x86: Rely on toolchain for relocatable code
Subsystems (22)
cover x86: Rely on toolchain for relocatable code
From: Ard Biesheuvel <[email protected]>
The x86_64 port has a number of historical quirks that result in a reliance on toolchain features that are either poorly specified or basically implementation details of the toolchain:
- the 'kernel' C model implemented by the compiler is intended for position dependent code residing in the 'negative' 2 GiB of the virtual address space, but is used to create a position independent executable (for virtual KASLR);
- the 'kernel' C model has other properties that are not written down anywhere, and may therefore deviate between compilers and versions, which now includes the Rust compilers too (e.g., use %gs not %fs for per-CPU references);
- the relocation format used to perform the PIE relocation at boot is complicated and non-standard, as it deals with 3 types of displacements, including 32-bit negative displacements for RIP-relative per-CPU references that are not subject to relocation fixups (as they are places in a separate, disjoint address space);
- the relocation table is generated from static relocation metadata taken from the ELF input objects into the linker, and these describe the input not the output - relaxations or other linker tweaks may result in a mismatch between the two, and GNU ld and LLD display different behavior here;
- this disjoint per-CPU address space requires elaborate hacks in the linker script and the startup code;
- some of the startup code executes from a 1:1 mapping of memory, where RIP-relative references are mandatory, whereas RIP-relative per-CPU variable references can only work correctly from the kernel virtual mapping (as they need to wrap around from the negative 2 GiB space into the 0x0 based per-CPU region);
The reason for this odd situation wrt per-CPU variable addressing is the fact that we rely on the user-space TLS arrangement for per-task stack cookies, and this was implemented using a fixed offset of 40 bytes from %GS. If we bump the minimum GCC version to 8.1, we can switch to symbol based stack cookie references, allowing the same arrangement to be adopted as on other architectures, i.e., where the CPU register carries the per-CPU offset, and UP or boot-time per-CPU references point into the per-CPU load area directly (using an offset of 0x0).
With that out of the way, we can untangle this whole thing, and replace the bespoke tooling and relocation formats with ordinary, linker generated relocation tables, using the RELR format that reduces the memory footprint of the relocation table by 20x. The compilers can efficiently generate position independent code these days, without unnecessary indirections via the Global Object Table (GOT) except for a handful of special cases (see the KVM patch for an example where a GOT-based indirection is the best choice for pushing the absolute address of a symbol onto the stack in a position independent manner when there are no free GPRs)
It also brings us much closer to the ordinary PIE relocation model used for most of user space, which is therefore much better supported and less likely to create problems as we increase the range of compilers and linkers that need to be supported.
Tested on GCC 8 - 14 and Clang 15 - 17, using EFI and bare metal boot using a variety of entry points (decompressor, EFI stub, XenPV, PVH)
Ard Biesheuvel (28):
x86/pvh: Call C code via the kernel virtual mapping
Documentation: Bump minimum GCC version to 8.1 x86/tools: Use mmap() to simplify relocs host tool x86/boot: Permit GOTPCREL relocations for x86_64 builds
x86: Define the stack protector guard symbol explicitly x86/percpu: Get rid of absolute per-CPU variable placement scripts/kallsyms: Avoid 0x0 as the relative base scripts/kallsyms: Remove support for absolute per-CPU variables x86/tools: Remove special relocation handling for per-CPU variables x86/xen: Avoid relocatable quantities in Xen ELF notes x86/pvh: Avoid absolute symbol references in .head.text x86/pm-trace: Use RIP-relative accesses for .tracedata x86/kvm: Use RIP-relative addressing
x86/rethook: Use RIP-relative reference for return address x86/sync_core: Use RIP-relative addressing
x86/entry_64: Use RIP-relative addressing
x86/hibernate: Prefer RIP-relative accesses
x86/boot/64: Determine VA/PA offset before entering C code x86/boot/64: Avoid intentional absolute symbol references in
.head.text
x64/acpi: Use PIC-compatible references in wakeup_64.S x86/head: Use PIC-compatible symbol references in startup code
asm-generic: Treat PIC .data.rel.ro sections as .rodata tools/objtool: Mark generated sections as writable tools/objtool: Treat indirect ftrace calls as direct calls
x86: Use PIE codegen for the core kernel
x86/boot: Implement support for ELF RELA/RELR relocations x86/kernel: Switch to PIE linking for the core kernel x86/tools: Drop x86_64 support from 'relocs' tool
Documentation/admin-guide/README.rst | 2 +- Documentation/arch/x86/zero-page.rst | 3 +- Documentation/process/changes.rst | 2 +- arch/x86/Kconfig | 3 +- arch/x86/Makefile | 22 +- arch/x86/boot/Makefile | 1 + arch/x86/boot/compressed/Makefile | 2 +- arch/x86/boot/compressed/misc.c | 16 +- arch/x86/entry/calling.h | 9 +- arch/x86/entry/entry_64.S | 12 +- arch/x86/entry/vdso/Makefile | 1 + arch/x86/include/asm/desc.h | 1 - arch/x86/include/asm/init.h | 2 +- arch/x86/include/asm/percpu.h | 22 - arch/x86/include/asm/pm-trace.h | 4 +- arch/x86/include/asm/processor.h | 14 +- arch/x86/include/asm/setup.h | 3 +- arch/x86/include/asm/stackprotector.h | 4 - arch/x86/include/asm/sync_core.h | 3 +- arch/x86/include/uapi/asm/bootparam.h | 2 +- arch/x86/kernel/acpi/wakeup_64.S | 11 +- arch/x86/kernel/head64.c | 76 +++- arch/x86/kernel/head_64.S | 40 +- arch/x86/kernel/irq_64.c | 1 - arch/x86/kernel/kvm.c | 8 +- arch/x86/kernel/relocate_kernel_64.S | 6 +- arch/x86/kernel/rethook.c | 3 +- arch/x86/kernel/setup_percpu.c | 9 +- arch/x86/kernel/vmlinux.lds.S | 75 ++-- arch/x86/platform/pvh/head.S | 57 ++- arch/x86/power/hibernate_asm_64.S | 4 +- arch/x86/realmode/rm/Makefile | 1 + arch/x86/tools/Makefile | 2 +- arch/x86/tools/relocs.c | 425 +++----------------- arch/x86/tools/relocs.h | 11 +- arch/x86/tools/relocs_64.c | 18 - arch/x86/tools/relocs_common.c | 11 +- arch/x86/xen/xen-head.S | 16 +- drivers/base/power/trace.c | 6 +- drivers/firmware/efi/libstub/x86-stub.c | 2 + include/asm-generic/vmlinux.lds.h | 10 +- include/linux/compiler.h | 2 +- init/Kconfig | 5 - kernel/kallsyms.c | 12 +- scripts/kallsyms.c | 53 +-- scripts/link-vmlinux.sh | 4 - tools/objtool/check.c | 43 +- tools/objtool/elf.c | 2 +- tools/objtool/include/objtool/special.h | 2 +- tools/perf/util/annotate.c | 4 +- 50 files changed, 380 insertions(+), 667 deletions(-) delete mode 100644 arch/x86/tools/relocs_64.c
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 1/28 x86/pvh: Call C code via the kernel virtual mapping
From: Ard Biesheuvel <[email protected]>
Calling C code via a different mapping than it was linked at is problematic, because the compiler assumes that RIP-relative and absolute symbol references are interchangeable. GCC in particular may use RIP-relative per-CPU variable references even when not using -fpic.
So call xen_prepare_pvh() via its kernel virtual mapping on x86_64, so that those RIP-relative references produce the correct values. This matches the pre-existing behavior for i386, which also invokes xen_prepare_pvh() via the kernel virtual mapping before invoking startup_32 with paging disabled again.
--- arch/x86/platform/pvh/head.S | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
arch/x86/platform/pvh/head.S +5 -1
@@ -101,7 +101,11 @@ SYM_CODE_START_LOCAL(pvh_start_xen) xor %edx, %edx wrmsr - call xen_prepare_pvh + /* Call xen_prepare_pvh() via the kernel virtual mapping */ + leaq xen_prepare_pvh(%rip), %rax + addq $__START_KERNEL_map, %rax + ANNOTATE_RETPOLINE_SAFE + call *%rax /* startup_64 expects boot_params in %rsi. */ mov $_pa(pvh_bootparams), %rsi
Signature · 1 line
2.46.0.792.g87dc391469-goog
Re: [RFC PATCH 01/28] x86/pvh: Call C code via the kernel virtual mapping
Quoted 7 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Calling C code via a different mapping than it was linked at is problematic, because the compiler assumes that RIP-relative and absolute symbol references are interchangeable. GCC in particular may use RIP-relative per-CPU variable references even when not using -fpic.
Quoted 8 lines
So call
xen_prepare_pvh()via its kernel virtual mapping onx86_64, so that those RIP-relative references produce the correct values. This matches the pre-existing behavior for i386, which also invokesxen_prepare_pvh()via the kernel virtual mapping before invokingstartup_32with paging disabled again.Fixes 7243b93345f7 ("xen/pvh: Bootstrap PVH guest")Signed-off-by Ard Biesheuvel <[email protected]>
I found that before this change xen_prepare_pvh() would call through some pv_ops function pointers into the kernel virtual mapping.
Regards,
Jason
RFC PATCH 2/28 Documentation: Bump minimum GCC version to 8.1
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Bump the minimum GCC version to 8.1 to gain unconditional support for referring to the per-task stack cookie using a symbol rather than relying on the fixed offset of 40 bytes from %GS, which requires elaborate hacks to support.
--- Documentation/admin-guide/README.rst | 2 +- Documentation/process/changes.rst | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
Documentation/admin-guide/README.rst +1 -1
@@ -259,7 +259,7 @@ Configuring the kernel Compiling the kernel -------------------- - - Make sure you have at least gcc 5.1 available. + - Make sure you have at least gcc 8.1 available. For more information, refer to :ref:`Documentation/process/changes.rst <changes>`. - Do a ``make`` to create a compressed kernel image. It is also possible to do
Documentation/process/changes.rst +1 -1
@@ -29,7 +29,7 @@ you probably needn't concern yourself with pcmciautils. ====================== =============== ======================================== Program Minimal version Command to check the version ====================== =============== ======================================== -GNU C 5.1 gcc --version +GNU C 8.1 gcc --version Clang/LLVM (optional) 13.0.1 clang --version Rust (optional) 1.78.0 rustc --version bindgen (optional) 0.65.1 bindgen --version
Signature · 1 line
2.46.0.792.g87dc391469-goog
Re: [RFC PATCH 02/28] Documentation: Bump minimum GCC version to 8.1
Quoted 7 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Bump the minimum GCC version to 8.1 to gain unconditional support for referring to the per-task stack cookie using a symbol rather than relying on the fixed offset of 40 bytes from %GS, which requires elaborate hacks to support.
Quoted 5 lines
Signed-off-by Ard Biesheuvel <[email protected]>--- Documentation/admin-guide/README.rst | 2 +- Documentation/process/changes.rst | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
As we discussed during plumbers, I think this is reasonable, both the gcc-8.1 version and the timing after the 6.12-LTS kernel.
We obviously need to go through all the other version checks to see what else can be cleaned up. I would suggest we also raise the binutils version to 2.30+, which is what RHEL8 shipped alongside gcc-8. I have not found other distros that use older binutils in combination with gcc-8 or higher, Debian 10 uses binutils-2.31.
I don't think we want to combine the additional cleanup with your series, but if we can agree on the version, we can do that in parallel.
FWIW, here are links to the last few times we discussed this, and there are already has a few other things that would benefit from more modern compilers:
https://lore.kernel.org/lkml/[email protected]/
https://lore.kernel.org/lkml/CAFULd4biN8FPRtU54Q0QywfBFvvWV-s1M3kWF9YOmozyAX9+ZQ@mail.gmail.com/
https://lore.kernel.org/lkml/CAK8P3a1Vt17Yry_gTQ0dwr7_tEoFhuec+mQzzKzFvZGD5Hrnow@mail.gmail.com/
Arnd
Mark Rutland
reviewerRe: [RFC PATCH 02/28] Documentation: Bump minimum GCC version to 8.1
Hi Arnd,
Quoted 20 lines · Arnd Bergmann
Quoted 7 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Bump the minimum GCC version to 8.1 to gain unconditional support for referring to the per-task stack cookie using a symbol rather than relying on the fixed offset of 40 bytes from %GS, which requires elaborate hacks to support.
Quoted 5 lines
Signed-off-by Ard Biesheuvel <[email protected]>--- Documentation/admin-guide/README.rst | 2 +- Documentation/process/changes.rst | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)Acked-by Arnd Bergmann <[email protected]>As we discussed during plumbers, I think this is reasonable, both the gcc-8.1 version and the timing after the 6.12-LTS kernel.
Quoted 9 lines
We obviously need to go through all the other version checks to see what else can be cleaned up. I would suggest we also raise the binutils version to 2.30+, which is what RHEL8 shipped alongside gcc-8. I have not found other distros that use older binutils in combination with gcc-8 or higher, Debian 10 uses binutils-2.31.
I don't think we want to combine the additional cleanup with your series, but if we can agree on the version, we can do that in parallel.
Were you planning to send patches to that effect, or did you want someone else to do that? I think we were largely agreed on making those changes, but it wasn't clear to me who was actually going to send patches, and I couldn't spot a subsequent thread on LKML.
Mark.
Arnd Bergmann
maintainerRe: [RFC PATCH 02/28] Documentation: Bump minimum GCC version to 8.1
Quoted 14 lines · Mark Rutland
Quoted 12 lines · Arnd Bergmann
Quoted 1 line · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
We obviously need to go through all the other version checks to see what else can be cleaned up. I would suggest we also raise the binutils version to 2.30+, which is what RHEL8 shipped alongside gcc-8. I have not found other distros that use older binutils in combination with gcc-8 or higher, Debian 10 uses binutils-2.31.
I don't think we want to combine the additional cleanup with your series, but if we can agree on the version, we can do that in parallel.
Quoted 4 lines
Were you planning to send patches to that effect, or did you want someone else to do that? I think we were largely agreed on making those changes, but it wasn't clear to me who was actually going to send patches, and I couldn't spot a subsequent thread on LKML.
I hadn't planned on doing that, but I could help (after my vacation). As Ard already posted the the patch for gcc, I was expecting that this one would get merged along with the other patches in the series.
Ard, what is the status of your series, is this likely to make it into 6.14, or should we have a separate patch to just raise the minimum gcc and binutils version independent of your work?
Arnd
Re: [RFC PATCH 02/28] Documentation: Bump minimum GCC version to 8.1
In reply to Ard BiesheuvelQuoted 3 lines · Ard Biesheuvel
Documentation/admin-guide/README.rst | 2 +- Documentation/process/changes.rst | 2 +-
This should update scripts/min-tool-version.sh too. With that:
As Arnd says, the cleanups can be done afterwards.
Cheers,
Miguel
Re: [RFC PATCH 02/28] Documentation: Bump minimum GCC version to 8.1
In reply to Ard BiesheuvelQuoted 7 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Bump the minimum GCC version to 8.1 to gain unconditional support for referring to the per-task stack cookie using a symbol rather than relying on the fixed offset of 40 bytes from %GS, which requires elaborate hacks to support.
Quoted 5 lines
Signed-off-by Ard Biesheuvel <[email protected]>--- Documentation/admin-guide/README.rst | 2 +- Documentation/process/changes.rst | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-)
I'd like this for arm64 and others too (for unconditional support for -fpatchable-function-entry), so FWIW:
I think you'll want to update scripts/min-tool-version.sh too; judging by the diff in the cover letter that's not handled elsehere in the series.
Mark.
Quoted 31 lines
Documentation/admin-guide/README.rst +1
-1@@ -259,7 +259,7 @@ Configuring the kernel Compiling the kernel -------------------- - - Make sure you have at least gcc 5.1 available. + - Make sure you have at least gcc 8.1 available. For more information, refer to :ref:`Documentation/process/changes.rst <changes>`. - Do a ``make`` to create a compressed kernel image. It is also possible to doDocumentation/process/changes.rst +1
-1@@ -29,7 +29,7 @@ you probably needn't concern yourself with pcmciautils. ====================== =============== ======================================== Program Minimal version Command to check the version ====================== =============== ======================================== -GNU C 5.1 gcc --version +GNU C 8.1 gcc --version Clang/LLVM (optional) 13.0.1 clang --version Rust (optional) 1.78.0 rustc --version bindgen (optional) 0.65.1 bindgen --versionSignature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 3/28 x86/tools: Use mmap() to simplify relocs host tool
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Instead of relying on fseek() and fread() to traverse the vmlinux file when processing the ELF relocations, mmap() the whole thing and use memcpy() or direct references where appropriate:
- the executable and section headers are byte swabbed before use if the host is big endian, so there, the copy is retained;
- the strtab and extended symtab are not byte swabbed so there, the copies are replaced with direct references into the mmap()'ed region.
This substantially simplifies the code, and makes it much easier to refer to other file contents directly. This will be used by a subsequent patch to handle GOTPCREL relocations.
--- arch/x86/tools/relocs.c | 145 ++++++++------------ arch/x86/tools/relocs.h | 2 + 2 files changed, 62 insertions(+), 85 deletions(-)
arch/x86/tools/relocs.c +60 -85
@@ -37,15 +37,17 @@ static struct relocs relocs64;#endif struct section { -Elf_Shdr shdr; -struct section *link; -Elf_Sym *symtab; -Elf32_Word *xsymtab; -Elf_Rel *reltab; -char *strtab; +Elf_Shdr shdr; +struct section *link; +Elf_Sym *symtab; +const Elf32_Word *xsymtab; +Elf_Rel *reltab; +const char *strtab; };static struct section *secs; +static const void *elf_image; +static const char *const sym_regex_kernel[S_NSYMTYPES ] = {/* * Following symbols have been audited. There values are constant and do
@@ -291,7 +293,7 @@ static Elf_Sym *sym_lookup(const char *symname)for (i =0 ; i < shnum; i++) {struct section *sec = &secs[i];long nsyms; -char *strtab; +const char *strtab;Elf_Sym *symtab;Elf_Sym *sym;
@@ -354,7 +356,7 @@ static uint64_t elf64_to_cpu(uint64_t val)static int sym_index (Elf_Sym *sym) {Elf_Sym *symtab = secs[shsymtabndx].symtab; -Elf32_Word *xsymtab = secs[shxsymtabndx].xsymtab; +const Elf32_Word *xsymtab = secs[shxsymtabndx].xsymtab;unsigned long offset;int index;
@@ -368,10 +370,9 @@ static int sym_index(Elf_Sym *sym)return elf32_to_cpu (xsymtab[index]); } -static void read_ehdr (FILE *fp) +static void read_ehdr (void ) { -if (fread (&ehdr,sizeof (ehdr),1 , fp) !=1 ) -die ("Cannot read ELF header: %s\n" ,strerror (errno)); +memcpy (&ehdr, elf_image,sizeof (ehdr));if (memcmp (ehdr.e_ident,ELFMAG ,SELFMAG ) !=0 )die ("No ELF magic\n" );if (ehdr.e_ident[EI_CLASS ] !=ELF_CLASS )
@@ -414,60 +415,48 @@ static void read_ehdr(FILE *fp)if (shnum ==SHN_UNDEF || shstrndx ==SHN_XINDEX ) { -Elf_Shdr shdr; - -if (fseek (fp, ehdr.e_shoff,SEEK_SET ) <0 ) -die ("Seek to %" FMT " failed: %s\n" , ehdr.e_shoff,strerror (errno)); - -if (fread (&shdr,sizeof (shdr),1 , fp) !=1 ) -die ("Cannot read initial ELF section header: %s\n" ,strerror (errno)); +const Elf_Shdr *shdr = elf_image + ehdr.e_shoff;if (shnum ==SHN_UNDEF ) - shnum =elf_xword_to_cpu (shdr.sh_size); + shnum =elf_xword_to_cpu (shdr->sh_size);if (shstrndx ==SHN_XINDEX ) - shstrndx =elf_word_to_cpu (shdr.sh_link); + shstrndx =elf_word_to_cpu (shdr->sh_link); }if (shstrndx >= shnum)die ("String table index out of bounds\n" ); } -static void read_shdrs (FILE *fp) +static void read_shdrs (void ) { +const Elf_Shdr *shdr = elf_image + ehdr.e_shoff;int i; -Elf_Shdr shdr; secs =calloc (shnum,sizeof (struct section ));if (!secs)die ("Unable to allocate %ld section headers\n" , shnum); -if (fseek (fp, ehdr.e_shoff,SEEK_SET ) <0 ) -die ("Seek to %" FMT " failed: %s\n" , ehdr.e_shoff,strerror (errno)); - -for (i =0 ; i < shnum; i++) { +for (i =0 ; i < shnum; i++, shdr++) {struct section *sec = &secs[i]; -if (fread (&shdr,sizeof (shdr),1 , fp) !=1 ) -die ("Cannot read ELF section headers %d/%ld: %s\n" , i, shnum,strerror (errno)); - - sec->shdr.sh_name =elf_word_to_cpu (shdr.sh_name); - sec->shdr.sh_type =elf_word_to_cpu (shdr.sh_type); - sec->shdr.sh_flags =elf_xword_to_cpu (shdr.sh_flags); - sec->shdr.sh_addr =elf_addr_to_cpu (shdr.sh_addr); - sec->shdr.sh_offset =elf_off_to_cpu (shdr.sh_offset); - sec->shdr.sh_size =elf_xword_to_cpu (shdr.sh_size); - sec->shdr.sh_link =elf_word_to_cpu (shdr.sh_link); - sec->shdr.sh_info =elf_word_to_cpu (shdr.sh_info); - sec->shdr.sh_addralign =elf_xword_to_cpu (shdr.sh_addralign); - sec->shdr.sh_entsize =elf_xword_to_cpu (shdr.sh_entsize); + sec->shdr.sh_name =elf_word_to_cpu (shdr->sh_name); + sec->shdr.sh_type =elf_word_to_cpu (shdr->sh_type); + sec->shdr.sh_flags =elf_xword_to_cpu (shdr->sh_flags); + sec->shdr.sh_addr =elf_addr_to_cpu (shdr->sh_addr); + sec->shdr.sh_offset =elf_off_to_cpu (shdr->sh_offset); + sec->shdr.sh_size =elf_xword_to_cpu (shdr->sh_size); + sec->shdr.sh_link =elf_word_to_cpu (shdr->sh_link); + sec->shdr.sh_info =elf_word_to_cpu (shdr->sh_info); + sec->shdr.sh_addralign =elf_xword_to_cpu (shdr->sh_addralign); + sec->shdr.sh_entsize =elf_xword_to_cpu (shdr->sh_entsize);if (sec->shdr.sh_link < shnum) sec->link = &secs[sec->shdr.sh_link]; } } -static void read_strtabs (FILE *fp) +static void read_strtabs (void ) {int i;
@@ -476,20 +465,11 @@ static void read_strtabs(FILE *fp)if (sec->shdr.sh_type !=SHT_STRTAB )continue ; - - sec->strtab =malloc (sec->shdr.sh_size); -if (!sec->strtab) -die ("malloc of %" FMT " bytes for strtab failed\n" , sec->shdr.sh_size); - -if (fseek (fp, sec->shdr.sh_offset,SEEK_SET ) <0 ) -die ("Seek to %" FMT " failed: %s\n" , sec->shdr.sh_offset,strerror (errno)); - -if (fread (sec->strtab,1 , sec->shdr.sh_size, fp) != sec->shdr.sh_size) -die ("Cannot read symbol table: %s\n" ,strerror (errno)); + sec->strtab = elf_image + sec->shdr.sh_offset; } } -static void read_symtabs (FILE *fp) +static void read_symtabs (void ) {int i, j;
@@ -499,16 +479,7 @@ static void read_symtabs(FILE *fp)switch (sec->shdr.sh_type) {case SHT_SYMTAB_SHNDX : - sec->xsymtab =malloc (sec->shdr.sh_size); -if (!sec->xsymtab) -die ("malloc of %" FMT " bytes for xsymtab failed\n" , sec->shdr.sh_size); - -if (fseek (fp, sec->shdr.sh_offset,SEEK_SET ) <0 ) -die ("Seek to %" FMT " failed: %s\n" , sec->shdr.sh_offset,strerror (errno)); - -if (fread (sec->xsymtab,1 , sec->shdr.sh_size, fp) != sec->shdr.sh_size) -die ("Cannot read extended symbol table: %s\n" ,strerror (errno)); - + sec->xsymtab = elf_image + sec->shdr.sh_offset; shxsymtabndx = i;continue ;
@@ -519,11 +490,7 @@ static void read_symtabs(FILE *fp)if (!sec->symtab)die ("malloc of %" FMT " bytes for symtab failed\n" , sec->shdr.sh_size); -if (fseek (fp, sec->shdr.sh_offset,SEEK_SET ) <0 ) -die ("Seek to %" FMT " failed: %s\n" , sec->shdr.sh_offset,strerror (errno)); - -if (fread (sec->symtab,1 , sec->shdr.sh_size, fp) != sec->shdr.sh_size) -die ("Cannot read symbol table: %s\n" ,strerror (errno)); +memcpy (sec->symtab, elf_image + sec->shdr.sh_offset, sec->shdr.sh_size);for (j =0 ; j < num_syms; j++) {Elf_Sym *sym = &sec->symtab[j];
@@ -543,12 +510,13 @@ static void read_symtabs(FILE *fp) } -static void read_relocs (FILE *fp) +static void read_relocs (void ) {int i, j;for (i =0 ; i < shnum; i++) {struct section *sec = &secs[i]; +const Elf_Rel *reltab = elf_image + sec->shdr.sh_offset;if (sec->shdr.sh_type !=SHT_REL_TYPE )continue ;
@@ -557,19 +525,12 @@ static void read_relocs(FILE *fp)if (!sec->reltab)die ("malloc of %" FMT " bytes for relocs failed\n" , sec->shdr.sh_size); -if (fseek (fp, sec->shdr.sh_offset,SEEK_SET ) <0 ) -die ("Seek to %" FMT " failed: %s\n" , sec->shdr.sh_offset,strerror (errno)); - -if (fread (sec->reltab,1 , sec->shdr.sh_size, fp) != sec->shdr.sh_size) -die ("Cannot read symbol table: %s\n" ,strerror (errno)); -for (j =0 ; j < sec->shdr.sh_size/sizeof (Elf_Rel); j++) {Elf_Rel *rel = &sec->reltab[j]; - - rel->r_offset =elf_addr_to_cpu (rel->r_offset); - rel->r_info =elf_xword_to_cpu (rel->r_info); + rel->r_offset =elf_addr_to_cpu (reltab[j].r_offset); + rel->r_info =elf_xword_to_cpu (reltab[j].r_info);#if (SHT_REL_TYPE ==SHT_RELA ) - rel->r_addend =elf_xword_to_cpu (rel->r_addend); + rel->r_addend =elf_xword_to_cpu (reltab[j].r_addend);#endif } }
@@ -591,7 +552,7 @@ static void print_absolute_symbols(void)for (i =0 ; i < shnum; i++) {struct section *sec = &secs[i]; -char *sym_strtab; +const char *sym_strtab;int j;if (sec->shdr.sh_type !=SHT_SYMTAB )
@@ -633,7 +594,7 @@ static void print_absolute_relocs(void)for (i =0 ; i < shnum; i++) {struct section *sec = &secs[i];struct section *sec_applies, *sec_symtab; -char *sym_strtab; +const char *sym_strtab;Elf_Sym *sh_symtab;int j;
@@ -725,7 +686,7 @@ static void walk_relocs(int (*process)(struct section *sec, Elf_Rel *rel,/* Walk through the relocations */ for (i =0 ; i < shnum; i++) { -char *sym_strtab; +const char *sym_strtab;Elf_Sym *sh_symtab;struct section *sec_applies, *sec_symtab;int j;
@@ -1177,12 +1138,24 @@ void process(FILE *fp, int use_real_mode, int as_text, int show_absolute_syms, int show_absolute_relocs, int show_reloc_info) { + int fd = fileno(fp); +struct stat sb; +void *p; + +if (fstat (fd, &sb)) +die ("fstat() failed\n" ); + + elf_image = p =mmap (NULL , sb.st_size,PROT_READ ,MAP_PRIVATE , fd,0 ); +if (p ==MAP_FAILED ) +die ("mmap() failed\n" ); +regex_init (use_real_mode); -read_ehdr (fp); -read_shdrs (fp); -read_strtabs (fp); -read_symtabs (fp); -read_relocs (fp); + +read_ehdr (); +read_shdrs (); +read_strtabs (); +read_symtabs (); +read_relocs ();if (ELF_BITS ==64 )percpu_init ();
@@ -1203,4 +1176,6 @@ void process(FILE *fp, int use_real_mode, int as_text, }emit_relocs (as_text, use_real_mode); + +munmap (p, sb.st_size); }
arch/x86/tools/relocs.h +2 -0
@@ -16,6 +16,8 @@#include <endian.h> #include <regex.h> #include <tools/le_byteshift.h> +#include <sys/stat.h> +#include <sys/mman.h> __attribute__((__format__ (printf,1 ,2 )))void die (char *fmt, ...) __attribute__((noreturn));
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 4/28 x86/boot: Permit GOTPCREL relocations for x86_64 builds
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Some of the early x86_64 startup code is written in C, and executes in the early 1:1 mapping of the kernel, which is not the address it was linked at, and this requires special care when accessing global variables. This is currently being dealt with on an ad-hoc basis, primarily in head64.c, using explicit pointer fixups, but it would be better to rely on the compiler for this, by using -fPIE to generate code that can run at any address, and uses RIP-relative accesses to refer to global variables.
