| 1 |
/* -*- linux-c -*- |
| 2 |
* Context Runtime Functions |
| 3 |
* Copyright (C) 2014 Red Hat Inc. |
| 4 |
* |
| 5 |
* This file is part of systemtap, and is free software. You can |
| 6 |
* redistribute it and/or modify it under the terms of the GNU General |
| 7 |
* Public License (GPL); either version 2, or (at your option) any |
| 8 |
* later version. |
| 9 |
*/ |
| 10 |
|
| 11 |
#ifndef _LINUX_RUNTIME_CONTEXT_H_ |
| 12 |
#define _LINUX_RUNTIME_CONTEXT_H_ |
| 13 |
|
| 14 |
/* Can't use a lock primitive for this because lock_acquire() has tracepoints */ |
| 15 |
static atomic_t _stp_contexts_busy_ctr = ATOMIC_INIT(0); |
| 16 |
|
| 17 |
/* Per-cpu data is always initialized with zero filled. */ |
| 18 |
static DEFINE_PER_CPU(struct context *, contexts); |
| 19 |
|
| 20 |
static int _stp_runtime_contexts_alloc(void) |
| 21 |
{ |
| 22 |
unsigned int cpu; |
| 23 |
|
| 24 |
/* We don't use for_aech_possible_cpu() here since the number of possible |
| 25 |
* CPUs may be very large even though there are many fewere online CPUs. |
| 26 |
* For example, VMWare guests usually have 128 possible CPUs while only |
| 27 |
* have a few online CPUs. Once the context structs were |
| 28 |
* allocated for online CPUs at this point, we will discard any context |
| 29 |
* fetching operations on any future online CPUs dynamically added |
| 30 |
* through the kernel's CPU hotplug feature. Memory allocations of the |
| 31 |
* context structs can only happen right here. |
| 32 |
*/ |
| 33 |
for_each_online_cpu(cpu) { |
| 34 |
/* Module init, so in user context, safe to use |
| 35 |
* "sleeping" allocation. */ |
| 36 |
struct context *c = _stp_vzalloc_node(sizeof (struct context), |
| 37 |
cpu_to_node(cpu)); |
| 38 |
if (unlikely(c == NULL)) { |
| 39 |
_stp_error ("context (size %lu per cpu) allocation failed", |
| 40 |
(unsigned long) sizeof (struct context)); |
| 41 |
return -ENOMEM; |
| 42 |
} |
| 43 |
per_cpu(contexts, cpu) = c; |
| 44 |
} |
| 45 |
return 0; |
| 46 |
} |
| 47 |
|
| 48 |
static bool _stp_runtime_context_trylock(void) |
| 49 |
{ |
| 50 |
bool locked; |
| 51 |
|
| 52 |
/* Need to disable preemption because of the smp_processor_id() call |
| 53 |
in _stp_runtime_get_context(). */ |
| 54 |
preempt_disable(); |
| 55 |
|
| 56 |
/* fast path to ignore new online CPUs without percpu context memory |
| 57 |
* allocations. this also serves as an extra safe guard for NULL context |
| 58 |
* pointers. */ |
| 59 |
if (unlikely(_stp_runtime_get_context() == NULL)) { |
| 60 |
preempt_enable_no_resched(); |
| 61 |
return false; |
| 62 |
} |
| 63 |
|
| 64 |
locked = atomic_add_unless(&_stp_contexts_busy_ctr, 1, INT_MAX); |
| 65 |
if (!locked) |
| 66 |
preempt_enable_no_resched(); |
| 67 |
|
| 68 |
return locked; |
| 69 |
} |
| 70 |
|
| 71 |
static void _stp_runtime_context_unlock(void) |
| 72 |
{ |
| 73 |
atomic_dec(&_stp_contexts_busy_ctr); |
| 74 |
preempt_enable_no_resched(); |
| 75 |
} |
| 76 |
|
| 77 |
/* We should be free of all probes by this time, but for example the timer for |
| 78 |
* _stp_ctl_work_callback may still be running and looking for contexts. We |
| 79 |
* use _stp_contexts_busy_ctr to be sure its safe to free them. */ |
| 80 |
static void _stp_runtime_contexts_free(void) |
| 81 |
{ |
| 82 |
unsigned int cpu; |
| 83 |
|
| 84 |
/* Sync to make sure existing readers are done */ |
| 85 |
while (atomic_cmpxchg(&_stp_contexts_busy_ctr, 0, INT_MAX)) |
| 86 |
cpu_relax(); |
| 87 |
|
| 88 |
/* Now we can actually free the contexts */ |
| 89 |
|
| 90 |
/* NB We cannot use the for_each_online_cpu() here since online |
| 91 |
* CPUs may get changed on-the-fly through the CPU hotplug feature |
| 92 |
* of the kernel. We only allocated the context structs on original |
| 93 |
* online CPUs when _stp_runtime_contexts_alloc() was called. And we |
| 94 |
