Clean up ARM compressed loader
PATCH 1/1 Clean up ARM compressed loader
The -Dstatic= stuff in the Makefile is a hack and subtly breaks the current inflate implementation, because the fixed inflate tables are included into a function and removing static places them in the stack, which bloats the stack and breaks references after the function returns. So get rid of the hack.
Instead, clean up the stub loader and unify it. The loader is now exactly the same regardless of whether you've enabled CONFIG_ZBOOT_ROM or not, and runs from RAM in PIC mode. CONFIG_ZBOOT_ROM simply enables an extra section of stub code that lets the zImage boot from ROM by copying itself to RAM first.
The old TEXT/BSS settings for the ROM mode are replaced by CONFIG_ZBOOT_ROM_START and CONFIG_ZBOOT_ROM_END. These simply define a ROM address space area that lets the loader know when it should copy itself to RAM. The zImage can be burned into ROM/Flash anywhere within this space - it's still relocatable. It will also run from RAM.
When running from ROM, the copy is done to the kernel base address. This gets overwritten with the decompressed kernel later by the existing relocation code that's already used for RAM zImages loaded where the final kernel should be. This isn't fully optimal, but the code needs to be there anyway for RAM use, and this way we don't have to worry about finding a suitable non-overlapping region of RAM to copy the zImage.
The disadvantage of copying zImage to RAM is requiring enough RAM to fit both the compressed and uncompressed kernel image. This shouldn't be a problem in practical systems.
This is untested on a real ROM platform, but has been tested using RAM as fake ROM.
--- arch/arm/Kconfig | 35 +++--- arch/arm/boot/Makefile | 2 +- arch/arm/boot/compressed/.gitignore | 1 - arch/arm/boot/compressed/Makefile | 27 +----- arch/arm/boot/compressed/head.S | 173 ++++++++++++++++++------------- arch/arm/boot/compressed/misc.c | 3 +- arch/arm/boot/compressed/vmlinux.lds | 71 +++++++++++++ arch/arm/boot/compressed/vmlinux.lds.in | 61 ----------- 8 files changed, 195 insertions(+), 178 deletions(-) create mode 100644 arch/arm/boot/compressed/vmlinux.lds delete mode 100644 arch/arm/boot/compressed/vmlinux.lds.in
arch/arm/Kconfig +18 -17
@@ -1259,38 +1259,39 @@ endmenu menu "Boot options" -# Compressed boot loader in ROM. Yes, we really want to ask about -# TEXT and BSS so we preserve their values in the config files. -config ZBOOT_ROM_TEXT - hex "Compressed ROM boot loader base address" +config ZBOOT_ROM_START + hex "ROM/flash start address (for compressed ROM boot)" default "0" help - The physical address at which the ROM-able zImage is to be - placed in the target. Platforms which normally make use of - ROM-able zImage formats normally set this to a suitable + The physical address of the ROM in the target. This setting is + used by the compressed loader to determine whether it is running + from ROM or RAM. This does not have to be the address of the zImage in + ROM; it can be the start of the ROM chip. Platforms which normally + make use of ROM-able zImage formats normally set this to a suitable value in their defconfig file. If ZBOOT_ROM is not enabled, this has no effect. -config ZBOOT_ROM_BSS - hex "Compressed ROM boot loader BSS address" +config ZBOOT_ROM_SIZE + hex "ROM/flash size (for compressed ROM boot)" default "0" help - The base address of an area of read/write memory in the target - for the ROM-able zImage which must be available while the - decompressor is running. It must be large enough to hold the - entire decompressed kernel plus an additional 128 KiB. - Platforms which normally make use of ROM-able zImage formats - normally set this to a suitable value in their defconfig file. + The size of the ROM chip in your target. You may place the zImage in + ROM anywhere within the span defined by ZBOOT_ROM_START and + ZBOOT_ROM_SIZE. Platforms which normally make use of ROM-able zImage + formats normally set this to a suitable value in their defconfig file. If ZBOOT_ROM is not enabled, this has no effect. config ZBOOT_ROM bool "Compressed boot loader in ROM/flash" - depends on ZBOOT_ROM_TEXT != ZBOOT_ROM_BSS + depends on ZBOOT_ROM_SIZE != 0 help Say Y here if you intend to execute your compressed kernel image - (zImage) directly from ROM or flash. If unsure, say N. + (zImage) directly from ROM or flash. You need to specify the ROM start + address and size. The resulting zImage will still boot from RAM. + + If unsure, say N. config CMDLINE string "Default kernel command string"
arch/arm/boot/Makefile +1 -1
@@ -65,7 +65,7 @@ quiet_cmd_uimage = UIMAGE $@ -n 'Linux-$(KERNELRELEASE)' -d $< $@ ifeq ($(CONFIG_ZBOOT_ROM),y) -$(obj)/uImage: LOADADDR=$(CONFIG_ZBOOT_ROM_TEXT) +$(obj)/uImage: LOADADDR=$(CONFIG_ZBOOT_ROM_START) else $(obj)/uImage: LOADADDR=$(ZRELADDR) endif
arch/arm/boot/compressed/.gitignore +0 -1
@@ -1,3 +1,2 @@ font.c piggy.gz -vmlinux.lds
arch/arm/boot/compressed/Makefile +2 -25
@@ -48,25 +48,12 @@ else endif endif -# -# We now have a PIC decompressor implementation. Decompressors running -# from RAM should not define ZTEXTADDR. Decompressors running directly -# from ROM or Flash must define ZTEXTADDR (preferably via the config) -# FIXME: Previous assignment to ztextaddr-y is lost here. See SHARK -ifeq ($(CONFIG_ZBOOT_ROM),y) -ZTEXTADDR := $(CONFIG_ZBOOT_ROM_TEXT) -ZBSSADDR := $(CONFIG_ZBOOT_ROM_BSS) -else -ZTEXTADDR := 0 -ZBSSADDR := ALIGN(4) -endif - -SEDFLAGS = s/TEXT_START/$(ZTEXTADDR)/;s/BSS_START/$(ZBSSADDR)/ +EXTRA_CFLAGS := -fpic -fno-builtin suffix_$(CONFIG_KERNEL_GZIP) = gzip suffix_$(CONFIG_KERNEL_LZO) = lzo -targets := vmlinux vmlinux.lds \ +targets := vmlinux \ piggy.$(suffix_y) piggy.$(suffix_y).o \ font.o font.c head.o misc.o $(OBJS)
@@ -75,7 +62,6 @@ ORIG_CFLAGS := $(KBUILD_CFLAGS) KBUILD_CFLAGS = $(subst -pg, , $(ORIG_CFLAGS)) endif -EXTRA_CFLAGS := -fpic -fno-builtin EXTRA_AFLAGS := -Wa,-march=all # Supply ZRELADDR, INITRD_PHYS and PARAMS_PHYS to the decompressor via
@@ -106,10 +92,6 @@ lib1funcs = $(obj)/lib1funcs.o $(obj)/lib1funcs.S: $(srctree)/arch/$(SRCARCH)/lib/lib1funcs.S FORCE $(call cmd,shipped) -# Don't allow any static data in misc.o, which -# would otherwise mess up our GOT table -CFLAGS_misc.o := -Dstatic= - $(obj)/vmlinux: $(obj)/vmlinux.lds $(obj)/$(HEAD) $(obj)/piggy.$(suffix_y).o \ $(addprefix $(obj)/, $(OBJS)) $(lib1funcs) FORCE $(call if_changed,ld)
