kernel/context_tracking.c

   1 /*
   2  * Context tracking: Probe on high level context boundaries such as kernel
   3  * and userspace. This includes syscalls and exceptions entry/exit.
   4  *
   5  * This is used by RCU to remove its dependency on the timer tick while a CPU
   6  * runs in userspace.
   7  *
   8  *  Started by Frederic Weisbecker:
   9  *
  10  * Copyright (C) 2012 Red Hat, Inc., Frederic Weisbecker <fweisbec@redhat.com>
  11  *
  12  * Many thanks to Gilad Ben-Yossef, Paul McKenney, Ingo Molnar, Andrew Morton,
  13  * Steven Rostedt, Peter Zijlstra for suggestions and improvements.
  14  *
  15  */
  16
  17 #include <linux/context_tracking.h>
  18 #include <linux/rcupdate.h>
  19 #include <linux/sched.h>
  20 #include <linux/hardirq.h>
  21 #include <linux/export.h>
  22 #include <linux/kprobes.h>
  23
  24 #define CREATE_TRACE_POINTS
  25 #include <trace/events/context_tracking.h>
  26
  27 struct static_key context_tracking_enabled = STATIC_KEY_INIT_FALSE;
  28 EXPORT_SYMBOL_GPL(context_tracking_enabled);
  29
  30 DEFINE_PER_CPU(struct context_tracking, context_tracking);
  31 EXPORT_SYMBOL_GPL(context_tracking);
  32
  33 static bool context_tracking_recursion_enter(void)
  34 {
  35         int recursion;
  36
  37         recursion = __this_cpu_inc_return(context_tracking.recursion);
  38         if (recursion == 1)
  39                 return true;
  40
  41         WARN_ONCE((recursion < 1), "Invalid context tracking recursion value %d\n", recursion);
  42         __this_cpu_dec(context_tracking.recursion);
  43
  44         return false;
  45 }
  46
  47 static void context_tracking_recursion_exit(void)
  48 {
  49         __this_cpu_dec(context_tracking.recursion);
  50 }
  51
  52 /**
  53  * context_tracking_enter - Inform the context tracking that the CPU is going
  54  *                          enter user or guest space mode.
  55  *
  56  * This function must be called right before we switch from the kernel
  57  * to user or guest space, when it's guaranteed the remaining kernel
  58  * instructions to execute won't use any RCU read side critical section
  59  * because this function sets RCU in extended quiescent state.
  60  */
  61 void context_tracking_enter(enum ctx_state state)
  62 {
  63         unsigned long flags;
  64
  65         /*
  66          * Some contexts may involve an exception occuring in an irq,
  67          * leading to that nesting:
  68          * rcu_irq_enter() rcu_user_exit() rcu_user_exit() rcu_irq_exit()
  69          * This would mess up the dyntick_nesting count though. And rcu_irq_*()
  70          * helpers are enough to protect RCU uses inside the exception. So
  71          * just return immediately if we detect we are in an IRQ.
  72          */
  73         if (in_interrupt())
  74                 return;
  75
  76         /* Kernel threads aren't supposed to go to userspace */
  77         WARN_ON_ONCE(!current->mm);
  78
  79         local_irq_save(flags);
  80         if (!context_tracking_recursion_enter())
  81                 goto out_irq_restore;
  82
  83         if ( __this_cpu_read(context_tracking.state) != state) {
  84                 if (__this_cpu_read(context_tracking.active)) {
  85                         /*
  86                          * At this stage, only low level arch entry code remains and
  87                          * then we'll run in userspace. We can assume there won't be
  88                          * any RCU read-side critical section until the next call to
  89                          * user_exit() or rcu_irq_enter(). Let's remove RCU's dependency
  90                          * on the tick.
  91                          */
  92                         if (state == CONTEXT_USER) {
  93                                 trace_user_enter(0);
  94                                 vtime_user_enter(current);
  95                         }
  96                         rcu_user_enter();
  97                 }
  98                 /*
  99                  * Even if context tracking is disabled on this CPU, because it's outside
 100                  * the full dynticks mask for example, we still have to keep track of the
 101                  * context transitions and states to prevent inconsistency on those of
 102                  * other CPUs.
