Add EhFrameHdr::lookup_and_parse function #316

roblabla · 2018-07-31T19:52:20Z

Currently, EhFrameHdr::lookup returns a Pointer, which is an absolute pointer which may (or may not) point to the correct FDE. This API is used by unwind-rs to speed up stack unwinding. Unfortunately, unwind-rs requires patches to gimli (see gimli-rs/unwind-rs#6) in order to make full use of this API. This is because there is no method to parse the returned Pointer into a FrameDescriptionEntry. Unwind makes parse_cfi_entry, and uses it as such:

let fdeptr = eh_frame_hdr.table().unwrap().lookup(address, bases).unwrap();
let fdeptr = match fdeptr {
    Pointer::Direct(x) => x,
    _ => unreachable!(),
};

let entry = gimli::parse_cfi_entry(bases, *sel, &mut EndianBuf::new(unsafe { std::slice::from_raw_parts(fdeptr as *const u8, 0x1000000) }, NativeEndian)).unwrap().unwrap();
let target_fde = match entry {
    CieOrFde::Fde(fde) => Some(fde.parse(|offset| sel.cie_from_offset(bases, offset))?),
    CieOrFde::Cie(_) => unimplemented!(), // return error here probably
};
if let Some(fde) = target_fde {
    if !fde.contains(address) {
        return Err(gimli::Error::NoUnwindInfoForAddress);
    }
}

As can be seen here, there is a big fat unsafe in the middle, which sucks. I believe there is a better solution.

The complicated part is that the fdeptr is absolute, and has no length associated to it. It is, however, guaranteed to always point to an eh_frame section. Furthermore, the eh_frame_hdr has a pointer to the start of this section. As such, I propose this solution, which allows us to get rid of the unsafety.

I additionally took the liberty of adding the guarantee that the returned FramePointer will be valid (there is no need to check contains()).

I am in the process of testing this, and the implementation and doc might need a bit of refinement.

CC @main--

coveralls · 2018-07-31T20:36:04Z

Coverage increased (+0.05%) to 84.94% when pulling 7bdfa0b on roblabla:lookup_and_parse into ed96598 on gimli-rs:master.

main--

I like your solution. If you revise the error handling I think this can be merged.

Although since this is based on code I wrote I certainly wouldn't mind another opinion from someone else.

main-- · 2018-07-31T21:23:46Z

src/cfi.rs

+
+        let entry = parse_cfi_entry(bases, frame.clone(), &mut input)?;
+        let target_fde = match entry {
+            Some(CieOrFde::Fde(fde)) => Some(fde.parse(|offset| frame.cie_from_offset(bases, offset))?),


This might be inefficient, one may want to cache CIEs.

(which, by the way, is the same reason why fde.parse() takes a closure)

I see. Should we return PartialFrameDescriptionEntry ?

Either this or also take a closure parameter (which has the advantage that we can keep the contains check here which is a potential footgun otherwise).

main-- · 2018-07-31T21:25:14Z

src/cfi.rs

+        let entry = parse_cfi_entry(bases, frame.clone(), &mut input)?;
+        let target_fde = match entry {
+            Some(CieOrFde::Fde(fde)) => Some(fde.parse(|offset| frame.cie_from_offset(bases, offset))?),
+            Some(CieOrFde::Cie(_)) => unimplemented!(),


I never was entirely sure about this (hence the unimplemented!() but I believe this should actually be unreachable.

Either way, none of these cases should panic - returning an error is the correct reaction for the gimli API.

It should be unreachable!() yeah, eh_frame_hdr points to FDEs only, not CIEs (that would make no sense).

I'll check which error makes more sense to return.

main-- · 2018-07-31T21:28:10Z

src/cfi.rs

+            None => None
+        };
+        match target_fde {
+            Some(ref fde) if fde.contains(address) => Ok((*fde).clone()),


The ref here is required because of the pattern guard, right? You can get around the awkwardness of cloning for no reason by moving this into the match above (since the other arms of that match are dead anyways).

main-- · 2018-07-31T21:30:05Z

src/cfi.rs

+
+        let eh_frame_ptr = match self.hdr.eh_frame_ptr() {
+            Pointer::Direct(x) => x,
+            _ => unreachable!(),


Not sure if it makes sense to deal with indirect pointers here. I certainly have never encountered any. Returning an error is probably fine.

roblabla · 2018-07-31T23:53:16Z

I'll add an example in the doc, and then I think this is ready to merge.

roblabla · 2018-08-01T01:48:46Z

I added a test and an example (and hopefully fixed compilation, woops). If it compiles and the tests pass, this should be ready to go.

