Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754309Ab2KGJs4 (ORCPT ); Wed, 7 Nov 2012 04:48:56 -0500 Received: from us02smtp1.synopsys.com ([198.182.60.75]:52601 "EHLO vaxjo.synopsys.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754243Ab2KGJso (ORCPT ); Wed, 7 Nov 2012 04:48:44 -0500 From: Vineet Gupta To: linux-arch@vger.kernel.org, linux-kernel@vger.kernel.org Cc: tglx@linutronix.de, arnd@arndb.de, Vineet Gupta Subject: [RFC PATCH v1 19/31] ARC: MMU Context Management Date: Wed, 7 Nov 2012 10:47:42 +0100 Message-Id: <1352281674-2186-20-git-send-email-vgupta@synopsys.com> X-Mailer: git-send-email 1.7.0.4 In-Reply-To: <1352281674-2186-1-git-send-email-vgupta@synopsys.com> References: <1352281674-2186-1-git-send-email-vgupta@synopsys.com> Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 10866 Lines: 318 ARC700 MMU provides for tagging TLB entries with a 8-bit ASID to avoid having to flush the TLB every task switch. It also allows for a quick way to invalidate all the TLB entries for task useful for: * COW sementics during fork() * task exit()ing Signed-off-by: Vineet Gupta --- arch/arc/include/asm/arcregs.h | 7 ++ arch/arc/include/asm/mmu.h | 23 ++++ arch/arc/include/asm/mmu_context.h | 209 ++++++++++++++++++++++++++++++++++++ arch/arc/mm/tlb.c | 23 ++++ 4 files changed, 262 insertions(+), 0 deletions(-) create mode 100644 arch/arc/include/asm/mmu.h create mode 100644 arch/arc/include/asm/mmu_context.h create mode 100644 arch/arc/mm/tlb.c diff --git a/arch/arc/include/asm/arcregs.h b/arch/arc/include/asm/arcregs.h index c6e2805..c12eb9b 100644 --- a/arch/arc/include/asm/arcregs.h +++ b/arch/arc/include/asm/arcregs.h @@ -85,6 +85,13 @@ #define DC_CTRL_INV_MODE_FLUSH 0x40 #define DC_CTRL_FLUSH_STATUS 0x100 +/* MMU Management regs */ +#define ARC_REG_PID 0x409 +#define ARC_REG_SCRATCH_DATA0 0x418 + +/* Bits in MMU PID register */ +#define MMU_ENABLE (1 << 31) /* Enable MMU for process */ + /* * Floating Pt Registers * Status regs are read-only (build-time) so need not be saved/restored diff --git a/arch/arc/include/asm/mmu.h b/arch/arc/include/asm/mmu.h new file mode 100644 index 0000000..56b0232 --- /dev/null +++ b/arch/arc/include/asm/mmu.h @@ -0,0 +1,23 @@ +/* + * 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. + */ + +#ifndef _ASM_ARC_MMU_H +#define _ASM_ARC_MMU_H + +#ifndef __ASSEMBLY__ + +typedef struct { + unsigned long asid; /* Pvt Addr-Space ID for mm */ +#ifdef CONFIG_ARC_TLB_DBG + struct task_struct *tsk; +#endif +} mm_context_t; + +#endif + +#endif diff --git a/arch/arc/include/asm/mmu_context.h b/arch/arc/include/asm/mmu_context.h new file mode 100644 index 0000000..d12f3de --- /dev/null +++ b/arch/arc/include/asm/mmu_context.h @@ -0,0 +1,209 @@ +/* + * 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: May 2011 + * -Refactored get_new_mmu_context( ) to only handle live-mm. + * retiring-mm handled in other hooks + * + * Vineetg: March 25th, 2008: Bug #92690 + * -Major rewrite of Core ASID allocation routine get_new_mmu_context + * + * Amit Bhor, Sameer Dhavale: Codito Technologies 2004 + */ + +#ifndef _ASM_ARC_MMU_CONTEXT_H +#define _ASM_ARC_MMU_CONTEXT_H + +#include +#include + +#include + +/* ARC700 ASID Management + * + * ARC MMU provides 8-bit ASID (0..255) to TAG TLB entries, allowing entries + * with same vaddr (different tasks) to co-exit. This provides for + * "Fast Context Switch" i.e. no TLB flush on ctxt-switch + * + * Linux assigns each task a unique ASID. A simple round-robin allocation + * of