@@ -120,10 +102,5 @@ $(obj)/piggy.$(suffix_y): $(obj)/../Image FORCE $(obj)/piggy.$(suffix_y).o: $(obj)/piggy.$(suffix_y) FORCE -CFLAGS_font.o := -Dstatic= - $(obj)/font.c: $(FONTC) $(call cmd,shipped) - -$(obj)/vmlinux.lds: $(obj)/vmlinux.lds.in arch/arm/boot/Makefile .config - @sed "$(SEDFLAGS)" < $< > $@
arch/arm/boot/compressed/head.S +102 -71
239 lines not on this page · raw message
arch/arm/boot/compressed/misc.c +1 -2
@@ -23,7 +23,6 @@ unsigned int __machine_arch_type;#include <linux/compiler.h> /* for inline */ #include <linux/types.h> /* for size_t */ #include <linux/stddef.h> /* for NULL */ -#include <asm/string.h> #include <linux/linkage.h> #include <asm/unaligned.h>
@@ -166,7 +165,7 @@ void __memzero (__ptr_t s, size_t n) *u.ucp++ =0 ; } -static inline __ptr_t memcpy(__ptr_t __dest, __const __ptr_t __src, +static __ptr_t memcpy(__ptr_t __dest, __const __ptr_t __src,size_t __n) {int i =0 ;
arch/arm/boot/compressed/vmlinux.lds new file +71 -0
71 lines not on this page · raw message
arch/arm/boot/compressed/vmlinux.lds.in deleted +0 -61
61 lines not on this page · raw message
Signature · 1 line
1.6.4.4
PATCH 1/1 Clean up ARM compressed loader
Hello,
Quoted 11 lines · Hector Martin
The -Dstatic= stuff in the
Makefileis a hack and subtly breaks the current inflate implementation, because the fixed inflate tables are included into a function and removing static places them in the stack, which bloats the stack and breaks references after the function returns. So get rid of the hack.Instead, clean up the stub loader and unify it. The loader is now exactly the same regardless of whether you've enabled
CONFIG_ZBOOT_ROMor not, and runs from RAM in PIC mode.CONFIG_ZBOOT_ROMsimply enables an extra section of stub code that lets the zImage boot from ROM by copying itself to RAM first.
This is a big regression as this introduces two new copies. And as CONFIG_ZBOOT_ROM isn't primary about saving RAM at the early boot stages but for fast booting this is not acceptable.
Best regards
Uwe
Signature · 2 lines
Pengutronix e.K. | Uwe Kleine-K?nig | Industrial Linux Solutions | http://www.pengutronix.de/ |
PATCH 1/1 Clean up ARM compressed loader
Quoted 3 lines · Uwe Kleine-K?nig
This is a big regression as this introduces two new copies. And as
CONFIG_ZBOOT_ROMisn't primary about saving RAM at the early boot stages but for fast booting this is not acceptable.
I'm not seeing any significant performance decrease due to these optimized copies, compared to the time spent actually uncompressing the kernel. In fact, it's quite possible that this will speed things up on some implementations (particularly simpler/older/slower ones), as doing a block-based copy from ROM to RAM followed by a RAM->RAM inflate can very well be faster than uncompressing straight from ROM, if the ROM bus is slower. Inflate causes many small accesses.
If this can be shown to be a real performance issue, it can be fixed by:
1) not copying the actual payload to RAM, instead passing the ROM address to the decompressor (you still need a gaping hole in RAM since you can't relocate the other sections up to fill the hole, but you avoid the copy)
2) keeping track of the true uncompressed kernel size (not just some 4x expansion guess), so the loader can be placed directly afterwards, leaving space for the kernel and avoiding the final relocation.
I'll be happy to make these changes if it really makes a difference, but I'm not convinced right now.
It's worth noting that while debugging this I found a bug in the inflate stuff that causes inflate to be *much* slower for certain targets, including some ARM ones (yet nobody appears to have complained). I'll be submitting a fix for this to LKML, as it affects the generic deflate implementation.
Signature · 2 lines
Hector Martin (hector at marcansoft.com) Public Key: http://www.marcansoft.com/marcan.asc
PATCH 1/1 Clean up ARM compressed loader
In reply to Hector MartinQuoted 5 lines · Hector Martin
Instead, clean up the stub loader and unify it. The loader is now exactly the same regardless of whether you've enabled
CONFIG_ZBOOT_ROMor not, and runs from RAM in PIC mode.CONFIG_ZBOOT_ROMsimply enables an extra section of stub code that lets the zImage boot from ROM by copying itself to RAM first.
Unacceptable.
The point of putting the decompressor in ROM is to save the time spent copying it to RAM first. Boot loaders are already perfectly capable of copying the image into RAM and calling it there.
Getting the decompressor to do that step first just moves the copy from the boot loader into the decompressor. It doesn't eliminate it, which is what ROM-based decompression is all about.
Basically, this patch makes the feature entirely pointless.
PATCH 1/1 Clean up ARM compressed loader
Quoted 13 lines · Russell King - ARM Linux
Quoted 5 lines · Hector Martin
Instead, clean up the stub loader and unify it. The loader is now exactly the same regardless of whether you've enabled
CONFIG_ZBOOT_ROMor not, and runs from RAM in PIC mode.CONFIG_ZBOOT_ROMsimply enables an extra section of stub code that lets the zImage boot from ROM by copying itself to RAM first.Unacceptable.
The point of putting the decompressor in ROM is to save the time spent copying it to RAM first. Boot loaders are already perfectly capable of copying the image into RAM and calling it there.
Quoted 5 lines
Getting the decompressor to do that step first just moves the copy from the boot loader into the decompressor. It doesn't eliminate it, which is what ROM-based decompression is all about.
Basically, this patch makes the feature entirely pointless.
What about simply not compiling the decompressor with -fPIC when using ZBOOT_ROM=y? That would certainly solve the problem with the only restriction that such kernel images won't be bootable from RAM which is probably an acceptable compromize.
The idea with ZBOOT_ROM is also to have an empty .data section so there is no data segment to copy. This is achieved by not having any writable global variable nor static local variables (unless they're marked const of course). We probably should add an assert statement in the linker script to ensure this is the case.
Nicolas
PATCH 1/1 Clean up ARM compressed loader
Quoted 4 lines · Nicolas Pitre
What about simply not compiling the decompressor with -fPIC when using ZBOOT_ROM=y? That would certainly solve the problem with the only restriction that such kernel images won't be bootable from RAM which is probably an acceptable compromize.
Unfortunately, that doesn't solve the stack-bashing with ZBOOT_ROM=n.
I'm afraid that the commit adding LZO support is going to have to be reverted until a better solution to this is found.
