Questions about querying map object information

Questions about querying map object information

[Hao Xiang .]

Hi everyone,

I am super new to bpf and the open source community in general. Please
bear with me asking some basic questions.
We are working on a bpf monitoring tool to track the CPU and memory
usage for all bpf programs loaded in the system. We were able to get
CPU usage per bpf program with the BPF_OBJ_GET_INFO_BY_ID syscall on a
bpf prog object. We are trying to do the same on a map object to query
for per map memory usage. The information returned from bpf_map_info
only contains things like max_entries, key_size, value_size, which can
be used to calculate estimated memory allocation size. But we are also
interested in knowing how much memory is actually being used by our
program. For instance, one of our bpf program uses a map with type
hashtable. The hashtable is created with a chunk of pre-allocated
memory based on the max_entres, key_size and value size. The
pre-allocated size is useful information to know but so is the current
number of entries in the hashtable. We used to run into a performance
issue where our bpf map's max_entries is set to be too small and we
end up totally exhausting the pre-allocated memory. So knowing things
like current entry count VS max entry count of a hashtable is useful
information for us.
With that being said, we have a few questions and hopefully we can get
some help from the community.
1) We couldn't find anything in bpf_map_info to give us the current
entry count of a hashtable. I read that bpf_map_info returns
information about a map object in general. So it makes total sense to
not have information of a particular map type. But is there an
existing place we can get the per map type information (eg, the
current entry count of a hashtable, the number of elements pushed to a
stack, etc)?
2) If there isn't an existing place to return map type specific
information, would it make sense to extend the structure bpf_map_info
with a union at the end and have that union to contain per map type
specific information?

Thanks, Hao

Re: Questions about querying map object information

[Hao Xiang .]

Ping...

Can someone please help to shed some light on this?

Thanks, Hao

On Sun, Jul 3, 2022 at 3:33 PM Hao Xiang . <hao.xiang@bytedance.com> wrote:
>
> Hi everyone,
>
> I am super new to bpf and the open source community in general. Please
> bear with me asking some basic questions.
> We are working on a bpf monitoring tool to track the CPU and memory
> usage for all bpf programs loaded in the system. We were able to get
> CPU usage per bpf program with the BPF_OBJ_GET_INFO_BY_ID syscall on a
> bpf prog object. We are trying to do the same on a map object to query
> for per map memory usage. The information returned from bpf_map_info
> only contains things like max_entries, key_size, value_size, which can
> be used to calculate estimated memory allocation size. But we are also
> interested in knowing how much memory is actually being used by our
> program. For instance, one of our bpf program uses a map with type
> hashtable. The hashtable is created with a chunk of pre-allocated
> memory based on the max_entres, key_size and value size. The
> pre-allocated size is useful information to know but so is the current
> number of entries in the hashtable. We used to run into a performance
> issue where our bpf map's max_entries is set to be too small and we
> end up totally exhausting the pre-allocated memory. So knowing things
> like current entry count VS max entry count of a hashtable is useful
> information for us.
> With that being said, we have a few questions and hopefully we can get
> some help from the community.
> 1) We couldn't find anything in bpf_map_info to give us the current
> entry count of a hashtable. I read that bpf_map_info returns
> information about a map object in general. So it makes total sense to
> not have information of a particular map type. But is there an
> existing place we can get the per map type information (eg, the
> current entry count of a hashtable, the number of elements pushed to a
> stack, etc)?
> 2) If there isn't an existing place to return map type specific
> information, would it make sense to extend the structure bpf_map_info
> with a union at the end and have that union to contain per map type
> specific information?
>
> Thanks, Hao

Re: Questions about querying map object information

[Hao Luo]

