Re: [PATCH bpf-next 2/5] bpf, fsverity: Add kfunc bpf_get_fsverity_digest

Re: [PATCH bpf-next 2/5] bpf, fsverity: Add kfunc bpf_get_fsverity_digest

[Eric Biggers]

On Fri, Oct 13, 2023 at 11:26:41AM -0700, Song Liu wrote:
> The kfunc can be used to read fsverity_digest, so that we can verify
> signature in BPF LSM.
> 
> This kfunc is added to fs/verity/measure.c because some data structure used
> in the function is private to fsverity (fs/verity/fsverity_private.h).
> 
> Signed-off-by: Song Liu <song@kernel.org>
> ---
>  fs/verity/measure.c | 66 +++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 66 insertions(+)
> 
> diff --git a/fs/verity/measure.c b/fs/verity/measure.c
> index eec5956141da..2d4b2e6f5a5d 100644
> --- a/fs/verity/measure.c
> +++ b/fs/verity/measure.c
> @@ -8,6 +8,8 @@
>  #include "fsverity_private.h"
>  
>  #include <linux/uaccess.h>
> +#include <linux/bpf.h>
> +#include <linux/btf.h>
>  
>  /**
>   * fsverity_ioctl_measure() - get a verity file's digest
> @@ -100,3 +102,67 @@ int fsverity_get_digest(struct inode *inode,
>  	return hash_alg->digest_size;
>  }
>  EXPORT_SYMBOL_GPL(fsverity_get_digest);
> +
> +/* bpf kfuncs */
> +__diag_push();
> +__diag_ignore_all("-Wmissing-prototypes",
> +		  "kfuncs which will be used in BPF programs");
> +
> +/**
> + * bpf_get_fsverity_digest: read fsverity digest of file
> + * @file: file to get digest from
> + * @digest_ptr: (out) dynptr for struct fsverity_digest
> + *
> + * Read fsverity_digest of *file* into *digest_ptr*.
> + *
> + * Return: 0 on success, a negative value on error.
> + */
> +__bpf_kfunc int bpf_get_fsverity_digest(struct file *file, struct bpf_dynptr_kern *digest_ptr)
> +{
> +	const struct inode *inode = file_inode(file);
> +	struct fsverity_digest *arg = digest_ptr->data;

What alignment is guaranteed here?

> +	const struct fsverity_info *vi;
> +	const struct fsverity_hash_alg *hash_alg;
> +	int out_digest_sz;
> +
> +	if (__bpf_dynptr_size(digest_ptr) < sizeof(struct fsverity_digest))
> +		return -EINVAL;
> +
> +	vi = fsverity_get_info(inode);
> +	if (!vi)
> +		return -ENODATA; /* not a verity file */
> +
> +	hash_alg = vi->tree_params.hash_alg;
> +
> +	arg->digest_algorithm = hash_alg - fsverity_hash_algs;
> +	arg->digest_size = hash_alg->digest_size;
> +
> +	out_digest_sz = __bpf_dynptr_size(digest_ptr) - sizeof(struct fsverity_digest);
> +
> +	/* copy digest */
> +	memcpy(arg->digest, vi->file_digest,  min_t(int, hash_alg->digest_size, out_digest_sz));
> +
> +	/* fill the extra buffer with zeros */
> +	memset(arg->digest + arg->digest_size, 0, out_digest_sz - hash_alg->digest_size);

Can't 'out_digest_sz - hash_alg->digest_size' underflow?

> +
> +	return 0;
> +}
> +
> +__diag_pop();
> +
> +BTF_SET8_START(fsverity_set)
> +BTF_ID_FLAGS(func, bpf_get_fsverity_digest, KF_SLEEPABLE)

Should it be sleepable?  Nothing in it sleeps, as far as I can tell.

> +BTF_SET8_END(fsverity_set)
> +
> +const struct btf_kfunc_id_set bpf_fsverity_set = {
> +	.owner = THIS_MODULE,
> +	.set = &fsverity_set,
> +};

static const?

> +
> +static int __init bpf_fsverity_init(void)
> +{
> +	return register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING,
> +					 &bpf_fsverity_set);
> +}
> +
> +late_initcall(bpf_fsverity_init);

Maybe this should be called by the existing fsverity_init() initcall instead of
having a brand new initcall just for this.

Also, doesn't this all need to be guarded by a kconfig such as CONFIG_BPF?

Also, it looks like I'm being signed up to maintain this.  This isn't a stable
UAPI, right?  No need to document this in Documentation/?

- Eric

Re: [PATCH bpf-next 2/5] bpf, fsverity: Add kfunc bpf_get_fsverity_digest

[Song Liu]

On Sun, Oct 15, 2023 at 12:07 AM Eric Biggers <ebiggers@kernel.org> wrote:
>
[...]
> > + */
> > +__bpf_kfunc int bpf_get_fsverity_digest(struct file *file, struct bpf_dynptr_kern *digest_ptr)
> > +{
> > +     const struct inode *inode = file_inode(file);
> > +     struct fsverity_digest *arg = digest_ptr->data;
>
> What alignment is guaranteed here?

drnptr doesn't not provide alignment guarantee for digest_ptr->data.
If we need alignment guarantee, we need to add it here.

>
> > +     const struct fsverity_info *vi;
> > +     const struct fsverity_hash_alg *hash_alg;
> > +     int out_digest_sz;
> > +
> > +     if (__bpf_dynptr_size(digest_ptr) < sizeof(struct fsverity_digest))
> > +             return -EINVAL;
> > +
> > +     vi = fsverity_get_info(inode);
> > +     if (!vi)
> > +             return -ENODATA; /* not a verity file */
> > +
> > +     hash_alg = vi->tree_params.hash_alg;
> > +
> > +     arg->digest_algorithm = hash_alg - fsverity_hash_algs;
> > +     arg->digest_size = hash_alg->digest_size;
> > +
> > +     out_digest_sz = __bpf_dynptr_size(digest_ptr) - sizeof(struct fsverity_digest);
> > +
> > +     /* copy digest */
> > +     memcpy(arg->digest, vi->file_digest,  min_t(int, hash_alg->digest_size, out_digest_sz));
> > +
> > +     /* fill the extra buffer with zeros */
> > +     memset(arg->digest + arg->digest_size, 0, out_digest_sz - hash_alg->digest_size);
>
> Can't 'out_digest_sz - hash_alg->digest_size' underflow?

Will fix this in the next version.

>
> > +
> > +     return 0;
> > +}
> > +
> > +__diag_pop();
> > +
> > +BTF_SET8_START(fsverity_set)
> > +BTF_ID_FLAGS(func, bpf_get_fsverity_digest, KF_SLEEPABLE)
>
> Should it be sleepable?  Nothing in it sleeps, as far as I can tell.

Indeed, we can remove sleepable requirement here.

>
> > +BTF_SET8_END(fsverity_set)
> > +
> > +const struct btf_kfunc_id_set bpf_fsverity_set = {
> > +     .owner = THIS_MODULE,
> > +     .set = &fsverity_set,
> > +};
>
> static const?

Will fix in v2.

>
> > +
> > +static int __init bpf_fsverity_init(void)
> > +{
> > +     return register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING,
> > +                                      &bpf_fsverity_set);
> > +}
> > +
> > +late_initcall(bpf_fsverity_init);
>
> Maybe this should be called by the existing fsverity_init() initcall instead of
> having a brand new initcall just for this.

Yeah, that would also work.

>
> Also, doesn't this all need to be guarded by a kconfig such as CONFIG_BPF?

Will add this in v2.

>
> Also, it looks like I'm being signed up to maintain this.  This isn't a stable
> UAPI, right?  No need to document this in Documentation/?

BPF kfuncs are not UAPI. They are as stable as exported symbols.
We do have some documents for BPF kfuncs, for example in
Documentation/bpf/cpumasks.rst.

