The sqpoll thread can be used for performing the napi busy poll in a
similar way that it does io polling for file systems supporting direct
access bypassing the page cache.
The other way that io_uring can be used for napi busy poll is by
calling io_uring_enter() to get events.
If the user specify a timeout value, it is distributed between polling
and sleeping by using the systemwide setting
/proc/sys/net/core/busy_poll.
The changes have been tested with this program:
https://github.com/lano1106/io_uring_udp_ping
and the result is:
Without sqpoll:
NAPI busy loop disabled:
rtt min/avg/max/mdev = 40.631/42.050/58.667/1.547 us
NAPI busy loop enabled:
rtt min/avg/max/mdev = 30.619/31.753/61.433/1.456 us
With sqpoll:
NAPI busy loop disabled:
rtt min/avg/max/mdev = 42.087/44.438/59.508/1.533 us
NAPI busy loop enabled:
rtt min/avg/max/mdev = 35.779/37.347/52.201/0.924 us
v2:
* Evaluate list_empty(&ctx->napi_list) outside io_napi_busy_loop() to keep
__io_sq_thread() execution as fast as possible
* In io_cqring_wait(), move up the sig block to avoid needless computation
if the block exits the function
* In io_cqring_wait(), protect ctx->napi_list from race condition by
splicing it into a local list
* In io_cqring_wait(), allow busy polling when uts is missing
* Fix kernel test robot issues
v3:
* Fix do_div() type mismatch warning
* Reduce uring_lock contention by creating a spinlock for protecting
napi_list
* Support correctly MULTISHOT poll requests
v4:
* Put back benchmark result in commit text
v5:
* Protect napi_list from concurrent access from io_workers threads
Olivier Langlois (2):
io_uring: minor io_cqring_wait() optimization
io_uring: Add support for napi_busy_poll
fs/io_uring.c | 248 ++++++++++++++++++++++++++++++++++++++++++++++++--
1 file changed, 239 insertions(+), 9 deletions(-)
--
2.35.1
Move up the block manipulating the sig variable to execute code
that may encounter an error and exit first before continuing
executing the rest of the function and avoid useless computations
Signed-off-by: Olivier Langlois <[email protected]>
---
fs/io_uring.c | 16 ++++++++--------
1 file changed, 8 insertions(+), 8 deletions(-)
diff --git a/fs/io_uring.c b/fs/io_uring.c
index 4715980e9015..f7b8df79a02b 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -7732,14 +7732,6 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
break;
} while (1);
- if (uts) {
- struct timespec64 ts;
-
- if (get_timespec64(&ts, uts))
- return -EFAULT;
- timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
- }
-
if (sig) {
#ifdef CONFIG_COMPAT
if (in_compat_syscall())
@@ -7753,6 +7745,14 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
return ret;
}
+ if (uts) {
+ struct timespec64 ts;
+
+ if (get_timespec64(&ts, uts))
+ return -EFAULT;
+ timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
+ }
+
init_waitqueue_func_entry(&iowq.wq, io_wake_function);
iowq.wq.private = current;
INIT_LIST_HEAD(&iowq.wq.entry);
--
2.35.1
The sqpoll thread can be used for performing the napi busy poll in a
similar way that it does io polling for file systems supporting direct
access bypassing the page cache.
The other way that io_uring can be used for napi busy poll is by
calling io_uring_enter() to get events.
If the user specify a timeout value, it is distributed between polling
and sleeping by using the systemwide setting
/proc/sys/net/core/busy_poll.