PATCH 1/1 Clean up ARM compressed loader
Quoted 7 lines · Russell King - ARM Linux
Quoted 4 lines · Nicolas Pitre
What about simply not compiling the decompressor with -fPIC when using ZBOOT_ROM=y? That would certainly solve the problem with the only restriction that such kernel images won't be bootable from RAM which is probably an acceptable compromize.
Unfortunately, that doesn't solve the stack-bashing with ZBOOT_ROM=n.
Yes it does, that's exactly what my first version of the patch did. Once you get rid of the partial relocation used for ROM builds (with split text/bss) you don't need -Dstatic=, and once you get rid of that you solve the stack-bashing. The ROM build becomes a bog standard non-relocatable ROM image (with the usual LMA/VMA linker script stuff to copy initialized data to RAM), and the RAM build becomes a bog standard relocatable image (a single contiguous blob including text/rodata/data/bss) that doesn't suffer from any issues when you move it around.
Signature · 2 lines
Hector Martin (hector at marcansoft.com) Public Key: http://www.marcansoft.com/marcan.asc
PATCH 1/1 Clean up ARM compressed loader
Quoted 9 lines · Hector Martin
Quoted 7 lines · Russell King - ARM Linux
Quoted 4 lines · Nicolas Pitre
What about simply not compiling the decompressor with -fPIC when using ZBOOT_ROM=y? That would certainly solve the problem with the only restriction that such kernel images won't be bootable from RAM which is probably an acceptable compromize.
Unfortunately, that doesn't solve the stack-bashing with ZBOOT_ROM=n.
Quoted 9 lines
Yes it does, that's exactly what my first version of the patch did. Once you get rid of the partial relocation used for ROM builds (with split text/bss) you don't need -Dstatic=, and once you get rid of that you solve the stack-bashing. The ROM build becomes a bog standard non-relocatable ROM image (with the usual LMA/VMA linker script stuff to copy initialized data to RAM), and the RAM build becomes a bog standard relocatable image (a single contiguous blob including text/rodata/data/bss) that doesn't suffer from any issues when you move it around.
Did you bother to read my previous reply explaining that this is not a hack for the toolchain? It sounds to me like you didn't.
PATCH 1/1 Clean up ARM compressed loader
Quoted 18 lines · Russell King - ARM Linux
Quoted 7 lines · Hector Martin
Quoted 6 lines · Russell King - ARM Linux
Quoted 4 lines · Nicolas Pitre
What about simply not compiling the decompressor with -fPIC when using ZBOOT_ROM=y? That would certainly solve the problem with the only restriction that such kernel images won't be bootable from RAM which is probably an acceptable compromize.
Unfortunately, that doesn't solve the stack-bashing with ZBOOT_ROM=n.
Quoted 9 lines
Yes it does, that's exactly what my first version of the patch did. Once you get rid of the partial relocation used for ROM builds (with split text/bss) you don't need -Dstatic=, and once you get rid of that you solve the stack-bashing. The ROM build becomes a bog standard non-relocatable ROM image (with the usual LMA/VMA linker script stuff to copy initialized data to RAM), and the RAM build becomes a bog standard relocatable image (a single contiguous blob including text/rodata/data/bss) that doesn't suffer from any issues when you move it around.
Quoted 2 lines
Did you bother to read my previous reply explaining that this is not a hack for the toolchain? It sounds to me like you didn't.
Are you new to C? It looks like you are seriously still claiming that -Dstatic= is A-OK under all circumstances. It very obviously isn't.
In your other e-mail, you're missing the point. You're arguing that compiler behavior regarding access to data (particularly that with static vs. global visibility) is well-defined enough that you can guarantee that GOT-relative accesses will not happen as long as you avoid static data. I'm still not convinced, but that's not my main point.
The problem is that -Dstatic= isn't a magical safe way of making objects globally visible, because the 'static' keyword in C means different things in different places. Specifically, it means a totally different thing inside functions. Instead of turning a static file-global array into a globally visible array, you're turning a static (as in non-stack-allocated) array inside a function into a stack-allocated array. And as no doubt you're aware, stack objects become undefined once the function exits, and references to them are invalid. Welcome to our inflate heisenbug:
static void zlib_fixedtables(struct inflate_state *state)
{
# include "inffixed.h" /* static blah {len,dist}fix[] = {...}; */
state->lencode = lenfix;
state->lenbits = 9;
state->distcode = distfix;
state->distbits = 5;}
Guess what happens when you remove the static keywords there.
Long story short, -Dstatic= isn't accomplishing what you want, because it's crude and doesn't handle all cases properly. If you want to keep the current model, you have to be willing to manually watch changes to the decompressor implementations so they do not violate your assumptions (i.e. never use static inside functions and rely on it, at least), because -Dstatic= will break valid C code. Under _all_ circumstances, no matter what your other compiler flags are or how you link and load stuff, -Dstatic= is guaranteed to break some valid C code, and that's what's happening here.
Sure, you could patch up the issue by moving the #include out of the function above, but then you'd better be willing to watch for anything like that anywhere in decompressor code used for ARM, because other kernel developers aren't going to know that ARM silently discards their static keywords. I'm not even sure this is the only broken code so far, there could be more.
Signature · 2 lines
Hector Martin (hector at marcansoft.com) Public Key: http://www.marcansoft.com/marcan.asc
PATCH 1/1 Clean up ARM compressed loader
Quoted 19 lines · Hector Martin
Quoted 18 lines · Russell King - ARM Linux
Quoted 16 lines · Hector Martin
Quoted 6 lines · Russell King - ARM Linux
Quoted 4 lines · Nicolas Pitre
What about simply not compiling the decompressor with -fPIC when using ZBOOT_ROM=y? That would certainly solve the problem with the only restriction that such kernel images won't be bootable from RAM which is probably an acceptable compromize.
Unfortunately, that doesn't solve the stack-bashing with ZBOOT_ROM=n.
Yes it does, that's exactly what my first version of the patch did. Once you get rid of the partial relocation used for ROM builds (with split text/bss) you don't need -Dstatic=, and once you get rid of that you solve the stack-bashing. The ROM build becomes a bog standard non-relocatable ROM image (with the usual LMA/VMA linker script stuff to copy initialized data to RAM), and the RAM build becomes a bog standard relocatable image (a single contiguous blob including text/rodata/data/bss) that doesn't suffer from any issues when you move it around.
Quoted 5 lines
Quoted 2 lines
Did you bother to read my previous reply explaining that this is not a hack for the toolchain? It sounds to me like you didn't.
Are you new to C? It looks like you are seriously still claiming that -Dstatic= is A-OK under all circumstances. It very obviously isn't.
I didn't see Russell make any such claim anywhere. We're talking about a very particular case here.
Quoted 10 lines
In your other e-mail, you're missing the point. You're arguing that compiler behavior regarding access to data (particularly that with static vs. global visibility) is well-defined enough that you can guarantee that GOT-relative accesses will not happen as long as you avoid static data. I'm still not convinced, but that's not my main point.
The problem is that -Dstatic= isn't a magical safe way of making objects globally visible, because the 'static' keyword in C means different things in different places. Specifically, it means a totally different thing inside functions. [...]
Yes we all know that.
