On Mon, Oct 12, 2020 at 5:24 PM Eric Dumazet <[email protected]> wrote:
>
>
>
> On 10/12/20 10:39 AM, Muchun Song wrote:
> > On Mon, Oct 12, 2020 at 3:42 PM Eric Dumazet <[email protected]> wrote:
> >>
> >> On Mon, Oct 12, 2020 at 6:22 AM Muchun Song <[email protected]> wrote:
> >>>
> >>> On Mon, Oct 12, 2020 at 2:39 AM Cong Wang <[email protected]> wrote:
> >>>>
> >>>> On Sat, Oct 10, 2020 at 3:39 AM Muchun Song <[email protected]> wrote:
> >>>>>
> >>>>> The amount of memory allocated to sockets buffer can become significant.
> >>>>> However, we do not display the amount of memory consumed by sockets
> >>>>> buffer. In this case, knowing where the memory is consumed by the kernel
> >>>>
> >>>> We do it via `ss -m`. Is it not sufficient? And if not, why not adding it there
> >>>> rather than /proc/meminfo?
> >>>
> >>> If the system has little free memory, we can know where the memory is via
> >>> /proc/meminfo. If a lot of memory is consumed by socket buffer, we cannot
> >>> know it when the Sock is not shown in the /proc/meminfo. If the unaware user
> >>> can't think of the socket buffer, naturally they will not `ss -m`. The
> >>> end result
> >>> is that we still don’t know where the memory is consumed. And we add the
> >>> Sock to the /proc/meminfo just like the memcg does('sock' item in the cgroup
> >>> v2 memory.stat). So I think that adding to /proc/meminfo is sufficient.
> >>>
> >>>>
> >>>>> static inline void __skb_frag_unref(skb_frag_t *frag)
> >>>>> {
> >>>>> - put_page(skb_frag_page(frag));
> >>>>> + struct page *page = skb_frag_page(frag);
> >>>>> +
> >>>>> + if (put_page_testzero(page)) {
> >>>>> + dec_sock_node_page_state(page);
> >>>>> + __put_page(page);
> >>>>> + }
> >>>>> }
> >>>>
> >>>> You mix socket page frag with skb frag at least, not sure this is exactly
> >>>> what you want, because clearly skb page frags are frequently used
> >>>> by network drivers rather than sockets.
> >>>>
> >>>> Also, which one matches this dec_sock_node_page_state()? Clearly
> >>>> not skb_fill_page_desc() or __skb_frag_ref().
> >>>
> >>> Yeah, we call inc_sock_node_page_state() in the skb_page_frag_refill().
> >>> So if someone gets the page returned by skb_page_frag_refill(), it must
> >>> put the page via __skb_frag_unref()/skb_frag_unref(). We use PG_private
> >>> to indicate that we need to dec the node page state when the refcount of
> >>> page reaches zero.
> >>>
> >>
> >> Pages can be transferred from pipe to socket, socket to pipe (splice()
> >> and zerocopy friends...)
> >>
> >> If you want to track TCP memory allocations, you always can look at
> >> /proc/net/sockstat,
> >> without adding yet another expensive memory accounting.
> >
> > The 'mem' item in the /proc/net/sockstat does not represent real
> > memory usage. This is just the total amount of charged memory.
> >
> > For example, if a task sends a 10-byte message, it only charges one
> > page to memcg. But the system may allocate 8 pages. Therefore, it
> > does not truly reflect the memory allocated by the above memory
> > allocation path. We can see the difference via the following message.
> >
> > cat /proc/net/sockstat
> > sockets: used 698
> > TCP: inuse 70 orphan 0 tw 617 alloc 134 mem 13
> > UDP: inuse 90 mem 4
> > UDPLITE: inuse 0
> > RAW: inuse 1
> > FRAG: inuse 0 memory 0
> >
> > cat /proc/meminfo | grep Sock
> > Sock: 13664 kB
> >
> > The /proc/net/sockstat only shows us that there are 17*4 kB TCP
> > memory allocations. But apply this patch, we can see that we truly
> > allocate 13664 kB(May be greater than this value because of per-cpu
> > stat cache). Of course the load of the example here is not high. In
> > some high load cases, I believe the difference here will be even
> > greater.
> >
>
> This is great, but you have not addressed my feedback.
>
> TCP memory allocations are bounded by /proc/sys/net/ipv4/tcp_mem
>
> Fact that the memory is forward allocated or not is a detail.
>
> If you think we must pre-allocate memory, instead of forward allocations,
> your patch does not address this. Adding one line per consumer in /proc/meminfo looks
> wrong to me.
I think that the consumer which consumes a lot of memory should be added
to the /proc/meminfo. This can help us know the user of large memory.
>
> If you do not want 9.37 % of physical memory being possibly used by TCP,
> just change /proc/sys/net/ipv4/tcp_mem accordingly ?
We are not complaining about TCP using too much memory, but how do
we know that TCP uses a lot of memory. When I firstly face this problem,
I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
If we can know the amount memory of the socket buffer via /proc/meminfo, we
may not need to spend a lot of time troubleshooting this problem. Not everyone
knows that a lot of memory may be used here. But I believe many people
should know /proc/meminfo to confirm memory users.
Thanks.
>
>
--
Yours,
Muchun
On Mon, Oct 12, 2020 at 2:53 AM Muchun Song <[email protected]> wrote:
> We are not complaining about TCP using too much memory, but how do
> we know that TCP uses a lot of memory. When I firstly face this problem,
> I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
> If we can know the amount memory of the socket buffer via /proc/meminfo, we
> may not need to spend a lot of time troubleshooting this problem. Not everyone
> knows that a lot of memory may be used here. But I believe many people
> should know /proc/meminfo to confirm memory users.
Well, I'd bet networking people know `ss -m` better than /proc/meminfo,
generally speaking.
The practice here is that if you want some networking-specific counters,
add it to where networking people know better, that is, `ss -m` or /proc/net/...