Hi Hao Xiang,

On Mon, Jul 11, 2022 at 3:50 PM Hao Xiang . <hao.xiang@bytedance.com> wrote:
>
> Ping...
>
> Can someone please help to shed some light on this?
>
> Thanks, Hao
>
> On Sun, Jul 3, 2022 at 3:33 PM Hao Xiang . <hao.xiang@bytedance.com> wrote:
> >
> > Hi everyone,
> >
> > I am super new to bpf and the open source community in general. Please
> > bear with me asking some basic questions.
> > We are working on a bpf monitoring tool to track the CPU and memory
> > usage for all bpf programs loaded in the system. We were able to get
> > CPU usage per bpf program with the BPF_OBJ_GET_INFO_BY_ID syscall on a
> > bpf prog object. We are trying to do the same on a map object to query
> > for per map memory usage. The information returned from bpf_map_info
> > only contains things like max_entries, key_size, value_size, which can
> > be used to calculate estimated memory allocation size. But we are also
> > interested in knowing how much memory is actually being used by our
> > program. For instance, one of our bpf program uses a map with type
> > hashtable. The hashtable is created with a chunk of pre-allocated
> > memory based on the max_entres, key_size and value size. The
> > pre-allocated size is useful information to know but so is the current
> > number of entries in the hashtable. We used to run into a performance
> > issue where our bpf map's max_entries is set to be too small and we
> > end up totally exhausting the pre-allocated memory. So knowing things
> > like current entry count VS max entry count of a hashtable is useful
> > information for us.
> > With that being said, we have a few questions and hopefully we can get
> > some help from the community.
> > 1) We couldn't find anything in bpf_map_info to give us the current
> > entry count of a hashtable. I read that bpf_map_info returns
> > information about a map object in general. So it makes total sense to
> > not have information of a particular map type. But is there an
> > existing place we can get the per map type information (eg, the
> > current entry count of a hashtable, the number of elements pushed to a
> > stack, etc)?

cc more BPF experts for their comments.

I agree with you that knowing the current space usage of a map is
quite helpful. In my understanding, a naive and inefficient way to
estimate space usage is iterating the map and counting the map
elements. That's doable, but may not be the best method though.
Regarding auto-adjusting map size, I remember Andrii talked about
resizable hash maps, maybe he can tell you more [1].

[1] http://vger.kernel.org/bpfconf2022_material/lsfmmbpf2022-algs.pdf

> > 2) If there isn't an existing place to return map type specific
> > information, would it make sense to extend the structure bpf_map_info
> > with a union at the end and have that union to contain per map type
> > specific information?
> >
> > Thanks, Hao

Re: [External] Re: Questions about querying map object information

[Hao Xiang .]

Hi Hao,

I am not talking to myself :-) Thanks for addressing the question.
Please see inline.

