: BPF_CGROUP_{GET,SET}SOCKOPT handling when optlen > PAGE_SIZE
Martin KaFai Lau
reviewerThe cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen > PAGE_SIZE (or optval_end - optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return -EFAULT to the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or used bpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the ctx->optlen before returning -EFAULT to the user space when ctx.optlen > max_optlen?
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.
Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?
Good point on chaining being broken because of this requirement :-/
With tracking, we need to be careful, because the following situation might be problematic:
Suppose setsockopt is larger than 4k, the program can rewrite some byte in the first 4k, not touch optlen and expect this to work. Currently, optlen=0 explicitly means "ignore whatever is in the bpf buffer and use the original one".
If we can have a tracking that catches situations like this - we should be able to drop that optlen=0 requirement.
IIRC, that's the only tricky part.
Martin KaFai Lau
reviewerQuoted 17 lines · Stanislav Fomichev
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?
Quoted 6 lines
Good point on chaining being broken because of this requirement :-/
With tracking, we need to be careful, because the following situation might be problematic:
Suppose setsockopt is larger than 4k, the program can rewrite some byte in the first 4k, not touch optlen and expect this to work.
If the bpf prog rewrites the first 4k, it must change the ctx.optlen to get it work. Otherwise, the kernel will return -EFAULT because the ctx.optlen is larger than the max_optlen (or optval_end - optval).
Quoted 4 lines
Currently, optlen=0 explicitly means "ignore whatever is in the bpf
buffer and use the original one" > If we can have a tracking that catches situations like this - we should be able to drop that optlen=0 requirement.
IIRC, that's the only tricky part.
Ah, I meant, in __cgroup_bpf_run_filter_setsockopt, use "!ctx.optlen_changed && ctx.optlen > max_optlen" test to imply "ignore whatever is in the bpf
buffer and use the original one". Add 'bool optlen_changed' to 'struct bpf_sockopt_kern' and set ctx.optlen_changed to true in cg_sockopt_convert_ctx_access() whenever there is BPF_WRITE to ctx.optlen. Would it work or may be I am still missing something in the writing first 4k case above?
Quoted 35 lines · Martin KaFai Lau
Quoted 17 lines · Stanislav Fomichev
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?Quoted 6 lines
Good point on chaining being broken because of this requirement :-/
With tracking, we need to be careful, because the following situation might be problematic:
Suppose setsockopt is larger than 4k, the program can rewrite some byte in the first 4k, not touch optlen and expect this to work.If the bpf prog rewrites the first 4k, it must change the ctx.optlen to get it work. Otherwise, the kernel will return
-EFAULTbecause the ctx.optlen is larger than themax_optlen(oroptval_end- optval).Quoted 4 lines
Currently, optlen=0 explicitly means "ignore whatever is in the bpf
buffer and use the original one" > If we can have a tracking that catches situations like this - we should be able to drop that optlen=0 requirement.
IIRC, that's the only tricky part.
Quoted 7 lines
Ah, I meant, in
__cgroup_bpf_run_filter_setsockopt, use "!ctx.optlen_changed && ctx.optlen >max_optlen" test to imply "ignore whatever is in the bpf buffer and use the original one". Add 'booloptlen_changed' to 'structbpf_sockopt_kern' and set ctx.optlen_changed to true incg_sockopt_convert_ctx_access()whenever there isBPF_WRITEto ctx.optlen. Would it work or may be I am still missing something in the writing first 4k case above?
What if the program wants to keep optlen as is? Here is the hypothetical case: ctx->optlen is 8k, we allocate/expose only the first 4k, the program does ctx->optval[0] = 0xff and doesn't change the optlen. It wants the rest of the payload to be passed as is with only the first byte changed.
The condition "!ctx.optlen_changed && ctx.optlen > max_optlen" is true, so, if we treat this as explicit optlen=0, we ignore the program's changes.
But this is not what the program has intended, right? It wants to amend something and pass the rest as is.
It seems like we need to have both optlen_changed and optval_changed. If both are false, we should be able to safely do optlen=0 equivalent. Tracking only optlen seems to be problematic?
