bpf: fix couple of netdevsim issues
Subsystems (5)
cover bpf: fix couple of netdevsim issues
The first one is found by the syzkaller and needs to make a proper distinction between offloaded and dev-bound cases.
The second one restores behavior where the offloaded/dev-bound programs with a dead netdev were not correctly skipped during BPF_PROG_GET_NEXT_ID.
Stanislav Fomichev (2):
netdevsim: don't accept device bound programs
bpf: bring back removal of dev-bound id from idr
drivers/net/netdevsim/bpf.c | 4 ++-- include/linux/bpf.h | 2 ++ kernel/bpf/offload.c | 3 +++ kernel/bpf/syscall.c | 15 +++++++++++---- 4 files changed, 18 insertions(+), 6 deletions(-)
Signature · 1 line
2.42.0.869.gea05f2083d-goog
PATCH 1/2 netdevsim: don't accept device bound programs
Commit 2b3486bc2d23 ("bpf: Introduce device-bound XDP programs") introduced device-bound programs by largely reusing existing offloading infrastructure. This changed the semantics of 'prog->aux->offload' a bit. Now, it's non-null for both offloaded and device-bound programs.
Instead of looking at 'prog->aux->offload' let's call bpf_prog_is_offloaded which should be true iff the program is offloaded and not merely device-bound.
--- drivers/net/netdevsim/bpf.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/net/netdevsim/bpf.c +2 -2
@@ -93,7 +93,7 @@ static void nsim_prog_set_loaded(struct bpf_prog *prog, bool loaded) {struct nsim_bpf_bound_prog *state; -if (!prog || !prog->aux->offload) +if (!prog || !bpf_prog_is_offloaded (prog->aux))return ; state = prog->aux->offload->dev_priv;
@@ -311,7 +311,7 @@ nsim_setup_prog_hw_checks(struct netdevsim *ns, struct netdev_bpf *bpf)if (!bpf->prog)return 0 ; -if (!bpf->prog->aux->offload) { +if (!bpf_prog_is_offloaded (bpf->prog->aux)) {NSIM_EA (bpf->extack,"xdpoffload of non-bound program" );return -EINVAL ; }
Signature · 1 line
2.42.0.869.gea05f2083d-goog
Re: [PATCH bpf-next 1/2] netdevsim: don't accept device bound programs
Quoted 43 lines · Stanislav Fomichev
Commit
2b3486bc2d23("bpf: Introduce device-bound XDP programs") introduced device-bound programs by largely reusing existing offloading infrastructure. This changed the semantics of 'prog->aux->offload' a bit. Now, it's non-null for both offloaded and device-bound programs.Instead of looking at '
prog->aux->offload' let's callbpf_prog_is_offloadedwhich should be true iff the program is offloaded and not merely device-bound.Cc Dipendra Khadka <[email protected]>Reported-by [email protected]Fixes 2b3486bc2d23 ("bpf: Introduce device-bound XDP programs")Signed-off-by Stanislav Fomichev <[email protected]>--- drivers/net/netdevsim/bpf.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)drivers/net/netdevsim/bpf.c +2
-2@@ -93,7 +93,7 @@ static void nsim_prog_set_loaded(struct bpf_prog *prog, bool loaded) {struct nsim_bpf_bound_prog *state; -if (!prog || !prog->aux->offload) +if (!prog || !bpf_prog_is_offloaded (prog->aux))return ; state = prog->aux->offload->dev_priv;@@ -311,7 +311,7 @@ nsim_setup_prog_hw_checks(struct netdevsim *ns, struct netdev_bpf *bpf)if (!bpf->prog)return 0 ; -if (!bpf->prog->aux->offload) { +if (!bpf_prog_is_offloaded (bpf->prog->aux)) {NSIM_EA (bpf->extack,"xdpoffload of non-bound program" );return -EINVAL ; }Signature · 1 line
2.42.0.869.gea05f2083d-goog
Forgot to CC netdev of these..
