arm64: Process management
authorCatalin Marinas <catalin.marinas@arm.com>
Mon, 5 Mar 2012 11:49:28 +0000 (11:49 +0000)
committerCatalin Marinas <catalin.marinas@arm.com>
Mon, 17 Sep 2012 12:41:58 +0000 (13:41 +0100)
The patch adds support for thread creation and context switching. The
context switching CPU specific code is introduced with the CPU support
patch (part of the arch/arm64/mm/proc.S file). AArch64 supports
ASID-tagged TLBs and the ASID can be either 8 or 16-bit wide (detectable
via the ID_AA64AFR0_EL1 register).

Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Acked-by: Tony Lindgren <tony@atomide.com>
Acked-by: Nicolas Pitre <nico@linaro.org>
Acked-by: Olof Johansson <olof@lixom.net>
Acked-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
arch/arm64/include/asm/mmu_context.h [new file with mode: 0644]
arch/arm64/include/asm/thread_info.h [new file with mode: 0644]
arch/arm64/kernel/process.c [new file with mode: 0644]
arch/arm64/mm/context.c [new file with mode: 0644]

diff --git a/arch/arm64/include/asm/mmu_context.h b/arch/arm64/include/asm/mmu_context.h
new file mode 100644 (file)
index 0000000..f68465d
--- /dev/null
@@ -0,0 +1,152 @@
+/*
+ * Based on arch/arm/include/asm/mmu_context.h
+ *
+ * Copyright (C) 1996 Russell King.
+ * Copyright (C) 2012 ARM Ltd.
+ *
+ * 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.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+#ifndef __ASM_MMU_CONTEXT_H
+#define __ASM_MMU_CONTEXT_H
+
+#include <linux/compiler.h>
+#include <linux/sched.h>
+
+#include <asm/cacheflush.h>
+#include <asm/proc-fns.h>
+#include <asm-generic/mm_hooks.h>
+#include <asm/cputype.h>
+#include <asm/pgtable.h>
+
+#define MAX_ASID_BITS  16
+
+extern unsigned int cpu_last_asid;
+
+void __init_new_context(struct task_struct *tsk, struct mm_struct *mm);
+void __new_context(struct mm_struct *mm);
+
+/*
+ * Set TTBR0 to empty_zero_page. No translations will be possible via TTBR0.
+ */
+static inline void cpu_set_reserved_ttbr0(void)
+{
+       unsigned long ttbr = page_to_phys(empty_zero_page);
+
+       asm(
+       "       msr     ttbr0_el1, %0                   // set TTBR0\n"
+       "       isb"
+       :
+       : "r" (ttbr));
+}
+
+static inline void switch_new_context(struct mm_struct *mm)
+{
+       unsigned long flags;
+
+       __new_context(mm);
+
+       local_irq_save(flags);
+       cpu_switch_mm(mm->pgd, mm);
+       local_irq_restore(flags);
+}
+
+static inline void check_and_switch_context(struct mm_struct *mm,
+                                           struct task_struct *tsk)
+{
+       /*
+        * Required during context switch to avoid speculative page table
+        * walking with the wrong TTBR.
+        */
+       cpu_set_reserved_ttbr0();
+
+       if (!((mm->context.id ^ cpu_last_asid) >> MAX_ASID_BITS))
+               /*
+                * The ASID is from the current generation, just switch to the
+                * new pgd. This condition is only true for calls from
+                * context_switch() and interrupts are already disabled.
+                */
+               cpu_switch_mm(mm->pgd, mm);
+       else if (irqs_disabled())
+               /*
+                * Defer the new ASID allocation until after the context
+                * switch critical region since __new_context() cannot be
+                * called with interrupts disabled.
+                */
+               set_ti_thread_flag(task_thread_info(tsk), TIF_SWITCH_MM);
+       else
+               /*
+                * That is a direct call to switch_mm() or activate_mm() with
+                * interrupts enabled and a new context.