The changes have been tested with this program:
https://github.com/lano1106/io_uring_udp_ping
and the result is:
Without sqpoll:
NAPI busy loop disabled:
rtt min/avg/max/mdev = 40.631/42.050/58.667/1.547 us
NAPI busy loop enabled:
rtt min/avg/max/mdev = 30.619/31.753/61.433/1.456 us
With sqpoll:
NAPI busy loop disabled:
rtt min/avg/max/mdev = 42.087/44.438/59.508/1.533 us
NAPI busy loop enabled:
rtt min/avg/max/mdev = 35.779/37.347/52.201/0.924 us
Co-developed-by: Hao Xu <[email protected]>
Signed-off-by: Hao Xu <[email protected]>
Signed-off-by: Olivier Langlois <[email protected]>
---
fs/io_uring.c | 232 +++++++++++++++++++++++++++++++++++++++++++++++++-
1 file changed, 231 insertions(+), 1 deletion(-)
diff --git a/fs/io_uring.c b/fs/io_uring.c
index f7b8df79a02b..82f306034761 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -63,6 +63,7 @@
#include <net/sock.h>
#include <net/af_unix.h>
#include <net/scm.h>
+#include <net/busy_poll.h>
#include <linux/anon_inodes.h>
#include <linux/sched/mm.h>
#include <linux/uaccess.h>
@@ -395,6 +396,11 @@ struct io_ring_ctx {
struct list_head sqd_list;
unsigned long check_cq_overflow;
+#ifdef CONFIG_NET_RX_BUSY_POLL
+ /* used to track busy poll napi_id */
+ struct list_head napi_list;
+ spinlock_t napi_lock; /* napi_list lock */
+#endif
struct {
unsigned cached_cq_tail;
@@ -1464,6 +1470,10 @@ static __cold struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
INIT_WQ_LIST(&ctx->locked_free_list);
INIT_DELAYED_WORK(&ctx->fallback_work, io_fallback_req_func);
INIT_WQ_LIST(&ctx->submit_state.compl_reqs);
+#ifdef CONFIG_NET_RX_BUSY_POLL
+ INIT_LIST_HEAD(&ctx->napi_list);
+ spin_lock_init(&ctx->napi_lock);
+#endif
return ctx;
err:
kfree(ctx->dummy_ubuf);
@@ -5399,6 +5409,108 @@ IO_NETOP_FN(send);
IO_NETOP_FN(recv);
#endif /* CONFIG_NET */
+#ifdef CONFIG_NET_RX_BUSY_POLL
+
+#define NAPI_TIMEOUT (60 * SEC_CONVERSION)
+
+struct napi_entry {
+ struct list_head list;
+ unsigned int napi_id;
+ unsigned long timeout;
+};
+
+/*
+ * Add busy poll NAPI ID from sk.
+ */
+static void io_add_napi(struct file *file, struct io_ring_ctx *ctx)
+{
+ unsigned int napi_id;
+ struct socket *sock;
+ struct sock *sk;
+ struct napi_entry *ne;
+
+ if (!net_busy_loop_on())
+ return;
+
+ sock = sock_from_file(file);
+ if (!sock)
+ return;
+
+ sk = sock->sk;
+ if (!sk)
+ return;
+
+ napi_id = READ_ONCE(sk->sk_napi_id);
+
+ /* Non-NAPI IDs can be rejected */
+ if (napi_id < MIN_NAPI_ID)
+ return;
+
+ spin_lock(&ctx->napi_lock);
+ list_for_each_entry(ne, &ctx->napi_list, list) {
+ if (ne->napi_id == napi_id) {
+ ne->timeout = jiffies + NAPI_TIMEOUT;
+ goto out;
+ }
+ }
+
+ ne = kmalloc(sizeof(*ne), GFP_NOWAIT);
+ if (!ne)
+ goto out;
+
+ ne->napi_id = napi_id;
+ ne->timeout = jiffies + NAPI_TIMEOUT;
+ list_add_tail(&ne->list, &ctx->napi_list);
+out:
+ spin_unlock(&ctx->napi_lock);
+}
+
+static inline void io_check_napi_entry_timeout(struct napi_entry *ne)
+{
+ if (time_after(jiffies, ne->timeout)) {
+ list_del(&ne->list);
+ kfree(ne);
+ }
+}
+
+/*
+ * Busy poll if globally on and supporting sockets found
+ */
+static bool io_napi_busy_loop(struct list_head *napi_list)
+{
+ struct napi_entry *ne, *n;
+
+ list_for_each_entry_safe(ne, n, napi_list, list) {
+ napi_busy_loop(ne->napi_id, NULL, NULL, true,
+ BUSY_POLL_BUDGET);
+ io_check_napi_entry_timeout(ne);
+ }
+ return !list_empty(napi_list);
+}
+
+static void io_free_napi_list(struct io_ring_ctx *ctx)
+{
+ spin_lock(&ctx->napi_lock);
+ while (!list_empty(&ctx->napi_list)) {
+ struct napi_entry *ne =
+ list_first_entry(&ctx->napi_list, struct napi_entry,
+ list);
+
+ list_del(&ne->list);
+ kfree(ne);
+ }
+ spin_unlock(&ctx->napi_lock);
+}
+#else
