Since commit 35fa91eed8 ("ARM: kernel: merge core and init PLTs"),
the ARM module PLT code allocates all PLT entries in a single core
section, since the overhead of having a separate init PLT section is
not justified by the small number of PLT entries usually required for
init code.
However, the core and init module regions are allocated independently,
and there is a corner case where the core region may be allocated from
the VMALLOC region if the dedicated module region is exhausted, but the
init region, being much smaller, can still be allocated from the module
region. This puts the PLT entries out of reach of the relocated branch
instructions, defeating the whole purpose of PLTs.
So split the core and init PLT regions, and name the latter ".init.plt"
so it gets allocated along with (and sufficiently close to) the .init
sections that it serves. Also, given that init PLT entries may need to
be emitted for branches that target the core module, modify the logic
that disregards defined symbols to only disregard symbols that are
defined in the same section.
Fixes: 35fa91eed8 ("ARM: kernel: merge core and init PLTs")
Cc: <stable@vger.kernel.org> # v4.9+
Reported-by: Angus Clark <angus@angusclark.org>
Tested-by: Angus Clark <angus@angusclark.org>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
Given that we now sort the relocation sections in a way that guarantees
that entries that can share a single PLT entry end up adjacently, there
is no a longer a need to go over the entire list to look for an existing
entry that matches our jump target. If such a match exists, it was the
last one to be emitted, so we can simply check the preceding slot.
Note that this will still work correctly in the [theoretical] presence of
call/jump relocations against SHN_UNDEF symbols with non-zero addends,
although not optimally. Since the relocations are presented in the same
order that we checked them for duplicates, any duplicates that we failed
to spot the first time around will be accounted for in the PLT allocation
so there is guaranteed to be sufficient space for them when actually
emitting the PLT.
For instance, the following sequence of relocations:
000004d8 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
000004fc 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
0000050e 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
00000520 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
00000532 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
00000544 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
00000556 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
00000568 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
0000057a 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
0000058c 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
0000059e 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
000005b0 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
000005c2 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
000005d4 00058b0a R_ARM_THM_CALL 00000000 warn_slowpath_null
may result in several PLT entries to be allocated, and also emitted, if
any of the entries in the middle refer to a Place that contains a non-zero
addend (i.e., one for all the preceding zero-addend relocations, one for
all the following zero-addend relocations, and one for the non-zero addend
relocation itself)
Tested-by: Jongsung Kim <neidhard.kim@lge.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
The PLT allocation routines try to establish an upper bound on the
number of PLT entries that will be required at relocation time, and
optimize this by disregarding duplicates (i.e., PLT entries that will
end up pointing to the same function). This is currently a O(n^2)
algorithm, but we can greatly simplify this by
- sorting the relocation section so that relocations that can use the
same PLT entry will be listed adjacently,
- disregard jump/call relocations with addends; these are highly unusual,
for relocations against SHN_UNDEF symbols, and so we can simply allocate
a PLT entry for each one we encounter, without trying to optimize away
duplicates.
Tested-by: Jongsung Kim <neidhard.kim@lge.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
When CONFIG_ARM_MODULE_PLTS is enabled, jump and call instructions in
modules no longer need to be within 16 MB (8 MB for Thumb2) of their
targets. If they are further away, a PLT entry will be generated on the
fly for each of them, which extends the range to the entire 32-bit
address space.
However, since these PLT entries will become the branch targets of the
original jump and call instructions, the PLT itself needs to be in
range, or we end up in the same situation we started in. Since the PLT
is in a separate section, this essentially means that all jumps and calls
inside the same module must be resolvable without PLT entries.
The PLT allocation code executes before the module itself is loaded in
its final location, and so it has to use a worst-case estimate for
which jumps and calls will require an entry in the PLT at relocation
time. As an optimization, this code deduplicates entries pointing to
the same symbol, using a O(n^2) algorithm. However, it does not take
the above into account, i.e., that PLT entries will only be needed for
jump and call relocations against symbols that are not defined in the
module.
So disregard relocations against symbols that are defined in the module
itself.
As an additional minor optimization, ignore input sections that lack
the SHF_EXECINSTR flag. Since jump and call relocations operate on
executable instructions only, there is no need to look in sections that
do not contain executable code.
Tested-by: Jongsung Kim <neidhard.kim@lge.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
The PLT code uses a separate .init.plt section to allocate PLT entries
for jump and call instructions in __init code. However, even for fairly
sizable modules like mac80211.ko, we only end up with a couple of PLT
entries in the .init section, and so we can simplify the code
significantly by emitting all PLT entries into the same section.
Tested-by: Jongsung Kim <neidhard.kim@lge.com>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Makes it easier to handle init vs core cleanly, though the change is
fairly invasive across random architectures.
It simplifies the rbtree code immediately, however, while keeping the
core data together in the same cachline (now iff the rbtree code is
enabled).
Acked-by: Peter Zijlstra <peterz@infradead.org>
Reviewed-by: Josh Poimboeuf <jpoimboe@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>
The new veneer support for loadable modules on ARM uses the
__opcode_to_mem_thumb32() function to count R_ARM_THM_CALL
and R_ARM_THM_JUMP24 relocations.
However, this function is not defined for big-endian kernels
on ARMv5 or before, causing a compile-time error:
arch/arm/kernel/module-plts.c: In function 'count_plts':
arch/arm/kernel/module-plts.c:124:9: error: implicit declaration of function '__opcode_to_mem_thumb32' [-Werror=implicit-function-declaration]
__opcode_to_mem_thumb32(0x07ff2fff)))
^
As we know that this part of the function is only needed for
Thumb2 kernels, and that those can never happen with BE32,
we can avoid the error by enclosing the code in an #ifdef.
Fixes: 7d485f647c ("ARM: 8220/1: allow modules outside of bl range")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Loading modules far away from the kernel in memory is problematic
because the 'bl' instruction only has limited reach, and modules are not
built with PLTs. Instead of using the -mlong-calls option (which affects
all compiler emitted bl instructions, but not the ones in assembler),
this patch allocates some additional space at module load time, and
populates it with PLT like veneers when encountering relocations that
are out of range.
This should work with all relocations against symbols exported by the
kernel, including those resulting from GCC generated implicit function
calls for ftrace etc.
The module memory size increases by about 5% on average, regardless of
whether any PLT entries were actually needed. However, due to the page
based rounding that occurs when allocating module memory, the average
memory footprint increase is negligible.
Reviewed-by: Nicolas Pitre <nico@linaro.org>
Signed-off-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
