net: xsk: update tx queue consumer immdiately after transmission
Subsystems (4)
PATCH 1/1 net: xsk: update tx queue consumer immdiately after transmission
From: Jason Xing <[email protected]>
For afxdp, the return value of sendto() syscall doesn't reflect how many descs handled in the kernel. One of use cases is that when user-space application tries to know the number of transmitted skbs and then decides if it continues to send, say, is it stopped due to max tx budget?
The following formular can be used after sending to learn how many skbs/descs the kernel takes care of:
tx_queue.consumers_before - tx_queue.consumers_after
Prior to the current patch, the consumer of tx queue is not immdiately updated at the end of each sendto syscall, which leads the consumer value out-of-dated from the perspective of user space. So this patch requires store operation to pass the cached value to the shared value to handle the problem.
--- net/xdp/xsk.c | 2 ++ 1 file changed, 2 insertions(+)
net/xdp/xsk.c +2 -0
@@ -856,6 +856,8 @@ static int __xsk_generic_xmit(struct sock *sk) } out: +__xskq_cons_release (xs->tx); +if (sent_frame)if (xsk_tx_writeable (xs)) sk->sk_write_space (sk);
Signature · 1 line
2.43.5
Quoted 28 lines · Jason Xing
From: Jason Xing <[email protected]>
For afxdp, the return value of
sendto()syscall doesn't reflect how many descs handled in the kernel. One of use cases is that when user-space application tries to know the number of transmitted skbs and then decides if it continues to send, say, is it stopped due to max tx budget?The following formular can be used after sending to learn how many skbs/descs the kernel takes care of:
tx_queue.consumers_before - tx_queue.consumers_after
Prior to the current patch, the consumer of tx queue is not immdiately updated at the end of each sendto syscall, which leads the consumer value out-of-dated from the perspective of user space. So this patch requires store operation to pass the cached value to the shared value to handle the problem.
Signed-off-by Jason Xing <[email protected]>--- net/xdp/xsk.c | 2 ++ 1 file changed, 2 insertions(+)net/xdp/xsk.c +0
-0
net/xdp/xsk.c static int __xsk_generic_xmit(struct sock *sk) · line 863
Quoted 8 lines
net/xdp/xsk.c +2
-0@@ -856,6 +856,8 @@ static int __xsk_generic_xmit(struct sock *sk) } out: +__xskq_cons_release (xs->tx); +if (sent_frame)if (xsk_tx_writeable (xs)) sk->sk_write_space (sk);
So for the "good" case we are going to write the cons twice? From xskq_cons_peek_desc and from here? Maybe make this __xskq_cons_release conditional ('if (err)')?
I also wonder whether we should add a test for that? Should be easy to verify by sending more than 32 packets. Is there a place in tools/testing/selftests/bpf/xskxceiver.c to add that?
Quoted 32 lines · Stanislav Fomichev
Quoted 30 lines · Jason Xing
From: Jason Xing <[email protected]>
For afxdp, the return value of
sendto()syscall doesn't reflect how many descs handled in the kernel. One of use cases is that when user-space application tries to know the number of transmitted skbs and then decides if it continues to send, say, is it stopped due to max tx budget?The following formular can be used after sending to learn how many skbs/descs the kernel takes care of:
tx_queue.consumers_before - tx_queue.consumers_after
Prior to the current patch, the consumer of tx queue is not immdiately updated at the end of each sendto syscall, which leads the consumer value out-of-dated from the perspective of user space. So this patch requires store operation to pass the cached value to the shared value to handle the problem.
Signed-off-by Jason Xing <[email protected]>--- net/xdp/xsk.c | 2 ++ 1 file changed, 2 insertions(+)net/xdp/xsk.c +0
-0@@ -856,6 +856,8 @@ static int __xsk_generic_xmit(struct sock *sk) }
Quoted 10 lines
net/xdp/xsk.c static int __xsk_generic_xmit(struct sock *sk) · line 863
Quoted 6 lines
net/xdp/xsk.c +2
-0out: +__xskq_cons_release (xs->tx); +if (sent_frame)if (xsk_tx_writeable (xs)) sk->sk_write_space (sk);So for the "good" case we are going to write the cons twice? From
xskq_cons_peek_descand from here? Maybe make this__xskq_cons_releaseconditional ('if (err)')?
One unlikely exception:
xskq_cons_peek_desc()->xskq_cons_read_desc()->xskq_cons_is_valid_desc()->return false;
?