While it is possible to avoid most GOT based symbol references that the compiler typically emits when running in -fPIE mode, by using 'hidden' visibility, there are cases where the compiler will always rely on the GOT, for instance, for weak external references (which may remain unsatisfied at link time).
This means the build may produce a small number of GOT entries nonetheless. So update the reloc processing host tool to add support for this, and place the GOT in the .text section rather than discard it.
Note that multiple GOT based references to the same symbol will share a single GOT entry, and so naively emitting a relocation for the GOT entry each time a reference to it is encountered could result in duplicates. Work around this by relying on the fact that the relocation lists are sorted, and deduplicate 64-bit relocations as they are emitted by comparing each entry with the previous one.
--- arch/x86/Makefile | 4 +++ arch/x86/kernel/vmlinux.lds.S | 5 +++ arch/x86/tools/relocs.c | 33 ++++++++++++++++++-- include/asm-generic/vmlinux.lds.h | 7 +++++ 4 files changed, 47 insertions(+), 2 deletions(-)
arch/x86/Makefile +4 -0
@@ -192,6 +192,10 @@ else KBUILD_CFLAGS += -mcmodel=kernel KBUILD_RUSTFLAGS += -Cno-redzone=y KBUILD_RUSTFLAGS += -Ccode-model=kernel + + # Don't emit relaxable GOTPCREL relocations + KBUILD_AFLAGS_KERNEL += -Wa,-mrelax-relocations=no + KBUILD_CFLAGS_KERNEL += -Wa,-mrelax-relocations=no endif #
arch/x86/kernel/vmlinux.lds.S +5 -0
@@ -20,6 +20,9 @@ #define RUNTIME_DISCARD_EXIT #define EMITS_PT_NOTE #define RO_EXCEPTION_TABLE_ALIGN 16 +#ifdef CONFIG_X86_64 +#define GOT_IN_RODATA +#endif #include <asm-generic/vmlinux.lds.h> #include <asm/asm-offsets.h>
@@ -464,10 +467,12 @@ SECTIONS * Sections that should stay zero sized, which is safer to * explicitly check instead of blindly discarding. */ +#ifdef CONFIG_X86_32 .got : { *(.got) *(.igot.*) } ASSERT(SIZEOF(.got) == 0, "Unexpected GOT entries detected!") +#endif .plt : { *(.plt) *(.plt.*) *(.iplt)
arch/x86/tools/relocs.c +31 -2
@@ -223,6 +223,8 @@ static const char *rel_type(unsigned type)REL_TYPE (R_X86_64_JUMP_SLOT ),REL_TYPE (R_X86_64_RELATIVE ),REL_TYPE (R_X86_64_GOTPCREL ), +REL_TYPE (R_X86_64_GOTPCRELX ), +REL_TYPE (R_X86_64_REX_GOTPCRELX ),REL_TYPE (R_X86_64_32 ),REL_TYPE (R_X86_64_32S ), REL_TYPE(R_X86_64_16),
@@ -843,6 +845,7 @@ static int do_reloc64(struct section *sec, Elf_Rel *rel, ElfW(Sym) *sym,case R_X86_64_32 :case R_X86_64_32S :case R_X86_64_64 : +case R_X86_64_GOTPCREL :/* * References to the percpu area don't need to be adjusted. */
@@ -861,6 +864,31 @@ static int do_reloc64(struct section *sec, Elf_Rel *rel, ElfW(Sym) *sym,break ; } +if (r_type ==R_X86_64_GOTPCREL ) { +Elf_Shdr *s = &secs[sec->shdr.sh_info].shdr; +unsigned file_off = offset - s->sh_addr + s->sh_offset; + +/* +* GOTPCREL relocations refer to instructions that load +* a 64-bit address via a 32-bit relative reference to +* the GOT. In this case, it is the GOT entry that +* needs to be fixed up, not the immediate offset in +* the opcode. Note that the linker will have applied an +* addend of -4 to compensate for the delta between the +* relocation offset and the value of RIP when the +* instruction executes, and this needs to be backed out +* again. (Addends other than -4 are permitted in +* principle, but make no sense in practice so they are +* not supported.) +*/ +if (rel->r_addend !=-4 ) { +die ("invalid addend (%ld) for %s relocation: %s\n" , + rel->r_addend,rel_type (r_type), symname); +break ; + } + offset +=4 + (int32_t )get_unaligned_le32 (elf_image + file_off); + } +/* * Relocation offsets for 64 bit kernels are output * as 32 bits and sign extended back to 64 bits when
@@ -870,7 +898,7 @@ static int do_reloc64(struct section *sec, Elf_Rel *rel, ElfW(Sym) *sym,if ((int32_t )offset != (int64_t )offset)die ("Relocation offset doesn't fit in 32 bits\n" ); -if (r_type ==R_X86_64_64 ) +if (r_type ==R_X86_64_64 || r_type ==R_X86_64_GOTPCREL )add_reloc (&relocs64, offset);else add_reloc (&relocs32, offset);
@@ -1085,7 +1113,8 @@ static void emit_relocs(int as_text, int use_real_mode)/* Now print each relocation */ for (i =0 ; i < relocs64.count; i++) -write_reloc (relocs64.offset[i], stdout); +if (!i || relocs64.offset[i] != relocs64.offset[i -1 ]) +write_reloc (relocs64.offset[i], stdout);/* Print a stop */ write_reloc (0 , stdout);
include/asm-generic/vmlinux.lds.h +7 -0
@@ -443,6 +443,12 @@#endif #endif +#ifdef GOT_IN_RODATA +#define GOT_RODATA *(.got .igot*) +#else +#define GOT_RODATA +#endif +/* * Read only Data */
@@ -454,6 +460,7 @@SCHED_DATA \RO_AFTER_INIT_DATA /* Read only after init */ \ . = ALIGN(8); \ + GOT_RODATA \ BOUNDED_SECTION_BY(__tracepoints_ptrs, ___tracepoints_ptrs) \ *(__tracepoints_strings)/* Tracepoints: strings */ \ } \
Signature · 1 line
2.46.0.792.g87dc391469-goog
Josh Poimboeuf
maintainerRe: [RFC PATCH 04/28] x86/boot: Permit GOTPCREL relocations for x86_64 builds
Quoted 12 lines · Ard Biesheuvel
arch/x86/tools/relocs.c +12
-0+if (r_type ==R_X86_64_GOTPCREL ) { +Elf_Shdr *s = &secs[sec->shdr.sh_info].shdr; +unsigned file_off = offset - s->sh_addr + s->sh_offset; + +/* +* GOTPCREL relocations refer to instructions that load +* a 64-bit address via a 32-bit relative reference to +* the GOT. In this case, it is the GOT entry that +* needs to be fixed up, not the immediate offset in +* the opcode. Note that the linker will have applied an +* addend of -4 to compensate for the delta between the +* relocation offset and the value of RIP when the
arch/x86/tools/relocs.c static int do_reloc64(struct section *sec, Elf_Rel *rel, ElfW(Sym) *sym, · line 888
Quoted 10 lines
arch/x86/tools/relocs.c +10
-0+* instruction executes, and this needs to be backed out +* again. (Addends other than -4 are permitted in +* principle, but make no sense in practice so they are +* not supported.) +*/ +if (rel->r_addend !=-4 ) { +die ("invalid addend (%ld) for %s relocation: %s\n" , + rel->r_addend,rel_type (r_type), symname); +break ; + }
For x86 PC-relative addressing, the addend is <reloc offset> - <subsequent insn offset>. So a PC-relative addend can be something other than -4 when the relocation applies to the middle of an instruction, e.g.:
5b381: 66 81 3d 00 00 00 00 01 06 cmpw $0x601,0x0(%rip) # 5b38a <generic_validate_add_page+0x4a> 5b384: R_X86_64_PC32 boot_cpu_data-0x6
5f283: 81 3d 00 00 00 00 ff ff ff 00 cmpl $0xffffff,0x0(%rip) # 5f28d <x86_acpi_suspend_lowlevel+0x9d> 5f285: R_X86_64_PC32 smpboot_control-0x8
72f67: c6 05 00 00 00 00 01 movb $0x1,0x0(%rip) # 72f6e <sched_itmt_update_handler+0x6e> 72f69: R_X86_64_PC32 x86_topology_update-0x5
Presumably that could also happen with R_X86_64_GOTPCREL?
Signature · 1 line
Josh
Ard Biesheuvel
maintainerRe: [RFC PATCH 04/28] x86/boot: Permit GOTPCREL relocations for x86_64 builds
Quoted 30 lines · Josh Poimboeuf
Quoted 12 lines · Ard Biesheuvel
arch/x86/tools/relocs.c +12
-0+if (r_type ==R_X86_64_GOTPCREL ) { +Elf_Shdr *s = &secs[sec->shdr.sh_info].shdr; +unsigned file_off = offset - s->sh_addr + s->sh_offset; + +/* +* GOTPCREL relocations refer to instructions that load +* a 64-bit address via a 32-bit relative reference to +* the GOT. In this case, it is the GOT entry that +* needs to be fixed up, not the immediate offset in +* the opcode. Note that the linker will have applied an +* addend of -4 to compensate for the delta between the +* relocation offset and the value of RIP when the arch/x86/tools/relocs.c static int do_reloc64(struct section *sec, Elf_Rel *rel, ElfW(Sym) *sym, · line 888
Quoted 10 lines
arch/x86/tools/relocs.c +10
-0+* instruction executes, and this needs to be backed out +* again. (Addends other than -4 are permitted in +* principle, but make no sense in practice so they are +* not supported.) +*/ +if (rel->r_addend !=-4 ) { +die ("invalid addend (%ld) for %s relocation: %s\n" , + rel->r_addend,rel_type (r_type), symname); +break ; + }For x86 PC-relative addressing, the addend is <reloc offset> - <subsequent insn offset>. So a PC-relative addend can be something other than -4 when the relocation applies to the middle of an instruction, e.g.:
Quoted 8 lines
5b381: 66 81 3d 00 00 00 00 01 06 cmpw $0x601,0x0(%rip) # 5b38a <generic_validate_add_page+0x4a> 5b384: R_X86_64_PC32 boot_cpu_data-0x6 5f283: 81 3d 00 00 00 00 ff ff ff 00 cmpl $0xffffff,0x0(%rip) # 5f28d <x86_acpi_suspend_lowlevel+0x9d> 5f285: R_X86_64_PC32 smpboot_control-0x8 72f67: c6 05 00 00 00 00 01 movb $0x1,0x0(%rip) # 72f6e <sched_itmt_update_handler+0x6e> 72f69: R_X86_64_PC32 x86_topology_update-0x5Presumably that could also happen with
R_X86_64_GOTPCREL?
In theory, yes.
But for the class of GOTPCREL relaxable instructions listed in the psABI, the addend is always -4, and these are the only ones we might expect from the compiler when using -fpic with 'hidden' visibility and/or -mdirect-extern-access. Note that the memory operand foo@GOTPCREL(%rip) produces the *address* of foo, and so it is always the source operand, appearing at the end of the encoding.
Alternatively, we might simply subtract the addend from 'offset' before applying the displacement from the opcode.
Note that this code gets removed again in the last patch, after switching to PIE linking.
RFC PATCH 5/28 x86: Define the stack protector guard symbol explicitly
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Specify the guard symbol for the stack cookie explicitly, rather than positioning it exactly 40 bytes into the per-CPU area. Doing so removes the need for the per-CPU region to be absolute rather than relative to the placement of the per-CPU template region in the kernel image, and this allows the special handling for absolute per-CPU symbols to be removed entirely.
This is a worthwhile cleanup in itself, but it is also a prerequisite for PIE codegen and PIE linking, which can replace our bespoke and rather clunky runtime relocation handling.
--- arch/x86/Makefile | 4 ++++ arch/x86/include/asm/init.h | 2 +- arch/x86/include/asm/processor.h | 11 +++-------- arch/x86/include/asm/stackprotector.h | 4 ---- tools/perf/util/annotate.c | 4 ++-- 5 files changed, 10 insertions(+), 15 deletions(-)
arch/x86/Makefile +4 -0
@@ -193,6 +193,10 @@ else KBUILD_RUSTFLAGS += -Cno-redzone=y KBUILD_RUSTFLAGS += -Ccode-model=kernel + ifeq ($(CONFIG_STACKPROTECTOR),y) + KBUILD_CFLAGS += -mstack-protector-guard-symbol=fixed_percpu_data + endif + # Don't emit relaxable GOTPCREL relocations KBUILD_AFLAGS_KERNEL += -Wa,-mrelax-relocations=no KBUILD_CFLAGS_KERNEL += -Wa,-mrelax-relocations=no
arch/x86/include/asm/init.h +1 -1
@@ -2,7 +2,7 @@ #ifndef _ASM_X86_INIT_H #define _ASM_X86_INIT_H -#define __head __section(".head.text") +#define __head __section(".head.text") __no_stack_protector struct x86_mapping_info { void *(*alloc_pgt_page)(void *); /* allocate buf for page table */
arch/x86/include/asm/processor.h +3 -8
@@ -402,14 +402,9 @@ struct irq_stack {#ifdef CONFIG_X86_64 struct fixed_percpu_data {/* -* GCC hardcodes the stack canary as %gs:40. Since the -* irq_stack is the object at %gs:0, we reserve the bottom -* 48 bytes of the irq stack for the canary. -* -* Once we are willing to require -mstack-protector-guard-symbol= -* support for x86_64 stackprotector, we can get rid of this. +* Since the irq_stack is the object at %gs:0, the bottom 8 bytes of +* the irq stack are reserved for the canary. */ -char gs_base[40 ];unsigned long stack_canary; };
@@ -418,7 +413,7 @@ DECLARE_INIT_PER_CPU(fixed_percpu_data);static inline unsigned long cpu_kernelmode_gs_base (int cpu) { -return (unsigned long )per_cpu (fixed_percpu_data.gs_base, cpu); +return (unsigned long )&per_cpu (fixed_percpu_data, cpu); } extern asmlinkage void entry_SYSCALL32_ignore(void);
arch/x86/include/asm/stackprotector.h +0 -4
@@ -51,10 +51,6 @@ static __always_inline void boot_init_stack_canary(void) {unsigned long canary =get_random_canary (); -#ifdef CONFIG_X86_64 -BUILD_BUG_ON (offsetof(struct fixed_percpu_data , stack_canary) !=40 ); -#endif - current->stack_canary = canary;#ifdef CONFIG_X86_64 this_cpu_write (fixed_percpu_data.stack_canary, canary);
tools/perf/util/annotate.c +2 -2
@@ -2485,10 +2485,10 @@ static bool is_stack_operation(struct arch *arch, struct disasm_line *dl)static bool is_stack_canary (struct arch *arch,struct annotated_op_loc *loc) { -/* On x86_64, %gs:40 is used for stack canary */ +/* On x86_64, %gs:0 is used for stack canary */ if (arch__is (arch,"x86" )) {if (loc->segment ==INSN_SEG_X86_GS && loc->imm && - loc->offset ==40 ) + loc->offset ==0 )return true; }
Signature · 1 line
2.46.0.792.g87dc391469-goog
Ian Rogers
reviewerRe: [RFC PATCH 05/28] x86: Define the stack protector guard symbol explicitly
Quoted 101 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Specify the guard symbol for the stack cookie explicitly, rather than positioning it exactly 40 bytes into the per-CPU area. Doing so removes the need for the per-CPU region to be absolute rather than relative to the placement of the per-CPU template region in the kernel image, and this allows the special handling for absolute per-CPU symbols to be removed entirely.
This is a worthwhile cleanup in itself, but it is also a prerequisite for PIE codegen and PIE linking, which can replace our bespoke and rather clunky runtime relocation handling.
Signed-off-by Ard Biesheuvel <[email protected]>--- arch/x86/Makefile | 4 ++++ arch/x86/include/asm/init.h | 2 +- arch/x86/include/asm/processor.h | 11 +++-------- arch/x86/include/asm/stackprotector.h | 4 ---- tools/perf/util/annotate.c | 4 ++-- 5 files changed, 10 insertions(+), 15 deletions(-)arch/x86/Makefile +4
-0@@ -193,6 +193,10 @@ else KBUILD_RUSTFLAGS += -Cno-redzone=y KBUILD_RUSTFLAGS += -Ccode-model=kernel + ifeq ($(CONFIG_STACKPROTECTOR),y) + KBUILD_CFLAGS += -mstack-protector-guard-symbol=fixed_percpu_data + endif + # Don't emit relaxable GOTPCREL relocations KBUILD_AFLAGS_KERNEL += -Wa,-mrelax-relocations=no KBUILD_CFLAGS_KERNEL += -Wa,-mrelax-relocations=noarch/x86/include/asm/init.h +1
-1@@ -2,7 +2,7 @@ #ifndef _ASM_X86_INIT_H #define _ASM_X86_INIT_H -#define __head __section(".head.text") +#define __head __section(".head.text") __no_stack_protector struct x86_mapping_info { void *(*alloc_pgt_page)(void *); /* allocate buf for page table */arch/x86/include/asm/processor.h +3
-8@@ -402,14 +402,9 @@ struct irq_stack {#ifdef CONFIG_X86_64 struct fixed_percpu_data {/* -* GCC hardcodes the stack canary as %gs:40. Since the -* irq_stack is the object at %gs:0, we reserve the bottom -* 48 bytes of the irq stack for the canary. -* -* Once we are willing to require -mstack-protector-guard-symbol= -* support for x86_64 stackprotector, we can get rid of this. +* Since the irq_stack is the object at %gs:0, the bottom 8 bytes of +* the irq stack are reserved for the canary. */ -char gs_base[40 ];unsigned long stack_canary; };@@ -418,7 +413,7 @@ DECLARE_INIT_PER_CPU(fixed_percpu_data);static inline unsigned long cpu_kernelmode_gs_base (int cpu) { -return (unsigned long )per_cpu (fixed_percpu_data.gs_base, cpu); +return (unsigned long )&per_cpu (fixed_percpu_data, cpu); } extern asmlinkage void entry_SYSCALL32_ignore(void);arch/x86/include/asm/stackprotector.h +0
-4@@ -51,10 +51,6 @@ static __always_inline void boot_init_stack_canary(void) {unsigned long canary =get_random_canary (); -#ifdef CONFIG_X86_64 -BUILD_BUG_ON (offsetof(struct fixed_percpu_data , stack_canary) !=40 ); -#endif - current->stack_canary = canary;#ifdef CONFIG_X86_64 this_cpu_write (fixed_percpu_data.stack_canary, canary);tools/perf/util/annotate.c +0
-0
tools/perf/util/annotate.c static bool is_stack_operation(struct arch *arch, struct disasm_line *dl) · line 2491
Quoted 9 lines
tools/perf/util/annotate.c +2
-2@@ -2485,10 +2485,10 @@ static bool is_stack_operation(struct arch *arch, struct disasm_line *dl)static bool is_stack_canary (struct arch *arch,struct annotated_op_loc *loc) { -/* On x86_64, %gs:40 is used for stack canary */ +/* On x86_64, %gs:0 is used for stack canary */ if (arch__is (arch,"x86" )) {if (loc->segment ==INSN_SEG_X86_GS && loc->imm && - loc->offset ==40 ) + loc->offset ==0 )
As a new perf tool can run on old kernels we may need to have this be something like:
(loc->offset == 40 /* pre v6.xx kernels */ || loc->offset == 0 /* v6.xx and later */ )
We could make this dependent on the kernel by processing the os_release string:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/env.h#n55
but that could well be more trouble than it is worth.
Thanks,
Ian
Quoted 7 lines
tools/perf/util/annotate.c +0
-1return true; } --2.46.0.792.g87dc391469-goog
Ard Biesheuvel
maintainerRe: [RFC PATCH 05/28] x86: Define the stack protector guard symbol explicitly
Quoted 25 lines · Ian Rogers
Quoted 23 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Specify the guard symbol for the stack cookie explicitly, rather than positioning it exactly 40 bytes into the per-CPU area. Doing so removes the need for the per-CPU region to be absolute rather than relative to the placement of the per-CPU template region in the kernel image, and this allows the special handling for absolute per-CPU symbols to be removed entirely.
This is a worthwhile cleanup in itself, but it is also a prerequisite for PIE codegen and PIE linking, which can replace our bespoke and rather clunky runtime relocation handling.
Signed-off-by Ard Biesheuvel <[email protected]>--- arch/x86/Makefile | 4 ++++ arch/x86/include/asm/init.h | 2 +- arch/x86/include/asm/processor.h | 11 +++-------- arch/x86/include/asm/stackprotector.h | 4 ---- tools/perf/util/annotate.c | 4 ++-- 5 files changed, 10 insertions(+), 15 deletions(-)
...
Quoted 20 lines
Quoted 5 lines
tools/perf/util/annotate.c +0
-0tools/perf/util/annotate.c static bool is_stack_operation(struct arch *arch, struct disasm_line *dl) · line 2491
Quoted 9 lines
tools/perf/util/annotate.c +2
-2@@ -2485,10 +2485,10 @@ static bool is_stack_operation(struct arch *arch, struct disasm_line *dl)static bool is_stack_canary (struct arch *arch,struct annotated_op_loc *loc) { -/* On x86_64, %gs:40 is used for stack canary */ +/* On x86_64, %gs:0 is used for stack canary */ if (arch__is (arch,"x86" )) {if (loc->segment ==INSN_SEG_X86_GS && loc->imm && - loc->offset ==40 ) + loc->offset ==0 )As a new perf tool can run on old kernels we may need to have this be something like:
(loc->offset== 40 /* pre v6.xx kernels */ ||loc->offset== 0 /* v6.xx and later */ )
Quoted 4 lines
We could make this dependent on the kernel by processing the
os_releasestring:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/env.h#n55
but that could well be more trouble than it is worth.