Re: [PATCH bpf-next 1/2] netdevsim: don't accept device bound programs
In reply to Stanislav FomichevQuoted 9 lines · On Mon, 13 Nov 2023 20:54:52 -0800 Stanislav Fomichev wrote:
Commit
2b3486bc2d23("bpf: Introduce device-bound XDP programs") introduced device-bound programs by largely reusing existing offloading infrastructure. This changed the semantics of 'prog->aux->offload' a bit. Now, it's non-null for both offloaded and device-bound programs.Instead of looking at '
prog->aux->offload' let's callbpf_prog_is_offloadedwhich should be true iff the program is offloaded and not merely device-bound.
PATCH 2/2 bpf: bring back removal of dev-bound id from idr
In reply to Stanislav FomichevCommit ef01f4e25c17 ("bpf: restore the ebpf program ID for BPF_AUDIT_UNLOAD and PERF_BPF_EVENT_PROG_UNLOAD") stopped removing program's id from idr when the offloaded/bound netdev goes away. I was supposed to take a look and check in [0], but apparently I did not.
The purpose of idr removal is to avoid BPF_PROG_GET_NEXT_ID returning stale ids for the programs that have a dead netdev. This functionality is verified by test_offload.py, but we don't run this test in the CI.
Introduce new bpf_prog_remove_from_idr which takes care of correctly dealing with potential double idr_remove() via separate skip_idr_remove flag in the aux.
Verified by running the test manually:test_offload.py: OK
0: https://lore.kernel.org/all/CAKH8qBtyR20ZWAc11z1-6pGb3Hd47AQUTbE_cfoktG59TqaJ7Q@mail.gmail.com/
--- include/linux/bpf.h | 2 ++ kernel/bpf/offload.c | 3 +++ kernel/bpf/syscall.c | 15 +++++++++++---- 3 files changed, 16 insertions(+), 4 deletions(-)
include/linux/bpf.h +2 -0
@@ -1414,6 +1414,7 @@ struct bpf_prog_aux {bool xdp_has_frags;bool exception_cb;bool exception_boundary; +bool skip_idr_remove;/* BTF_KIND_FUNC_PROTO for valid attach_btf_id */ const struct btf_type *attach_func_proto;/* function name for valid attach_btf_id */
@@ -2049,6 +2050,7 @@ void bpf_prog_inc(struct bpf_prog *prog);struct bpf_prog *__must_check bpf_prog_inc_not_zero (struct bpf_prog *prog);void bpf_prog_put (struct bpf_prog *prog); +void bpf_prog_remove_from_idr (struct bpf_prog *prog);void bpf_prog_free_id (struct bpf_prog *prog);void bpf_map_free_id (struct bpf_map *map);
kernel/bpf/offload.c +3 -0
@@ -112,6 +112,9 @@ static void __bpf_prog_offload_destroy(struct bpf_prog *prog)if (offload->dev_state) offload->offdev->ops->destroy (prog); +/* Make sure BPF_PROG_GET_NEXT_ID can't find this dead program */ +bpf_prog_remove_from_idr (prog); +list_del_init (&offload->offloads);kfree (offload); prog->aux->offload =NULL ;
kernel/bpf/syscall.c +11 -4
@@ -2083,10 +2083,19 @@ static int bpf_prog_alloc_id(struct bpf_prog *prog)return id >0 ?0 : id; } -void bpf_prog_free_id (struct bpf_prog *prog) +void bpf_prog_remove_from_idr (struct bpf_prog *prog) {unsigned long flags; +spin_lock_irqsave (&prog_idr_lock, flags); +if (!prog->aux->skip_idr_remove) +idr_remove (&prog_idr, prog->aux->id); + prog->aux->skip_idr_remove =1 ; +spin_unlock_irqrestore (&prog_idr_lock, flags); +} + +void bpf_prog_free_id (struct bpf_prog *prog) +{/* cBPF to eBPF migrations are currently not in the idr store. * Offloaded programs are removed from the store when their device * disappears - even if someone grabs an fd to them they are unusable,
@@ -2095,10 +2104,8 @@ void bpf_prog_free_id(struct bpf_prog *prog)if (!prog->aux->id)return ; -spin_lock_irqsave (&prog_idr_lock, flags); -idr_remove (&prog_idr, prog->aux->id); +bpf_prog_remove_from_idr (prog); prog->aux->id =0 ; -spin_unlock_irqrestore (&prog_idr_lock, flags); }static void __bpf_prog_put_rcu (struct rcu_head *rcu)
