[bpf-next,0/6] bpf: add BPF_MAP_DUMP command to dump more than one entry per call

This introduces a new command to retrieve multiple number of entries
from a bpf map.

This new command can be executed from the existing BPF syscall as
follows:

err =  bpf(BPF_MAP_DUMP, union bpf_attr *attr, u32 size)
using attr->dump.map_fd, attr->dump.prev_key, attr->dump.buf,
attr->dump.buf_len
returns zero or negative error, and populates buf and buf_len on
succees

This implementation is wrapping the existing bpf methods:
map_get_next_key and map_lookup_elem

Note that this implementation can be extended later to do dump and
delete by extending map_lookup_and_delete_elem (currently it only works
for bpf queue/stack maps) and either use a new flag in map_dump or a new
command map_dump_and_delete. 

Results show that even with a 1-elem_size buffer, it runs ~40 faster
than the current implementation, improvements of ~85% are reported when
the buffer size is increased, although, after the buffer size is around
5% of the total number of entries there's no huge difference in
increasing it.

Tested:
Tried different size buffers to handle case where the bulk is bigger, or
the elements to retrieve are less than the existing ones, all runs read
a map of 100K entries. Below are the results(in ns) from the different
runs:

buf_len_1:       69038725 entry-by-entry: 112384424 improvement
38.569134
buf_len_2:       40897447 entry-by-entry: 111030546 improvement
63.165590
buf_len_230:     13652714 entry-by-entry: 111694058 improvement
87.776687
buf_len_5000:    13576271 entry-by-entry: 111101169 improvement
87.780263
buf_len_73000:   14694343 entry-by-entry: 111740162 improvement
86.849542
buf_len_100000:  13745969 entry-by-entry: 114151991 improvement
87.958187
buf_len_234567:  14329834 entry-by-entry: 114427589 improvement
87.476941

The series of patches are split as follows:

- First patch move some map_lookup_elem logic into 2 fucntions to
deduplicate code: bpf_map_value_size and bpf_map_copy_value
- Second patch introduce map_dump function
- Third patch syncs tools linux headers
- Fourth patch adds libbpf support
- Last two patches adds tests

RFC Changelog:

- remove wrong usage of attr.flags
- move map_fd to remove hole after it

v3:
- add explanation of the API in the commit message
- fix masked errors and return them to user
- copy last_key from return buf into prev_key if it was provided
- run perf test with kpti and retpoline mitigations

v2:
- use proper bpf-next tag

Brian Vazquez (6):
  bpf: add bpf_map_value_size and bp_map_copy_value helper functions
  bpf: add BPF_MAP_DUMP command to dump more than one entry per call
  bpf: keep bpf.h in sync with tools/
  libbpf: support BPF_MAP_DUMP command
  selftests/bpf: test BPF_MAP_DUMP command on a bpf hashmap
  selftests/bpf: add test to measure performance of BPF_MAP_DUMP

 include/uapi/linux/bpf.h                |   9 +
 kernel/bpf/syscall.c                    | 251 ++++++++++++++++++------
 tools/include/uapi/linux/bpf.h          |   9 +
 tools/lib/bpf/bpf.c                     |  28 +++
 tools/lib/bpf/bpf.h                     |   4 +
 tools/lib/bpf/libbpf.map                |   2 +
 tools/testing/selftests/bpf/test_maps.c | 148 +++++++++++++-
 7 files changed, 388 insertions(+), 63 deletions(-)

On Wed, Jul 24, 2019 at 10:09 AM Brian Vazquez <brianvv@google.com> wrote:
>
> This introduces a new command to retrieve multiple number of entries
> from a bpf map.
>
> This new command can be executed from the existing BPF syscall as
> follows:
>
> err =  bpf(BPF_MAP_DUMP, union bpf_attr *attr, u32 size)
> using attr->dump.map_fd, attr->dump.prev_key, attr->dump.buf,
> attr->dump.buf_len
> returns zero or negative error, and populates buf and buf_len on
> succees
>
> This implementation is wrapping the existing bpf methods:
> map_get_next_key and map_lookup_elem
>
> Note that this implementation can be extended later to do dump and
> delete by extending map_lookup_and_delete_elem (currently it only works
> for bpf queue/stack maps) and either use a new flag in map_dump or a new
> command map_dump_and_delete.
>
> Results show that even with a 1-elem_size buffer, it runs ~40 faster

Why is the new command 40% faster with 1-elem_size buffer?

