the ultimate goal of this change is to get all code used in libc.a out
of dynlink.c, so that the dynamic linker code can be moved to its own
tree and object files in the src tree can all be shared between libc.a
and libc.so.
this is possible with the new build system that allows src/*/$(ARCH)/*
files which do not shadow a file in the parent directory, and yields a
more logical organization. eventually it will be possible to remove
arch/*/src from the build system.
if two or more threads accessed tls in a dso that was loaded after
the threads were created, then __tls_get_new could do out-of-bound
memory access (leading to segfault).
accidentally byte count was used instead of element count when
the new dtv pointer was computed. (dso->new_dtv is (void**).)
it is rare that the same dso provides dtv for several threads,
the crash was not observed in practice, but possible to trigger.
commit ad1cd43a86 eliminated
preprocessor-level omission of references to the init/fini array
symbols from object files going into libc.so. the references are weak,
and the intent was that the linker would resolve them to zero in
libc.so, but instead it leaves undefined references that could be
satisfied at runtime. normally these references would be harmless,
since the code using them does not even get executed, but some older
binutils versions produce a linking error: when linking a program
against libc.so, ld first tries to use the hidden init/fini array
symbols produced by the linker script to satisfy the references in
libc.so, then produces an error because the definitions are hidden.
ideally ld would have already provided definitions of these symbols
when linking libc.so, but the linker script for -shared omits them.
to avoid this situation, the dynamic linker now provides its own dummy
definitions of the init/fini array symbols for libc.so. since they are
hidden, everything binds at ld time and no references remain in the
dynamic symbol table. with modern binutils and --gc-sections, both
the dummy empty array objects and the code referencing them get
dropped at link time, anyway.
the _init and _fini symbols are also switched back to using weak
definitions rather than weak references since the latter behave
somewhat problematically in general, and the weak definition approach
was known to work well.
the nommu kernel shares memory when it can anyway for private
read-only maps, but semantically the map should be private. this can
make a difference when debugging breakpoints are to be used, in which
case the kernel may need to ensure that the mapping is not shared.
the new behavior matches how the kernel FDPIC loader maps the main
program and/or program interpreter (dynamic linker) binary.
this both allows removal of some of the main remaining uses of the
SHARED macro and clears one obstacle to static-linked dlopen support,
which may be added at some point in the future.
specialized single-TLS-module versions of __copy_tls and __reset_tls
are removed and replaced with code adapted from their dynamic-linked
versions, capable of operating on a whole chain of TLS modules, and
use of the dynamic linker's DSO chain (which contains large struct dso
objects) by these functions is replaced with a new chain of struct
tls_module objects containing only the information needed for
implementing TLS. this may also yield some performance benefit
initializing TLS for a new thread when a large number of modules
without TLS have been loaded, since since there is no need to walk
structures for modules without TLS.
use weak definitions that the dynamic linker can override instead of
preprocessor conditionals on SHARED so that the same libc start and
exit code can be used for both static and dynamic linking.
on linux/nommu, non-writable private mappings of files may actually
use memory shared with other processes or the fs cache. the old nommu
loader code (used when mmap with MAP_FIXED fails) simply wrote over
top of the original file mapping, possibly clobbering this shared
memory. no such breakage was observed in practice, but it should have
been possible.
the new code starts by mapping anonymous writable memory on archs that
might support nommu, then maps load segments over top of it, falling
back to read if MAP_FIXED fails. we use an anonymous map rather than a
writable file map to avoid reading more data from disk than needed.
since pages cannot be loaded lazily on fault, in case of large
data/bss, mapping the full file may read a lot of data that will
subsequently be thrown away when processing additional LOAD segments.
as a result, we cannot skip the first LOAD segment when operating in
this mode.
these changes affect only non-FDPIC nommu support.
these files are all accepted as legacy arm syntax when producing arm
code, but legacy syntax cannot be used for producing thumb2 with
access to the full ISA. even after switching to UAL, some asm source
files contain instructions which are not valid in thumb mode, so these
will need to be addressed separately.
when a library being loaded has bss (i.e. data segment with
p_memsz>p_filesz), this region needs to be zeroed with a combination
of memset and/or mmap. the regular ELF loader always did this but the
FDPIC code path omitted it, leading to objects in bss having
uninitialized/junk contents.
when determining which module an address belongs to, all function
descriptor ranges must be checked first, in case the allocated memory
falls inside another module's memory range.
