> - fdget_raw is only used inside fs/, so it doesn't look right to skip the layers. Sorry, I mistakenly thought that as long as it is non-internel.h it can be accessed externally. Would you please give me more details about how to deal with this situation or is there any documents explain this? I hope to learn more then fix this. > - accessing current->fs is not always correct, so the code should somehow > check that it's ok to do so, but I'm not sure if (in_irq()) would be enough. I'll check more about this these days, then determine the solution. > - some implementations of d_dname do sleep. For example: dmabuffs_dname. > Though it seems to me that it's a bug in that particular FS. But I'd like > to hear clear yes from VFS experts that fdget_raw() + d_path() is ok > from preempt_disabled section. > The other alternative is to wait for sleepable and preemptible BPF programs to > appear. Which is probably a month or so away. Then all these issues will > disappear. Sorry for didn't check the whole exist callback functions, gave an imprecise conclusion. I will do my best to learn more, I hope I can make better contributions to bpf in the future. > The other alternative is to wait for sleepable and preemptible BPF programs to > appear. Which is probably a month or so away. Then all these issues will > disappear. I think wait for sleepable and preemptible BPF programs to appear is a better way to compatible with all kinds implementations of d_dname. Thank you for providing these valuable suggestions and information. Alexei Starovoitov <alexei.starovoitov@xxxxxxxxx> 于2019年11月6日周三 上午6:19写道： > > On Sun, Nov 03, 2019 at 02:54:17AM -0500, Wenbo Zhang wrote: > > When people want to identify which file system files are being opened, > > read, and written to, they can use this helper with file descriptor as > > input to achieve this goal. Other pseudo filesystems are also supported. > > > > This requirement is mainly discussed here: > > > > https://github.com/iovisor/bcc/issues/237 > > > > v6->v7: > > - fix missing signed-off-by line > > > > v5->v6: addressed Andrii's feedback > > - avoid unnecessary goto end by having two explicit returns > > > > v4->v5: addressed Andrii and Daniel's feedback > > - rename bpf_fd2path to bpf_get_file_path to be consistent with other > > helper's names > > - when fdget_raw fails, set ret to -EBADF instead of -EINVAL > > - remove fdput from fdget_raw's error path > > - use IS_ERR instead of IS_ERR_OR_NULL as d_path ether returns a pointer > > into the buffer or an error code if the path was too long > > - modify the normal path's return value to return copied string length > > including NUL > > - update this helper description's Return bits. > > > > v3->v4: addressed Daniel's feedback > > - fix missing fdput() > > - move fd2path from kernel/bpf/trace.c to kernel/trace/bpf_trace.c > > - move fd2path's test code to another patch > > - add comment to explain why use fdget_raw instead of fdget > > > > v2->v3: addressed Yonghong's feedback > > - remove unnecessary LOCKDOWN_BPF_READ > > - refactor error handling section for enhanced readability > > - provide a test case in tools/testing/selftests/bpf > > > > v1->v2: addressed Daniel's feedback > > - fix backward compatibility > > - add this helper description > > - fix signed-off name > > > > Signed-off-by: Wenbo Zhang <ethercflow@xxxxxxxxx> > > --- > > include/uapi/linux/bpf.h | 15 ++++++++++- > > kernel/trace/bpf_trace.c | 48 ++++++++++++++++++++++++++++++++++ > > tools/include/uapi/linux/bpf.h | 15 ++++++++++- > > 3 files changed, 76 insertions(+), 2 deletions(-) > > > > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h > > index a6bf19dabaab..d618a914c6fe 100644 > > --- a/include/uapi/linux/bpf.h > > +++ b/include/uapi/linux/bpf.h > > @@ -2777,6 +2777,18 @@ union bpf_attr { > > * restricted to raw_tracepoint bpf programs. > > * Return > > * 0 on success, or a negative error in case of failure. > > + * > > + * int bpf_get_file_path(char *path, u32 size, int fd) > > + * Description > > + * Get **file** atrribute from the current task by *fd*, then call > > + * **d_path** to get it's absolute path and copy it as string into > > + * *path* of *size*. The **path** also support pseudo filesystems > > + * (whether or not it can be mounted). The *size* must be strictly > > + * positive. On success, the helper makes sure that the *path* is > > + * NUL-terminated. On failure, it is filled with zeroes. > > + * Return > > + * On success, returns the length of the copied string INCLUDING > > + * the trailing NUL, or a negative error in case of failure. > > */ > > #define __BPF_FUNC_MAPPER(FN) \ > > FN(unspec), \ > > @@ -2890,7 +2902,8 @@ union bpf_attr { > > FN(sk_storage_delete), \ > > FN(send_signal), \ > > FN(tcp_gen_syncookie), \ > > - FN(skb_output), > > + FN(skb_output), \ > > + FN(get_file_path), > > > > /* integer value in 'imm' field of BPF_CALL instruction selects which helper > > * function eBPF program intends to call > > diff --git a/kernel/trace/bpf_trace.c b/kernel/trace/bpf_trace.c > > index f50bf19f7a05..41be1c5989af 100644 > > --- a/kernel/trace/bpf_trace.c > > +++ b/kernel/trace/bpf_trace.c > > @@ -683,6 +683,52 @@ static const struct bpf_func_proto bpf_send_signal_proto = { > > .arg1_type = ARG_ANYTHING, > > }; > > > > +BPF_CALL_3(bpf_get_file_path, char *, dst, u32, size, int, fd) > > +{ > > + struct fd f; > > + char *p; > > + int ret = -EBADF; > > + > > + /* Use fdget_raw instead of fdget to support O_PATH, and > > + * fdget_raw doesn't have any sleepable code, so it's ok > > + * to be here. > > + */ > > + f = fdget_raw(fd); > > + if (!f.file) > > + goto error; > > + > > + /* d_path doesn't have any sleepable code, so it's ok to > > + * be here. But it uses the current macro to get fs_struct > > + * (current->fs). So this helper shouldn't be called in > > + * interrupt context. > > + */ > > + p = d_path(&f.file->f_path, dst, size); > > + if (IS_ERR(p)) { > > + ret = PTR_ERR(p); > > + fdput(f); > > + goto error; > > + } > > This is definitely very useful helper that bpf tracing community has > been asking for long time, but I have few concerns with implementation: > - fdget_raw is only used inside fs/, so it doesn't look right to skip the layers. > - accessing current->fs is not always correct, so the code should somehow > check that it's ok to do so, but I'm not sure if (in_irq()) would be enough. > - some implementations of d_dname do sleep. For example: dmabuffs_dname. > Though it seems to me that it's a bug in that particular FS. But I'd like > to hear clear yes from VFS experts that fdget_raw() + d_path() is ok > from preempt_disabled section. > > The other alternative is to wait for sleepable and preemptible BPF programs to > appear. Which is probably a month or so away. Then all these issues will > disappear. >