+static inline void io_add_napi(struct file *file, struct io_ring_ctx *ctx)
+{
+}
+
+static inline void io_free_napi_list(struct io_ring_ctx *ctx)
+{
+}
+#endif /* CONFIG_NET_RX_BUSY_POLL */
+
struct io_poll_table {
struct poll_table_struct pt;
struct io_kiocb *req;
@@ -5545,6 +5657,7 @@ static int io_poll_check_events(struct io_kiocb *req)
if (unlikely(!filled))
return -ECANCELED;
io_cqring_ev_posted(ctx);
+ io_add_napi(req->file, ctx);
} else if (req->result) {
return 0;
}
@@ -5777,6 +5890,7 @@ static int __io_arm_poll_handler(struct io_kiocb *req,
__io_poll_execute(req, mask);
return 0;
}
+ io_add_napi(req->file, req->ctx);
/*
* Release ownership. If someone tried to queue a tw while it was
@@ -7519,7 +7633,13 @@ static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
!(ctx->flags & IORING_SETUP_R_DISABLED))
ret = io_submit_sqes(ctx, to_submit);
mutex_unlock(&ctx->uring_lock);
-
+#ifdef CONFIG_NET_RX_BUSY_POLL
+ spin_lock(&ctx->napi_lock);
+ if (!list_empty(&ctx->napi_list) &&
+ io_napi_busy_loop(&ctx->napi_list))
+ ++ret;
+ spin_unlock(&ctx->napi_lock);
+#endif
if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
wake_up(&ctx->sqo_sq_wait);
if (creds)
@@ -7650,6 +7770,9 @@ struct io_wait_queue {
struct io_ring_ctx *ctx;
unsigned cq_tail;
unsigned nr_timeouts;
+#ifdef CONFIG_NET_RX_BUSY_POLL
+ unsigned busy_poll_to;
+#endif
};
static inline bool io_should_wake(struct io_wait_queue *iowq)
@@ -7711,6 +7834,87 @@ static inline int io_cqring_wait_schedule(struct io_ring_ctx *ctx,
return 1;
}
+#ifdef CONFIG_NET_RX_BUSY_POLL
+static void io_adjust_busy_loop_timeout(struct timespec64 *ts,
+ struct io_wait_queue *iowq)
+{
+ unsigned busy_poll_to = READ_ONCE(sysctl_net_busy_poll);
+ struct timespec64 pollto = ns_to_timespec64(1000 * (s64)busy_poll_to);
+
+ if (timespec64_compare(ts, &pollto) > 0) {
+ *ts = timespec64_sub(*ts, pollto);
+ iowq->busy_poll_to = busy_poll_to;
+ } else {
+ u64 to = timespec64_to_ns(ts);
+
+ do_div(to, 1000);
+ iowq->busy_poll_to = to;
+ ts->tv_sec = 0;
+ ts->tv_nsec = 0;
+ }
+}
+
+static inline bool io_busy_loop_timeout(unsigned long start_time,
+ unsigned long bp_usec)
+{
+ if (bp_usec) {
+ unsigned long end_time = start_time + bp_usec;
+ unsigned long now = busy_loop_current_time();
+
+ return time_after(now, end_time);
+ }
+ return true;
+}
+
+static bool io_busy_loop_end(void *p, unsigned long start_time)
+{
+ struct io_wait_queue *iowq = p;
+
+ return signal_pending(current) ||
+ io_should_wake(iowq) ||
+ io_busy_loop_timeout(start_time, iowq->busy_poll_to);
+}
+
+static void io_blocking_napi_busy_loop(struct list_head *napi_list,
+ struct io_wait_queue *iowq)
+{
+ unsigned long start_time =
+ list_is_singular(napi_list) ? 0 :
+ busy_loop_current_time();
+
+ do {
+ if (list_is_singular(napi_list)) {
+ struct napi_entry *ne =
+ list_first_entry(napi_list,
+ struct napi_entry, list);
+
+ napi_busy_loop(ne->napi_id, io_busy_loop_end, iowq,
+ true, BUSY_POLL_BUDGET);
+ io_check_napi_entry_timeout(ne);
+ break;
+ }
+ } while (io_napi_busy_loop(napi_list) &&
+ !io_busy_loop_end(iowq, start_time));
+}
+
+static void io_putback_napi_list(struct io_ring_ctx *ctx,
+ struct list_head *napi_list)
+{
+ struct napi_entry *cne, *lne;
+
+ spin_lock(&ctx->napi_lock);
+ list_for_each_entry(cne, &ctx->napi_list, list)
+ list_for_each_entry(lne, napi_list, list)
+ if (cne->napi_id == lne->napi_id) {
+ list_del(&lne->list);
+ kfree(lne);
+ break;
+ }
+ list_splice(napi_list, &ctx->napi_list);
+ spin_unlock(&ctx->napi_lock);
+}
+#endif /* CONFIG_NET_RX_BUSY_POLL */
+
/*
* Wait until events become available, if we don't already have some. The
* application must reap them itself, as they reside on the shared cq ring.