Yeah. I also wonder what the purpose of this feature is. At the end of this series, the stack cookie will no longer be at a fixed offset of %GS anyway, and so perf will not be able to identify it in the same manner. So it is probably better to just leave this in place, as the %gs:0 case will not exist in the field (assuming that the series lands all at once).
Any idea why this deviates from other architectures? Is x86_64 the only arch that needs to identify stack canary accesses in perf? We could rename the symbol to something identifiable, and do it across all architectures, if this really serves a need (and assuming that perf has insight into the symbol table).
Ian Rogers
reviewerRe: [RFC PATCH 05/28] x86: Define the stack protector guard symbol explicitly
Quoted 60 lines · Ard Biesheuvel
Quoted 25 lines · Ian Rogers
Quoted 23 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Specify the guard symbol for the stack cookie explicitly, rather than positioning it exactly 40 bytes into the per-CPU area. Doing so removes the need for the per-CPU region to be absolute rather than relative to the placement of the per-CPU template region in the kernel image, and this allows the special handling for absolute per-CPU symbols to be removed entirely.
This is a worthwhile cleanup in itself, but it is also a prerequisite for PIE codegen and PIE linking, which can replace our bespoke and rather clunky runtime relocation handling.
Signed-off-by Ard Biesheuvel <[email protected]>--- arch/x86/Makefile | 4 ++++ arch/x86/include/asm/init.h | 2 +- arch/x86/include/asm/processor.h | 11 +++-------- arch/x86/include/asm/stackprotector.h | 4 ---- tools/perf/util/annotate.c | 4 ++-- 5 files changed, 10 insertions(+), 15 deletions(-)...
Quoted 20 lines
Quoted 5 lines
tools/perf/util/annotate.c +0
-0tools/perf/util/annotate.c static bool is_stack_operation(struct arch *arch, struct disasm_line *dl) · line 2491
Quoted 9 lines
tools/perf/util/annotate.c +2
-2@@ -2485,10 +2485,10 @@ static bool is_stack_operation(struct arch *arch, struct disasm_line *dl)static bool is_stack_canary (struct arch *arch,struct annotated_op_loc *loc) { -/* On x86_64, %gs:40 is used for stack canary */ +/* On x86_64, %gs:0 is used for stack canary */ if (arch__is (arch,"x86" )) {if (loc->segment ==INSN_SEG_X86_GS && loc->imm && - loc->offset ==40 ) + loc->offset ==0 )As a new perf tool can run on old kernels we may need to have this be something like:
(loc->offset== 40 /* pre v6.xx kernels */ ||loc->offset== 0 /* v6.xx and later */ )Quoted 4 lines
We could make this dependent on the kernel by processing the
os_releasestring:
https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/util/env.h#n55
but that could well be more trouble than it is worth.Yeah. I also wonder what the purpose of this feature is. At the end of this series, the stack cookie will no longer be at a fixed offset of %GS anyway, and so perf will not be able to identify it in the same manner. So it is probably better to just leave this in place, as the %gs:0 case will not exist in the field (assuming that the series lands all at once).
Quoted 5 lines
Any idea why this deviates from other architectures? Is
x86_64the only arch that needs to identify stack canary accesses in perf? We could rename the symbol to something identifiable, and do it across all architectures, if this really serves a need (and assuming that perf has insight into the symbol table).
This is relatively new work coming from Namhyung for data type profiling and I believe is pretty much just x86 at the moment - although the ever awesome IBM made contributions for PowerPC. The data type profiling is trying to classify memory accesses which is why it cares about the stack canary instruction, the particular encoding shouldn't matter.
Thanks,
Ian
Re: [RFC PATCH 05/28] x86: Define the stack protector guard symbol explicitly
In reply to Ard BiesheuvelQuoted 10 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Specify the guard symbol for the stack cookie explicitly, rather than positioning it exactly 40 bytes into the per-CPU area. Doing so removes the need for the per-CPU region to be absolute rather than relative to the placement of the per-CPU template region in the kernel image, and this allows the special handling for absolute per-CPU symbols to be removed entirely.
Quoted 3 lines
This is a worthwhile cleanup in itself, but it is also a prerequisite for PIE codegen and PIE linking, which can replace our bespoke and rather clunky runtime relocation handling.
I would like to point out a series that converted the stack protector guard symbol to a normal percpu variable [1], so there was no need to assume anything about the location of the guard symbol.
[1] "[PATCH v4 00/16] x86-64: Stack protector and percpu improvements"
https://lore.kernel.org/lkml/[email protected]/
Uros.
Quoted 103 lines
Signed-off-by Ard Biesheuvel <[email protected]>--- arch/x86/Makefile | 4 ++++ arch/x86/include/asm/init.h | 2 +- arch/x86/include/asm/processor.h | 11 +++-------- arch/x86/include/asm/stackprotector.h | 4 ---- tools/perf/util/annotate.c | 4 ++-- 5 files changed, 10 insertions(+), 15 deletions(-)arch/x86/Makefile +4
-0@@ -193,6 +193,10 @@ else KBUILD_RUSTFLAGS += -Cno-redzone=y KBUILD_RUSTFLAGS += -Ccode-model=kernel + ifeq ($(CONFIG_STACKPROTECTOR),y) + KBUILD_CFLAGS += -mstack-protector-guard-symbol=fixed_percpu_data + endif + # Don't emit relaxable GOTPCREL relocations KBUILD_AFLAGS_KERNEL += -Wa,-mrelax-relocations=no KBUILD_CFLAGS_KERNEL += -Wa,-mrelax-relocations=noarch/x86/include/asm/init.h +1
-1@@ -2,7 +2,7 @@ #ifndef _ASM_X86_INIT_H #define _ASM_X86_INIT_H -#define __head __section(".head.text") +#define __head __section(".head.text") __no_stack_protector struct x86_mapping_info { void *(*alloc_pgt_page)(void *); /* allocate buf for page table */arch/x86/include/asm/processor.h +3
-8@@ -402,14 +402,9 @@ struct irq_stack {#ifdef CONFIG_X86_64 struct fixed_percpu_data {/* -* GCC hardcodes the stack canary as %gs:40. Since the -* irq_stack is the object at %gs:0, we reserve the bottom -* 48 bytes of the irq stack for the canary. -* -* Once we are willing to require -mstack-protector-guard-symbol= -* support for x86_64 stackprotector, we can get rid of this. +* Since the irq_stack is the object at %gs:0, the bottom 8 bytes of +* the irq stack are reserved for the canary. */ -char gs_base[40 ];unsigned long stack_canary; };@@ -418,7 +413,7 @@ DECLARE_INIT_PER_CPU(fixed_percpu_data);static inline unsigned long cpu_kernelmode_gs_base (int cpu) { -return (unsigned long )per_cpu (fixed_percpu_data.gs_base, cpu); +return (unsigned long )&per_cpu (fixed_percpu_data, cpu); } extern asmlinkage void entry_SYSCALL32_ignore(void);arch/x86/include/asm/stackprotector.h +0
-4@@ -51,10 +51,6 @@ static __always_inline void boot_init_stack_canary(void) {unsigned long canary =get_random_canary (); -#ifdef CONFIG_X86_64 -BUILD_BUG_ON (offsetof(struct fixed_percpu_data , stack_canary) !=40 ); -#endif - current->stack_canary = canary;#ifdef CONFIG_X86_64 this_cpu_write (fixed_percpu_data.stack_canary, canary);tools/perf/util/annotate.c +2
-3@@ -2485,10 +2485,10 @@ static bool is_stack_operation(struct arch *arch, struct disasm_line *dl)static bool is_stack_canary (struct arch *arch,struct annotated_op_loc *loc) { -/* On x86_64, %gs:40 is used for stack canary */ +/* On x86_64, %gs:0 is used for stack canary */ if (arch__is (arch,"x86" )) {if (loc->segment ==INSN_SEG_X86_GS && loc->imm && - loc->offset ==40 ) + loc->offset ==0 )return true; } --2.46.0.792.g87dc391469-goog
Re: [RFC PATCH 05/28] x86: Define the stack protector guard symbol explicitly
Quoted 16 lines · Uros Bizjak
Quoted 10 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Specify the guard symbol for the stack cookie explicitly, rather than positioning it exactly 40 bytes into the per-CPU area. Doing so removes the need for the per-CPU region to be absolute rather than relative to the placement of the per-CPU template region in the kernel image, and this allows the special handling for absolute per-CPU symbols to be removed entirely.
Quoted 3 lines
This is a worthwhile cleanup in itself, but it is also a prerequisite for PIE codegen and PIE linking, which can replace our bespoke and rather clunky runtime relocation handling.
Quoted 8 lines
I would like to point out a series that converted the stack protector guard symbol to a normal percpu variable [1], so there was no need to assume anything about the location of the guard symbol.
[1] "[PATCH v4 00/16] x86-64: Stack protector and percpu improvements"
https://lore.kernel.org/lkml/[email protected]/Uros.
I plan on resubmitting that series sometime after the 6.12 merge window closes. As I recall from the last version, it was decided to wait until after the next LTS release to raise the minimum GCC version to 8.1 and avoid the need to be compatible with the old stack protector layout.
Brian Gerst
Ard Biesheuvel
maintainerRe: [RFC PATCH 05/28] x86: Define the stack protector guard symbol explicitly
Quoted 25 lines · Brian Gerst
Quoted 23 lines · Uros Bizjak
Quoted 10 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Specify the guard symbol for the stack cookie explicitly, rather than positioning it exactly 40 bytes into the per-CPU area. Doing so removes the need for the per-CPU region to be absolute rather than relative to the placement of the per-CPU template region in the kernel image, and this allows the special handling for absolute per-CPU symbols to be removed entirely.
Quoted 3 lines
This is a worthwhile cleanup in itself, but it is also a prerequisite for PIE codegen and PIE linking, which can replace our bespoke and rather clunky runtime relocation handling.
I would like to point out a series that converted the stack protector guard symbol to a normal percpu variable [1], so there was no need to assume anything about the location of the guard symbol.
[1] "[PATCH v4 00/16] x86-64: Stack protector and percpu improvements"
https://lore.kernel.org/lkml/[email protected]/
Quoted 8 lines
Quoted 1 line
Uros.
I plan on resubmitting that series sometime after the 6.12 merge window closes. As I recall from the last version, it was decided to wait until after the next LTS release to raise the minimum GCC version to 8.1 and avoid the need to be compatible with the old stack protector layout.
Hi Brian,
I'd be more than happy to compare notes on that - I wasn't aware of your intentions here, or I would have reached out before sending this RFC.
There are two things that you would need to address for Clang support to work correctly:
- the workaround I cc'ed you on the other day [0],
- a workaround for the module loader so it tolerates the GOTPCRELX relocations that Clang emits [1]
[0] https://lore.kernel.org/all/[email protected]/
[1] https://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git/commit/?id=a18121aabbdd
Re: [RFC PATCH 05/28] x86: Define the stack protector guard symbol explicitly
Quoted 50 lines · Ard Biesheuvel
Quoted 25 lines · Brian Gerst
Quoted 23 lines · Uros Bizjak
Quoted 10 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Specify the guard symbol for the stack cookie explicitly, rather than positioning it exactly 40 bytes into the per-CPU area. Doing so removes the need for the per-CPU region to be absolute rather than relative to the placement of the per-CPU template region in the kernel image, and this allows the special handling for absolute per-CPU symbols to be removed entirely.
Quoted 3 lines
This is a worthwhile cleanup in itself, but it is also a prerequisite for PIE codegen and PIE linking, which can replace our bespoke and rather clunky runtime relocation handling.
I would like to point out a series that converted the stack protector guard symbol to a normal percpu variable [1], so there was no need to assume anything about the location of the guard symbol.
[1] "[PATCH v4 00/16] x86-64: Stack protector and percpu improvements"
https://lore.kernel.org/lkml/[email protected]/Quoted 8 lines
Quoted 1 line
Uros.
I plan on resubmitting that series sometime after the 6.12 merge window closes. As I recall from the last version, it was decided to wait until after the next LTS release to raise the minimum GCC version to 8.1 and avoid the need to be compatible with the old stack protector layout.
Hi Brian,
I'd be more than happy to compare notes on that - I wasn't aware of your intentions here, or I would have reached out before sending this RFC.
There are two things that you would need to address for Clang support to work correctly:
- the workaround I cc'ed you on the other day [0],
- a workaround for the module loader so it tolerates the GOTPCRELX relocations that Clang emits [1]
Quoted 2 lines
[0] https://lore.kernel.org/all/[email protected]/
[1] https://git.kernel.org/pub/scm/linux/kernel/git/ardb/linux.git/commit/?id=a18121aabbdd
The first patch should be applied independently as a bug fix, since it already affects the 32-bit build with clang.
I don't have an environment with an older clang compiler to test the second patch, but I'll assume it will be necessary. I did run into an issue with the GOTPCRELX relocations before [1], but I thought it was just an objtool issue and didn't do more testing to know if modules were broken or not.
Brian Gerst
Re: [RFC PATCH 05/28] x86: Define the stack protector guard symbol explicitly
In reply to Ard BiesheuvelQuoted 28 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Specify the guard symbol for the stack cookie explicitly, rather than positioning it exactly 40 bytes into the per-CPU area. Doing so removes the need for the per-CPU region to be absolute rather than relative to the placement of the per-CPU template region in the kernel image, and this allows the special handling for absolute per-CPU symbols to be removed entirely.
This is a worthwhile cleanup in itself, but it is also a prerequisite for PIE codegen and PIE linking, which can replace our bespoke and rather clunky runtime relocation handling.
Signed-off-by Ard Biesheuvel <[email protected]>--- arch/x86/Makefile | 4 ++++ arch/x86/include/asm/init.h | 2 +- arch/x86/include/asm/processor.h | 11 +++-------- arch/x86/include/asm/stackprotector.h | 4 ---- tools/perf/util/annotate.c | 4 ++-- 5 files changed, 10 insertions(+), 15 deletions(-)arch/x86/Makefile +0
-0
arch/x86/Makefile else · line 197
Quoted 5 lines
arch/x86/Makefile +2
-0@@ -193,6 +193,10 @@ else KBUILD_RUSTFLAGS += -Cno-redzone=y KBUILD_RUSTFLAGS += -Ccode-model=kernel + ifeq ($(CONFIG_STACKPROTECTOR),y) + KBUILD_CFLAGS += -mstack-protector-guard-symbol=fixed_percpu_data
Looking at:
arch/x86/Makefile else · line 199
Quoted 2 lines
arch/x86/Makefile +2
-0+ * Since the irq_stack is the object at %gs:0, the bottom 8 bytes of + * the irq stack are reserved for the canary.
Please note that %gs:0 can also be achieved with
-mstack-protector-guard-offset=0
Uros.
RFC PATCH 6/28 x86/percpu: Get rid of absolute per-CPU variable placement
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
For historic reasons, per-CPU symbols on x86_64 are emitted in an address space that is disjoint from the ordinary kernel VA space, starting at address 0x0. This splits a per-CPU symbol reference into a base plus offset, where the base is programmed into the GS segment register.
This deviates from the usual approach adopted by other SMP architectures, where the base is a reference to the variable in the kernel image's per-CPU template area, and the offset is a per-CPU value accounting for the displacement of that particular CPU's per-CPU region with respect to the template area. This gives per-CPU variable references a range that is identical to ordinary references, and requires no special handling for the startup code, as the offset will simply be 0x0 up until the point where per-CPU variables are initialized properly.
The x86_64 approach was needed to accommodate per-task stack protector cookies, which used to live at a fixed offset of GS+40, requiring GS to be treated as a base register. This is no longer the case, though, and so GS can be repurposed as a true per-CPU offset, adopting the same strategy as other architectures.
This also removes the need for linker tricks to emit the per-CPU ELF segment at a different virtual address. It also means RIP-relative per-CPU variables no longer need to be relocated in the opposite direction when KASLR is applied, which was necessary because the 0x0 based per-CPU region remains in place even when the kernel is moved around.
--- arch/x86/include/asm/desc.h | 1 - arch/x86/include/asm/percpu.h | 22 -------------- arch/x86/include/asm/processor.h | 5 ++-- arch/x86/kernel/head64.c | 2 +- arch/x86/kernel/head_64.S | 12 ++------ arch/x86/kernel/irq_64.c | 1 - arch/x86/kernel/setup_percpu.c | 9 +----- arch/x86/kernel/vmlinux.lds.S | 30 -------------------- arch/x86/platform/pvh/head.S | 6 ++-- arch/x86/tools/relocs.c | 8 +----- arch/x86/xen/xen-head.S | 10 ++----- init/Kconfig | 1 - 12 files changed, 13 insertions(+), 94 deletions(-)
arch/x86/include/asm/desc.h +0 -1
@@ -46,7 +46,6 @@ struct gdt_page { } __attribute__((aligned(PAGE_SIZE))); DECLARE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page); -DECLARE_INIT_PER_CPU (gdt_page);/* Provide the original GDT */ static inline struct desc_struct *get_cpu_gdt_rw (unsigned int cpu)
arch/x86/include/asm/percpu.h +0 -22
@@ -20,12 +20,6 @@#define PER_CPU_VAR (var) __percpu(var)__percpu_rel -#ifdef CONFIG_X86_64_SMP -# define INIT_PER_CPU_VAR (var) init_per_cpu__##var -#else -# define INIT_PER_CPU_VAR (var) var -#endif -#else /* !__ASSEMBLY__: */ #include <linux/build_bug.h>
@@ -97,22 +91,6 @@#define __percpu_arg (x) __percpu_prefix "%" #x#define __force_percpu_arg (x) __force_percpu_prefix "%" #x -/* -* Initialized pointers to per-CPU variables needed for the boot -* processor need to use these macros to get the proper address -* offset from __per_cpu_load on SMP. -* -* There also must be an entry in vmlinux_64.lds.S -*/ -#define DECLARE_INIT_PER_CPU (var) \ - extern typeof(var) init_per_cpu_var(var) - -#ifdef CONFIG_X86_64_SMP -# define init_per_cpu_var (var) init_per_cpu__##var -#else -# define init_per_cpu_var (var) var -#endif -/* * For arch-specific code, we can use direct single-insn ops (they * don't give an lvalue though).
arch/x86/include/asm/processor.h +3 -2
@@ -409,11 +409,12 @@ struct fixed_percpu_data { }; DECLARE_PER_CPU_FIRST(struct fixed_percpu_data, fixed_percpu_data) __visible; -DECLARE_INIT_PER_CPU (fixed_percpu_data);static inline unsigned long cpu_kernelmode_gs_base (int cpu) { -return (unsigned long )&per_cpu (fixed_percpu_data, cpu); +extern unsigned long __per_cpu_offset[]; + +return IS_ENABLED (CONFIG_SMP ) ? __per_cpu_offset[cpu] :0 ; } extern asmlinkage void entry_SYSCALL32_ignore(void);
arch/x86/kernel/head64.c +1 -1
@@ -559,7 +559,7 @@ void early_setup_idt(void)*/ void __head startup_64_setup_gdt_idt(void) { -struct desc_struct *gdt = (void *)(__force unsigned long )init_per_cpu_var (gdt_page.gdt); +struct desc_struct *gdt = (void *)(__force unsigned long )gdt_page.gdt;void *handler =NULL ;struct desc_ptr startup_gdt_descr = {
arch/x86/kernel/head_64.S +3 -9
@@ -68,11 +68,10 @@ SYM_CODE_START_NOALIGN(startup_64) /* Set up the stack for verify_cpu() */ leaq __top_init_kernel_stack(%rip), %rsp - /* Setup GSBASE to allow stack canary access for C code */ + /* Clear %gs so early per-CPU references target the per-CPU load area */ movl $MSR_GS_BASE, %ecx - leaq INIT_PER_CPU_VAR(fixed_percpu_data)(%rip), %rdx - movl %edx, %eax - shrq $32, %rdx + xorl %eax, %eax + cdq wrmsr call startup_64_setup_gdt_idt
@@ -361,15 +360,10 @@ SYM_INNER_LABEL(common_startup_64, SYM_L_LOCAL) /* Set up %gs. * - * The base of %gs always points to fixed_percpu_data. If the - * stack protector canary is enabled, it is located at %gs:40. * Note that, on SMP, the boot cpu uses init data section until * the per cpu areas are set up. */ movl $MSR_GS_BASE,%ecx -#ifndef CONFIG_SMP - leaq INIT_PER_CPU_VAR(fixed_percpu_data)(%rip), %rdx -#endif movl %edx, %eax shrq $32, %rdx wrmsr
arch/x86/kernel/irq_64.c +0 -1
@@ -27,7 +27,6 @@#include <asm/apic.h> DEFINE_PER_CPU_PAGE_ALIGNED(struct irq_stack, irq_stack_backing_store) __visible; -DECLARE_INIT_PER_CPU (irq_stack_backing_store);#ifdef CONFIG_VMAP_STACK /*
arch/x86/kernel/setup_percpu.c +1 -8
@@ -23,17 +23,10 @@#include <asm/cpumask.h> #include <asm/cpu.h> -#ifdef CONFIG_X86_64 -#define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load) -#else -#define BOOT_PERCPU_OFFSET 0 -#endif - -DEFINE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET; +DEFINE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off) = 0;EXPORT_PER_CPU_SYMBOL (this_cpu_off);unsigned long __per_cpu_offset[NR_CPUS ]__ro_after_init = { - [0 ...NR_CPUS -1 ] =BOOT_PERCPU_OFFSET , };EXPORT_SYMBOL (__per_cpu_offset);
arch/x86/kernel/vmlinux.lds.S +0 -30
@@ -103,9 +103,6 @@ PHDRS { text PT_LOAD FLAGS(5); /* R_E */ data PT_LOAD FLAGS(6); /* RW_ */ #ifdef CONFIG_X86_64 -#ifdef CONFIG_SMP - percpu PT_LOAD FLAGS(6); /* RW_ */ -#endif init PT_LOAD FLAGS(7); /* RWE */ #endif note PT_NOTE FLAGS(0); /* ___ */
@@ -225,17 +222,6 @@ SECTIONS __init_begin = .; /* paired with __init_end */ } -#if defined(CONFIG_X86_64) && defined(CONFIG_SMP) - /* - * percpu offsets are zero-based on SMP. PERCPU_VADDR() changes the - * output PHDR, so the next output section - .init.text - should - * start another segment - init. - */ - PERCPU_VADDR(INTERNODE_CACHE_BYTES, 0, :percpu) - ASSERT(SIZEOF(.data..percpu) < CONFIG_PHYSICAL_START, - "per-CPU data too large - increase CONFIG_PHYSICAL_START") -#endif - INIT_TEXT_SECTION(PAGE_SIZE) #ifdef CONFIG_X86_64 :init
@@ -356,9 +342,7 @@ SECTIONS EXIT_DATA } -#if !defined(CONFIG_X86_64) || !defined(CONFIG_SMP) PERCPU_SECTION(INTERNODE_CACHE_BYTES) -#endif RUNTIME_CONST(shift, d_hash_shift) RUNTIME_CONST(ptr, dentry_hashtable)
@@ -497,20 +481,6 @@ SECTIONS "kernel image bigger than KERNEL_IMAGE_SIZE"); #ifdef CONFIG_X86_64 -/* - * Per-cpu symbols which need to be offset from __per_cpu_load - * for the boot processor. - */ -#define INIT_PER_CPU(x) init_per_cpu__##x = ABSOLUTE(x) + __per_cpu_load -INIT_PER_CPU(gdt_page); -INIT_PER_CPU(fixed_percpu_data); -INIT_PER_CPU(irq_stack_backing_store); - -#ifdef CONFIG_SMP -. = ASSERT((fixed_percpu_data == 0), - "fixed_percpu_data is not at start of per-cpu area"); -#endif - #ifdef CONFIG_MITIGATION_UNRET_ENTRY . = ASSERT((retbleed_return_thunk & 0x3f) == 0, "retbleed_return_thunk not cacheline-aligned"); #endif
arch/x86/platform/pvh/head.S +2 -4
@@ -95,9 +95,9 @@ SYM_CODE_START_LOCAL(pvh_start_xen) /* 64-bit entry point. */ .code64 1: - /* Set base address in stack canary descriptor. */ + /* Clear %gs so early per-CPU references target the per-CPU load area */ mov $MSR_GS_BASE,%ecx - mov $_pa(canary), %eax + xor %eax, %eax xor %edx, %edx wrmsr
@@ -161,8 +161,6 @@ SYM_DATA_START_LOCAL(gdt_start) SYM_DATA_END_LABEL(gdt_start, SYM_L_LOCAL, gdt_end) .balign 16 -SYM_DATA_LOCAL(canary, .fill 48, 1, 0) - SYM_DATA_START_LOCAL(early_stack) .fill BOOT_STACK_SIZE, 1, 0 SYM_DATA_END_LABEL(early_stack, SYM_L_LOCAL, early_stack_end)
arch/x86/tools/relocs.c +1 -7
@@ -88,7 +88,6 @@ static const char * const sym_regex_kernel[S_NSYMTYPES] = { "(jiffies|jiffies_64)|"#if ELF_BITS ==64 "__per_cpu_load|" -"init_per_cpu__.*|" "__end_rodata_hpage_align|"#endif "__vvar_page|"
@@ -785,10 +784,6 @@ static void percpu_init(void)* The GNU linker incorrectly associates: * __init_begin * __per_cpu_load -* -* The "gold" linker incorrectly associates: -* init_per_cpu__fixed_percpu_data -* init_per_cpu__gdt_page */ static int is_percpu_sym (ElfW (Sym) *sym,const char *symname) {
@@ -796,8 +791,7 @@ static int is_percpu_sym(ElfW(Sym) *sym, const char *symname)return (shndx == per_cpu_shndx) &&strcmp (symname,"__init_begin" ) && -strcmp (symname,"__per_cpu_load" ) && -strncmp (symname,"init_per_cpu_" ,13 ); +strcmp (symname,"__per_cpu_load" ); }
arch/x86/xen/xen-head.S +2 -8
@@ -51,15 +51,9 @@ SYM_CODE_START(startup_xen) leaq __top_init_kernel_stack(%rip), %rsp - /* Set up %gs. - * - * The base of %gs always points to fixed_percpu_data. If the - * stack protector canary is enabled, it is located at %gs:40. - * Note that, on SMP, the boot cpu uses init data section until - * the per cpu areas are set up. - */ + /* Clear %gs so early per-CPU references target the per-CPU load area */ movl $MSR_GS_BASE,%ecx - movq $INIT_PER_CPU_VAR(fixed_percpu_data),%rax + xorl %eax, %eax cdq wrmsr
init/Kconfig +0 -1
@@ -1838,7 +1838,6 @@ config KALLSYMS_ALL config KALLSYMS_ABSOLUTE_PERCPU bool depends on KALLSYMS - default X86_64 && SMP # end of the "standard kernel features (expert users)" menu
Signature · 1 line
2.46.0.792.g87dc391469-goog
Re: [RFC PATCH 06/28] x86/percpu: Get rid of absolute per-CPU variable placement
Quoted 6 lines · Ard Biesheuvel
The
x86_64approach was needed to accommodate per-task stack protector cookies, which used to live at a fixed offset of GS+40, requiring GS to be treated as a base register. This is no longer the case, though, and so GS can be repurposed as a true per-CPU offset, adopting the same strategy as other architectures.