Do you have a recommendation or preference on where we should
document this? AFAICT, we can either add it to fsverity.rst or somewhere
in Documentation/bpf/.

Thanks,
Song

Re: [PATCH bpf-next 2/5] bpf, fsverity: Add kfunc bpf_get_fsverity_digest

[Eric Biggers]

On Mon, Oct 16, 2023 at 01:10:40PM -0700, Song Liu wrote:
> On Sun, Oct 15, 2023 at 12:07 AM Eric Biggers <ebiggers@kernel.org> wrote:
> >
> [...]
> > > + */
> > > +__bpf_kfunc int bpf_get_fsverity_digest(struct file *file, struct bpf_dynptr_kern *digest_ptr)
> > > +{
> > > +     const struct inode *inode = file_inode(file);
> > > +     struct fsverity_digest *arg = digest_ptr->data;
> >
> > What alignment is guaranteed here?
> 
> drnptr doesn't not provide alignment guarantee for digest_ptr->data.
> If we need alignment guarantee, we need to add it here.

So technically it's wrong to cast it to struct fsverity_digest, then.

> >
> > Also, it looks like I'm being signed up to maintain this.  This isn't a stable
> > UAPI, right?  No need to document this in Documentation/?
> 
> BPF kfuncs are not UAPI. They are as stable as exported symbols.
> We do have some documents for BPF kfuncs, for example in
> Documentation/bpf/cpumasks.rst.
> 
> Do you have a recommendation or preference on where we should
> document this? AFAICT, we can either add it to fsverity.rst or somewhere
> in Documentation/bpf/.

The BPF documentation seems like the right place.

- Eric

Re: [PATCH bpf-next 2/5] bpf, fsverity: Add kfunc bpf_get_fsverity_digest

[Song Liu]

On Mon, Oct 16, 2023 at 8:12 PM Eric Biggers <ebiggers@kernel.org> wrote:
>
> On Mon, Oct 16, 2023 at 01:10:40PM -0700, Song Liu wrote:
> > On Sun, Oct 15, 2023 at 12:07 AM Eric Biggers <ebiggers@kernel.org> wrote:
> > >
> > [...]
> > > > + */
> > > > +__bpf_kfunc int bpf_get_fsverity_digest(struct file *file, struct bpf_dynptr_kern *digest_ptr)
> > > > +{
> > > > +     const struct inode *inode = file_inode(file);
> > > > +     struct fsverity_digest *arg = digest_ptr->data;
> > >
> > > What alignment is guaranteed here?
> >
> > drnptr doesn't not provide alignment guarantee for digest_ptr->data.
> > If we need alignment guarantee, we need to add it here.
>
> So technically it's wrong to cast it to struct fsverity_digest, then.

We can enforce alignment here. Would __aligned(2) be sufficient?

Thanks,
Song

Re: [PATCH bpf-next 2/5] bpf, fsverity: Add kfunc bpf_get_fsverity_digest

[Eric Biggers]

On Mon, Oct 16, 2023 at 10:35:16PM -0700, Song Liu wrote:
> On Mon, Oct 16, 2023 at 8:12 PM Eric Biggers <ebiggers@kernel.org> wrote:
> >
> > On Mon, Oct 16, 2023 at 01:10:40PM -0700, Song Liu wrote:
> > > On Sun, Oct 15, 2023 at 12:07 AM Eric Biggers <ebiggers@kernel.org> wrote:
> > > >
> > > [...]
> > > > > + */
> > > > > +__bpf_kfunc int bpf_get_fsverity_digest(struct file *file, struct bpf_dynptr_kern *digest_ptr)
> > > > > +{
> > > > > +     const struct inode *inode = file_inode(file);
> > > > > +     struct fsverity_digest *arg = digest_ptr->data;
> > > >
> > > > What alignment is guaranteed here?
> > >
> > > drnptr doesn't not provide alignment guarantee for digest_ptr->data.
> > > If we need alignment guarantee, we need to add it here.
> >
> > So technically it's wrong to cast it to struct fsverity_digest, then.
> 
> We can enforce alignment here. Would __aligned(2) be sufficient?
> 

Do you mean something like the following:

	if (!IS_ALIGNED((uintptr_t)digest_ptr->data, __alignof__(*arg)))
		return -EINVAL;

Re: [PATCH bpf-next 2/5] bpf, fsverity: Add kfunc bpf_get_fsverity_digest

[Song Liu]

> On Oct 16, 2023, at 10:46 PM, Eric Biggers <ebiggers@kernel.org> wrote:
> 
> On Mon, Oct 16, 2023 at 10:35:16PM -0700, Song Liu wrote:
>> On Mon, Oct 16, 2023 at 8:12 PM Eric Biggers <ebiggers@kernel.org> wrote:
>>> 
>>> On Mon, Oct 16, 2023 at 01:10:40PM -0700, Song Liu wrote:
>>>> On Sun, Oct 15, 2023 at 12:07 AM Eric Biggers <ebiggers@kernel.org> wrote:
>>>>> 
>>>> [...]
>>>>>> + */
>>>>>> +__bpf_kfunc int bpf_get_fsverity_digest(struct file *file, struct bpf_dynptr_kern *digest_ptr)
>>>>>> +{
>>>>>> +     const struct inode *inode = file_inode(file);
>>>>>> +     struct fsverity_digest *arg = digest_ptr->data;
>>>>> 
>>>>> What alignment is guaranteed here?
>>>> 
>>>> drnptr doesn't not provide alignment guarantee for digest_ptr->data.
>>>> If we need alignment guarantee, we need to add it here.
>>> 
>>> So technically it's wrong to cast it to struct fsverity_digest, then.
>> 
>> We can enforce alignment here. Would __aligned(2) be sufficient?
>> 
> 
> Do you mean something like the following:
> 
> if (!IS_ALIGNED((uintptr_t)digest_ptr->data, __alignof__(*arg)))
> return -EINVAL;

I was thinking about hard-coding the alignment requirement. 
__alignof__ is much better. Thanks for the suggestion!

Song

Re: [PATCH bpf-next 2/5] bpf, fsverity: Add kfunc bpf_get_fsverity_digest

[Andrii Nakryiko]

On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
>
> The kfunc can be used to read fsverity_digest, so that we can verify
> signature in BPF LSM.
>
> This kfunc is added to fs/verity/measure.c because some data structure used
> in the function is private to fsverity (fs/verity/fsverity_private.h).
>
> Signed-off-by: Song Liu <song@kernel.org>
> ---
>  fs/verity/measure.c | 66 +++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 66 insertions(+)
>
> diff --git a/fs/verity/measure.c b/fs/verity/measure.c
> index eec5956141da..2d4b2e6f5a5d 100644
> --- a/fs/verity/measure.c
> +++ b/fs/verity/measure.c
> @@ -8,6 +8,8 @@
>  #include "fsverity_private.h"
>
>  #include <linux/uaccess.h>
> +#include <linux/bpf.h>
> +#include <linux/btf.h>
>
>  /**
>   * fsverity_ioctl_measure() - get a verity file's digest
> @@ -100,3 +102,67 @@ int fsverity_get_digest(struct inode *inode,
>         return hash_alg->digest_size;
>  }
>  EXPORT_SYMBOL_GPL(fsverity_get_digest);
> +
> +/* bpf kfuncs */
> +__diag_push();
> +__diag_ignore_all("-Wmissing-prototypes",
> +                 "kfuncs which will be used in BPF programs");
> +
> +/**
> + * bpf_get_fsverity_digest: read fsverity digest of file
> + * @file: file to get digest from
> + * @digest_ptr: (out) dynptr for struct fsverity_digest
> + *
> + * Read fsverity_digest of *file* into *digest_ptr*.
> + *
> + * Return: 0 on success, a negative value on error.
> + */
> +__bpf_kfunc int bpf_get_fsverity_digest(struct file *file, struct bpf_dynptr_kern *digest_ptr)
> +{
> +       const struct inode *inode = file_inode(file);
> +       struct fsverity_digest *arg = digest_ptr->data;

this can be null

I think we need some internal helpers that are similar to
bpf_dynptr_slice() that would handle invalid dynptr cases, as well as
abstract away potentially non-contiguous memory dynptr points to.
WDYT?