dladdr itself must also check addresses against function descriptors
before doing a best-match search against the symbol table. even when
doing the latter (e.g. for code addresses obtained from mcontext_t),
also check whether the best-match was a function, and if so, replace
the result with a function descriptor address. which is the nominal
"base address" of the function and which the caller needs if it
intends to subsequently call the matching function.
since commits 2907afb8db and
6fc30c2493, __dls2 is no longer called
via symbol lookup, but instead uses relative addressing that needs to
be resolved at link time. on some linker versions, and/or if
-Bsymbolic-functions is not used, the linker may leave behind a
dynamic relocation, which is not suitable for bootstrapping the
dynamic linker, if the reference to __dls2 is marked hidden but the
definition is not actually hidden. correcting the definition to use
hidden visibility fixes the problem.
the static-PIE entry point rcrt1 was likewise affected and is also
fixed by this patch.
lookup the dso an address falls in based on the loadmap and not just a
base/length. fix the main app's fake loadmap used when loaded by a
non-fdpic-aware loader so that it does not cover the whole memory
space.
function descriptor addresses are also matched for future use by
dladdr, but reverse lookups of function descriptors via dladdr have
not been implemented yet. some revisions may be needed in the future
once reclaim_gaps supports fdpic, so that function descriptors
allocated in reclaimed heap space do not get detected as belonging to
the module whose gaps they were allocated in.
previously these resolved to the code address rather than the address
of the function descriptor.
the conditions for accepting or rejecting symbols are quite
inconsistent between the different points in the dynamic linker code
where such decisions are made. this commit attempts to be at least as
correct as anything already there, but does not improve consistency.
it has been tested to correctly avoid symbols that are merely
references to functions defined in other modules, at least in simple
usage, but at some point all symbol lookup logic should be reviewed
and refactored/unified.
this change is needed to be compatible with fdpic, where some of the
main application's relocations may be performed as part of the crt1
entry point. if we call init functions before passing control, these
relocations will not yet have been performed, and the init code will
potentially make use of invalid pointers.
conceptually, no code provided by the application or third-party
libraries should run before the application entry point. the
difference is not observable to programs using the crt1 we provide,
but it could come into play if custom entry point code is used, so
it's better to be doing this right anyway.
previously, the normal ELF library loading code was used even for
fdpic, so only the kernel-loaded dynamic linker and main app could
benefit from separate placement of segments and shared text.
at this point not all functionality is complete. the dynamic linker
itself, and main app if it is also loaded by the kernel, take
advantage of fdpic and do not need constant displacement between
segments, but additional libraries loaded by the dynamic linker follow
normal ELF semantics for mapping still. this fully works, but does not
admit shared text on nommu.
in terms of actual functional correctness, dlsym's results are
presently incorrect for function symbols, RTLD_NEXT fails to identify
the caller correctly, and dladdr fails almost entirely.
with the dynamic linker entry point working, support for static pie is
automatically included, but linking the main application as ET_DYN
(pie) probably does not make sense for fdpic anyway. ET_EXEC is
equally relocatable but more efficient at representing relocations.
the fdpic code will need to count symbols, and it may be useful
elsewhere in the future too. counting is trivial as long as sysv hash
is present, but for gnu-hash-only libraries it's complex.
the behavior of the count is changed slightly: we now include symbols
that are not accessible by the gnu hash table in the count. this may
make dladdr slightly slower. if this is a problem, dladdr can subtract
out the part that should not be accessible. unlike in the old code,
subtracting this out is easy even in the fast path where sysv hash is
available too.
these are in do_relocs. the first one was omitted in commit
301335a80b because it slightly changes
code (using dso->base rather than cached local var base) and would
have prevented easy verification. the other was an oversight.
for ordinary ELF with fixed segment displacements, load address
computation is simply adding the base load address. but for FDPIC,
each segment has its own load address, and virtual addresses need to
be adjusted according to the segment they fall in. abstracting this
computation is the first step to making the dynamic linker ready for
FDPIC.
for this first commit, a macro is used rather than a function in order
to facilitate correctness checking. I have verified that the generated
code does not change on my i386 build.
this new generic version of the stage-2 function lookup should work
for any arch where static data is accessible via got-relative or
pc-relative addressing, using approximately the technique described in
the log message for commit 2907afb8db.