Quoted 17 lines
static void
zlib_fixedtables(struct inflate_state *state){
# include "
inffixed.h" /* static blah {len,dist}fix[] = {...}; */state->lencode = lenfix; state->lenbits = 9; state->distcode = distfix; state->distbits = 5;}
Guess what happens when you remove the static keywords there.
Quoted 9 lines
Long story short, -Dstatic= isn't accomplishing what you want, because it's crude and doesn't handle all cases properly. If you want to keep the current model, you have to be willing to manually watch changes to the decompressor implementations so they do not violate your assumptions (i.e. never use static inside functions and rely on it, at least), because -Dstatic= will break valid C code. Under _all_ circumstances, no matter what your other compiler flags are or how you link and load stuff, -Dstatic= is guaranteed to break some valid C code, and that's what's happening here.
Obviously. But that's the core of the argument: we made sure that the previous implementation didn't use any static within functions. Why do you need to do that? This is usually a bad idea anyway as this makes the function non reentrant.
Quoted 6 lines
Sure, you could patch up the issue by moving the #include out of the function above, but then you'd better be willing to watch for anything like that anywhere in decompressor code used for ARM, because other kernel developers aren't going to know that ARM silently discards their static keywords. I'm not even sure this is the only broken code so far, there could be more.
Two possibilities that I can see:
Put this in a common header file for the decompressor code:
/* ARM wants to redefine this sometimes */
#ifndef static_func
#define static_func
#endif
Then declare your functions with static_func, and ARM can use -Dstatic_func.
Or, we give up the ability to load in RAM a ROMable zImage.
Personally I'd favor first solution.
Nicolas
PATCH 1/1 Clean up ARM compressed loader
Quoted 6 lines · Nicolas Pitre
Put this in a common header file for the decompressor code:
/* ARM wants to redefine this sometimes */
#ifndefstatic_func
#definestatic_func
#endif
Obviously I meant
#define static_func static
Nicolas
PATCH 1/1 Clean up ARM compressed loader
In reply to Nicolas PitreQuoted 4 lines · Nicolas Pitre
Obviously. But that's the core of the argument: we made sure that the previous implementation didn't use any static within functions. Why do you need to do that? This is usually a bad idea anyway as this makes the function non reentrant.
Indeed - but having static data within a function doesn't make it non- reentrant - having static data modified by the function can make it non-reentrant.
I wonder if we can get around this problem a slightly different way - instead of building the decompressor into misc.c, build it as a separate PIC object without the -Dstatic= stuff, and verify that it contains no read-write data.
Since these decompressors are designed to be used within the multi-threaded environment of the kernel, they shouldn't have any read-write data within them (or if they do, they'd better use some locking - which would make them unusable for decompressors.)
Then we can tailor our misc.c not to use static data (which is the real problem - not the functions being static), and the -Dstatic= hack will go away.
PATCH 1/1 Clean up ARM compressed loader
Quoted 4 lines · Russell King - ARM Linux
I wonder if we can get around this problem a slightly different way - instead of building the decompressor into
misc.c, build it as a separate PIC object without the -Dstatic= stuff, and verify that it contains no read-write data.
You're still going to have issues if it uses read-only data and accesses it via an offset from the GOT, though.
Quoted 3 lines
Then we can tailor our
misc.cnot to use static data (which is the real problem - not the functions being static), and the -Dstatic= hack will go away.
If you can programmatically ensure that the decompressors do not use anything but GOT entries to access data, that's certainly a better solution than what we have now (I don't mean actually "fixing" them as in -Dstatic=, just something that makes the build fail if this is not the case so someone else can go and make the required changes). Maybe building the object and then grepping the relocations for offenders, or something like that.
However, considering my second patch can be adapted to avoid the two extra copies, I still think that's a cleaner solution. All I have to do is move things around in the linker script so the piggy is at the end of the zImage, avoid copying that the first time around (and instead pass the old pointer to ROM to the decompressor), and introduce the real decompressed uncompressed kernel size to the mix so I can relocate the decompressor to ZRELADDR+kernel_size (and fix the relocation check code so it uses the real size, not some 4x guess), which would avoid the second copy.
Signature · 2 lines
Hector Martin (hector at marcansoft.com) Public Key: http://www.marcansoft.com/marcan.asc
PATCH 1/1 Clean up ARM compressed loader
Quoted 7 lines · Hector Martin
If you can programmatically ensure that the decompressors do not use anything but GOT entries to access data, that's certainly a better solution than what we have now (I don't mean actually "fixing" them as in -Dstatic=, just something that makes the build fail if this is not the case so someone else can go and make the required changes). Maybe building the object and then grepping the relocations for offenders, or something like that.
Russell made a patch which simply ends up discarding .data section.
Quoted 9 lines
However, considering my second patch can be adapted to avoid the two extra copies, I still think that's a cleaner solution. All I have to do is move things around in the linker script so the piggy is at the end of the zImage, avoid copying that the first time around (and instead pass the old pointer to ROM to the decompressor), and introduce the real decompressed uncompressed kernel size to the mix so I can relocate the decompressor to ZRELADDR+kernel_size (and fix the relocation check code so it uses the real size, not some 4x guess), which would avoid the second copy.
You are still copying the actual decompressor code which is less optimal than what we already have. The current solution, as I explained already, involves _zero_ copying.
Nicolas
PATCH 1/1 Clean up ARM compressed loader
Quoted 11 lines · Nicolas Pitre
Quoted 7 lines · Hector Martin
If you can programmatically ensure that the decompressors do not use anything but GOT entries to access data, that's certainly a better solution than what we have now (I don't mean actually "fixing" them as in -Dstatic=, just something that makes the build fail if this is not the case so someone else can go and make the required changes). Maybe building the object and then grepping the relocations for offenders, or something like that.
Russell made a patch which simply ends up discarding .data section.
You still have the issue of the .bss section. Though I must admit this would be rare, a decompressor using the .bss section would still cause silent breakage. And I'm still not convinced that compiler behavior is defined such that this cannot break in the future.
Quoted 3 lines
You are still copying the actual decompressor code which is less optimal than what we already have. The current solution, as I explained already, involves _zero_ copying.
The overhead of copying the decompressor is negligible. We're talking somewhere along the lines of 0.1% of the time spent inside the decompressor, worst case.
Signature · 2 lines
Hector Martin (hector at marcansoft.com) Public Key: http://www.marcansoft.com/marcan.asc
PATCH 1/1 Clean up ARM compressed loader
Quoted 11 lines · Hector Martin
Quoted 10 lines · Nicolas Pitre
Quoted 7 lines · Hector Martin
If you can programmatically ensure that the decompressors do not use anything but GOT entries to access data, that's certainly a better solution than what we have now (I don't mean actually "fixing" them as in -Dstatic=, just something that makes the build fail if this is not the case so someone else can go and make the required changes). Maybe building the object and then grepping the relocations for offenders, or something like that.
Quoted 5 lines
Quoted 1 line
Russell made a patch which simply ends up discarding .data section.
You still have the issue of the .bss section. Though I must admit this would be rare, a decompressor using the .bss section would still cause silent breakage.