> +       const struct fsverity_info *vi;
> +       const struct fsverity_hash_alg *hash_alg;
> +       int out_digest_sz;
> +
> +       if (__bpf_dynptr_size(digest_ptr) < sizeof(struct fsverity_digest))
> +               return -EINVAL;
> +
> +       vi = fsverity_get_info(inode);
> +       if (!vi)
> +               return -ENODATA; /* not a verity file */
> +
> +       hash_alg = vi->tree_params.hash_alg;
> +
> +       arg->digest_algorithm = hash_alg - fsverity_hash_algs;
> +       arg->digest_size = hash_alg->digest_size;
> +
> +       out_digest_sz = __bpf_dynptr_size(digest_ptr) - sizeof(struct fsverity_digest);
> +
> +       /* copy digest */
> +       memcpy(arg->digest, vi->file_digest,  min_t(int, hash_alg->digest_size, out_digest_sz));
> +
> +       /* fill the extra buffer with zeros */
> +       memset(arg->digest + arg->digest_size, 0, out_digest_sz - hash_alg->digest_size);
> +
> +       return 0;
> +}
> +
> +__diag_pop();
> +
> +BTF_SET8_START(fsverity_set)
> +BTF_ID_FLAGS(func, bpf_get_fsverity_digest, KF_SLEEPABLE)
> +BTF_SET8_END(fsverity_set)
> +
> +const struct btf_kfunc_id_set bpf_fsverity_set = {
> +       .owner = THIS_MODULE,
> +       .set = &fsverity_set,
> +};
> +
> +static int __init bpf_fsverity_init(void)
> +{
> +       return register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING,
> +                                        &bpf_fsverity_set);
> +}
> +
> +late_initcall(bpf_fsverity_init);
> --
> 2.34.1
>

Re: [PATCH bpf-next 5/5] selftests/bpf: Add test that use fsverity and xattr to sign a file

[Alexei Starovoitov]

On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
>
> +#define MAGIC_SIZE 8
> +char digest[MAGIC_SIZE + sizeof(struct fsverity_digest) + SHA256_DIGEST_SIZE];
> +
> +__u32 monitored_pid;
> +char xattr_name[] = "user.sig";
> +char sig[MAX_SIG_SIZE];
> +__u32 sig_size;
> +__u32 user_keyring_serial;
> +
> +SEC("lsm.s/file_open")
> +int BPF_PROG(test_file_open, struct file *f)
> +{
> +       struct bpf_dynptr digest_ptr, sig_ptr, name_ptr;
> +       struct bpf_key *trusted_keyring;
> +       __u32 pid;
> +       int ret;
> +
> +       pid = bpf_get_current_pid_tgid() >> 32;
> +       if (pid != monitored_pid)
> +               return 0;
> +
> +       /* digest_ptr points to fsverity_digest */
> +       bpf_dynptr_from_mem(digest + MAGIC_SIZE, sizeof(digest) - MAGIC_SIZE, 0, &digest_ptr);
> +
> +       ret = bpf_get_fsverity_digest(f, &digest_ptr);
> +       /* No verity, allow access */
> +       if (ret < 0)
> +               return 0;
> +
> +       /* Move digest_ptr to fsverity_formatted_digest */
> +       bpf_dynptr_from_mem(digest, sizeof(digest), 0, &digest_ptr);
> +
> +       /* Read signature from xattr */
> +       bpf_dynptr_from_mem(sig, sizeof(sig), 0, &sig_ptr);
> +       bpf_dynptr_from_mem(xattr_name, sizeof(xattr_name), 0, &name_ptr);
> +       ret = bpf_get_file_xattr(f, &name_ptr, &sig_ptr);
> +       /* No signature, reject access */
> +       if (ret < 0)
> +               return -EPERM;
> +
> +       trusted_keyring = bpf_lookup_user_key(user_keyring_serial, 0);
> +       if (!trusted_keyring)
> +               return -ENOENT;
> +
> +       /* Verify signature */
> +       ret = bpf_verify_pkcs7_signature(&digest_ptr, &sig_ptr, trusted_keyring);
> +
> +       bpf_key_put(trusted_keyring);
> +       return ret;
> +}

I think the UX is cumbersome.
Putting a NULL terminated string into dynptr not only a source
code ugliness, but it adds run-time overhead too.
We better add proper C style string support in the verifier,
so that bpf_get_file_xattr() can look like normal C.

Re: [PATCH bpf-next 5/5] selftests/bpf: Add test that use fsverity and xattr to sign a file

[Song Liu]

> On Oct 17, 2023, at 12:08 PM, Alexei Starovoitov <alexei.starovoitov@gmail.com> wrote:
> 
> On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
>> 
>> +#define MAGIC_SIZE 8
>> +char digest[MAGIC_SIZE + sizeof(struct fsverity_digest) + SHA256_DIGEST_SIZE];
>> +
>> +__u32 monitored_pid;
>> +char xattr_name[] = "user.sig";
>> +char sig[MAX_SIG_SIZE];
>> +__u32 sig_size;
>> +__u32 user_keyring_serial;
>> +
>> +SEC("lsm.s/file_open")
>> +int BPF_PROG(test_file_open, struct file *f)
>> +{
>> +       struct bpf_dynptr digest_ptr, sig_ptr, name_ptr;
>> +       struct bpf_key *trusted_keyring;
>> +       __u32 pid;
>> +       int ret;
>> +
>> +       pid = bpf_get_current_pid_tgid() >> 32;
>> +       if (pid != monitored_pid)
>> +               return 0;
>> +
>> +       /* digest_ptr points to fsverity_digest */
>> +       bpf_dynptr_from_mem(digest + MAGIC_SIZE, sizeof(digest) - MAGIC_SIZE, 0, &digest_ptr);
>> +
>> +       ret = bpf_get_fsverity_digest(f, &digest_ptr);
>> +       /* No verity, allow access */
>> +       if (ret < 0)
>> +               return 0;
>> +
>> +       /* Move digest_ptr to fsverity_formatted_digest */
>> +       bpf_dynptr_from_mem(digest, sizeof(digest), 0, &digest_ptr);
>> +
>> +       /* Read signature from xattr */
>> +       bpf_dynptr_from_mem(sig, sizeof(sig), 0, &sig_ptr);
>> +       bpf_dynptr_from_mem(xattr_name, sizeof(xattr_name), 0, &name_ptr);
>> +       ret = bpf_get_file_xattr(f, &name_ptr, &sig_ptr);
>> +       /* No signature, reject access */
>> +       if (ret < 0)
>> +               return -EPERM;
>> +
>> +       trusted_keyring = bpf_lookup_user_key(user_keyring_serial, 0);
>> +       if (!trusted_keyring)
>> +               return -ENOENT;
>> +
>> +       /* Verify signature */
>> +       ret = bpf_verify_pkcs7_signature(&digest_ptr, &sig_ptr, trusted_keyring);
>> +
>> +       bpf_key_put(trusted_keyring);
>> +       return ret;
>> +}
> 
> I think the UX is cumbersome.
> Putting a NULL terminated string into dynptr not only a source
> code ugliness, but it adds run-time overhead too.
> We better add proper C style string support in the verifier,
> so that bpf_get_file_xattr() can look like normal C.

That will indeed look much better. Let me check what do we need 
make this happen.