Martin KaFai Lau
reviewerQuoted 44 lines · Stanislav Fomichev
Quoted 35 lines · Martin KaFai Lau
Quoted 17 lines · Stanislav Fomichev
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?Quoted 6 lines
Good point on chaining being broken because of this requirement :-/
With tracking, we need to be careful, because the following situation might be problematic:
Suppose setsockopt is larger than 4k, the program can rewrite some byte in the first 4k, not touch optlen and expect this to work.If the bpf prog rewrites the first 4k, it must change the ctx.optlen to get it work. Otherwise, the kernel will return
-EFAULTbecause the ctx.optlen is larger than themax_optlen(oroptval_end- optval).Quoted 4 lines
Currently, optlen=0 explicitly means "ignore whatever is in the bpf
buffer and use the original one" > If we can have a tracking that catches situations like this - we should be able to drop that optlen=0 requirement.
IIRC, that's the only tricky part.Quoted 7 lines
Ah, I meant, in
__cgroup_bpf_run_filter_setsockopt, use "!ctx.optlen_changed && ctx.optlen >max_optlen" test to imply "ignore whatever is in the bpf buffer and use the original one". Add 'booloptlen_changed' to 'structbpf_sockopt_kern' and set ctx.optlen_changed to true incg_sockopt_convert_ctx_access()whenever there isBPF_WRITEto ctx.optlen. Would it work or may be I am still missing something in the writing first 4k case above?
Quoted 5 lines
What if the program wants to keep optlen as is? Here is the hypothetical case:
ctx->optlenis 8k, we allocate/expose only the first 4k, the program does ctx->optval[0] =0xffand doesn't change the optlen. It wants the rest of the payload to be passed as is with only the first byte changed.
I think we are talking about the same case but we may have different understanding on how the current __cgroup_bpf_run_filter_setsockopt() is handling it.
I don't see the current kernel supports this now. If the bpf prog does not change the ctx->optlen from 8k to something that is <= 4k, the kernel will just return -EFAULT in here, no?
else if (ctx.optlen /* 8k */ > max_optlen /* 4k */ || ctx.optlen < -1) {
/* optlen is out of bounds */
ret = -EFAULT;
}or you meant the future change needs to consider this new case and also support gluing the first 4k (that was exposed to the bpf prog) with the second 4k (that was not exposed to the bpf prog)?
Quoted 9 lines
The condition "!ctx.optlen_changed && ctx.optlen >
max_optlen" is true, so, if we treat this as explicit optlen=0, we ignore the program's changes.
But this is not what the program has intended, right? It wants to amend something and pass the rest as is.It seems like we need to have both
optlen_changedandoptval_changed. If both are false, we should be able to safely do optlen=0 equivalent. Tracking only optlen seems to be problematic?
Quoted 59 lines · Martin KaFai Lau
Quoted 44 lines · Stanislav Fomichev
Quoted 35 lines · Martin KaFai Lau
Quoted 17 lines · Stanislav Fomichev
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?Quoted 6 lines
Good point on chaining being broken because of this requirement :-/
With tracking, we need to be careful, because the following situation might be problematic:
Suppose setsockopt is larger than 4k, the program can rewrite some byte in the first 4k, not touch optlen and expect this to work.If the bpf prog rewrites the first 4k, it must change the ctx.optlen to get it work. Otherwise, the kernel will return
-EFAULTbecause the ctx.optlen is larger than themax_optlen(oroptval_end- optval).Quoted 4 lines
Currently, optlen=0 explicitly means "ignore whatever is in the bpf
buffer and use the original one" > If we can have a tracking that catches situations like this - we should be able to drop that optlen=0 requirement.
IIRC, that's the only tricky part.Quoted 7 lines
Ah, I meant, in
__cgroup_bpf_run_filter_setsockopt, use "!ctx.optlen_changed && ctx.optlen >max_optlen" test to imply "ignore whatever is in the bpf buffer and use the original one". Add 'booloptlen_changed' to 'structbpf_sockopt_kern' and set ctx.optlen_changed to true incg_sockopt_convert_ctx_access()whenever there isBPF_WRITEto ctx.optlen. Would it work or may be I am still missing something in the writing first 4k case above?Quoted 5 lines
What if the program wants to keep optlen as is? Here is the hypothetical case:
ctx->optlenis 8k, we allocate/expose only the first 4k, the program does ctx->optval[0] =0xffand doesn't change the optlen. It wants the rest of the payload to be passed as is with only the first byte changed.I think we are talking about the same case but we may have different understanding on how the current
__cgroup_bpf_run_filter_setsockopt()is handling it.I don't see the current kernel supports this now. If the bpf prog does not change the
ctx->optlenfrom 8k to something that is <= 4k, the kernel will just return-EFAULTin here, no?
