diff options
Diffstat (limited to 'arch/arc/mm/tlbex.S')
| -rw-r--r-- | arch/arc/mm/tlbex.S | 248 |
1 files changed, 112 insertions, 136 deletions
diff --git a/arch/arc/mm/tlbex.S b/arch/arc/mm/tlbex.S index 3357d26ffe5..79bfc81358c 100644 --- a/arch/arc/mm/tlbex.S +++ b/arch/arc/mm/tlbex.S @@ -39,22 +39,41 @@ #include <linux/linkage.h> #include <asm/entry.h> -#include <asm/tlb.h> +#include <asm/mmu.h> #include <asm/pgtable.h> #include <asm/arcregs.h> #include <asm/cache.h> #include <asm/processor.h> -#if (CONFIG_ARC_MMU_VER == 1) #include <asm/tlb-mmu1.h> -#endif -;-------------------------------------------------------------------------- -; scratch memory to save the registers (r0-r3) used to code TLB refill Handler -; For details refer to comments before TLBMISS_FREEUP_REGS below +;----------------------------------------------------------------- +; ARC700 Exception Handling doesn't auto-switch stack and it only provides +; ONE scratch AUX reg "ARC_REG_SCRATCH_DATA0" +; +; For Non-SMP, the scratch AUX reg is repurposed to cache task PGD, so a +; "global" is used to free-up FIRST core reg to be able to code the rest of +; exception prologue (IRQ auto-disabled on Exceptions, so it's IRQ-safe). +; Since the Fast Path TLB Miss handler is coded with 4 regs, the remaining 3 +; need to be saved as well by extending the "global" to be 4 words. Hence +; ".size ex_saved_reg1, 16" +; [All of this dance is to avoid stack switching for each TLB Miss, since we +; only need to save only a handful of regs, as opposed to complete reg file] +; +; For ARC700 SMP, the "global" obviously can't be used for free up the FIRST +; core reg as it will not be SMP safe. +; Thus scratch AUX reg is used (and no longer used to cache task PGD). +; To save the rest of 3 regs - per cpu, the global is made "per-cpu". +; Epilogue thus has to locate the "per-cpu" storage for regs. +; To avoid cache line bouncing the per-cpu global is aligned/sized per +; L1_CACHE_SHIFT, despite fundamentally needing to be 12 bytes only. Hence +; ".size ex_saved_reg1, (CONFIG_NR_CPUS << L1_CACHE_SHIFT)" + +; As simple as that.... ;-------------------------------------------------------------------------- +; scratch memory to save [r0-r3] used to code TLB refill Handler ARCFP_DATA ex_saved_reg1 - .align 1 << L1_CACHE_SHIFT ; IMP: Must be Cache Line aligned + .align 1 << L1_CACHE_SHIFT .type ex_saved_reg1, @object #ifdef CONFIG_SMP .size ex_saved_reg1, (CONFIG_NR_CPUS << L1_CACHE_SHIFT) @@ -66,6 +85,44 @@ ex_saved_reg1: .zero 16 #endif +.macro TLBMISS_FREEUP_REGS +#ifdef CONFIG_SMP + sr r0, [ARC_REG_SCRATCH_DATA0] ; freeup r0 to code with + GET_CPU_ID r0 ; get to per cpu scratch mem, + lsl r0, r0, L1_CACHE_SHIFT ; cache line wide per cpu + add r0, @ex_saved_reg1, r0 +#else + st r0, [@ex_saved_reg1] + mov_s r0, @ex_saved_reg1 +#endif + st_s r1, [r0, 4] + st_s r2, [r0, 8] + st_s r3, [r0, 12] + + ; VERIFY if the ASID in MMU-PID Reg is same as + ; one in Linux data structures + + tlb_paranoid_check_asm +.endm + +.macro TLBMISS_RESTORE_REGS +#ifdef CONFIG_SMP + GET_CPU_ID r0 ; get to per cpu scratch mem + lsl r0, r0, L1_CACHE_SHIFT ; each is cache line wide + add r0, @ex_saved_reg1, r0 + ld_s r3, [r0,12] + ld_s r2, [r0, 8] + ld_s r1, [r0, 4] + lr