Quoted 6 lines
This also removes the need for linker tricks to emit the per-CPU ELF segment at a different virtual address. It also means RIP-relative per-CPU variables no longer need to be relocated in the opposite direction when KASLR is applied, which was necessary because the
0x0based per-CPU region remains in place even when the kernel is moved around.
Looks like a good cleanup. Hope it does not break anything that relies on structures %GS points to.
RFC PATCH 7/28 scripts/kallsyms: Avoid 0x0 as the relative base
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
In some cases, LLVM's lld linker may emit the following symbol into the symbol table
0000000000000000 ? _GLOBAL_OFFSET_TABLE_
and its presence throws off the relative base logic in kallsyms. Since 0x0 is never a valid relative base, just ignore it.
--- scripts/kallsyms.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
scripts/kallsyms.c +1 -1
@@ -747,7 +747,7 @@ static void record_relative_base(void)unsigned int i;for (i =0 ; i < table_cnt; i++) -if (!symbol_absolute (table[i])) { +if (table[i]->addr && !symbol_absolute (table[i])) {/* * The table is sorted by address. * Take the first non-absolute symbol value.
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 8/28 scripts/kallsyms: Remove support for absolute per-CPU variables
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
SMP on x86_64 no longer needs absolute per-CPU variables, so this support can be dropped from kallsyms as well, as no other architectures rely on this functionality.
--- init/Kconfig | 4 -- kernel/kallsyms.c | 12 +---- scripts/kallsyms.c | 51 +++----------------- scripts/link-vmlinux.sh | 4 -- 4 files changed, 9 insertions(+), 62 deletions(-)
init/Kconfig +0 -4
@@ -1835,10 +1835,6 @@ config KALLSYMS_ALL Say N unless you really need all symbols, or kernel live patching. -config KALLSYMS_ABSOLUTE_PERCPU - bool - depends on KALLSYMS - # end of the "standard kernel features (expert users)" menu config ARCH_HAS_MEMBARRIER_CALLBACKS
kernel/kallsyms.c +2 -10
@@ -148,16 +148,8 @@ static unsigned int get_symbol_offset(unsigned long pos)unsigned long kallsyms_sym_address (int idx) { -/* values are unsigned offsets if --absolute-percpu is not in effect */ -if (!IS_ENABLED (CONFIG_KALLSYMS_ABSOLUTE_PERCPU )) -return kallsyms_relative_base + (u32 )kallsyms_offsets[idx]; - -/* ...otherwise, positive offsets are absolute values */ -if (kallsyms_offsets[idx] >=0 ) -return kallsyms_offsets[idx]; - -/* ...and negative offsets are relative to kallsyms_relative_base - 1 */ -return kallsyms_relative_base -1 - kallsyms_offsets[idx]; +/* values are unsigned offsets */ +return kallsyms_relative_base + (u32 )kallsyms_offsets[idx]; }static unsigned int get_symbol_seq (int index)
scripts/kallsyms.c +7 -44
120 lines not on this page · raw message
scripts/link-vmlinux.sh +0 -4
@@ -140,10 +140,6 @@ kallsyms() kallsymopt="${kallsymopt} --all-symbols" fi - if is_enabled CONFIG_KALLSYMS_ABSOLUTE_PERCPU; then - kallsymopt="${kallsymopt} --absolute-percpu" - fi - info KSYMS "${2}.S" scripts/kallsyms ${kallsymopt} "${1}" > "${2}.S"
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 9/28 x86/tools: Remove special relocation handling for per-CPU variables
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Due to the placement of per-CPU variables in a special, 0x0 based disjoint memory segment in the ELF binary, the KASLR relocation tool needed to perform special processing for references to such variables, as they were not affected by KASLR displacement.
This meant that absolute references could be ignored, and RIP-relative references had to be compensated for KASLR, by applying the same offset but negated.
None of this is necessary any longer, so remove this handling from the relocation host tool.
--- arch/x86/boot/compressed/misc.c | 14 +-- arch/x86/tools/relocs.c | 130 +------------------- 2 files changed, 2 insertions(+), 142 deletions(-)
arch/x86/boot/compressed/misc.c +1 -13
@@ -235,7 +235,7 @@ static void handle_relocations(void *output, unsigned long output_len,/* * Process relocations: 32 bit relocations first then 64 bit after. -* Three sets of binary relocations are added to the end of the kernel +* Two sets of binary relocations are added to the end of the kernel * before compression. Each relocation table entry is the kernel * address of the location which needs to be updated stored as a * 32-bit value which is sign extended to 64 bits.
@@ -245,8 +245,6 @@ static void handle_relocations(void *output, unsigned long output_len,* kernel bits... * 0 - zero terminator for 64 bit relocations * 64 bit relocation repeated -* 0 - zero terminator for inverse 32 bit relocations -* 32 bit inverse relocation repeated * 0 - zero terminator for 32 bit relocations * 32 bit relocation repeated *
@@ -267,16 +265,6 @@ static void handle_relocations(void *output, unsigned long output_len,long extended = *reloc; extended += map; - ptr = (unsigned long )extended; -if (ptr < min_addr || ptr > max_addr) -error ("inverse 32-bit relocation outside of kernel!\n" ); - - *(int32_t *)ptr -= delta; - } -for (reloc--; *reloc; reloc--) { -long extended = *reloc; - extended += map; - ptr = (unsigned long )extended;if (ptr < min_addr || ptr > max_addr)error ("64-bit relocation outside of kernel!\n" );
arch/x86/tools/relocs.c +1 -129
195 lines not on this page · raw message
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 10/28 x86/xen: Avoid relocatable quantities in Xen ELF notes
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Xen puts virtual and physical addresses into ELF notes that are treated by the linker as relocatable by default. Doing so is not only pointless, given that the ELF notes are only intended for consumption by Xen before the kernel boots. It is also a KASLR leak, given that the kernel's ELF notes are exposed via the world readable /sys/kernel/notes.
So emit these constants in a way that prevents the linker from marking them as relocatable. This involves place-relative relocations (which subtract their own virtual address from the symbol value) and linker provided absolute symbols that add the address of the place to the desired value.
--- arch/x86/kernel/vmlinux.lds.S | 13 +++++++++++++ arch/x86/platform/pvh/head.S | 6 +++--- arch/x86/tools/relocs.c | 1 + arch/x86/xen/xen-head.S | 6 ++++-- 4 files changed, 21 insertions(+), 5 deletions(-)
arch/x86/kernel/vmlinux.lds.S +13 -0
@@ -111,6 +111,19 @@ PHDRS { SECTIONS { . = __START_KERNEL; + +#ifdef CONFIG_XEN_PV +xen_elfnote_entry_offset = + ABSOLUTE(xen_elfnote_entry) + ABSOLUTE(startup_xen); +xen_elfnote_hypercall_page_offset = + ABSOLUTE(xen_elfnote_hypercall_page) + ABSOLUTE(hypercall_page); +#endif + +#ifdef CONFIG_PVH +xen_elfnote_phys32_entry_offset = + ABSOLUTE(xen_elfnote_phys32_entry) + ABSOLUTE(pvh_start_xen - LOAD_OFFSET); +#endif + #ifdef CONFIG_X86_32 phys_startup_32 = ABSOLUTE(startup_32 - LOAD_OFFSET); #else
arch/x86/platform/pvh/head.S +3 -3
@@ -50,7 +50,7 @@ #define PVH_CS_SEL (PVH_GDT_ENTRY_CS * 8) #define PVH_DS_SEL (PVH_GDT_ENTRY_DS * 8) -SYM_CODE_START_LOCAL(pvh_start_xen) +SYM_CODE_START(pvh_start_xen) UNWIND_HINT_END_OF_STACK cld
@@ -165,5 +165,5 @@ SYM_DATA_START_LOCAL(early_stack) .fill BOOT_STACK_SIZE, 1, 0 SYM_DATA_END_LABEL(early_stack, SYM_L_LOCAL, early_stack_end) - ELFNOTE(Xen, XEN_ELFNOTE_PHYS32_ENTRY, - _ASM_PTR (pvh_start_xen - __START_KERNEL_map)) + ELFNOTE(Xen, XEN_ELFNOTE_PHYS32_ENTRY, .global xen_elfnote_phys32_entry; + xen_elfnote_phys32_entry: _ASM_PTR xen_elfnote_phys32_entry_offset - .)
arch/x86/tools/relocs.c +1 -0
@@ -57,6 +57,7 @@ static const char * const sym_regex_kernel[S_NSYMTYPES] = { [S_ABS] ="^(xen_irq_disable_direct_reloc$|" "xen_save_fl_direct_reloc$|" +"xen_elfnote_.+_offset$|" "VDSO|" "__kcfi_typeid_|" "__crc_)",
arch/x86/xen/xen-head.S +4 -2
@@ -88,7 +88,8 @@ SYM_CODE_END(xen_cpu_bringup_again) ELFNOTE(Xen, XEN_ELFNOTE_VIRT_BASE, _ASM_PTR __START_KERNEL_map) /* Map the p2m table to a 512GB-aligned user address. */ ELFNOTE(Xen, XEN_ELFNOTE_INIT_P2M, .quad (PUD_SIZE * PTRS_PER_PUD)) - ELFNOTE(Xen, XEN_ELFNOTE_ENTRY, _ASM_PTR startup_xen) + ELFNOTE(Xen, XEN_ELFNOTE_ENTRY, .globl xen_elfnote_entry; + xen_elfnote_entry: _ASM_PTR xen_elfnote_entry_offset - .) ELFNOTE(Xen, XEN_ELFNOTE_FEATURES, .ascii "!writable_page_tables") ELFNOTE(Xen, XEN_ELFNOTE_PAE_MODE, .asciz "yes") ELFNOTE(Xen, XEN_ELFNOTE_L1_MFN_VALID,
@@ -109,7 +110,8 @@ SYM_CODE_END(xen_cpu_bringup_again) #else # define FEATURES_DOM0 0 #endif - ELFNOTE(Xen, XEN_ELFNOTE_HYPERCALL_PAGE, _ASM_PTR hypercall_page) + ELFNOTE(Xen, XEN_ELFNOTE_HYPERCALL_PAGE, .globl xen_elfnote_hypercall_page; + xen_elfnote_hypercall_page: _ASM_PTR xen_elfnote_hypercall_page_offset - .) ELFNOTE(Xen, XEN_ELFNOTE_SUPPORTED_FEATURES, .long FEATURES_PV | FEATURES_PVH | FEATURES_DOM0) ELFNOTE(Xen, XEN_ELFNOTE_LOADER, .asciz "generic")
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 11/28 x86/pvh: Avoid absolute symbol references in .head.text
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
The .head.text section contains code that may execute from a different address than it was linked at. This is fragile, given that the x86 ABI can refer to global symbols via absolute or relative references, and the toolchain assumes that these are interchangeable, which they are not in this particular case.
In the case of the PVH code, there are some additional complications:
- the absolute references are in 32-bit code, which get emitted with R_X86_64_32 relocations, and these are not permitted in PIE code;
- the code in question is not actually relocatable: it can only run correctly from the physical load address specified in the ELF note.
So rewrite the code to only rely on relative symbol references: these are always 32-bits wide, even in 64-bit code, and are resolved by the linker at build time.
--- arch/x86/platform/pvh/head.S | 39 ++++++++++++++------ 1 file changed, 27 insertions(+), 12 deletions(-)
arch/x86/platform/pvh/head.S +27 -12
@@ -54,7 +54,20 @@ SYM_CODE_START(pvh_start_xen) UNWIND_HINT_END_OF_STACK cld - lgdt (_pa(gdt)) + /* + * This is position dependent code that can only execute correctly from + * the physical address that the kernel was linked to run at. Use the + * symbols emitted for the ELF note to construct the build time physical + * address of pvh_start_xen(), without relying on absolute 32-bit ELF + * relocations, as these are not supported by the linker when running in + * -pie mode, and should be avoided in .head.text in general. + */ +0: mov $xen_elfnote_phys32_entry_offset - 0b, %ebp + sub $xen_elfnote_phys32_entry - 0b, %ebp + + lea (gdt - pvh_start_xen)(%ebp), %eax + add %eax, 2(%eax) + lgdt (%eax) mov $PVH_DS_SEL,%eax mov %eax,%ds
@@ -62,14 +75,14 @@ SYM_CODE_START(pvh_start_xen) mov %eax,%ss /* Stash hvm_start_info. */ - mov $_pa(pvh_start_info), %edi + lea (pvh_start_info - pvh_start_xen)(%ebp), %edi mov %ebx, %esi - mov _pa(pvh_start_info_sz), %ecx + mov (pvh_start_info_sz - pvh_start_xen)(%ebp), %ecx shr $2,%ecx rep movsl - mov $_pa(early_stack_end), %esp + lea (early_stack_end - pvh_start_xen)(%ebp), %esp /* Enable PAE mode. */ mov %cr4, %eax
@@ -84,17 +97,21 @@ SYM_CODE_START(pvh_start_xen) wrmsr /* Enable pre-constructed page tables. */ - mov $_pa(init_top_pgt), %eax + lea (init_top_pgt - pvh_start_xen)(%ebp), %eax mov %eax, %cr3 mov $(X86_CR0_PG | X86_CR0_PE), %eax mov %eax, %cr0 /* Jump to 64-bit mode. */ - ljmp $PVH_CS_SEL, $_pa(1f) + lea (1f - pvh_start_xen)(%ebp), %eax + push $PVH_CS_SEL + push %eax + lret /* 64-bit entry point. */ .code64 1: + UNWIND_HINT_END_OF_STACK /* Clear %gs so early per-CPU references target the per-CPU load area */ mov $MSR_GS_BASE,%ecx xor %eax, %eax
@@ -108,10 +125,8 @@ SYM_CODE_START(pvh_start_xen) call *%rax /* startup_64 expects boot_params in %rsi. */ - mov $_pa(pvh_bootparams), %rsi - mov $_pa(startup_64), %rax - ANNOTATE_RETPOLINE_SAFE - jmp *%rax + lea pvh_bootparams(%rip), %rsi + jmp startup_64 #else /* CONFIG_X86_64 */
@@ -146,8 +161,8 @@ SYM_CODE_END(pvh_start_xen) .section ".init.data","aw" .balign 8 SYM_DATA_START_LOCAL(gdt) - .word gdt_end - gdt_start - .long _pa(gdt_start) + .word gdt_end - gdt_start - 1 + .long gdt_start - gdt .word 0 SYM_DATA_END(gdt) SYM_DATA_START_LOCAL(gdt_start)
Signature · 1 line
2.46.0.792.g87dc391469-goog
Re: [RFC PATCH 11/28] x86/pvh: Avoid absolute symbol references in .head.text
Hi Ard,
Quoted 14 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
The .head.text section contains code that may execute from a different address than it was linked at. This is fragile, given that the x86 ABI can refer to global symbols via absolute or relative references, and the toolchain assumes that these are interchangeable, which they are not in this particular case.
In the case of the PVH code, there are some additional complications:
- the absolute references are in 32-bit code, which get emitted withR_X86_64_32relocations, and these are not permitted in PIE code;
- the code in question is not actually relocatable: it can only run correctly from the physical load address specified in the ELF note.
Quoted 5 lines
So rewrite the code to only rely on relative symbol references: these are always 32-bits wide, even in 64-bit code, and are resolved by the linker at build time.
Signed-off-by Ard Biesheuvel <[email protected]>
Juergen queued up my patches to make the PVH entry point position independent (5 commits):
https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=linux-next
My commit that corresponds to this patch of yours is:
https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/commit/?h=linux-next&id=1db29f99edb056d8445876292f53a63459142309
(There are more changes to handle adjusting the page tables.)
Regards,
Jason
Ard Biesheuvel
maintainerRe: [RFC PATCH 11/28] x86/pvh: Avoid absolute symbol references in .head.text
Quoted 24 lines · Jason Andryuk
Hi Ard,
Quoted 14 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
The .head.text section contains code that may execute from a different address than it was linked at. This is fragile, given that the x86 ABI can refer to global symbols via absolute or relative references, and the toolchain assumes that these are interchangeable, which they are not in this particular case.
In the case of the PVH code, there are some additional complications:
- the absolute references are in 32-bit code, which get emitted withR_X86_64_32relocations, and these are not permitted in PIE code;
- the code in question is not actually relocatable: it can only run correctly from the physical load address specified in the ELF note.Quoted 5 lines
So rewrite the code to only rely on relative symbol references: these are always 32-bits wide, even in 64-bit code, and are resolved by the linker at build time.
Signed-off-by Ard Biesheuvel <[email protected]>
Quoted 9 lines
Juergen queued up my patches to make the PVH entry point position independent (5 commits):
https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=linux-nextMy commit that corresponds to this patch of yours is:
https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/commit/?h=linux-next&id=1db29f99edb056d8445876292f53a63459142309(There are more changes to handle adjusting the page tables.)
Thanks for the head's up. Those changes look quite similar, so I guess I should just rebase my stuff onto the xen tree.
The only thing that I would like to keep from my version is
+ lea (gdt - pvh_start_xen)(%ebp), %eax
+ add %eax, 2(%eax)
+ lgdt (%eax)
and
- .word gdt_end - gdt_start
- .long _pa(gdt_start)
+ .word gdt_end - gdt_start - 1
+ .long gdt_start - gdt
The first line is a bugfix, btw, so perhaps I should send that out separately. But my series relies on all 32-bit absolute symbol references being removed, since the linker rejects those when running in PIE mode, and so the second line is needed to get rid of the _pa() there.
Re: [RFC PATCH 11/28] x86/pvh: Avoid absolute symbol references in .head.text
Quoted 35 lines · Ard Biesheuvel
Quoted 24 lines · Jason Andryuk
Hi Ard,
Quoted 11 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
The .head.text section contains code that may execute from a different address than it was linked at. This is fragile, given that the x86 ABI can refer to global symbols via absolute or relative references, and the toolchain assumes that these are interchangeable, which they are not in this particular case.
In the case of the PVH code, there are some additional complications:
- the absolute references are in 32-bit code, which get emitted withR_X86_64_32 relocations, and these are not permitted in PIE code;Quoted 8 lines
- the code in question is not actually relocatable: it can only run correctly from the physical load address specified in the ELF note.
So rewrite the code to only rely on relative symbol references: these are always 32-bits wide, even in 64-bit code, and are resolved by the linker at build time.
Signed-off-by Ard Biesheuvel <[email protected]>Quoted 9 lines
Juergen queued up my patches to make the PVH entry point position independent (5 commits):
https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/log/?h=linux-nextMy commit that corresponds to this patch of yours is:
https://git.kernel.org/pub/scm/linux/kernel/git/xen/tip.git/commit/?h=linux-next&id=1db29f99edb056d8445876292f53a63459142309(There are more changes to handle adjusting the page tables.)
Quoted 6 lines
Thanks for the head's up. Those changes look quite similar, so I guess I should just rebase my stuff onto the xen tree.
The only thing that I would like to keep from my version is
+ lea (gdt - pvh_start_xen)(%ebp), %eax
If you rebase on top of the xen tree, using rva() would match the rest of the code:
lea rva(gdt)(%ebp), %eax
Quoted 10 lines
+ add %eax, 2(%eax)
+ lgdt (%eax)and
- .word
gdt_end-gdt_start
- .long_pa(gdt_start)
+ .wordgdt_end-gdt_start- 1
+ .longgdt_start- gdt
Quoted 5 lines
The first line is a bugfix, btw, so perhaps I should send that out separately. But my series relies on all 32-bit absolute symbol references being removed, since the linker rejects those when running in PIE mode, and so the second line is needed to get rid of the
_pa()there.
Sounds good.
Regards,
Jason
RFC PATCH 12/28 x86/pm-trace: Use RIP-relative accesses for .tracedata
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Use RIP-relative accesses and 32-bit offsets for .tracedata, to avoid the need for relocation fixups at boot time.
--- arch/x86/include/asm/pm-trace.h | 4 ++-- drivers/base/power/trace.c | 6 +++--- 2 files changed, 5 insertions(+), 5 deletions(-)
arch/x86/include/asm/pm-trace.h +2 -2
@@ -8,10 +8,10 @@ do { \ if (pm_trace_enabled) { \ const void *tracedata; \ - asm volatile(_ASM_MOV " $1f,%0\n" \ + asm volatile("lea " _ASM_RIP(1f) ", %0\n" \ ".section .tracedata,\"a\"\n" \ "1:\t.word %c1\n\t" \ - _ASM_PTR " %c2\n" \ + ".long %c2 - .\n" \ ".previous" \ :"=r" (tracedata) \ : "i" (__LINE__), "i" (__FILE__)); \
drivers/base/power/trace.c +3 -3
@@ -167,7 +167,7 @@ EXPORT_SYMBOL(set_trace_device);void generate_pm_trace (const void *tracedata,unsigned int user) {unsigned short lineno = *(unsigned short *)tracedata; -const char *file = *(const char **)(tracedata +2 ); +const char *file =offset_to_ptr ((int *)(tracedata +2 ));unsigned int user_hash_value, file_hash_value;if (!x86_platform.legacy.rtc)
@@ -187,9 +187,9 @@ static int show_file_hash(unsigned int value) match =0 ;for (tracedata = __tracedata_start ; tracedata < __tracedata_end ; - tracedata +=2 +sizeof (unsigned long )) { + tracedata +=2 +sizeof (int )) {unsigned short lineno = *(unsigned short *)tracedata; -const char *file = *(const char **)(tracedata +2 ); +const char *file =offset_to_ptr ((int *)(tracedata +2 ));unsigned int hash =hash_string (lineno, file,FILEHASH );if (hash != value)continue ;
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 13/28 x86/kvm: Use RIP-relative addressing
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Avoid absolute references in code, which require fixing up at boot time, and replace them with RIP-relative ones. In this particular case, due to the register pressure, they cannot be avoided entirely, so one absolute reference is retained but the resulting reference via the GOT is compatible with running the linker in PIE mode.
--- arch/x86/kernel/kvm.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
arch/x86/kernel/kvm.c +5 -3
@@ -800,9 +800,11 @@ extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);* Hand-optimize version for x86-64 to avoid 8 64-bit register saving and * restoring to/from the stack. */ -#define PV_VCPU_PREEMPTED_ASM \ - "movq __per_cpu_offset(,%rdi,8), %rax\n\t" \ - "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax)\n\t" \ +#define PV_VCPU_PREEMPTED_ASM \ + "leaq __per_cpu_offset(%rip), %rax \n\t" \ + "movq (%rax,%rdi,8), %rax \n\t" \ + "addq steal_time@GOTPCREL(%rip), %rax \n\t" \ + "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "(%rax) \n\t" \ "setne %al\n\t" DEFINE_ASM_FUNC(__raw_callee_save___kvm_vcpu_is_preempted,
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 14/28 x86/rethook: Use RIP-relative reference for return address
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Instead of pushing an immediate absolute address, which is incompatible with PIE codegen or linking, use a LEA instruction to take the address into a register.
