diff options
Diffstat (limited to 'arch/arc/mm/tlbex.S')
| -rw-r--r-- | arch/arc/mm/tlbex.S | 408 |
1 files changed, 408 insertions, 0 deletions
diff --git a/arch/arc/mm/tlbex.S b/arch/arc/mm/tlbex.S new file mode 100644 index 0000000..9df765d --- /dev/null +++ b/arch/arc/mm/tlbex.S @@ -0,0 +1,408 @@ +/* + * TLB Exception Handling for ARC + * + * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com) + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + * + * Vineetg: April 2011 : + * -MMU v1: moved out legacy code into a seperate file + * -MMU v3: PD{0,1} bits layout changed: They don't overlap anymore, + * helps avoid a shift when preparing PD0 from PTE + * + * Vineetg: July 2009 + * -For MMU V2, we need not do heuristics at the time of commiting a D-TLB + * entry, so that it doesn't knock out it's I-TLB entry + * -Some more fine tuning: + * bmsk instead of add, asl.cc instead of branch, delay slot utilise etc + * + * Vineetg: July 2009 + * -Practically rewrote the I/D TLB Miss handlers + * Now 40 and 135 instructions a peice as compared to 131 and 449 resp. + * Hence Leaner by 1.5 K + * Used Conditional arithmetic to replace excessive branching + * Also used short instructions wherever possible + * + * Vineetg: Aug 13th 2008 + * -Passing ECR (Exception Cause REG) to do_page_fault( ) for printing + * more information in case of a Fatality + * + * Vineetg: March 25th Bug #92690 + * -Added Debug Code to check if sw-ASID == hw-ASID + + * Rahul Trivedi, Amit Bhor: Codito Technologies 2004 + */ + + .cpu A7 + +#include <linux/linkage.h> +#include <asm/entry.h> +#include <asm/tlb.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 +;-------------------------------------------------------------------------- + +ARCFP_DATA ex_saved_reg1 + .align 1 << L1_CACHE_SHIFT ; IMP: Must be Cache Line aligned + .type ex_saved_reg1, @object +#ifdef CONFIG_SMP + .size ex_saved_reg1, (CONFIG_NR_CPUS << L1_CACHE_SHIFT) +ex_saved_reg1: + .zero (CONFIG_NR_CPUS << L1_CACHE_SHIFT) +#else + .size ex_saved_reg1, 16 +ex_saved_reg1: + .zero 16 +#endif + +;============================================================================ +; Troubleshooting Stuff +;============================================================================ + +; Linux keeps ASID (Address Space ID) in task->active_mm->context.asid +; When Creating TLB Entries, instead of doing 3 dependent loads from memory, +; we use the MMU PID Reg to get current ASID. +; 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 + +#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] + + lr r1, [ARC_REG_PID] + and r1, r1, 0xFF + breq r1, r0, 5f + + ; 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 + + ; 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 + j print_asid_mismatch + +5: ; ASIDs match so proceed normally + nop + +#endif + +.endm + +;============================================================================ +;TLB Miss handling Code +;============================================================================ + +;----------------------------------------------------------------------------- +; This macro does the page-table lookup for the faulting address. +; OUT: r0 = PTE faulted on, r1 = ptr to PTE, r2 = Faulting V-address +.macro LOAD_FAULT_PTE + + lr r2, [efa] + +#ifndef CONFIG_SMP + lr r1, [ARC_REG_SCRATCH_DATA0] ; current pgd +#else + GET_CURR_TASK_ON_CPU r1 + ld r1, [r1, TASK_ACT_MM] + ld r1, [r1, MM_PGD] +#endif + + lsr r0, r2, PGDIR_SHIFT ; Bits for indexing into PGD + ld.as r1, [r1, r0] ; PGD entry corresp to faulting addr + and.f r1, r1, PAGE_MASK ; Ignoring protection and other flags + ; contains Ptr to Page Table + bz.d do_slow_path_pf ; if no Page Table, do page fault + + ; Get the PTE entry: The idea is + ; (1) x = addr >> PAGE_SHIFT -> masks page-off bits from @fault-addr + ; (2) y = x & (PTRS_PER_PTE - 1) -> to get index + ; (3) z = pgtbl[y] + ; To avoid the multiply by in end, we do the -2, <<2 below + + lsr r0, r2, (PAGE_SHIFT - 2) + and r0, r0, ( (PTRS_PER_PTE - 1) << 2) + ld.aw