The arena allocator incrementally applies RW mprotect to a PROT_NONE
mapping. Use PROT_MAX to ensure the pointers derived from the original
mapping have RW capability permission.
Specifies the prot flags a mapping may gain via mprotect or MAP_FIXED.
On CHERI targets this is used to get capability with more permissions
than the original mmap protection would imply.
This works around a gcc issue where it const folds inf/inf into nan,
preventing the invalid exception signal to be raised.
(x-x)/(x-x) is more robust against optimizations and works for x==nan
too.
The issue should be fixed in gcc-11.3.0 and gcc-12, but glibc supports
older compilers.
Add more cap_ hooks to implement narrowing without depending on a
global capability covering the heap. Either recording every
narrowed capability in a lookup table or recording every mapping
used for the heap are supported. The morello implmentation uses
a lookup table for now.
The lookup table adds memory overhead, failure paths and locks.
Recording and removing entries from the lookup table must be done
carefully in realloc so on failure the old pointer is usable and
on success the old pointer is immediately reusable concurrently.
The locks require fork hooks so malloc works in multi-threaded
fork child.
Public interfaces return pointers with narrow bounds, this internally
requires bumping the size and alignment requirement of allocations so
the bounds are representible.
When pointers with narrow bounds need to be turned back to have wide
bounds (free, realloc), the pointer is rederived from DDC. (So this
patch relies on DDC to cover all heap memory with RW permission.)
Allocations above the mmap threshold waste memory for alignment and
realloc often falls back to the inefficient alloc, copy, free sequence
instead of mremap or other inplace solution.
__libc_free must only be used for memory given out by __libc_malloc
and similar public apis, but tcache stores a cache of already freed
pointers and itself is allocated using internal malloc apis. Strong
double free detection in __libc_free breaks tcache_thread_shutdown,
so use a cut down version of free to reset tcache entries.
In dl_iterate_phdr phdr is the only capability passed to the callback
that may be used to derive pointers of the elf module, so ensure it
has wide bounds.
Used internally for r_debug tests, but with the assumption that
the return value can be dereferenced, so change the prototype
and return a valid capability.
Also used in pldd, where we only support purecap abi processes.
The dlpi_addr field is a capability that has value l_addr, but we can
only do this for libraries (ET_DYN) where l_addr == l_map_start,
otherwise we return l_addr which is normally 0 then (ET_EXEC) so the
caller can detect and special case it.
For now l_addr != 0 and l_addr != l_map_start case is not supported.
Note: this api may be used by the unwinder to find and read .eh_frame
data.
TODO: dlpi_addr could be address only, but requires unwinder update
and agreement about the abi.
Pointers are no longer derived from l_addr, but l_map_start (RX) and
l_rw_start (RW) so it does not have to be a capability.
This also allows removing hacks where l_addr was derived from DDC.
There is no traditional TLS support in morello that would explicitly
call __tls_get_addr, but the libc uses it internally and the returned
pointer escapes to user code. So bound the pointers according to
the tls symbol size instead of doing so in each caller.
(Affects dlsym and dynamic TLSDESC.)
It has to return a pointer that can be dereferenced, so it must be
derived correctly from RX and RW capabilities.
Try to have tight object bounds and seal function symbols.
All symbol addresses can be derived from the RX capability of the
module (l_map_start). For RW object symbols pointer will have to
be rederived from l_rw_start.
The dynamic section of an executable needs to be written to set the
DT_DEBUG entry for debuggers (unless the target has some other place
to store r_debug). For this reason we make l_ld writable whenever
the dynamic section is writable.
The ld.so l_ld is kept RX, since it does not have DT_DEBUG.
(Note: relocating the dynamic section is not allowed on cheri and
that's the only other reason glibc would write to it.)
The l_map_start and l_rw_start of the ld.so and exe comes from the auxv
since they are normally mapped by the kernel. Some generic code had to
be modified so l_map_start is propagated and not overwritten when it is
recomputed.
The l_rw_range should exclude the relro region, but in libc.so and
ld.so this does not work: symbols are accessed before relro is applied
and then the permission should be writable.
Use a new hook to do the rtld bootstrap map base address and root
capability setup on CHERI.
This will be needed to use separate per module RX and RW root caps.
For each module keep an RX and an RW root capability. Use the existing
l_map_start for RX (covering all load segments) and add l_rw_start for
RW (covering all writable load segments).
For relocation processing, we also need individual RW ranges to decide
which objects need to be derived from RW and RX capabilities. In
practice most modules have exactly one RW segment and it's unlikely
that any module needs more than four distinct ranges to tightly cover
the RW mappings.
Only added on CHERI targets so always has to be used behind ifdef.
The purecap ELF entry is special: passes separate argc, argv, envp,
auxv in registers instead of on the stack.
The ldso internal _dl_start still expects continuous argc, argv, envp,
auxv, so that's emulated.
For each module there will be separate RW and RX capabilities that
cover the writable and all load segments respectively.
Prepare the relative reloc processing in static start code for such
separate capabilities.
The prototype of __libc_start_main is changed to
void
__libc_start_main (int main (int, char **, char **, void *),
int argc, char **argv, char **envp, void *auxv,
void rtld_fini (void), void *sp);
so envp is passed down separately and the unused init, fini args are
dropped.
Previously the entry address was fixed up to be a capability before
using it so l_entry could be ElfW(Addr), but the code is simpler and
more consistent if l_entry is a capability throughout:
The AT_ENTRY auxv entry is specified to be a capability and a number
if internal l_entry usage is simpler if it is elfptr_t.
This is a temporary workaround.
length is rounded up to pagesize and don't use exact bound (bounds
will be larger if exact value is not representable).
TODO: kernel should do this
nl_langinfo_l ignores its locale argument with NL_CURRENT_INDIRECT
which is wrong when that argument does not match the current thread's
locale.
upstream glibc is not tested with static linking so this is not found.
__correctly_grouped_prefixmb only worked with thousands_len == 1,
otherwise it read past the end of cp or thousands.
Avoid OOB access by considering thousands_len when initializing cp.
On morello with strict bounds checking this fixes
FAIL: stdlib/tst-strtod4
FAIL: stdlib/tst-strtod5i
both of which set cs_CZ.UTF-8 locale that has 3 byte thousands_len.