--- arch/x86/kernel/rethook.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
arch/x86/kernel/rethook.c +2 -1
@@ -27,7 +27,8 @@ asm(#ifdef CONFIG_X86_64 ANNOTATE_NOENDBR /* This is only jumped from ret instruction */ /* Push a fake return address to tell the unwinder it's a rethook. */ -" pushq $arch_rethook_trampoline\n" +" leaq arch_rethook_trampoline(%rip), %rdi\n" +" pushq %rdi\n" UNWIND_HINT_FUNC " pushq $" __stringify(__KERNEL_DS) "\n"/* Save the 'sp - 16', this will be fixed later. */
Signature · 1 line
2.46.0.792.g87dc391469-goog
Re: [RFC PATCH 14/28] x86/rethook: Use RIP-relative reference for return address
Quoted 4 lines · Ard Biesheuvel
Instead of pushing an immediate absolute address, which is incompatible with PIE codegen or linking, use a LEA instruction to take the address into a register.
I don't think you can do this - it corrupts %rdi.
Yes, the code uses %rdi later, but that's inside the SAVE_REGS_STRING / RESTORE_REGS_STRING area.
And we do have special calling conventions that aren't the regular ones, so %rdi might actually be used elsewhere. For example, __get_user_X and __put_user_X all have magical calling conventions:
they don't actually use %rdi, but part of the calling convention is that the unused registers aren't modified.
Of course, I'm not actually sure you can probe those and trigger this issue, but it all makes me think it's broken.
And it's entirely possible that I'm wrong for some reason, but this just _looks_ very very wrong to me.
I think you can do this with a "pushq mem" instead, and put the relocation into the memory location.
Linus
Ard Biesheuvel
maintainerRe: [RFC PATCH 14/28] x86/rethook: Use RIP-relative reference for return address
Quoted 12 lines · Linus Torvalds
Quoted 4 lines · Ard Biesheuvel
Instead of pushing an immediate absolute address, which is incompatible with PIE codegen or linking, use a LEA instruction to take the address into a register.
I don't think you can do this - it corrupts %rdi.
Yes, the code uses %rdi later, but that's inside the
SAVE_REGS_STRING/RESTORE_REGS_STRINGarea.
Oops, I missed that.
Quoted 6 lines
And we do have special calling conventions that aren't the regular ones, so %rdi might actually be used elsewhere. For example,
__get_user_Xand__put_user_Xall have magical calling conventions:
they don't actually use %rdi, but part of the calling convention is that the unused registers aren't modified.
Quoted 9 lines
Of course, I'm not actually sure you can probe those and trigger this issue, but it all makes me think it's broken.
And it's entirely possible that I'm wrong for some reason, but this just _looks_ very very wrong to me.
I think you can do this with a "pushq mem" instead, and put the relocation into the memory location.
I'll change this into
pushq arch_rethook_trampoline@GOTPCREL(%rip)
which I had originally. I was trying to avoid the load from memory, but that obviously only works if the register is not live.
Re: [RFC PATCH 14/28] x86/rethook: Use RIP-relative reference for return address
Quoted 17 lines · Ard Biesheuvel
Quoted 6 lines · Linus Torvalds
And we do have special calling conventions that aren't the regular ones, so %rdi might actually be used elsewhere. For example,
__get_user_Xand__put_user_Xall have magical calling conventions:
they don't actually use %rdi, but part of the calling convention is that the unused registers aren't modified.Quoted 9 lines
Of course, I'm not actually sure you can probe those and trigger this issue, but it all makes me think it's broken.
And it's entirely possible that I'm wrong for some reason, but this just _looks_ very very wrong to me.
I think you can do this with a "pushq mem" instead, and put the relocation into the memory location.
Quoted 6 lines
I'll change this into
pushq arch_rethook_trampoline@GOTPCREL(%rip)
which I had originally. I was trying to avoid the load from memory, but that obviously only works if the register is not live.
But does that work? Won't that will push the 8 bytes from the start of arch_rethook_trampoline, when what's wanted is simply the address of arch_rethook_trampoline itself.
~Andrew
Re: [RFC PATCH 14/28] x86/rethook: Use RIP-relative reference for return address
Quoted 20 lines · Andrew Cooper
Quoted 19 lines · Ard Biesheuvel
Quoted 6 lines · Linus Torvalds
And we do have special calling conventions that aren't the regular ones, so %rdi might actually be used elsewhere. For example,
__get_user_Xand__put_user_Xall have magical calling conventions:
they don't actually use %rdi, but part of the calling convention is that the unused registers aren't modified.Quoted 9 lines
Of course, I'm not actually sure you can probe those and trigger this issue, but it all makes me think it's broken.
And it's entirely possible that I'm wrong for some reason, but this just _looks_ very very wrong to me.
I think you can do this with a "pushq mem" instead, and put the relocation into the memory location.
I'll change this into
Quoted 8 lines
Quoted 4 lines
pushq arch_rethook_trampoline@GOTPCREL(%rip)
which I had originally. I was trying to avoid the load from memory, but that obviously only works if the register is not live.
But does that work? Won't that will push the 8 bytes from the start of
arch_rethook_trampoline, when what's wanted is simply the address ofarch_rethook_trampolineitself.
What you describe is
pushq arch_rethook_trampoline(%rip)
The @GOTPCREL makes the PUSH access an item from the GOT, and that item is arch_rethook_trampoline's address.
Jan
RFC PATCH 15/28 x86/sync_core: Use RIP-relative addressing
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Use RIP-relative accesses and avoid fixups at runtime.
--- arch/x86/include/asm/sync_core.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
arch/x86/include/asm/sync_core.h +2 -1
@@ -31,7 +31,8 @@ static inline void iret_to_self(void)"pushfq\n\t" "mov %%cs, %0\n\t" "pushq %q0\n\t" -"pushq $1f\n\t" +"leaq 1f(%%rip), %q0\n\t" +"pushq %q0\n\t" "iretq\n\t" "1:" : "=&r" (tmp), ASM_CALL_CONSTRAINT : : "cc", "memory");
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 16/28 x86/entry_64: Use RIP-relative addressing
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Fix up a couple of occurrences in the x86_64 entry code where we take the absolute address of a symbol while we could use RIP-relative addressing just the same. This avoids relocation fixups at boot for these quantities.
--- arch/x86/entry/calling.h | 9 +++++---- arch/x86/entry/entry_64.S | 12 +++++++----- 2 files changed, 12 insertions(+), 9 deletions(-)
arch/x86/entry/calling.h +5 -4
@@ -375,8 +375,8 @@ For 32-bit we have the following conventions - kernel is built with .endm .macro SAVE_AND_SET_GSBASE scratch_reg:req save_reg:req + GET_PERCPU_BASE \scratch_reg \save_reg rdgsbase \save_reg - GET_PERCPU_BASE \scratch_reg wrgsbase \scratch_reg .endm
@@ -412,15 +412,16 @@ For 32-bit we have the following conventions - kernel is built with* Thus the kernel would consume a guest's TSC_AUX if an NMI arrives * while running KVM's run loop. */ -.macro GET_PERCPU_BASE reg:req +.macro GET_PERCPU_BASE reg:req scratch:req LOAD_CPU_AND_NODE_SEG_LIMIT \reg andq $VDSO_CPUNODE_MASK, \reg - movq __per_cpu_offset(, \reg, 8), \reg + leaq __per_cpu_offset(%rip), \scratch + movq (\scratch, \reg, 8), \reg .endm #else -.macro GET_PERCPU_BASE reg:req +.macro GET_PERCPU_BASE reg:req scratch:req movq pcpu_unit_offsets(%rip), \reg .endm
arch/x86/entry/entry_64.S +7 -5
@@ -1038,7 +1038,8 @@ SYM_CODE_START(error_entry) movl %ecx, %eax /* zero extend */ cmpq %rax, RIP+8(%rsp) je .Lbstep_iret - cmpq $.Lgs_change, RIP+8(%rsp) + leaq .Lgs_change(%rip), %rcx + cmpq %rcx, RIP+8(%rsp) jne .Lerror_entry_done_lfence /*
@@ -1250,10 +1251,10 @@ SYM_CODE_START(asm_exc_nmi) * the outer NMI. */ - movq $repeat_nmi, %rdx + leaq repeat_nmi(%rip), %rdx cmpq 8(%rsp), %rdx ja 1f - movq $end_repeat_nmi, %rdx + leaq end_repeat_nmi(%rip), %rdx cmpq 8(%rsp), %rdx ja nested_nmi_out 1:
@@ -1307,7 +1308,8 @@ nested_nmi: pushq %rdx pushfq pushq $__KERNEL_CS - pushq $repeat_nmi + leaq repeat_nmi(%rip), %rdx + pushq %rdx /* Put stack back */ addq $(6*8), %rsp
@@ -1346,7 +1348,7 @@ first_nmi: addq $8, (%rsp) /* Fix up RSP */ pushfq /* RFLAGS */ pushq $__KERNEL_CS /* CS */ - pushq $1f /* RIP */ + pushq 1f@GOTPCREL(%rip) /* RIP */ iretq /* continues at repeat_nmi below */ UNWIND_HINT_IRET_REGS 1:
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 17/28 x86/hibernate: Prefer RIP-relative accesses
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Replace some absolute symbol references with RIP-relative ones, so we don't need to fix them up at boot.
--- arch/x86/power/hibernate_asm_64.S | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
arch/x86/power/hibernate_asm_64.S +2 -2
@@ -39,7 +39,7 @@ SYM_FUNC_START(restore_registers) movq %rax, %cr4; # turn PGE back on /* We don't restore %rax, it must be 0 anyway */ - movq $saved_context, %rax + leaq saved_context(%rip), %rax movq pt_regs_sp(%rax), %rsp movq pt_regs_bp(%rax), %rbp movq pt_regs_si(%rax), %rsi
@@ -70,7 +70,7 @@ SYM_FUNC_START(restore_registers) SYM_FUNC_END(restore_registers) SYM_FUNC_START(swsusp_arch_suspend) - movq $saved_context, %rax + leaq saved_context(%rip), %rax movq %rsp, pt_regs_sp(%rax) movq %rbp, pt_regs_bp(%rax) movq %rsi, pt_regs_si(%rax)
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 18/28 x86/boot/64: Determine VA/PA offset before entering C code
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Implicit absolute symbol references (e.g., taking the address of a global variable) must be avoided in the C code that runs from the early 1:1 mapping of the kernel, given that this is a practice that violates assumptions on the part of the toolchain. I.e., RIP-relative and absolute references are expected to produce the same values, and so the compiler is free to choose either. However, the code currently assumes that RIP-relative references are never emitted here.
So an explicit virtual-to-physical offset needs to be used instead to derive the kernel virtual addresses of _text and _end, instead of simply taking the addresses and assuming that the compiler will not choose to use a RIP-relative references in this particular case.
Currently, phys_base is already used to perform such calculations, but it is derived from the kernel virtual address of _text, which is taken using an implicit absolute symbol reference. So instead, derive this VA-to-PA offset in asm code, using the kernel VA of common_startup_64 (which we already keep in a global variable for other reasons), and pass it to the C startup code.
--- arch/x86/include/asm/setup.h | 2 +- arch/x86/kernel/head64.c | 8 +++++--- arch/x86/kernel/head_64.S | 9 ++++++++- 3 files changed, 14 insertions(+), 5 deletions(-)
arch/x86/include/asm/setup.h +1 -1
@@ -49,7 +49,7 @@ extern unsigned long saved_video_mode;extern void reserve_standard_io_resources (void );extern void i386_reserve_resources (void ); -extern unsigned long __startup_64 (unsigned long physaddr,struct boot_params *bp); +extern unsigned long __startup_64 (unsigned long p2v_offset,struct boot_params *bp);extern void startup_64_setup_gdt_idt (void );extern void early_setup_idt (void );extern void __init do_early_exception (struct pt_regs *regs,int trapnr);
arch/x86/kernel/head64.c +5 -3
@@ -138,12 +138,14 @@ static unsigned long __head sme_postprocess_startup(struct boot_params *bp, pmdv* doesn't have to generate PC-relative relocations when accessing globals from * that function. Clang actually does not generate them, which leads to * boot-time crashes. To work around this problem, every global pointer must -* be accessed using RIP_REL_REF(). +* be accessed using RIP_REL_REF(). Kernel virtual addresses can be determined +* by subtracting p2v_offset from the RIP-relative address. */ -unsigned long __head __startup_64 (unsigned long physaddr, +unsigned long __head __startup_64 (unsigned long p2v_offset,struct boot_params *bp) {pmd_t (*early_pgts)[PTRS_PER_PMD ] =RIP_REL_REF (early_dynamic_pgts); +unsigned long physaddr = (unsigned long )&RIP_REL_REF (_text);unsigned long pgtable_flags;unsigned long load_delta;pgdval_t *pgd;
@@ -163,7 +165,7 @@ unsigned long __head __startup_64(unsigned long physaddr,* Compute the delta between the address I am compiled to run at * and the address I am actually running at. */ - load_delta = physaddr - (unsigned long )(_text - __START_KERNEL_map); + load_delta = __START_KERNEL_map + p2v_offset;RIP_REL_REF (phys_base) = load_delta;/* Is the address not 2M aligned? */
arch/x86/kernel/head_64.S +8 -1
@@ -99,13 +99,20 @@ SYM_CODE_START_NOALIGN(startup_64) /* Sanitize CPU configuration */ call verify_cpu + /* + * Use the 1:1 physical and kernel virtual addresses of + * common_startup_64 to determine the physical-to-virtual offset, and + * pass it as the first argument to __startup_64(). + */ + leaq common_startup_64(%rip), %rdi + subq 0f(%rip), %rdi + /* * Perform pagetable fixups. Additionally, if SME is active, encrypt * the kernel and retrieve the modifier (SME encryption mask if SME * is active) to be added to the initial pgdir entry that will be * programmed into CR3. */ - leaq _text(%rip), %rdi movq %r15, %rsi call __startup_64
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 19/28 x86/boot/64: Avoid intentional absolute symbol references in .head.text
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
The code in .head.text executes from a 1:1 mapping and cannot generally refer to global variables using their kernel virtual addresses. However, there are some occurrences of such references that are valid: the kernel virtual addresses of _text and _end are needed to populate the page tables correctly, and some other section markers are used in a similar way.
To avoid the need for making exceptions to the rule that .head.text must not contain any absolute symbol references, derive these addresses from the RIP-relative 1:1 mapped physical addresses, which can be safely determined using RIP_REL_REF().
--- arch/x86/kernel/head64.c | 30 ++++++++++++-------- 1 file changed, 18 insertions(+), 12 deletions(-)
arch/x86/kernel/head64.c +18 -12
@@ -91,9 +91,11 @@ static inline bool check_la57_support(void)return true; } -static unsigned long __head sme_postprocess_startup (struct boot_params *bp,pmdval_t *pmd) +static unsigned long __head sme_postprocess_startup (struct boot_params *bp, +pmdval_t *pmd, +unsigned long p2v_offset) { -unsigned long vaddr, vaddr_end; +unsigned long paddr, paddr_end;int i;/* Encrypt the kernel and related (if SME is active) */
@@ -106,10 +108,10 @@ static unsigned long __head sme_postprocess_startup(struct boot_params *bp, pmdv* attribute. */ if (sme_get_me_mask ()) { - vaddr = (unsigned long )__start_bss_decrypted; - vaddr_end = (unsigned long )__end_bss_decrypted; + paddr = (unsigned long )&RIP_REL_REF (__start_bss_decrypted); + paddr_end = (unsigned long )&RIP_REL_REF (__end_bss_decrypted); -for (; vaddr < vaddr_end; vaddr +=PMD_SIZE ) { +for (; paddr < paddr_end; paddr +=PMD_SIZE ) {/* * On SNP, transition the page to shared in the RMP table so that * it is consistent with the page table attribute change.
@@ -118,11 +120,11 @@ static unsigned long __head sme_postprocess_startup(struct boot_params *bp, pmdv* mapping (kernel .text). PVALIDATE, by way of * early_snp_set_memory_shared(), requires a valid virtual * address but the kernel is currently running off of the identity -* mapping so use __pa() to get a *currently* valid virtual address. +* mapping so use the PA to get a *currently* valid virtual address. */ -early_snp_set_memory_shared (__pa (vaddr),__pa (vaddr),PTRS_PER_PMD ); +early_snp_set_memory_shared (paddr, paddr,PTRS_PER_PMD ); - i =pmd_index (vaddr); + i =pmd_index (paddr - p2v_offset); pmd[i] -=sme_get_me_mask (); } }
@@ -146,6 +148,7 @@ unsigned long __head __startup_64(unsigned long p2v_offset, {pmd_t (*early_pgts)[PTRS_PER_PMD ] =RIP_REL_REF (early_dynamic_pgts);unsigned long physaddr = (unsigned long )&RIP_REL_REF (_text); +unsigned long va_text, va_end;unsigned long pgtable_flags;unsigned long load_delta;pgdval_t *pgd;
@@ -172,6 +175,9 @@ unsigned long __head __startup_64(unsigned long p2v_offset,if (load_delta & ~PMD_MASK )for (;;); + va_text = physaddr - p2v_offset; + va_end = (unsigned long )&RIP_REL_REF (_end) - p2v_offset; +/* Include the SME encryption mask in the fixup value */ load_delta +=sme_get_me_mask ();
@@ -232,7 +238,7 @@ unsigned long __head __startup_64(unsigned long p2v_offset, pmd_entry +=sme_get_me_mask (); pmd_entry += physaddr; -for (i =0 ; i <DIV_ROUND_UP (_end - _text,PMD_SIZE ); i++) { +for (i =0 ; i <DIV_ROUND_UP (va_end - va_text,PMD_SIZE ); i++) {int idx = i + (physaddr >>PMD_SHIFT ); pmd[idx %PTRS_PER_PMD ] = pmd_entry + i *PMD_SIZE ;
@@ -257,11 +263,11 @@ unsigned long __head __startup_64(unsigned long p2v_offset, pmd = &RIP_REL_REF (level2_kernel_pgt)->pmd;/* invalidate pages before the kernel image */ -for (i =0 ; i <pmd_index ((unsigned long )_text); i++) +for (i =0 ; i <pmd_index (va_text); i++) pmd[i] &= ~_PAGE_PRESENT;/* fixup pages that are part of the kernel image */ -for (; i <=pmd_index ((unsigned long )_end); i++) +for (; i <=pmd_index (va_end); i++)if (pmd[i] & _PAGE_PRESENT) pmd[i] += load_delta;
@@ -269,7 +275,7 @@ unsigned long __head __startup_64(unsigned long p2v_offset,for (; i <PTRS_PER_PMD ; i++) pmd[i] &= ~_PAGE_PRESENT; -return sme_postprocess_startup (bp, pmd); +return sme_postprocess_startup (bp, pmd, p2v_offset); }/* Wipe all early page tables except for the kernel symbol map */
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 20/28 x64/acpi: Use PIC-compatible references in wakeup_64.S
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Use ordinary RIP-relative references to make the code compatible with running the linker in PIE mode.
Note that wakeup_long64() runs in the kernel's ordinary virtual mapping so there is no need to record the address of .Lresume_point in a global variable. And fix the comment while at it.
--- arch/x86/kernel/acpi/wakeup_64.S | 11 ++++------- 1 file changed, 4 insertions(+), 7 deletions(-)
arch/x86/kernel/acpi/wakeup_64.S +4 -7
@@ -14,7 +14,7 @@ .code64 /* - * Hooray, we are in Long 64-bit mode (but still running in low memory) + * Hooray, we are in Long 64-bit mode */ SYM_FUNC_START(wakeup_long64) movq saved_magic(%rip), %rax
@@ -40,7 +40,7 @@ SYM_FUNC_START(wakeup_long64) movq saved_rsi(%rip), %rsi movq saved_rbp(%rip), %rbp - movq saved_rip(%rip), %rax + leaq .Lresume_point(%rip), %rax ANNOTATE_RETPOLINE_SAFE jmp *%rax SYM_FUNC_END(wakeup_long64)
@@ -51,7 +51,7 @@ SYM_FUNC_START(do_suspend_lowlevel) xorl %eax, %eax call save_processor_state - movq $saved_context, %rax + leaq saved_context(%rip), %rax movq %rsp, pt_regs_sp(%rax) movq %rbp, pt_regs_bp(%rax) movq %rsi, pt_regs_si(%rax)
@@ -70,8 +70,6 @@ SYM_FUNC_START(do_suspend_lowlevel) pushfq popq pt_regs_flags(%rax) - movq $.Lresume_point, saved_rip(%rip) - movq %rsp, saved_rsp(%rip) movq %rbp, saved_rbp(%rip) movq %rbx, saved_rbx(%rip)
@@ -88,7 +86,7 @@ SYM_FUNC_START(do_suspend_lowlevel) .align 4 .Lresume_point: /* We don't restore %rax, it must be 0 anyway */ - movq $saved_context, %rax + leaq saved_context(%rip), %rax movq saved_context_cr4(%rax), %rbx movq %rbx, %cr4 movq saved_context_cr3(%rax), %rbx
@@ -137,7 +135,6 @@ saved_rsi: .quad 0 saved_rdi: .quad 0 saved_rbx: .quad 0 -saved_rip: .quad 0 saved_rsp: .quad 0 SYM_DATA(saved_magic, .quad 0)
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 21/28 x86/head: Use PIC-compatible symbol references in startup code
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Use RIP-relative symbol references to make them compatible with running the linker in PIE mode.
--- arch/x86/kernel/head_64.S | 14 +++++++++----- arch/x86/kernel/relocate_kernel_64.S | 6 ++++-- 2 files changed, 13 insertions(+), 7 deletions(-)
arch/x86/kernel/head_64.S +9 -5
@@ -182,8 +182,9 @@ SYM_INNER_LABEL(secondary_startup_64_no_verify, SYM_L_GLOBAL) xorl %r15d, %r15d /* Derive the runtime physical address of init_top_pgt[] */ - movq phys_base(%rip), %rax - addq $(init_top_pgt - __START_KERNEL_map), %rax + leaq init_top_pgt(%rip), %rax + subq $__START_KERNEL_map, %rax + addq phys_base(%rip), %rax /* * Retrieve the modifier (SME encryption mask if SME is active) to be
@@ -314,7 +315,8 @@ SYM_INNER_LABEL(common_startup_64, SYM_L_LOCAL) .Lsetup_cpu: /* Get the per cpu offset for the given CPU# which is in ECX */ - movq __per_cpu_offset(,%rcx,8), %rdx + leaq __per_cpu_offset(%rip), %rdx + movq (%rdx,%rcx,8), %rdx #else xorl %edx, %edx /* zero-extended to clear all of RDX */ #endif /* CONFIG_SMP */
@@ -325,7 +327,8 @@ SYM_INNER_LABEL(common_startup_64, SYM_L_LOCAL) * * RDX contains the per-cpu offset */ - movq pcpu_hot + X86_current_task(%rdx), %rax + leaq pcpu_hot + X86_current_task(%rip), %rax + movq (%rax,%rdx), %rax movq TASK_threadsp(%rax), %rsp /*
@@ -346,7 +349,8 @@ SYM_INNER_LABEL(common_startup_64, SYM_L_LOCAL) */ subq $16, %rsp movw $(GDT_SIZE-1), (%rsp) - leaq gdt_page(%rdx), %rax + leaq gdt_page(%rip), %rax + addq %rdx, %rax movq %rax, 2(%rsp) lgdt (%rsp) addq $16, %rsp
arch/x86/kernel/relocate_kernel_64.S +4 -2
@@ -106,6 +106,9 @@ SYM_CODE_START_NOALIGN(relocate_kernel) /* setup a new stack at the end of the physical control page */ lea PAGE_SIZE(%r8), %rsp + /* take the virtual address of virtual_mapped() before jumping */ + leaq virtual_mapped(%rip), %r14 + /* jump to identity mapped page */ addq $(identity_mapped - relocate_kernel), %r8 pushq %r8
@@ -225,8 +228,7 @@ SYM_CODE_START_LOCAL_NOALIGN(identity_mapped) movq %rax, %cr3 lea PAGE_SIZE(%r8), %rsp call swap_pages - movq $virtual_mapped, %rax - pushq %rax + pushq %r14 ANNOTATE_UNRET_SAFE ret int3
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 22/28 asm-generic: Treat PIC .data.rel.ro sections as .rodata
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
When running the compiler in PIC/PIE mode, it will emit data objects that are 'const' in the context of the program into the .data.rel.ro section if they contain absolute addresses of statically allocated global objects. This helps the dynamic loader distinguish between objects that are truly const from objects that will need to be fixed up by the loader before starting the program.
This is not a concern for the kernel, but it does mean those .data.rel.ro input sections need to be handled. So treat them as .rodata.
It also means some explicit uses of .rodata for global structures containing absolute addresses need to be changed to .data.rel.ro to prevent the linker from warning about incompatible section flags.