H/w ASID is done using software tracker @asid_cache. + * When it reaches max 255, the allocation cycle starts afresh by flushing + * the entire TLB and wrapping ASID back to zero. + * + * For book-keeping, Linux uses a couple of data-structures: + * -mm_struct has an @asid field to keep a note of task's ASID (needed at the + * time of say switch_mm( ) + * -An array of mm structs @asid_mm_map[] for asid->mm the reverse mapping, + * given an ASID, finding the mm struct associated. + * + * The round-robin allocation algorithm allows for ASID stealing. + * If asid tracker is at "x-1", a new req will allocate "x", even if "x" was + * already assigned to another (switched-out) task. Obviously the prev owner + * is marked with an invalid ASID to make it request for a new ASID when it + * gets scheduled next time. However its TLB entries (with ASID "x") could + * exist, which must be cleared before the same ASID is used by the new owner. + * Flushing them would be plausible but costly solution. Instead we force a + * allocation policy quirk, which ensures that a stolen ASID won't have any + * TLB entries associates, alleviating the need to flush. + * The quirk essentially is not allowing ASID allocated in prev cycle + * to be used past a roll-over in the next cycle. + * When this happens (i.e. task ASID > asid tracker), task needs to refresh + * its ASID, aligning it to current value of tracker. If the task doesn't get + * scheduled past a roll-over, hence its ASID is not yet realigned with + * tracker, such ASID is anyways safely reusable because it is + * gauranteed that TLB entries with that ASID wont exist. + */ + +#define FIRST_ASID 0 +#define MAX_ASID 255 /* 8 bit PID field in PID Aux reg */ +#define NO_ASID (MAX_ASID + 1) /* ASID Not alloc to mmu ctxt */ +#define NUM_ASID ((MAX_ASID - FIRST_ASID) + 1) + +/* ASID to mm struct mapping */ +extern struct mm_struct *asid_mm_map[NUM_ASID + 1]; + +extern int asid_cache; + +/* + * Assign a new ASID to task. If the task already has an ASID, it is + * relinquished. + */ +static inline void get_new_mmu_context(struct mm_struct *mm) +{ + struct mm_struct *prev_owner; + unsigned long flags; + + local_irq_save(flags); + + /* + * Relinquish the currently owned ASID (if any). + * Doing unconditionally saves a cmp-n-branch; for already unused + * ASID slot, the value was/remains NULL + */ + asid_mm_map[mm->context.asid] = (struct mm_struct *)NULL; + + /* move to new ASID */ + if (++asid_cache > MAX_ASID) { /* ASID roll-over */ + asid_cache = FIRST_ASID; + flush_tlb_all(); + } + + /* + * Is next ASID already owned by some-one else (we are stealing it). + * If so, let the orig owner be aware of this, so when it runs, it + * asks for a brand new ASID. This would only happen for a long-lived + * task with ASID from prev allocation cycle (before ASID roll-over). + * + * This might look wrong - if we are re-using some other task's ASID, + * won't we use it's stale TLB entries too. Actually switch_mm( ) takes + * care of such a case: it ensures that task with ASID from prev alloc + * cycle, when scheduled will refresh it's ASID: see switch_mm( ) below + * The stealing scenario described here will only happen if that task + * didn't get a chance to refresh it's ASID - implying stale entries + * won't exist. + */ + prev_owner = asid_mm_map[asid_cache]; + if (prev_owner) + prev_owner->context.asid = NO_ASID; + + /* Assign new ASID to tsk */ + asid_mm_map[asid_cache] = mm; + mm->context.asid = asid_cache; + +#ifdef CONFIG_ARC_TLB_DBG + pr_info("ARC_TLB_DBG: NewMM=0x%x OldMM=0x%x task_struct=0x%x Task: %s," + " pid:%u, assigned asid:%lu\n", + (unsigned int)mm, (unsigned int)prev_owner, + (unsigned int)(mm->context.tsk), (mm->context.tsk)->comm, + (mm->context.tsk)->pid, mm->context.asid); +#endif + + write_aux_reg(ARC_REG_PID, asid_cache | MMU_ENABLE); + + local_irq_restore(flags); +} + +/* + * Initialize the context related info for a new mm_struct + * instance. + */ +static inline int +init_new_context(struct task_struct *tsk, struct mm_struct *mm) +{ + mm->context.asid = NO_ASID; +#ifdef CONFIG_ARC_TLB_DBG + mm->context.tsk = tsk; +#endif + return 0; +} + +/* Prepare the MMU for task: setup PID reg with allocated ASID + If task doesn't have an ASID (never alloc or stolen, get a new ASID) +*/ +static inline void switch_mm(struct mm_struct *prev, struct mm_struct *next, + struct task_struct *tsk) +{ + /* PGD cached in MMU reg to avoid 3 mem lookups: task->mm->pgd */ + write_aux_reg(ARC_REG_SCRATCH_DATA0, next->pgd); + + /* + * Get a new ASID if task doesn't have a valid one. Possible when + * -task never had an ASID (fresh after fork) + * -it's ASID was stolen - past an ASID roll-over. + * -There's a third obscure scenario (if this task is running for the + * first time afer an ASID rollover), where despite having a valid + * ASID, we force a get for new ASID - see comments at top. + * + * Both the non-alloc scenario and first-use-after-rollover can be + * detected using the single condition below: NO_ASID = 256 + * while asid_cache is always a valid ASID value (0-255). + */ + if (next->context.asid > asid_cache) { + get_new_mmu_context(next); + } else { + /* + * XXX: This will never happen given the chks above + * BUG_ON(next->context.asid > MAX_ASID); + */ + write_aux_reg(ARC_REG_PID, next->context.asid | MMU_ENABLE); + } + +} + +static inline void destroy_context(struct mm_struct *mm) +{ + unsigned long flags; + + local_irq_save(flags); + + asid_mm_map[mm->context.asid] = NULL; + mm->context.asid = NO_ASID; + + local_irq_restore(flags); +} + +/* it seemed that deactivate_mm( ) is a reasonable place to do book-keeping + * for retiring-mm. However destroy_context( ) still needs to do that because + * between mm_release( ) = >deactive_mm( ) and + * mmput => .. => __mmdrop( ) => destroy_context( ) + * there is a good chance that task gets sched-out/in, making it's ASID valid + * again (this teased me for a whole day). + */ +#define deactivate_mm(tsk, mm) do { } while (0) + +static inline void activate_mm(struct mm_struct *prev, struct mm_struct *next) +{ + write_aux_reg(ARC_REG_SCRATCH_DATA0, next->pgd); + + /* Unconditionally get a new ASID */ + get_new_mmu_context(next); + +} + +#define enter_lazy_tlb(mm, tsk) + +#endif /* __ASM_ARC_MMU_CONTEXT_H */ diff --git a/arch/arc/mm/tlb.c b/arch/arc/mm/tlb.c new file mode 100644 index 0000000..f1edae2 --- /dev/null +++ b/arch/arc/mm/tlb.c @@ -0,0 +1,23 @@ +/* + * TLB Management (flush/create/diagnostics) for ARC700 + * + * 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. + */ + +#include +#include +#include +#include + +/* A copy of the ASID from the PID reg is kept in asid_cache */ +int asid_cache = FIRST_ASID; + +/* ASID to mm struct mapping. We have one extra entry corresponding to + * NO_ASID to save us a compare when clearing the mm entry for old asid + * see get_new_mmu_context (asm-arc/mmu_context.h) + */ +struct mm_struct *asid_mm_map[NUM_ASID + 1]; -- 1.7.4.1 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/