Or maybe the problem you described is not specific to networking at all,
there must be some other places where pages are allocated but not charged.
If so, adding a general mm counter in /proc/meminfo makes sense, but
it won't be specific to networking.
Thanks.
On Tue, Oct 13, 2020 at 6:12 AM Cong Wang <[email protected]> wrote:
>
> On Mon, Oct 12, 2020 at 2:53 AM Muchun Song <[email protected]> wrote:
> > We are not complaining about TCP using too much memory, but how do
> > we know that TCP uses a lot of memory. When I firstly face this problem,
> > I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
> > If we can know the amount memory of the socket buffer via /proc/meminfo, we
> > may not need to spend a lot of time troubleshooting this problem. Not everyone
> > knows that a lot of memory may be used here. But I believe many people
> > should know /proc/meminfo to confirm memory users.
>
> Well, I'd bet networking people know `ss -m` better than /proc/meminfo,
I agree with you. But if someone(not networking people) faces the same
problem. He may suspect that there is a memory leak or think that a certain
driver allocates memory but has no statistics. He only saw the memory
disappeared via /proc/meminfo.
> generally speaking.
>
> The practice here is that if you want some networking-specific counters,
> add it to where networking people know better, that is, `ss -m` or /proc/net/...
>
> Or maybe the problem you described is not specific to networking at all,
> there must be some other places where pages are allocated but not charged.
Yeah, it is not charged. The allocation path is as follows. This allocation
consumes 25GB memory on our server. And it belongs to the network core.
Thanks.
__alloc_pages_nodemask+0x11d/0x290
skb_page_frag_refill+0x68/0xf0
sk_page_frag_refill+0x19/0x70
tcp_sendmsg_locked+0x2f4/0xd10
tcp_sendmsg+0x29/0xa0
sock_sendmsg+0x30/0x40
sock_write_iter+0x8f/0x100
__vfs_write+0x10b/0x190
vfs_write+0xb0/0x190
ksys_write+0x5a/0xd0
do_syscall_64+0x5d/0x110
entry_SYSCALL_64_after_hwframe+0x44/0xa9
> If so, adding a general mm counter in /proc/meminfo makes sense, but
> it won't be specific to networking.
>
> Thanks.
--
Yours,
Muchun
On Mon, Oct 12, 2020 at 05:53:01PM +0800, Muchun Song wrote:
> On Mon, Oct 12, 2020 at 5:24 PM Eric Dumazet <[email protected]> wrote:
> >
> > On 10/12/20 10:39 AM, Muchun Song wrote:
> > > On Mon, Oct 12, 2020 at 3:42 PM Eric Dumazet <[email protected]> wrote:
> > >>
> > >> On Mon, Oct 12, 2020 at 6:22 AM Muchun Song <[email protected]> wrote:
> > >>>
> > >>> On Mon, Oct 12, 2020 at 2:39 AM Cong Wang <[email protected]> wrote:
> > >>>>
> > >>>> On Sat, Oct 10, 2020 at 3:39 AM Muchun Song <[email protected]> wrote:
> > >>>>>
> > >>>>> The amount of memory allocated to sockets buffer can become significant.
> > >>>>> However, we do not display the amount of memory consumed by sockets
> > >>>>> buffer. In this case, knowing where the memory is consumed by the kernel
> > >>>>
> > >>>> We do it via `ss -m`. Is it not sufficient? And if not, why not adding it there
> > >>>> rather than /proc/meminfo?
> > >>>
> > >>> If the system has little free memory, we can know where the memory is via
> > >>> /proc/meminfo. If a lot of memory is consumed by socket buffer, we cannot
> > >>> know it when the Sock is not shown in the /proc/meminfo. If the unaware user
> > >>> can't think of the socket buffer, naturally they will not `ss -m`. The
> > >>> end result
> > >>> is that we still don’t know where the memory is consumed. And we add the
> > >>> Sock to the /proc/meminfo just like the memcg does('sock' item in the cgroup
> > >>> v2 memory.stat). So I think that adding to /proc/meminfo is sufficient.
> > >>>
> > >>>>
> > >>>>> static inline void __skb_frag_unref(skb_frag_t *frag)
> > >>>>> {
> > >>>>> - put_page(skb_frag_page(frag));
> > >>>>> + struct page *page = skb_frag_page(frag);
> > >>>>> +
> > >>>>> + if (put_page_testzero(page)) {
> > >>>>> + dec_sock_node_page_state(page);
> > >>>>> + __put_page(page);
> > >>>>> + }
> > >>>>> }
> > >>>>
> > >>>> You mix socket page frag with skb frag at least, not sure this is exactly
> > >>>> what you want, because clearly skb page frags are frequently used
> > >>>> by network drivers rather than sockets.
> > >>>>
> > >>>> Also, which one matches this dec_sock_node_page_state()? Clearly
> > >>>> not skb_fill_page_desc() or __skb_frag_ref().
> > >>>
> > >>> Yeah, we call inc_sock_node_page_state() in the skb_page_frag_refill().
> > >>> So if someone gets the page returned by skb_page_frag_refill(), it must
> > >>> put the page via __skb_frag_unref()/skb_frag_unref(). We use PG_private
> > >>> to indicate that we need to dec the node page state when the refcount of
> > >>> page reaches zero.
> > >>>
> > >>
> > >> Pages can be transferred from pipe to socket, socket to pipe (splice()
> > >> and zerocopy friends...)
> > >>
> > >> If you want to track TCP memory allocations, you always can look at
> > >> /proc/net/sockstat,
> > >> without adding yet another expensive memory accounting.
> > >
> > > The 'mem' item in the /proc/net/sockstat does not represent real
> > > memory usage. This is just the total amount of charged memory.
> > >
> > > For example, if a task sends a 10-byte message, it only charges one
> > > page to memcg. But the system may allocate 8 pages. Therefore, it
> > > does not truly reflect the memory allocated by the above memory
> > > allocation path. We can see the difference via the following message.