since all the mips-like archs that need got slots fo access static
data have already transitioned to the new stage chaining scheme, the
old dynamic symbol lookup code is now removed.
aarch64, arm, and sh have not yet transitioned; with this commit, they
are now using the new generic code.
previously, the call into stage 2 was made by looking up the symbol
name "__dls2" (which was chosen short to be easy to look up) from the
dynamic symbol table. this was no problem for the dynamic linker,
since it always exports all its symbols. in the case of the static pie
entry point, however, the dynamic symbol table does not contain the
necessary symbol unless -rdynamic/-E was used when linking. this
linking requirement is a major obstacle both to practical use of
static-pie as a nommu binary format (since it greatly enlarges the
file) and to upstream toolchain support for static-pie (adding -E to
default linking specs is not reasonable).
this patch replaces the runtime symbolic lookup with a link-time
lookup via an inline asm fragment, which reloc.h is responsible for
providing. in this initial commit, the asm is provided only for i386,
and the old lookup code is left in place as a fallback for archs that
have not yet transitioned.
modifying crt_arch.h to pass the stage-2 function pointer as an
argument was considered as an alternative, but such an approach would
not be compatible with fdpic, where it's impossible to compute
function pointers without already having performed relocations. it was
also deemed desirable to keep crt_arch.h as simple/minimal as
possible.
in principle, archs with pc-relative or got-relative addressing of
static variables could instead load the stage-2 function pointer from
a static volatile object. that does not work for fdpic, and is not
safe against reordering on mips-like archs that use got slots even for
static functions, but it's a valid on i386 and many others, and could
provide a reasonable default implementation in the future.
The callers need to check the value of the pointer anyway, so make
them pass the pointer to gnu_lookup instead of reloading it there.
Reorder gnu_lookup arguments so that always-used ones are listed
first. GCC can choose a calling convention with arguments in registers
(e.g. up to 3 arguments in eax, ecx, edx on x86), but cannot reorder
the arguments for static functions.
Introduce gnu_lookup_filtered and use it to speed up symbol lookups in
find_sym (do_dlsym is left as is, based on an expectation that
frequently dlsym queries will use a dlopen handle rather than
RTLD_NEXT or RTLD_DEFAULT, and will not need to look at more than one
DSO).
the TLS ABI spec for mips, powerpc, and some other (presently
unsupported) RISC archs has the return value of __tls_get_addr offset
by +0x8000 and the result of DTPOFF relocations offset by -0x8000. I
had previously assumed this part of the ABI was actually just an
implementation detail, since the adjustments cancel out. however, when
the local dynamic model is used for accessing TLS that's known to be
in the same DSO, either of the following may happen:
1. the -0x8000 offset may already be applied to the argument structure
passed to __tls_get_addr at ld time, without any opportunity for
runtime relocations.
2. __tls_get_addr may be used with a zero offset argument to obtain a
base address for the module's TLS, to which the caller then applies
immediate offsets for individual objects accessed using the local
dynamic model. since the immediate offsets have the -0x8000 adjustment
applied to them, the base address they use needs to include the
+0x8000 offset.
it would be possible, but more complex, to store the pointers in the
dtv[] array with the +0x8000 offset pre-applied, to avoid the runtime
cost of adding 0x8000 on each call to __tls_get_addr. this change
could be made later if measurements show that it would help.
previously, loading of additional libraries beyond libc/ldso did not
work on nommu kernels, nor did loading programs via invocation of the
dynamic linker as a command.
the arm atomics/TLS runtime selection code is called from
__set_thread_area and depends on having libc.auxv and __hwcap
available. commit 71f099cb7d moved the
first call to __set_thread_area to the top of dynamic linking stage 3,
before this data is made available, causing the runtime detection code
to always see __hwcap as zero and thereby select the atomics/TLS
implementations based on kuser helper.
upcoming work on superh will use similar runtime detection.
ideally this early-init code should be cleanly refactored and shared
between the dynamic linker and static-linked startup.
commit f3ddd17380 inadvertently removed
the early check for "none" type relocations, causing the address
dso->base+0 to be dereferenced to obtain an addend. shared libraries,
(including libc.so) and PIE executables were unaffected, since their
base addresses are the actual address of their mappings and are
readable. non-PIE main executables, however, have a base address of 0
because their load addresses are absolute and not offset at load time.
in practice none-type relocations do not arise with toolchains that
are in use except on mips, and on mips it's moderately rare for a
non-PIE executable to have a relocation table, since the mips-specific
got processing serves in its place for most purposes.
commit f3ddd17380 introduced early
relocations and subsequent reprocessing as part of the dynamic linker
bootstrap overhaul, to allow use of arbitrary libc functions before
the main application and libraries are loaded, but only reprocessed
GOT/PLT relocation types.
commit c093e2e820 added reprocessing of
non-GOT/PLT relocations to fix an actual regression that was observed
on powerpc, but only for RELA format tables with out-of-line addends.