What breakage? We do use the .bss right now. But .bss can be fixed to some absolute address in RAM with no further handling besides clearing
I found the following at the top of arch/arm/boot/compressed/misc.c:
* Nicolas Pitre <[email protected]> 1999/04/14 :
* For this code to run directly from Flash, all constant variables must * be marked with 'const' and all other variables initialized at run-time * only. This way all non constant variables will end up in the bss segment, * which should point to addresses in RAM and cleared to 0 on start. * This allows for a much quicker boot time.
Quoted 2 lines
And I'm still not convinced that compiler behavior is defined such that this cannot break in the future.
Seems to me that we're relying on compiler conventions that have been around for decades already. What do you expect to break?
Quoted 7 lines
Quoted 3 lines
You are still copying the actual decompressor code which is less optimal than what we already have. The current solution, as I explained already, involves _zero_ copying.
The overhead of copying the decompressor is negligible. We're talking somewhere along the lines of 0.1% of the time spent inside the decompressor, worst case.
Possible. And granted this whole scheme was elaborated in 1999 when CPUs were much slower and the kernel much smaller relative to the decompressor, etc. But again that "just works" already, without the copy overhead, and the code needed to make the copy (I know it's only 4 asm instructions), and the complexity needed to care about the actual decompressed kernel size, etc. etc.
Nicolas
PATCH 1/1 Clean up ARM compressed loader
Quoted 9 lines · Nicolas Pitre
Quoted 5 lines · Hector Martin
Quoted 1 line · Nicolas Pitre
Russell made a patch which simply ends up discarding .data section.
You still have the issue of the .bss section. Though I must admit this would be rare, a decompressor using the .bss section would still cause silent breakage.
What breakage? We do use the .bss right now. But .bss can be fixed to some absolute address in RAM with no further handling besides clearing
static data in BSS can end up addressed with a GOT-relative offset (even if you put it at "some absolute address" - this is still PIC code). That
will fail once you move around the text/rodata/GOT blob when running
the ROM code from RAM, since it'll push the BSS section somewhere else (potentially not RAM). So you still need to ensure that all BSS data is globally visible.
Quoted 8 lines
I found the following at the top of
arch/arm/boot/compressed/misc.c:* Nicolas Pitre <[email protected]> 1999/04/14 :
* For this code to run directly from Flash, all constant variables must * be marked with 'const' and all other variables initialized at run-time * only. This way all non constant variables will end up in the bss segment, * which should point to addresses in RAM and cleared to 0 on start. * This allows for a much quicker boot time.
This is a different issue, the part where so far the code hasn't been copying data. This is unrelated to the issue that -Dstatic= tries to solve, which has to do with the way data can be located relative to the GOT.
Quoted 5 lines
Quoted 2 lines
And I'm still not convinced that compiler behavior is defined such that this cannot break in the future.
Seems to me that we're relying on compiler conventions that have been around for decades already. What do you expect to break?
Compilers are getting smarter, and code relying on old assumptions that may no longer be true (because compilers optimize better or do things differently) is precisely that which is most liable to break. The only way you can properly guarantee things will work is by sticking to actual standards and doing what the compiler expects you to do. Moving around text/rodata without moving around bss/(maybe data) isn't something allowed by the compiler. Sure, you can hack around it like we're doing here, but you're relying on your hacks correctly preventing the compiler from doing something valid yet which breaks your assumptions.
Quoted 6 lines
Possible. And granted this whole scheme was elaborated in 1999 when CPUs were much slower and the kernel much smaller relative to the decompressor, etc. But again that "just works" already, without the copy overhead, and the code needed to make the copy (I know it's only 4 asm instructions), and the complexity needed to care about the actual decompressed kernel size, etc. etc.
Well, if you're into speed anyway, caring about the actual kernel size is desirable because it avoids a totally useless copy when the zImage is loaded into RAM somewhere near the end of where the kernel would be. The guesstimate in the code will think it'll overlap and move things around unnecessarily. Worse, if the kernel ends up being highly compressible for whatever reason, it could overflow the 4x guesstimate and make the whole compressed loader fail again if it happens to be around there.
But this discussion is rather pointless because at this stage I highly doubt Russell is going to change his mind.
Signature · 2 lines
Hector Martin (hector at marcansoft.com) Public Key: http://www.marcansoft.com/marcan.asc
PATCH 1/1 Clean up ARM compressed loader
Quoted 9 lines · Hector Martin
Compilers are getting smarter, and code relying on old assumptions that may no longer be true (because compilers optimize better or do things differently) is precisely that which is most liable to break. The only way you can properly guarantee things will work is by sticking to actual standards and doing what the compiler expects you to do. Moving around text/rodata without moving around bss/(maybe data) isn't something allowed by the compiler. Sure, you can hack around it like we're doing here, but you're relying on your hacks correctly preventing the compiler from doing something valid yet which breaks your assumptions.
Sure. And so far we've used a naive hack with -Dstatic. Russell's patch is now removing the hack and making sure the restrictions we count on are respected wrt .data section. If we need to ensure no static .bss variable are used we might just add this in the Makefile:
[ "$(nm compressed/vmlinux | grep " b ")" != "" ]
Quoted 7 lines
Well, if you're into speed anyway, caring about the actual kernel size is desirable because it avoids a totally useless copy when the zImage is loaded into RAM somewhere near the end of where the kernel would be. The guesstimate in the code will think it'll overlap and move things around unnecessarily. Worse, if the kernel ends up being highly compressible for whatever reason, it could overflow the 4x guesstimate and make the whole compressed loader fail again if it happens to be around there.
Those are valid points, and rather orthogonal to the ROMable decompressor issue. Better not to mix them together.
And I vaguely remember trying to do just that a couple years ago:
getting hold of the decompressed kernel size and feeding that into the decompressor code. I somehow didn't come up with a satisfactory solution and never revisited it since.
Quoted 2 lines
But this discussion is rather pointless because at this stage I highly doubt Russell is going to change his mind.
If you produce a patch for the above, and it is clean and portable, then I suspect he would just merge it. Only at that point we could reconsider the ROMable decompressor issue.
Nicolas
PATCH 1/1 Clean up ARM compressed loader
Quoted 7 lines · Nicolas Pitre
Quoted 2 lines · Hector Martin
But this discussion is rather pointless because at this stage I highly doubt Russell is going to change his mind.
If you produce a patch for the above, and it is clean and portable, then I suspect he would just merge it. Only at that point we could reconsider the ROMable decompressor issue.
Hector's arguments have been based around denying various well defined features of the ELF shared library support, and it would be wreckless of me to accept changes based upon denials.
PATCH 1/1 Clean up ARM compressed loader
Quoted 5 lines · Russell King - ARM Linux
Quoted 4 lines · Nicolas Pitre
Quoted 2 lines · Hector Martin
But this discussion is rather pointless because at this stage I highly doubt Russell is going to change his mind.
Quoted 7 lines
Quoted 3 lines
If you produce a patch for the above, and it is clean and portable, then I suspect he would just merge it. Only at that point we could reconsider the ROMable decompressor issue.
Hector's arguments have been based around denying various well defined features of the ELF shared library support, and it would be wreckless of me to accept changes based upon denials.
I think we've pretty much settled this now.