> > >
> > > cat /proc/net/sockstat
> > > sockets: used 698
> > > TCP: inuse 70 orphan 0 tw 617 alloc 134 mem 13
> > > UDP: inuse 90 mem 4
> > > UDPLITE: inuse 0
> > > RAW: inuse 1
> > > FRAG: inuse 0 memory 0
> > >
> > > cat /proc/meminfo | grep Sock
> > > Sock: 13664 kB
> > >
> > > The /proc/net/sockstat only shows us that there are 17*4 kB TCP
> > > memory allocations. But apply this patch, we can see that we truly
> > > allocate 13664 kB(May be greater than this value because of per-cpu
> > > stat cache). Of course the load of the example here is not high. In
> > > some high load cases, I believe the difference here will be even
> > > greater.
> > >
> >
> > This is great, but you have not addressed my feedback.
> >
> > TCP memory allocations are bounded by /proc/sys/net/ipv4/tcp_mem
> >
> > Fact that the memory is forward allocated or not is a detail.
> >
> > If you think we must pre-allocate memory, instead of forward allocations,
> > your patch does not address this. Adding one line per consumer in /proc/meminfo looks
> > wrong to me.
>
> I think that the consumer which consumes a lot of memory should be added
> to the /proc/meminfo. This can help us know the user of large memory.
>
> >
> > If you do not want 9.37 % of physical memory being possibly used by TCP,
> > just change /proc/sys/net/ipv4/tcp_mem accordingly ?
>
> We are not complaining about TCP using too much memory, but how do
> we know that TCP uses a lot of memory. When I firstly face this problem,
> I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
> If we can know the amount memory of the socket buffer via /proc/meminfo, we
> may not need to spend a lot of time troubleshooting this problem. Not everyone
> knows that a lot of memory may be used here. But I believe many people
> should know /proc/meminfo to confirm memory users.
If I undestand correctly, the problem you are trying to solve is to
simplify troubleshooting of memory usage for people who may not be aware
that networking stack can be a large memory consumer.
For that a paragraph in 'man 5 proc' maybe a good start:
From ddbcf38576d1a2b0e36fe25a27350d566759b664 Mon Sep 17 00:00:00 2001
From: Mike Rapoport <[email protected]>
Date: Tue, 13 Oct 2020 11:07:35 +0300
Subject: [PATCH] proc.5: meminfo: add not anout network stack memory
consumption
Signed-off-by: Mike Rapoport <[email protected]>
---
man5/proc.5 | 8 ++++++++
1 file changed, 8 insertions(+)
diff --git a/man5/proc.5 b/man5/proc.5
index ed309380b..8414676f1 100644
--- a/man5/proc.5
+++ b/man5/proc.5
@@ -3478,6 +3478,14 @@ Except as noted below,
all of the fields have been present since at least Linux 2.6.0.
Some fields are displayed only if the kernel was configured
with various options; those dependencies are noted in the list.
+.IP
+Note that significant part of memory allocated by the network stack
+is not accounted in the file.
+The memory consumption of the network stack can be queried
+using
+.IR /proc/net/sockstat
+or
+.BR ss (8)
.RS
.TP
.IR MemTotal " %lu"
--
2.25.4
On 10/13/20 1:09 AM, Mike Rapoport wrote:
> On Mon, Oct 12, 2020 at 05:53:01PM +0800, Muchun Song wrote:
>> On Mon, Oct 12, 2020 at 5:24 PM Eric Dumazet <[email protected]> wrote:
>>>
>>> On 10/12/20 10:39 AM, Muchun Song wrote:
>>>> On Mon, Oct 12, 2020 at 3:42 PM Eric Dumazet <[email protected]> wrote:
>>>>>
>>>>> On Mon, Oct 12, 2020 at 6:22 AM Muchun Song <[email protected]> wrote:
>>>>>>
>>>>>> On Mon, Oct 12, 2020 at 2:39 AM Cong Wang <[email protected]> wrote:
>>>>>>>
>>>>>>> On Sat, Oct 10, 2020 at 3:39 AM Muchun Song <[email protected]> wrote:
>>>>>>>>
>>>>>>>> The amount of memory allocated to sockets buffer can become significant.
>>>>>>>> However, we do not display the amount of memory consumed by sockets
>>>>>>>> buffer. In this case, knowing where the memory is consumed by the kernel
>>>>>>>
>>>>>>> We do it via `ss -m`. Is it not sufficient? And if not, why not adding it there
>>>>>>> rather than /proc/meminfo?
>>>>>>
>>>>>> If the system has little free memory, we can know where the memory is via
>>>>>> /proc/meminfo. If a lot of memory is consumed by socket buffer, we cannot
>>>>>> know it when the Sock is not shown in the /proc/meminfo. If the unaware user
>>>>>> can't think of the socket buffer, naturally they will not `ss -m`. The
>>>>>> end result
>>>>>> is that we still don’t know where the memory is consumed. And we add the
>>>>>> Sock to the /proc/meminfo just like the memcg does('sock' item in the cgroup
>>>>>> v2 memory.stat). So I think that adding to /proc/meminfo is sufficient.
>>>>>>
>>>>>>>
>>>>>>>> static inline void __skb_frag_unref(skb_frag_t *frag)
>>>>>>>> {
>>>>>>>> - put_page(skb_frag_page(frag));
>>>>>>>> + struct page *page = skb_frag_page(frag);
>>>>>>>> +
>>>>>>>> + if (put_page_testzero(page)) {
>>>>>>>> + dec_sock_node_page_state(page);
>>>>>>>> + __put_page(page);
>>>>>>>> + }
>>>>>>>> }
>>>>>>>
>>>>>>> You mix socket page frag with skb frag at least, not sure this is exactly
>>>>>>> what you want, because clearly skb page frags are frequently used
>>>>>>> by network drivers rather than sockets.
>>>>>>>
>>>>>>> Also, which one matches this dec_sock_node_page_state()? Clearly
>>>>>>> not skb_fill_page_desc() or __skb_frag_ref().
