Commit 76b1e9b9 authored by Eliezer Tamir's avatar Eliezer Tamir Committed by David S. Miller

net/fs: change busy poll time accounting

Suggested by Linus:
Changed time accounting for busy-poll:
- Make it microsecond based.
- Use unsigned longs.
- Revert back to use time_after instead of time_in_range.
Reorder poll/select busy loop conditions:
- Clear busy_flag after one time we can't busy-poll.
- Only init busy_end if we actually are going to busy-poll.
Added one more missing need_resched() test.
Signed-off-by: default avatarEliezer Tamir <eliezer.tamir@linux.intel.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
parent cbf55001
...@@ -403,8 +403,7 @@ int do_select(int n, fd_set_bits *fds, struct timespec *end_time) ...@@ -403,8 +403,7 @@ int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
int retval, i, timed_out = 0; int retval, i, timed_out = 0;
unsigned long slack = 0; unsigned long slack = 0;
unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0; unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
u64 busy_start = busy_loop_start_time(busy_flag); unsigned long busy_end = 0;
u64 busy_end = busy_loop_end_time();
rcu_read_lock(); rcu_read_lock();
retval = max_select_fd(n, fds); retval = max_select_fd(n, fds);
...@@ -506,9 +505,15 @@ int do_select(int n, fd_set_bits *fds, struct timespec *end_time) ...@@ -506,9 +505,15 @@ int do_select(int n, fd_set_bits *fds, struct timespec *end_time)
} }
/* only if found POLL_BUSY_LOOP sockets && not out of time */ /* only if found POLL_BUSY_LOOP sockets && not out of time */
if (!need_resched() && can_busy_loop && if (can_busy_loop && !need_resched()) {
busy_loop_range(busy_start, busy_end)) if (!busy_end) {
continue; busy_end = busy_loop_end_time();
continue;
}
if (!busy_loop_timeout(busy_end))
continue;
}
busy_flag = 0;
/* /*
* If this is the first loop and we have a timeout * If this is the first loop and we have a timeout
...@@ -780,9 +785,7 @@ static int do_poll(unsigned int nfds, struct poll_list *list, ...@@ -780,9 +785,7 @@ static int do_poll(unsigned int nfds, struct poll_list *list,
int timed_out = 0, count = 0; int timed_out = 0, count = 0;
unsigned long slack = 0; unsigned long slack = 0;
unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0; unsigned int busy_flag = net_busy_loop_on() ? POLL_BUSY_LOOP : 0;
u64 busy_start = busy_loop_start_time(busy_flag); unsigned long busy_end = 0;
u64 busy_end = busy_loop_end_time();
/* Optimise the no-wait case */ /* Optimise the no-wait case */
if (end_time && !end_time->tv_sec && !end_time->tv_nsec) { if (end_time && !end_time->tv_sec && !end_time->tv_nsec) {
...@@ -834,9 +837,15 @@ static int do_poll(unsigned int nfds, struct poll_list *list, ...@@ -834,9 +837,15 @@ static int do_poll(unsigned int nfds, struct poll_list *list,
break; break;
/* only if found POLL_BUSY_LOOP sockets && not out of time */ /* only if found POLL_BUSY_LOOP sockets && not out of time */
if (!need_resched() && can_busy_loop && if (can_busy_loop && !need_resched()) {
busy_loop_range(busy_start, busy_end)) if (!busy_end) {
continue; busy_end = busy_loop_end_time();
continue;
}
if (!busy_loop_timeout(busy_end))
continue;
}
busy_flag = 0;
/* /*
* If this is the first loop and we have a timeout * If this is the first loop and we have a timeout
......