It is however true that the decompressor code doesn't have the decompressed kernel size and that a 4x expansion is assumed which is not optimal. This alone, without mixing it up with other issues, should be sufficient for considering an eventual patch from Hector if his fix is well done. That's what I was referring to above, and I think the ROMable decompressor structure change can be postponed until real issues with it actually arise.
Nicolas
PATCH 1/1 Clean up ARM compressed loader
In reply to Nicolas PitreQuoted 6 lines · Nicolas Pitre
Quoted 2 lines · Hector Martin
And I'm still not convinced that compiler behavior is defined such that this cannot break in the future.
Seems to me that we're relying on compiler conventions that have been around for decades already. What do you expect to break?
Sigh. It's well defined behaviour - it's part of the ELF ABI. There's nothing undefined or spurious about it, and it's not going away any time soon.
For example, my C library has this:
00ca5d60 B __environ
as part of it's BSS segment. If I do:
#include <stdio.h>
char **__environ;
int main() {
printf("&__environ = %p\n", &__environ);
printf("__environ = %p\n", __environ);
printf("__environ[0] = %s\n", __environ[0]);
return 0;
}I get:
080496a0 B __environ
as part of the program. So, as soon as I run this program, there exists two '__environ' variables - one in the program and one in the C library.
Let's see what happens when I run the program:
&__environ = 0x80496a0__environ = 0xbfa01c1c
__environ[0] = HOSTNAME=rmk-PC
That's definitely the one in the application program. How can the program, which clearly has this variable as part of its BSS segment independent of the C library, and which never initializes this BSS variable, end up with data in there?
It's well defined behaviour brought about by global data and the GOT, where the shared library is pointed to the version in the user program rather than its own BSS - independent of the rest of the shared library's text, data, and bss segments. When the shared library reads or writes '__environ', it does it via the GOT which is redirected _at run time_ to point at the application binary copy by the dynamic linker.
And now try changing that 'char **__environ' to be extern, and see what effect that has:
080496e0 B __environ@@GLIBC_2.0
Yes, still part of the application binary (this time versioned with the C library version), and the C library still finds that instead of its own copy.
But, build an application binary which doesn't reference __environ, and the version in the C library's BSS segment will be used instead.
This is not undefined compiler behaviour. This is well defined behaviour, and it's precisely this behaviour that we're using. We're directing the *global* data references to some other part of memory independent of the rest of the decompressor image.
So, rather than wasting my time with this crap, can we please move on?
I guess I could've pulled several people's trees during the time it's taken to write this email which shouldn't have been necessary.
PATCH 1/1 Clean up ARM compressed loader
In reply to Russell King - ARM LinuxQuoted 3 lines · Russell King - ARM Linux
Then we can tailor our
misc.cnot to use static data (which is the real problem - not the functions being static), and the -Dstatic= hack will go away.
And here's the patch - not run tested yet, but compile-tested for gzip and LZO in both ROM=y and ROM=n modes.
ARM: Eliminate decompressor -Dstatic= PIC hack
We used to build decompressors with -Dstatic= to avoid any local data being generated. The problem is that local data generates GOTOFF relocations, which means we can't relocate the data relative to the text segment.
Global data, on the other hand, goes through the GOT, and can be relocated anywhere.
Unfortunately, with the new decompressors, this presents a problem since they declare static data within functions, and this leads to stack overflow.
Fix this by separating out the decompressor code into a separate file, and removing 'static' from BSS data in misc.c.
Also, discard the .data section - this means that should we end up with read/write initialized data, the decompressor will fail to link and the problem will be obvious.
arch/arm/boot/compressed/Makefile | 6 +-- arch/arm/boot/compressed/decompress.c | 45 +++++++++++++ arch/arm/boot/compressed/misc.c | 109 +++--------------------------- arch/arm/boot/compressed/vmlinux.lds.in | 2 +- 4 files changed, 58 insertions(+), 104 deletions(-)
arch/arm/boot/compressed/Makefile +1 -5
@@ -5,7 +5,7 @@ # HEAD = head.o -OBJS = misc.o +OBJS = misc.o decompress.o FONTC = $(srctree)/drivers/video/console/font_acorn_8x8.c #
@@ -106,10 +106,6 @@ lib1funcs = $(obj)/lib1funcs.o $(obj)/lib1funcs.S: $(srctree)/arch/$(SRCARCH)/lib/lib1funcs.S FORCE $(call cmd,shipped) -# Don't allow any static data in misc.o, which -# would otherwise mess up our GOT table -CFLAGS_misc.o := -Dstatic= - $(obj)/vmlinux: $(obj)/vmlinux.lds $(obj)/$(HEAD) $(obj)/piggy.$(suffix_y).o \ $(addprefix $(obj)/, $(OBJS)) $(lib1funcs) FORCE $(call if_changed,ld)
arch/arm/boot/compressed/decompress.c new file +45 -0
45 lines not on this page · raw message
arch/arm/boot/compressed/misc.c +11 -98
149 lines not on this page · raw message
arch/arm/boot/compressed/vmlinux.lds.in +1 -1
@@ -14,6 +14,7 @@ SECTIONS /DISCARD/ : { *(.ARM.exidx*) *(.ARM.extab*) + *(.data) } . = TEXT_START;
@@ -40,7 +41,6 @@ SECTIONS .got : { *(.got) } _got_end = .; .got.plt : { *(.got.plt) } - .data : { *(.data) } _edata = .; . = BSS_START;
PATCH 1/1 Clean up ARM compressed loader
Quoted 17 lines · Russell King - ARM Linux
And here's the patch - not run tested yet, but compile-tested for gzip and LZO in both ROM=y and ROM=n modes.
ARM: Eliminate decompressor -Dstatic= PIC hack
We used to build decompressors with -Dstatic= to avoid any local data being generated. The problem is that local data generates GOTOFF relocations, which means we can't relocate the data relative to the text segment.
Global data, on the other hand, goes through the GOT, and can be relocated anywhere.
Unfortunately, with the new decompressors, this presents a problem since they declare static data within functions, and this leads to stack overflow.
Quoted 8 lines
Fix this by separating out the decompressor code into a separate file, and removing 'static' from BSS data in
misc.c.Also, discard the .data section - this means that should we end up with read/write initialized data, the decompressor will fail to link and the problem will be obvious.
Signed-off-by Russell King <[email protected]>
However...
arch/arm/boot/compressed/vmlinux.lds.in SECTIONS · line 18
Quoted 6 lines
arch/arm/boot/compressed/vmlinux.lds.in +1
-0@@ -14,6 +14,7 @@ SECTIONS /DISCARD/ : { *(.ARM.exidx*) *(.ARM.extab*) + *(.data) }
I'd put a comment there, otherwise lots of people will go WTF when coming across this.
Nicolas
PATCH 1/1 Clean up ARM compressed loader
Quoted 2 lines · Nicolas Pitre
I'd put a comment there, otherwise lots of people will go WTF when coming across this.
Added, thanks.
PATCH 1/1 Clean up ARM compressed loader
In reply to Nicolas PitreQuoted 11 lines · Nicolas Pitre
However...