* cannot allocate new memory from within this context. */ |
| 95 |
for_each_possible_cpu(cpu) { |
| 96 |
struct context *c = per_cpu(contexts, cpu); |
| 97 |
if (likely(c)) |
| 98 |
_stp_vfree(c); |
| 99 |
} |
| 100 |
} |
| 101 |
|
| 102 |
static inline struct context * _stp_runtime_get_context(void) |
| 103 |
{ |
| 104 |
return per_cpu(contexts, smp_processor_id()); |
| 105 |
} |
| 106 |
|
| 107 |
static struct context * _stp_runtime_entryfn_get_context(void) |
| 108 |
{ |
| 109 |
struct context* __restrict__ c = NULL; |
| 110 |
|
| 111 |
if (!_stp_runtime_context_trylock()) |
| 112 |
return NULL; |
| 113 |
|
| 114 |
c = _stp_runtime_get_context(); |
| 115 |
if (c != NULL) { |
| 116 |
if (!atomic_cmpxchg(&c->busy, 0, 1)) { |
| 117 |
// NB: Notice we're not releasing _stp_contexts_busy_ctr |
| 118 |
// here. We exepect the calling code to call |
| 119 |
// _stp_runtime_entryfn_get_context() and |
| 120 |
// _stp_runtime_entryfn_put_context() as a |
| 121 |
// pair. |
| 122 |
return c; |
| 123 |
} |
| 124 |
} |
| 125 |
_stp_runtime_context_unlock(); |
| 126 |
return NULL; |
| 127 |
} |
| 128 |
|
| 129 |
static inline void _stp_runtime_entryfn_put_context(struct context *c) |
| 130 |
{ |
| 131 |
if (c) { |
| 132 |
atomic_set(&c->busy, 0); |
| 133 |
_stp_runtime_context_unlock(); |
| 134 |
} |
| 135 |
} |
| 136 |
|
| 137 |
static void _stp_runtime_context_wait(void) |
| 138 |
{ |
| 139 |
int holdon; |
| 140 |
unsigned long hold_start; |
| 141 |
int hold_index; |
| 142 |
|
| 143 |
hold_start = jiffies; |
| 144 |
hold_index = -1; |
| 145 |
do { |
| 146 |
int i; |
| 147 |
|
| 148 |
holdon = 0; |
| 149 |
if (!_stp_runtime_context_trylock()) |
| 150 |
break; |
| 151 |
|
| 152 |
for_each_possible_cpu(i) { |
| 153 |
struct context *c = per_cpu(contexts, i); |
| 154 |
if (c != NULL |
| 155 |
&& atomic_read (& c->busy)) { |
| 156 |
holdon = 1; |
| 157 |
|
| 158 |
/* Just in case things are really |
| 159 |
* stuck, let's print some diagnostics. */ |
| 160 |
if (time_after(jiffies, hold_start + HZ) // > 1 second |
| 161 |
&& (i > hold_index)) { // not already printed |
| 162 |
hold_index = i; |
| 163 |
printk(KERN_ERR "%s context[%d] stuck: %s\n", THIS_MODULE->name, i, c->probe_point); |
| 164 |
} |
| 165 |
} |
| 166 |
} |
| 167 |
_stp_runtime_context_unlock(); |
| 168 |
|
| 169 |
/* |
| 170 |
* Just in case things are really really stuck, a |
| 171 |
* handler probably suffered a fault, and the kernel |
| 172 |
* probably killed a task/thread already. We can't be |
| 173 |
* quite sure in what state everything is in, however |
| 174 |
* auxiliary stuff like kprobes / uprobes / locks have |
| 175 |
* already been unregistered. So it's *probably* safe |
| 176 |
* to pretend/assume/hope everything is OK, and let |
| 177 |
* the cleanup finish. |
| 178 |
* |
| 179 |
* In the worst case, there may occur a fault, as a |
| 180 |
* genuinely running probe handler tries to access |
| 181 |
* script globals (about to be freed), or something |
| 182 |
* accesses module memory (about to be unloaded). |
| 183 |
* This is sometimes stinky, so the alternative |
| 184 |
* (default) is to change from a livelock to a |
| 185 |
* livelock that sleeps awhile. |
| 186 |
*/ |
| 187 |
#ifdef STAP_OVERRIDE_STUCK_CONTEXT |
| 188 |
if (time_after(jiffies, hold_start + HZ*10)) { // > 10 seconds |
| 189 |
printk(KERN_ERR "%s overriding stuck context to allow module shutdown.", THIS_MODULE->name); |
| 190 |
holdon = 0; // allow loop to exit |
| 191 |
} |
| 192 |
#else |
| 193 |
/* at least stop sucking down the staprun cpu */ |
| 194 |
msleep(250); |
| 195 |
#endif |
| 196 |
|
| 197 |
/* NB: we run at least one of these during the |
| 198 |
* shutdown sequence: */ |
| 199 |
yield(); /* aka schedule() and then some */ |
| 200 |
} while (holdon); |
| 201 |
} |
| 202 |
|
| 203 |
#endif /* _LINUX_RUNTIME_CONTEXT_H_ */ |