Thanks,
Song

Re: [PATCH bpf-next 1/5] bpf: Add kfunc bpf_get_file_xattr

[Andrii Nakryiko]

On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
>
> This kfunc can be used to read xattr of a file.
>
> Since vfs_getxattr() requires null-terminated string as input "name", a new
> helper bpf_dynptr_is_string() is added to check the input before calling
> vfs_getxattr().
>
> Signed-off-by: Song Liu <song@kernel.org>
> ---
>  include/linux/bpf.h      | 12 +++++++++++
>  kernel/trace/bpf_trace.c | 44 ++++++++++++++++++++++++++++++++++++++++
>  2 files changed, 56 insertions(+)
>
> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> index 61bde4520f5c..f14fae45e13d 100644
> --- a/include/linux/bpf.h
> +++ b/include/linux/bpf.h
> @@ -2472,6 +2472,13 @@ static inline bool has_current_bpf_ctx(void)
>         return !!current->bpf_ctx;
>  }
>
> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)

is_zero_terminated would be more accurate? though there is nothing
really dynptr-specific here...

> +{
> +       char *str = ptr->data;
> +
> +       return str[__bpf_dynptr_size(ptr) - 1] == '\0';
> +}
> +
>  void notrace bpf_prog_inc_misses_counter(struct bpf_prog *prog);
>
>  void bpf_dynptr_init(struct bpf_dynptr_kern *ptr, void *data,
> @@ -2708,6 +2715,11 @@ static inline bool has_current_bpf_ctx(void)
>         return false;
>  }
>
> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
> +{
> +       return false;
> +}
> +
>  static inline void bpf_prog_inc_misses_counter(struct bpf_prog *prog)
>  {
>  }
> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> index df697c74d519..946268574e05 100644
> --- a/kernel/trace/bpf_trace.c
> +++ b/kernel/trace/bpf_trace.c
> @@ -24,6 +24,7 @@
>  #include <linux/key.h>
>  #include <linux/verification.h>
>  #include <linux/namei.h>
> +#include <linux/fileattr.h>
>
>  #include <net/bpf_sk_storage.h>
>
> @@ -1429,6 +1430,49 @@ static int __init bpf_key_sig_kfuncs_init(void)
>  late_initcall(bpf_key_sig_kfuncs_init);
>  #endif /* CONFIG_KEYS */
>
> +/* filesystem kfuncs */
> +__diag_push();
> +__diag_ignore_all("-Wmissing-prototypes",
> +                 "kfuncs which will be used in BPF programs");
> +
> +/**
> + * bpf_get_file_xattr - get xattr of a file
> + * @name_ptr: name of the xattr
> + * @value_ptr: output buffer of the xattr value
> + *
> + * Get xattr *name_ptr* of *file* and store the output in *value_ptr*.
> + *
> + * Return: 0 on success, a negative value on error.
> + */
> +__bpf_kfunc int bpf_get_file_xattr(struct file *file, struct bpf_dynptr_kern *name_ptr,
> +                                  struct bpf_dynptr_kern *value_ptr)
> +{
> +       if (!bpf_dynptr_is_string(name_ptr))
> +               return -EINVAL;

so dynptr can be invalid and name_ptr->data will be NULL, you should
account for that

and there could also be special dynptrs that don't have contiguous
memory region, so somehow you'd need to take care of that as well

> +
> +       return vfs_getxattr(mnt_idmap(file->f_path.mnt), file_dentry(file), name_ptr->data,
> +                           value_ptr->data, __bpf_dynptr_size(value_ptr));
> +}
> +
> +__diag_pop();
> +
> +BTF_SET8_START(fs_kfunc_set)
> +BTF_ID_FLAGS(func, bpf_get_file_xattr, KF_SLEEPABLE)
> +BTF_SET8_END(fs_kfunc_set)
> +
> +const struct btf_kfunc_id_set bpf_fs_kfunc_set = {
> +       .owner = THIS_MODULE,
> +       .set = &fs_kfunc_set,
> +};
> +
> +static int __init bpf_fs_kfuncs_init(void)
> +{
> +       return register_btf_kfunc_id_set(BPF_PROG_TYPE_TRACING,
> +                                        &bpf_fs_kfunc_set);
> +}
> +
> +late_initcall(bpf_fs_kfuncs_init);
> +
>  static const struct bpf_func_proto *
>  bpf_tracing_func_proto(enum bpf_func_id func_id, const struct bpf_prog *prog)
>  {
> --
> 2.34.1
>

Re: [PATCH bpf-next 1/5] bpf: Add kfunc bpf_get_file_xattr

[Song Liu]

> On Oct 17, 2023, at 11:58 AM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> 
> On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
>> 
>> This kfunc can be used to read xattr of a file.
>> 
>> Since vfs_getxattr() requires null-terminated string as input "name", a new
>> helper bpf_dynptr_is_string() is added to check the input before calling
>> vfs_getxattr().
>> 
>> Signed-off-by: Song Liu <song@kernel.org>
>> ---
>> include/linux/bpf.h      | 12 +++++++++++
>> kernel/trace/bpf_trace.c | 44 ++++++++++++++++++++++++++++++++++++++++
>> 2 files changed, 56 insertions(+)
>> 
>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>> index 61bde4520f5c..f14fae45e13d 100644
>> --- a/include/linux/bpf.h
>> +++ b/include/linux/bpf.h
>> @@ -2472,6 +2472,13 @@ static inline bool has_current_bpf_ctx(void)
>>        return !!current->bpf_ctx;
>> }
>> 
>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
> 
> is_zero_terminated would be more accurate? though there is nothing
> really dynptr-specific here...

is_zero_terminated sounds better. 

> 
>> +{
>> +       char *str = ptr->data;
>> +
>> +       return str[__bpf_dynptr_size(ptr) - 1] == '\0';
>> +}
>> +
>> void notrace bpf_prog_inc_misses_counter(struct bpf_prog *prog);
>> 
>> void bpf_dynptr_init(struct bpf_dynptr_kern *ptr, void *data,
>> @@ -2708,6 +2715,11 @@ static inline bool has_current_bpf_ctx(void)
>>        return false;
>> }
>> 
>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
>> +{
>> +       return false;
>> +}
>> +
>> static inline void bpf_prog_inc_misses_counter(struct bpf_prog *prog)
>> {
>> }
>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>> index df697c74d519..946268574e05 100644
>> --- a/kernel/trace/bpf_trace.c
>> +++ b/kernel/trace/bpf_trace.c
>> @@ -24,6 +24,7 @@
>> #include <linux/key.h>
>> #include <linux/verification.h>
>> #include <linux/namei.h>
>> +#include <linux/fileattr.h>
>> 
>> #include <net/bpf_sk_storage.h>
>> 
>> @@ -1429,6 +1430,49 @@ static int __init bpf_key_sig_kfuncs_init(void)
>> late_initcall(bpf_key_sig_kfuncs_init);
>> #endif /* CONFIG_KEYS */
>> 
>> +/* filesystem kfuncs */
>> +__diag_push();
>> +__diag_ignore_all("-Wmissing-prototypes",
>> +                 "kfuncs which will be used in BPF programs");
>> +
>> +/**
>> + * bpf_get_file_xattr - get xattr of a file
>> + * @name_ptr: name of the xattr
>> + * @value_ptr: output buffer of the xattr value
>> + *
>> + * Get xattr *name_ptr* of *file* and store the output in *value_ptr*.
>> + *
>> + * Return: 0 on success, a negative value on error.
>> + */
>> +__bpf_kfunc int bpf_get_file_xattr(struct file *file, struct bpf_dynptr_kern *name_ptr,
>> +                                  struct bpf_dynptr_kern *value_ptr)
>> +{
>> +       if (!bpf_dynptr_is_string(name_ptr))
>> +               return -EINVAL;
> 
> so dynptr can be invalid and name_ptr->data will be NULL, you should
> account for that

We can add a NULL check (or size check) here. 

> 
> and there could also be special dynptrs that don't have contiguous
> memory region, so somehow you'd need to take care of that as well

We can require the dynptr to be BPF_DYNPTR_TYPE_LOCAL. I don't think
we need this for dynptr of skb or xdp. Would this be sufficient?