REL table (inline addends at the relocation address) reprocessing is
trickier because the first relocation pass clobbers the addends.
this patch extends symbolic relocation reprocessing for libc/ldso to
support all relocation types, whether REL or RELA format tables are
used. it is believed not to alter behavior on any existing archs for
the current dynamic linker and libc code. the motivations for this
change are consistency and future-proofing. it ensures that behavior
does not differ depending on whether REL or RELA tables are used,
which could lead to undetected arch-specific bugs. it also ensures
that, if in the future code depending on additional relocation types
is added to libc.so, either at the source level or as part of the
compiler runtime that gets pulled in (for example, soft-float with TLS
for fenv), the new code will work properly.
the implementation concept is simple: stage 2 of the dynamic linker
counts the number of symbolic relocations in the libc/ldso REL table
and allocates a VLA to save their addends into; stage 3 then uses the
saved addends in place of the inline ones which were clobbered. for
stack safety, a hard limit (currently 4k) is imposed on the number of
such addends; this should be a couple orders of magnitude larger than
the actual need. this number is not a runtime variable that could
break fail-safety; it is constant for a given libc.so build.
this move eliminates a duplicate "by-hand" symbol lookup loop from the
stage-1 code and replaces it with a call to find_sym, which can be
used once we're in stage 2. it reduces the size of the stage 1 code,
which is helpful because stage 1 will become the crt start file for
static-PIE executables, and it will allow stage 3 to access stage 2's
automatic storage, which will be important in an upcoming commit.
the outer-loop approach made sense when we were also processing
DT_JMPREL, which might be in REL or RELA form, to avoid major code
duplication. commit 09db855b35 removed
processing of DT_JMPREL, and in the remaining two tables, the format
(REL or RELA) is known by the name of the table. simply writing two
versions of the loop results in smaller and simpler code.
the DT_JMPREL relocation table necessarily consists entirely of
JMP_SLOT (REL_PLT in internal nomenclature) relocations, which are
symbolic; they cannot be resolved in stage 1, so there is no point in
processing them.
this fixes a regression on powerpc that was introduced in commit
f3ddd17380. global data accesses on
powerpc seem to be using a translation-unit-local GOT filled via
R_PPC_ADDR32 relocations rather than R_PPC_GLOB_DAT. being a non-GOT
relocation type, these were not reprocessed after adding the main
application and its libraries to the chain, causing libc code not to
see copy relocations in the main program, and therefore to use the
pre-copy-relocation addresses for global data objects (like environ).
the motivation for the dynamic linker only reprocessing GOT/PLT
relocation types in stage 3 is that these types always have a zero
addend, making them safe to process again even if the storage for the
addend has been clobbered. other relocation types which can be used
for address constants in initialized data objects may have non-zero
addends which will be clobbered during the first pass of relocation
processing if they're stored inline (REL form) rather than out-of-line
(RELA form).
powerpc generally uses only RELA, so this patch is sufficient to fix
the regression in practice, but is not fully general, and would not
suffice if an alternate toolchain generated REL for powerpc.
the allocating path which can fail is for dynamic TLS, which can only
occur at runtime, and the check for runtime was already made in the
outer conditional.
commit 637dd2d383 introduced the checks
for RTLD_DEFAULT and RTLD_NEXT here, claiming they fixed a regression,
but the above conditional block clearly already covered these cases,
and removing the checks produces no difference in the generated code.
this fixes truncation of error messages containing long pathnames or
symbol names.
the dlerror state was previously required by POSIX to be global. the
resolution of bug 97 relaxed the requirements to allow thread-safe
implementations of dlerror with thread-local state and message buffer.
these functions are never called directly; only their addresses are
used, so PLT indirections should never happen unless a broken
application tries to redefine them, but it's still best to make them
hidden.
Rich Felker