+                */
+               switch_new_context(mm);
+}
+
+#define init_new_context(tsk,mm)       (__init_new_context(tsk,mm),0)
+#define destroy_context(mm)            do { } while(0)
+
+#define finish_arch_post_lock_switch \
+       finish_arch_post_lock_switch
+static inline void finish_arch_post_lock_switch(void)
+{
+       if (test_and_clear_thread_flag(TIF_SWITCH_MM)) {
+               struct mm_struct *mm = current->mm;
+               unsigned long flags;
+
+               __new_context(mm);
+
+               local_irq_save(flags);
+               cpu_switch_mm(mm->pgd, mm);
+               local_irq_restore(flags);
+       }
+}
+
+/*
+ * This is called when "tsk" is about to enter lazy TLB mode.
+ *
+ * mm:  describes the currently active mm context
+ * tsk: task which is entering lazy tlb
+ * cpu: cpu number which is entering lazy tlb
+ *
+ * tsk->mm will be NULL
+ */
+static inline void
+enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk)
+{
+}
+
+/*
+ * This is the actual mm switch as far as the scheduler
+ * is concerned.  No registers are touched.  We avoid
+ * calling the CPU specific function when the mm hasn't
+ * actually changed.
+ */
+static inline void
+switch_mm(struct mm_struct *prev, struct mm_struct *next,
+         struct task_struct *tsk)
+{
+       unsigned int cpu = smp_processor_id();
+
+#ifdef CONFIG_SMP
+       /* check for possible thread migration */
+       if (!cpumask_empty(mm_cpumask(next)) &&
+           !cpumask_test_cpu(cpu, mm_cpumask(next)))
+               __flush_icache_all();
+#endif
+       if (!cpumask_test_and_set_cpu(cpu, mm_cpumask(next)) || prev != next)
+               check_and_switch_context(next, tsk);
+}
+
+#define deactivate_mm(tsk,mm)  do { } while (0)
+#define activate_mm(prev,next) switch_mm(prev, next, NULL)
+
+#endif
diff --git a/arch/arm64/include/asm/thread_info.h b/arch/arm64/include/asm/thread_info.h
new file mode 100644 (file)
index 0000000..3659e46
--- /dev/null
@@ -0,0 +1,127 @@
+/*
+ * Based on arch/arm/include/asm/thread_info.h
+ *
+ * Copyright (C) 2002 Russell King.
+ * Copyright (C) 2012 ARM Ltd.
+ *
+ * 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.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+#ifndef __ASM_THREAD_INFO_H
+#define __ASM_THREAD_INFO_H
+
+#ifdef __KERNEL__
+
+#include <linux/compiler.h>
+
+#ifndef CONFIG_ARM64_64K_PAGES
+#define THREAD_SIZE_ORDER      1
+#endif
+
+#define THREAD_SIZE            8192
+#define THREAD_START_SP                (THREAD_SIZE - 16)
+
+#ifndef __ASSEMBLY__
+
+struct task_struct;
+struct exec_domain;
+
+#include <asm/types.h>
+
+typedef unsigned long mm_segment_t;
+
+/*
+ * low level task data that entry.S needs immediate access to.
+ * __switch_to() assumes cpu_context follows immediately after cpu_domain.