Thanks,
Song

> 
>> +
>> +       return vfs_getxattr(mnt_idmap(file->f_path.mnt), file_dentry(file), name_ptr->data,
>> +                           value_ptr->data, __bpf_dynptr_size(value_ptr));
>> +}

Re: [PATCH bpf-next 1/5] bpf: Add kfunc bpf_get_file_xattr

[Andrii Nakryiko]

On Tue, Oct 17, 2023 at 1:31 PM Song Liu <songliubraving@meta.com> wrote:
>
>
>
> > On Oct 17, 2023, at 11:58 AM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> >
> > On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
> >>
> >> This kfunc can be used to read xattr of a file.
> >>
> >> Since vfs_getxattr() requires null-terminated string as input "name", a new
> >> helper bpf_dynptr_is_string() is added to check the input before calling
> >> vfs_getxattr().
> >>
> >> Signed-off-by: Song Liu <song@kernel.org>
> >> ---
> >> include/linux/bpf.h      | 12 +++++++++++
> >> kernel/trace/bpf_trace.c | 44 ++++++++++++++++++++++++++++++++++++++++
> >> 2 files changed, 56 insertions(+)
> >>
> >> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> >> index 61bde4520f5c..f14fae45e13d 100644
> >> --- a/include/linux/bpf.h
> >> +++ b/include/linux/bpf.h
> >> @@ -2472,6 +2472,13 @@ static inline bool has_current_bpf_ctx(void)
> >>        return !!current->bpf_ctx;
> >> }
> >>
> >> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
> >
> > is_zero_terminated would be more accurate? though there is nothing
> > really dynptr-specific here...
>
> is_zero_terminated sounds better.
>
> >
> >> +{
> >> +       char *str = ptr->data;
> >> +
> >> +       return str[__bpf_dynptr_size(ptr) - 1] == '\0';
> >> +}
> >> +
> >> void notrace bpf_prog_inc_misses_counter(struct bpf_prog *prog);
> >>
> >> void bpf_dynptr_init(struct bpf_dynptr_kern *ptr, void *data,
> >> @@ -2708,6 +2715,11 @@ static inline bool has_current_bpf_ctx(void)
> >>        return false;
> >> }
> >>
> >> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
> >> +{
> >> +       return false;
> >> +}
> >> +
> >> static inline void bpf_prog_inc_misses_counter(struct bpf_prog *prog)
> >> {
> >> }
> >> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> >> index df697c74d519..946268574e05 100644
> >> --- a/kernel/trace/bpf_trace.c
> >> +++ b/kernel/trace/bpf_trace.c
> >> @@ -24,6 +24,7 @@
> >> #include <linux/key.h>
> >> #include <linux/verification.h>
> >> #include <linux/namei.h>
> >> +#include <linux/fileattr.h>
> >>
> >> #include <net/bpf_sk_storage.h>
> >>
> >> @@ -1429,6 +1430,49 @@ static int __init bpf_key_sig_kfuncs_init(void)
> >> late_initcall(bpf_key_sig_kfuncs_init);
> >> #endif /* CONFIG_KEYS */
> >>
> >> +/* filesystem kfuncs */
> >> +__diag_push();
> >> +__diag_ignore_all("-Wmissing-prototypes",
> >> +                 "kfuncs which will be used in BPF programs");
> >> +
> >> +/**
> >> + * bpf_get_file_xattr - get xattr of a file
> >> + * @name_ptr: name of the xattr
> >> + * @value_ptr: output buffer of the xattr value
> >> + *
> >> + * Get xattr *name_ptr* of *file* and store the output in *value_ptr*.
> >> + *
> >> + * Return: 0 on success, a negative value on error.
> >> + */
> >> +__bpf_kfunc int bpf_get_file_xattr(struct file *file, struct bpf_dynptr_kern *name_ptr,
> >> +                                  struct bpf_dynptr_kern *value_ptr)
> >> +{
> >> +       if (!bpf_dynptr_is_string(name_ptr))
> >> +               return -EINVAL;
> >
> > so dynptr can be invalid and name_ptr->data will be NULL, you should
> > account for that
>
> We can add a NULL check (or size check) here.

there must be some helper to check if dynptr is valid, let's use that
instead of NULL checks

>
> >
> > and there could also be special dynptrs that don't have contiguous
> > memory region, so somehow you'd need to take care of that as well
>
> We can require the dynptr to be BPF_DYNPTR_TYPE_LOCAL. I don't think
> we need this for dynptr of skb or xdp. Would this be sufficient?

well, to keep thing simple we can have a simple internal helper API
that will tell if it's safe to assume that dynptr memory is contiguous
and it's ok to use dynptr memory. But still, you shouldn't access data
pointer directly, there must be some helper for that. Please check. It
has to take into account offset and stuff like that.


Also, and separately from that, we should think about providing a
bpf_dynptr_slice()-like helper that will accept a fixed-sized
temporary buffer and return pointer to either actual memory or copy
non-contiguous memory into that buffer. That will make sure you can
use any dynptr as a source of data, and only pay the price of memory
copy in rare cases where it's necessary

>
> Thanks,
> Song
>
> >
> >> +
> >> +       return vfs_getxattr(mnt_idmap(file->f_path.mnt), file_dentry(file), name_ptr->data,
> >> +                           value_ptr->data, __bpf_dynptr_size(value_ptr));
> >> +}
>

Re: [PATCH bpf-next 1/5] bpf: Add kfunc bpf_get_file_xattr

[Song Liu]

> On Oct 17, 2023, at 2:52 PM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> 
> On Tue, Oct 17, 2023 at 1:31 PM Song Liu <songliubraving@meta.com> wrote:
>> 
>> 
>> 
>>> On Oct 17, 2023, at 11:58 AM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
>>> 
>>> On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
>>>> 
>>>> This kfunc can be used to read xattr of a file.
>>>> 
>>>> Since vfs_getxattr() requires null-terminated string as input "name", a new
>>>> helper bpf_dynptr_is_string() is added to check the input before calling
>>>> vfs_getxattr().
>>>> 
>>>> Signed-off-by: Song Liu <song@kernel.org>
>>>> ---
>>>> include/linux/bpf.h      | 12 +++++++++++
>>>> kernel/trace/bpf_trace.c | 44 ++++++++++++++++++++++++++++++++++++++++
>>>> 2 files changed, 56 insertions(+)
>>>> 
>>>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>>>> index 61bde4520f5c..f14fae45e13d 100644
>>>> --- a/include/linux/bpf.h
>>>> +++ b/include/linux/bpf.h
>>>> @@ -2472,6 +2472,13 @@ static inline bool has_current_bpf_ctx(void)
>>>>       return !!current->bpf_ctx;
>>>> }
>>>> 
>>>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
>>> 
>>> is_zero_terminated would be more accurate? though there is nothing
>>> really dynptr-specific here...
>> 
>> is_zero_terminated sounds better.
>> 
>>> 
>>>> +{
>>>> +       char *str = ptr->data;
>>>> +
>>>> +       return str[__bpf_dynptr_size(ptr) - 1] == '\0';
>>>> +}
>>>> +
>>>> void notrace bpf_prog_inc_misses_counter(struct bpf_prog *prog);
>>>> 
>>>> void bpf_dynptr_init(struct bpf_dynptr_kern *ptr, void *data,
>>>> @@ -2708,6 +2715,11 @@ static inline bool has_current_bpf_ctx(void)
>>>>       return false;
>>>> }
>>>> 
>>>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
>>>> +{
>>>> +       return false;
>>>> +}
>>>> +
>>>> static inline void bpf_prog_inc_misses_counter(struct bpf_prog *prog)
>>>> {
>>>> }
>>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>>> index df697c74d519..946268574e05 100644
>>>> --- a/kernel/trace/bpf_trace.c
>>>> +++ b/kernel/trace/bpf_trace.c
>>>> @@ -24,6 +24,7 @@
>>>> #include <linux/key.h>
>>>> #include <linux/verification.h>
>>>> #include <linux/namei.h>
>>>> +#include <linux/fileattr.h>
>>>> 
>>>> #include <net/bpf_sk_storage.h>
>>>> 
>>>> @@ -1429,6 +1430,49 @@ static int __init bpf_key_sig_kfuncs_init(void)
>>>> late_initcall(bpf_key_sig_kfuncs_init);
>>>> #endif /* CONFIG_KEYS */
>>>> 
>>>> +/* filesystem kfuncs */
>>>> +__diag_push();
>>>> +__diag_ignore_all("-Wmissing-prototypes",
>>>> +                 "kfuncs which will be used in BPF programs");
>>>> +
>>>> +/**
>>>> + * bpf_get_file_xattr - get xattr of a file
>>>> + * @name_ptr: name of the xattr
>>>> + * @value_ptr: output buffer of the xattr value
>>>> + *
>>>> + * Get xattr *name_ptr* of *file* and store the output in *value_ptr*.
>>>> + *
>>>> + * Return: 0 on success, a negative value on error.
>>>> + */
>>>> +__bpf_kfunc int bpf_get_file_xattr(struct file *file, struct bpf_dynptr_kern *name_ptr,
>>>> +                                  struct bpf_dynptr_kern *value_ptr)
>>>> +{
>>>> +       if (!bpf_dynptr_is_string(name_ptr))
>>>> +               return -EINVAL;
>>> 
>>> so dynptr can be invalid and name_ptr->data will be NULL, you should
>>> account for that
>> 
>> We can add a NULL check (or size check) here.
> 
> there must be some helper to check if dynptr is valid, let's use that
> instead of NULL checks

