On Fri, Nov 1, 2024 at 1:28 PM Juntong Deng <juntong.deng@xxxxxxxxxxx> wrote:
>
> On 2024/11/1 19:08, Andrii Nakryiko wrote:
> > On Tue, Oct 29, 2024 at 5:17 PM Juntong Deng <juntong.deng@xxxxxxxxxxx> wrote:
> >>
> >> This patch adds struct file related CRIB kfuncs.
> >>
> >> bpf_fget_task() is used to get a pointer to the struct file
> >> corresponding to the task file descriptor. Note that this function
> >> acquires a reference to struct file.
> >>
> >> bpf_get_file_ops_type() is used to determine what exactly this file
> >> is based on the file operations, such as socket, eventfd, timerfd,
> >> pipe, etc, in order to perform different checkpoint/restore processing
> >> for different file types. This function currently has only one return
> >> value, FILE_OPS_UNKNOWN, but will increase with the file types that
> >> CRIB supports for checkpoint/restore.
> >>
> >> Signed-off-by: Juntong Deng <juntong.deng@xxxxxxxxxxx>
> >> ---
> >> kernel/bpf/crib/crib.c | 4 ++++
> >> kernel/bpf/crib/files.c | 44 +++++++++++++++++++++++++++++++++++++++++
> >> 2 files changed, 48 insertions(+)
> >>
> >
> > Please CC Christian Brauner and fs mailing list
> > (linux-fsdevel@xxxxxxxxxxxxxxx, both cc'ed) on changes like this (this
> > entire patch set)
> >
>
> Thanks for your reply!
>
> I will CC Christian Brauner and fs mailing list in the next
> patch series.
>
> >> diff --git a/kernel/bpf/crib/crib.c b/kernel/bpf/crib/crib.c
> >> index e6536ee9a845..78ddd19d5693 100644
> >> --- a/kernel/bpf/crib/crib.c
> >> +++ b/kernel/bpf/crib/crib.c
> >> @@ -14,6 +14,10 @@ BTF_ID_FLAGS(func, bpf_iter_task_file_next, KF_ITER_NEXT | KF_RET_NULL)
> >> BTF_ID_FLAGS(func, bpf_iter_task_file_get_fd)
> >> BTF_ID_FLAGS(func, bpf_iter_task_file_destroy, KF_ITER_DESTROY)
> >>
> >> +BTF_ID_FLAGS(func, bpf_fget_task, KF_ACQUIRE | KF_TRUSTED_ARGS | KF_RET_NULL)
> >> +BTF_ID_FLAGS(func, bpf_get_file_ops_type, KF_TRUSTED_ARGS)
> >> +BTF_ID_FLAGS(func, bpf_put_file, KF_RELEASE)
> >> +
> >> BTF_KFUNCS_END(bpf_crib_kfuncs)
> >>
> >> static const struct btf_kfunc_id_set bpf_crib_kfunc_set = {
> >> diff --git a/kernel/bpf/crib/files.c b/kernel/bpf/crib/files.c
> >> index ececf150303f..8e0e29877359 100644
> >> --- a/kernel/bpf/crib/files.c
> >> +++ b/kernel/bpf/crib/files.c
> >> @@ -5,6 +5,14 @@
> >> #include <linux/fdtable.h>
> >> #include <linux/net.h>
> >>
> >> +/**
> >> + * This enum will grow with the file types that CRIB supports for
> >> + * checkpoint/restore.
> >> + */
> >> +enum {
> >> + FILE_OPS_UNKNOWN = 0
> >> +};
> >> +
> >> struct bpf_iter_task_file {
> >> __u64 __opaque[3];
> >> } __aligned(8);
> >> @@ -102,4 +110,40 @@ __bpf_kfunc void bpf_iter_task_file_destroy(struct bpf_iter_task_file *it)
> >> fput(kit->file);
> >> }
> >>
> >> +/**
> >> + * bpf_fget_task() - Get a pointer to the struct file corresponding to
> >> + * the task file descriptor
> >> + *
> >> + * Note that this function acquires a reference to struct file.
> >> + *
> >> + * @task: the specified struct task_struct
> >> + * @fd: the file descriptor
> >> + *
> >> + * @returns the corresponding struct file pointer if found,
> >> + * otherwise returns NULL
> >> + */
> >> +__bpf_kfunc struct file *bpf_fget_task(struct task_struct *task, unsigned int fd)
> >> +{
> >> + struct file *file;
> >> +
> >> + file = fget_task(task, fd);
> >> + return file;
> >> +}
> >> +
> >> +/**
> >> + * bpf_get_file_ops_type() - Determine what exactly this file is based on
> >> + * the file operations, such as socket, eventfd, timerfd, pipe, etc
> >> + *
> >> + * This function will grow with the file types that CRIB supports for
> >> + * checkpoint/restore.
> >> + *
> >> + * @file: a pointer to the struct file
> >> + *
> >> + * @returns the file operations type
> >> + */
> >> +__bpf_kfunc unsigned int bpf_get_file_ops_type(struct file *file)
> >> +{
> >> + return FILE_OPS_UNKNOWN;
> >> +}
> >> +
> >
> > this is not very supportable, users can do the same by accessing
> > file->f_op and comparing it to a set of known struct file_operations
> > references.
> >
>
> Yes, users can access file->f_op, but there seems to be no way for
> users to get references to struct file_operations for the various file
> types? For example, how does a user get a reference to socket_file_ops?
See [0]. Libbpf will find it for the BPF program from kallsyms.
[0] https://github.com/torvalds/linux/blob/master/tools/testing/selftests/bpf/progs/test_ksyms.c#L13-L18
>
> Also, currently the struct file_operations for most of the file types
> are static, and I cannot even get a reference to them in
> crib/files.c directly.
>
> My future plan is to add functions like is_socket_file_ops to the
> corresponding files (e.g. net/socket.c).
>
> >> __bpf_kfunc_end_defs();
> >> --
> >> 2.39.5
> >>
>