r0, [r1, r0] ; get PTE and PTE ptr for fault addr +#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] +1: +#endif + +.endm + +;----------------------------------------------------------------- +; Convert Linux PTE entry into TLB entry +; A one-word PTE entry is programmed as two-word TLB Entry [PD0:PD1] in mmu +; 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 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 + + or r3, r3, r2 ; S | vaddr | {sasid|asid} + sr r3,[ARC_REG_TLBPD0] ; rewrite PD0 +.endm + +;----------------------------------------------------------------- +; Commit the TLB entry into MMU + +.macro COMMIT_ENTRY_TO_MMU + + /* Get free TLB slot: Set = computed from vaddr, way = random */ + sr TLBGetIndex, [ARC_REG_TLBCOMMAND] + + /* Commit the Write */ +#if (CONFIG_ARC_MMU_VER >= 2) /* introduced in v2 */ + sr TLBWriteNI, [ARC_REG_TLBCOMMAND] +#else + sr TLBWrite, [ARC_REG_TLBCOMMAND] +#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 + +;----------------------------------------------------------------------------- +; I-TLB Miss Exception Handler +;----------------------------------------------------------------------------- + +ARC_ENTRY EV_TLBMissI + + TLBMISS_FREEUP_REGS + +#ifdef CONFIG_ARC_DBG_TLB_MISS_COUNT + ld r0, [@numitlb] + add r0, r0, 1 + st r0, [@numitlb] +#endif + + ;---------------------------------------------------------------- + ; Get the PTE corresponding to V-addr accessed + LOAD_FAULT_PTE + + ;---------------------------------------------------------------- + ; VERIFY_PTE: Check if PTE permissions approp for executing code + cmp_s r2, VMALLOC_START + mov.lo r2, (_PAGE_PRESENT | _PAGE_READ | _PAGE_EXECUTE) + mov.hs r2, (_PAGE_PRESENT | _PAGE_K_READ | _PAGE_K_EXECUTE) + + 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 + + CONV_PTE_TO_TLB + COMMIT_ENTRY_TO_MMU + TLBMISS_RESTORE_REGS + rtie + +ARC_EXIT EV_TLBMissI + +;----------------------------------------------------------------------------- +; D-TLB Miss Exception Handler +;----------------------------------------------------------------------------- + +ARC_ENTRY EV_TLBMissD + + TLBMISS_FREEUP_REGS + +#ifdef CONFIG_ARC_DBG_TLB_MISS_COUNT + ld r0, [@numdtlb] + add r0, r0, 1 + st r0, [@numdtlb] +#endif + + ;---------------------------------------------------------------- + ; Get the PTE corresponding to V-addr accessed + ; If PTE exists, it will setup, r0 = PTE, r1 = Ptr to PTE + LOAD_FAULT_PTE + + ;---------------------------------------------------------------- + ; VERIFY_PTE: Chk if PTE permissions approp for data access (R/W/R+W) + + mov_s r2, 0 + lr r3, [ecr] + btst_s r3, ECR_C_BIT_DTLB_LD_MISS ; Read Access + 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_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 + + ; 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 + brne.d r3, r2, do_slow_path_pf ; is ((pte & flags_test) == flags_test) + + ;---------------------------------------------------------------- + ; UPDATE_PTE: Let Linux VM know that page was accessed/dirty + lr r3, [ecr] + or r0, r0, (_PAGE_PRESENT | _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 + + CONV_PTE_TO_TLB + +#if (CONFIG_ARC_MMU_VER == 1) + ; MMU with 2 way set assoc J-TLB, needs some help in pathetic case of + ; memcpy where 3 parties contend for 2 ways, ensuing a livelock. + ; But only for old MMU or one with Metal Fix + TLB_WRITE_HEURISTICS +#endif + + COMMIT_ENTRY_TO_MMU + TLBMISS_RESTORE_REGS + rtie + +;-------- Common routine to call Linux Page Fault Handler ----------- +do_slow_path_pf: + + ; Restore the 4-scratch regs saved by fast path miss handler + TLBMISS_RESTORE_REGS + + ; 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 + + ; ------- 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 + + ; 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 + ; exceptions get re-enable + + FAKE_RET_FROM_EXCPN r9 + + bl do_page_fault + b ret_from_exception + +ARC_EXIT EV_TLBMissD + +ARC_ENTRY EV_TLBMissB ; Bogus entry to measure sz of DTLBMiss hdlr |