On Mon, Jul 11, 2022 at 4:35 PM Hao Luo <haoluo@google.com> wrote:
>
> Hi Hao Xiang,
>
> On Mon, Jul 11, 2022 at 3:50 PM Hao Xiang . <hao.xiang@bytedance.com> wrote:
> >
> > Ping...
> >
> > Can someone please help to shed some light on this?
> >
> > Thanks, Hao
> >
> > On Sun, Jul 3, 2022 at 3:33 PM Hao Xiang . <hao.xiang@bytedance.com> wrote:
> > >
> > > Hi everyone,
> > >
> > > I am super new to bpf and the open source community in general. Please
> > > bear with me asking some basic questions.
> > > We are working on a bpf monitoring tool to track the CPU and memory
> > > usage for all bpf programs loaded in the system. We were able to get
> > > CPU usage per bpf program with the BPF_OBJ_GET_INFO_BY_ID syscall on a
> > > bpf prog object. We are trying to do the same on a map object to query
> > > for per map memory usage. The information returned from bpf_map_info
> > > only contains things like max_entries, key_size, value_size, which can
> > > be used to calculate estimated memory allocation size. But we are also
> > > interested in knowing how much memory is actually being used by our
> > > program. For instance, one of our bpf program uses a map with type
> > > hashtable. The hashtable is created with a chunk of pre-allocated
> > > memory based on the max_entres, key_size and value size. The
> > > pre-allocated size is useful information to know but so is the current
> > > number of entries in the hashtable. We used to run into a performance
> > > issue where our bpf map's max_entries is set to be too small and we
> > > end up totally exhausting the pre-allocated memory. So knowing things
> > > like current entry count VS max entry count of a hashtable is useful
> > > information for us.
> > > With that being said, we have a few questions and hopefully we can get
> > > some help from the community.
> > > 1) We couldn't find anything in bpf_map_info to give us the current
> > > entry count of a hashtable. I read that bpf_map_info returns
> > > information about a map object in general. So it makes total sense to
> > > not have information of a particular map type. But is there an
> > > existing place we can get the per map type information (eg, the
> > > current entry count of a hashtable, the number of elements pushed to a
> > > stack, etc)?
>
> cc more BPF experts for their comments.
>
> > I agree with you that knowing the current space usage of a map is
> > quite helpful. In my understanding, a naive and inefficient way to
> > estimate space usage is iterating the map and counting the map
> > elements. That's doable, but may not be the best method though.
> > Regarding auto-adjusting map size, I remember Andrii talked about
> > resizable hash maps, maybe he can tell you more [1].
>
> I agree it's doable to get the map count by iterating through the map. I just
> looked through the code and realized that it's quite heavy doing it that way.
> Basically we will have to make N bpf syscalls where N = used count. Our monitor
> program wants to retrieve the current usage in a higher frequency and the underlying
> cost looks pretty significant. Other than that, is it expected that a user process who
> can do BPF_OBJ_GET_INFO_BY_FD requires the same privilege as someone
> who can do BPF_MAP_GET_NEXT_KEY? I would expect that only the process
> talking to the bpf prog can update/delete/get keys of a map but a process who only
> wants to monitor the prog/map requires less privilege?
>
>
> [1] http://vger.kernel.org/bpfconf2022_material/lsfmmbpf2022-algs.pdf
>
> > > 2) If there isn't an existing place to return map type specific
> > > information, would it make sense to extend the structure bpf_map_info
> > > with a union at the end and have that union to contain per map type
> > > specific information?
> > >
> > > Thanks, Hao

Re: Questions about querying map object information

[Jiri Olsa]

On Mon, Jul 11, 2022 at 04:35:19PM -0700, Hao Luo wrote:
> Hi Hao Xiang,
> 
> On Mon, Jul 11, 2022 at 3:50 PM Hao Xiang . <hao.xiang@bytedance.com> wrote:
> >
> > Ping...
> >
> > Can someone please help to shed some light on this?
> >
> > Thanks, Hao
> >
> > On Sun, Jul 3, 2022 at 3:33 PM Hao Xiang . <hao.xiang@bytedance.com> wrote:
> > >
> > > Hi everyone,
> > >
> > > I am super new to bpf and the open source community in general. Please
> > > bear with me asking some basic questions.
> > > We are working on a bpf monitoring tool to track the CPU and memory
> > > usage for all bpf programs loaded in the system. We were able to get
> > > CPU usage per bpf program with the BPF_OBJ_GET_INFO_BY_ID syscall on a
> > > bpf prog object. We are trying to do the same on a map object to query
> > > for per map memory usage. The information returned from bpf_map_info
> > > only contains things like max_entries, key_size, value_size, which can
> > > be used to calculate estimated memory allocation size. But we are also
> > > interested in knowing how much memory is actually being used by our
> > > program. For instance, one of our bpf program uses a map with type
> > > hashtable. The hashtable is created with a chunk of pre-allocated
> > > memory based on the max_entres, key_size and value size. The
> > > pre-allocated size is useful information to know but so is the current
> > > number of entries in the hashtable. We used to run into a performance
> > > issue where our bpf map's max_entries is set to be too small and we
> > > end up totally exhausting the pre-allocated memory. So knowing things
> > > like current entry count VS max entry count of a hashtable is useful
> > > information for us.
> > > With that being said, we have a few questions and hopefully we can get
> > > some help from the community.
> > > 1) We couldn't find anything in bpf_map_info to give us the current
> > > entry count of a hashtable. I read that bpf_map_info returns
> > > information about a map object in general. So it makes total sense to
> > > not have information of a particular map type. But is there an
> > > existing place we can get the per map type information (eg, the
> > > current entry count of a hashtable, the number of elements pushed to a
> > > stack, etc)?
> 
> cc more BPF experts for their comments.
> 
> I agree with you that knowing the current space usage of a map is
> quite helpful. In my understanding, a naive and inefficient way to
> estimate space usage is iterating the map and counting the map
> elements. That's doable, but may not be the best method though.
> Regarding auto-adjusting map size, I remember Andrii talked about
> resizable hash maps, maybe he can tell you more [1].