>>>>>>
>>>>>> Yeah, we call inc_sock_node_page_state() in the skb_page_frag_refill().
>>>>>> So if someone gets the page returned by skb_page_frag_refill(), it must
>>>>>> put the page via __skb_frag_unref()/skb_frag_unref(). We use PG_private
>>>>>> to indicate that we need to dec the node page state when the refcount of
>>>>>> page reaches zero.
>>>>>>
>>>>>
>>>>> Pages can be transferred from pipe to socket, socket to pipe (splice()
>>>>> and zerocopy friends...)
>>>>>
>>>>> If you want to track TCP memory allocations, you always can look at
>>>>> /proc/net/sockstat,
>>>>> without adding yet another expensive memory accounting.
>>>>
>>>> The 'mem' item in the /proc/net/sockstat does not represent real
>>>> memory usage. This is just the total amount of charged memory.
>>>>
>>>> For example, if a task sends a 10-byte message, it only charges one
>>>> page to memcg. But the system may allocate 8 pages. Therefore, it
>>>> does not truly reflect the memory allocated by the above memory
>>>> allocation path. We can see the difference via the following message.
>>>>
>>>> cat /proc/net/sockstat
>>>> sockets: used 698
>>>> TCP: inuse 70 orphan 0 tw 617 alloc 134 mem 13
>>>> UDP: inuse 90 mem 4
>>>> UDPLITE: inuse 0
>>>> RAW: inuse 1
>>>> FRAG: inuse 0 memory 0
>>>>
>>>> cat /proc/meminfo | grep Sock
>>>> Sock: 13664 kB
>>>>
>>>> The /proc/net/sockstat only shows us that there are 17*4 kB TCP
>>>> memory allocations. But apply this patch, we can see that we truly
>>>> allocate 13664 kB(May be greater than this value because of per-cpu
>>>> stat cache). Of course the load of the example here is not high. In
>>>> some high load cases, I believe the difference here will be even
>>>> greater.
>>>>
>>>
>>> This is great, but you have not addressed my feedback.
>>>
>>> TCP memory allocations are bounded by /proc/sys/net/ipv4/tcp_mem
>>>
>>> Fact that the memory is forward allocated or not is a detail.
>>>
>>> If you think we must pre-allocate memory, instead of forward allocations,
>>> your patch does not address this. Adding one line per consumer in /proc/meminfo looks
>>> wrong to me.
>>
>> I think that the consumer which consumes a lot of memory should be added
>> to the /proc/meminfo. This can help us know the user of large memory.
>>
>>>
>>> If you do not want 9.37 % of physical memory being possibly used by TCP,
>>> just change /proc/sys/net/ipv4/tcp_mem accordingly ?
>>
>> We are not complaining about TCP using too much memory, but how do
>> we know that TCP uses a lot of memory. When I firstly face this problem,
>> I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
>> If we can know the amount memory of the socket buffer via /proc/meminfo, we
>> may not need to spend a lot of time troubleshooting this problem. Not everyone
>> knows that a lot of memory may be used here. But I believe many people
>> should know /proc/meminfo to confirm memory users.
>
> If I undestand correctly, the problem you are trying to solve is to
> simplify troubleshooting of memory usage for people who may not be aware
> that networking stack can be a large memory consumer.
>
> For that a paragraph in 'man 5 proc' maybe a good start:
>
>>From ddbcf38576d1a2b0e36fe25a27350d566759b664 Mon Sep 17 00:00:00 2001
> From: Mike Rapoport <[email protected]>
> Date: Tue, 13 Oct 2020 11:07:35 +0300
> Subject: [PATCH] proc.5: meminfo: add not anout network stack memory
> consumption
>
> Signed-off-by: Mike Rapoport <[email protected]>
> ---
> man5/proc.5 | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/man5/proc.5 b/man5/proc.5
> index ed309380b..8414676f1 100644
> --- a/man5/proc.5
> +++ b/man5/proc.5
> @@ -3478,6 +3478,14 @@ Except as noted below,
> all of the fields have been present since at least Linux 2.6.0.
> Some fields are displayed only if the kernel was configured
> with various options; those dependencies are noted in the list.
> +.IP
> +Note that significant part of memory allocated by the network stack
> +is not accounted in the file.
> +The memory consumption of the network stack can be queried
> +using
> +.IR /proc/net/sockstat
> +or
> +.BR ss (8)
> .RS
> .TP
> .IR MemTotal " %lu"
Hi Mike,
Could you tell us what units those values are in?
or is that already explained somewhere else?
thanks.
--
~Randy
On 10/12/20 11:53 AM, Muchun Song wrote:
> We are not complaining about TCP using too much memory, but how do
> we know that TCP uses a lot of memory. When I firstly face this problem,
> I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
> If we can know the amount memory of the socket buffer via /proc/meminfo, we
> may not need to spend a lot of time troubleshooting this problem. Not everyone
> knows that a lot of memory may be used here. But I believe many people
> should know /proc/meminfo to confirm memory users.
Adding yet another operations in networking fast path is a high cost to pay
just to add one extra line in /proc/meminfo, while /proc/net/sockstat
is already a good proxy, with per protocol details, instead of a single bucket.
I reiterate that zero copy would make this counter out of sync,
unless special support is added (adding yet another operations ?)
Also your patch does not address gazillions of page allocations from drivers
in RX path.
Here at Google the majority of networking mm usage when hosts are under stress
is in RX path, when out of order queues start to grow in TCP sockets.
Allocations in TX path were greatly reduced and optimally sized with the introduction
of /proc/sys/net/ipv4/tcp_notsent_lowat.
We have gazillions of put_page()/__free_page()/__free_pages()/alloc_page()/... all
over the places, adding yet another tracking of "this page is used by networking stacks"
is going to be quite a big project.
I thought memcg was a better goal in the long run, lets focus on it.