r0, [ARC_REG_SCRATCH_DATA0] +#else + mov_s r0, @ex_saved_reg1 + ld_s r3, [r0,12] + ld_s r2, [r0, 8] + ld_s r1, [r0, 4] + ld_s r0, [r0] +#endif +.endm + ;============================================================================ ; Troubleshooting Stuff ;============================================================================ @@ -76,34 +133,35 @@ ex_saved_reg1: ; In bizzare scenrios SW and HW ASID can get out-of-sync which is trouble. ; So we try to detect this in TLB Mis shandler - -.macro DBG_ASID_MISMATCH +.macro tlb_paranoid_check_asm #ifdef CONFIG_ARC_DBG_TLB_PARANOIA - ; make sure h/w ASID is same as s/w ASID - GET_CURR_TASK_ON_CPU r3 ld r0, [r3, TASK_ACT_MM] ld r0, [r0, MM_CTXT+MM_CTXT_ASID] + breq r0, 0, 55f ; Error if no ASID allocated lr r1, [ARC_REG_PID] and r1, r1, 0xFF - breq r1, r0, 5f + and r2, r0, 0xFF ; MMU PID bits only for comparison + breq r1, r2, 5f + +55: ; Error if H/w and S/w ASID don't match, but NOT if in kernel mode - lr r0, [erstatus] - bbit0 r0, STATUS_U_BIT, 5f + lr r2, [erstatus] + bbit0 r2, STATUS_U_BIT, 5f ; We sure are in troubled waters, Flag the error, but to do so ; need to switch to kernel mode stack to call error routine GET_TSK_STACK_BASE r3, sp ; Call printk to shoutout aloud - mov r0, 1 + mov r2, 1 j print_asid_mismatch -5: ; ASIDs match so proceed normally +5: ; ASIDs match so proceed normally nop #endif @@ -147,9 +205,9 @@ ex_saved_reg1: #ifdef CONFIG_ARC_DBG_TLB_MISS_COUNT and.f 0, r0, _PAGE_PRESENT bz 1f - ld r2, [num_pte_not_present] - add r2, r2, 1 - st r2, [num_pte_not_present] + ld r3, [num_pte_not_present] + add r3, r3, 1 + st r3, [num_pte_not_present] 1: #endif @@ -161,13 +219,17 @@ ex_saved_reg1: ; IN: r0 = PTE, r1 = ptr to PTE .macro CONV_PTE_TO_TLB - and r3, r0, PTE_BITS_IN_PD1 ; Extract permission flags+PFN from PTE - sr r3, [ARC_REG_TLBPD1] ; these go in PD1 + and r3, r0, PTE_BITS_RWX ; r w x + lsl r2, r3, 3 ; r w x 0 0 0 + and.f 0, r0, _PAGE_GLOBAL + or.z r2, r2, r3 ; r w x r w x + + and r3, r0, PTE_BITS_NON_RWX_IN_PD1 ; Extract PFN+cache bits from PTE + or r3, r3, r2 + + sr r3, [ARC_REG_TLBPD1] ; these go in PD1 and r2, r0, PTE_BITS_IN_PD0 ; Extract other PTE flags: (V)alid, (G)lb -#if (CONFIG_ARC_MMU_VER <= 2) /* Neednot be done with v3 onwards */ - lsr r2, r2 ; shift PTE flags to match layout in PD0 -#endif lr r3,[ARC_REG_TLBPD0] ; MMU prepares PD0 with vaddr and asid @@ -191,68 +253,6 @@ ex_saved_reg1: #endif .endm -;----------------------------------------------------------------- -; ARC700 Exception Handling doesn't auto-switch stack and it only provides -; ONE scratch AUX reg "ARC_REG_SCRATCH_DATA0" -; -; For Non-SMP, the scratch AUX reg is repurposed to cache task PGD, so a -; "global" is used to free-up FIRST core reg to be able to code the rest of -; exception prologue (IRQ auto-disabled on Exceptions, so it's IRQ-safe). -; Since the Fast Path TLB Miss handler is coded with 4 regs, the remaining 3 -; need to be saved as well by extending the "global" to be 4 words. Hence -; ".size ex_saved_reg1, 16" -; [All of this dance is to avoid stack switching for each TLB Miss, since we -; only need to save only a handful of regs, as opposed to complete reg file] -; -; For ARC700 SMP, the "global" obviously can't be used for free up the FIRST -; core