Quoted 14 lines
else if (ctx.optlen /* 8k */ > max_optlen /* 4k */ || ctx.optlen < -1) { /* optlen is out of bounds */ ret = -EFAULT; }or you meant the future change needs to consider this new case and also support gluing the first 4k (that was exposed to the bpf prog) with the second 4k (that was not exposed to the bpf prog)?
Quoted 5 lines
The condition "!ctx.optlen_changed && ctx.optlen >
max_optlen" is true, so, if we treat this as explicit optlen=0, we ignore the program's changes.
But this is not what the program has intended, right? It wants to amend something and pass the rest as is.
Right, I'm not talking about how it's handled now. Now optlen > max_optlen triggers EFAULT.
But in the future, if we add tracking, we want 'optlen > max_optlen' to behave as explicit 'optlen = 0' as long as the user hasn't changed the optlen _and_ also hasn't changed anything in the buffer.
Quoted 4 lines
Quoted 3 lines
It seems like we need to have both
optlen_changedandoptval_changed. If both are false, we should be able to safely do optlen=0 equivalent. Tracking only optlen seems to be problematic?
Martin KaFai Lau
reviewerQuoted 75 lines · Stanislav Fomichev
Quoted 59 lines · Martin KaFai Lau
Quoted 44 lines · Stanislav Fomichev
Quoted 35 lines · Martin KaFai Lau
Quoted 17 lines · Stanislav Fomichev
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?Quoted 6 lines
Good point on chaining being broken because of this requirement :-/
With tracking, we need to be careful, because the following situation might be problematic:
Suppose setsockopt is larger than 4k, the program can rewrite some byte in the first 4k, not touch optlen and expect this to work.If the bpf prog rewrites the first 4k, it must change the ctx.optlen to get it work. Otherwise, the kernel will return
-EFAULTbecause the ctx.optlen is larger than themax_optlen(oroptval_end- optval).Quoted 4 lines
Currently, optlen=0 explicitly means "ignore whatever is in the bpf
buffer and use the original one" > If we can have a tracking that catches situations like this - we should be able to drop that optlen=0 requirement.
IIRC, that's the only tricky part.Quoted 7 lines
Ah, I meant, in
__cgroup_bpf_run_filter_setsockopt, use "!ctx.optlen_changed && ctx.optlen >max_optlen" test to imply "ignore whatever is in the bpf buffer and use the original one". Add 'booloptlen_changed' to 'structbpf_sockopt_kern' and set ctx.optlen_changed to true incg_sockopt_convert_ctx_access()whenever there isBPF_WRITEto ctx.optlen. Would it work or may be I am still missing something in the writing first 4k case above?Quoted 5 lines
What if the program wants to keep optlen as is? Here is the hypothetical case:
ctx->optlenis 8k, we allocate/expose only the first 4k, the program does ctx->optval[0] =0xffand doesn't change the optlen. It wants the rest of the payload to be passed as is with only the first byte changed.I think we are talking about the same case but we may have different understanding on how the current
__cgroup_bpf_run_filter_setsockopt()is handling it.I don't see the current kernel supports this now. If the bpf prog does not change the
ctx->optlenfrom 8k to something that is <= 4k, the kernel will just return-EFAULTin here, no?Quoted 14 lines
else if (ctx.optlen /* 8k */ > max_optlen /* 4k */ || ctx.optlen < -1) { /* optlen is out of bounds */ ret = -EFAULT; }or you meant the future change needs to consider this new case and also support gluing the first 4k (that was exposed to the bpf prog) with the second 4k (that was not exposed to the bpf prog)?
Quoted 5 lines
The condition "!ctx.optlen_changed && ctx.optlen >
max_optlen" is true, so, if we treat this as explicit optlen=0, we ignore the program's changes.
But this is not what the program has intended, right? It wants to amend something and pass the rest as is.
Quoted 5 lines
Right, I'm not talking about how it's handled now. Now optlen >
max_optlentriggers EFAULT.
But in the future, if we add tracking, we want 'optlen >max_optlen' to behave as explicit 'optlen = 0' as long as the user hasn't changed the optlen _and_ also hasn't changed anything in the buffer.
Ah, ic.
Tracking the runtime buffer change will be hard as of the current state through the ctx->optval. I don't think we need to track that either. If the existing bpf prog wants the changed buf to be used, it must have changed the optlen already. Thus, tracking optlen only should be as good.