Signature · 1 line
2.42.0.869.gea05f2083d-goog
Re: [PATCH bpf-next 2/2] bpf: bring back removal of dev-bound id from idr
Quoted 100 lines · Stanislav Fomichev
Commit
ef01f4e25c17("bpf: restore the ebpf program ID forBPF_AUDIT_UNLOADandPERF_BPF_EVENT_PROG_UNLOAD") stopped removing program's id from idr when the offloaded/bound netdev goes away. I was supposed to take a look and check in [0], but apparently I did not.The purpose of idr removal is to avoid
BPF_PROG_GET_NEXT_IDreturning stale ids for the programs that have a dead netdev. This functionality is verified bytest_offload.py, but we don't run this test in the CI.Introduce new
bpf_prog_remove_from_idrwhich takes care of correctly dealing with potential doubleidr_remove()via separateskip_idr_removeflag in the aux.Verified by running the test manually:
test_offload.py: OK0: https://lore.kernel.org/all/CAKH8qBtyR20ZWAc11z1-6pGb3Hd47AQUTbE_cfoktG59TqaJ7Q@mail.gmail.com/
Fixes ef01f4e25c17 ("bpf: restore the ebpf program ID for BPF_AUDIT_UNLOAD and PERF_BPF_EVENT_PROG_UNLOAD")Signed-off-by Stanislav Fomichev <[email protected]>--- include/linux/bpf.h | 2 ++ kernel/bpf/offload.c | 3 +++ kernel/bpf/syscall.c | 15 +++++++++++---- 3 files changed, 16 insertions(+), 4 deletions(-)include/linux/bpf.h +2
-0@@ -1414,6 +1414,7 @@ struct bpf_prog_aux {bool xdp_has_frags;bool exception_cb;bool exception_boundary; +bool skip_idr_remove;/* BTF_KIND_FUNC_PROTO for valid attach_btf_id */ const struct btf_type *attach_func_proto;/* function name for valid attach_btf_id */ @@ -2049,6 +2050,7 @@ void bpf_prog_inc(struct bpf_prog *prog);struct bpf_prog *__must_check bpf_prog_inc_not_zero (struct bpf_prog *prog);void bpf_prog_put (struct bpf_prog *prog); +void bpf_prog_remove_from_idr (struct bpf_prog *prog);void bpf_prog_free_id (struct bpf_prog *prog);void bpf_map_free_id (struct bpf_map *map);kernel/bpf/offload.c +3
-0@@ -112,6 +112,9 @@ static void __bpf_prog_offload_destroy(struct bpf_prog *prog)if (offload->dev_state) offload->offdev->ops->destroy (prog); +/* Make sure BPF_PROG_GET_NEXT_ID can't find this dead program */ +bpf_prog_remove_from_idr (prog); +list_del_init (&offload->offloads);kfree (offload); prog->aux->offload =NULL ;kernel/bpf/syscall.c +11
-4@@ -2083,10 +2083,19 @@ static int bpf_prog_alloc_id(struct bpf_prog *prog)return id >0 ?0 : id; } -void bpf_prog_free_id (struct bpf_prog *prog) +void bpf_prog_remove_from_idr (struct bpf_prog *prog) {unsigned long flags; +spin_lock_irqsave (&prog_idr_lock, flags); +if (!prog->aux->skip_idr_remove) +idr_remove (&prog_idr, prog->aux->id); + prog->aux->skip_idr_remove =1 ; +spin_unlock_irqrestore (&prog_idr_lock, flags); +} + +void bpf_prog_free_id (struct bpf_prog *prog) +{/* cBPF to eBPF migrations are currently not in the idr store. * Offloaded programs are removed from the store when their device * disappears - even if someone grabs an fd to them they are unusable, @@ -2095,10 +2104,8 @@ void bpf_prog_free_id(struct bpf_prog *prog)if (!prog->aux->id)return ; -spin_lock_irqsave (&prog_idr_lock, flags); -idr_remove (&prog_idr, prog->aux->id); +bpf_prog_remove_from_idr (prog); prog->aux->id =0 ; -spin_unlock_irqrestore (&prog_idr_lock, flags); }static void __bpf_prog_put_rcu (struct rcu_head *rcu)Signature · 1 line
2.42.0.869.gea05f2083d-goog
Same here, should have been CC'ed to netdev.