reg as it will not be SMP safe. -; Thus scratch AUX reg is used (and no longer used to cache task PGD). -; To save the rest of 3 regs - per cpu, the global is made "per-cpu". -; Epilogue thus has to locate the "per-cpu" storage for regs. -; To avoid cache line bouncing the per-cpu global is aligned/sized per -; L1_CACHE_SHIFT, despite fundamentally needing to be 12 bytes only. Hence -; ".size ex_saved_reg1, (CONFIG_NR_CPUS << L1_CACHE_SHIFT)" - -; As simple as that.... - -.macro TLBMISS_FREEUP_REGS -#ifdef CONFIG_SMP - sr r0, [ARC_REG_SCRATCH_DATA0] ; freeup r0 to code with - GET_CPU_ID r0 ; get to per cpu scratch mem, - lsl r0, r0, L1_CACHE_SHIFT ; cache line wide per cpu - add r0, @ex_saved_reg1, r0 -#else - st r0, [@ex_saved_reg1] - mov_s r0, @ex_saved_reg1 -#endif - st_s r1, [r0, 4] - st_s r2, [r0, 8] - st_s r3, [r0, 12] - - ; VERIFY if the ASID in MMU-PID Reg is same as - ; one in Linux data structures - - DBG_ASID_MISMATCH -.endm - -;----------------------------------------------------------------- -.macro TLBMISS_RESTORE_REGS -#ifdef CONFIG_SMP - GET_CPU_ID r0 ; get to per cpu scratch mem - lsl r0, r0, L1_CACHE_SHIFT ; each is cache line wide - add r0, @ex_saved_reg1, r0 - ld_s r3, [r0,12] - ld_s r2, [r0, 8] - ld_s r1, [r0, 4] - lr r0, [ARC_REG_SCRATCH_DATA0] -#else - mov_s r0, @ex_saved_reg1 - ld_s r3, [r0,12] - ld_s r2, [r0, 8] - ld_s r1, [r0, 4] - ld_s r0, [r0] -#endif -.endm ARCFP_CODE ;Fast Path Code, candidate for ICCM @@ -260,7 +260,7 @@ ARCFP_CODE ;Fast Path Code, candidate for ICCM ; I-TLB Miss Exception Handler ;----------------------------------------------------------------------------- -ARC_ENTRY EV_TLBMissI +ENTRY(EV_TLBMissI) TLBMISS_FREEUP_REGS @@ -271,35 +271,35 @@ ARC_ENTRY EV_TLBMissI #endif ;---------------------------------------------------------------- - ; Get the PTE corresponding to V-addr accessed + ; Get the PTE corresponding to V-addr accessed, r2 is setup with EFA LOAD_FAULT_PTE ;---------------------------------------------------------------- ; VERIFY_PTE: Check if PTE permissions approp for executing code cmp_s r2, VMALLOC_START - mov.lo r2, (_PAGE_PRESENT | _PAGE_U_READ | _PAGE_U_EXECUTE) - mov.hs r2, (_PAGE_PRESENT | _PAGE_K_READ | _PAGE_K_EXECUTE) + mov_s r2, (_PAGE_PRESENT | _PAGE_EXECUTE) + or.hs r2, r2, _PAGE_GLOBAL and r3, r0, r2 ; Mask out NON Flag bits from PTE xor.f r3, r3, r2 ; check ( ( pte & flags_test ) == flags_test ) bnz do_slow_path_pf ; Let Linux VM know that the page was accessed - or r0, r0, (_PAGE_PRESENT | _PAGE_ACCESSED) ; set Accessed Bit - st_s r0, [r1] ; Write back PTE + or r0, r0, _PAGE_ACCESSED ; set Accessed Bit + st_s r0, [r1] ; Write back PTE CONV_PTE_TO_TLB COMMIT_ENTRY_TO_MMU TLBMISS_RESTORE_REGS rtie -ARC_EXIT EV_TLBMissI +END(EV_TLBMissI) ;----------------------------------------------------------------------------- ; D-TLB Miss Exception Handler ;----------------------------------------------------------------------------- -ARC_ENTRY EV_TLBMissD +ENTRY(EV_TLBMissD) TLBMISS_FREEUP_REGS @@ -311,32 +311,27 @@ ARC_ENTRY EV_TLBMissD ;---------------------------------------------------------------- ; Get the PTE corresponding to V-addr accessed - ; If PTE exists, it will setup, r0 = PTE, r1 = Ptr to PTE + ; If PTE exists, it will