+ */
+struct thread_info {
+       unsigned long           flags;          /* low level flags */
+       mm_segment_t            addr_limit;     /* address limit */
+       struct task_struct      *task;          /* main task structure */
+       struct exec_domain      *exec_domain;   /* execution domain */
+       struct restart_block    restart_block;
+       int                     preempt_count;  /* 0 => preemptable, <0 => bug */
+       int                     cpu;            /* cpu */
+};
+
+#define INIT_THREAD_INFO(tsk)                                          \
+{                                                                      \
+       .task           = &tsk,                                         \
+       .exec_domain    = &default_exec_domain,                         \
+       .flags          = 0,                                            \
+       .preempt_count  = INIT_PREEMPT_COUNT,                           \
+       .addr_limit     = KERNEL_DS,                                    \
+       .restart_block  = {                                             \
+               .fn     = do_no_restart_syscall,                        \
+       },                                                              \
+}
+
+#define init_thread_info       (init_thread_union.thread_info)
+#define init_stack             (init_thread_union.stack)
+
+/*
+ * how to get the thread information struct from C
+ */
+static inline struct thread_info *current_thread_info(void) __attribute_const__;
+
+static inline struct thread_info *current_thread_info(void)
+{
+       register unsigned long sp asm ("sp");
+       return (struct thread_info *)(sp & ~(THREAD_SIZE - 1));
+}
+
+#define thread_saved_pc(tsk)   \
+       ((unsigned long)(tsk->thread.cpu_context.pc))
+#define thread_saved_sp(tsk)   \
+       ((unsigned long)(tsk->thread.cpu_context.sp))
+#define thread_saved_fp(tsk)   \
+       ((unsigned long)(tsk->thread.cpu_context.fp))
+
+#endif
+
+/*
+ * We use bit 30 of the preempt_count to indicate that kernel
+ * preemption is occurring.  See <asm/hardirq.h>.
+ */
+#define PREEMPT_ACTIVE 0x40000000
+
+/*
+ * thread information flags:
+ *  TIF_SYSCALL_TRACE  - syscall trace active
+ *  TIF_SIGPENDING     - signal pending
+ *  TIF_NEED_RESCHED   - rescheduling necessary
+ *  TIF_NOTIFY_RESUME  - callback before returning to user
+ *  TIF_USEDFPU                - FPU was used by this task this quantum (SMP)
+ *  TIF_POLLING_NRFLAG - true if poll_idle() is polling TIF_NEED_RESCHED
+ */
+#define TIF_SIGPENDING         0
+#define TIF_NEED_RESCHED       1
+#define TIF_NOTIFY_RESUME      2       /* callback before returning to user */
+#define TIF_SYSCALL_TRACE      8
+#define TIF_POLLING_NRFLAG     16
+#define TIF_MEMDIE             18      /* is terminating due to OOM killer */
+#define TIF_FREEZE             19
+#define TIF_RESTORE_SIGMASK    20
+#define TIF_SINGLESTEP         21
+#define TIF_32BIT              22      /* 32bit process */
+#define TIF_SWITCH_MM          23      /* deferred switch_mm */
+
+#define _TIF_SIGPENDING                (1 << TIF_SIGPENDING)
+#define _TIF_NEED_RESCHED      (1 << TIF_NEED_RESCHED)
+#define _TIF_NOTIFY_RESUME     (1 << TIF_NOTIFY_RESUME)
+#define _TIF_32BIT             (1 << TIF_32BIT)
+
+#define _TIF_WORK_MASK         (_TIF_NEED_RESCHED | _TIF_SIGPENDING | \
+                                _TIF_NOTIFY_RESUME)
+
+#endif /* __KERNEL__ */
+#endif /* __ASM_THREAD_INFO_H */
diff --git a/arch/arm64/kernel/process.c b/arch/arm64/kernel/process.c
new file mode 100644 (file)
index 0000000..f22965e
--- /dev/null
@@ -0,0 +1,408 @@
+/*
+ * Based on arch/arm/kernel/process.c
+ *
+ * Original Copyright (C) 1995  Linus Torvalds
+ * Copyright (C) 1996-2000 Russell King - Converted to ARM.
+ * Copyright (C) 2012 ARM Ltd.
+ *
+ * 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.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include <stdarg.h>
+
+#include <linux/export.h>
+#include <linux/sched.h>
+#include <linux/kernel.h>
+#include <linux/mm.h>
+#include <linux/stddef.h>
+#include <linux/unistd.h>
+#include <linux/user.h>
+#include <linux/delay.h>
+#include <linux/reboot.h>
+#include <linux/interrupt.h>
+#include <linux/kallsyms.h>
+#include <linux/init.h>
+#include <linux/cpu.h>
+#include <linux/elfcore.h>
+#include <linux/pm.h>
+#include <linux/tick.h>
+#include <linux/utsname.h>
+#include <linux/uaccess.h>
+#include <linux/random.h>
+#include <linux/hw_breakpoint.h>
+#include <linux/personality.h>
+#include <linux/notifier.h>
+
+#include <asm/compat.h>
+#include <asm/cacheflush.h>
+#include <asm/processor.h>
+#include <asm/stacktrace.h>
+#include <asm/fpsimd.h>
+
+static void setup_restart(void)
+{
+       /*
+        * Tell the mm system that we are going to reboot -
+        * we may need it to insert some 1:1 mappings so that
+        * soft boot works.