Martin KaFai Lau
reviewerRe: [PATCH bpf-next 2/2] bpf: bring back removal of dev-bound id from idr
In reply to Stanislav FomichevQuoted 7 lines · Stanislav Fomichev
Commit
ef01f4e25c17("bpf: restore the ebpf program ID forBPF_AUDIT_UNLOADandPERF_BPF_EVENT_PROG_UNLOAD") stopped removing program's id from idr when the offloaded/bound netdev goes away. I was supposed to take a look and check in [0], but apparently I did not.The purpose of idr removal is to avoid
BPF_PROG_GET_NEXT_IDreturning stale ids for the programs that have a dead netdev. This functionality
What may be wrong if BPF_PROG_GET_NEXT_ID returns the id?
e.g. If the prog is pinned somewhere, it may be useful to know a prog is still loaded in the system.
Does the fixes mean to be for the bpf tree instead?
Quoted 39 lines
is verified by
test_offload.py, but we don't run this test in the CI.Introduce new
bpf_prog_remove_from_idrwhich takes care of correctly dealing with potential doubleidr_remove()via separateskip_idr_removeflag in the aux.Verified by running the test manually:
test_offload.py: OK0: https://lore.kernel.org/all/CAKH8qBtyR20ZWAc11z1-6pGb3Hd47AQUTbE_cfoktG59TqaJ7Q@mail.gmail.com/
Fixes ef01f4e25c17 ("bpf: restore the ebpf program ID for BPF_AUDIT_UNLOAD and PERF_BPF_EVENT_PROG_UNLOAD")Signed-off-by Stanislav Fomichev <[email protected]>--- include/linux/bpf.h | 2 ++ kernel/bpf/offload.c | 3 +++ kernel/bpf/syscall.c | 15 +++++++++++---- 3 files changed, 16 insertions(+), 4 deletions(-)include/linux/bpf.h +2
-0@@ -1414,6 +1414,7 @@ struct bpf_prog_aux {bool xdp_has_frags;bool exception_cb;bool exception_boundary; +bool skip_idr_remove;/* BTF_KIND_FUNC_PROTO for valid attach_btf_id */ const struct btf_type *attach_func_proto;/* function name for valid attach_btf_id */ @@ -2049,6 +2050,7 @@ void bpf_prog_inc(struct bpf_prog *prog);struct bpf_prog *__must_check bpf_prog_inc_not_zero (struct bpf_prog *prog);void bpf_prog_put (struct bpf_prog *prog); +void bpf_prog_remove_from_idr (struct bpf_prog *prog);void bpf_prog_free_id (struct bpf_prog *prog);void bpf_map_free_id (struct bpf_map *map);
Daniel Borkmann
maintainerRe: [PATCH bpf-next 2/2] bpf: bring back removal of dev-bound id from idr
Quoted 11 lines · Martin KaFai Lau
Quoted 7 lines · Stanislav Fomichev
Commit
ef01f4e25c17("bpf: restore the ebpf program ID forBPF_AUDIT_UNLOADandPERF_BPF_EVENT_PROG_UNLOAD") stopped removing program's id from idr when the offloaded/bound netdev goes away. I was supposed to take a look and check in [0], but apparently I did not.The purpose of idr removal is to avoid
BPF_PROG_GET_NEXT_IDreturning stale ids for the programs that have a dead netdev. This functionalityWhat may be wrong if
BPF_PROG_GET_NEXT_IDreturns the id?
e.g. If the prog is pinned somewhere, it may be useful to know a prog is still loaded in the system.
