[Dwarf-Discuss] DW_AT_segment and relocation

[Michael Eager]

On 9/27/19 1:29 AM, Jayvee Neumann wrote:
> Thank you Michael and Ron,
> 
> what I am trying to do is write a consumer that is able to handle images 
> from several targets: i386, AMD64, and also ARM. The target images will 
> be loaded and running (and debugged during run-time).
> I want to find a way to handle addresses in a simple way, abstracting 
> from the underlying architecture. So my first Idea was giving each 
> address a segment that can either be zero or an expression.
> 
> That lead to the problem I mentioned in the first email. I think this is 
> best explained with an example. If the image (e.g. a DLL or a shared 
> library) has a preferred loading address of 0x1000, this address is used 
> in DWARF. But the image could be loaded at e.g. 0x2000 (or any other 
> address). The consumer has to adjust all the pc values in DWARF in that 
> case. A function whose DWARF information say is loaded from e.g. 0x10E8 
> to 0x1120 is actually loaded from 0x20E8 to 0x2120. An attribute 
> DW_AT_low_pc with the value 0x1500 would represent the address 0x2500 
> when the image is actually loaded. But an address like Segment:Offset 
> (e.g. CS:0x0123) would be valid during run-time without any adjustment, 
> since the segment is computed from registers during run-time and thus 
> necessarily adjusted to the load-time addresses and not build-time 
> addresses. So the rule of thumb would be: Do not adjust segmented 
> addresses in the consumer. Or have I misunderstood something about this 
> matter?

When it reads the debug info, a debugger has to go through the same 
relocation process as happens when the executable or library is loaded. 
Every relocatable reference needs to be adjusted to account for the 
situation you describe, where the code is loaded at a different offset.

DW_AT_segment really cannot be used to do this.  There is (in general) 
no register or other expression which will perform this relocation for 
you, since the computation depends on information (e. g., the load 
addresses for each piece of the executable) which may not be accessible 
to the program.

Unless you are planning to support 286 or 386 in 16-bit mode in one of 
the segmented memory schemes used on these processors, you do not need 
support for DW_AT_segment.

> What is kind of counter intuitive is the fact that the DWARF5 standard 
> says: if DW_AT_segment is specified at any parent DIE it is valid at its 
> subsequent child DIEs. Consumer libraries like libdwarf have no function 
> that allows for determining the parent DIE if only the child DIE is 
> known. It would be necessary to go through all DIEs and check which one 
> is an ancestor of the DIE in question. So this little sentence implies a 
> lot of difficulties for a consumer.

I haven't looked at how debuggers like GDB handle segments in a very 
long time.  I believe that when the DWARF info is read and used to 
create an internal representation, a current segment value is maintained 
as the debug info is parsed recursively.  This is set as you descend the 
parse tree, pushing previous values, and popped when ascend the tree.

Almost all architectures, and specifically the ARM and AMD64 
architectures you mention, use linear, not segmented, address spaces.

> 
> Kind regards,
> Jayvee
> 
> Am Do., 26. Sept. 2019 um 23:12?Uhr schrieb Michael Eager 
> <eager at eagerm.com <mailto:eager at eagerm.com>>:
> 
>     On 9/26/19 3:33 AM, Jayvee Neumann via Dwarf-Discuss wrote:
>      > Dear DWARF experts,
>      >
>      > I have a question regarding the attribute DW_AT_segment. I do not
>     quite
>      > understand how to handle it, yet.
>      > It can appear in a DIE (or its parent) whenever DW_AT_low_pc,
>      > DW_AT_high_pc, DW_AT_ranges, DW_AT_entry_pc, or a location
>     description
>      > that evaluates to an Address are used.
>      > It can contain a location expression itself. So it can depend not
>     only
>      > on compile-time information but also on run-time information.
>      >
>      > Assume, I have an attribute DW_AT_low_pc. If there is no
>     DW_AT_segment,
>      > this attribute is simply the low_pc as it was determined during
>      > compile-time. While debugging, this value has to be relocated if the
>      > image had been relocated. If there is a DW_AT_segment expression is
>      > relocation still necessary? Evaluating the expression could also
>      > incorporate the usage of run-time information and relocated
>     addresses
>      > (i.e. the DS/CS/SS registers).
>      >
>      > I hope I was able to articulate my question well, English is not my
>      > native language.
>      >
>      > Kind Regards
>      > Jayvee
> 
>     Your question is quite clear, but we don't know the context.? What
>     architecture are you working with and what ar you trying to do?
> 
>     As Ron mentioned, AT_segment is intended to support architectures like
>     the X86 where an address is composed of a [segment,offset] pair.? Most
>     architectures do not use segments.
> 
>     -- 
>     Michael Eager eager at eagerm.com <mailto:eager at eagerm.com>
>     1960 Park Blvd., Palo Alto, CA 94306
> 

-- 
Michael Eager    eager at eagerm.com
1960 Park Blvd., Palo Alto, CA 94306