| Message ID | 20171124021208.13419-1-jakub.kicinski@netronome.com |
|---|---|
| Headers | show |
| Series | bpf: pass ifindex for bpf offload | expand |
On Thu, 23 Nov 2017 18:11:57 -0800 Jakub Kicinski <jakub.kicinski@netronome.com> wrote: > Hi! > > This series allows us to pass ifindex automatically when we > set up TC cls_bpf or XDP offload. There is a fair bit of > refactoring to separate the parse and load stages of lib/bpf.c. > In case of TC the skip_sw flag may come after the program > arguments (e.g. "bpf obj prog.o da skip_sw"), so we can't > just load the program as we parse the arguments. Note that > this impacts only loading of the program, all other supported > methods of finding a program (pinned, bytecode, bytefile-file) > are handled as previously, the load call will do nothing for > them. > > To simplify the implementation f_bpf and m_bpf will no longer > allow specifying programs multiple times. Device ifindex is > also resolved before running filter-specific code. > > > Jakub Kicinski (11): > bpf: pass program type in struct bpf_cfg_in > bpf: keep parsed program mode in struct bpf_cfg_in > bpf: allocate opcode table in struct bpf_cfg_in > bpf: split parse from program loading > bpf: rename bpf_parse_common() to bpf_parse_and_load_common() > bpf: expose bpf_parse_common() and bpf_load_common() > bpf: allow loading programs for a specific ifindex > {f,m}_bpf: don't allow specifying multiple bpf programs > tc_filter: resolve device name before parsing filter > f_bpf: communicate ifindex for eBPF offload > iplink: communicate ifindex for xdp offload > > include/bpf_util.h | 25 +++++++- > ip/iplink.c | 4 +- > ip/iplink_xdp.c | 13 ++++- > ip/iproute_lwtunnel.c | 3 +- > ip/xdp.h | 4 +- > lib/bpf.c | 155 ++++++++++++++++++++++++++++++-------------------- > tc/f_bpf.c | 18 +++++- > tc/m_bpf.c | 6 +- > tc/tc_filter.c | 50 ++++++++-------- > 9 files changed, 178 insertions(+), 100 deletions(-) > Looks good, applied. Thanks Jakub
Hi Stephen,
after merge of master into net-next branch the build is now broken
for BPF loader as follows:
# make
lib
make[1]: Entering directory '/home/darkstar/iproute2/lib'
CC libgenl.o
CC ll_map.o
CC libnetlink.o
libnetlink.c:120:2: warning: #warning "libmnl required for error support" [-Wcpp]
#warning "libmnl required for error support"
^~~~~~~
AR libnetlink.a
CC utils.o
CC rt_names.o
CC ll_types.o
CC ll_proto.o
CC ll_addr.o
CC inet_proto.o
CC namespace.o
CC json_writer.o
CC json_print.o
CC names.o
CC color.o
CC bpf.o
bpf.c: In function ‘bpf_prog_load_dev’:
bpf.c:1080:6: error: ‘union bpf_attr’ has no member named ‘prog_ifindex’; did you mean ‘prog_id’?
attr.prog_ifindex = ifindex;
^
../config.mk:39: recipe for target 'bpf.o' failed
make[1]: *** [bpf.o] Error 1
make[1]: Leaving directory '/home/darkstar/iproute2/lib'
Makefile:59: recipe for target 'all' failed
make: *** [all] Error 2
Can you rebase kernel headers to fix it, in net-next branch they
are not up to date which is why the build error occurs in combination
with below merged series?
Thanks,
Daniel
On 11/24/2017 03:11 AM, Jakub Kicinski wrote:
> Hi!
>
> This series allows us to pass ifindex automatically when we
> set up TC cls_bpf or XDP offload. There is a fair bit of
> refactoring to separate the parse and load stages of lib/bpf.c.
> In case of TC the skip_sw flag may come after the program
> arguments (e.g. "bpf obj prog.o da skip_sw"), so we can't
> just load the program as we parse the arguments. Note that
> this impacts only loading of the program, all other supported
> methods of finding a program (pinned, bytecode, bytefile-file)
> are handled as previously, the load call will do nothing for
> them.
>
> To simplify the implementation f_bpf and m_bpf will no longer
> allow specifying programs multiple times. Device ifindex is
> also resolved before running filter-specific code.
>
>
> Jakub Kicinski (11):
> bpf: pass program type in struct bpf_cfg_in
> bpf: keep parsed program mode in struct bpf_cfg_in
> bpf: allocate opcode table in struct bpf_cfg_in
> bpf: split parse from program loading
> bpf: rename bpf_parse_common() to bpf_parse_and_load_common()
> bpf: expose bpf_parse_common() and bpf_load_common()
> bpf: allow loading programs for a specific ifindex
> {f,m}_bpf: don't allow specifying multiple bpf programs
> tc_filter: resolve device name before parsing filter
> f_bpf: communicate ifindex for eBPF offload
> iplink: communicate ifindex for xdp offload
>
> include/bpf_util.h | 25 +++++++-
> ip/iplink.c | 4 +-
> ip/iplink_xdp.c | 13 ++++-
> ip/iproute_lwtunnel.c | 3 +-
> ip/xdp.h | 4 +-
> lib/bpf.c | 155 ++++++++++++++++++++++++++++++--------------------
> tc/f_bpf.c | 18 +++++-
> tc/m_bpf.c | 6 +-
> tc/tc_filter.c | 50 ++++++++--------
> 9 files changed, 178 insertions(+), 100 deletions(-)
>
On Wed, 29 Nov 2017 02:05:09 +0100 Daniel Borkmann <daniel@iogearbox.net> wrote: > Hi Stephen, > > after merge of master into net-next branch the build is now broken > for BPF loader as follows: > > # make > > lib > make[1]: Entering directory '/home/darkstar/iproute2/lib' > CC libgenl.o > CC ll_map.o > CC libnetlink.o > libnetlink.c:120:2: warning: #warning "libmnl required for error support" [-Wcpp] > #warning "libmnl required for error support" > ^~~~~~~ > AR libnetlink.a > CC utils.o > CC rt_names.o > CC ll_types.o > CC ll_proto.o > CC ll_addr.o > CC inet_proto.o > CC namespace.o > CC json_writer.o > CC json_print.o > CC names.o > CC color.o > CC bpf.o > bpf.c: In function ‘bpf_prog_load_dev’: > bpf.c:1080:6: error: ‘union bpf_attr’ has no member named ‘prog_ifindex’; did you mean ‘prog_id’? > attr.prog_ifindex = ifindex; > ^ > ../config.mk:39: recipe for target 'bpf.o' failed > make[1]: *** [bpf.o] Error 1 > make[1]: Leaving directory '/home/darkstar/iproute2/lib' > Makefile:59: recipe for target 'all' failed > make: *** [all] Error 2 > > Can you rebase kernel headers to fix it, in net-next branch they > are not up to date which is why the build error occurs in combination > with below merged series? > > Thanks, > Daniel > > > On 11/24/2017 03:11 AM, Jakub Kicinski wrote: > > Hi! > > > > This series allows us to pass ifindex automatically when we > > set up TC cls_bpf or XDP offload. There is a fair bit of > > refactoring to separate the parse and load stages of lib/bpf.c. > > In case of TC the skip_sw flag may come after the program > > arguments (e.g. "bpf obj prog.o da skip_sw"), so we can't > > just load the program as we parse the arguments. Note that > > this impacts only loading of the program, all other supported > > methods of finding a program (pinned, bytecode, bytefile-file) > > are handled as previously, the load call will do nothing for > > them. > > > > To simplify the implementation f_bpf and m_bpf will no longer > > allow specifying programs multiple times. Device ifindex is > > also resolved before running filter-specific code. > > > > > > Jakub Kicinski (11): > > bpf: pass program type in struct bpf_cfg_in > > bpf: keep parsed program mode in struct bpf_cfg_in > > bpf: allocate opcode table in struct bpf_cfg_in > > bpf: split parse from program loading > > bpf: rename bpf_parse_common() to bpf_parse_and_load_common() > > bpf: expose bpf_parse_common() and bpf_load_common() > > bpf: allow loading programs for a specific ifindex > > {f,m}_bpf: don't allow specifying multiple bpf programs > > tc_filter: resolve device name before parsing filter > > f_bpf: communicate ifindex for eBPF offload > > iplink: communicate ifindex for xdp offload > > > > include/bpf_util.h | 25 +++++++- > > ip/iplink.c | 4 +- > > ip/iplink_xdp.c | 13 ++++- > > ip/iproute_lwtunnel.c | 3 +- > > ip/xdp.h | 4 +- > > lib/bpf.c | 155 ++++++++++++++++++++++++++++++-------------------- > > tc/f_bpf.c | 18 +++++- > > tc/m_bpf.c | 6 +- > > tc/tc_filter.c | 50 ++++++++-------- > > 9 files changed, 178 insertions(+), 100 deletions(-) > > > Are you using your own uapi headers? On current master branch... $ git grep prog_ifindex include/uapi/linux/bpf.h: __u32 prog_ifindex; /* ifindex of netdev to prep for */ lib/bpf.c: attr.prog_ifindex = ifindex;
On Wed, 29 Nov 2017 02:05:09 +0100 Daniel Borkmann <daniel@iogearbox.net> wrote: > Hi Stephen, > > after merge of master into net-next branch the build is now broken > for BPF loader as follows: > > # make > > lib > make[1]: Entering directory '/home/darkstar/iproute2/lib' > CC libgenl.o > CC ll_map.o > CC libnetlink.o > libnetlink.c:120:2: warning: #warning "libmnl required for error support" [-Wcpp] > #warning "libmnl required for error support" > ^~~~~~~ > AR libnetlink.a > CC utils.o > CC rt_names.o > CC ll_types.o > CC ll_proto.o > CC ll_addr.o > CC inet_proto.o > CC namespace.o > CC json_writer.o > CC json_print.o > CC names.o > CC color.o > CC bpf.o > bpf.c: In function ‘bpf_prog_load_dev’: > bpf.c:1080:6: error: ‘union bpf_attr’ has no member named ‘prog_ifindex’; did you mean ‘prog_id’? > attr.prog_ifindex = ifindex; > ^ > ../config.mk:39: recipe for target 'bpf.o' failed > make[1]: *** [bpf.o] Error 1 > make[1]: Leaving directory '/home/darkstar/iproute2/lib' > Makefile:59: recipe for target 'all' failed > make: *** [all] Error 2 > > Can you rebase kernel headers to fix it, in net-next branch they > are not up to date which is why the build error occurs in combination > with below merged series? > > Thanks, > Daniel > > > On 11/24/2017 03:11 AM, Jakub Kicinski wrote: > > Hi! > > > > This series allows us to pass ifindex automatically when we > > set up TC cls_bpf or XDP offload. There is a fair bit of > > refactoring to separate the parse and load stages of lib/bpf.c. > > In case of TC the skip_sw flag may come after the program > > arguments (e.g. "bpf obj prog.o da skip_sw"), so we can't > > just load the program as we parse the arguments. Note that > > this impacts only loading of the program, all other supported > > methods of finding a program (pinned, bytecode, bytefile-file) > > are handled as previously, the load call will do nothing for > > them. > > > > To simplify the implementation f_bpf and m_bpf will no longer > > allow specifying programs multiple times. Device ifindex is > > also resolved before running filter-specific code. > > > > > > Jakub Kicinski (11): > > bpf: pass program type in struct bpf_cfg_in > > bpf: keep parsed program mode in struct bpf_cfg_in > > bpf: allocate opcode table in struct bpf_cfg_in > > bpf: split parse from program loading > > bpf: rename bpf_parse_common() to bpf_parse_and_load_common() > > bpf: expose bpf_parse_common() and bpf_load_common() > > bpf: allow loading programs for a specific ifindex > > {f,m}_bpf: don't allow specifying multiple bpf programs > > tc_filter: resolve device name before parsing filter > > f_bpf: communicate ifindex for eBPF offload > > iplink: communicate ifindex for xdp offload > > > > include/bpf_util.h | 25 +++++++- > > ip/iplink.c | 4 +- > > ip/iplink_xdp.c | 13 ++++- > > ip/iproute_lwtunnel.c | 3 +- > > ip/xdp.h | 4 +- > > lib/bpf.c | 155 ++++++++++++++++++++++++++++++-------------------- > > tc/f_bpf.c | 18 +++++- > > tc/m_bpf.c | 6 +- > > tc/tc_filter.c | 50 ++++++++-------- > > 9 files changed, 178 insertions(+), 100 deletions(-) > > > Sorry, net-next branch had out of date headers. Now fixed.
On 11/29/2017 03:17 AM, Stephen Hemminger wrote:
[...]
> Sorry, net-next branch had out of date headers. Now fixed.
Yeah, master worked fine. Thanks for the fix!
Hi! This series allows us to pass ifindex automatically when we set up TC cls_bpf or XDP offload. There is a fair bit of refactoring to separate the parse and load stages of lib/bpf.c. In case of TC the skip_sw flag may come after the program arguments (e.g. "bpf obj prog.o da skip_sw"), so we can't just load the program as we parse the arguments. Note that this impacts only loading of the program, all other supported methods of finding a program (pinned, bytecode, bytefile-file) are handled as previously, the load call will do nothing for them. To simplify the implementation f_bpf and m_bpf will no longer allow specifying programs multiple times. Device ifindex is also resolved before running filter-specific code. Jakub Kicinski (11): bpf: pass program type in struct bpf_cfg_in bpf: keep parsed program mode in struct bpf_cfg_in bpf: allocate opcode table in struct bpf_cfg_in bpf: split parse from program loading bpf: rename bpf_parse_common() to bpf_parse_and_load_common() bpf: expose bpf_parse_common() and bpf_load_common() bpf: allow loading programs for a specific ifindex {f,m}_bpf: don't allow specifying multiple bpf programs tc_filter: resolve device name before parsing filter f_bpf: communicate ifindex for eBPF offload iplink: communicate ifindex for xdp offload include/bpf_util.h | 25 +++++++- ip/iplink.c | 4 +- ip/iplink_xdp.c | 13 ++++- ip/iproute_lwtunnel.c | 3 +- ip/xdp.h | 4 +- lib/bpf.c | 155 ++++++++++++++++++++++++++++++-------------------- tc/f_bpf.c | 18 +++++- tc/m_bpf.c | 6 +- tc/tc_filter.c | 50 ++++++++-------- 9 files changed, 178 insertions(+), 100 deletions(-)