Wouldn't this strictly speaking provide an invalid id (== 0) upon unload back to audit - see the bpf_audit_prog(prog, BPF_AUDIT_UNLOAD) call location?
Quoted 1 line
Does the fixes mean to be for the bpf tree instead?
+1 given syzbot report, I'll take the first one in for now.
Quoted 39 lines
Quoted 37 lines
is verified by
test_offload.py, but we don't run this test in the CI.Introduce new
bpf_prog_remove_from_idrwhich takes care of correctly dealing with potential doubleidr_remove()via separateskip_idr_removeflag in the aux.Verified by running the test manually:
test_offload.py: OK0: https://lore.kernel.org/all/CAKH8qBtyR20ZWAc11z1-6pGb3Hd47AQUTbE_cfoktG59TqaJ7Q@mail.gmail.com/
Fixes ef01f4e25c17 ("bpf: restore the ebpf program ID for BPF_AUDIT_UNLOAD and PERF_BPF_EVENT_PROG_UNLOAD")Signed-off-by Stanislav Fomichev <[email protected]>---
include/linux/bpf.h | 2 ++
kernel/bpf/offload.c| 3 +++
kernel/bpf/syscall.c| 15 +++++++++++----3 files changed, 16 insertions(+), 4 deletions(-)include/linux/bpf.h +2
-0@@ -1414,6 +1414,7 @@ struct bpf_prog_aux { bool xdp_has_frags; bool exception_cb; bool exception_boundary; + bool skip_idr_remove; /* BTF_KIND_FUNC_PROTO for valid attach_btf_id */ const struct btf_type *attach_func_proto; /* function name for valid attach_btf_id */@@ -2049,6 +2050,7 @@ void bpf_prog_inc(struct bpf_prog *prog); struct bpf_prog * __must_check bpf_prog_inc_not_zero(struct bpf_prog *prog); void bpf_prog_put(struct bpf_prog *prog); +void bpf_prog_remove_from_idr (struct bpf_prog *prog); void bpf_prog_free_id(struct bpf_prog *prog); void bpf_map_free_id(struct bpf_map *map);
Re: [PATCH bpf-next 2/2] bpf: bring back removal of dev-bound id from idr
Quoted 7 lines · Daniel Borkmann
Quoted 6 lines · Martin KaFai Lau
Quoted 5 lines · Stanislav Fomichev
Commit
ef01f4e25c17("bpf: restore the ebpf program ID forBPF_AUDIT_UNLOADandPERF_BPF_EVENT_PROG_UNLOAD") stopped removing program's id from idr when the offloaded/bound netdev goes away. I was supposed to take a look and check in [0], but apparently I did not.
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines
The purpose of idr removal is to avoid
BPF_PROG_GET_NEXT_IDreturning stale ids for the programs that have a dead netdev. This functionalityWhat may be wrong if
BPF_PROG_GET_NEXT_IDreturns the id?
e.g. If the prog is pinned somewhere, it may be useful to know a prog is still loaded in the system.
bpftool is a bit spooked by those prog ids currently: calling GET_INFO_BY_ID on those programs returns ENODEV. So we can keep those ids around, but need some tweaks on the bpftool in this case. LMK if any of you prefer this option.
Quoted 2 lines
Wouldn't this strictly speaking provide an invalid id (== 0) upon unload back to audit - see the
bpf_audit_prog(prog, BPF_AUDIT_UNLOAD)call location?
Removing from idr shouldn't affect bpf_audit_prog, right? bpf_audit_prog is using prog->aux->id for its purposes, so as long as we are not resetting this value - we're good.