@@ -7723,6 +7927,9 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
struct io_rings *rings = ctx->rings;
ktime_t timeout = KTIME_MAX;
int ret;
+#ifdef CONFIG_NET_RX_BUSY_POLL
+ LIST_HEAD(local_napi_list);
+#endif
do {
io_cqring_overflow_flush(ctx);
@@ -7745,13 +7952,29 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
return ret;
}
+#ifdef CONFIG_NET_RX_BUSY_POLL
+ iowq.busy_poll_to = 0;
+ if (!(ctx->flags & IORING_SETUP_SQPOLL)) {
+ spin_lock(&ctx->napi_lock);
+ list_splice_init(&ctx->napi_list, &local_napi_list);
+ spin_unlock(&ctx->napi_lock);
+ }
+#endif
if (uts) {
struct timespec64 ts;
if (get_timespec64(&ts, uts))
return -EFAULT;
+#ifdef CONFIG_NET_RX_BUSY_POLL
+ if (!list_empty(&local_napi_list))
+ io_adjust_busy_loop_timeout(&ts, &iowq);
+#endif
timeout = ktime_add_ns(timespec64_to_ktime(ts), ktime_get_ns());
}
+#ifdef CONFIG_NET_RX_BUSY_POLL
+ else if (!list_empty(&local_napi_list))
+ iowq.busy_poll_to = READ_ONCE(sysctl_net_busy_poll);
+#endif
init_waitqueue_func_entry(&iowq.wq, io_wake_function);
iowq.wq.private = current;
@@ -7761,6 +7984,12 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events,
iowq.cq_tail = READ_ONCE(ctx->rings->cq.head) + min_events;
trace_io_uring_cqring_wait(ctx, min_events);
+#ifdef CONFIG_NET_RX_BUSY_POLL
+ if (iowq.busy_poll_to)
+ io_blocking_napi_busy_loop(&local_napi_list, &iowq);
+ if (!list_empty(&local_napi_list))
+ io_putback_napi_list(ctx, &local_napi_list);
+#endif
do {
/* if we can't even flush overflow, don't wait for more */
if (!io_cqring_overflow_flush(ctx)) {
@@ -9483,6 +9712,7 @@ static __cold void io_ring_ctx_free(struct io_ring_ctx *ctx)
io_req_caches_free(ctx);
if (ctx->hash_map)
io_wq_put_hash(ctx->hash_map);
+ io_free_napi_list(ctx);
kfree(ctx->cancel_hash);
kfree(ctx->dummy_ubuf);
kfree(ctx);
--
2.35.1
On 3/8/22 3:17 PM, Olivier Langlois wrote:
> Move up the block manipulating the sig variable to execute code
> that may encounter an error and exit first before continuing
> executing the rest of the function and avoid useless computations
I don't think this is worthwhile doing. If you're hitting an error
in any of them, it's by definition not the fast path.
--
Jens Axboe
On 3/8/22 3:17 PM, Olivier Langlois wrote:
> The sqpoll thread can be used for performing the napi busy poll in a
> similar way that it does io polling for file systems supporting direct
> access bypassing the page cache.
>
> The other way that io_uring can be used for napi busy poll is by
> calling io_uring_enter() to get events.
>
> If the user specify a timeout value, it is distributed between polling
> and sleeping by using the systemwide setting
> /proc/sys/net/core/busy_poll.
I think we should get this queued up, but it doesn't apply to
for-5.18/io_uring at all. I can fix it up, but just curious what you
tested against?
--
Jens Axboe
On Tue, 2022-03-08 at 17:54 -0700, Jens Axboe wrote:
> On 3/8/22 3:17 PM, Olivier Langlois wrote:
> > Move up the block manipulating the sig variable to execute code
> > that may encounter an error and exit first before continuing
> > executing the rest of the function and avoid useless computations
>
> I don't think this is worthwhile doing. If you're hitting an error
> in any of them, it's by definition not the fast path.
>
Well, by itself it is not a big improvement but it is still an
improvement.
but most importantly, it has to be considered in the context of the
current patchset because in patch #2, the following step is to
1. acquire the napi spin lock
2. splice the context napi list into a local one.
3. release the lock
If this patch is not in place before patch #2, you would need undo all
that before returning from the sig block which would make the function
bigger when all that is completely avoidable by accepting this patch...
Both patches were together in v1 but I decided to break them apart
thinking that this was the right thing to do...