Review SECTIONS · line 18
Quoted 6 lines
(quoted diff) +1
-0@@ -14,6 +14,7 @@ SECTIONS /DISCARD/ : { *(.ARM.exidx*) *(.ARM.extab*) + *(.data) }I'd put a comment there, otherwise lots of people will go WTF when coming across this.
Unfortunately, the decompressor in ROM mode has been subtly broken for a while - since 2552fc2.
This assumes 'sp' is the end of the image, which for ROM=n, is fine. However, for ROM=y, this most definitely is wrong, so this patch is needed.
arch/arm/boot/compressed/head.S +24 -26
@@ -170,8 +170,8 @@ not_angel: .text adr r0, LC0 - ARM( ldmia r0, {r1, r2, r3, r4, r5, r6, ip, sp} ) - THUMB( ldmia r0, {r1, r2, r3, r4, r5, r6, ip} ) + ARM( ldmia r0, {r1, r2, r3, r4, r5, r6, r11, ip, sp}) + THUMB( ldmia r0, {r1, r2, r3, r4, r5, r6, r11, ip} ) THUMB( ldr sp, [r0, #28] ) subs r0, r0, r1 @ calculate the delta offset
@@ -182,12 +182,14 @@ not_angel: /* * We're running at a different address. We need to fix * up various pointers: - * r5 - zImage base address - * r6 - GOT start + * r5 - zImage base address (_start) + * r6 - end of zImage text (_etext) + * r11 - GOT start * ip - GOT end */ add r5, r5, r0 add r6, r6, r0 + add r11, r11, r0 add ip, ip, r0 #ifndef CONFIG_ZBOOT_ROM
@@ -205,10 +207,10 @@ not_angel: /* * Relocate all entries in the GOT table. */ -1: ldr r1, [r6, #0] @ relocate entries in the GOT +1: ldr r1, [r11, #0] @ relocate entries in the GOT add r1, r1, r0 @ table. This fixes up the - str r1, [r6], #4 @ C references. - cmp r6, ip + str r1, [r11], #4 @ C references. + cmp r11, ip blo 1b #else
@@ -216,12 +218,12 @@ not_angel: * Relocate entries in the GOT table. We only relocate * the entries that are outside the (relocated) BSS region. */ -1: ldr r1, [r6, #0] @ relocate entries in the GOT +1: ldr r1, [r11, #0] @ relocate entries in the GOT cmp r1, r2 @ entry < bss_start || cmphs r3, r1 @ _end < entry addlo r1, r1, r0 @ table. This fixes up the - str r1, [r6], #4 @ C references. - cmp r6, ip + str r1, [r11], #4 @ C references. + cmp r11, ip blo 1b #endif
@@ -247,6 +249,7 @@ not_relocated: mov r0, #0 * Check to see if we will overwrite ourselves. * r4 = final kernel address * r5 = start of this image + * r6 = end of this image's text segment * r2 = end of malloc space (and therefore this image) * We basically want: * r4 >= r2 -> OK
@@ -254,8 +257,8 @@ not_relocated: mov r0, #0 */ cmp r4, r2 bhs wont_overwrite - sub r3, sp, r5 @ > compressed kernel size - add r0, r4, r3, lsl #2 @ allow for 4x expansion + sub r3, r6, r5 @ (_etext-_start) =~ compressed + add r0, r4, r3, lsl #2 @ kernel size, allow for 4x expansion cmp r0, r5 bls wont_overwrite
@@ -271,7 +274,6 @@ not_relocated: mov r0, #0 * r1-r3 = unused * r4 = kernel execution address * r5 = decompressed kernel start - * r6 = processor ID * r7 = architecture ID * r8 = atags pointer * r9-r12,r14 = corrupted
@@ -312,7 +314,8 @@ LC0: .word LC0 @ r1 .word _end @ r3 .word zreladdr @ r4 .word _start @ r5 - .word _got_start @ r6 + .word _end @ r6 + .word _got_start @ r11 .word _got_end @ ip .word user_stack+4096 @ sp LC1: .word reloc_end - reloc_start
@@ -336,7 +339,6 @@ params: ldr r0, =params_phys * * On entry, * r4 = kernel execution address - * r6 = processor ID * r7 = architecture number * r8 = atags pointer * r9 = run-time address of "start" (???)
@@ -542,7 +544,6 @@ __common_mmu_cache_on: * r1-r3 = unused * r4 = kernel execution address * r5 = decompressed kernel start - * r6 = processor ID * r7 = architecture ID * r8 = atags pointer * r9-r12,r14 = corrupted
@@ -581,19 +582,19 @@ call_kernel: bl cache_clean_flush * r1 = corrupted * r2 = corrupted * r3 = block offset - * r6 = corrupted + * r9 = corrupted * r12 = corrupted */ call_cache_fn: adr r12, proc_types #ifdef CONFIG_CPU_CP15 - mrc p15, 0, r6, c0, c0 @ get processor ID + mrc p15, 0, r9, c0, c0 @ get processor ID #else - ldr r6, =CONFIG_PROCESSOR_ID + ldr r9, =CONFIG_PROCESSOR_ID #endif 1: ldr r1, [r12, #0] @ get value ldr r2, [r12, #4] @ get mask - eor r1, r1, r6 @ (real ^ match) + eor r1, r1, r9 @ (real ^ match) tst r1, r2 @ & mask ARM( addeq pc, r12, r3 ) @ call cache function THUMB( addeq r12, r3 )
@@ -778,8 +779,7 @@ proc_types: * Turn off the Cache and MMU. ARMv3 does not support * reading the control register, but ARMv4 does. * - * On entry, r6 = processor ID - * On exit, r0, r1, r2, r3, r12 corrupted + * On exit, r0, r1, r2, r3, r9, r12 corrupted * This routine must preserve: r4, r6, r7 */ .align 5
@@ -852,10 +852,8 @@ __armv3_mmu_cache_off: /* * Clean and flush the cache to maintain consistency. * - * On entry, - * r6 = processor ID * On exit, - * r1, r2, r3, r11, r12 corrupted + * r1, r2, r3, r9, r11, r12 corrupted * This routine must preserve: * r0, r4, r5, r6, r7 */
@@ -967,7 +965,7 @@ __armv4_mmu_cache_flush: mov r2, #64*1024 @ default: 32K dcache size (*2) mov r11, #32 @ default: 32 byte line size mrc p15, 0, r3, c0, c0, 1 @ read cache type - teq r3, r6 @ cache ID register present? + teq r3, r9 @ cache ID register present? beq no_cache_id mov r1, r3, lsr #18 and r1, r1, #7
PATCH 1/1 Clean up ARM compressed loader
arch/arm/boot/compressed/head.S LC0: .word LC0 @ r1 · line 317
Quoted 2 lines · Russell King - ARM Linux
arch/arm/boot/compressed/head.S +1
-1- .word _got_start @ r6 + .word _end @ r6
Grr, that should be _etext.
PATCH 1/1 Clean up ARM compressed loader
In reply to Nicolas PitreQuoted 32 lines · Nicolas Pitre
Quoted 20 lines · Hector Martin
Quoted 17 lines · Russell King - ARM Linux
Quoted 7 lines · Hector Martin
Quoted 6 lines · Russell King - ARM Linux
Quoted 4 lines · Nicolas Pitre
What about simply not compiling the decompressor with -fPIC when using ZBOOT_ROM=y? That would certainly solve the problem with the only restriction that such kernel images won't be bootable from RAM which is probably an acceptable compromize.