If the bpf prog is depending on the kernel to do implicit -EFAULT like this, yes, it will break even the buffer change is tracked.
if (ctx->optlen > ctx->optval_end - ctx->optval)
return 1; /* 0 will be -EPERM, so 1 here to make kernel return -EFAULT for
us */
I would argue that it is more like a surprise than a feature if the bpf prog depends on ctx.optlen > max_optlen (only for the > 4k case though) to do an implicit reject (through EFAULT) instead of directly using the 'return 0' or bpf_set_retval() which is exactly how it should be done to reject other "normal" integer optval.
I am also not sure how useful it is to expose partial data to the bpf prog and have a way for the bpf prog to tell the kernel to join the remaining. Instead, it would be more useful to have API for the bpf prog to have access to the whole data instead.
Quoted 4 lines
Quoted 4 lines
Quoted 3 lines
It seems like we need to have both
optlen_changedandoptval_changed. If both are false, we should be able to safely do optlen=0 equivalent. Tracking only optlen seems to be problematic?
Quoted 83 lines · Martin KaFai Lau
Quoted 75 lines · Stanislav Fomichev
Quoted 59 lines · Martin KaFai Lau
Quoted 44 lines · Stanislav Fomichev
Quoted 35 lines · Martin KaFai Lau
Quoted 17 lines · Stanislav Fomichev
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?Quoted 6 lines
Good point on chaining being broken because of this requirement :-/
With tracking, we need to be careful, because the following situation might be problematic:
Suppose setsockopt is larger than 4k, the program can rewrite some byte in the first 4k, not touch optlen and expect this to work.If the bpf prog rewrites the first 4k, it must change the ctx.optlen to get it work. Otherwise, the kernel will return
-EFAULTbecause the ctx.optlen is larger than themax_optlen(oroptval_end- optval).Quoted 4 lines
Currently, optlen=0 explicitly means "ignore whatever is in the bpf
buffer and use the original one" > If we can have a tracking that catches situations like this - we should be able to drop that optlen=0 requirement.
IIRC, that's the only tricky part.Quoted 7 lines
Ah, I meant, in
__cgroup_bpf_run_filter_setsockopt, use "!ctx.optlen_changed && ctx.optlen >max_optlen" test to imply "ignore whatever is in the bpf buffer and use the original one". Add 'booloptlen_changed' to 'structbpf_sockopt_kern' and set ctx.optlen_changed to true incg_sockopt_convert_ctx_access()whenever there isBPF_WRITEto ctx.optlen. Would it work or may be I am still missing something in the writing first 4k case above?Quoted 5 lines
What if the program wants to keep optlen as is? Here is the hypothetical case:
ctx->optlenis 8k, we allocate/expose only the first 4k, the program does ctx->optval[0] =0xffand doesn't change the optlen. It wants the rest of the payload to be passed as is with only the first byte changed.I think we are talking about the same case but we may have different understanding on how the current
__cgroup_bpf_run_filter_setsockopt()is handling it.I don't see the current kernel supports this now. If the bpf prog does not change the
ctx->optlenfrom 8k to something that is <= 4k, the kernel will just return-EFAULTin here, no?Quoted 14 lines
else if (ctx.optlen /* 8k */ > max_optlen /* 4k */ || ctx.optlen < -1) { /* optlen is out of bounds */ ret = -EFAULT; }or you meant the future change needs to consider this new case and also support gluing the first 4k (that was exposed to the bpf prog) with the second 4k (that was not exposed to the bpf prog)?
Quoted 5 lines
The condition "!ctx.optlen_changed && ctx.optlen >
max_optlen" is true, so, if we treat this as explicit optlen=0, we ignore the program's changes.
But this is not what the program has intended, right? It wants to amend something and pass the rest as is.Quoted 5 lines
Right, I'm not talking about how it's handled now. Now optlen >
max_optlentriggers EFAULT.
But in the future, if we add tracking, we want 'optlen >max_optlen' to behave as explicit 'optlen = 0' as long as the user hasn't changed the optlen _and_ also hasn't changed anything in the buffer.
Quoted 6 lines
Ah, ic.
Tracking the runtime buffer change will be hard as of the current state through the
ctx->optval. I don't think we need to track that either. If the existing bpf prog wants the changed buf to be used, it must have changed the optlen already. Thus, tracking optlen only should be as good.
I might be still missing something on why tracking optlen is enough?
Consider this BPF program:
SEC("cgroup/setsockopt")
int _setsockopt(struct bpf_sockopt *ctx)
{
__u8 *optval_end = ctx->optval_end;
__u8 *optval = ctx->optval;
if (optval + 1 > optval_end) return 0;
optval[0] = 0xff;
return 1;
}And the userspace doing the following:
__u32 buf[4096*2] = {};
ret = setsockopt(fd, SOME_LEVEL, SOME_OPTLEN, &buf, sizeof(buf));
Right now, without explicit 'optlen = 0' in the BPF program, we'll get -1/EFAULT here (unarguably, this is a bad interface, but still better than ignoring program's buf?).
