On some CPU architecures it's possible to use MMU pages of different
sizes, for example on ARC or ARM. And while for user-space
applications the page size is supposed to be transparent, there's
still some use of that extra information. In particular it's possible
to align data structures or code/data sections on page boundary, etc.
For these tricks to become possible tools which pack data (think of
the linker, like GNU "ld") need to be informed of the page size to
be considered.
Obviously, there're some sane defaults which are being used most of
the time, so we even think about that peculiarity, but when non-default
value needs to be used, GNU "ld" accepts 2 properties related to page
size:
-z common-page-size=XXX
-z max-page-size=YYY
And while in thery those might be different (but always "common" <= "max"),
and that might make sense if we build for some unknown platfrom,
in case of Buildroot when we build entire target's filesystem and so
know exactly the configuration we're targeting to, we may safely assume
"common-page-size"="max-page-size".
See a lengthy discussion in this thread [1].
Fixes:
http://autobuild.buildroot.net/results/c8b2f331c98453670cd982558144c4fd84674a3d/ (uclibc)
http://autobuild.buildroot.net/results/3a22f7aac38145b26c549254b819f87329e7a77e/ (glibc)
And while at it, recover use of "XX-page-size" for ARC, as with [2]
moving page size selection in the generic code we've got unexpected
override for ARC (note "=", but not "+="):
--------------------->8--------------------
ARCH_TOOLCHAIN_WRAPPER_OPTS = -matomic
--------------------->8--------------------
[1] https://lists.buildroot.org/pipermail/buildroot/2022-July/646176.html
[2] https://git.buildroot.net/buildroot/commit/?id=dcb74db89e74e512e36b32cea6f574a1a1ca84c4
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
This commit is based on earlier work from Łukasz Stelmach
<l.stelmach@samsung.com> to add support for different page sizes on
ARM64.
In his initial submission, Łukasz took an approach similar to this
one, i.e make it ARM64-specific. Following the feedback on the mailing
list, his second version [1] tried to generalize the logic to
configure the page size between architectures. But the general
consensus during the review process was that there wasn't much to
generalize in the end.
So, this new iteration is back to a simpler approach:
* We have new options in Config.in.arm to configure the page
size. Only 4 KB and 64 KB are supported, because our testing in
Qemu and real hardware has not allowed to get a successful setup
for 16 KB pages. We can always re-add support for 16 KB later if
that is resolved.
* The logic to define the ARCH_TOOLCHAIN_WRAPPER_OPTS options is
moved from the ARC-specific file to arch/arch.mk, and extended to
cover ARM64.
* The appropriate logic in uclibc.mk and linux.mk is added to tweak
the relevant configuration options.
* A test case is added in the runtime test infrastructure to test
building and booting under Qemu a 64 KB configuration, with all 3 C
libraries.
For the regular configuration of 4 KB pages, this commit makes one
functional change: on ARM64, -Wl,-z,max-page-size=4096 is now passed in
the compiler flags of the wrapper.
[1] https://patchwork.ozlabs.org/project/buildroot/list/?series=275452
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Signed-off-by: Arnout Vandecappelle (Essensium/Mind) <arnout@mind.be>
Back in the day we relied on a default value that used to be 8KiB
and it worked perfectly fine for ARC's default 8KiB page as well as
4 KiB ones, but not for 16 KiB, see [1] for more details.
So that we fixed by setting "max-page-size" if 16KiB pages are in use by
commit d024d369b8 ("arch/arc: Accommodate 16 KiB MMU pages").
But as Yann very rightfully mentioned here [2] we should be setting this
thing explicitly for all page sizes because:
1. Defaults might change unexpectedly
2. Explicitly set stuff is better understood
3. We act similarly to all settings but not only addressing some corner cases
[1] https://git.buildroot.org/buildroot/commit/?id=d024d369b82d2d3d9d4d75489c19e9488202bca0
[2] https://patchwork.ozlabs.org/patch/1212544/#2330647
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
Cc: Yann E. MORIN <yann.morin.1998@free.fr>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
ARC processors are known for its configurability and one of those
configurable things is MMU page size which might be set to any
power of two from 4 KiB to 16 MiB, though in the Linux kernel we
only support 4, 8 and 16 KiB due to practical considerations.
And the most used setting is 8 KiB thus GNU LD assumes maximum
page size is 8 KiB by default and while this works for smaller
pages (it's OK to align segments by larger value it will be still
peoperly aligned) this breaks execution of user-space apps on HW
with larger pages because Elf sections might very well span across
allocated pages and thus make executable broken.
Simplest example:
------------------------------------>8-----------------------------------
$ arc-linux-gcc test.c
$ arc-linux-readelf --segments a.out
Program Headers:
Type Offset VirtAddr PhysAddr FileSiz MemSiz Flg Align
...
LOAD 0x000000 0x00010000 0x00010000 0x003e8 0x003e8 R E 0x2000 <-- See
LOAD 0x001f24 0x00013f24 0x00013f24 0x000f0 0x0010c RW 0x2000
------------------------------------>8-----------------------------------
Fortunately we may override default page size settings with "max-page-size"
linker option this way:
------------------------------------>8-----------------------------------
$ arc-linux-gcc test.c -Wl,-z,max-page-size=16384
$ arc-linux-readelf --segments a.out
Elf file type is EXEC (Executable file)
Entry point 0x102c4
There are 8 program headers, starting at offset 52
Program Headers:
Type Offset VirtAddr PhysAddr FileSiz MemSiz Flg Align
...
LOAD 0x000000 0x00010000 0x00010000 0x003e8 0x003e8 R E 0x4000 <-- See
LOAD 0x001f24 0x00015f24 0x00015f24 0x000f0 0x0010c RW 0x4000
------------------------------------>8-----------------------------------
Which we implement with that change.
Signed-off-by: Alexey Brodkin <abrodkin@synopsys.com>
[yann.morin.1998@free.fr: fix comment: s/8196/8192/]
Signed-off-by: Yann E. MORIN <yann.morin.1998@free.fr>
As reported by Alexey in:
https://patchwork.ozlabs.org/patch/1087480/https://patchwork.ozlabs.org/patch/1087471/
when BR2_ARC_ATOMIC_EXT is enabled, -matomic needs to always be passed
to the compiler to allow atomic instructions to be used. So instead of
passing them through the command-line CFLAGS, we enforce them in the
toolchain wrapper directly.
Reported-by: Alexey Brodkin <Alexey.Brodkin@synopsys.com>
Signed-off-by: Yann E. MORIN <yann.morin.1998@free.fr>
Cc: Alexey Brodkin <Alexey.Brodkin@synopsys.com>
Cc: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
Cc: Arnout Vandecappelle <arnout@mind.be>
Acked-by: Alexey Brodkin <abrodkin@synopsys.com>
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@bootlin.com>