Yeah, we can use bpf_dynptr_is_null(). 

> 
>> 
>>> 
>>> and there could also be special dynptrs that don't have contiguous
>>> memory region, so somehow you'd need to take care of that as well
>> 
>> We can require the dynptr to be BPF_DYNPTR_TYPE_LOCAL. I don't think
>> we need this for dynptr of skb or xdp. Would this be sufficient?
> 
> well, to keep thing simple we can have a simple internal helper API
> that will tell if it's safe to assume that dynptr memory is contiguous
> and it's ok to use dynptr memory. But still, you shouldn't access data
> pointer directly, there must be some helper for that. Please check. It
> has to take into account offset and stuff like that.

Yeah, we can use bpf_dynptr_write(), which is a helper (not kfunc). 

> 
> Also, and separately from that, we should think about providing a
> bpf_dynptr_slice()-like helper that will accept a fixed-sized
> temporary buffer and return pointer to either actual memory or copy
> non-contiguous memory into that buffer. That will make sure you can
> use any dynptr as a source of data, and only pay the price of memory
> copy in rare cases where it's necessary

I don't quite follow here. Currently, we have 

bpf_dynptr_data()
bpf_dynptr_slice()
bpf_dynptr_slice_rdwr()
bpf_dynptr_write()

AFAICT, they are sufficient to cover existing use cases (and the new 
use case we are adding in this set). What's the new kfunc are you
thinking about?

Thanks,
Song

Re: [PATCH bpf-next 1/5] bpf: Add kfunc bpf_get_file_xattr

[Andrii Nakryiko]

On Tue, Oct 17, 2023 at 3:16 PM Song Liu <songliubraving@meta.com> wrote:
>
>
>
> > On Oct 17, 2023, at 2:52 PM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> >
> > On Tue, Oct 17, 2023 at 1:31 PM Song Liu <songliubraving@meta.com> wrote:
> >>
> >>
> >>
> >>> On Oct 17, 2023, at 11:58 AM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> >>>
> >>> On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
> >>>>
> >>>> This kfunc can be used to read xattr of a file.
> >>>>
> >>>> Since vfs_getxattr() requires null-terminated string as input "name", a new
> >>>> helper bpf_dynptr_is_string() is added to check the input before calling
> >>>> vfs_getxattr().
> >>>>
> >>>> Signed-off-by: Song Liu <song@kernel.org>
> >>>> ---
> >>>> include/linux/bpf.h      | 12 +++++++++++
> >>>> kernel/trace/bpf_trace.c | 44 ++++++++++++++++++++++++++++++++++++++++
> >>>> 2 files changed, 56 insertions(+)
> >>>>
> >>>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
> >>>> index 61bde4520f5c..f14fae45e13d 100644
> >>>> --- a/include/linux/bpf.h
> >>>> +++ b/include/linux/bpf.h
> >>>> @@ -2472,6 +2472,13 @@ static inline bool has_current_bpf_ctx(void)
> >>>>       return !!current->bpf_ctx;
> >>>> }
> >>>>
> >>>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
> >>>
> >>> is_zero_terminated would be more accurate? though there is nothing
> >>> really dynptr-specific here...
> >>
> >> is_zero_terminated sounds better.
> >>
> >>>
> >>>> +{
> >>>> +       char *str = ptr->data;
> >>>> +
> >>>> +       return str[__bpf_dynptr_size(ptr) - 1] == '\0';
> >>>> +}
> >>>> +
> >>>> void notrace bpf_prog_inc_misses_counter(struct bpf_prog *prog);
> >>>>
> >>>> void bpf_dynptr_init(struct bpf_dynptr_kern *ptr, void *data,
> >>>> @@ -2708,6 +2715,11 @@ static inline bool has_current_bpf_ctx(void)
> >>>>       return false;
> >>>> }
> >>>>
> >>>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
> >>>> +{
> >>>> +       return false;
> >>>> +}
> >>>> +
> >>>> static inline void bpf_prog_inc_misses_counter(struct bpf_prog *prog)
> >>>> {
> >>>> }
> >>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
> >>>> index df697c74d519..946268574e05 100644
> >>>> --- a/kernel/trace/bpf_trace.c
> >>>> +++ b/kernel/trace/bpf_trace.c
> >>>> @@ -24,6 +24,7 @@
> >>>> #include <linux/key.h>
> >>>> #include <linux/verification.h>
> >>>> #include <linux/namei.h>
> >>>> +#include <linux/fileattr.h>
> >>>>
> >>>> #include <net/bpf_sk_storage.h>
> >>>>
> >>>> @@ -1429,6 +1430,49 @@ static int __init bpf_key_sig_kfuncs_init(void)
> >>>> late_initcall(bpf_key_sig_kfuncs_init);
> >>>> #endif /* CONFIG_KEYS */
> >>>>
> >>>> +/* filesystem kfuncs */
> >>>> +__diag_push();
> >>>> +__diag_ignore_all("-Wmissing-prototypes",
> >>>> +                 "kfuncs which will be used in BPF programs");
> >>>> +
> >>>> +/**
> >>>> + * bpf_get_file_xattr - get xattr of a file
> >>>> + * @name_ptr: name of the xattr
> >>>> + * @value_ptr: output buffer of the xattr value
> >>>> + *
> >>>> + * Get xattr *name_ptr* of *file* and store the output in *value_ptr*.
> >>>> + *
> >>>> + * Return: 0 on success, a negative value on error.
> >>>> + */
> >>>> +__bpf_kfunc int bpf_get_file_xattr(struct file *file, struct bpf_dynptr_kern *name_ptr,
> >>>> +                                  struct bpf_dynptr_kern *value_ptr)
> >>>> +{
> >>>> +       if (!bpf_dynptr_is_string(name_ptr))
> >>>> +               return -EINVAL;
> >>>
> >>> so dynptr can be invalid and name_ptr->data will be NULL, you should
> >>> account for that
> >>
> >> We can add a NULL check (or size check) here.
> >
> > there must be some helper to check if dynptr is valid, let's use that
> > instead of NULL checks
>
> Yeah, we can use bpf_dynptr_is_null().
>
> >
> >>
> >>>
> >>> and there could also be special dynptrs that don't have contiguous
> >>> memory region, so somehow you'd need to take care of that as well
> >>
> >> We can require the dynptr to be BPF_DYNPTR_TYPE_LOCAL. I don't think
> >> we need this for dynptr of skb or xdp. Would this be sufficient?
> >
> > well, to keep thing simple we can have a simple internal helper API
> > that will tell if it's safe to assume that dynptr memory is contiguous
> > and it's ok to use dynptr memory. But still, you shouldn't access data
> > pointer directly, there must be some helper for that. Please check. It
> > has to take into account offset and stuff like that.
>
> Yeah, we can use bpf_dynptr_write(), which is a helper (not kfunc).
>
> >
> > Also, and separately from that, we should think about providing a
> > bpf_dynptr_slice()-like helper that will accept a fixed-sized
> > temporary buffer and return pointer to either actual memory or copy
> > non-contiguous memory into that buffer. That will make sure you can
> > use any dynptr as a source of data, and only pay the price of memory
> > copy in rare cases where it's necessary
>
> I don't quite follow here. Currently, we have
>
> bpf_dynptr_data()
> bpf_dynptr_slice()
> bpf_dynptr_slice_rdwr()
> bpf_dynptr_write()
>
> AFAICT, they are sufficient to cover existing use cases (and the new
> use case we are adding in this set). What's the new kfunc are you
> thinking about?