If we track only optlen in the program, we'd get success, but the changed buffer will be ignored by the kernel. (what am I missing here?)
Quoted 6 lines
If the bpf prog is depending on the kernel to do implicit
-EFAULTlike this, yes, it will break even the buffer change is tracked.if (
ctx->optlen>ctx->optval_end-ctx->optval)
return 1; /* 0 will be-EPERM, so 1 here to make kernel return-EFAULTfor
us */
[..]
Quoted 5 lines
I would argue that it is more like a surprise than a feature if the bpf prog depends on ctx.optlen >
max_optlen(only for the > 4k case though) to do an implicit reject (through EFAULT) instead of directly using the 'return 0' orbpf_set_retval()which is exactly how it should be done to reject other "normal" integer optval.
That all comes from the issue above. We want to have a contract with the bpf program: when optlen>4k, it has to do something with the optlen (set it to 0 to ignore, set it to <4096 to pass to the kernel). It can't just change something in the 4k of the exposed buffer and assume this data will be passed to the kernel.
Quoted 4 lines
I am also not sure how useful it is to expose partial data to the bpf prog and have a way for the bpf prog to tell the kernel to join the remaining. Instead, it would be more useful to have API for the bpf prog to have access to the whole data instead.
That seems like a better way to go? We didn't do that initially because the data is in the __user memory and we can't pass it to bpf; we had to do this extra copy/allocation :-( I think we decided against copying everything because this can be abused due to no sane limit on the setsockopt value size. Nothing prevents userspace from passing a huge buffer when doing, say, SO_MARK; the kernel will read the first int and be happy with it.
Quoted 5 lines
Quoted 4 lines
Quoted 4 lines
Quoted 3 lines
It seems like we need to have both
optlen_changedandoptval_changed. If both are false, we should be able to safely do optlen=0 equivalent. Tracking only optlen seems to be problematic?
Martin KaFai Lau
reviewerQuoted 112 lines · Stanislav Fomichev
Quoted 83 lines · Martin KaFai Lau
Quoted 75 lines · Stanislav Fomichev
Quoted 59 lines · Martin KaFai Lau
Quoted 44 lines · Stanislav Fomichev
Quoted 35 lines · Martin KaFai Lau
Quoted 17 lines · Stanislav Fomichev
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?Quoted 6 lines
Good point on chaining being broken because of this requirement :-/
With tracking, we need to be careful, because the following situation might be problematic:
Suppose setsockopt is larger than 4k, the program can rewrite some byte in the first 4k, not touch optlen and expect this to work.If the bpf prog rewrites the first 4k, it must change the ctx.optlen to get it work. Otherwise, the kernel will return
-EFAULTbecause the ctx.optlen is larger than themax_optlen(oroptval_end- optval).Quoted 4 lines
Currently, optlen=0 explicitly means "ignore whatever is in the bpf
buffer and use the original one" > If we can have a tracking that catches situations like this - we should be able to drop that optlen=0 requirement.
IIRC, that's the only tricky part.Quoted 7 lines
Ah, I meant, in
__cgroup_bpf_run_filter_setsockopt, use "!ctx.optlen_changed && ctx.optlen >max_optlen" test to imply "ignore whatever is in the bpf buffer and use the original one". Add 'booloptlen_changed' to 'structbpf_sockopt_kern' and set ctx.optlen_changed to true incg_sockopt_convert_ctx_access()whenever there isBPF_WRITEto ctx.optlen. Would it work or may be I am still missing something in the writing first 4k case above?Quoted 5 lines
What if the program wants to keep optlen as is? Here is the hypothetical case:
ctx->optlenis 8k, we allocate/expose only the first 4k, the program does ctx->optval[0] =0xffand doesn't change the optlen. It wants the rest of the payload to be passed as is with only the first byte changed.I think we are talking about the same case but we may have different understanding on how the current
__cgroup_bpf_run_filter_setsockopt()is handling it.I don't see the current kernel supports this now. If the bpf prog does not change the
ctx->optlenfrom 8k to something that is <= 4k, the kernel will just return-EFAULTin here, no?Quoted 14 lines
else if (ctx.optlen /* 8k */ > max_optlen /* 4k */ || ctx.optlen < -1) { /* optlen is out of bounds */ ret = -EFAULT; }or you meant the future change needs to consider this new case and also support gluing the first 4k (that was exposed to the bpf prog) with the second 4k (that was not exposed to the bpf prog)?