On Tue, Oct 13, 2020 at 07:43:59AM -0700, Randy Dunlap wrote:
> On 10/13/20 1:09 AM, Mike Rapoport wrote:
> > On Mon, Oct 12, 2020 at 05:53:01PM +0800, Muchun Song wrote:
> >> On Mon, Oct 12, 2020 at 5:24 PM Eric Dumazet <[email protected]> wrote:
> >>>
> >>> On 10/12/20 10:39 AM, Muchun Song wrote:
> >>>> On Mon, Oct 12, 2020 at 3:42 PM Eric Dumazet <[email protected]> wrote:
> >>
> >> We are not complaining about TCP using too much memory, but how do
> >> we know that TCP uses a lot of memory. When I firstly face this problem,
> >> I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
> >> If we can know the amount memory of the socket buffer via /proc/meminfo, we
> >> may not need to spend a lot of time troubleshooting this problem. Not everyone
> >> knows that a lot of memory may be used here. But I believe many people
> >> should know /proc/meminfo to confirm memory users.
> >
> > If I undestand correctly, the problem you are trying to solve is to
> > simplify troubleshooting of memory usage for people who may not be aware
> > that networking stack can be a large memory consumer.
> >
> > For that a paragraph in 'man 5 proc' maybe a good start:
> >
> >>From ddbcf38576d1a2b0e36fe25a27350d566759b664 Mon Sep 17 00:00:00 2001
> > From: Mike Rapoport <[email protected]>
> > Date: Tue, 13 Oct 2020 11:07:35 +0300
> > Subject: [PATCH] proc.5: meminfo: add not anout network stack memory
> > consumption
> >
> > Signed-off-by: Mike Rapoport <[email protected]>
> > ---
> > man5/proc.5 | 8 ++++++++
> > 1 file changed, 8 insertions(+)
> >
> > diff --git a/man5/proc.5 b/man5/proc.5
> > index ed309380b..8414676f1 100644
> > --- a/man5/proc.5
> > +++ b/man5/proc.5
> > @@ -3478,6 +3478,14 @@ Except as noted below,
> > all of the fields have been present since at least Linux 2.6.0.
> > Some fields are displayed only if the kernel was configured
> > with various options; those dependencies are noted in the list.
> > +.IP
> > +Note that significant part of memory allocated by the network stack
> > +is not accounted in the file.
> > +The memory consumption of the network stack can be queried
> > +using
> > +.IR /proc/net/sockstat
> > +or
> > +.BR ss (8)
> > .RS
> > .TP
> > .IR MemTotal " %lu"
>
> Hi Mike,
>
> Could you tell us what units those values are in?
> or is that already explained somewhere else?
It is described a few lines above and anyway, "MemTotal" is a part of
the diff context ;-)
> thanks.
> --
> ~Randy
>
>
--
Sincerely yours,
Mike.
On 10/13/20 8:12 AM, Mike Rapoport wrote:
> On Tue, Oct 13, 2020 at 07:43:59AM -0700, Randy Dunlap wrote:
>> On 10/13/20 1:09 AM, Mike Rapoport wrote:
>>> On Mon, Oct 12, 2020 at 05:53:01PM +0800, Muchun Song wrote:
>>>> On Mon, Oct 12, 2020 at 5:24 PM Eric Dumazet <[email protected]> wrote:
>>>>>
>>>>> On 10/12/20 10:39 AM, Muchun Song wrote:
>>>>>> On Mon, Oct 12, 2020 at 3:42 PM Eric Dumazet <[email protected]> wrote:
>>>>
>>>> We are not complaining about TCP using too much memory, but how do
>>>> we know that TCP uses a lot of memory. When I firstly face this problem,
>>>> I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
>>>> If we can know the amount memory of the socket buffer via /proc/meminfo, we
>>>> may not need to spend a lot of time troubleshooting this problem. Not everyone
>>>> knows that a lot of memory may be used here. But I believe many people
>>>> should know /proc/meminfo to confirm memory users.
>>>
>>> If I undestand correctly, the problem you are trying to solve is to
>>> simplify troubleshooting of memory usage for people who may not be aware
>>> that networking stack can be a large memory consumer.
>>>
>>> For that a paragraph in 'man 5 proc' maybe a good start:
>>>
>>> >From ddbcf38576d1a2b0e36fe25a27350d566759b664 Mon Sep 17 00:00:00 2001
>>> From: Mike Rapoport <[email protected]>
>>> Date: Tue, 13 Oct 2020 11:07:35 +0300
>>> Subject: [PATCH] proc.5: meminfo: add not anout network stack memory
>>> consumption
>>>
>>> Signed-off-by: Mike Rapoport <[email protected]>
>>> ---
>>> man5/proc.5 | 8 ++++++++
>>> 1 file changed, 8 insertions(+)
>>>
>>> diff --git a/man5/proc.5 b/man5/proc.5
>>> index ed309380b..8414676f1 100644
>>> --- a/man5/proc.5
>>> +++ b/man5/proc.5
>>> @@ -3478,6 +3478,14 @@ Except as noted below,
>>> all of the fields have been present since at least Linux 2.6.0.
>>> Some fields are displayed only if the kernel was configured
>>> with various options; those dependencies are noted in the list.
>>> +.IP
>>> +Note that significant part of memory allocated by the network stack
>>> +is not accounted in the file.
>>> +The memory consumption of the network stack can be queried
>>> +using
>>> +.IR /proc/net/sockstat
>>> +or
>>> +.BR ss (8)
>>> .RS
>>> .TP
>>> .IR MemTotal " %lu"
>>
>> Hi Mike,
>>
>> Could you tell us what units those values are in?
>> or is that already explained somewhere else?
>
> It is described a few lines above and anyway, "MemTotal" is a part of
> the diff context ;-)
with no units AFAICT.