--- include/asm-generic/vmlinux.lds.h | 2 +- include/linux/compiler.h | 2 +- scripts/kallsyms.c | 2 +- tools/objtool/check.c | 11 ++++++----- tools/objtool/include/objtool/special.h | 2 +- 5 files changed, 10 insertions(+), 9 deletions(-)
include/asm-generic/vmlinux.lds.h +1 -1
@@ -456,7 +456,7 @@ . = ALIGN((align)); \ .rodata : AT(ADDR(.rodata) - LOAD_OFFSET) { \ __start_rodata = .; \ - *(.rodata) *(.rodata.*) \ + *(.rodata .rodata.* .data.rel.ro*) \ SCHED_DATA \RO_AFTER_INIT_DATA /* Read only after init */ \ . = ALIGN(8); \
include/linux/compiler.h +1 -1
@@ -133,7 +133,7 @@ void ftrace_likely_update(struct ftrace_likely_data *f, int val,#define annotate_unreachable () __annotate_unreachable(__COUNTER__)/* Annotate a C jump table to allow objtool to follow the code flow */ -#define __annotate_jump_table __section(".rodata..c_jump_table") +#define __annotate_jump_table __section(".data.rel.ro.c_jump_table")#else /* !CONFIG_OBJTOOL */ #define annotate_reachable ()
scripts/kallsyms.c +1 -1
@@ -357,7 +357,7 @@ static void write_src(void)printf ("#define ALGN .balign 4\n" );printf ("#endif\n" ); -printf ("\t.section .rodata, \"a\"\n" ); +printf ("\t.section .data.rel.ro, \"a\"\n" );output_label ("kallsyms_num_syms" );printf ("\t.long\t%u\n" , table_cnt);
tools/objtool/check.c +6 -5
@@ -2575,15 +2575,16 @@ static void mark_rodata(struct objtool_file *file)* Search for the following rodata sections, each of which can * potentially contain jump tables: * -* - .rodata: can contain GCC switch tables -* - .rodata.<func>: same, if -fdata-sections is being used -* - .rodata..c_jump_table: contains C annotated jump tables +* - .rodata .data.rel.ro : can contain GCC switch tables +* - .rodata.<func> .data.rel.ro.<func> : same, if -fdata-sections is being used +* - .data.rel.ro.c_jump_table : contains C annotated jump tables * * .rodata.str1.* sections are ignored; they don't contain jump tables. */ for_each_sec (file, sec) { -if (!strncmp (sec->name,".rodata" ,7 ) && - !strstr (sec->name,".str1." )) { +if ((!strncmp (sec->name,".rodata" ,7 ) && + !strstr (sec->name,".str1." )) || + !strncmp (sec->name,".data.rel.ro" ,12 )) { sec->rodata = true; found = true; }
tools/objtool/include/objtool/special.h +1 -1
@@ -10,7 +10,7 @@#include <objtool/check.h> #include <objtool/elf.h> -#define C_JUMP_TABLE_SECTION ".rodata..c_jump_table" +#define C_JUMP_TABLE_SECTION ".data.rel.ro.c_jump_table"struct special_alt {struct list_head list;
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 23/28 tools/objtool: Mark generated sections as writable
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
objtool generates ELF sections such as __mcount_loc, which carry absolute symbol references that need to be fixed up at boot time, based on the actual virtual placement of the kernel binary.
This involves writing to the section at boot time, and in some cases (e.g., when using --pie and -z text), the lld linker is more pedantic about this, and complains about absolute relocations operating on read-only sections.
None of this actually matters for vmlinux, which manages its own mapping permissions, and so we can just set the SHF_WRITE flag on those sections to make the linker happy.
--- tools/objtool/elf.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
tools/objtool/elf.c +1 -1
@@ -1142,7 +1142,7 @@ struct section *elf_create_section(struct elf *elf, const char *name, sec->sh.sh_entsize = entsize; sec->sh.sh_type =SHT_PROGBITS ; sec->sh.sh_addralign =1 ; - sec->sh.sh_flags =SHF_ALLOC ; + sec->sh.sh_flags =SHF_ALLOC |SHF_WRITE ;/* Add section name to .shstrtab (or .strtab for Clang) */ shstrtab =find_section_by_name (elf,".shstrtab" );
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 24/28 tools/objtool: Treat indirect ftrace calls as direct calls
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
In some cases, the compiler may rely on indirect calls using GOT slots as memory operands to emit function calls. This leaves it up to the linker to relax the call to a direct call if possible, i.e., if the destination address is known at link time and in range, which may not be the case when building shared libraries for user space.
On x86, this may happen when building in PIC mode with ftrace enabled, and given that vmlinux is a fully linked binary, this relaxation is always possible, and therefore mandatory per the x86_64 psABI.
This means that the indirect calls to __fentry__ that are observeable in vmlinux.o will have been converted to direct calls in vmlinux, and can be treated as such by objtool.
--- tools/objtool/check.c | 32 ++++++++++++++++++-- 1 file changed, 30 insertions(+), 2 deletions(-)
tools/objtool/check.c +30 -2
@@ -1696,11 +1696,39 @@ static int add_call_destinations(struct objtool_file *file)struct reloc *reloc;for_each_insn (file, insn) { -if (insn->type !=INSN_CALL ) +if (insn->type !=INSN_CALL && + insn->type !=INSN_CALL_DYNAMIC )continue ; reloc =insn_reloc (file, insn); -if (!reloc) { +if (insn->type ==INSN_CALL_DYNAMIC ) { +if (!reloc) +continue ; + +/* +* GCC 13 and older on x86 will always emit the call to +* __fentry__ using a relaxable GOT-based symbol +* reference when operating in PIC mode, i.e., +* +* call *0x0(%rip) +* R_X86_64_GOTPCRELX __fentry__-0x4 +* +* where it is left up to the linker to relax this into +* +* call __fentry__ +* nop +* +* if __fentry__ turns out to be DSO local, which is +* always the case for vmlinux. Given that this +* relaxation is mandatory per the x86_64 psABI, these +* calls can simply be treated as direct calls. +*/ +if (arch_ftrace_match (reloc->sym->name)) { + insn->type =INSN_CALL ; +add_call_dest (file, insn, reloc->sym, false); + } + + }else if (!reloc) { dest_off =arch_jump_destination (insn); dest =find_call_destination (insn->sec, dest_off);
Signature · 1 line
2.46.0.792.g87dc391469-goog
Josh Poimboeuf
maintainerRe: [RFC PATCH 24/28] tools/objtool: Treat indirect ftrace calls as direct calls
Quoted 16 lines · Ard Biesheuvel
tools/objtool/check.c +16
-0+if (insn->type ==INSN_CALL_DYNAMIC ) { +if (!reloc) +continue ; + +/* +* GCC 13 and older on x86 will always emit the call to +* __fentry__ using a relaxable GOT-based symbol +* reference when operating in PIC mode, i.e., +* +* call *0x0(%rip) +* R_X86_64_GOTPCRELX __fentry__-0x4 +* +* where it is left up to the linker to relax this into +* +* call __fentry__ +* nop
tools/objtool/check.c static int add_call_destinations(struct objtool_file *file) · line 1729
Quoted 10 lines
tools/objtool/check.c +10
-0+* +* if __fentry__ turns out to be DSO local, which is +* always the case for vmlinux. Given that this +* relaxation is mandatory per the x86_64 psABI, these +* calls can simply be treated as direct calls. +*/ +if (arch_ftrace_match (reloc->sym->name)) { + insn->type =INSN_CALL ; +add_call_dest (file, insn, reloc->sym, false); + }
Can the compiler also do this for non-fentry direct calls? If so would it make sense to generalize this by converting all INSN_CALL_DYNAMIC+reloc to INSN_CALL?
And maybe something similar for add_jump_destinations().
Signature · 1 line
Josh
Ard Biesheuvel
maintainerRe: [RFC PATCH 24/28] tools/objtool: Treat indirect ftrace calls as direct calls
Quoted 29 lines · Josh Poimboeuf
Quoted 16 lines · Ard Biesheuvel
tools/objtool/check.c +16
-0+if (insn->type ==INSN_CALL_DYNAMIC ) { +if (!reloc) +continue ; + +/* +* GCC 13 and older on x86 will always emit the call to +* __fentry__ using a relaxable GOT-based symbol +* reference when operating in PIC mode, i.e., +* +* call *0x0(%rip) +* R_X86_64_GOTPCRELX __fentry__-0x4 +* +* where it is left up to the linker to relax this into +* +* call __fentry__ +* nop tools/objtool/check.c static int add_call_destinations(struct objtool_file *file) · line 1729
Quoted 10 lines
tools/objtool/check.c +10
-0+* +* if __fentry__ turns out to be DSO local, which is +* always the case for vmlinux. Given that this +* relaxation is mandatory per the x86_64 psABI, these +* calls can simply be treated as direct calls. +*/ +if (arch_ftrace_match (reloc->sym->name)) { + insn->type =INSN_CALL ; +add_call_dest (file, insn, reloc->sym, false); + }
Quoted 1 line
Can the compiler also do this for non-fentry direct calls?
No, it is essentially an oversight in GCC that this happens at all, and I fixed it [0] for GCC 14, i.e., to honour -mdirect-extern-access when emitting these calls.
But even without that, it is peculiar at the very least that the compiler would emit GOT based indirect calls at all.
Instead of
call *__fentry__@GOTPCREL(%rip)
it should simply emit
call __fentry__@PLT
and leave it up to the linker to resolve this directly or lazily/eagerly via a PLT jump (assuming -fno-plt is not being used)
Quoted 6 lines
If so would
it make sense to generalize this by converting all INSN_CALL_DYNAMIC+reloc toINSN_CALL?And maybe something similar for
add_jump_destinations().
I suppose that the pattern INSN_CALL_DYNAMIC+reloc is unambiguous, and can therefore always be treated as INSN_CALL. But I don't anticipate any other occurrences here, and if they do exist, they indicate some other weirdness in the compiler, so perhaps it is better not to add general support for these.
[0] https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=bde21de1205c0456f6df68c950fb7ee631fcfa93
RFC PATCH 25/28 x86: Use PIE codegen for the core kernel
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
As an intermediate step towards enabling PIE linking for the 64-bit x86 kernel, enable PIE codegen for all objects that are linked into the kernel proper.
This substantially reduces the number of relocations that need to be processed when booting a relocatable KASLR kernel.
Before (size in bytes of the reloc table):
797372 arch/x86/boot/compressed/vmlinux.relocs
After:
400252 arch/x86/boot/compressed/vmlinux.relocs
--- arch/x86/Makefile | 11 ++++++++++- arch/x86/boot/Makefile | 1 + arch/x86/boot/compressed/Makefile | 2 +- arch/x86/entry/vdso/Makefile | 1 + arch/x86/realmode/rm/Makefile | 1 + include/asm-generic/vmlinux.lds.h | 1 + 6 files changed, 15 insertions(+), 2 deletions(-)
arch/x86/Makefile +10 -1
@@ -193,13 +193,22 @@ else KBUILD_RUSTFLAGS += -Cno-redzone=y KBUILD_RUSTFLAGS += -Ccode-model=kernel + PIE_CFLAGS-y := -fpie -mcmodel=small \ + -include $(srctree)/include/linux/hidden.h + + PIE_CFLAGS-$(CONFIG_CC_IS_GCC) += $(call cc-option.-mdirect-extern-access) + PIE_CFLAGS-$(CONFIG_CC_IS_CLANG) += -fdirect-access-external-data + ifeq ($(CONFIG_STACKPROTECTOR),y) KBUILD_CFLAGS += -mstack-protector-guard-symbol=fixed_percpu_data + + # the 'small' C model defaults to %fs + PIE_CFLAGS-$(CONFIG_SMP) += -mstack-protector-guard-reg=gs endif # Don't emit relaxable GOTPCREL relocations KBUILD_AFLAGS_KERNEL += -Wa,-mrelax-relocations=no - KBUILD_CFLAGS_KERNEL += -Wa,-mrelax-relocations=no + KBUILD_CFLAGS_KERNEL += -Wa,-mrelax-relocations=no $(PIE_CFLAGS-y) endif #
arch/x86/boot/Makefile +1 -0
@@ -57,6 +57,7 @@ KBUILD_AFLAGS := $(KBUILD_CFLAGS) -D__ASSEMBLY__ KBUILD_CFLAGS += $(call cc-option,-fmacro-prefix-map=$(srctree)/=) KBUILD_CFLAGS += -fno-asynchronous-unwind-tables KBUILD_CFLAGS += $(CONFIG_CC_IMPLICIT_FALLTHROUGH) +KBUILD_CFLAGS_KERNEL := $(obj)/bzImage: asflags-y := $(SVGA_MODE)
arch/x86/boot/compressed/Makefile +1 -1
@@ -73,7 +73,7 @@ LDFLAGS_vmlinux += -T hostprogs := mkpiggy HOST_EXTRACFLAGS += -I$(srctree)/tools/include -sed-voffset := -e 's/^\([0-9a-fA-F]*\) [ABCDGRSTVW] \(_text\|__start_rodata\|__bss_start\|_end\)$$/\#define VO_\2 _AC(0x\1,UL)/p' +sed-voffset := -e 's/^\([0-9a-fA-F]*\) [ABbCDdGRSTtVW] \(_text\|__start_rodata\|__bss_start\|_end\)$$/\#define VO_\2 _AC(0x\1,UL)/p' quiet_cmd_voffset = VOFFSET $@ cmd_voffset = $(NM) $< | sed -n $(sed-voffset) > $@
arch/x86/entry/vdso/Makefile +1 -0
@@ -141,6 +141,7 @@ endif endif $(obj)/vdso32.so.dbg: KBUILD_CFLAGS = $(KBUILD_CFLAGS_32) +$(obj)/vdso32.so.dbg: KBUILD_CFLAGS_KERNEL := $(obj)/vdso32.so.dbg: $(obj)/vdso32/vdso32.lds $(vobjs32) FORCE $(call if_changed,vdso_and_check)
arch/x86/realmode/rm/Makefile +1 -0
@@ -67,3 +67,4 @@ KBUILD_CFLAGS := $(REALMODE_CFLAGS) -D_SETUP -D_WAKEUP \ -I$(srctree)/arch/x86/boot KBUILD_AFLAGS := $(KBUILD_CFLAGS) -D__ASSEMBLY__ KBUILD_CFLAGS += -fno-asynchronous-unwind-tables +KBUILD_CFLAGS_KERNEL :=
include/asm-generic/vmlinux.lds.h +1 -0
@@ -349,6 +349,7 @@ *(DATA_MAIN) \ *(.data..decrypted) \ *(.ref.data) \ + *(.data.rel*) \ *(.data..shared_aligned) /* percpu related */ \ *(.data.unlikely) \ __start_once = .; \
Signature · 1 line
2.46.0.792.g87dc391469-goog
Re: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 10 lines · Ard Biesheuvel
This substantially reduces the number of relocations that need to be processed when booting a relocatable KASLR kernel.
Before (size in bytes of the reloc table):
797372
arch/x86/boot/compressed/vmlinux.relocsAfter:
400252
arch/x86/boot/compressed/vmlinux.relocs
I don't know why anybody would care about the size of the relocation table?
What matters is what it does to general performance.
Traditionally even on x86-64 PIC/E has a cost and the kernel model was intended to avoid that.
From my perspective this patch kit doesn't fix a real problem, it's all risk of performance regression with no gain.
-Andi
Ard Biesheuvel
maintainerRe: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 7 lines · Andi Kleen
Quoted 5 lines · Ard Biesheuvel
This substantially reduces the number of relocations that need to be processed when booting a relocatable KASLR kernel.
Before (size in bytes of the reloc table):
Quoted 8 lines
Quoted 5 lines
797372
arch/x86/boot/compressed/vmlinux.relocsAfter:
400252
arch/x86/boot/compressed/vmlinux.relocsI don't know why anybody would care about the size of the relocation table?
Fair point.
Quoted 5 lines
What matters is what it does to general performance.
Traditionally even on x86-64 PIC/E has a cost and the kernel model was intended to avoid that.
Is the x86_64 kernel C model specified anywhere, to your knowledge?
Quoted 3 lines
From my perspective this patch kit doesn't fix a real problem, it's all risk of performance regression with no gain.
It's all in the cover letter and the commit logs so I won't rehash it here, but I understand that your priorities may be different from mine.
I'll provide some numbers about the impact on code size. Are there any other performance related aspects that you think might be impacted by the use of position independent code generation?
Re: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 7 lines · Ard Biesheuvel
Quoted 5 lines
What matters is what it does to general performance.
Traditionally even on x86-64 PIC/E has a cost and the kernel model was intended to avoid that.
Is the
x86_64kernel C model specified anywhere, to your knowledge?
The basics are in the ABI. Maybe some of the details of TLS / stack protector are missing (I guess that could be fixed, adding HJ)
Some of the motivation was also in early papers like
https://www.ucw.cz/~hubicka/papers/amd64/amd64.html
I'm copying Honza Hubicka who did the original work.
Quoted 8 lines
Quoted 4 lines
From my perspective this patch kit doesn't fix a real problem, it's all risk of performance regression with no gain.
It's all in the cover letter and the commit logs so I won't rehash it here, but I understand that your priorities may be different from mine.
It sounded fairly nebulous to me. If Linux wanted to support a third tool chain and it didn't support the kernel model yet it would be somehow easier. Apart from the kernel model likely being one of the minor issues in such a endeavour, I don't see a third tool chain othan than gcc and llvm anywhere on the horizon?
Quoted 4 lines
I'll provide some numbers about the impact on code size. Are there any other performance related aspects that you think might be impacted by the use of position independent code generation?
Code size isn't a sufficient metric either.
Linux sometimes goes to great length for small gains, for example there was a huge effort to avoid frame pointers, even though it's a small percentage delta. PIC could well be larger than frame pointers.
You need to run it with some real workloads, e.g. some of the kernel oriented workloads in 0day or phoronix, and see if there are performance regressions.
Unfortunately for an intrusive change like this this might also vary for different CPUs, so may need some more coverage.
-Andi
Ard Biesheuvel
maintainerRe: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 11 lines · Andi Kleen
Quoted 7 lines · Ard Biesheuvel
Quoted 5 lines
What matters is what it does to general performance.
Traditionally even on x86-64 PIC/E has a cost and the kernel model was intended to avoid that.
Is the
x86_64kernel C model specified anywhere, to your knowledge?
Quoted 8 lines
The basics are in the ABI. Maybe some of the details of TLS / stack protector are missing (I guess that could be fixed, adding HJ)
Some of the motivation was also in early papers like
https://www.ucw.cz/~hubicka/papers/amd64/amd64.htmlI'm copying Honza Hubicka who did the original work.
Thanks.
So the psABI has
Kernel code model:
The kernel of an operating system is usually rather small but runs in the negative half of the address space.
plus an explanation on the ranges of symbolic references.
The problem here is that it is inherently a position dependent code model, where 'the virtual address of code executed is known at link time' as per the psABI.
We currently violate that assumption in two different ways:
- the decompressor applies KASLR by using the static ELF relocations (which are intended for consumption by the linker not the loader) and use them to relocate the executable to its randomized virtual address; those static relocations can go out of sync with the actual code when relaxations are applied;
- some of the startup code is now written in C, but is called via the 1:1 mapping; absolute symbol references don't work in that context, and we rely on faith and a whole pile of hacks to ensure that this does not break.
Quoted 9 lines
Quoted 8 lines
Quoted 4 lines
From my perspective this patch kit doesn't fix a real problem, it's all risk of performance regression with no gain.
It's all in the cover letter and the commit logs so I won't rehash it here, but I understand that your priorities may be different from mine.
Quoted 6 lines
It sounded fairly nebulous to me. If Linux wanted to support a third tool chain and it didn't support the kernel model yet it would be somehow easier. Apart from the kernel model likely being one of the minor issues in such a endeavour, I don't see a third tool chain othan than gcc and llvm anywhere on the horizon?
I was referring to Rust, which is llvm based but will also have a GCC based alternative (gcc-rs) in the future. I am aware that these will most likely reuse most of the existing backends, where these concerns have the most impact, but they still need wider consideration than they used to in the past.
So the tl;dr for the rationale behind this series is that it is better to use a code model and a relocation model that
a) matches the reality of how our code operates, and
b) deviates as little as possible from how code is generally constructed for user space.
On top of that, it would be better to use relocation metadata that is intended for consumption by a runtime loader, rather than rolling our own based on a relocation format that is intended for consumption by the build time linker.
Quoted 11 lines
Quoted 4 lines
I'll provide some numbers about the impact on code size. Are there any other performance related aspects that you think might be impacted by the use of position independent code generation?
Code size isn't a sufficient metric either.
Linux sometimes goes to great length for small gains, for example there was a huge effort to avoid frame pointers, even though it's a small percentage delta. PIC could well be larger than frame pointers.
Quoted 7 lines
You need to run it with some real workloads, e.g. some of the kernel oriented workloads in 0day or phoronix, and see if there are performance regressions.
Unfortunately for an intrusive change like this this might also vary for different CPUs, so may need some more coverage.
I'll spend some time to look into this in more detail (and get some help internally at Google to measure the real-world impact)
H. Peter Anvin
reviewerRe: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
In reply to Ard BiesheuvelQuoted 9 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
As an intermediate step towards enabling PIE linking for the 64-bit x86 kernel, enable PIE codegen for all objects that are linked into the kernel proper.
This substantially reduces the number of relocations that need to be processed when booting a relocatable KASLR kernel.
This really seems like going completely backwards to me.
You are imposing a more restrictive code model on the kernel, optimizing for boot time in a way that will exert a permanent cost on the running kernel.
There is a *huge* difference between the kernel and user space here:
KERNEL MEMORY IS PERMANENTLY ALLOCATED, AND IS NEVER SHARED.
Dirtying user pages requires them to be unshared and dirty, which is undesirable. Kernel pages are *always* unshared and dirty.
Quoted 4 lines
It also brings us much closer to the ordinary PIE relocation model used for most of user space, which is therefore much better supported and less likely to create problems as we increase the range of compilers and linkers that need to be supported.
We have been resisting *for ages* making the kernel worse to accomodate broken compilers. We don't "need" to support more compilers -- we need the compilers to support us. We have working compilers; any new compiler that wants to play should be expected to work correctly.
-hpa
Ard Biesheuvel
maintainerRe: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Hi Peter,
Thanks for taking a look.
Quoted 12 lines · H. Peter Anvin
Quoted 9 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
As an intermediate step towards enabling PIE linking for the 64-bit x86 kernel, enable PIE codegen for all objects that are linked into the kernel proper.
This substantially reduces the number of relocations that need to be processed when booting a relocatable KASLR kernel.
Quoted 6 lines
This really seems like going completely backwards to me.
You are imposing a more restrictive code model on the kernel, optimizing for boot time in a way that will exert a permanent cost on the running kernel.
Fair point about the boot time. This is not the only concern, though, and arguably the least important one.
As I responded to Andi before, it is also about using a code model and relocation model that matches the reality of how the code is executed:
- the early C code runs from the 1:1 mapping, and needs special hacks to accommodate this
- KASLR runs the kernel from a different virtual address than the one we told the linker about
Quoted 4 lines
There is a *huge* difference between the kernel and user space here:
KERNEL MEMORY IS PERMANENTLY ALLOCATED, AND IS NEVER SHARED.
No need to shout.
Quoted 3 lines
Dirtying user pages requires them to be unshared and dirty, which is undesirable. Kernel pages are *always* unshared and dirty.
I guess you are referring to the use of a GOT? That is a valid concern, but it does not apply here. With hidden visibility and compiler command line options like -mdirect-access-extern, all emitted symbol references are direct. Disallowing text relocations could be trivially enabled with this series if desired, and actually helps avoid the tricky bugs we keep fixing in the early startup code that executes from the 1:1 mapping (the C code in .head.text)
So it mostly comes down to minor differences in addressing modes, e.g.,
movq $sym, %reg
actually uses more bytes than
leaq sym(%rip), %reg
whereas
movq sym, %reg
and
movq sym(%rip), %reg
are the same length.
OTOH, indexing a statically allocated global array like
movl array(,%reg1,4), %reg2
will be converted into
leaq array(%rip), %reg2
movl (%reg2,%reg1,4), %reg2
and is therefore less efficient in terms of code footprint. But in general, the x86_64 ISA and psABI are quite flexible in this regard, and extrapolating from past experiences with PIC code on i386 is not really justified here.
As Andi also pointed out, what ultimately matters is the performance, as well as code size where it impacts performance, through the I-cache footprint. I'll do some testing before reposting, and maybe not bother if the impact is negative.
Quoted 10 lines
Quoted 4 lines
It also brings us much closer to the ordinary PIE relocation model used for most of user space, which is therefore much better supported and less likely to create problems as we increase the range of compilers and linkers that need to be supported.
We have been resisting *for ages* making the kernel worse to accomodate broken compilers. We don't "need" to support more compilers -- we need the compilers to support us. We have working compilers; any new compiler that wants to play should be expected to work correctly.
We are in a much better place now than we were before in that regard, which is actually how this effort came about: instead of lying to the compiler, and maintaining our own pile of scripts and relocation tools, we can just do what other arches are doing in Linux, and let the toolchain do it for us.
Re: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 5 lines · Ard Biesheuvel
I guess you are referring to the use of a GOT? That is a valid concern, but it does not apply here. With hidden visibility and compiler command line options like -mdirect-access-extern, all emitted symbol references are direct.
I absolutely hate GOT entries. We definitely shouldn't ever do anything that causes them on x86-64.
I'd much rather just do boot-time relocation, and I don't think the "we run code at a different location than we told the linker" is an arghument against it.
Please, let's make sure we never have any of the global offset table horror.
Yes, yes, you can't avoid them on other architectures.
That said, doing changes like changing "mov $sym" to "lea sym(%rip)" I feel are a complete no-brainer and should be done regardless of any other code generation issues.
Let's not do relocation for no good reason.
Linus
Ard Biesheuvel
maintainerRe: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 11 lines · Linus Torvalds
Quoted 5 lines · Ard Biesheuvel
I guess you are referring to the use of a GOT? That is a valid concern, but it does not apply here. With hidden visibility and compiler command line options like -mdirect-access-extern, all emitted symbol references are direct.
I absolutely hate GOT entries. We definitely shouldn't ever do anything that causes them on x86-64.
Quoted 8 lines
I'd much rather just do boot-time relocation, and I don't think the "we run code at a different location than we told the linker" is an arghument against it.
Please, let's make sure we never have any of the global offset table horror.
Yes, yes, you can't avoid them on other architectures.
