm) do { } while (0) |
| @@ -123,23 +126,20 @@ static inline void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm, str | |||
| 123 | /* Activate a new MM instance for the current task. */ | 126 | /* Activate a new MM instance for the current task. */ |
| 124 | static inline void activate_mm(struct mm_struct *active_mm, struct mm_struct *mm) | 127 | static inline void activate_mm(struct mm_struct *active_mm, struct mm_struct *mm) |
| 125 | { | 128 | { |
| 129 | unsigned long flags; | ||
| 126 | int cpu; | 130 | int cpu; |
| 127 | 131 | ||
| 128 | /* Note: page_table_lock is used here to serialize switch_mm | 132 | spin_lock_irqsave(&mm->context.lock, flags); |
| 129 | * and activate_mm, and their calls to get_new_mmu_context. | ||
| 130 | * This use of page_table_lock is unrelated to its other uses. | ||
| 131 | */ | ||
| 132 | spin_lock(&mm->page_table_lock); | ||
| 133 | if (!CTX_VALID(mm->context)) | 133 | if (!CTX_VALID(mm->context)) |
| 134 | get_new_mmu_context(mm); | 134 | get_new_mmu_context(mm); |
| 135 | cpu = smp_processor_id(); | 135 | cpu = smp_processor_id(); |
| 136 | if (!cpu_isset(cpu, mm->cpu_vm_mask)) | 136 | if (!cpu_isset(cpu, mm->cpu_vm_mask)) |
| 137 | cpu_set(cpu, mm->cpu_vm_mask); | 137 | cpu_set(cpu, mm->cpu_vm_mask); |
| 138 | spin_unlock(&mm->page_table_lock); | ||
| 139 | 138 | ||
| 140 | load_secondary_context(mm); | 139 | load_secondary_context(mm); |
| 141 | __flush_tlb_mm(CTX_HWBITS(mm->context), SECONDARY_CONTEXT); | 140 | __flush_tlb_mm(CTX_HWBITS(mm->context), SECONDARY_CONTEXT); |
| 142 | reload_tlbmiss_state(current, mm); | 141 | tsb_context_switch(mm); |
| 142 | spin_unlock_irqrestore(&mm->context.lock, flags); | ||
| 143 | } | 143 | } |
| 144 | 144 | ||
| 145 | #endif /* !(__ASSEMBLY__) */ | 145 | #endif /* !(__ASSEMBLY__) */ |
