| Age | Commit message (Collapse) | Author |
|---|
|
|
|
is needed if ldso should use libcs malloc whenever possible.
Fix RTLD_LAZY propagation to RTLD_NOW relocation when requested by
libdl.
|
|
PPC32, SPARC32/64 and S390 includes the PLT in its RELA size. This caused ldso
to always do unlazy relocation of the JMPRELs. This patch fixes it.
|
|
This patch makes -fpic work for PCC and optimzes the relcation by moving the cache
flushing stuff to JMP relocs only. Actually PPCs ldso can only handle small
GOT tables(<=8192 entries)anyhow, so it makes little sense to compile PPC with -fPIC.
libuClibc shrunk from 340724 to 330780 bytes with -fpic.
|
|
|
|
Hi yet again :)
in dl-startup.c when performing boot strap relocation the following test
exists to make sure that only "_dl_" symbols are relocated:
/* We only do a partial dynamic linking right now. The user
is not supposed to define any symbols that start with a
'_dl', so we can do this with confidence. */
if (!symname || !_dl_symbol(symname)) {
continue;
}
However on PPC(and the other archs as well I suspect) all symbols are
"_dl_" symbols so the test is never true. The test can be removed and the
whole loop simplified(smaller). This also makes it possible to
simplify elfinterp.c
This remove the scanning of ldso.so relocs, making relocation faster.
I have tested this on PPC and it works well.
Do you think this optimization will work for the other arches as well?
I can't see why not.
Jocke
* Tested on x86, arm, mipsel, and powerpc by Erik and works nicely
-Erik
|
|
shared libraries with -fPIC
|
|
useful syscall failure diagnostics.
|
|
|
|
Hi
I just noticed that gcc has an "b"(Address base register) operand that
will match all "r" registers but r0. It is a better fix then adding
r0 to the clobber list.
What do you think?
|
|
Begin converting some big ugly macros to inline functions
instead
|
|
|
|
Hi it is me again.
This is the latest ldso patch. the NEW weak symbol handling works now
with a little special handling in _dl_find_hash(). You get to chose
if you want the new or old handling :)
There was 2 missing _dl_check_if_named_library_is_loaded() calls in _dlopen().
I then disabled the _dl_check_if_named_library_is_loaded() in dl-elf.c since
it is rendundant.
Question, why does some _dl_linux_resolver(), like i386, have 2 calls
to _dl_find_hash()? I think that is wrong, isn't it?
I really hope you can check this out soon ...
|
|
|
|
to be arch specific to work properly.
|
|
to make it easier to treat it specially while not bothering the
rest of the code with the same constraints.
|
|
I still don't have a good handle on why and when the different
scope values should be used.
|
|
> Not there yet, but the interfaces are much closer now...
> Heading to bed though, since its 5:30am. :-)
This works, but I discovered something "funny". For all
relocs but COPY you can use scope instead of scope->dyn->symbol_scope
and it is much faster to do that. Search for "Funny" in the patch
to see what I mean. Probably I probably broke something, please let me
know if it works for you too.
I think I am done with ldso now. Has been fun and I hope to
actually use uClibc/busybox soon in a project.
Ohh, there is one thing left, double sized PLT entries, but maybe these
aren't needed in uClibc?
Sweet dreams
Jocke
|
|
_dl_parse_lazy_relocation_information() and _dl_parse_copy_information()
so they are all consistant, allowing for future consolidation.
Trim some trailing whitespace as well.
|
|
Here is the cleaned up laze reloc patch.
Summary:
- Minor cleanup.
- disable the "if (finaladdr <= 0x01fffffc || finaladdr >= 0xfe000000)"
test since it almost never triggered.
- Optimized the lazy relocs handling.
Would be great if you could commit ASAP.
Jocke
|
|
|
|
Removed R_PPC_NONE and R_PPC_REL24 as these does not seem to be used.
Corrected R_PPC_ADDR32 and R_PPC_JMP_SLOT
|
|
|
|
Hi again
I have done some serious clenup of powerpc/elfinterp.c
- Make it look more like glibc.