+        */
+       setup_mm_for_reboot();
+
+       /* Clean and invalidate caches */
+       flush_cache_all();
+
+       /* Turn D-cache off */
+       cpu_cache_off();
+
+       /* Push out any further dirty data, and ensure cache is empty */
+       flush_cache_all();
+}
+
+void soft_restart(unsigned long addr)
+{
+       setup_restart();
+       cpu_reset(addr);
+}
+
+/*
+ * Function pointers to optional machine specific functions
+ */
+void (*pm_power_off)(void);
+EXPORT_SYMBOL_GPL(pm_power_off);
+
+void (*pm_restart)(const char *cmd);
+EXPORT_SYMBOL_GPL(pm_restart);
+
+
+/*
+ * This is our default idle handler.
+ */
+static void default_idle(void)
+{
+       /*
+        * This should do all the clock switching and wait for interrupt
+        * tricks
+        */
+       cpu_do_idle();
+       local_irq_enable();
+}
+
+void (*pm_idle)(void) = default_idle;
+EXPORT_SYMBOL_GPL(pm_idle);
+
+/*
+ * The idle thread, has rather strange semantics for calling pm_idle,
+ * but this is what x86 does and we need to do the same, so that
+ * things like cpuidle get called in the same way.  The only difference
+ * is that we always respect 'hlt_counter' to prevent low power idle.
+ */
+void cpu_idle(void)
+{
+       local_fiq_enable();
+
+       /* endless idle loop with no priority at all */
+       while (1) {
+               tick_nohz_idle_enter();
+               rcu_idle_enter();
+               while (!need_resched()) {
+                       /*
+                        * We need to disable interrupts here to ensure
+                        * we don't miss a wakeup call.
+                        */
+                       local_irq_disable();
+                       if (!need_resched()) {
+                               stop_critical_timings();
+                               pm_idle();
+                               start_critical_timings();
+                               /*
+                                * pm_idle functions should always return
+                                * with IRQs enabled.
+                                */
+                               WARN_ON(irqs_disabled());
+                       } else {
+                               local_irq_enable();
+                       }
+               }
+               rcu_idle_exit();
+               tick_nohz_idle_exit();
+               schedule_preempt_disabled();
+       }
+}
+
+void machine_shutdown(void)
+{
+#ifdef CONFIG_SMP
+       smp_send_stop();
+#endif
+}
+
+void machine_halt(void)
+{
+       machine_shutdown();
+       while (1);
+}
+
+void machine_power_off(void)
+{
+       machine_shutdown();
+       if (pm_power_off)
+               pm_power_off();
+}
+
+void machine_restart(char *cmd)
+{
+       machine_shutdown();
+
+       /* Disable interrupts first */
+       local_irq_disable();
+       local_fiq_disable();
+
+       /* Now call the architecture specific reboot code. */
+       if (pm_restart)
+               pm_restart(cmd);
+
+       /*
+        * Whoops - the architecture was unable to reboot.
+        */
+       printk("Reboot failed -- System halted\n");
+       while (1);
+}
+
+void __show_regs(struct pt_regs *regs)
+{
+       int i;
+
+       printk("CPU: %d    %s  (%s %.*s)\n",
+               raw_smp_processor_id(), print_tainted(),
+               init_utsname()->release,
+               (int)strcspn(init_utsname()->version, " "),
+               init_utsname()->version);
+       print_symbol("PC is at %s\n", instruction_pointer(regs));
+       print_symbol("LR is at %s\n", regs->regs[30]);
+       printk("pc : [<%016llx>] lr : [<%016llx>] pstate: %08llx\n",
+              regs->pc, regs->regs[30], regs->pstate);
+       printk("sp : %016llx\n", regs->sp);
+       for (i = 29; i >= 0; i--) {
+               printk("x%-2d: %016llx ", i, regs->regs[i]);
+               if (i % 2 == 0)
+                       printk("\n");
+       }
+       printk("\n");
+}
+
+void show_regs(struct pt_regs * regs)
+{
+       printk("\n");
+       printk("Pid: %d, comm: %20s\n", task_pid_nr(current), current->comm);
+       __show_regs(regs);
+}
+
+/*
+ * Free current thread data structures etc..