Martin KaFai Lau
reviewerRe: [PATCH bpf-next 2/2] bpf: bring back removal of dev-bound id from idr
Quoted 8 lines · Stanislav Fomichev
Quoted 8 lines
Quoted 8 lines
Quoted 7 lines
Commit
ef01f4e25c17("bpf: restore the ebpf program ID forBPF_AUDIT_UNLOADandPERF_BPF_EVENT_PROG_UNLOAD") stopped removing program's id from idr when the offloaded/bound netdev goes away. I was supposed to take a look and check in [0], but apparently I did not.The purpose of idr removal is to avoid
BPF_PROG_GET_NEXT_IDreturning stale ids for the programs that have a dead netdev. This functionality
Quoted 7 lines
Quoted 2 lines
Quoted 2 lines
What may be wrong if
BPF_PROG_GET_NEXT_IDreturns the id?
e.g. If the prog is pinned somewhere, it may be useful to know a prog is still loaded in the system.bpftool is a bit spooked by those prog ids currently: calling
GET_INFO_BY_IDon those programs returns ENODEV. So we can keep those ids around, but need some tweaks on the bpftool in this case. LMK if any of you prefer this option.
I think it is in general useful to improve 'bpftool prog show' to keep going for the next prog id if possible. May be print an error message after the prog id and then keep going for the next prog id?
Re: [PATCH bpf-next 2/2] bpf: bring back removal of dev-bound id from idr
Quoted 13 lines · Martin KaFai Lau
Quoted 11 lines · Stanislav Fomichev
Quoted 10 lines
Quoted 10 lines
Quoted 7 lines
Commit
ef01f4e25c17("bpf: restore the ebpf program ID forBPF_AUDIT_UNLOADandPERF_BPF_EVENT_PROG_UNLOAD") stopped removing program's id from idr when the offloaded/bound netdev goes away. I was supposed to take a look and check in [0], but apparently I did not.The purpose of idr removal is to avoid
BPF_PROG_GET_NEXT_IDreturning stale ids for the programs that have a dead netdev. This functionalityWhat may be wrong if
BPF_PROG_GET_NEXT_IDreturns the id?
e.g. If the prog is pinned somewhere, it may be useful to know a prog is still loaded in the system.
Quoted 8 lines
Quoted 4 lines
bpftool is a bit spooked by those prog ids currently: calling
GET_INFO_BY_IDon those programs returns ENODEV. So we can keep those ids around, but need some tweaks on the bpftool in this case. LMK if any of you prefer this option.I think it is in general useful to improve 'bpftool prog show' to keep going for the next prog id if possible. May be print an error message after the prog id and then keep going for the next prog id?
Replied with a v2 where I mark those progs as 'orphaned'!
Daniel Borkmann
maintainerRe: [PATCH bpf-next 2/2] bpf: bring back removal of dev-bound id from idr
Quoted 18 lines · Stanislav Fomichev
Quoted 17 lines · Martin KaFai Lau
Quoted 15 lines · Stanislav Fomichev
Quoted 10 lines
Quoted 10 lines
Quoted 7 lines
Commit
ef01f4e25c17("bpf: restore the ebpf program ID forBPF_AUDIT_UNLOADandPERF_BPF_EVENT_PROG_UNLOAD") stopped removing program's id from idr when the offloaded/bound netdev goes away. I was supposed to take a look and check in [0], but apparently I did not.The purpose of idr removal is to avoid
BPF_PROG_GET_NEXT_IDreturning stale ids for the programs that have a dead netdev. This functionalityWhat may be wrong if
BPF_PROG_GET_NEXT_IDreturns the id?
e.g. If the prog is pinned somewhere, it may be useful to know a prog is still loaded in the system.bpftool is a bit spooked by those prog ids currently: calling
GET_INFO_BY_IDon those programs returns ENODEV. So we can keep those ids around, but need some tweaks on the bpftool in this case. LMK if any of you prefer this option.
Quoted 5 lines
Quoted 3 lines
I think it is in general useful to improve 'bpftool prog show' to keep going for the next prog id if possible. May be print an error message after the prog id and then keep going for the next prog id?
Replied with a v2 where I mark those progs as 'orphaned'!
Sg, we could perhaps do something similar for netdev detached links.