Quoted 5 lines
The condition "!ctx.optlen_changed && ctx.optlen >
max_optlen" is true, so, if we treat this as explicit optlen=0, we ignore the program's changes.
But this is not what the program has intended, right? It wants to amend something and pass the rest as is.Quoted 5 lines
Right, I'm not talking about how it's handled now. Now optlen >
max_optlentriggers EFAULT.
But in the future, if we add tracking, we want 'optlen >max_optlen' to behave as explicit 'optlen = 0' as long as the user hasn't changed the optlen _and_ also hasn't changed anything in the buffer.Quoted 6 lines
Ah, ic.
Tracking the runtime buffer change will be hard as of the current state through the
ctx->optval. I don't think we need to track that either. If the existing bpf prog wants the changed buf to be used, it must have changed the optlen already. Thus, tracking optlen only should be as good.I might be still missing something on why tracking optlen is enough?
Consider this BPF program:
SEC("cgroup/setsockopt")
int_setsockopt(struct bpf_sockopt *ctx){ __u8 *optval_end = ctx->optval_end; __u8 *optval = ctx->optval; if (optval + 1 > optval_end) return 0; optval[0] = 0xff; return 1; }And the userspace doing the following:
__u32buf[4096*2] = {};
ret =setsockopt(fd, SOME_LEVEL, SOME_OPTLEN, &buf, sizeof(buf));
Quoted 6 lines
Right now, without explicit 'optlen = 0' in the BPF program, we'll get -1/EFAULT here (unarguably, this is a bad interface, but still better than ignoring program's buf?).
If we track only optlen in the program, we'd get success, but the changed buffer will be ignored by the kernel. (what am I missing here?)
Right, this will break if the bpf prog depends on this -EFAULT behavior in anyway. Similar to my example below, tracking the buffer change still won't be enough because we don't know the intention of the bpf prog (changed but forgot to update optlen or it does want to return -EFAULT).
After these few examples in the thread, I think this optlen and buffer tracking does not seem to be a tangible path to solve it. It seems like it is only papering around it.
Quoted 27 lines
Quoted 7 lines
If the bpf prog is depending on the kernel to do implicit
-EFAULTlike this, yes, it will break even the buffer change is tracked.if (
ctx->optlen>ctx->optval_end-ctx->optval)
return 1; /* 0 will be-EPERM, so 1 here to make kernel return-EFAULTfor
us */[..]
Quoted 5 lines
I would argue that it is more like a surprise than a feature if the bpf prog depends on ctx.optlen >
max_optlen(only for the > 4k case though) to do an implicit reject (through EFAULT) instead of directly using the 'return 0' orbpf_set_retval()which is exactly how it should be done to reject other "normal" integer optval.That all comes from the issue above. We want to have a contract with the bpf program: when optlen>4k, it has to do something with the optlen (set it to 0 to ignore, set it to <4096 to pass to the kernel). It can't just change something in the 4k of the exposed buffer and assume this data will be passed to the kernel.
Quoted 4 lines
I am also not sure how useful it is to expose partial data to the bpf prog and have a way for the bpf prog to tell the kernel to join the remaining. Instead, it would be more useful to have API for the bpf prog to have access to the whole data instead.
Quoted 7 lines
That seems like a better way to go? We didn't do that initially because the data is in the
__usermemory and we can't pass it to bpf; we had to do this extra copy/allocation :-( I think we decided against copying everything because this can be abused due to no sane limit on the setsockopt value size. Nothing prevents userspace from passing a huge buffer when doing, say,SO_MARK; the kernel will read the first int and be happy with it.
yeah, may be one thing for the future API is to avoid the pre allocation. There is bpf_copy_from_user but it needs to be sleepable.
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Quoted 5 lines
Quoted 4 lines
It seems like we need to have both
optlen_changedandoptval_changed. If both are false, we should be able to safely do optlen=0 equivalent. Tracking only optlen seems to be problematic?
