aboutsummaryrefslogtreecommitdiffstats
path: root/fs/timerfd.c
diff options
context:
space:
mode:
authorThomas Gleixner <tglx@linutronix.de>2011-05-20 10:18:50 -0400
committerThomas Gleixner <tglx@linutronix.de>2011-05-23 07:59:53 -0400
commit9ec2690758a5467f24beb301cca5098078073bba (patch)
treee5bc78f690d12635a56460ea6f54b49318221dc8 /fs/timerfd.c
parent250f972d85effad5b6e10da4bbd877e6a4b503b6 (diff)
timerfd: Manage cancelable timers in timerfd
Peter is concerned about the extra scan of CLOCK_REALTIME_COS in the timer interrupt. Yes, I did not think about it, because the solution was so elegant. I didn't like the extra list in timerfd when it was proposed some time ago, but with a rcu based list the list walk it's less horrible than the original global lock, which was held over the list iteration. Requested-by: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Peter Zijlstra <peterz@infradead.org>
Diffstat (limited to 'fs/timerfd.c')
-rw-r--r--fs/timerfd.c105
1 files changed, 74 insertions, 31 deletions
diff --git a/fs/timerfd.c b/fs/timerfd.c
index 7e14c9e7c4ee..f67acbdda5e8 100644
--- a/fs/timerfd.c
+++ b/fs/timerfd.c
@@ -22,6 +22,7 @@
22#include <linux/anon_inodes.h> 22#include <linux/anon_inodes.h>
23#include <linux/timerfd.h> 23#include <linux/timerfd.h>
24#include <linux/syscalls.h> 24#include <linux/syscalls.h>
25#include <linux/rcupdate.h>
25 26
26struct timerfd_ctx { 27struct timerfd_ctx {
27 struct hrtimer tmr; 28 struct hrtimer tmr;
@@ -31,9 +32,14 @@ struct timerfd_ctx {
31 u64 ticks; 32 u64 ticks;
32 int expired; 33 int expired;
33 int clockid; 34 int clockid;
35 struct rcu_head rcu;
36 struct list_head clist;
34 bool might_cancel; 37 bool might_cancel;
35}; 38};
36 39
40static LIST_HEAD(cancel_list);
41static DEFINE_SPINLOCK(cancel_lock);
42
37/* 43/*
38 * This gets called when the timer event triggers. We set the "expired" 44 * This gets called when the timer event triggers. We set the "expired"
39 * flag, but we do not re-arm the timer (in case it's necessary, 45 * flag, but we do not re-arm the timer (in case it's necessary,
@@ -53,28 +59,69 @@ static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr)
53 return HRTIMER_NORESTART; 59 return HRTIMER_NORESTART;
54} 60}
55 61
56static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx) 62/*
63 * Called when the clock was set to cancel the timers in the cancel
64 * list.
65 */
66void timerfd_clock_was_set(void)
57{ 67{
58 ktime_t remaining; 68 ktime_t moffs = ktime_get_monotonic_offset();
69 struct timerfd_ctx *ctx;
70 unsigned long flags;
59 71
60 remaining = hrtimer_expires_remaining(&ctx->tmr); 72 rcu_read_lock();
61 return remaining.tv64 < 0 ? ktime_set(0, 0): remaining; 73 list_for_each_entry_rcu(ctx, &cancel_list, clist) {
74 if (!ctx->might_cancel)
75 continue;
76 spin_lock_irqsave(&ctx->wqh.lock, flags);
77 if (ctx->moffs.tv64 != moffs.tv64) {
78 ctx->moffs.tv64 = KTIME_MAX;
79 wake_up_locked(&ctx->wqh);
80 }
81 spin_unlock_irqrestore(&ctx->wqh.lock, flags);
82 }
83 rcu_read_unlock();
62} 84}
63 85
64static bool timerfd_canceled(struct timerfd_ctx *ctx) 86static void timerfd_remove_cancel(struct timerfd_ctx *ctx)
65{ 87{
66 ktime_t moffs; 88 if (ctx->might_cancel) {
89 ctx->might_cancel = false;
90 spin_lock(&cancel_lock);
91 list_del_rcu(&ctx->clist);
92 spin_unlock(&cancel_lock);
93 }
94}
67 95
68 if (!ctx->might_cancel) 96static bool timerfd_canceled(struct timerfd_ctx *ctx)
97{
98 if (!ctx->might_cancel || ctx->moffs.tv64 != KTIME_MAX)
69 return false; 99 return false;