But I was unclear. I wasn't referring to /proc/meminfo, but instead
to /proc/net/sockstat and its units:
sockets: used 1224
TCP: inuse 11 orphan 1 tw 1 alloc 26 mem 3
UDP: inuse 4 mem 2
UDPLITE: inuse 0
RAW: inuse 0
FRAG: inuse 0 memory 0
E.g., for TCP and UDP, are those socket counts or some unit of memory?
If units of memory, what unit size?
thanks.
--
~Randy
On Tue, Oct 13, 2020 at 4:09 PM Mike Rapoport <[email protected]> wrote:
>
> On Mon, Oct 12, 2020 at 05:53:01PM +0800, Muchun Song wrote:
> > On Mon, Oct 12, 2020 at 5:24 PM Eric Dumazet <[email protected]> wrote:
> > >
> > > On 10/12/20 10:39 AM, Muchun Song wrote:
> > > > On Mon, Oct 12, 2020 at 3:42 PM Eric Dumazet <[email protected]> wrote:
> > > >>
> > > >> On Mon, Oct 12, 2020 at 6:22 AM Muchun Song <[email protected]> wrote:
> > > >>>
> > > >>> On Mon, Oct 12, 2020 at 2:39 AM Cong Wang <[email protected]> wrote:
> > > >>>>
> > > >>>> On Sat, Oct 10, 2020 at 3:39 AM Muchun Song <[email protected]> wrote:
> > > >>>>>
> > > >>>>> The amount of memory allocated to sockets buffer can become significant.
> > > >>>>> However, we do not display the amount of memory consumed by sockets
> > > >>>>> buffer. In this case, knowing where the memory is consumed by the kernel
> > > >>>>
> > > >>>> We do it via `ss -m`. Is it not sufficient? And if not, why not adding it there
> > > >>>> rather than /proc/meminfo?
> > > >>>
> > > >>> If the system has little free memory, we can know where the memory is via
> > > >>> /proc/meminfo. If a lot of memory is consumed by socket buffer, we cannot
> > > >>> know it when the Sock is not shown in the /proc/meminfo. If the unaware user
> > > >>> can't think of the socket buffer, naturally they will not `ss -m`. The
> > > >>> end result
> > > >>> is that we still don’t know where the memory is consumed. And we add the
> > > >>> Sock to the /proc/meminfo just like the memcg does('sock' item in the cgroup
> > > >>> v2 memory.stat). So I think that adding to /proc/meminfo is sufficient.
> > > >>>
> > > >>>>
> > > >>>>> static inline void __skb_frag_unref(skb_frag_t *frag)
> > > >>>>> {
> > > >>>>> - put_page(skb_frag_page(frag));
> > > >>>>> + struct page *page = skb_frag_page(frag);
> > > >>>>> +
> > > >>>>> + if (put_page_testzero(page)) {
> > > >>>>> + dec_sock_node_page_state(page);
> > > >>>>> + __put_page(page);
> > > >>>>> + }
> > > >>>>> }
> > > >>>>
> > > >>>> You mix socket page frag with skb frag at least, not sure this is exactly
> > > >>>> what you want, because clearly skb page frags are frequently used
> > > >>>> by network drivers rather than sockets.
> > > >>>>
> > > >>>> Also, which one matches this dec_sock_node_page_state()? Clearly
> > > >>>> not skb_fill_page_desc() or __skb_frag_ref().
> > > >>>
> > > >>> Yeah, we call inc_sock_node_page_state() in the skb_page_frag_refill().
> > > >>> So if someone gets the page returned by skb_page_frag_refill(), it must
> > > >>> put the page via __skb_frag_unref()/skb_frag_unref(). We use PG_private
> > > >>> to indicate that we need to dec the node page state when the refcount of
> > > >>> page reaches zero.
> > > >>>
> > > >>
> > > >> Pages can be transferred from pipe to socket, socket to pipe (splice()
> > > >> and zerocopy friends...)
> > > >>
> > > >> If you want to track TCP memory allocations, you always can look at
> > > >> /proc/net/sockstat,
> > > >> without adding yet another expensive memory accounting.
> > > >
> > > > The 'mem' item in the /proc/net/sockstat does not represent real
> > > > memory usage. This is just the total amount of charged memory.
> > > >
> > > > For example, if a task sends a 10-byte message, it only charges one
> > > > page to memcg. But the system may allocate 8 pages. Therefore, it
> > > > does not truly reflect the memory allocated by the above memory
> > > > allocation path. We can see the difference via the following message.
> > > >
> > > > cat /proc/net/sockstat
> > > > sockets: used 698
> > > > TCP: inuse 70 orphan 0 tw 617 alloc 134 mem 13
> > > > UDP: inuse 90 mem 4
> > > > UDPLITE: inuse 0
> > > > RAW: inuse 1
> > > > FRAG: inuse 0 memory 0
> > > >
> > > > cat /proc/meminfo | grep Sock
> > > > Sock: 13664 kB
> > > >
> > > > The /proc/net/sockstat only shows us that there are 17*4 kB TCP
> > > > memory allocations. But apply this patch, we can see that we truly
> > > > allocate 13664 kB(May be greater than this value because of per-cpu
> > > > stat cache). Of course the load of the example here is not high. In
> > > > some high load cases, I believe the difference here will be even
> > > > greater.
> > > >
> > >
> > > This is great, but you have not addressed my feedback.
> > >
> > > TCP memory allocations are bounded by /proc/sys/net/ipv4/tcp_mem
> > >
> > > Fact that the memory is forward allocated or not is a detail.
> > >
> > > If you think we must pre-allocate memory, instead of forward allocations,
> > > your patch does not address this. Adding one line per consumer in /proc/meminfo looks
> > > wrong to me.
> >
> > I think that the consumer which consumes a lot of memory should be added
> > to the /proc/meminfo. This can help us know the user of large memory.
> >
> > >
> > > If you do not want 9.37 % of physical memory being possibly used by TCP,
> > > just change /proc/sys/net/ipv4/tcp_mem accordingly ?