philipc

I like this solution too.

philipc · 2018-08-01T07:17:51Z

src/cfi.rs

+        let mut input = &mut frame.section().clone();
+        input.skip(offset)?;
+
+        let entry = parse_cfi_entry(bases, frame.clone(), &mut input)?;


Unneeded clone.

philipc · 2018-08-01T07:21:01Z

src/cfi.rs

+            Some(CieOrFde::Fde(fde)) => Ok(fde.parse(cb)?),
+            Some(CieOrFde::Cie(_)) => Err(Error::NotFdePointer),
+            None => Err(Error::NoUnwindInfoForAddress)
+        }?;


Style nit: I think I prefer returns in the match arms instead.

philipc · 2018-08-01T07:25:11Z

src/cfi.rs

+    /// # let bases = BaseAddresses::default()
+    /// #     .set_cfi(address_of_cfi_section_in_memory)
+    /// #     .set_text(address_of_text_section_in_memory)
+    /// #     .set_data(address_of_data_section_in_memory);


Simpler to let bases = unimplemented!();?

roblabla · 2018-08-01T13:35:11Z

All ready.

philipc · 2018-08-06T07:26:47Z

src/cfi.rs

+    where
+        F: FnMut(EhFrameOffset<R::Offset>) -> Result<CommonInformationEntry<EhFrame<R>, R, R::Offset>>
+    {
+        let fdeptr = self.lookup(address, bases)?;


Is it possible for this to be encoded as DW_EH_PE_pcrel? If so, we'll get this wrong, since bases.cfi will be for the eh_frame section.

Add EhFrameHdr::lookup_and_parse function

a33221b

main-- suggested changes Jul 31, 2018

View reviewed changes

Proper error handling, allow caching of CIE

2d1e95f

roblabla added 4 commits August 1, 2018 01:55

Add back the check to make sure the FDE contains the address

7a2c897

Simple example

502b365

Add unit test

7470bb1

fixup! Proper error handling, allow caching of CIE

f879d80

philipc requested changes Aug 1, 2018

View reviewed changes

Implement style nits

7bdfa0b

roblabla mentioned this pull request Aug 1, 2018

Use upstream gimli gimli-rs/unwind-rs#17

Merged

main-- approved these changes Aug 1, 2018

View reviewed changes

roblabla changed the title ~~[WIP] Add EhFrameHdr::lookup_and_parse function~~ Add EhFrameHdr::lookup_and_parse function Aug 1, 2018

philipc approved these changes Aug 1, 2018

View reviewed changes

philipc merged commit 1e063cb into gimli-rs:master Aug 1, 2018

philipc reviewed Aug 6, 2018

View reviewed changes

philipc mentioned this pull request Aug 31, 2018

UnwindTableRow should have a reference to its FDE #311

Closed

philipc mentioned this pull request Dec 31, 2018

EhHdrTable::lookup requires a base address to .eh_frame_hdr in BaseAdresses::data instead of a base address to .data as the docs suggest #350

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add EhFrameHdr::lookup_and_parse function #316

Add EhFrameHdr::lookup_and_parse function #316

roblabla commented Jul 31, 2018 •

edited

Loading

coveralls commented Jul 31, 2018 •

edited

Loading

main-- left a comment

main-- Jul 31, 2018

main-- Jul 31, 2018

roblabla Jul 31, 2018

main-- Jul 31, 2018

main-- Jul 31, 2018

roblabla Jul 31, 2018 •

edited

Loading

main-- Jul 31, 2018

main-- Jul 31, 2018

roblabla commented Jul 31, 2018 •

edited

Loading

roblabla commented Aug 1, 2018

philipc left a comment

philipc Aug 1, 2018

philipc Aug 1, 2018

philipc Aug 1, 2018

roblabla commented Aug 1, 2018

philipc Aug 6, 2018

Add EhFrameHdr::lookup_and_parse function #316

Add EhFrameHdr::lookup_and_parse function #316

Conversation

roblabla commented Jul 31, 2018 • edited Loading

coveralls commented Jul 31, 2018 • edited Loading

main-- left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

roblabla Jul 31, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

roblabla commented Jul 31, 2018 • edited Loading

roblabla commented Aug 1, 2018

philipc left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

roblabla commented Aug 1, 2018

Choose a reason for hiding this comment

roblabla commented Jul 31, 2018 •

edited

Loading

coveralls commented Jul 31, 2018 •

edited

Loading

roblabla Jul 31, 2018 •

edited

Loading

roblabla commented Jul 31, 2018 •

edited

Loading