...@@ -47,7 +47,7 @@ static inline bool net_busy_loop_on(void) ...@@ -47,7 +47,7 @@ static inline bool net_busy_loop_on(void)
* we only care that the average is bounded * we only care that the average is bounded
*/ */
#ifdef CONFIG_DEBUG_PREEMPT #ifdef CONFIG_DEBUG_PREEMPT
static inline u64 busy_loop_sched_clock(void) static inline u64 busy_loop_us_clock(void)
{ {
u64 rc; u64 rc;
...@@ -55,37 +55,24 @@ static inline u64 busy_loop_sched_clock(void) ...@@ -55,37 +55,24 @@ static inline u64 busy_loop_sched_clock(void)
rc = sched_clock(); rc = sched_clock();
preempt_enable_no_resched_notrace(); preempt_enable_no_resched_notrace();
return rc; return rc >> 10;
} }
#else /* CONFIG_DEBUG_PREEMPT */ #else /* CONFIG_DEBUG_PREEMPT */
static inline u64 busy_loop_sched_clock(void) static inline u64 busy_loop_us_clock(void)
{ {
return sched_clock(); return sched_clock() >> 10;
} }
#endif /* CONFIG_DEBUG_PREEMPT */ #endif /* CONFIG_DEBUG_PREEMPT */
/* we don't mind a ~2.5% imprecision so <<10 instead of *1000 static inline unsigned long sk_busy_loop_end_time(struct sock *sk)
* sk->sk_ll_usec is a u_int so this can't overflow
*/
static inline u64 sk_busy_loop_end_time(struct sock *sk)
{ {
return (u64)ACCESS_ONCE(sk->sk_ll_usec) << 10; return busy_loop_us_clock() + ACCESS_ONCE(sk->sk_ll_usec);
} }
/* in poll/select we use the global sysctl_net_ll_poll value /* in poll/select we use the global sysctl_net_ll_poll value */
* only call sched_clock() if enabled static inline unsigned long busy_loop_end_time(void)
*/
static inline u64 busy_loop_end_time(void)
{
return (u64)ACCESS_ONCE(sysctl_net_ll_poll) << 10;
}
/* if flag is not set we don't need to know the time
* so we want to avoid a potentially expensive sched_clock()
*/
static inline u64 busy_loop_start_time(unsigned int flag)
{ {
return flag ? busy_loop_sched_clock() : 0; return busy_loop_us_clock() + ACCESS_ONCE(sysctl_net_ll_poll);
} }
static inline bool sk_can_busy_loop(struct sock *sk) static inline bool sk_can_busy_loop(struct sock *sk)
...@@ -94,12 +81,12 @@ static inline bool sk_can_busy_loop(struct sock *sk) ...@@ -94,12 +81,12 @@ static inline bool sk_can_busy_loop(struct sock *sk)
!need_resched() && !signal_pending(current); !need_resched() && !signal_pending(current);
} }
/* careful! time_in_range64 will evaluate now twice */
static inline bool busy_loop_range(u64 start_time, u64 run_time) static inline bool busy_loop_timeout(unsigned long end_time)
{ {
u64 now = busy_loop_sched_clock(); unsigned long now = busy_loop_us_clock();
return time_in_range64(now, start_time, start_time + run_time); return time_after(now, end_time);
} }
/* when used in sock_poll() nonblock is known at compile time to be true /* when used in sock_poll() nonblock is known at compile time to be true
...@@ -107,8 +94,7 @@ static inline bool busy_loop_range(u64 start_time, u64 run_time) ...@@ -107,8 +94,7 @@ static inline bool busy_loop_range(u64 start_time, u64 run_time)
*/ */
static inline bool sk_busy_loop(struct sock *sk, int nonblock) static inline bool sk_busy_loop(struct sock *sk, int nonblock)
{ {
u64 start_time = busy_loop_start_time(!nonblock); unsigned long end_time = !nonblock ? sk_busy_loop_end_time(sk) : 0;
u64 end_time = sk_busy_loop_end_time(sk);
const struct net_device_ops *ops; const struct net_device_ops *ops;
struct napi_struct *napi; struct napi_struct *napi;
int rc = false; int rc = false;
...@@ -139,7 +125,7 @@ static inline bool sk_busy_loop(struct sock *sk, int nonblock) ...@@ -139,7 +125,7 @@ static inline bool sk_busy_loop(struct sock *sk, int nonblock)
LINUX_MIB_LOWLATENCYRXPACKETS, rc); LINUX_MIB_LOWLATENCYRXPACKETS, rc);
} while (!nonblock && skb_queue_empty(&sk->sk_receive_queue) && } while (!nonblock && skb_queue_empty(&sk->sk_receive_queue) &&
busy_loop_range(start_time, end_time)); !need_resched() && !busy_loop_timeout(end_time));
rc = !skb_queue_empty(&sk->sk_receive_queue); rc = !skb_queue_empty(&sk->sk_receive_queue);
out: out:
...@@ -165,12 +151,7 @@ static inline unsigned long net_busy_loop_on(void) ...@@ -165,12 +151,7 @@ static inline unsigned long net_busy_loop_on(void)
return 0; return 0;
} }
static inline u64 busy_loop_start_time(unsigned int flag) static inline unsigned long busy_loop_end_time(void)
{
return 0;
}
static inline u64 busy_loop_end_time(void)
{ {
return 0; return 0;
} }
...@@ -193,9 +174,9 @@ static inline void sk_mark_ll(struct sock *sk, struct sk_buff *skb) ...@@ -193,9 +174,9 @@ static inline void sk_mark_ll(struct sock *sk, struct sk_buff *skb)
{ {
} }
static inline bool busy_loop_range(u64 start_time, u64 run_time) static inline bool busy_loop_timeout(unsigned long end_time)
{ {
return false; return true;
} }
#endif /* CONFIG_NET_LL_RX_POLL */ #endif /* CONFIG_NET_LL_RX_POLL */
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment