eBPF updates
Subsystems (5)
Daniel Borkmann
maintainercover eBPF updates
Two small eBPF helper additions to better match up with ancillary classic BPF functionality.
Daniel Borkmann (2):
ebpf: add prandom helper for packet sampling
ebpf: add helper for obtaining current processor id
include/linux/bpf.h | 3 +++ include/uapi/linux/bpf.h | 2 ++ kernel/bpf/core.c | 3 +++ kernel/bpf/helpers.c | 25 +++++++++++++++++++++++++ net/core/filter.c | 4 ++++ 5 files changed, 37 insertions(+)
Signature · 1 line
1.9.3
Daniel Borkmann
maintainerPATCH 1/2 ebpf: add prandom helper for packet sampling
This work is similar to commit 4cd3675ebf74 ("filter: added BPF random opcode") and adds a possibility for packet sampling in eBPF.
Currently, this is only possible in classic BPF and useful to combine sampling with f.e. packet sockets, possible also with tc.
Given classic BPF operates in 32bit space only, we leave the option to the eBPF application developer whether he needs a single or double word wide random number since eBPF is in 64bit space.
Example function proto-type looks like:
static u64 (*get_rnd)(bool dw) = (void *) BPF_FUNC_random;
--- include/linux/bpf.h | 2 ++ include/uapi/linux/bpf.h | 1 + kernel/bpf/core.c | 2 ++ kernel/bpf/helpers.c | 13 +++++++++++++ net/core/filter.c | 2 ++ 5 files changed, 20 insertions(+)
include/linux/bpf.h +2 -0
@@ -154,4 +154,6 @@ extern const struct bpf_func_proto bpf_map_lookup_elem_proto;extern const struct bpf_func_proto bpf_map_update_elem_proto;extern const struct bpf_func_proto bpf_map_delete_elem_proto; +extern const struct bpf_func_proto bpf_random_proto; +#endif /* _LINUX_BPF_H */
include/uapi/linux/bpf.h +1 -0
@@ -165,6 +165,7 @@ enum bpf_func_id { BPF_FUNC_map_lookup_elem,/* void *map_lookup_elem(&map, &key) */ BPF_FUNC_map_update_elem,/* int map_update_elem(&map, &key, &value, flags) */ BPF_FUNC_map_delete_elem,/* int map_delete_elem(&map, &key) */ + BPF_FUNC_random,/* prandom() as u32 or u64 */ __BPF_FUNC_MAX_ID, };
kernel/bpf/core.c +2 -0
@@ -661,6 +661,8 @@ const struct bpf_func_proto bpf_map_lookup_elem_proto __weak;const struct bpf_func_proto bpf_map_update_elem_proto__weak ;const struct bpf_func_proto bpf_map_delete_elem_proto__weak ; +const struct bpf_func_proto bpf_random_proto__weak ; +/* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call * skb_copy_bits(), so provide a weak definition of it for NET-less config. */
kernel/bpf/helpers.c +13 -0
@@ -11,6 +11,7 @@*/ #include <linux/bpf.h> #include <linux/rcupdate.h> +#include <linux/random.h> /* If kernel subsystem is allowing eBPF programs to call this function, * inside its own verifier_ops->get_func_proto() callback it should return
@@ -87,3 +88,15 @@ const struct bpf_func_proto bpf_map_delete_elem_proto = { .arg1_type = ARG_CONST_MAP_PTR, .arg2_type = ARG_PTR_TO_MAP_KEY, }; + +static u64 bpf_random (u64 dw,u64 r2,u64 r3,u64 r4,u64 r5) +{ +return prandom_u32 () | (dw ? ((u64 )prandom_u32 () <<32 ) :0 ); +} + +const struct bpf_func_proto bpf_random_proto = { + .func = bpf_random, + .gpl_only = false, + .ret_type =RET_INTEGER , + .arg1_type =ARG_ANYTHING , +};
net/core/filter.c +2 -0
@@ -1139,6 +1139,8 @@ sk_filter_func_proto(enum bpf_func_id func_id)return &bpf_map_update_elem_proto;case BPF_FUNC_map_delete_elem:return &bpf_map_delete_elem_proto; +case BPF_FUNC_random: +return &bpf_random_proto;default :return NULL ; }
Signature · 1 line
1.9.3
Re: [PATCH net-next 1/2] ebpf: add prandom helper for packet sampling
include/uapi/linux/bpf.h enum bpf_func_id { · line 168
Quoted 10 lines · Daniel Borkmann
include/uapi/linux/bpf.h +1
-0@@ -165,6 +165,7 @@ enum bpf_func_id { BPF_FUNC_map_lookup_elem,/* void *map_lookup_elem(&map, &key) */ BPF_FUNC_map_update_elem,/* int map_update_elem(&map, &key, &value, flags) */ BPF_FUNC_map_delete_elem,/* int map_delete_elem(&map, &key) */ + BPF_FUNC_random, /* prandom() as u32 or u64 */
can you rename it to BPF_FUNC_get_random_value ?
I'd like names to be as descriptive as possible.
User space can pick any name it likes anyway:
static u64 (*my_foo_func)() = (void *) BPF_FUNC_get_random_value;
Also it will make the following less obscure:
include/uapi/linux/bpf.h enum bpf_func_id { · line 169
Quoted 1 line
include/uapi/linux/bpf.h +1
-0+extern const struct bpf_func_proto bpf_random_proto;
'bpf_get_random_value_proto' looks better then 'bpf_random_proto' ;)
include/uapi/linux/bpf.h enum bpf_func_id {
Quoted 3 lines
include/uapi/linux/bpf.h +3
-0+static u64 bpf_random (u64 dw,u64 r2,u64 r3,u64 r4,u64 r5) +{ +return prandom_u32 () | (dw ? ((u64 )prandom_u32 () <<32 ) :0 );
can you make first argument to be a flag instead of bool ? 0 to return u32, 1 to return u64 and >=2 to be reserved and return -EINVAL?
This way we can extend it later. For example 2 might return 8 truly random bytes from get_random_bytes_arch()
include/uapi/linux/bpf.h
Quoted 3 lines
include/uapi/linux/bpf.h +3
-0+const struct bpf_func_proto bpf_random_proto = { + .func = bpf_random, + .gpl_only = false,
non-gpl allowed?
well I guess prandom_u32() is non-gpl either.
Daniel Borkmann
maintainerRe: [PATCH net-next 1/2] ebpf: add prandom helper for packet sampling
Quoted 14 lines · Alexei Starovoitov
include/uapi/linux/bpf.h enum bpf_func_id { · line 168
Quoted 10 lines · Daniel Borkmann
include/uapi/linux/bpf.h +1
-0@@ -165,6 +165,7 @@ enum bpf_func_id { BPF_FUNC_map_lookup_elem,/* void *map_lookup_elem(&map, &key) */ BPF_FUNC_map_update_elem,/* int map_update_elem(&map, &key, &value, flags) */ BPF_FUNC_map_delete_elem,/* int map_delete_elem(&map, &key) */ + BPF_FUNC_random, /* prandom() as u32 or u64 */can you rename it to
BPF_FUNC_get_random_value? I'd like names to be as descriptive as possible.
Ok, sure.
Quoted 8 lines
User space can pick any name it likes anyway:
static u64 (*my_foo_func)() = (void *)BPF_FUNC_get_random_value;Also it will make the following less obscure:
include/uapi/linux/bpf.h enum bpf_func_id { · line 169
Quoted 1 line
include/uapi/linux/bpf.h +1
-0+extern const struct bpf_func_proto bpf_random_proto;'
bpf_get_random_value_proto' looks better then 'bpf_random_proto' ;)
Quoted 9 lines
include/uapi/linux/bpf.h enum bpf_func_id {
Quoted 3 lines
include/uapi/linux/bpf.h +3
-0+static u64 bpf_random (u64 dw,u64 r2,u64 r3,u64 r4,u64 r5) +{ +return prandom_u32 () | (dw ? ((u64 )prandom_u32 () <<32 ) :0 );can you make first argument to be a flag instead of bool ? 0 to return u32, 1 to return u64 and >=2 to be reserved and return
-EINVAL?
This way we can extend it later. For example 2 might return 8 truly random bytes fromget_random_bytes_arch()
Yes, that's better.
Quoted 6 lines
include/uapi/linux/bpf.h
Quoted 3 lines
include/uapi/linux/bpf.h +3
-0+const struct bpf_func_proto bpf_random_proto = { + .func = bpf_random, + .gpl_only = false,non-gpl allowed?
well I guessprandom_u32()is non-gpl either.
It can be used by anyone, yes. If there are strong opinions to make this GPL-only, I don't mind, but I think it might be exaggerated.
Will respin.
Thanks,
Daniel
Daniel Borkmann
maintainerRe: [PATCH net-next 1/2] ebpf: add prandom helper for packet sampling
Quoted 1 line · Daniel Borkmann
On 03/13/2015 05:58 PM, Alexei Starovoitov wrote:
...
Quoted 7 lines
Quoted 7 lines
kernel/bpf/helpers.c const struct bpf_func_proto bpf_map_delete_elem_proto = { · line 94
Quoted 3 lines
kernel/bpf/helpers.c +3
-0+static u64 bpf_random (u64 dw,u64 r2,u64 r3,u64 r4,u64 r5) +{ +return prandom_u32 () | (dw ? ((u64 )prandom_u32 () <<32 ) :0 );can you make first argument to be a flag instead of bool ? 0 to return u32, 1 to return u64 and >=2 to be reserved and return
-EINVAL?
...
Quoted 2 lines
Quoted 2 lines
This way we can extend it later. For example 2 might return 8 truly random bytes from
get_random_bytes_arch()
Hmm, just got back to this thinking about it a bit more. I don't like that we would return with a negative error here. An eBPF application would then need to check for errors as well to make sure, and the signature returns max u64. So _theoretically_, -EINVAL could also come as a result from prandom() albeit unlikely. That would make the interface a bit ugly, imho.
I think the interface should be as simple as possible for obtaining a random number. I think that if we would want to support something like get_random_bytes_arch(), it would be better to do so in a new interface, where the name also more clearly indicates where the random data comes from.
That said, I'd rather prefer to leave the code as is, but go with the rename, maybe even BPF_FUNC_get_prandom_val so it indicates it's pseudo-random as opposed to, say, something like BPF_FUNC_get_srandom_val if we need that one day.
Cheers,
Daniel
Re: [PATCH net-next 1/2] ebpf: add prandom helper for packet sampling
Quoted 27 lines · Daniel Borkmann
Quoted 1 line · Daniel Borkmann
On 03/13/2015 05:58 PM, Alexei Starovoitov wrote:
...
Quoted 7 lines
Quoted 7 lines
kernel/bpf/helpers.c const struct bpf_func_proto bpf_map_delete_elem_proto = { · line 94
Quoted 3 lines
kernel/bpf/helpers.c +3
-0+static u64 bpf_random (u64 dw,u64 r2,u64 r3,u64 r4,u64 r5) +{ +return prandom_u32 () | (dw ? ((u64 )prandom_u32 () <<32 ) :0 );can you make first argument to be a flag instead of bool ? 0 to return u32, 1 to return u64 and >=2 to be reserved and return
-EINVAL?...
Quoted 2 lines
Quoted 2 lines
This way we can extend it later. For example 2 might return 8 truly random bytes from
get_random_bytes_arch()Hmm, just got back to this thinking about it a bit more. I don't like that we would return with a negative error here. An eBPF application would then need to check for errors as well to make sure, and the signature returns max u64. So _theoretically_,
-EINVALcould also come as a result fromprandom()albeit unlikely. That would make the interface a bit ugly, imho.I think the interface should be as simple as possible for obtaining a random number. I think that if we would want to support something like
get_random_bytes_arch(), it would be better to do so in a new interface, where the name also more clearly indicates where the random data comes from.
Quoted 4 lines
That said, I'd rather prefer to leave the code as is, but go with the rename, maybe even
BPF_FUNC_get_prandom_valso it indicates it's pseudo-random as opposed to, say, something likeBPF_FUNC_get_srandom_valif we need that one day.
fair point.
I was also thinking about it more and I believe that 'bool' flag is redundant.
If program wants to have 64-bit value it should just call this helper twice. It actually will be faster, since such 64-bit value will be without branches whereas 'dw' flag introduces if condition in critical path that will be executed for every packet. So my vote is for BPF_FUNC_get_prandom_u32 that returns prandom_u32() and nothing else.
Daniel Borkmann
maintainerRe: [PATCH net-next 1/2] ebpf: add prandom helper for packet sampling
On 03/13/2015 11:53 PM, Alexei Starovoitov wrote:
...
Quoted 2 lines
So my vote is for
BPF_FUNC_get_prandom_u32that returnsprandom_u32()and nothing else.
I'll go for that. Occam's razor. ;)
Daniel Borkmann
maintainerPATCH 2/2 ebpf: add helper for obtaining current processor id
In reply to Daniel BorkmannThis patch adds the possibility to obtain raw_smp_processor_id() in eBPF. Currently, this is only possible in classic BPF where commit da2033c28226 ("filter: add SKF_AD_RXHASH and SKF_AD_CPU") has added facilities for this.
Single threaded packet socket applications (f.e. tcpdump, netsniff-ng) e.g. writing to pcap files and making use of eBPF can use this helper function to spread flows after RPS dispatch to several instances. It would also allow to track per CPU statistics (or data) with eBPF maps.