setup, r0 = PTE, r1 = Ptr to PTE, r2 = EFA LOAD_FAULT_PTE ;---------------------------------------------------------------- ; VERIFY_PTE: Chk if PTE permissions approp for data access (R/W/R+W) - mov_s r2, 0 + cmp_s r2, VMALLOC_START + mov_s r2, _PAGE_PRESENT ; common bit for K/U PTE + or.hs r2, r2, _PAGE_GLOBAL ; kernel PTE only + + ; Linux PTE [RWX] bits are semantically overloaded: + ; -If PAGE_GLOBAL set, they refer to kernel-only flags (vmalloc) + ; -Otherwise they are user-mode permissions, and those are exactly + ; same for kernel mode as well (e.g. copy_(to|from)_user) + lr r3, [ecr] btst_s r3, ECR_C_BIT_DTLB_LD_MISS ; Read Access - or.nz r2, r2, _PAGE_U_READ ; chk for Read flag in PTE + or.nz r2, r2, _PAGE_READ ; chk for Read flag in PTE btst_s r3, ECR_C_BIT_DTLB_ST_MISS ; Write Access - or.nz r2, r2, _PAGE_U_WRITE ; chk for Write flag in PTE - ; Above laddering takes care of XCHG access - ; which is both Read and Write - - ; If kernel mode access, ; make _PAGE_xx flags as _PAGE_K_xx - ; For copy_(to|from)_user, despite exception taken in kernel mode, - ; this code is not hit, because EFA would still be the user mode - ; address (EFA < 0x6000_0000). - ; This code is for legit kernel mode faults, vmalloc specifically - ; (EFA: 0x7000_0000 to 0x7FFF_FFFF) - - lr r3, [efa] - cmp r3, VMALLOC_START - 1 ; If kernel mode access - asl.hi r2, r2, 3 ; make _PAGE_xx flags as _PAGE_K_xx - or r2, r2, _PAGE_PRESENT ; Common flag for K/U mode + or.nz r2, r2, _PAGE_WRITE ; chk for Write flag in PTE + ; Above laddering takes care of XCHG access (both R and W) ; By now, r2 setup with all the Flags we need to check in PTE and r3, r0, r2 ; Mask out NON Flag bits from PTE @@ -345,7 +340,7 @@ ARC_ENTRY EV_TLBMissD ;---------------------------------------------------------------- ; UPDATE_PTE: Let Linux VM know that page was accessed/dirty lr r3, [ecr] - or r0, r0, (_PAGE_PRESENT | _PAGE_ACCESSED) ; Accessed bit always + or r0, r0, _PAGE_ACCESSED ; Accessed bit always btst_s r3, ECR_C_BIT_DTLB_ST_MISS ; See if it was a Write Access ? or.nz r0, r0, _PAGE_MODIFIED ; if Write, set Dirty bit as well st_s r0, [r1] ; Write back PTE @@ -371,28 +366,11 @@ do_slow_path_pf: ; Slow path TLB Miss handled as a regular ARC Exception ; (stack switching / save the complete reg-file). - ; That requires freeing up r9 - EXCPN_PROLOG_FREEUP_REG r9 - - lr r9, [erstatus] - - SWITCH_TO_KERNEL_STK - SAVE_ALL_SYS + EXCEPTION_PROLOGUE ; ------- setup args for Linux Page fault Hanlder --------- - mov_s r0, sp - lr r2, [efa] - lr r3, [ecr] - - ; Both st and ex imply WRITE access of some sort, hence do_page_fault( ) - ; invoked with write=1 for DTLB-st/ex Miss and write=0 for ITLB miss or - ; DTLB-ld Miss - ; DTLB Miss Cause code is ld = 0x01 , st = 0x02, ex = 0x03 - ; Following code uses that fact that st/ex have one bit in common - - btst_s r3, ECR_C_BIT_DTLB_ST_MISS - mov.z r1, 0 - mov.nz r1, 1 + mov_s r1, sp + lr r0, [efa] ; We don't want exceptions to be disabled while the fault is handled. ; Now that we have saved the context we return from exception hence @@ -403,6 +381,4 @@ do_slow_path_pf: bl do_page_fault b ret_from_exception -ARC_EXIT EV_TLBMissD - -ARC_ENTRY EV_TLBMissB ; Bogus entry to measure sz of DTLBMiss hdlr +END(EV_TLBMissD) |