- Performance modifictions.
- Fixed a few bugs in _dl_do_reloc(). These seem not to have affected
anything, but this how glibc do it.
Jocke
|
|
The current behavior of the powerpc boot1_arch.h seems somewhat broken.
Currently room is made on the stack pointer for the link register, but the link
register is never actually pushed onto it. glibc bears the following comments:
/* Call _dl_start with one parameter pointing at argc */
mr r3,r1
/* (we have to frob the stack pointer a bit to allow room for
_dl_start to save the link register). */
followed by the -16 add to r1. Despite the fact that r1 is modified, the link
register is never actually pushed onto r1, thus the adjustment is completely
superfluous.
There's two possible fixes for this, either saving the link register in the way
that glibc does, or getting rid of the r1 adjustment. As I'm not sure if saving
the link register will actually break the _dl_boot2 callin, both options will
probably want to be played with.
The following bit of inline assembly in the attached patch builds cleanly for
me with gcc 3.3 on darwin, but I'm not able to test it any further beyond that.
|
|
> Is there anything I can do/provide that would assist in the solving of
> this problem on PowerPC? I'm still concerned about my lack of flash space...
Try this path. It fixes the recent ldso problems for me on my mpc860 board.
You should be able to use ldso without the dcbx patch to the kernel now.
Please report back.
Jocke
|
|
For sh64 we need implicit access to the symtab, primarily to get at the
->st_other value. This presently isn't possible, as PERFORM_BOOTSTRAP_RELOC()
is invoked as such:
PERFORM_BOOTSTRAP_RELOC(rpnt, reloc_addr, symbol_addr, load_addr);
while we can easily get the symtab_index value from rpnt->r_info, this still
doesn't buy us easy access to the actual table. As such, I've modified
PERFORM_BOOTSTRAP_RELOC() to take an additional SYMTAB argument. Most
architectures aren't going to care about this, but unfortunately we don't
have any other options for sh64.
The following patch fixes up the API for what we need for sh64, and updates
the other architectures appropriately.
|
|
The patch touches a minor (well, not that minor, but perhaps only
rarely encountered) bug in the powerpc dynamic linker.
The problem is that addi is called in inline assembly, but there is no
restriction on the second argument. In powerpc assembler, if the
second argument to addi is r0, it is taken as the value 0, not the
contents of r0. This happened to me, making the stack pointer 0 on
the invocation on the application.
The patch is against 0.9.22, but there didn't seem to be any changes
to the relevant section in 0.9.23.
|
|
This is just a wild guess, but you could try this to see if it fixes
Richards problem:
|
|
I think I messed up a little in my latest patch to Erik. Can you try
this on top of CVS(which I think you have already)
Jocke
And later writes:
Hi Erik
I just saw something that might be a problem.
The "delta" variable is signed and
the "delta" calculations, such as delta = PLT_LONGBRANCH_ENTRY_WORDS*4 - (insn_addr-plt_addr+4),
are supposed to be unsigned.
Jocke
|
|
Comparing glibc with uClibc makes me think that the delta calculations are
wrong here. Comparing some more I still think there are a
data_words[index] assignments missing. Here is a path that has both the
data_words[index] and the above delta calclations.
This also fixes a terribly obvious bug, also spotted by Joakim, which Erik
introduced when he copied things from the i386 ldso code.
With this patch applied, things now seem to be working perfectly!
|
|
Hi again
Back at work. Here is a patch that fixes the 2 errors I found yesterday.
I have excluded the "data_words[index]" part for now.
|
|
|
|
Oops, found another ppc 8xx bug.
8xx CPUs may need this as well to work:
|
|
> Very interesting. Do you have any suggestions for how
> we could fix our powerpc shared library loader
Removing those instr. comes with a very big performance
penalty. To flush the dcache you will have read up to 8KB
dummy data and to invalidate the icache you will have to
execute up to 16KB nops. I don't know of any other way from
user space.
hmm, actually I think it will work reliable to perform a
store to the same page(s) as the dcbst/icbi will act on. That
way you will make the DTLB Error happen(if any) prior to the
dcbst/icbi. The worst thing that can happen then is a regular
DTLB Miss and that works for dcbst/icbi.