> >
> > We are not complaining about TCP using too much memory, but how do
> > we know that TCP uses a lot of memory. When I firstly face this problem,
> > I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
> > If we can know the amount memory of the socket buffer via /proc/meminfo, we
> > may not need to spend a lot of time troubleshooting this problem. Not everyone
> > knows that a lot of memory may be used here. But I believe many people
> > should know /proc/meminfo to confirm memory users.
>
> If I undestand correctly, the problem you are trying to solve is to
> simplify troubleshooting of memory usage for people who may not be aware
> that networking stack can be a large memory consumer.
Yeah, you are right. Although the information provided by /proc/net/sockstat
is not accurate, it can also provide some valuable information. I think that it
might be better if we can add a total amount socket buffer to /proc/meminfo.
The amount socket buffer statistics can be from /proc/net/sockstat directly.
Thanks.
>
> For that a paragraph in 'man 5 proc' maybe a good start:
>
> From ddbcf38576d1a2b0e36fe25a27350d566759b664 Mon Sep 17 00:00:00 2001
> From: Mike Rapoport <[email protected]>
> Date: Tue, 13 Oct 2020 11:07:35 +0300
> Subject: [PATCH] proc.5: meminfo: add not anout network stack memory
> consumption
>
> Signed-off-by: Mike Rapoport <[email protected]>
> ---
> man5/proc.5 | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/man5/proc.5 b/man5/proc.5
> index ed309380b..8414676f1 100644
> --- a/man5/proc.5
> +++ b/man5/proc.5
> @@ -3478,6 +3478,14 @@ Except as noted below,
> all of the fields have been present since at least Linux 2.6.0.
> Some fields are displayed only if the kernel was configured
> with various options; those dependencies are noted in the list.
> +.IP
> +Note that significant part of memory allocated by the network stack
> +is not accounted in the file.
> +The memory consumption of the network stack can be queried
> +using
> +.IR /proc/net/sockstat
> +or
> +.BR ss (8)
> .RS
> .TP
> .IR MemTotal " %lu"
> --
> 2.25.4
>
>
--
Yours,
Muchun
On Tue, Oct 13, 2020 at 08:21:13AM -0700, Randy Dunlap wrote:
> On 10/13/20 8:12 AM, Mike Rapoport wrote:
> > On Tue, Oct 13, 2020 at 07:43:59AM -0700, Randy Dunlap wrote:
> >> On 10/13/20 1:09 AM, Mike Rapoport wrote:
> >>> On Mon, Oct 12, 2020 at 05:53:01PM +0800, Muchun Song wrote:
> >>>> On Mon, Oct 12, 2020 at 5:24 PM Eric Dumazet <[email protected]> wrote:
> >>>>>
> >>>>> On 10/12/20 10:39 AM, Muchun Song wrote:
> >>>>>> On Mon, Oct 12, 2020 at 3:42 PM Eric Dumazet <[email protected]> wrote:
> >>>>
> >>>> We are not complaining about TCP using too much memory, but how do
> >>>> we know that TCP uses a lot of memory. When I firstly face this problem,
> >>>> I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
> >>>> If we can know the amount memory of the socket buffer via /proc/meminfo, we
> >>>> may not need to spend a lot of time troubleshooting this problem. Not everyone
> >>>> knows that a lot of memory may be used here. But I believe many people
> >>>> should know /proc/meminfo to confirm memory users.
> >>>
> >>> If I undestand correctly, the problem you are trying to solve is to
> >>> simplify troubleshooting of memory usage for people who may not be aware
> >>> that networking stack can be a large memory consumer.
> >>>
> >>> For that a paragraph in 'man 5 proc' maybe a good start:
> >>>
> >>> >From ddbcf38576d1a2b0e36fe25a27350d566759b664 Mon Sep 17 00:00:00 2001
> >>> From: Mike Rapoport <[email protected]>
> >>> Date: Tue, 13 Oct 2020 11:07:35 +0300
> >>> Subject: [PATCH] proc.5: meminfo: add not anout network stack memory
> >>> consumption
> >>>
> >>> Signed-off-by: Mike Rapoport <[email protected]>
> >>> ---
> >>> man5/proc.5 | 8 ++++++++
> >>> 1 file changed, 8 insertions(+)
> >>>
> >>> diff --git a/man5/proc.5 b/man5/proc.5
> >>> index ed309380b..8414676f1 100644
> >>> --- a/man5/proc.5
> >>> +++ b/man5/proc.5
> >>> @@ -3478,6 +3478,14 @@ Except as noted below,
> >>> all of the fields have been present since at least Linux 2.6.0.
> >>> Some fields are displayed only if the kernel was configured
> >>> with various options; those dependencies are noted in the list.
> >>> +.IP
> >>> +Note that significant part of memory allocated by the network stack
> >>> +is not accounted in the file.
> >>> +The memory consumption of the network stack can be queried
> >>> +using
> >>> +.IR /proc/net/sockstat
> >>> +or
> >>> +.BR ss (8)
> >>> .RS
> >>> .TP
> >>> .IR MemTotal " %lu"
> >>
> >> Hi Mike,
> >>
> >> Could you tell us what units those values are in?
> >> or is that already explained somewhere else?
> >
> > It is described a few lines above and anyway, "MemTotal" is a part of
> > the diff context ;-)
>
> with no units AFAICT.
>
> But I was unclear. I wasn't referring to /proc/meminfo, but instead
> to /proc/net/sockstat and its units:
>
> sockets: used 1224
> TCP: inuse 11 orphan 1 tw 1 alloc 26 mem 3
> UDP: inuse 4 mem 2
> UDPLITE: inuse 0
> RAW: inuse 0
> FRAG: inuse 0 memory 0
>
> E.g., for TCP and UDP, are those socket counts or some unit of memory?
> If units of memory, what unit size?
Ah, these are in 4k pages, AFAIU.
And, as it seems /proc/net/sockstat lacks a description in proc.5 at
all...
> thanks.
> --
> ~Randy
>
>
--
Sincerely yours,
Mike.