+ */
+void exit_thread(void)
+{
+}
+
+void flush_thread(void)
+{
+       fpsimd_flush_thread();
+       flush_ptrace_hw_breakpoint(current);
+}
+
+void release_thread(struct task_struct *dead_task)
+{
+}
+
+int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
+{
+       fpsimd_save_state(&current->thread.fpsimd_state);
+       *dst = *src;
+       return 0;
+}
+
+asmlinkage void ret_from_fork(void) asm("ret_from_fork");
+
+int copy_thread(unsigned long clone_flags, unsigned long stack_start,
+               unsigned long stk_sz, struct task_struct *p,
+               struct pt_regs *regs)
+{
+       struct pt_regs *childregs = task_pt_regs(p);
+       unsigned long tls = p->thread.tp_value;
+
+       *childregs = *regs;
+       childregs->regs[0] = 0;
+
+       if (is_compat_thread(task_thread_info(p)))
+               childregs->compat_sp = stack_start;
+       else {
+               /*
+                * Read the current TLS pointer from tpidr_el0 as it may be
+                * out-of-sync with the saved value.
+                */
+               asm("mrs %0, tpidr_el0" : "=r" (tls));
+               childregs->sp = stack_start;
+       }
+
+       memset(&p->thread.cpu_context, 0, sizeof(struct cpu_context));
+       p->thread.cpu_context.sp = (unsigned long)childregs;
+       p->thread.cpu_context.pc = (unsigned long)ret_from_fork;
+
+       /* If a TLS pointer was passed to clone, use that for the new thread. */
+       if (clone_flags & CLONE_SETTLS)
+               tls = regs->regs[3];
+       p->thread.tp_value = tls;
+
+       ptrace_hw_copy_thread(p);
+
+       return 0;
+}
+
+static void tls_thread_switch(struct task_struct *next)
+{
+       unsigned long tpidr, tpidrro;
+
+       if (!is_compat_task()) {
+               asm("mrs %0, tpidr_el0" : "=r" (tpidr));
+               current->thread.tp_value = tpidr;
+       }
+
+       if (is_compat_thread(task_thread_info(next))) {
+               tpidr = 0;
+               tpidrro = next->thread.tp_value;
+       } else {
+               tpidr = next->thread.tp_value;
+               tpidrro = 0;
+       }
+
+       asm(
+       "       msr     tpidr_el0, %0\n"
+       "       msr     tpidrro_el0, %1"
+       : : "r" (tpidr), "r" (tpidrro));
+}
+
+/*
+ * Thread switching.
+ */
+struct task_struct *__switch_to(struct task_struct *prev,
+                               struct task_struct *next)
+{
+       struct task_struct *last;
+
+       fpsimd_thread_switch(next);
+       tls_thread_switch(next);
+       hw_breakpoint_thread_switch(next);
+
+       /* the actual thread switch */
+       last = cpu_switch_to(prev, next);
+
+       return last;
+}
+
+/*
+ * Fill in the task's elfregs structure for a core dump.
+ */
+int dump_task_regs(struct task_struct *t, elf_gregset_t *elfregs)
+{
+       elf_core_copy_regs(elfregs, task_pt_regs(t));
+       return 1;
+}
+
+/*
+ * fill in the fpe structure for a core dump...