 103                  * If a task triggers an exception in userspace, sleep on the exception
 104                  * handler and then migrate to another CPU, that new CPU must know where
 105                  * the exception returns by the time we call exception_exit().
 106                  * This information can only be provided by the previous CPU when it called
 107                  * exception_enter().
 108                  * OTOH we can spare the calls to vtime and RCU when context_tracking.active
 109                  * is false because we know that CPU is not tickless.
 110                  */
 111                 __this_cpu_write(context_tracking.state, state);
 112         }
 113         context_tracking_recursion_exit();
 114 out_irq_restore:
 115         local_irq_restore(flags);
 116 }
 117 NOKPROBE_SYMBOL(context_tracking_enter);
 118 EXPORT_SYMBOL_GPL(context_tracking_enter);
 119
 120 void context_tracking_user_enter(void)
 121 {
 122         user_enter();
 123 }
 124 NOKPROBE_SYMBOL(context_tracking_user_enter);
 125
 126 /**
 127  * context_tracking_exit - Inform the context tracking that the CPU is
 128  *                         exiting user or guest mode and entering the kernel.
 129  *
 130  * This function must be called after we entered the kernel from user or
 131  * guest space before any use of RCU read side critical section. This
 132  * potentially include any high level kernel code like syscalls, exceptions,
 133  * signal handling, etc...
 134  *
 135  * This call supports re-entrancy. This way it can be called from any exception
 136  * handler without needing to know if we came from userspace or not.
 137  */
 138 void context_tracking_exit(enum ctx_state state)
 139 {
 140         unsigned long flags;
 141
 142         if (in_interrupt())
 143                 return;
 144
 145         local_irq_save(flags);
 146         if (!context_tracking_recursion_enter())
 147                 goto out_irq_restore;
 148
 149         if (__this_cpu_read(context_tracking.state) == state) {
 150                 if (__this_cpu_read(context_tracking.active)) {
 151                         /*
 152                          * We are going to run code that may use RCU. Inform
 153                          * RCU core about that (ie: we may need the tick again).
 154                          */
 155                         rcu_user_exit();
 156                         if (state == CONTEXT_USER) {
 157                                 vtime_user_exit(current);
 158                                 trace_user_exit(0);
 159                         }
 160                 }
 161                 __this_cpu_write(context_tracking.state, CONTEXT_KERNEL);
 162         }
 163         context_tracking_recursion_exit();
 164 out_irq_restore:
 165         local_irq_restore(flags);
 166 }
 167 NOKPROBE_SYMBOL(context_tracking_exit);
 168 EXPORT_SYMBOL_GPL(context_tracking_exit);
 169
 170 void context_tracking_user_exit(void)
 171 {
 172         user_exit();
 173 }
 174 NOKPROBE_SYMBOL(context_tracking_user_exit);
 175
 176 void __init context_tracking_cpu_set(int cpu)
 177 {
 178         static __initdata bool initialized = false;
 179
 180         if (!per_cpu(context_tracking.active, cpu)) {
 181                 per_cpu(context_tracking.active, cpu) = true;
 182                 static_key_slow_inc(&context_tracking_enabled);
 183         }
 184
 185         if (initialized)
 186                 return;
 187
 188         /*
 189          * Set TIF_NOHZ to init/0 and let it propagate to all tasks through fork
 190          * This assumes that init is the only task at this early boot stage.
 191          */
 192         set_tsk_thread_flag(&init_task, TIF_NOHZ);
 193         WARN_ON_ONCE(!tasklist_empty());
 194
 195         initialized = true;
 196 }
 197
 198 #ifdef CONFIG_CONTEXT_TRACKING_FORCE
 199 void __init context_tracking_init(void)
 200 {
 201         int cpu;
 202
 203         for_each_possible_cpu(cpu)
 204                 context_tracking_cpu_set(cpu);
 205 }
 206 #endif