GCC/binutils never needs them. GCC 13 and older will emit some horrible indirect fentry hook calls via the GOT, but those are NOPed out by objtool anyway, so that is easily fixable.
Clang/LLD is slightly trickier, because Clang emits relaxable GOTPCREL relocations, but LLD doesn't update the relocations emitted via --emit-relocs, so the relocs tool gets confused. This is one of the reasons I had for proposing to simply switch to PIE linking, and let the linker deal with all of that. Note that Clang may emit these even when not generating PIC/PIE code at all.
So this is the reason I wanted to add support for GOTPCREL relocations in the relocs tool - it is really quite trivial to do, and makes our jobs slightly easier when dealing with these compiler quirks. The alternative would be to teach objtool how to relax 'movq foo@GOTPCREL(%rip)' into 'leaq foo(%rip)' - these are GOTPCREL relaxations described in the x86_64 psABI for ELF, which is why compilers are assuming more and more that emitting these is fine even without -fpic, given that the linker is supposed to elide them if possible.
Note that there are 1 or 2 cases in the asm code where it is actually quite useful to refer to the address of foo as 'foo@GOTPCREL(%rip)' in instructions that take memory operands, but those individual cases are easily converted to something else if even having a GOT with just 2 entries is a dealbreaker for you.
Quoted 4 lines
That said, doing changes like changing "mov $sym" to "lea sym(%rip)" I feel are a complete no-brainer and should be done regardless of any other code generation issues.
Yes, this is the primary reason I ended up looking into this in the first place. Earlier this year, we ended up having to introduce RIP_REL_REF() to emit those RIP-relative references explicitly, in order to prevent the C code that is called via the early 1:1 mapping from exploding. The amount of C code called in that manner has been growing steadily over time with the introduction of 5-level paging and SEV-SNP and TDX support, which need to play all kinds of tricks before the normal kernel mappings are created.
Compiling with -fpie and linking with --pie -z text produces an executable that is guaranteed to have only RIP-relative references in the .text segment, removing the need for RIP_REL_REF entirely (it already does nothing when __pic__ is #define'd).
H. Peter Anvin
reviewerRe: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 5 lines · Ard Biesheuvel
Quoted 4 lines
That said, doing changes like changing "mov $sym" to "lea sym(%rip)" I feel are a complete no-brainer and should be done regardless of any other code generation issues.
Quoted 9 lines
Yes, this is the primary reason I ended up looking into this in the first place. Earlier this year, we ended up having to introduce
RIP_REL_REF()to emit those RIP-relative references explicitly, in order to prevent the C code that is called via the early 1:1 mapping from exploding. The amount of C code called in that manner has been growing steadily over time with the introduction of 5-level paging and SEV-SNP and TDX support, which need to play all kinds of tricks before the normal kernel mappings are created.
movq $sym to leaq sym(%rip) which you said ought to be smaller (and in reality appears to be the same size, 7 bytes) seems like a no-brainer and can be treated as a code quality issue -- in other words, file bug reports against gcc and clang.
Quoted 4 lines
Compiling with -fpie and linking with --pie -z text produces an executable that is guaranteed to have only RIP-relative references in the .text segment, removing the need for
RIP_REL_REFentirely (it already does nothing when__pic__is #define'd).
But -fpie has a considerable cost; specifically when we have indexed references, as in that case the base pointer needs to be manifest in a register, *and* it takes up a register slot in the EA, which may end converting one instruction into three.
Now, the "kernel" memory model is defined in the ABI document, but there is nothing that prevents us from making updates to it if we need to; e.g. the statement that movq $sym can be used is undesirable, of course.
-hpa
Re: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 17 lines · H. Peter Anvin
Quoted 5 lines · Ard Biesheuvel
Quoted 4 lines
That said, doing changes like changing "mov $sym" to "lea sym(%rip)" I feel are a complete no-brainer and should be done regardless of any other code generation issues.
Quoted 9 lines
Yes, this is the primary reason I ended up looking into this in the first place. Earlier this year, we ended up having to introduce
RIP_REL_REF()to emit those RIP-relative references explicitly, in order to prevent the C code that is called via the early 1:1 mapping from exploding. The amount of C code called in that manner has been growing steadily over time with the introduction of 5-level paging and SEV-SNP and TDX support, which need to play all kinds of tricks before the normal kernel mappings are created.
Quoted 4 lines
movq $sym to leaq sym(%rip) which you said ought to be smaller (and in reality appears to be the same size, 7 bytes) seems like a no-brainer and can be treated as a code quality issue -- in other words, file bug reports against gcc and clang.
It is the kernel assembly source that should be converted to rip-relative form, gcc (and probably clang) have nothing with it.
Uros.
H. Peter Anvin
reviewerRe: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 9 lines · Uros Bizjak
Quoted 5 lines
movq $sym to leaq sym(%rip) which you said ought to be smaller (and in reality appears to be the same size, 7 bytes) seems like a no-brainer and can be treated as a code quality issue -- in other words, file bug reports against gcc and clang.
It is the kernel assembly source that should be converted to rip-relative form, gcc (and probably clang) have nothing with it.
Sadly, that is not correct; neither gcc nor clang uses lea:
-hpa
gcc version 14.2.1 20240912 (Red Hat 14.2.1-3) (GCC)
hpa@tazenda:/tmp$ cat foo.c int foobar;
int *where_is_foobar(void)
{
return &foobar;
}hpa@tazenda:/tmp$ gcc -mcmodel=kernel -O2 -c -o foo.o foo.c hpa@tazenda:/tmp$ objdump -dr foo.o
foo.o: file format elf64-x86-64
Disassembly of section .text:
0000000000000000 <where_is_foobar>:
0: 48 c7 c0 00 00 00 00 mov $0x0,%rax
3: R_X86_64_32S foobar
7: c3 ret
clang version 18.1.8 (Fedora 18.1.8-1.fc40)
hpa@tazenda:/tmp$ clang -mcmodel=kernel -O2 -c -o foo.o foo.c hpa@tazenda:/tmp$ objdump -dr foo.o
foo.o: file format elf64-x86-64
Disassembly of section .text:
0000000000000000 <where_is_foobar>:
0: 48 c7 c0 00 00 00 00 mov $0x0,%rax
3: R_X86_64_32S foobar
7: c3 ret
Re: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 2 lines · H. Peter Anvin
Sadly, that is not correct; neither gcc nor clang uses lea:
Looking around, this may be intentional. At least according to Agner, several cores do better at "mov immediate" compared to "lea".
Eg a RIP-relative LEA on Zen 2 gets a throughput of two per cycle, but a "MOV r,i" gets four. That got fixed in Zen 3 and later, but apparently Intel had similar issues (Ivy Bridge: 1 LEA per cycle, vs 3 "mov i,r". Haswell is 1:4).
Of course, Agner's tables are good, but not necessarily always the whole story. There are other instruction tables on the internet (eg uops.info) with possibly more info.
And in reality, I would expect it to be a complete non-issue with any OoO engine and real code, because you are very seldom ALU limited particularly when there aren't any data dependencies.
But a RIP-relative LEA does seem to put a *bit* more pressure on the core resources, so the compilers are may be right to pick a "mov".
Linus
Re: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 5 lines · Linus Torvalds
Quoted 2 lines · H. Peter Anvin
Sadly, that is not correct; neither gcc nor clang uses lea:
Quoted 7 lines
Looking around, this may be intentional. At least according to Agner, several cores do better at "mov immediate" compared to "lea".
Eg a RIP-relative LEA on Zen 2 gets a throughput of two per cycle, but a "MOV r,i" gets four. That got fixed in Zen 3 and later, but apparently Intel had similar issues (Ivy Bridge: 1 LEA per cycle, vs 3 "mov i,r". Haswell is 1:4).
Yes, this is the case. I just missed your reply when replying to Peter's mail with a not so precise answer.
Uros.
Re: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
In reply to H. Peter AnvinQuoted 22 lines · H. Peter Anvin
Quoted 9 lines · Uros Bizjak
Quoted 5 lines
movq $sym to leaq sym(%rip) which you said ought to be smaller (and in reality appears to be the same size, 7 bytes) seems like a no-brainer and can be treated as a code quality issue -- in other words, file bug reports against gcc and clang.
It is the kernel assembly source that should be converted to rip-relative form, gcc (and probably clang) have nothing with it.
Sadly, that is not correct; neither gcc nor clang uses lea:
-hpa
gcc version 14.2.1 20240912 (Red Hat 14.2.1-3) (GCC)
hpa@tazenda:/tmp$ cat foo.c int foobar;
Quoted 6 lines
int
*where_is_foobar(void){ return &foobar; }hpa@tazenda:/tmp$ gcc -mcmodel=kernel -O2 -c -o foo.o foo.c
Indeed, but my reply was in the context of -fpie, which guarantees RIP relative access. IOW, the compiler will always generate sym(%rip) with -fpie, but (obviously) can't change assembly code in the kernel when the PIE is requested.
Otherwise, MOV $immediate, %reg is faster when PIE is not required, which is the case with -mcmodel=kernel. IIRC, LEA with %rip had some performance issues, which may not be the case anymore with newer processors.
Due to the non-negligible impact of PIE, perhaps some kind of CONFIG_PIE config definition should be introduced, so the assembly code would be able to choose optimal asm sequence when PIE and non-PIE is requested?
Uros.
RE: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
...
Quoted 4 lines
Due to the non-negligible impact of PIE, perhaps some kind of
CONFIG_PIEconfig definition should be introduced, so the assembly code would be able to choose optimal asm sequence when PIE and non-PIE is requested?
I wouldn't have thought that performance mattered in the asm code that runs during startup?
While x86-84 code (ignoring data references) is pretty much always position independent, the same isn't true of all architectures. Some (at least Nios-II) only have absolute call instructions. So you can't really move to pic code globally.
You'd also want 'bad' pic code that contained some fixups that needed the code patching.
(Which you really don't want for a shared library.) Otherwise you get an extra instruction for non-trivial data accesses.
Thinking....
Doesn't the code generated for -fpic assume that the dynamic loader has processed the relocations before it is run?
But the kernel startup code is running before they can have been done? So even if that C code were 'pic' it could still contain things that are invalid (probably arrays of pointers?).
So you lose one set of bugs and gain another.
David
-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
Re: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 9 lines · David Laight
...
Quoted 4 lines
Due to the non-negligible impact of PIE, perhaps some kind of
CONFIG_PIEconfig definition should be introduced, so the assembly code would be able to choose optimal asm sequence when PIE and non-PIE is requested?I wouldn't have thought that performance mattered in the asm code that runs during startup?
No, not the code that runs only once, where performance impact can be tolerated.
This one:
https://lore.kernel.org/lkml/[email protected]/
Uros.
H. Peter Anvin
reviewerRe: [RFC PATCH 25/28] x86: Use PIE codegen for the core kernel
Quoted 11 lines · Uros Bizjak
Quoted 9 lines · David Laight
...
Quoted 4 lines
Due to the non-negligible impact of PIE, perhaps some kind of
CONFIG_PIEconfig definition should be introduced, so the assembly code would be able to choose optimal asm sequence when PIE and non-PIE is requested?I wouldn't have thought that performance mattered in the asm code that runs during startup?
Quoted 8 lines
No, not the code that runs only once, where performance impact can be tolerated.
This one:
https://lore.kernel.org/lkml/[email protected]/
Uros.
Yeah, running the kernel proper as PIE seems like a lose all around. The decompressor, ELF stub, etc, are of course a different matter entirely (and at least the latter can't rely on the small or kernel memory models anyway.)
RFC PATCH 26/28 x86/boot: Implement support for ELF RELA/RELR relocations
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Add support for standard dynamic ELF relocations to perform the virtual relocation of the core kernel at boot. The RELR format results in a 10x reduction in memory footprint of the relocation data, and can be generated by the linker directly. This removes the need for
a) a host tool 'relocs' and a bespoke, clunky relocation table format where the table is simply concatenated to the vmlinux payload when building the decompressor;
b) dependence on the --emit-relocs linker switch, which dumps static, intermediate build time relocations into the ELF binary, to be subsequently used as runtime relocations.
The latter is especially problematic, as linkers may apply relaxations that result in the code going out of sync with the static relocation that annotated it in the input. This requires additional work on the part of the linker to update the static relocation, which is not even possible in all cases. Therefore, it is much better to consume a runtime, dynamic relocation format in the way it was intended.
This will require switching to linking vmlinux in PIE mode - this is implemented in a subsequent patch.
--- Documentation/arch/x86/zero-page.rst | 3 +- arch/x86/Kconfig | 1 + arch/x86/include/asm/setup.h | 1 + arch/x86/include/uapi/asm/bootparam.h | 2 +- arch/x86/kernel/head64.c | 36 ++++++++++++++++++++ arch/x86/kernel/head_64.S | 5 +++ arch/x86/kernel/vmlinux.lds.S | 24 +++++++++---- 7 files changed, 64 insertions(+), 8 deletions(-)
Documentation/arch/x86/zero-page.rst +2 -1
@@ -3,7 +3,7 @@ ========= Zero Page ========= -The additional fields in struct boot_params as a part of 32-bit boot +The additional fields in struct boot_params as a part of 32/64-bit boot protocol of kernel. These should be filled by bootloader or 16-bit real-mode setup code of the kernel. References/settings to it mainly are in::
@@ -20,6 +20,7 @@ Offset/Size Proto Name Meaning 060/010 ALL ist_info Intel SpeedStep (IST) BIOS support information (struct ist_info) 070/008 ALL acpi_rsdp_addr Physical address of ACPI RSDP table +078/008 64-bit kaslr_va_shift Virtual kASLR displacement of the core kernel 080/010 ALL hd0_info hd0 disk parameter, OBSOLETE!! 090/010 ALL hd1_info hd1 disk parameter, OBSOLETE!! 0A0/010 ALL sys_desc_table System description table (struct sys_desc_table),
arch/x86/Kconfig +1 -0
@@ -26,6 +26,7 @@ config X86_64 depends on 64BIT # Options that are inherently 64-bit kernel only: select ARCH_HAS_GIGANTIC_PAGE + select ARCH_HAS_RELR select ARCH_SUPPORTS_INT128 if CC_HAS_INT128 select ARCH_SUPPORTS_PER_VMA_LOCK select ARCH_SUPPORTS_HUGE_PFNMAP if TRANSPARENT_HUGEPAGE
arch/x86/include/asm/setup.h +1 -0
@@ -51,6 +51,7 @@ extern void reserve_standard_io_resources(void);extern void i386_reserve_resources (void );extern unsigned long __startup_64 (unsigned long p2v_offset,struct boot_params *bp);extern void startup_64_setup_gdt_idt (void ); +extern void startup_64_apply_relocations (struct boot_params *bp);extern void early_setup_idt (void );extern void __init do_early_exception (struct pt_regs *regs,int trapnr);
arch/x86/include/uapi/asm/bootparam.h +1 -1
@@ -120,7 +120,7 @@ struct boot_params {__u64 tboot_addr;/* 0x058 */ struct ist_info ist_info;/* 0x060 */ __u64 acpi_rsdp_addr;/* 0x070 */ -__u8 _pad3[8 ];/* 0x078 */ +__u64 kaslr_va_shift;/* 0x078 */ __u8 hd0_info[16 ];/* obsolete! */ /* 0x080 */ __u8 hd1_info[16 ];/* obsolete! */ /* 0x090 */ struct sys_desc_table sys_desc_table;/* obsolete! */ /* 0x0a0 */
arch/x86/kernel/head64.c +36 -0
@@ -20,6 +20,7 @@#include <linux/io.h> #include <linux/memblock.h> #include <linux/cc_platform.h> +#include <linux/elf.h> #include <linux/pgtable.h> #include <asm/asm.h>
@@ -588,3 +589,38 @@ void __head startup_64_setup_gdt_idt(void)startup_64_load_idt (handler); } + +#ifdef CONFIG_RELOCATABLE +void __head startup_64_apply_relocations(struct boot_params *bp) +{ +extern const Elf64_Rela __rela_start[], __rela_end[]; +extern const u64 __relr_start[], __relr_end[]; +u64 va_offset = (u64 )RIP_REL_REF (_text) - __START_KERNEL; +u64 va_shift = bp->kaslr_va_shift; +u64 *place =NULL ; + +if (!va_shift) +return ; + +for (const Elf64_Rela *r = __rela_start; r < __rela_end; r++) { +if (ELF64_R_TYPE (r->r_info) !=R_X86_64_RELATIVE ) +continue ; + + place = (u64 *)(r->r_offset + va_offset); + *place += va_shift; + } + +for (const u64 *rel = __relr_start; rel < __relr_end; rel++) { +if ((*rel &1 ) ==0 ) { + place = (u64 *)(*rel + va_offset); + *place++ += va_shift; +continue ; + } + +for (u64 *p = place, r = *rel >>1 ; r; p++, r >>=1 ) +if (r &1 ) + *p += va_shift; + place +=63 ; + } +} +#endif
arch/x86/kernel/head_64.S +5 -0
@@ -74,6 +74,11 @@ SYM_CODE_START_NOALIGN(startup_64) cdq wrmsr +#ifdef CONFIG_RELOCATABLE + movq %r15, %rdi + call startup_64_apply_relocations +#endif + call startup_64_setup_gdt_idt /* Now switch to __KERNEL_CS so IRET works reliably */
arch/x86/kernel/vmlinux.lds.S +18 -6
@@ -240,6 +240,18 @@ xen_elfnote_phys32_entry_offset = :init #endif + .init.rela : { + __rela_start = .; + *(.rela.*) *(.rela_*) + __rela_end = .; + } + + .init.relr : { + __relr_start = .; + *(.relr.*) + __relr_end = .; + } + /* * Section for code used exclusively before alternatives are run. All * references to such code must be patched out by alternatives, normally
@@ -469,12 +481,6 @@ xen_elfnote_phys32_entry_offset = *(.got) *(.igot.*) } ASSERT(SIZEOF(.got) == 0, "Unexpected GOT entries detected!") -#endif - - .plt : { - *(.plt) *(.plt.*) *(.iplt) - } - ASSERT(SIZEOF(.plt) == 0, "Unexpected run-time procedure linkages detected!") .rel.dyn : { *(.rel.*) *(.rel_*)
@@ -485,6 +491,12 @@ xen_elfnote_phys32_entry_offset = *(.rela.*) *(.rela_*) } ASSERT(SIZEOF(.rela.dyn) == 0, "Unexpected run-time relocations (.rela) detected!") +#endif + + .plt : { + *(.plt) *(.plt.*) *(.iplt) + } + ASSERT(SIZEOF(.plt) == 0, "Unexpected run-time procedure linkages detected!") } /*
Signature · 1 line
2.46.0.792.g87dc391469-goog
RFC PATCH 27/28 x86/kernel: Switch to PIE linking for the core kernel
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
Build the kernel as a Position Independent Executable (PIE). This results in more efficient relocation processing for the virtual displacement of the kernel (for KASLR). More importantly, it instructs the linker to generate what is actually needed (a program that can be moved around in memory before execution), which is better than having to rely on the linker to create a position dependent binary that happens to tolerate being moved around after poking it in exactly the right manner.
Note that this means that all codegen should be compatible with PIE, including Rust objects, so this needs to switch to the small code model with the PIE relocation model as well.
--- arch/x86/Kconfig | 2 +- arch/x86/Makefile | 11 +++++++---- arch/x86/boot/compressed/misc.c | 2 ++ arch/x86/kernel/vmlinux.lds.S | 5 +++++ drivers/firmware/efi/libstub/x86-stub.c | 2 ++ 5 files changed, 17 insertions(+), 5 deletions(-)
arch/x86/Kconfig +1 -1
@@ -2187,7 +2187,7 @@ config RANDOMIZE_BASE # Relocation on x86 needs some additional build support config X86_NEED_RELOCS def_bool y - depends on RANDOMIZE_BASE || (X86_32 && RELOCATABLE) + depends on X86_32 && RELOCATABLE config PHYSICAL_ALIGN hex "Alignment value to which kernel should be aligned"
arch/x86/Makefile +7 -4
@@ -206,9 +206,8 @@ else PIE_CFLAGS-$(CONFIG_SMP) += -mstack-protector-guard-reg=gs endif - # Don't emit relaxable GOTPCREL relocations - KBUILD_AFLAGS_KERNEL += -Wa,-mrelax-relocations=no - KBUILD_CFLAGS_KERNEL += -Wa,-mrelax-relocations=no $(PIE_CFLAGS-y) + KBUILD_CFLAGS_KERNEL += $(PIE_CFLAGS-y) + KBUILD_RUSTFLAGS_KERNEL += -Ccode-model=small -Crelocation-model=pie endif #
@@ -264,12 +263,16 @@ else LDFLAGS_vmlinux := endif +ifdef CONFIG_X86_64 +ldflags-pie-$(CONFIG_LD_IS_LLD) := --apply-dynamic-relocs +ldflags-pie-$(CONFIG_LD_IS_BFD) := -z call-nop=suffix-nop +LDFLAGS_vmlinux += --pie -z text $(ldflags-pie-y) + # # The 64-bit kernel must be aligned to 2MB. Pass -z max-page-size=0x200000 to # the linker to force 2MB page size regardless of the default page size used # by the linker. # -ifdef CONFIG_X86_64 LDFLAGS_vmlinux += -z max-page-size=0x200000 endif
arch/x86/boot/compressed/misc.c +2 -0
@@ -495,6 +495,8 @@ asmlinkage __visible void *extract_kernel(void *rmode, unsigned char *output)error ("Destination virtual address changed when not relocatable" );#endif + boot_params_ptr->kaslr_va_shift = virt_addr -LOAD_PHYSICAL_ADDR ; +debug_putstr ("\nDecompressing Linux... " );if (init_unaccepted_memory ()) {
arch/x86/kernel/vmlinux.lds.S +5 -0
@@ -459,6 +459,11 @@ xen_elfnote_phys32_entry_offset = DISCARDS + /DISCARD/ : { + *(.dynsym .gnu.hash .hash .dynamic .dynstr) + *(.interp .dynbss .eh_frame .sframe) + } + /* * Make sure that the .got.plt is either completely empty or it * contains only the lazy dispatch entries.
drivers/firmware/efi/libstub/x86-stub.c +2 -0
@@ -912,6 +912,8 @@ static efi_status_t efi_decompress_kernel(unsigned long *kernel_entry)if (status !=EFI_SUCCESS )return status; + boot_params_ptr->kaslr_va_shift = virt_addr -LOAD_PHYSICAL_ADDR ; + entry =decompress_kernel ((void *)addr, virt_addr, error);if (entry ==ULONG_MAX ) {efi_free (alloc_size, addr);
Signature · 1 line
2.46.0.792.g87dc391469-goog
Re: [RFC PATCH 27/28] x86/kernel: Switch to PIE linking for the core kernel
Quoted 11 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Build the kernel as a Position Independent Executable (PIE). This results in more efficient relocation processing for the virtual displacement of the kernel (for KASLR). More importantly, it instructs the linker to generate what is actually needed (a program that can be moved around in memory before execution), which is better than having to rely on the linker to create a position dependent binary that happens to tolerate being moved around after poking it in exactly the right manner.
Quoted 3 lines
Note that this means that all codegen should be compatible with PIE, including Rust objects, so this needs to switch to the small code model with the PIE relocation model as well.
I think that related to this work is the patch series [1] that introduces the changes necessary to build the kernel as Position Independent Executable (PIE) on x86_64 [1]. There are some more places that need to be adapted for PIE. The patch series also introduces objtool functionality to add validation for x86 PIE.
[1] "[PATCH RFC 00/43] x86/pie: Make kernel image's virtual address flexible"
https://lore.kernel.org/lkml/[email protected]/
Uros.