You will have to lookout for if dcbst/icbi crosses a page
boundary. Then you will have to perform a store to both
pages.
Jocke
# And again later writes:
Hi again
I think I know what the problem is. The
PPC_DCBST;PPC_SYNC;PPC_ICBI;PPC_ISYNC sequence is executed
even if no modification has been done i some cases:
_dl_linux_resolver(), the last else has no store for insns[0].
these is a insns[1] = OPCODE_B(delta - 4) that
does not have a PPC_DCBST.
_dl_do_lazy_reloc(), for R_PPC_NONE there is no store.
for R_PPC_JMP_SLOT there is a
insns[1] = OPCODE_B(delta)that does not
have a PPC_DCBST.
_dl_do_reloc(), for R_PPC_COPY there is no store.
for R_PPC_JMP_SLOT there is a
reloc_addr[1] = OPCODE_B(delta) that does not
have a PPC_DCBST.
_dl_init_got(), I THINK that the
PPC_DCBST(plt);
PPC_DCBST(plt+4);
PPC_DCBST(plt+8);
PPC_SYNC;
PPC_ICBI(plt);
PPC_ICBI(plt+4);
PPC_ICBI(plt+8);
PPC_ISYNC;
is off a bit. The address range does not match the sum
of the plt[] and tramp[] address range.
Jocke
# And then later added the comment:
I think that the tramp[] part should be included in the
PPC_DCBST/PPC_ICBI sequence. Then you have to add entries for
plt+12 and plt+16. If the tramp[] part should be excluded,
then all is well.
Jocke
|
|
|
|
|
|
|
|
it compiles properly with gcc 3.3.
-Erik
|
|
some other minor warnings.
|
|
Ok, now i got it after a day of work.
I have had a look into glibc and found the following:
sysdeps/powerpc/dl-machine.c:
...
case R_PPC_JMP_SLOT:
/* It used to be that elf_machine_fixup_plt was used here,
but that doesn't work when ld.so relocates itself
for the second time. On the bright side, there's
no need to worry about thread-safety here. */
{
Elf32_Sword delta = finaladdr - (Elf32_Word) reloc_addr;
...
The comment made me suspicious. The same position in uClibc looks like this:
ldso/ldso/powerpc/elfinterp.c:
...
case R_PPC_JMP_SLOT:
{
unsigned long targ_addr = (unsigned long)_dl_linux_resolve;
int delta = targ_addr - (unsigned long)reloc_addr;
...
When I change it to the following it works:
...
case R_PPC_JMP_SLOT:
{
unsigned long targ_addr = *reloc_addr;
int delta = targ_addr - (unsigned long)reloc_addr;
...
I hope it will not break anything. Can anyone review this change and
commit it into CVS?
thanks,
ron
|
|
|
|
Move all configuration options into the new config system.
-Erik
|
|
-Erik
|
|
shared library support. This also adds some cleaner error
handling, which I (Erik) then ported over to x86 and arm.
In addition Stefan added the following fixes:
- in hash.c was the lvalue handling of global library functions wrong.
To fix this I had to change the prototype of _dl_find_hash. (==> TIS and
ELF spec. Vers. 1.2)
- in ldso.c was the order of the .init sections calls wrong. Before we call
the initialization code of a library we have to check that all dependend
libraries are already initialized. This can easily made by calling it in the
revers loading order. For this I added a previous pointer chain.
- in ldso.c the ELF magics wasn't checked fo PPC, MIPS and SH architecture
|
|
-Erik
|
|
out page alignment issues. As we we were assuming 4k pages, which
need not be the case...
|
|
I am very pleased to announce that the MIPS dynamic linker/loader
for uClibc is now working. It works on big and little endian
platforms.
A few minor changes were needed to avoid breaking ldd, and since this
makes some non-trivial changes, I have tested on x86, arm, and powerpc
to be sure thoese arches didn't get broken. Excellent work Steven!
|
|
is encountered...
-Erik
|
|
|