Quoted 165 lines · Martin KaFai Lau
Quoted 112 lines · Stanislav Fomichev
Quoted 83 lines · Martin KaFai Lau
Quoted 75 lines · Stanislav Fomichev
Quoted 59 lines · Martin KaFai Lau
Quoted 44 lines · Stanislav Fomichev
Quoted 35 lines · Martin KaFai Lau
Quoted 17 lines · Stanislav Fomichev
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?Quoted 6 lines
Good point on chaining being broken because of this requirement :-/
With tracking, we need to be careful, because the following situation might be problematic:
Suppose setsockopt is larger than 4k, the program can rewrite some byte in the first 4k, not touch optlen and expect this to work.If the bpf prog rewrites the first 4k, it must change the ctx.optlen to get it work. Otherwise, the kernel will return
-EFAULTbecause the ctx.optlen is larger than themax_optlen(oroptval_end- optval).Quoted 4 lines
Currently, optlen=0 explicitly means "ignore whatever is in the bpf
buffer and use the original one" > If we can have a tracking that catches situations like this - we should be able to drop that optlen=0 requirement.
IIRC, that's the only tricky part.Quoted 7 lines
Ah, I meant, in
__cgroup_bpf_run_filter_setsockopt, use "!ctx.optlen_changed && ctx.optlen >max_optlen" test to imply "ignore whatever is in the bpf buffer and use the original one". Add 'booloptlen_changed' to 'structbpf_sockopt_kern' and set ctx.optlen_changed to true incg_sockopt_convert_ctx_access()whenever there isBPF_WRITEto ctx.optlen. Would it work or may be I am still missing something in the writing first 4k case above?Quoted 5 lines
What if the program wants to keep optlen as is? Here is the hypothetical case:
ctx->optlenis 8k, we allocate/expose only the first 4k, the program does ctx->optval[0] =0xffand doesn't change the optlen. It wants the rest of the payload to be passed as is with only the first byte changed.I think we are talking about the same case but we may have different understanding on how the current
__cgroup_bpf_run_filter_setsockopt()is handling it.I don't see the current kernel supports this now. If the bpf prog does not change the
ctx->optlenfrom 8k to something that is <= 4k, the kernel will just return-EFAULTin here, no?Quoted 14 lines
else if (ctx.optlen /* 8k */ > max_optlen /* 4k */ || ctx.optlen < -1) { /* optlen is out of bounds */ ret = -EFAULT; }or you meant the future change needs to consider this new case and also support gluing the first 4k (that was exposed to the bpf prog) with the second 4k (that was not exposed to the bpf prog)?
Quoted 5 lines
The condition "!ctx.optlen_changed && ctx.optlen >
max_optlen" is true, so, if we treat this as explicit optlen=0, we ignore the program's changes.
But this is not what the program has intended, right? It wants to amend something and pass the rest as is.Quoted 5 lines
Right, I'm not talking about how it's handled now. Now optlen >
max_optlentriggers EFAULT.
But in the future, if we add tracking, we want 'optlen >max_optlen' to behave as explicit 'optlen = 0' as long as the user hasn't changed the optlen _and_ also hasn't changed anything in the buffer.Quoted 6 lines
Ah, ic.
Tracking the runtime buffer change will be hard as of the current state through the
ctx->optval. I don't think we need to track that either. If the existing bpf prog wants the changed buf to be used, it must have changed the optlen already. Thus, tracking optlen only should be as good.I might be still missing something on why tracking optlen is enough?
Consider this BPF program:
SEC("cgroup/setsockopt")
int_setsockopt(struct bpf_sockopt *ctx){ __u8 *optval_end = ctx->optval_end; __u8 *optval = ctx->optval; if (optval + 1 > optval_end) return 0; optval[0] = 0xff; return 1; }And the userspace doing the following:
__u32buf[4096*2] = {};
ret =setsockopt(fd, SOME_LEVEL, SOME_OPTLEN, &buf, sizeof(buf));Quoted 6 lines
Right now, without explicit 'optlen = 0' in the BPF program, we'll get -1/EFAULT here (unarguably, this is a bad interface, but still better than ignoring program's buf?).
If we track only optlen in the program, we'd get success, but the changed buffer will be ignored by the kernel. (what am I missing here?)Right, this will break if the bpf prog depends on this
-EFAULTbehavior in anyway. Similar to my example below, tracking the buffer change still won't be enough because we don't know the intention of the bpf prog (changed but forgot to update optlen or it does want to return-EFAULT).After these few examples in the thread, I think this optlen and buffer tracking does not seem to be a tangible path to solve it. It seems like it is only papering around it.
Quoted 27 lines
Quoted 7 lines
If the bpf prog is depending on the kernel to do implicit
-EFAULTlike this, yes, it will break even the buffer change is tracked.if (
ctx->optlen>ctx->optval_end-ctx->optval)
return 1; /* 0 will be-EPERM, so 1 here to make kernel return-EFAULTfor
us */[..]