> than the current implementation, improvements of ~85% are reported when
> the buffer size is increased, although, after the buffer size is around
> 5% of the total number of entries there's no huge difference in
> increasing it.
>
> Tested:
> Tried different size buffers to handle case where the bulk is bigger, or
> the elements to retrieve are less than the existing ones, all runs read
> a map of 100K entries. Below are the results(in ns) from the different
> runs:
>
> buf_len_1:       69038725 entry-by-entry: 112384424 improvement
> 38.569134
> buf_len_2:       40897447 entry-by-entry: 111030546 improvement
> 63.165590
> buf_len_230:     13652714 entry-by-entry: 111694058 improvement
> 87.776687
> buf_len_5000:    13576271 entry-by-entry: 111101169 improvement
> 87.780263
> buf_len_73000:   14694343 entry-by-entry: 111740162 improvement
> 86.849542
> buf_len_100000:  13745969 entry-by-entry: 114151991 improvement
> 87.958187
> buf_len_234567:  14329834 entry-by-entry: 114427589 improvement
> 87.476941

It took me a while to figure out the meaning of 87.476941. It is probably
a good idea to say 87.5% instead.

Thanks,
Song

On Wed, Jul 24, 2019 at 12:20 PM Song Liu <liu.song.a23@gmail.com> wrote:
>
> On Wed, Jul 24, 2019 at 10:09 AM Brian Vazquez <brianvv@google.com> wrote:
> >
> > This introduces a new command to retrieve multiple number of entries
> > from a bpf map.
> >
> > This new command can be executed from the existing BPF syscall as
> > follows:
> >
> > err =  bpf(BPF_MAP_DUMP, union bpf_attr *attr, u32 size)
> > using attr->dump.map_fd, attr->dump.prev_key, attr->dump.buf,
> > attr->dump.buf_len
> > returns zero or negative error, and populates buf and buf_len on
> > succees
> >
> > This implementation is wrapping the existing bpf methods:
> > map_get_next_key and map_lookup_elem
> >
> > Note that this implementation can be extended later to do dump and
> > delete by extending map_lookup_and_delete_elem (currently it only works
> > for bpf queue/stack maps) and either use a new flag in map_dump or a new
> > command map_dump_and_delete.
> >
> > Results show that even with a 1-elem_size buffer, it runs ~40 faster
>
> Why is the new command 40% faster with 1-elem_size buffer?

The test is using a really simple map structure: uint64_t key,val.
Which makes the lookup_elem logic faster since it doesn't spend too
much time copying the value. My conclusion with the 40% was that this
new implementation only needs 1 syscall per elem compare to the 2
syscalls that we needed with previous implementation so in this
particular case the number of ops that we are doing is almost halved.
I did one experiment increasing the value_size (448*64B) and it was
only 14% faster with 1-elem_size buffer.

> > than the current implementation, improvements of ~85% are reported when
> > the buffer size is increased, although, after the buffer size is around
> > 5% of the total number of entries there's no huge difference in
> > increasing it.
> >
> > Tested:
> > Tried different size buffers to handle case where the bulk is bigger, or
> > the elements to retrieve are less than the existing ones, all runs read
> > a map of 100K entries. Below are the results(in ns) from the different
> > runs:
> >
> > buf_len_1:       69038725 entry-by-entry: 112384424 improvement
> > 38.569134
> > buf_len_2:       40897447 entry-by-entry: 111030546 improvement
> > 63.165590
> > buf_len_230:     13652714 entry-by-entry: 111694058 improvement
> > 87.776687
> > buf_len_5000:    13576271 entry-by-entry: 111101169 improvement
> > 87.780263
> > buf_len_73000:   14694343 entry-by-entry: 111740162 improvement
> > 86.849542
> > buf_len_100000:  13745969 entry-by-entry: 114151991 improvement
> > 87.958187
> > buf_len_234567:  14329834 entry-by-entry: 114427589 improvement
> > 87.476941
>
> It took me a while to figure out the meaning of 87.476941. It is probably
> a good idea to say 87.5% instead.

right, will change it in next version.