There are still two possible 'return false' in xskq_cons_peek_desc() while so far I didn't spot a single one happening.
Admittedly, your suggestion covers the majority of normal good ones. I can adjust it as you said.
Quoted 4 lines
I also wonder whether we should add a test for that? Should be easy to verify by sending more than 32 packets. Is there a place in
tools/testing/selftests/bpf/xskxceiver.cto add that?
Well, sorry, if it's not required, please don't force me to do so :S The patch is only one simple update of the consumer that is shared between user-space and kernel.
Thanks,
Jason
Maciej Fijalkowski
maintainerQuoted 31 lines · Stanislav Fomichev
Quoted 30 lines · Jason Xing
From: Jason Xing <[email protected]>
For afxdp, the return value of
sendto()syscall doesn't reflect how many descs handled in the kernel. One of use cases is that when user-space application tries to know the number of transmitted skbs and then decides if it continues to send, say, is it stopped due to max tx budget?The following formular can be used after sending to learn how many skbs/descs the kernel takes care of:
tx_queue.consumers_before - tx_queue.consumers_after
Prior to the current patch, the consumer of tx queue is not immdiately updated at the end of each sendto syscall, which leads the consumer value out-of-dated from the perspective of user space. So this patch requires store operation to pass the cached value to the shared value to handle the problem.
Signed-off-by Jason Xing <[email protected]>--- net/xdp/xsk.c | 2 ++ 1 file changed, 2 insertions(+)net/xdp/xsk.c +0
-0@@ -856,6 +856,8 @@ static int __xsk_generic_xmit(struct sock *sk) }
Quoted 10 lines
net/xdp/xsk.c static int __xsk_generic_xmit(struct sock *sk) · line 863
Quoted 6 lines
net/xdp/xsk.c +2
-0out: +__xskq_cons_release (xs->tx); +if (sent_frame)if (xsk_tx_writeable (xs)) sk->sk_write_space (sk);So for the "good" case we are going to write the cons twice? From
xskq_cons_peek_descand from here? Maybe make this__xskq_cons_releaseconditional ('if (err)')?
this patch updates a global state of producer whereas generic xmit loop updates local value. this global state is also updated within peeking function.
from quick look patch seems to be correct however my mind is in vacation mode so i'll take a second look on monday.
Quoted 4 lines
I also wonder whether we should add a test for that? Should be easy to verify by sending more than 32 packets. Is there a place in
tools/testing/selftests/bpf/xskxceiver.cto add that?
Quoted 44 lines · Maciej Fijalkowski
Quoted 31 lines · Stanislav Fomichev
Quoted 30 lines · Jason Xing
From: Jason Xing <[email protected]>
For afxdp, the return value of
sendto()syscall doesn't reflect how many descs handled in the kernel. One of use cases is that when user-space application tries to know the number of transmitted skbs and then decides if it continues to send, say, is it stopped due to max tx budget?The following formular can be used after sending to learn how many skbs/descs the kernel takes care of:
tx_queue.consumers_before - tx_queue.consumers_after
Prior to the current patch, the consumer of tx queue is not immdiately updated at the end of each sendto syscall, which leads the consumer value out-of-dated from the perspective of user space. So this patch requires store operation to pass the cached value to the shared value to handle the problem.
Signed-off-by Jason Xing <[email protected]>--- net/xdp/xsk.c | 2 ++ 1 file changed, 2 insertions(+)net/xdp/xsk.c +0
-0@@ -856,6 +856,8 @@ static int __xsk_generic_xmit(struct sock *sk) }Quoted 10 lines
net/xdp/xsk.c static int __xsk_generic_xmit(struct sock *sk) · line 863
Quoted 6 lines
net/xdp/xsk.c +2
-0out: +__xskq_cons_release (xs->tx); +if (sent_frame)if (xsk_tx_writeable (xs)) sk->sk_write_space (sk);So for the "good" case we are going to write the cons twice? From
xskq_cons_peek_descand from here? Maybe make this__xskq_cons_releaseconditional ('if (err)')?
Quoted 3 lines
this patch updates a global state of producer whereas generic xmit loop updates local value. this global state is also updated within peeking function.
Stanislav also pointed out the normal/majority of good cases. I will filter out the good case then.
Quoted 3 lines
from quick look patch seems to be correct however my mind is in vacation mode so i'll take a second look on monday.
Thanks. I'm very sure that the line this patch introduces can be helpful because I manually printk the delta to verify before/after __xskq_cons_release(xs->tx); and then spot a few numbers larger than zero during a simple test.