we keep entries count for hash map, maybe we could just add
'current count' to bpf_map_info, it seems generic enough

jirka

> 
> [1] http://vger.kernel.org/bpfconf2022_material/lsfmmbpf2022-algs.pdf
> 
> > > 2) If there isn't an existing place to return map type specific
> > > information, would it make sense to extend the structure bpf_map_info
> > > with a union at the end and have that union to contain per map type
> > > specific information?
> > >
> > > Thanks, Hao

Re: Questions about querying map object information

[Alexei Starovoitov]

On Tue, Jul 12, 2022 at 1:53 AM Jiri Olsa <olsajiri@gmail.com> wrote:
>
> On Mon, Jul 11, 2022 at 04:35:19PM -0700, Hao Luo wrote:
> > Hi Hao Xiang,
> >
> > On Mon, Jul 11, 2022 at 3:50 PM Hao Xiang . <hao.xiang@bytedance.com> wrote:
> > >
> > > Ping...
> > >
> > > Can someone please help to shed some light on this?
> > >
> > > Thanks, Hao
> > >
> > > On Sun, Jul 3, 2022 at 3:33 PM Hao Xiang . <hao.xiang@bytedance.com> wrote:
> > > >
> > > > Hi everyone,
> > > >
> > > > I am super new to bpf and the open source community in general. Please
> > > > bear with me asking some basic questions.
> > > > We are working on a bpf monitoring tool to track the CPU and memory
> > > > usage for all bpf programs loaded in the system. We were able to get
> > > > CPU usage per bpf program with the BPF_OBJ_GET_INFO_BY_ID syscall on a
> > > > bpf prog object. We are trying to do the same on a map object to query
> > > > for per map memory usage. The information returned from bpf_map_info
> > > > only contains things like max_entries, key_size, value_size, which can
> > > > be used to calculate estimated memory allocation size. But we are also
> > > > interested in knowing how much memory is actually being used by our
> > > > program. For instance, one of our bpf program uses a map with type
> > > > hashtable. The hashtable is created with a chunk of pre-allocated
> > > > memory based on the max_entres, key_size and value size. The
> > > > pre-allocated size is useful information to know but so is the current
> > > > number of entries in the hashtable. We used to run into a performance
> > > > issue where our bpf map's max_entries is set to be too small and we
> > > > end up totally exhausting the pre-allocated memory. So knowing things
> > > > like current entry count VS max entry count of a hashtable is useful
> > > > information for us.
> > > > With that being said, we have a few questions and hopefully we can get
> > > > some help from the community.
> > > > 1) We couldn't find anything in bpf_map_info to give us the current
> > > > entry count of a hashtable. I read that bpf_map_info returns
> > > > information about a map object in general. So it makes total sense to
> > > > not have information of a particular map type. But is there an
> > > > existing place we can get the per map type information (eg, the
> > > > current entry count of a hashtable, the number of elements pushed to a
> > > > stack, etc)?
> >
> > cc more BPF experts for their comments.
> >
> > I agree with you that knowing the current space usage of a map is
> > quite helpful. In my understanding, a naive and inefficient way to
> > estimate space usage is iterating the map and counting the map
> > elements. That's doable, but may not be the best method though.
> > Regarding auto-adjusting map size, I remember Andrii talked about
> > resizable hash maps, maybe he can tell you more [1].
>
> we keep entries count for hash map, maybe we could just add
> 'current count' to bpf_map_info, it seems generic enough

We actually don't keep a count for the default case of prealloc.