100 ctx->moffs = ktime_get_monotonic_offset();
101 return true;
102}
70 103
71 moffs = ktime_get_monotonic_offset(); 104static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags)
105{
106 if (ctx->clockid == CLOCK_REALTIME && (flags & TFD_TIMER_ABSTIME) &&
107 (flags & TFD_TIMER_CANCEL_ON_SET)) {
108 if (!ctx->might_cancel) {
109 ctx->might_cancel = true;
110 spin_lock(&cancel_lock);
111 list_add_rcu(&ctx->clist, &cancel_list);
112 spin_unlock(&cancel_lock);
113 }
114 } else if (ctx->might_cancel) {
115 timerfd_remove_cancel(ctx);
116 }
117}
72 118
73 if (moffs.tv64 == ctx->moffs.tv64) 119static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx)
74 return false; 120{
121 ktime_t remaining;
75 122
76 ctx->moffs = moffs; 123 remaining = hrtimer_expires_remaining(&ctx->tmr);
77 return true; 124 return remaining.tv64 < 0 ? ktime_set(0, 0): remaining;
78} 125}
79 126
80static int timerfd_setup(struct timerfd_ctx *ctx, int flags, 127static int timerfd_setup(struct timerfd_ctx *ctx, int flags,
@@ -87,13 +134,6 @@ static int timerfd_setup(struct timerfd_ctx *ctx, int flags,
87 htmode = (flags & TFD_TIMER_ABSTIME) ? 134 htmode = (flags & TFD_TIMER_ABSTIME) ?
88 HRTIMER_MODE_ABS: HRTIMER_MODE_REL; 135 HRTIMER_MODE_ABS: HRTIMER_MODE_REL;
89 136
90 ctx->might_cancel = false;
91 if (htmode == HRTIMER_MODE_ABS && ctx->clockid == CLOCK_REALTIME &&
92 (flags & TFD_TIMER_CANCELON_SET)) {
93 clockid = CLOCK_REALTIME_COS;
94 ctx->might_cancel = true;
95 }
96
97 texp = timespec_to_ktime(ktmr->it_value); 137 texp = timespec_to_ktime(ktmr->it_value);
98 ctx->expired = 0; 138 ctx->expired = 0;
99 ctx->ticks = 0; 139 ctx->ticks = 0;
@@ -113,8 +153,9 @@ static int timerfd_release(struct inode *inode, struct file *file)
113{ 153{
114 struct timerfd_ctx *ctx = file->private_data; 154 struct timerfd_ctx *ctx = file->private_data;
115 155
156 timerfd_remove_cancel(ctx);
116 hrtimer_cancel(&ctx->tmr); 157 hrtimer_cancel(&ctx->tmr);
117 kfree(ctx); 158 kfree_rcu(ctx, rcu);
118 return 0; 159 return 0;
119} 160}
120 161
@@ -149,20 +190,20 @@ static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
149 else 190 else
150 res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks); 191 res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
151 192
193 /*
194 * If clock has changed, we do not care about the
195 * ticks and we do not rearm the timer. Userspace must
196 * reevaluate anyway.
197 */
198 if (timerfd_canceled(ctx)) {
199 ctx->ticks = 0;
200 ctx->expired = 0;
201 res = -ECANCELED;
202 }
203
152 if (ctx->ticks) { 204 if (ctx->ticks) {
153 ticks = ctx->ticks; 205 ticks = ctx->ticks;
154 206
155 /*
156 * If clock has changed, we do not care about the
157 * ticks and we do not rearm the timer. Userspace must
158 * reevaluate anyway.
159 */
160 if (timerfd_canceled(ctx)) {
161 ticks = 0;
162 ctx->expired = 0;
163 res = -ECANCELED;
164 }
165
166 if (ctx->expired && ctx->tintv.tv64) { 207 if (ctx->expired && ctx->tintv.tv64) {
167 /* 208 /*
168 * If tintv.tv64 != 0, this is a periodic timer that 209 * If tintv.tv64 != 0, this is a periodic timer that
@@ -258,6 +299,8 @@ SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
258 return PTR_ERR(file); 299 return PTR_ERR(file);
259 ctx = file->private_data; 300 ctx = file->private_data;
260 301
302 timerfd_setup_cancel(ctx, flags);
303
261 /* 304 /*
262 * We need to stop the existing timer before reprogramming 305 * We need to stop the existing timer before reprogramming
263 * it to the new values. 306 * it to the new values.