Thanks,
Jason
Quoted 5 lines
Quoted 5 lines
I also wonder whether we should add a test for that? Should be easy to verify by sending more than 32 packets. Is there a place in
tools/testing/selftests/bpf/xskxceiver.cto add that?
Quoted 57 lines · Jason Xing
Quoted 32 lines · Stanislav Fomichev
Quoted 30 lines · Jason Xing
From: Jason Xing <[email protected]>
For afxdp, the return value of
sendto()syscall doesn't reflect how many descs handled in the kernel. One of use cases is that when user-space application tries to know the number of transmitted skbs and then decides if it continues to send, say, is it stopped due to max tx budget?The following formular can be used after sending to learn how many skbs/descs the kernel takes care of:
tx_queue.consumers_before - tx_queue.consumers_after
Prior to the current patch, the consumer of tx queue is not immdiately updated at the end of each sendto syscall, which leads the consumer value out-of-dated from the perspective of user space. So this patch requires store operation to pass the cached value to the shared value to handle the problem.
Signed-off-by Jason Xing <[email protected]>--- net/xdp/xsk.c | 2 ++ 1 file changed, 2 insertions(+)net/xdp/xsk.c +0
-0@@ -856,6 +856,8 @@ static int __xsk_generic_xmit(struct sock *sk) }Quoted 10 lines
net/xdp/xsk.c static int __xsk_generic_xmit(struct sock *sk) · line 863
Quoted 6 lines
net/xdp/xsk.c +2
-0out: +__xskq_cons_release (xs->tx); +if (sent_frame)if (xsk_tx_writeable (xs)) sk->sk_write_space (sk);So for the "good" case we are going to write the cons twice? From
xskq_cons_peek_descand from here? Maybe make this__xskq_cons_releaseconditional ('if (err)')?One unlikely exception:
xskq_cons_peek_desc()->xskq_cons_read_desc()->xskq_cons_is_valid_desc()->return false;
?There are still two possible 'return false' in
xskq_cons_peek_desc()while so far I didn't spot a single one happening.Admittedly, your suggestion covers the majority of normal good ones. I can adjust it as you said.
Quoted 1 line
Quoted 7 lines
Quoted 3 lines
I also wonder whether we should add a test for that? Should be easy to verify by sending more than 32 packets. Is there a place in
tools/testing/selftests/bpf/xskxceiver.cto add that?Well, sorry, if it's not required, please don't force me to do so :S The patch is only one simple update of the consumer that is shared between user-space and kernel.
My suspicion is that the same issue exists for the zc case. So would be nice to test it and fix it as well :-p
Quoted 63 lines · Stanislav Fomichev
Quoted 61 lines · Jason Xing
Quoted 32 lines · Stanislav Fomichev
Quoted 30 lines · Jason Xing
From: Jason Xing <[email protected]>
For afxdp, the return value of
sendto()syscall doesn't reflect how many descs handled in the kernel. One of use cases is that when user-space application tries to know the number of transmitted skbs and then decides if it continues to send, say, is it stopped due to max tx budget?The following formular can be used after sending to learn how many skbs/descs the kernel takes care of:
tx_queue.consumers_before - tx_queue.consumers_after
Prior to the current patch, the consumer of tx queue is not immdiately updated at the end of each sendto syscall, which leads the consumer value out-of-dated from the perspective of user space. So this patch requires store operation to pass the cached value to the shared value to handle the problem.
Signed-off-by Jason Xing <[email protected]>--- net/xdp/xsk.c | 2 ++ 1 file changed, 2 insertions(+)net/xdp/xsk.c +0
-0@@ -856,6 +856,8 @@ static int __xsk_generic_xmit(struct sock *sk) }Quoted 10 lines
net/xdp/xsk.c static int __xsk_generic_xmit(struct sock *sk) · line 863
Quoted 6 lines
net/xdp/xsk.c +2
-0out: +__xskq_cons_release (xs->tx); +if (sent_frame)if (xsk_tx_writeable (xs)) sk->sk_write_space (sk);So for the "good" case we are going to write the cons twice? From
xskq_cons_peek_descand from here? Maybe make this__xskq_cons_releaseconditional ('if (err)')?One unlikely exception:
xskq_cons_peek_desc()->xskq_cons_read_desc()->xskq_cons_is_valid_desc()->return false;
?There are still two possible 'return false' in
xskq_cons_peek_desc()while so far I didn't spot a single one happening.Admittedly, your suggestion covers the majority of normal good ones. I can adjust it as you said.