On 10/13/20 10:09 AM, Mike Rapoport wrote:
>> We are not complaining about TCP using too much memory, but how do
>> we know that TCP uses a lot of memory. When I firstly face this problem,
>> I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
>> If we can know the amount memory of the socket buffer via /proc/meminfo, we
>> may not need to spend a lot of time troubleshooting this problem. Not everyone
>> knows that a lot of memory may be used here. But I believe many people
>> should know /proc/meminfo to confirm memory users.
> If I undestand correctly, the problem you are trying to solve is to
> simplify troubleshooting of memory usage for people who may not be aware
> that networking stack can be a large memory consumer.
>
> For that a paragraph in 'man 5 proc' maybe a good start:
Yeah. Another major consumer that I've seen at some point was xfs buffers. And
there might be others, and adding everything to /proc/meminfo is not feasible. I
have once proposed adding a counter called "Unaccounted:" which would at least
tell the user easily if a significant portion is occupied by memory not
explained by the other meminfo counters, and look for trends (increase =
potential memory leak?). For specific prominent consumers not covered by meminfo
but that have some kind of internal counters, we could document where to look,
such as /proc/net/sockstat or maybe create some /proc/ or /sys directory with
file per consumer so that it's still easy to check, but without the overhead of
global counters and bloated /proc/meminfo?
> From ddbcf38576d1a2b0e36fe25a27350d566759b664 Mon Sep 17 00:00:00 2001
> From: Mike Rapoport<[email protected]>
> Date: Tue, 13 Oct 2020 11:07:35 +0300
> Subject: [PATCH] proc.5: meminfo: add not anout network stack memory
> consumption
>
> Signed-off-by: Mike Rapoport<[email protected]>
> ---
> man5/proc.5 | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/man5/proc.5 b/man5/proc.5
> index ed309380b..8414676f1 100644
> --- a/man5/proc.5
> +++ b/man5/proc.5
> @@ -3478,6 +3478,14 @@ Except as noted below,
> all of the fields have been present since at least Linux 2.6.0.
> Some fields are displayed only if the kernel was configured
> with various options; those dependencies are noted in the list.
> +.IP
> +Note that significant part of memory allocated by the network stack
> +is not accounted in the file.
> +The memory consumption of the network stack can be queried
> +using
> +.IR /proc/net/sockstat
> +or
> +.BR ss (8)
> .RS
> .TP
> .IR MemTotal " %lu"
> -- 2.25.4
On Fri, Oct 16, 2020 at 05:38:26PM +0200, Vlastimil Babka wrote:
> On 10/13/20 10:09 AM, Mike Rapoport wrote:
> > > We are not complaining about TCP using too much memory, but how do
> > > we know that TCP uses a lot of memory. When I firstly face this problem,
> > > I do not know who uses the 25GB memory and it is not shown in the /proc/meminfo.
> > > If we can know the amount memory of the socket buffer via /proc/meminfo, we
> > > may not need to spend a lot of time troubleshooting this problem. Not everyone
> > > knows that a lot of memory may be used here. But I believe many people
> > > should know /proc/meminfo to confirm memory users.
> > If I undestand correctly, the problem you are trying to solve is to
> > simplify troubleshooting of memory usage for people who may not be aware
> > that networking stack can be a large memory consumer.
> >
> > For that a paragraph in 'man 5 proc' maybe a good start:
>
> Yeah. Another major consumer that I've seen at some point was xfs buffers.
As well, there are other various type of memory consumers in embedded world,
depending on the features what they supprted, too. They often tempted to add
the memory consumption into /proc/meminfo or /proc/vmstat, too to get
memory visibility.
> And there might be others, and adding everything to /proc/meminfo is not
> feasible. I have once proposed adding a counter called "Unaccounted:" which
> would at least tell the user easily if a significant portion is occupied by
> memory not explained by the other meminfo counters, and look for trends
> (increase = potential memory leak?). For specific prominent consumers not
> covered by meminfo but that have some kind of internal counters, we could
> document where to look, such as /proc/net/sockstat or maybe create some
> /proc/ or /sys directory with file per consumer so that it's still easy to
> check, but without the overhead of global counters and bloated
> /proc/meminfo?
What have in my mind is to support simple general sysfs infra from MM for
driver/subysstems rather than creating each own memory stat. The API
could support flexible accounting like just global memory consumption and/or
consmption by key(e.g,. pid or each own special) for the detail.
So, they are all shown under /sys/kernel/mm/misc/ with detail as well as
/proc/meminfo with simple line for global.
Furthermore, I'd like to plug the infra into OOM message so once OOM occurs,
we could print each own hidden memory usage from driver side if the driver
believe they could be memory hogger. It would make easier to detect
such kinds of leak or hogging as well as better maintainace.
CCed: Paolo Bonzini
On Fri, Oct 16, 2020 at 1:53 PM Minchan Kim <[email protected]> wrote:
[snip]
> > And there might be others, and adding everything to /proc/meminfo is not
> > feasible. I have once proposed adding a counter called "Unaccounted:" which
> > would at least tell the user easily if a significant portion is occupied by
> > memory not explained by the other meminfo counters, and look for trends
> > (increase = potential memory leak?). For specific prominent consumers not
> > covered by meminfo but that have some kind of internal counters, we could
> > document where to look, such as /proc/net/sockstat or maybe create some
> > /proc/ or /sys directory with file per consumer so that it's still easy to
> > check, but without the overhead of global counters and bloated
> > /proc/meminfo?
>
> What have in my mind is to support simple general sysfs infra from MM for
> driver/subysstems rather than creating each own memory stat. The API
> could support flexible accounting like just global memory consumption and/or
> consmption by key(e.g,. pid or each own special) for the detail.
>
> So, they are all shown under /sys/kernel/mm/misc/ with detail as well as
> /proc/meminfo with simple line for global.
This reminds me of statsfs [1]. I am wondering if this can be another
useful use-case for statsfs.
[1] https://lkml.org/lkml/2020/5/26/332