From 2c3c1048746a4622d8c89a29670120dc8fab93c4 Mon Sep 17 00:00:00 2001 From: Daniel Baumann Date: Sun, 7 Apr 2024 20:49:45 +0200 Subject: Adding upstream version 6.1.76. Signed-off-by: Daniel Baumann --- arch/arc/mm/tlbex.S | 378 ++++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 378 insertions(+) create mode 100644 arch/arc/mm/tlbex.S (limited to 'arch/arc/mm/tlbex.S') diff --git a/arch/arc/mm/tlbex.S b/arch/arc/mm/tlbex.S new file mode 100644 index 000000000..e054780a8 --- /dev/null +++ b/arch/arc/mm/tlbex.S @@ -0,0 +1,378 @@ +/* SPDX-License-Identifier: GPL-2.0-only */ +/* + * TLB Exception Handling for ARC + * + * Copyright (C) 2004, 2007-2010, 2011-2012 Synopsys, Inc. (www.synopsys.com) + * + * 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 + */ + +#include +#include +#include +#include +#include +#include +#include + +#ifdef CONFIG_ISA_ARCOMPACT +;----------------------------------------------------------------- +; 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 + .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 + +.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, + asl 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] +.endm + +.macro TLBMISS_RESTORE_REGS +#ifdef CONFIG_SMP + GET_CPU_ID r0 ; get to per cpu scratch mem + asl 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 + +#else /* ARCv2 */ + +.macro TLBMISS_FREEUP_REGS +#ifdef CONFIG_ARC_HAS_LL64 + std r0, [sp, -16] + std r2, [sp, -8] +#else + PUSH r0 + PUSH r1 + PUSH r2 + PUSH r3 +#endif +.endm + +.macro TLBMISS_RESTORE_REGS +#ifdef CONFIG_ARC_HAS_LL64 + ldd r0, [sp, -16] + ldd r2, [sp, -8] +#else + POP r3 + POP r2 + POP r1 + POP r0 +#endif +.endm + +#endif + +;============================================================================ +;TLB Miss handling Code +;============================================================================ + +#ifndef PMD_SHIFT +#define PMD_SHIFT PUD_SHIFT +#endif + +#ifndef PUD_SHIFT +#define PUD_SHIFT PGDIR_SHIFT +#endif + +;----------------------------------------------------------------------------- +; 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] + +#ifdef CONFIG_ISA_ARCV2 + 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 r3, [r1, r0] ; PGD entry corresp to faulting addr + tst r3, r3 + bz do_slow_path_pf ; if no Page Table, do page fault + +#if CONFIG_PGTABLE_LEVELS > 3 + lsr r0, r2, PUD_SHIFT ; Bits for indexing into PUD + and r0, r0, (PTRS_PER_PUD - 1) + ld.as r1, [r3, r0] ; PMD entry + tst r1, r1 + bz do_slow_path_pf + mov r3, r1 +#endif + +#if CONFIG_PGTABLE_LEVELS > 2 + lsr r0, r2, PMD_SHIFT ; Bits for indexing into PMD + and r0, r0, (PTRS_PER_PMD - 1) + ld.as r1, [r3, r0] ; PMD entry + tst r1, r1 + bz do_slow_path_pf + mov r3, r1 +#endif + +#ifdef CONFIG_TRANSPARENT_HUGEPAGE + and.f 0, r3, _PAGE_HW_SZ ; Is this Huge PMD (thp) + add2.nz r1, r1, r0 + bnz.d 2f ; YES: PGD == PMD has THP PTE: stop pgd walk + mov.nz r0, r3 + +#endif + and r1, r3, PAGE_MASK + + ; 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 * 4) + +#ifdef CONFIG_ARC_HAS_PAE40 +#define PTE_SIZE_LOG 3 /* 8 == 2 ^ 3 */ +#else +#define PTE_SIZE_LOG 2 /* 4 == 2 ^ 2 */ +#endif + + ; multiply in step (3) above avoided by shifting lesser in step (1) + lsr r0, r2, ( PAGE_SHIFT - PTE_SIZE_LOG ) + and r0, r0, ( (PTRS_PER_PTE - 1) << PTE_SIZE_LOG ) + ld.aw r0, [r1, r0] ; r0: PTE (lower word only for PAE40) + ; r1: PTE ptr + +2: + +.endm + +;----------------------------------------------------------------- +; Convert Linux PTE entry into TLB entry +; A one-word PTE entry is programmed as two-word TLB Entry [PD0:PD1] in mmu +; (for PAE40, two-words PTE, while three-word TLB Entry [PD0:PD1:PD1HI]) +; IN: r0 = PTE, r1 = ptr to PTE + +.macro CONV_PTE_TO_TLB + and r3, r0, PTE_BITS_RWX ; r w x + asl r2, r3, 3 ; Kr Kw Kx 0 0 0 (GLOBAL, kernel only) + and.f 0, r0, _PAGE_GLOBAL + or.z r2, r2, r3 ; Kr Kw Kx Ur Uw Ux (!GLOBAL, user page) + + and r3, r0, PTE_BITS_NON_RWX_IN_PD1 ; Extract PFN+cache bits from PTE + or r3, r3, r2 + + sr r3, [ARC_REG_TLBPD1] ; paddr[31..13] | Kr Kw Kx Ur Uw Ux | C +#ifdef CONFIG_ARC_HAS_PAE40 + ld r3, [r1, 4] ; paddr[39..32] + sr r3, [ARC_REG_TLBPD1HI] +#endif + + and r2, r0, PTE_BITS_IN_PD0 ; Extract other PTE flags: (V)alid, (G)lb + + 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 +#ifdef CONFIG_ARC_MMU_V3 + + /* Get free TLB slot: Set = computed from vaddr, way = random */ + sr TLBGetIndex, [ARC_REG_TLBCOMMAND] + + /* Commit the Write */ + sr TLBWriteNI, [ARC_REG_TLBCOMMAND] + +#else + sr TLBInsertEntry, [ARC_REG_TLBCOMMAND] +#endif + +88: +.endm + + +ARCFP_CODE ;Fast Path Code, candidate for ICCM + +;----------------------------------------------------------------------------- +; I-TLB Miss Exception Handler +;----------------------------------------------------------------------------- + +ENTRY(EV_TLBMissI) + + TLBMISS_FREEUP_REGS + + ;---------------------------------------------------------------- + ; 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_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_ACCESSED ; set Accessed Bit + st_s r0, [r1] ; Write back PTE + + CONV_PTE_TO_TLB + COMMIT_ENTRY_TO_MMU + TLBMISS_RESTORE_REGS +EV_TLBMissI_fast_ret: ; additional label for VDK OS-kit instrumentation + rtie + +END(EV_TLBMissI) + +;----------------------------------------------------------------------------- +; D-TLB Miss Exception Handler +;----------------------------------------------------------------------------- + +ENTRY(EV_TLBMissD) + + TLBMISS_FREEUP_REGS + + ;---------------------------------------------------------------- + ; Get the PTE corresponding to V-addr accessed + ; 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) + + 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_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 (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 + 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 + or r0, r0, _PAGE_ACCESSED ; Accessed bit always + or.nz r0, r0, _PAGE_DIRTY ; if Write, set Dirty bit as well + st_s r0, [r1] ; Write back PTE + + CONV_PTE_TO_TLB + + COMMIT_ENTRY_TO_MMU + TLBMISS_RESTORE_REGS +EV_TLBMissD_fast_ret: ; additional label for VDK OS-kit instrumentation + rtie + +;-------- Common routine to call Linux Page Fault Handler ----------- +do_slow_path_pf: + +#ifdef CONFIG_ISA_ARCV2 + ; Set Z flag if exception in U mode. Hardware micro-ops do this on any + ; taken interrupt/exception, and thus is already the case at the entry + ; above, but ensuing code would have already clobbered. + ; EXCEPTION_PROLOGUE called in slow path, relies on correct Z flag set + + lr r2, [erstatus] + and r2, r2, STATUS_U_MASK + bxor.f 0, r2, STATUS_U_BIT +#endif + + ; 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). + b call_do_page_fault +END(EV_TLBMissD) -- cgit v1.2.3