Quoted 4 lines
I also wonder whether we should add a test for that? Should be easy to verify by sending more than 32 packets. Is there a place in
tools/testing/selftests/bpf/xskxceiver.cto add that?
Quoted 6 lines
Quoted 3 lines
Well, sorry, if it's not required, please don't force me to do so :S The patch is only one simple update of the consumer that is shared between user-space and kernel.
My suspicion is that the same issue exists for the zc case. So would be nice to test it and fix it as well :-p
Oh, well, I will take a look at how the selftest works in the next few days.
Allow me to ask the question that you asked me before: even though I didn't see the necessity to set the max budget for zc mode (just because I didn't spot it happening), would it be better if we separate both of them because it's an uAPI interface. IIUC, if the setsockopt is set, we will not separate it any more in the future?
Or we can keep using the hardcoded value (32) in the zc mode like before and __only__ touch the copy mode? Then if someone or I found the significance of making it tunable, then another parameter of setsockopt can be added? Does it make sense?
Thanks,
Jason
Quoted 10 lines
Allow me to ask the question that you asked me before: even though I didn't see the necessity to set the max budget for zc mode (just because I didn't spot it happening), would it be better if we separate both of them because it's an uAPI interface. IIUC, if the setsockopt is set, we will not separate it any more in the future?
Or we can keep using the hardcoded value (32) in the zc mode like before and
__only__touch the copy mode? Then if someone or I found the significance of making it tunable, then another parameter of setsockopt can be added? Does it make sense?
I found I replied to a wrong thread. Let me copy&paste there instead.
Thanks,
Jason
Quoted 63 lines · Stanislav Fomichev
Quoted 61 lines · Jason Xing
Quoted 32 lines · Stanislav Fomichev
Quoted 30 lines · Jason Xing
From: Jason Xing <[email protected]>
For afxdp, the return value of
sendto()syscall doesn't reflect how many descs handled in the kernel. One of use cases is that when user-space application tries to know the number of transmitted skbs and then decides if it continues to send, say, is it stopped due to max tx budget?The following formular can be used after sending to learn how many skbs/descs the kernel takes care of:
tx_queue.consumers_before - tx_queue.consumers_after
Prior to the current patch, the consumer of tx queue is not immdiately updated at the end of each sendto syscall, which leads the consumer value out-of-dated from the perspective of user space. So this patch requires store operation to pass the cached value to the shared value to handle the problem.
Signed-off-by Jason Xing <[email protected]>--- net/xdp/xsk.c | 2 ++ 1 file changed, 2 insertions(+)net/xdp/xsk.c +0
-0@@ -856,6 +856,8 @@ static int __xsk_generic_xmit(struct sock *sk) }Quoted 10 lines
net/xdp/xsk.c static int __xsk_generic_xmit(struct sock *sk) · line 863
Quoted 6 lines
net/xdp/xsk.c +2
-0out: +__xskq_cons_release (xs->tx); +if (sent_frame)if (xsk_tx_writeable (xs)) sk->sk_write_space (sk);So for the "good" case we are going to write the cons twice? From
xskq_cons_peek_descand from here? Maybe make this__xskq_cons_releaseconditional ('if (err)')?One unlikely exception:
xskq_cons_peek_desc()->xskq_cons_read_desc()->xskq_cons_is_valid_desc()->return false;
?There are still two possible 'return false' in
xskq_cons_peek_desc()while so far I didn't spot a single one happening.Admittedly, your suggestion covers the majority of normal good ones. I can adjust it as you said.
Quoted 4 lines
I also wonder whether we should add a test for that? Should be easy to verify by sending more than 32 packets. Is there a place in
tools/testing/selftests/bpf/xskxceiver.cto add that?
Quoted 6 lines
Quoted 3 lines
Well, sorry, if it's not required, please don't force me to do so :S The patch is only one simple update of the consumer that is shared between user-space and kernel.
My suspicion is that the same issue exists for the zc case. So would be nice to test it and fix it as well :-p
After digging into the logic around xsk_tx_peek_desc(), I can say that at the end of every caller of xsk_tx_peek_desc(), there is always a xsk_tx_release() function that used to update the local consumer to the global state of consumer. So for the zero copy mode, no need to change at all :)
I will soon send the v2 with the 'if (error)' statement in the __xsk_generic_xmit().
Thanks,
Jason