Unfortunately, that doesn't solve the stack-bashing with ZBOOT_ROM=n.
Quoted 9 lines
Yes it does, that's exactly what my first version of the patch did. Once you get rid of the partial relocation used for ROM builds (with split text/bss) you don't need -Dstatic=, and once you get rid of that you solve the stack-bashing. The ROM build becomes a bog standard non-relocatable ROM image (with the usual LMA/VMA linker script stuff to copy initialized data to RAM), and the RAM build becomes a bog standard relocatable image (a single contiguous blob including text/rodata/data/bss) that doesn't suffer from any issues when you move it around.
Quoted 2 lines
Did you bother to read my previous reply explaining that this is not a hack for the toolchain? It sounds to me like you didn't.
Quoted 9 lines
Long story short, -Dstatic= isn't accomplishing what you want, because it's crude and doesn't handle all cases properly. If you want to keep the current model, you have to be willing to manually watch changes to the decompressor implementations so they do not violate your assumptions (i.e. never use static inside functions and rely on it, at least), because -Dstatic= will break valid C code. Under _all_ circumstances, no matter what your other compiler flags are or how you link and load stuff, -Dstatic= is guaranteed to break some valid C code, and that's what's happening here.
Quoted 4 lines
Obviously. But that's the core of the argument: we made sure that the previous implementation didn't use any static within functions. Why do you need to do that? This is usually a bad idea anyway as this makes the function non reentrant.
Static inside functions doesn't imply being non-reentrant (any more than using global variables does). In this case the data arrays are also const so it doesn't matter. But even if they weren't it wouldn't imply non-reentrancy. For example, you could have the function generate the tables on the fly on the first call, and add a spinlock if you want it to be reentrant (in fact, I'm pretty sure the tables could optionally be generated on the fly in the original zlib version, though that part was removed for the kernel).
Quoted 10 lines
Two possibilities that I can see:
Put this in a common header file for the decompressor code:
/* ARM wants to redefine this sometimes */
#ifndefstatic_func
#definestatic_func
#endif
Quoted 4 lines
Then declare your functions with
static_func, and ARM can use -Dstatic_func.Or, we give up the ability to load in RAM a ROMable zImage.
I can also fix the second version of the patch to get rid of the two extra copies of the kernel.
Signature · 2 lines
Hector Martin (hector at marcansoft.com) Public Key: http://www.marcansoft.com/marcan.asc
PATCH 1/1 Clean up ARM compressed loader
Quoted 8 lines · Hector Martin
Quoted 4 lines · Nicolas Pitre
Obviously. But that's the core of the argument: we made sure that the previous implementation didn't use any static within functions. Why do you need to do that? This is usually a bad idea anyway as this makes the function non reentrant.
Static inside functions doesn't imply being non-reentrant (any more than using global variables does).
Usually it does, unless as you mention those are used read-only. I think we don't have to teach each other how a C compiler works, right?
But just to make sure we're on the same page: if you need a global variable, or a static variable within a function, it is preferable if they're truly read-only by marking them const. This has the advantage of not having to bother with .data relocation in RAM when booting a kernel directly from ROM as they get allocated to the .rodata section which is included in .text by default by the linker.
Nicolas
PATCH 1/1 Clean up ARM compressed loader
In reply to Hector MartinQuoted 19 lines · Hector Martin
Quoted 18 lines · Russell King - ARM Linux
Quoted 16 lines · Hector Martin
Quoted 6 lines · Russell King - ARM Linux
Quoted 4 lines · Nicolas Pitre
What about simply not compiling the decompressor with -fPIC when using ZBOOT_ROM=y? That would certainly solve the problem with the only restriction that such kernel images won't be bootable from RAM which is probably an acceptable compromize.
Unfortunately, that doesn't solve the stack-bashing with ZBOOT_ROM=n.
Yes it does, that's exactly what my first version of the patch did. Once you get rid of the partial relocation used for ROM builds (with split text/bss) you don't need -Dstatic=, and once you get rid of that you solve the stack-bashing. The ROM build becomes a bog standard non-relocatable ROM image (with the usual LMA/VMA linker script stuff to copy initialized data to RAM), and the RAM build becomes a bog standard relocatable image (a single contiguous blob including text/rodata/data/bss) that doesn't suffer from any issues when you move it around.
Quoted 5 lines
Quoted 2 lines
Did you bother to read my previous reply explaining that this is not a hack for the toolchain? It sounds to me like you didn't.
Are you new to C? It looks like you are seriously still claiming that -Dstatic= is A-OK under all circumstances. It very obviously isn't.
Clearly you didn't (a) read my mail and (b) understand it.
And if you're going to be that insulting, I see no point discussing this further with you.
PATCH 1/1 Clean up ARM compressed loader
In reply to Hector MartinQuoted 9 lines · Hector Martin
Quoted 7 lines · Russell King - ARM Linux
Quoted 4 lines · Nicolas Pitre
What about simply not compiling the decompressor with -fPIC when using ZBOOT_ROM=y? That would certainly solve the problem with the only restriction that such kernel images won't be bootable from RAM which is probably an acceptable compromize.
Unfortunately, that doesn't solve the stack-bashing with ZBOOT_ROM=n.
Quoted 9 lines
Yes it does, that's exactly what my first version of the patch did. Once you get rid of the partial relocation used for ROM builds (with split text/bss) you don't need -Dstatic=, and once you get rid of that you solve the stack-bashing. The ROM build becomes a bog standard non-relocatable ROM image (with the usual LMA/VMA linker script stuff to copy initialized data to RAM), and the RAM build becomes a bog standard relocatable image (a single contiguous blob including text/rodata/data/bss) that doesn't suffer from any issues when you move it around.
OK... Well, just to put things in perspective: the CONFIG_ZBOOT_ROM was invented by myself almost 12 years ago. No split LMA/VMA was needed because the code didn't use any writable global variable. The idea is to have the boot loader branch directly to the decompressor stored in flash (those were the days when NOR flash was more popular than NAND) and then the decompressor simply get the deflated data directly from ROM and stores the inflated kernel directly in RAM at its final address. No copy what so ever was involved, and with the CPU I and D caches enabled and burst mode properly configured in the flash you get the fastest boot possible. At least until I added support to do eXecute-In-Place from flash with the kernel that is (yes, I'm guilty for that as well). So that has to continue working otherwise it is a regression.
Then, Russell added the ability for the decompressor to be PIC. And he even managed for a ZBOOT_ROM kernel to just work if you decided to load it in RAM and execute it from there. But for that to work the -Dstatic hack was needed. Admitedly, this hack and the ability to run a ZBOOT_ROM kernel from anywhere in RAM is nice but not strictly needed. But no one could argue that this didn't work. Claiming that this is an abuse of the compiler is beside the point: it has worked for years and now _you_ are the one who wants to change the rules, not the GCC people.
Now I apparently missed something in the thread... What is this stack bashing about? Admitedly I didn't look at any of the patches (yet).
Nicolas