+ */
+int dump_fpu (struct pt_regs *regs, struct user_fp *fp)
+{
+       return 0;
+}
+EXPORT_SYMBOL(dump_fpu);
+
+/*
+ * Shuffle the argument into the correct register before calling the
+ * thread function.  x1 is the thread argument, x2 is the pointer to
+ * the thread function, and x3 points to the exit function.
+ */
+extern void kernel_thread_helper(void);
+asm(   ".section .text\n"
+"      .align\n"
+"      .type   kernel_thread_helper, #function\n"
+"kernel_thread_helper:\n"
+"      mov     x0, x1\n"
+"      mov     x30, x3\n"
+"      br      x2\n"
+"      .size   kernel_thread_helper, . - kernel_thread_helper\n"
+"      .previous");
+
+#define kernel_thread_exit     do_exit
+
+/*
+ * Create a kernel thread.
+ */
+pid_t kernel_thread(int (*fn)(void *), void *arg, unsigned long flags)
+{
+       struct pt_regs regs;
+
+       memset(&regs, 0, sizeof(regs));
+
+       regs.regs[1] = (unsigned long)arg;
+       regs.regs[2] = (unsigned long)fn;
+       regs.regs[3] = (unsigned long)kernel_thread_exit;
+       regs.pc = (unsigned long)kernel_thread_helper;
+       regs.pstate = PSR_MODE_EL1h;
+
+       return do_fork(flags|CLONE_VM|CLONE_UNTRACED, 0, &regs, 0, NULL, NULL);
+}
+EXPORT_SYMBOL(kernel_thread);
+
+unsigned long get_wchan(struct task_struct *p)
+{
+       struct stackframe frame;
+       int count = 0;
+       if (!p || p == current || p->state == TASK_RUNNING)
+               return 0;
+
+       frame.fp = thread_saved_fp(p);
+       frame.sp = thread_saved_sp(p);
+       frame.pc = thread_saved_pc(p);
+       do {
+               int ret = unwind_frame(&frame);
+               if (ret < 0)
+                       return 0;
+               if (!in_sched_functions(frame.pc))
+                       return frame.pc;
+       } while (count ++ < 16);
+       return 0;
+}
+
+unsigned long arch_align_stack(unsigned long sp)
+{
+       if (!(current->personality & ADDR_NO_RANDOMIZE) && randomize_va_space)
+               sp -= get_random_int() & ~PAGE_MASK;
+       return sp & ~0xf;
+}
+
+static unsigned long randomize_base(unsigned long base)
+{
+       unsigned long range_end = base + (STACK_RND_MASK << PAGE_SHIFT) + 1;
+       return randomize_range(base, range_end, 0) ? : base;
+}
+
+unsigned long arch_randomize_brk(struct mm_struct *mm)
+{
+       return randomize_base(mm->brk);
+}
+
+unsigned long randomize_et_dyn(unsigned long base)
+{
+       return randomize_base(base);
+}
diff --git a/arch/arm64/mm/context.c b/arch/arm64/mm/context.c
new file mode 100644 (file)
index 0000000..baa758d
--- /dev/null
@@ -0,0 +1,159 @@
+/*
+ * Based on arch/arm/mm/context.c
+ *
+ * Copyright (C) 2002-2003 Deep Blue Solutions Ltd, all rights reserved.
+ * Copyright (C) 2012 ARM Ltd.
+ *
+ * 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.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program.  If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include <linux/init.h>
+#include <linux/sched.h>
+#include <linux/mm.h>
+#include <linux/smp.h>
+#include <linux/percpu.h>
+
+#include <asm/mmu_context.h>
+#include <asm/tlbflush.h>
+#include <asm/cachetype.h>
+
+#define asid_bits(reg) \
+       (((read_cpuid(ID_AA64MMFR0_EL1) & 0xf0) >> 2) + 8)
+
+#define ASID_FIRST_VERSION     (1 << MAX_ASID_BITS)
+
+static DEFINE_RAW_SPINLOCK(cpu_asid_lock);
+unsigned int cpu_last_asid = ASID_FIRST_VERSION;
+
+/*
+ * We fork()ed a process, and we need a new context for the child to run in.