I wasn't talking about kfuncs, but rather just internal helpers to be
used by other kfuncs when working with dynptrs as input arguments.

>
> Thanks,
> Song
>
>

Re: [PATCH bpf-next 1/5] bpf: Add kfunc bpf_get_file_xattr

[Song Liu]

> On Oct 17, 2023, at 3:40 PM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> 
> On Tue, Oct 17, 2023 at 3:16 PM Song Liu <songliubraving@meta.com> wrote:
>> 
>> 
>> 
>>> On Oct 17, 2023, at 2:52 PM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
>>> 
>>> On Tue, Oct 17, 2023 at 1:31 PM Song Liu <songliubraving@meta.com> wrote:
>>>> 
>>>> 
>>>> 
>>>>> On Oct 17, 2023, at 11:58 AM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
>>>>> 
>>>>> On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
>>>>>> 
>>>>>> This kfunc can be used to read xattr of a file.
>>>>>> 
>>>>>> Since vfs_getxattr() requires null-terminated string as input "name", a new
>>>>>> helper bpf_dynptr_is_string() is added to check the input before calling
>>>>>> vfs_getxattr().
>>>>>> 
>>>>>> Signed-off-by: Song Liu <song@kernel.org>
>>>>>> ---
>>>>>> include/linux/bpf.h      | 12 +++++++++++
>>>>>> kernel/trace/bpf_trace.c | 44 ++++++++++++++++++++++++++++++++++++++++
>>>>>> 2 files changed, 56 insertions(+)
>>>>>> 
>>>>>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>>>>>> index 61bde4520f5c..f14fae45e13d 100644
>>>>>> --- a/include/linux/bpf.h
>>>>>> +++ b/include/linux/bpf.h
>>>>>> @@ -2472,6 +2472,13 @@ static inline bool has_current_bpf_ctx(void)
>>>>>>      return !!current->bpf_ctx;
>>>>>> }
>>>>>> 
>>>>>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
>>>>> 
>>>>> is_zero_terminated would be more accurate? though there is nothing
>>>>> really dynptr-specific here...
>>>> 
>>>> is_zero_terminated sounds better.
>>>> 
>>>>> 
>>>>>> +{
>>>>>> +       char *str = ptr->data;
>>>>>> +
>>>>>> +       return str[__bpf_dynptr_size(ptr) - 1] == '\0';
>>>>>> +}
>>>>>> +
>>>>>> void notrace bpf_prog_inc_misses_counter(struct bpf_prog *prog);
>>>>>> 
>>>>>> void bpf_dynptr_init(struct bpf_dynptr_kern *ptr, void *data,
>>>>>> @@ -2708,6 +2715,11 @@ static inline bool has_current_bpf_ctx(void)
>>>>>>      return false;
>>>>>> }
>>>>>> 
>>>>>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
>>>>>> +{
>>>>>> +       return false;
>>>>>> +}
>>>>>> +
>>>>>> static inline void bpf_prog_inc_misses_counter(struct bpf_prog *prog)
>>>>>> {
>>>>>> }
>>>>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>>>>> index df697c74d519..946268574e05 100644
>>>>>> --- a/kernel/trace/bpf_trace.c
>>>>>> +++ b/kernel/trace/bpf_trace.c
>>>>>> @@ -24,6 +24,7 @@
>>>>>> #include <linux/key.h>
>>>>>> #include <linux/verification.h>
>>>>>> #include <linux/namei.h>
>>>>>> +#include <linux/fileattr.h>
>>>>>> 
>>>>>> #include <net/bpf_sk_storage.h>
>>>>>> 
>>>>>> @@ -1429,6 +1430,49 @@ static int __init bpf_key_sig_kfuncs_init(void)
>>>>>> late_initcall(bpf_key_sig_kfuncs_init);
>>>>>> #endif /* CONFIG_KEYS */
>>>>>> 
>>>>>> +/* filesystem kfuncs */
>>>>>> +__diag_push();
>>>>>> +__diag_ignore_all("-Wmissing-prototypes",
>>>>>> +                 "kfuncs which will be used in BPF programs");
>>>>>> +
>>>>>> +/**
>>>>>> + * bpf_get_file_xattr - get xattr of a file
>>>>>> + * @name_ptr: name of the xattr
>>>>>> + * @value_ptr: output buffer of the xattr value
>>>>>> + *
>>>>>> + * Get xattr *name_ptr* of *file* and store the output in *value_ptr*.
>>>>>> + *
>>>>>> + * Return: 0 on success, a negative value on error.
>>>>>> + */
>>>>>> +__bpf_kfunc int bpf_get_file_xattr(struct file *file, struct bpf_dynptr_kern *name_ptr,
>>>>>> +                                  struct bpf_dynptr_kern *value_ptr)
>>>>>> +{
>>>>>> +       if (!bpf_dynptr_is_string(name_ptr))
>>>>>> +               return -EINVAL;
>>>>> 
>>>>> so dynptr can be invalid and name_ptr->data will be NULL, you should
>>>>> account for that
>>>> 
>>>> We can add a NULL check (or size check) here.
>>> 
>>> there must be some helper to check if dynptr is valid, let's use that
>>> instead of NULL checks
>> 
>> Yeah, we can use bpf_dynptr_is_null().
>> 
>>> 
>>>> 
>>>>> 
>>>>> and there could also be special dynptrs that don't have contiguous
>>>>> memory region, so somehow you'd need to take care of that as well
>>>> 
>>>> We can require the dynptr to be BPF_DYNPTR_TYPE_LOCAL. I don't think
>>>> we need this for dynptr of skb or xdp. Would this be sufficient?
>>> 
>>> well, to keep thing simple we can have a simple internal helper API
>>> that will tell if it's safe to assume that dynptr memory is contiguous
>>> and it's ok to use dynptr memory. But still, you shouldn't access data
>>> pointer directly, there must be some helper for that. Please check. It
>>> has to take into account offset and stuff like that.
>> 
>> Yeah, we can use bpf_dynptr_write(), which is a helper (not kfunc).
>> 
>>> 
>>> Also, and separately from that, we should think about providing a
>>> bpf_dynptr_slice()-like helper that will accept a fixed-sized
>>> temporary buffer and return pointer to either actual memory or copy
>>> non-contiguous memory into that buffer. That will make sure you can
>>> use any dynptr as a source of data, and only pay the price of memory
>>> copy in rare cases where it's necessary
>> 
>> I don't quite follow here. Currently, we have
>> 
>> bpf_dynptr_data()
>> bpf_dynptr_slice()
>> bpf_dynptr_slice_rdwr()
>> bpf_dynptr_write()
>> 
>> AFAICT, they are sufficient to cover existing use cases (and the new
>> use case we are adding in this set). What's the new kfunc are you
>> thinking about?
> 
> I wasn't talking about kfuncs, but rather just internal helpers to be
> used by other kfuncs when working with dynptrs as input arguments.