Quoted 102 lines
Signed-off-by Ard Biesheuvel <[email protected]>--- arch/x86/Kconfig | 2 +- arch/x86/Makefile | 11 +++++++---- arch/x86/boot/compressed/misc.c | 2 ++ arch/x86/kernel/vmlinux.lds.S | 5 +++++ drivers/firmware/efi/libstub/x86-stub.c | 2 ++ 5 files changed, 17 insertions(+), 5 deletions(-)arch/x86/Kconfig +1
-1@@ -2187,7 +2187,7 @@ config RANDOMIZE_BASE # Relocation on x86 needs some additional build support config X86_NEED_RELOCS def_bool y - depends on RANDOMIZE_BASE || (X86_32 && RELOCATABLE) + depends on X86_32 && RELOCATABLE config PHYSICAL_ALIGN hex "Alignment value to which kernel should be aligned"arch/x86/Makefile +7
-4@@ -206,9 +206,8 @@ else PIE_CFLAGS-$(CONFIG_SMP) += -mstack-protector-guard-reg=gs endif - # Don't emit relaxable GOTPCREL relocations - KBUILD_AFLAGS_KERNEL += -Wa,-mrelax-relocations=no - KBUILD_CFLAGS_KERNEL += -Wa,-mrelax-relocations=no $(PIE_CFLAGS-y) + KBUILD_CFLAGS_KERNEL += $(PIE_CFLAGS-y) + KBUILD_RUSTFLAGS_KERNEL += -Ccode-model=small -Crelocation-model=pie endif #@@ -264,12 +263,16 @@ else LDFLAGS_vmlinux := endif +ifdef CONFIG_X86_64 +ldflags-pie-$(CONFIG_LD_IS_LLD) := --apply-dynamic-relocs +ldflags-pie-$(CONFIG_LD_IS_BFD) := -z call-nop=suffix-nop +LDFLAGS_vmlinux += --pie -z text $(ldflags-pie-y) + # # The 64-bit kernel must be aligned to 2MB. Pass -z max-page-size=0x200000 to # the linker to force 2MB page size regardless of the default page size used # by the linker. # -ifdef CONFIG_X86_64 LDFLAGS_vmlinux += -z max-page-size=0x200000 endifarch/x86/boot/compressed/misc.c +2
-0@@ -495,6 +495,8 @@ asmlinkage __visible void *extract_kernel(void *rmode, unsigned char *output)error ("Destination virtual address changed when not relocatable" );#endif + boot_params_ptr->kaslr_va_shift = virt_addr -LOAD_PHYSICAL_ADDR ; +debug_putstr ("\nDecompressing Linux... " );if (init_unaccepted_memory ()) {arch/x86/kernel/vmlinux.lds.S +5
-0@@ -459,6 +459,11 @@ xen_elfnote_phys32_entry_offset = DISCARDS + /DISCARD/ : { + *(.dynsym .gnu.hash .hash .dynamic .dynstr) + *(.interp .dynbss .eh_frame .sframe) + } + /* * Make sure that the .got.plt is either completely empty or it * contains only the lazy dispatch entries.drivers/firmware/efi/libstub/x86-stub.c +2
-1@@ -912,6 +912,8 @@ static efi_status_t efi_decompress_kernel(unsigned long *kernel_entry)if (status !=EFI_SUCCESS )return status; + boot_params_ptr->kaslr_va_shift = virt_addr -LOAD_PHYSICAL_ADDR ; + entry =decompress_kernel ((void *)addr, virt_addr, error);if (entry ==ULONG_MAX ) {efi_free (alloc_size, addr); --2.46.0.792.g87dc391469-goog
Ard Biesheuvel
maintainerRe: [RFC PATCH 27/28] x86/kernel: Switch to PIE linking for the core kernel
Quoted 17 lines · Uros Bizjak
Quoted 11 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Build the kernel as a Position Independent Executable (PIE). This results in more efficient relocation processing for the virtual displacement of the kernel (for KASLR). More importantly, it instructs the linker to generate what is actually needed (a program that can be moved around in memory before execution), which is better than having to rely on the linker to create a position dependent binary that happens to tolerate being moved around after poking it in exactly the right manner.
Quoted 3 lines
Note that this means that all codegen should be compatible with PIE, including Rust objects, so this needs to switch to the small code model with the PIE relocation model as well.
Quoted 9 lines
I think that related to this work is the patch series [1] that introduces the changes necessary to build the kernel as Position Independent Executable (PIE) on
x86_64[1]. There are some more places that need to be adapted for PIE. The patch series also introduces objtool functionality to add validation for x86 PIE.[1] "[PATCH RFC 00/43] x86/pie: Make kernel image's virtual address flexible"
https://lore.kernel.org/lkml/[email protected]/
Hi Uros,
I am aware of that discussion, as I took part in it as well.
I don't think any of those changes are actually needed now - did you notice anything in particular that is missing?
Re: [RFC PATCH 27/28] x86/kernel: Switch to PIE linking for the core kernel
Quoted 29 lines · Ard Biesheuvel
Quoted 17 lines · Uros Bizjak
Quoted 11 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Build the kernel as a Position Independent Executable (PIE). This results in more efficient relocation processing for the virtual displacement of the kernel (for KASLR). More importantly, it instructs the linker to generate what is actually needed (a program that can be moved around in memory before execution), which is better than having to rely on the linker to create a position dependent binary that happens to tolerate being moved around after poking it in exactly the right manner.
Quoted 3 lines
Note that this means that all codegen should be compatible with PIE, including Rust objects, so this needs to switch to the small code model with the PIE relocation model as well.
Quoted 9 lines
I think that related to this work is the patch series [1] that introduces the changes necessary to build the kernel as Position Independent Executable (PIE) on
x86_64[1]. There are some more places that need to be adapted for PIE. The patch series also introduces objtool functionality to add validation for x86 PIE.[1] "[PATCH RFC 00/43] x86/pie: Make kernel image's virtual address flexible"
https://lore.kernel.org/lkml/[email protected]/
Quoted 6 lines
Hi Uros,
I am aware of that discussion, as I took part in it as well.
I don't think any of those changes are actually needed now - did you notice anything in particular that is missing?
Some time ago I went through the kernel sources and proposed several patches that changed all trivial occurrences of non-RIP addresses to RIP ones. The work was partially based on the mentioned patch series, and I remember, I left some of them out [e.g. 1], because they required a temporary variable. Also, there was discussion about ftrace [2], where no solution was found.
Looking through your series, I didn't find some of the non-RIP -> RIP changes proposed by the original series (especially the ftrace part), and noticed that there is no objtool validator proposed to ensure that all generated code is indeed PIE compatible.
Speaking of non-RIP -> RIP changes that require a temporary - would it be beneficial to make a macro that would use the RIP form only when #ifdef CONFIG_X86_PIE? That would avoid code size increase when PIE is not needed.
[1] https://lore.kernel.org/lkml/a0b69f3fac1834c05f960b916cc6eb0004cdffbf.1682673543.git.houwenlong.hwl@antgroup.com/
[2] https://lore.kernel.org/lkml/[email protected]/
[3] https://lore.kernel.org/lkml/226af8c63c5bfa361763dd041a997ee84fe926cf.1682673543.git.houwenlong.hwl@antgroup.com/
Thanks and best regards,
Uros.
Ard Biesheuvel
maintainerRe: [RFC PATCH 27/28] x86/kernel: Switch to PIE linking for the core kernel
Quoted 35 lines · Uros Bizjak
Quoted 33 lines · Ard Biesheuvel
Quoted 17 lines · Uros Bizjak
Quoted 11 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Build the kernel as a Position Independent Executable (PIE). This results in more efficient relocation processing for the virtual displacement of the kernel (for KASLR). More importantly, it instructs the linker to generate what is actually needed (a program that can be moved around in memory before execution), which is better than having to rely on the linker to create a position dependent binary that happens to tolerate being moved around after poking it in exactly the right manner.
Quoted 3 lines
Note that this means that all codegen should be compatible with PIE, including Rust objects, so this needs to switch to the small code model with the PIE relocation model as well.
Quoted 9 lines
I think that related to this work is the patch series [1] that introduces the changes necessary to build the kernel as Position Independent Executable (PIE) on
x86_64[1]. There are some more places that need to be adapted for PIE. The patch series also introduces objtool functionality to add validation for x86 PIE.[1] "[PATCH RFC 00/43] x86/pie: Make kernel image's virtual address flexible"
https://lore.kernel.org/lkml/[email protected]/Hi Uros,
I am aware of that discussion, as I took part in it as well.
Quoted 8 lines
Quoted 2 lines
I don't think any of those changes are actually needed now - did you notice anything in particular that is missing?
Some time ago I went through the kernel sources and proposed several patches that changed all trivial occurrences of non-RIP addresses to RIP ones. The work was partially based on the mentioned patch series, and I remember, I left some of them out [e.g. 1], because they required a temporary variable.
I have a similar patch in my series, but the DEBUG_ENTRY code just uses
pushf 1f@GOTPCREL(%rip)
so no temporaries are needed.
Quoted 3 lines
Also, there was discussion about ftrace
[2], where no solution was found.
When linking with -z call-nop=suffix-nop, the __fentry__ call via the GOT will be relaxed by the linker into a 5 byte call followed by a 1 byte NOP, so I don't think we need to do anything special here. It might mean we currently lose -mnop-mcount until we find a solution for that in the compiler. In case you remember, I contributed and you merged a GCC patch that makes the __fentry__ emission logic honour -fdirect-access-external-data which should help here. This landed in GCC 14.
Quoted 5 lines
Looking through your series, I didn't find some of the non-RIP -> RIP changes proposed by the original series (especially the ftrace part), and noticed that there is no objtool validator proposed to ensure that all generated code is indeed PIE compatible.
What would be the point of that? The linker will complain and throw an error if the code cannot be converted into a PIE executable, so I don't think we need objtool's help for that.
Quoted 5 lines
Speaking of non-RIP -> RIP changes that require a temporary - would it be beneficial to make a macro that would use the RIP form only when #ifdef
CONFIG_X86_PIE? That would avoid code size increase when PIE is not needed.
This series does not make the PIE support configurable. Do you think the code size increase is a concern if all GOT based symbol references are elided, e.g, via -fdirect-access-external-data?
Re: [RFC PATCH 27/28] x86/kernel: Switch to PIE linking for the core kernel
Quoted 71 lines · Ard Biesheuvel
Quoted 35 lines · Uros Bizjak
Quoted 33 lines · Ard Biesheuvel
Quoted 17 lines · Uros Bizjak
Quoted 11 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Build the kernel as a Position Independent Executable (PIE). This results in more efficient relocation processing for the virtual displacement of the kernel (for KASLR). More importantly, it instructs the linker to generate what is actually needed (a program that can be moved around in memory before execution), which is better than having to rely on the linker to create a position dependent binary that happens to tolerate being moved around after poking it in exactly the right manner.
Quoted 3 lines
Note that this means that all codegen should be compatible with PIE, including Rust objects, so this needs to switch to the small code model with the PIE relocation model as well.
Quoted 9 lines
I think that related to this work is the patch series [1] that introduces the changes necessary to build the kernel as Position Independent Executable (PIE) on
x86_64[1]. There are some more places that need to be adapted for PIE. The patch series also introduces objtool functionality to add validation for x86 PIE.[1] "[PATCH RFC 00/43] x86/pie: Make kernel image's virtual address flexible"
https://lore.kernel.org/lkml/[email protected]/Hi Uros,
I am aware of that discussion, as I took part in it as well.
Quoted 8 lines
Quoted 2 lines
I don't think any of those changes are actually needed now - did you notice anything in particular that is missing?
Some time ago I went through the kernel sources and proposed several patches that changed all trivial occurrences of non-RIP addresses to RIP ones. The work was partially based on the mentioned patch series, and I remember, I left some of them out [e.g. 1], because they required a temporary variable.
I have a similar patch in my series, but the
DEBUG_ENTRYcode just usespushf 1f@GOTPCREL(%rip)
so no temporaries are needed.
Quoted 3 lines
Also, there was discussion about ftrace
[2], where no solution was found.When linking with -z call-nop=suffix-nop, the
__fentry__call via the GOT will be relaxed by the linker into a 5 byte call followed by a 1 byte NOP, so I don't think we need to do anything special here. It might mean we currently lose -mnop-mcount until we find a solution for that in the compiler. In case you remember, I contributed and you merged a GCC patch that makes the__fentry__emission logic honour -fdirect-access-external-data which should help here. This landed in GCC 14.Quoted 5 lines
Looking through your series, I didn't find some of the non-RIP -> RIP changes proposed by the original series (especially the ftrace part), and noticed that there is no objtool validator proposed to ensure that all generated code is indeed PIE compatible.
Quoted 3 lines
What would be the point of that? The linker will complain and throw an error if the code cannot be converted into a PIE executable, so I don't think we need objtool's help for that.
Indeed.
Quoted 9 lines
Quoted 5 lines
Speaking of non-RIP -> RIP changes that require a temporary - would it be beneficial to make a macro that would use the RIP form only when #ifdef
CONFIG_X86_PIE? That would avoid code size increase when PIE is not needed.This series does not make the PIE support configurable. Do you think the code size increase is a concern if all GOT based symbol references are elided, e.g, via -fdirect-access-external-data?
I was looking at the code size measurement of the original patch series (perhaps these are not relevant with your series) and I think 2.2% - 2.4% code size increase can be problematic. Can you perhaps provide new code size increase measurements with your patches applied?
Thanks and BR,
Uros.
Re: [RFC PATCH 27/28] x86/kernel: Switch to PIE linking for the core kernel
In reply to Ard BiesheuvelQuoted 104 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Build the kernel as a Position Independent Executable (PIE). This results in more efficient relocation processing for the virtual displacement of the kernel (for KASLR). More importantly, it instructs the linker to generate what is actually needed (a program that can be moved around in memory before execution), which is better than having to rely on the linker to create a position dependent binary that happens to tolerate being moved around after poking it in exactly the right manner.
Note that this means that all codegen should be compatible with PIE, including Rust objects, so this needs to switch to the small code model with the PIE relocation model as well.
Signed-off-by Ard Biesheuvel <[email protected]>--- arch/x86/Kconfig | 2 +- arch/x86/Makefile | 11 +++++++---- arch/x86/boot/compressed/misc.c | 2 ++ arch/x86/kernel/vmlinux.lds.S | 5 +++++ drivers/firmware/efi/libstub/x86-stub.c | 2 ++ 5 files changed, 17 insertions(+), 5 deletions(-)arch/x86/Kconfig +1
-1@@ -2187,7 +2187,7 @@ config RANDOMIZE_BASE # Relocation on x86 needs some additional build support config X86_NEED_RELOCS def_bool y - depends on RANDOMIZE_BASE || (X86_32 && RELOCATABLE) + depends on X86_32 && RELOCATABLE config PHYSICAL_ALIGN hex "Alignment value to which kernel should be aligned"arch/x86/Makefile +7
-4@@ -206,9 +206,8 @@ else PIE_CFLAGS-$(CONFIG_SMP) += -mstack-protector-guard-reg=gs endif - # Don't emit relaxable GOTPCREL relocations - KBUILD_AFLAGS_KERNEL += -Wa,-mrelax-relocations=no - KBUILD_CFLAGS_KERNEL += -Wa,-mrelax-relocations=no $(PIE_CFLAGS-y) + KBUILD_CFLAGS_KERNEL += $(PIE_CFLAGS-y) + KBUILD_RUSTFLAGS_KERNEL += -Ccode-model=small -Crelocation-model=pie endif #@@ -264,12 +263,16 @@ else LDFLAGS_vmlinux := endif +ifdef CONFIG_X86_64 +ldflags-pie-$(CONFIG_LD_IS_LLD) := --apply-dynamic-relocs +ldflags-pie-$(CONFIG_LD_IS_BFD) := -z call-nop=suffix-nop +LDFLAGS_vmlinux += --pie -z text $(ldflags-pie-y) + # # The 64-bit kernel must be aligned to 2MB. Pass -z max-page-size=0x200000 to # the linker to force 2MB page size regardless of the default page size used # by the linker. # -ifdef CONFIG_X86_64 LDFLAGS_vmlinux += -z max-page-size=0x200000 endifarch/x86/boot/compressed/misc.c +2
-0@@ -495,6 +495,8 @@ asmlinkage __visible void *extract_kernel(void *rmode, unsigned char *output)error ("Destination virtual address changed when not relocatable" );#endif + boot_params_ptr->kaslr_va_shift = virt_addr -LOAD_PHYSICAL_ADDR ; +debug_putstr ("\nDecompressing Linux... " );if (init_unaccepted_memory ()) {arch/x86/kernel/vmlinux.lds.S +5
-0@@ -459,6 +459,11 @@ xen_elfnote_phys32_entry_offset = DISCARDS + /DISCARD/ : { + *(.dynsym .gnu.hash .hash .dynamic .dynstr) + *(.interp .dynbss .eh_frame .sframe) + } + /* * Make sure that the .got.plt is either completely empty or it * contains only the lazy dispatch entries.drivers/firmware/efi/libstub/x86-stub.c +0
-0
drivers/firmware/efi/libstub/x86-stub.c static efi_status_t efi_decompress_kernel(unsigned long *kernel_entry) · line 919
Quoted 8 lines
drivers/firmware/efi/libstub/x86-stub.c +2
-0@@ -912,6 +912,8 @@ static efi_status_t efi_decompress_kernel(unsigned long *kernel_entry)if (status !=EFI_SUCCESS )return status; + boot_params_ptr->kaslr_va_shift = virt_addr -LOAD_PHYSICAL_ADDR ; + entry =decompress_kernel ((void *)addr, virt_addr, error);if (entry ==ULONG_MAX ) {efi_free (alloc_size, addr);
This patch causes a build failure here (on 64-bit):
LD .tmp_vmlinux2 NM .tmp_vmlinux2.syms KSYMS .tmp_vmlinux2.kallsyms.S AS .tmp_vmlinux2.kallsyms.o LD vmlinux BTFIDS vmlinux
WARN: resolve_btfids: unresolved symbol bpf_lsm_key_free
FAILED elf_update(WRITE): invalid section entry size
make[5]: *** [scripts/Makefile.vmlinux:34: vmlinux] Error 255
make[5]: *** Deleting file 'vmlinux'
make[4]: *** [Makefile:1153: vmlinux] Error 2
make[3]: *** [debian/rules:74: build-arch] Error 2
dpkg-buildpackage: error: make -f debian/rules binary subprocess returned exit status 2 make[2]: *** [scripts/Makefile.package:121: bindeb-pkg] Error 2
make[1]: *** [/home/opc/linux-mainline-worktree2/Makefile:1544: bindeb-pkg] Error 2
make: *** [Makefile:224: __sub-make] Error 2
The parent commit builds fine. With V=1:
+ ldflags='-m elf_x86_64 -z noexecstack --pie -z text -z call-nop=suffix-nop -z max-page-size=0x200000 --build-id=sha1 --orphan-handling=warn --script=./arch/x86/kernel/vmlinux.lds -Map=vmlinux.map'
+ ld -m elf_x86_64 -z noexecstack --pie -z text -z call-nop=suffix-nop -z max-page-size=0x200000 --build-id=sha1 --orphan-handling=warn --script=./arch/x86/kernel/vmlinux.lds -Map=vmlinux.map -o vmlinux --whole-archive vmlinux.a .vmlinux.export.o init/version-timestamp.o --no-whole-archive --start-group --end-group .tmp_vmlinux2.kallsyms.o .tmp_vmlinux1.btf.o
+ is_enabled CONFIG_DEBUG_INFO_BTF
+ grep -q '^CONFIG_DEBUG_INFO_BTF=y' include/config/auto.conf
+ info BTFIDS vmlinux
+ printf ' %-7s %s\n' BTFIDS vmlinux
BTFIDS vmlinux
+ ./tools/bpf/resolve_btfids/resolve_btfids vmlinux
WARN: resolve_btfids: unresolved symbol bpf_lsm_key_free
FAILED elf_update(WRITE): invalid section entry size
I can send the full config off-list if necessary, but looks like it might be enough to set CONFIG_DEBUG_INFO_BTF=y.
Vegard
Ard Biesheuvel
maintainerRe: [RFC PATCH 27/28] x86/kernel: Switch to PIE linking for the core kernel
Quoted 26 lines · Vegard Nossum
Quoted 23 lines · Ard Biesheuvel
From: Ard Biesheuvel <[email protected]>
Build the kernel as a Position Independent Executable (PIE). This results in more efficient relocation processing for the virtual displacement of the kernel (for KASLR). More importantly, it instructs the linker to generate what is actually needed (a program that can be moved around in memory before execution), which is better than having to rely on the linker to create a position dependent binary that happens to tolerate being moved around after poking it in exactly the right manner.
Note that this means that all codegen should be compatible with PIE, including Rust objects, so this needs to switch to the small code model with the PIE relocation model as well.
Signed-off-by Ard Biesheuvel <[email protected]>--- arch/x86/Kconfig | 2 +- arch/x86/Makefile | 11 +++++++---- arch/x86/boot/compressed/misc.c | 2 ++ arch/x86/kernel/vmlinux.lds.S | 5 +++++ drivers/firmware/efi/libstub/x86-stub.c | 2 ++ 5 files changed, 17 insertions(+), 5 deletions(-)
...
Quoted 43 lines
This patch causes a build failure here (on 64-bit):
LD .tmp_vmlinux2 NM .tmp_vmlinux2.syms KSYMS .tmp_vmlinux2.kallsyms.S AS .tmp_vmlinux2.kallsyms.o LD vmlinux BTFIDS vmlinuxWARN:
resolve_btfids: unresolved symbolbpf_lsm_key_freeFAILEDelf_update(WRITE): invalid section entry size make[5]: *** [scripts/Makefile.vmlinux:34: vmlinux] Error 255 make[5]: *** Deleting file 'vmlinux'
make[4]: *** [Makefile:1153: vmlinux] Error 2
make[3]: *** [debian/rules:74: build-arch] Error 2
dpkg-buildpackage: error: make -f debian/rules binary subprocess returned exit status 2
make[2]: *** [scripts/Makefile.package:121: bindeb-pkg] Error 2 make[1]: *** [/home/opc/linux-mainline-worktree2/Makefile:1544:
bindeb-pkg] Error 2
make: *** [Makefile:224: __sub-make] Error 2The parent commit builds fine. With V=1:
+ ldflags='-m
elf_x86_64-z noexecstack --pie -z text -z call-nop=suffix-nop -z max-page-size=0x200000 --build-id=sha1 --orphan-handling=warn --script=./arch/x86/kernel/vmlinux.lds(quoted diff) +1
-5-Map=vmlinux.map' + ld -m elf_x86_64 -z noexecstack --pie -z text -z call-nop=suffix-nop -z max-page-size=0x200000 --build-id=sha1 --orphan-handling=warn --script=./arch/x86/kernel/vmlinux.lds -Map=vmlinux.map -o vmlinux --whole-archive vmlinux.a .vmlinux.export.o init/version-timestamp.o --no-whole-archive --start-group --end-group .tmp_vmlinux2.kallsyms.o.tmp_vmlinux1.btf.o
+is_enabledCONFIG_DEBUG_INFO_BTF
+ grep -q '^CONFIG_DEBUG_INFO_BTF=y'include/config/auto.conf
+ info BTFIDS vmlinux
+ printf ' %-7s %s\n' BTFIDS vmlinux
BTFIDS vmlinux
+./tools/bpf/resolve_btfids/resolve_btfidsvmlinux
WARN:resolve_btfids: unresolved symbolbpf_lsm_key_freeFAILEDelf_update(WRITE): invalid section entry size
Quoted 3 lines
I can send the full config off-list if necessary, but looks like it might be enough to set
CONFIG_DEBUG_INFO_BTF=y.
Thanks for the report. Turns out that adding the GOT to .rodata bumps the section's sh_entsize to 8, and libelf complains if the section size is not a multiple of the entry size.
I'll include a fix in the next revision.
RFC PATCH 28/28 x86/tools: Drop x86_64 support from 'relocs' tool
In reply to Ard BiesheuvelFrom: Ard Biesheuvel <[email protected]>
The relocs tool is no longer used on vmlinux, which is the only 64-bit ELF executable that it used to operate on in the 64-bit build. (It is still used for parts of the decompressor)
So drop the 64-bit handling - it is dead code now.
--- arch/x86/tools/Makefile | 2 +- arch/x86/tools/relocs.c | 178 +------------------- arch/x86/tools/relocs.h | 9 +- arch/x86/tools/relocs_64.c | 18 -- arch/x86/tools/relocs_common.c | 11 +- 5 files changed, 9 insertions(+), 209 deletions(-)
arch/x86/tools/Makefile +1 -1
@@ -40,7 +40,7 @@ $(obj)/insn_sanity.o: $(srctree)/tools/arch/x86/lib/insn.c $(srctree)/tools/arch HOST_EXTRACFLAGS += -I$(srctree)/tools/include hostprogs += relocs -relocs-objs := relocs_32.o relocs_64.o relocs_common.o +relocs-objs := relocs_32.o relocs_common.o PHONY += relocs relocs: $(obj)/relocs @:
arch/x86/tools/relocs.c +2 -176
265 lines not on this page · raw message
arch/x86/tools/relocs.h +3 -6
@@ -32,10 +32,7 @@ enum symtype {S_NSYMTYPES }; -void process_32 (FILE *fp,int use_real_mode,int as_text, -int show_absolute_syms,int show_absolute_relocs, -int show_reloc_info); -void process_64 (FILE *fp,int use_real_mode,int as_text, -int show_absolute_syms,int show_absolute_relocs, -int show_reloc_info); +void process (FILE *fp,int use_real_mode,int as_text, +int show_absolute_syms,int show_absolute_relocs, +int show_reloc_info);#endif /* RELOCS_H */
arch/x86/tools/relocs_64.c deleted +0 -18
@@ -1,18 +0,0 @@ -// SPDX-License-Identifier: GPL-2.0 -#include "relocs.h" - -#define ELF_BITS 64 - -#define ELF_MACHINE EM_X86_64 -#define ELF_MACHINE_NAME "x86_64" -#define SHT_REL_TYPE SHT_RELA -#define Elf_Rel Elf64_Rela - -#define ELF_CLASS ELFCLASS64 -#define ELF_R_SYM (val) ELF64_R_SYM(val) -#define ELF_R_TYPE (val) ELF64_R_TYPE(val) -#define ELF_ST_TYPE (o) ELF64_ST_TYPE(o) -#define ELF_ST_BIND (o) ELF64_ST_BIND(o) -#define ELF_ST_VISIBILITY (o) ELF64_ST_VISIBILITY(o) - -#include "relocs.c"
arch/x86/tools/relocs_common.c +3 -8
@@ -72,14 +72,9 @@ int main(int argc, char **argv)die ("Cannot read %s: %s" , fname,strerror (errno)); }rewind (fp); -if (e_ident[EI_CLASS ] ==ELFCLASS64 ) -process_64 (fp, use_real_mode, as_text, - show_absolute_syms, show_absolute_relocs, - show_reloc_info); -else -process_32 (fp, use_real_mode, as_text, - show_absolute_syms, show_absolute_relocs, - show_reloc_info); +process (fp, use_real_mode, as_text, + show_absolute_syms, show_absolute_relocs, + show_reloc_info);fclose (fp);return 0 ; }
Signature · 1 line
2.46.0.792.g87dc391469-goog