+ */
+void __init_new_context(struct task_struct *tsk, struct mm_struct *mm)
+{
+       mm->context.id = 0;
+       raw_spin_lock_init(&mm->context.id_lock);
+}
+
+static void flush_context(void)
+{
+       /* set the reserved TTBR0 before flushing the TLB */
+       cpu_set_reserved_ttbr0();
+       flush_tlb_all();
+       if (icache_is_aivivt())
+               __flush_icache_all();
+}
+
+#ifdef CONFIG_SMP
+
+static void set_mm_context(struct mm_struct *mm, unsigned int asid)
+{
+       unsigned long flags;
+
+       /*
+        * Locking needed for multi-threaded applications where the same
+        * mm->context.id could be set from different CPUs during the
+        * broadcast. This function is also called via IPI so the
+        * mm->context.id_lock has to be IRQ-safe.
+        */
+       raw_spin_lock_irqsave(&mm->context.id_lock, flags);
+       if (likely((mm->context.id ^ cpu_last_asid) >> MAX_ASID_BITS)) {
+               /*
+                * Old version of ASID found. Set the new one and reset
+                * mm_cpumask(mm).
+                */
+               mm->context.id = asid;
+               cpumask_clear(mm_cpumask(mm));
+       }
+       raw_spin_unlock_irqrestore(&mm->context.id_lock, flags);
+
+       /*
+        * Set the mm_cpumask(mm) bit for the current CPU.
+        */
+       cpumask_set_cpu(smp_processor_id(), mm_cpumask(mm));
+}
+
+/*
+ * Reset the ASID on the current CPU. This function call is broadcast from the
+ * CPU handling the ASID rollover and holding cpu_asid_lock.
+ */
+static void reset_context(void *info)
+{
+       unsigned int asid;
+       unsigned int cpu = smp_processor_id();
+       struct mm_struct *mm = current->active_mm;
+
+       smp_rmb();
+       asid = cpu_last_asid + cpu;
+
+       flush_context();
+       set_mm_context(mm, asid);
+
+       /* set the new ASID */
+       cpu_switch_mm(mm->pgd, mm);
+}
+
+#else
+
+static inline void set_mm_context(struct mm_struct *mm, unsigned int asid)
+{
+       mm->context.id = asid;
+       cpumask_copy(mm_cpumask(mm), cpumask_of(smp_processor_id()));
+}
+
+#endif
+
+void __new_context(struct mm_struct *mm)
+{
+       unsigned int asid;
+       unsigned int bits = asid_bits();
+
+       raw_spin_lock(&cpu_asid_lock);
+#ifdef CONFIG_SMP
+       /*
+        * Check the ASID again, in case the change was broadcast from another
+        * CPU before we acquired the lock.
+        */
+       if (!unlikely((mm->context.id ^ cpu_last_asid) >> MAX_ASID_BITS)) {
+               cpumask_set_cpu(smp_processor_id(), mm_cpumask(mm));
+               raw_spin_unlock(&cpu_asid_lock);
+               return;
+       }
+#endif
+       /*
+        * At this point, it is guaranteed that the current mm (with an old
+        * ASID) isn't active on any other CPU since the ASIDs are changed
+        * simultaneously via IPI.
+        */
+       asid = ++cpu_last_asid;
+
+       /*
+        * If we've used up all our ASIDs, we need to start a new version and
+        * flush the TLB.
+        */
+       if (unlikely((asid & ((1 << bits) - 1)) == 0)) {
+               /* increment the ASID version */
+               cpu_last_asid += (1 << MAX_ASID_BITS) - (1 << bits);
+               if (cpu_last_asid == 0)
+                       cpu_last_asid = ASID_FIRST_VERSION;
+               asid = cpu_last_asid + smp_processor_id();
+               flush_context();
+#ifdef CONFIG_SMP
+               smp_wmb();
+               smp_call_function(reset_context, NULL, 1);
+#endif
+               cpu_last_asid += NR_CPUS - 1;
+       }
+
+       set_mm_context(mm, asid);
+       raw_spin_unlock(&cpu_asid_lock);
+}