Quoted 5 lines
I would argue that it is more like a surprise than a feature if the bpf prog depends on ctx.optlen >
max_optlen(only for the > 4k case though) to do an implicit reject (through EFAULT) instead of directly using the 'return 0' orbpf_set_retval()which is exactly how it should be done to reject other "normal" integer optval.That all comes from the issue above. We want to have a contract with the bpf program: when optlen>4k, it has to do something with the optlen (set it to 0 to ignore, set it to <4096 to pass to the kernel). It can't just change something in the 4k of the exposed buffer and assume this data will be passed to the kernel.
Quoted 4 lines
I am also not sure how useful it is to expose partial data to the bpf prog and have a way for the bpf prog to tell the kernel to join the remaining. Instead, it would be more useful to have API for the bpf prog to have access to the whole data instead.
Quoted 7 lines
That seems like a better way to go? We didn't do that initially because the data is in the
__usermemory and we can't pass it to bpf; we had to do this extra copy/allocation :-( I think we decided against copying everything because this can be abused due to no sane limit on the setsockopt value size. Nothing prevents userspace from passing a huge buffer when doing, say,SO_MARK; the kernel will read the first int and be happy with it.
Quoted 2 lines
yeah, may be one thing for the future API is to avoid the pre allocation. There is
bpf_copy_from_userbut it needs to be sleepable.
+1, having a sleepable version might be cleaner alternative
Quoted 7 lines
Quoted 6 lines
Quoted 6 lines
Quoted 5 lines
Quoted 5 lines
Quoted 4 lines
It seems like we need to have both
optlen_changedandoptval_changed. If both are false, we should be able to safely do optlen=0 equivalent. Tracking only optlen seems to be problematic?
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.
Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?
Maybe optlen + optval/optval_end could be replaced with dynptr? If we do a new type of dynptr (DYNPTR_CTXBUF or something like that), we can implement tracking of whether it was ever modified through bpf_dynptr_write() or if direct memory access was ever used (was bpf_dynptr_data() called). Not sure how you'd know if bpf_dynptr_data() was used to modify data, though (this is where bpf_dynptr_data_rdonly() vs bpf_dynptr_data() would be helpful, perhaps). But just a seed of an idea, maybe you guys can somehow fit it here?
Martin KaFai Lau
reviewerQuoted 17 lines · Andrii Nakryiko
Quoted 12 lines · Martin KaFai Lau
The cgroup-bpf {get,set}sockopt prog is useful to change the optname behavior. The bpf prog usually just handles a few specific optnames and ignores most others. For the optnames that it ignores, it usually does not need to change the optlen. The exception is when optlen >
PAGE_SIZE(oroptval_end- optval). The bpf prog needs to set the optlen to 0 for this case or else the kernel will return-EFAULTto the userspace. It is usually not what the bpf prog wants because the bpf prog only expects error returning to userspace when it has explicitly 'return 0;' or usedbpf_set_retval(). If a bpf prog always changes optlen for optnames that it does not care to 0, it may risk if the latter bpf prog in the same cgroup may want to change/look-at it.Quoted 3 lines
Would like to explore if there is an easier way for the bpf prog to handle it. eg. does it make sense to track if the bpf prog has changed the
ctx->optlenbefore returning-EFAULTto the user space when ctx.optlen >max_optlen?
Quoted 9 lines
Maybe optlen + optval/optval_end could be replaced with dynptr? If we do a new type of dynptr (
DYNPTR_CTXBUFor something like that), we can implement tracking of whether it was ever modified throughbpf_dynptr_write()or if direct memory access was ever used (wasbpf_dynptr_data()called). Not sure how you'd know ifbpf_dynptr_data()was used to modify data, though (this is wherebpf_dynptr_data_rdonly()vsbpf_dynptr_data()would be helpful, perhaps). But just a seed of an idea, maybe you guys can somehow fit it here?
Yep, dynptr can be used here. May be one dynptr specifically for sockptr_t. The dynptr will have the __user pointer first to avoid a big (and potentially bogus) kernel buffer pre allocation. Then only read and write it through the bpf_dynptr_read() and write(). Since it is __user pointer, it won't be directly accessible through data slice with bpf_dynptr_data(). It should not be a perf issue, most of the common optval is an integer. The worst common case is the 16 bytes tcp-cc name. The bpf prog has to be sleepable first though which I think will be a useful thing to have.