AFAICT, kfuncs can call other kfuncs, for example bpf_dynptr_slice_rdwr()
calls bpf_dynptr_slice(). This is limited to the same file at the moment, 
since we do not expose kfuncs in headers. 


Thanks,
Song

Re: [PATCH bpf-next 1/5] bpf: Add kfunc bpf_get_file_xattr

[Hou Tao]

Hi,

On 10/18/2023 4:31 AM, Song Liu wrote:
>
>> On Oct 17, 2023, at 11:58 AM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
>>
>> On Fri, Oct 13, 2023 at 11:29 AM Song Liu <song@kernel.org> wrote:
>>> This kfunc can be used to read xattr of a file.
>>>
>>> Since vfs_getxattr() requires null-terminated string as input "name", a new
>>> helper bpf_dynptr_is_string() is added to check the input before calling
>>> vfs_getxattr().
>>>
>>> Signed-off-by: Song Liu <song@kernel.org>
>>> ---
>>> include/linux/bpf.h      | 12 +++++++++++
>>> kernel/trace/bpf_trace.c | 44 ++++++++++++++++++++++++++++++++++++++++
>>> 2 files changed, 56 insertions(+)
>>>
>>> diff --git a/include/linux/bpf.h b/include/linux/bpf.h
>>> index 61bde4520f5c..f14fae45e13d 100644
>>> --- a/include/linux/bpf.h
>>> +++ b/include/linux/bpf.h
>>> @@ -2472,6 +2472,13 @@ static inline bool has_current_bpf_ctx(void)
>>>        return !!current->bpf_ctx;
>>> }
>>>
>>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
>> is_zero_terminated would be more accurate? though there is nothing
>> really dynptr-specific here...
> is_zero_terminated sounds better. 
>
>>> +{
>>> +       char *str = ptr->data;
>>> +
>>> +       return str[__bpf_dynptr_size(ptr) - 1] == '\0';
>>> +}
>>> +
>>> void notrace bpf_prog_inc_misses_counter(struct bpf_prog *prog);
>>>
>>> void bpf_dynptr_init(struct bpf_dynptr_kern *ptr, void *data,
>>> @@ -2708,6 +2715,11 @@ static inline bool has_current_bpf_ctx(void)
>>>        return false;
>>> }
>>>
>>> +static inline bool bpf_dynptr_is_string(struct bpf_dynptr_kern *ptr)
>>> +{
>>> +       return false;
>>> +}
>>> +
>>> static inline void bpf_prog_inc_misses_counter(struct bpf_prog *prog)
>>> {
>>> }
>>> diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c
>>> index df697c74d519..946268574e05 100644
>>> --- a/kernel/trace/bpf_trace.c
>>> +++ b/kernel/trace/bpf_trace.c
>>> @@ -24,6 +24,7 @@
>>> #include <linux/key.h>
>>> #include <linux/verification.h>
>>> #include <linux/namei.h>
>>> +#include <linux/fileattr.h>
>>>
>>> #include <net/bpf_sk_storage.h>
>>>
>>> @@ -1429,6 +1430,49 @@ static int __init bpf_key_sig_kfuncs_init(void)
>>> late_initcall(bpf_key_sig_kfuncs_init);
>>> #endif /* CONFIG_KEYS */
>>>
>>> +/* filesystem kfuncs */
>>> +__diag_push();
>>> +__diag_ignore_all("-Wmissing-prototypes",
>>> +                 "kfuncs which will be used in BPF programs");
>>> +
>>> +/**
>>> + * bpf_get_file_xattr - get xattr of a file
>>> + * @name_ptr: name of the xattr
>>> + * @value_ptr: output buffer of the xattr value
>>> + *
>>> + * Get xattr *name_ptr* of *file* and store the output in *value_ptr*.
>>> + *
>>> + * Return: 0 on success, a negative value on error.
>>> + */
>>> +__bpf_kfunc int bpf_get_file_xattr(struct file *file, struct bpf_dynptr_kern *name_ptr,
>>> +                                  struct bpf_dynptr_kern *value_ptr)
>>> +{
>>> +       if (!bpf_dynptr_is_string(name_ptr))
>>> +               return -EINVAL;
>> so dynptr can be invalid and name_ptr->data will be NULL, you should
>> account for that
> We can add a NULL check (or size check) here. 
>
>> and there could also be special dynptrs that don't have contiguous
>> memory region, so somehow you'd need to take care of that as well
> We can require the dynptr to be BPF_DYNPTR_TYPE_LOCAL. I don't think
> we need this for dynptr of skb or xdp. Would this be sufficient?

I think bpf_dynptr_is_rdonly() is also needed. Because the content of
dynptr may be modified by other bpf program and the zero-terminated
condition will not true. As suggested by Alexei, add string support in
verifier is a better choice.
>
> Thanks,
> Song
>
>>> +
>>> +       return vfs_getxattr(mnt_idmap(file->f_path.mnt), file_dentry(file), name_ptr->data,
>>> +                           value_ptr->data, __bpf_dynptr_size(value_ptr));
>>> +}

Re: [PATCH bpf-next 1/5] bpf: Add kfunc bpf_get_file_xattr

[Alexei Starovoitov]

On Fri, Oct 13, 2023 at 11:30 AM Song Liu <song@kernel.org> wrote:
> +__bpf_kfunc int bpf_get_file_xattr(struct file *file, struct bpf_dynptr_kern *name_ptr,
> +                                  struct bpf_dynptr_kern *value_ptr)
> +{
> +       if (!bpf_dynptr_is_string(name_ptr))
> +               return -EINVAL;
> +
> +       return vfs_getxattr(mnt_idmap(file->f_path.mnt), file_dentry(file), name_ptr->data,
> +                           value_ptr->data, __bpf_dynptr_size(value_ptr));
> +}
> +
> +__diag_pop();
> +
> +BTF_SET8_START(fs_kfunc_set)
> +BTF_ID_FLAGS(func, bpf_get_file_xattr, KF_SLEEPABLE)

I suspect it needs to be allowlisted too.
Sleepable might not be enough.

KP proposed such kfunc in the past and there were recursion issues.

KP,
do you remember the details?

Re: [PATCH bpf-next 1/5] bpf: Add kfunc bpf_get_file_xattr

[KP Singh]

On Tue, Oct 17, 2023 at 9:11 PM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> On Fri, Oct 13, 2023 at 11:30 AM Song Liu <song@kernel.org> wrote:
> > +__bpf_kfunc int bpf_get_file_xattr(struct file *file, struct bpf_dynptr_kern *name_ptr,
> > +                                  struct bpf_dynptr_kern *value_ptr)
> > +{
> > +       if (!bpf_dynptr_is_string(name_ptr))
> > +               return -EINVAL;
> > +
> > +       return vfs_getxattr(mnt_idmap(file->f_path.mnt), file_dentry(file), name_ptr->data,
> > +                           value_ptr->data, __bpf_dynptr_size(value_ptr));
> > +}
> > +
> > +__diag_pop();
> > +
> > +BTF_SET8_START(fs_kfunc_set)
> > +BTF_ID_FLAGS(func, bpf_get_file_xattr, KF_SLEEPABLE)
>
> I suspect it needs to be allowlisted too.
> Sleepable might not be enough.
>
> KP proposed such kfunc in the past and there were recursion issues.
>
> KP,
> do you remember the details?

yeah, have a look at Al's reply:

https://lore.kernel.org/bpf/Yrs4+ThR4ACb5eD%2F@ZenIV/

it can create deadlocks and potentially UAFs (similar to the situation
Jann mentioned). This will need to be allowlisted.