On Tue, 2022-03-08 at 17:54 -0700, Jens Axboe wrote:
> On 3/8/22 3:17 PM, Olivier Langlois wrote:
> > The sqpoll thread can be used for performing the napi busy poll in
> > a
> > similar way that it does io polling for file systems supporting
> > direct
> > access bypassing the page cache.
> >
> > The other way that io_uring can be used for napi busy poll is by
> > calling io_uring_enter() to get events.
> >
> > If the user specify a timeout value, it is distributed between
> > polling
> > and sleeping by using the systemwide setting
> > /proc/sys/net/core/busy_poll.
>
> I think we should get this queued up, but it doesn't apply to
> for-5.18/io_uring at all. I can fix it up, but just curious what you
> tested against?
>
Hi Jens,
I did wrote the patch from
git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
My testing systems are based on 5.16. I have backported the patch and
compiled 5.16.12 with the patch for my testing.
sorry if I didn't use the right repo...
On 3/8/22 10:36 PM, Olivier Langlois wrote:
> On Tue, 2022-03-08 at 17:54 -0700, Jens Axboe wrote:
>> On 3/8/22 3:17 PM, Olivier Langlois wrote:
>>> Move up the block manipulating the sig variable to execute code
>>> that may encounter an error and exit first before continuing
>>> executing the rest of the function and avoid useless computations
>>
>> I don't think this is worthwhile doing. If you're hitting an error
>> in any of them, it's by definition not the fast path.
>>
> Well, by itself it is not a big improvement but it is still an
> improvement.
>
> but most importantly, it has to be considered in the context of the
> current patchset because in patch #2, the following step is to
>
> 1. acquire the napi spin lock
> 2. splice the context napi list into a local one.
> 3. release the lock
>
> If this patch is not in place before patch #2, you would need undo all
> that before returning from the sig block which would make the function
> bigger when all that is completely avoidable by accepting this patch...
>
> Both patches were together in v1 but I decided to break them apart
> thinking that this was the right thing to do...
Yeah ok, it does make sense in the context of patch 2. Thanks!
--
Jens Axboe
On 3/8/22 10:47 PM, Olivier Langlois wrote:
> On Tue, 2022-03-08 at 17:54 -0700, Jens Axboe wrote:
>> On 3/8/22 3:17 PM, Olivier Langlois wrote:
>>> The sqpoll thread can be used for performing the napi busy poll in
>>> a
>>> similar way that it does io polling for file systems supporting
>>> direct
>>> access bypassing the page cache.
>>>
>>> The other way that io_uring can be used for napi busy poll is by
>>> calling io_uring_enter() to get events.
>>>
>>> If the user specify a timeout value, it is distributed between
>>> polling
>>> and sleeping by using the systemwide setting
>>> /proc/sys/net/core/busy_poll.
>>
>> I think we should get this queued up, but it doesn't apply to
>> for-5.18/io_uring at all. I can fix it up, but just curious what you
>> tested against?
>>
> Hi Jens,
>
> I did wrote the patch from
> git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
>
> My testing systems are based on 5.16. I have backported the patch and
> compiled 5.16.12 with the patch for my testing.
>
> sorry if I didn't use the right repo...
Generally, since the patch is for 5.18, you'd want to base it on my
for-5.18/io_uring branch as then it would apply on top of what is
already pending. But I'll see if I can shoe horn it, unfortunately we'll
hit a merge error but it'll be minor.
--
Jens Axboe
On 3/8/22 3:17 PM, Olivier Langlois wrote:
> The sqpoll thread can be used for performing the napi busy poll in a
> similar way that it does io polling for file systems supporting direct
> access bypassing the page cache.
>
> The other way that io_uring can be used for napi busy poll is by
> calling io_uring_enter() to get events.
>
> If the user specify a timeout value, it is distributed between polling
> and sleeping by using the systemwide setting
> /proc/sys/net/core/busy_poll.
>
> The changes have been tested with this program:
> https://github.com/lano1106/io_uring_udp_ping
>
> and the result is:
> Without sqpoll:
> NAPI busy loop disabled:
> rtt min/avg/max/mdev = 40.631/42.050/58.667/1.547 us
> NAPI busy loop enabled:
> rtt min/avg/max/mdev = 30.619/31.753/61.433/1.456 us
>
> With sqpoll:
> NAPI busy loop disabled:
> rtt min/avg/max/mdev = 42.087/44.438/59.508/1.533 us
> NAPI busy loop enabled:
> rtt min/avg/max/mdev = 35.779/37.347/52.201/0.924 us
Applied, thanks.
--
Jens Axboe