Example function proto-type looks like:
unsigned int (*get_cpu_id)(void) = (void *) BPF_FUNC_cpu_id;
--- include/linux/bpf.h | 1 + include/uapi/linux/bpf.h | 1 + kernel/bpf/core.c | 1 + kernel/bpf/helpers.c | 12 ++++++++++++ net/core/filter.c | 2 ++ 5 files changed, 17 insertions(+)
include/linux/bpf.h +1 -0
@@ -155,5 +155,6 @@ extern const struct bpf_func_proto bpf_map_update_elem_proto;extern const struct bpf_func_proto bpf_map_delete_elem_proto;extern const struct bpf_func_proto bpf_random_proto; +extern const struct bpf_func_proto bpf_cpu_id_proto;#endif /* _LINUX_BPF_H */
include/uapi/linux/bpf.h +1 -0
@@ -166,6 +166,7 @@ enum bpf_func_id { BPF_FUNC_map_update_elem,/* int map_update_elem(&map, &key, &value, flags) */ BPF_FUNC_map_delete_elem,/* int map_delete_elem(&map, &key) */ BPF_FUNC_random,/* prandom() as u32 or u64 */ + BPF_FUNC_cpu_id,/* raw_smp_processor_id() */ __BPF_FUNC_MAX_ID, };
kernel/bpf/core.c +1 -0
@@ -662,6 +662,7 @@ const struct bpf_func_proto bpf_map_update_elem_proto __weak;const struct bpf_func_proto bpf_map_delete_elem_proto__weak ;const struct bpf_func_proto bpf_random_proto__weak ; +const struct bpf_func_proto bpf_cpu_id_proto__weak ;/* To execute LD_ABS/LD_IND instructions __bpf_prog_run() may call * skb_copy_bits(), so provide a weak definition of it for NET-less config.
kernel/bpf/helpers.c +12 -0
@@ -12,6 +12,7 @@#include <linux/bpf.h> #include <linux/rcupdate.h> #include <linux/random.h> +#include <linux/smp.h> /* If kernel subsystem is allowing eBPF programs to call this function, * inside its own verifier_ops->get_func_proto() callback it should return
@@ -100,3 +101,14 @@ const struct bpf_func_proto bpf_random_proto = { .ret_type = RET_INTEGER, .arg1_type = ARG_ANYTHING, }; + +static u64 bpf_cpu_id (u64 r1,u64 r2,u64 r3,u64 r4,u64 r5) +{ +return raw_smp_processor_id (); +} + +const struct bpf_func_proto bpf_cpu_id_proto = { + .func = bpf_cpu_id, + .gpl_only = false, + .ret_type =RET_INTEGER , +};
net/core/filter.c +2 -0
@@ -1141,6 +1141,8 @@ sk_filter_func_proto(enum bpf_func_id func_id)return &bpf_map_delete_elem_proto;case BPF_FUNC_random:return &bpf_random_proto; +case BPF_FUNC_cpu_id: +return &bpf_cpu_id_proto;default :return NULL ; }
Signature · 1 line
1.9.3
Re: [PATCH net-next 2/2] ebpf: add helper for obtaining current processor id
Quoted 9 lines · Daniel Borkmann
This patch adds the possibility to obtain
raw_smp_processor_id()in eBPF. Currently, this is only possible in classic BPF where commitda2033c28226("filter: addSKF_AD_RXHASHandSKF_AD_CPU") has added facilities for this.Single threaded packet socket applications (f.e. tcpdump, netsniff-ng) e.g. writing to pcap files and making use of eBPF can use this helper function to spread flows after RPS dispatch to several instances. It would also allow to track per CPU statistics (or data) with eBPF maps.
I'm not sure whether classic extension is still used by Eric as it was envisioned in that old commit, but I do like addition of it for eBPF, since it allows a poor men's per-cpu data structures via maps :)
include/uapi/linux/bpf.h enum bpf_func_id { · line 169
Quoted 9 lines
include/uapi/linux/bpf.h +1
-0@@ -166,6 +166,7 @@ enum bpf_func_id { BPF_FUNC_map_update_elem,/* int map_update_elem(&map, &key, &value, flags) */ BPF_FUNC_map_delete_elem,/* int map_delete_elem(&map, &key) */ BPF_FUNC_random,/* prandom() as u32 or u64 */ + BPF_FUNC_cpu_id, /* raw_smp_processor_id() */
same comment as in other patch:
could you rename it to BPF_FUNC_get_smp_processor_id ?
Daniel Borkmann
maintainerRe: [PATCH net-next 2/2] ebpf: add helper for obtaining current processor id
On 03/13/2015 06:06 PM, Alexei Starovoitov wrote:
...
Quoted 2 lines
same comment as in other patch:
could you rename it toBPF_FUNC_get_smp_processor_id?
Will do, yep.
From: Daniel Borkmann <[email protected]>
Date: Fri, 13 Mar 2015 14:22:02 +0100
Quoted 2 lines
Two small eBPF helper additions to better match up with ancillary classic BPF functionality.
Applied, thanks.