[lld] r244691 - COFF: Align sections to 512-byte boundaries on disk.
Sean Silva via llvm-commits
llvm-commits at lists.llvm.org
Thu Aug 13 16:04:01 PDT 2015
On Thu, Aug 13, 2015 at 12:51 AM, Rui Ueyama <ruiu at google.com> wrote:
> I think I do understand how the paging mechanism works. :) We are talking
> about different things. My question is why you think a file offset must be
> at a 4K boundary in order to map it efficiently to memory. To me you seems
> to be claiming that mmap(0, /*length*/4096, PROT_READ|PROT_EXEC, 0, SomeFD,
> /*file offset*/5120) is much inefficient than mmap(0, /*length*/4096,
> PROT_READ|PROT_EXEC, 0, SomeFD, /*file offset*/4096) because of the file
> offset of the former mmap call is not a multiple of 4096. And I'm saying
> that that's not true.
>
The paging layer generally has the file already in physical memory before
you ask it to map it. When it originally read the file off disk, how would
it know what alignment the file should have in physical memory? The reality
is that the alignment in the file is the alignment in physical memory. If
the paging system had to be aware of having different parts of the file
mapped in at arbitrary alignments it would be much more complicated (in
practice, it will either be a hard error, or will force the kernel to
explicitly make a copy, but the paging system won't do this automatically).
Also, keep in mind that 512 byte "sector size" has almost nothing to do
with how modern kernels/hardware do IO. It is a historical thing.
-- Sean Silva
>
> On Thu, Aug 13, 2015 at 4:10 PM, Sean Silva <chisophugis at gmail.com> wrote:
>
>>
>>
>> On Wed, Aug 12, 2015 at 10:59 PM, Rui Ueyama <ruiu at google.com> wrote:
>>
>>> What I don't understand is that why the offset from the beginning of the
>>> file must be multiple of page size in order to avoid full copy. Windows
>>> requires all sections to be aligned at least 4K in memory and 512 bytes on
>>> file, and I don't see any problem there.
>>>
>>> Let's say we have two sections, A and B, whose sizes are 1024B and
>>> 4096B, respectively. We also assume that A's offset from the beginning of
>>> file is 4096, and B's 5120. The loader can map offset 4096 to 8192 of the
>>> file to some page, and 5120 to 9216 to other page. Why can't that?
>>>
>>
>> From the kernel's perspective of mapping memory (on x86), memory is
>> divided into aligned 4K pieces. 5120 % 4096 == 1024, so in order to map it
>> at an address that is 4K aligned, it must do a full memmove in order to
>> move all the memory by 1024 bytes so that it is 4K aligned. This image
>> maybe helps to understand how a 32-bit x86 CPU understands a virtual memory
>> address:
>> https://upload.wikimedia.org/wikipedia/commons/8/8e/X86_Paging_4K.svg
>>
>> IIRC the resources I learned from are:
>> http://duartes.org/gustavo/blog/post/how-the-kernel-manages-your-memory/
>> http://duartes.org/gustavo/blog/post/the-thing-king/
>> (that web page has many other very, *very* good posts. A list can be seen
>> at: http://duartes.org/gustavo/blog/category/internals/)
>>
>> I think you will find that understanding virtual memory (and TLB) will
>> greatly help you optimize LLD, since many operations in LLD have very high
>> pressure on the virtual memory system.
>>
>> -- Sean Silva
>>
>>
>>>
>>> On Thu, Aug 13, 2015 at 2:44 PM, Sean Silva <chisophugis at gmail.com>
>>> wrote:
>>>
>>>>
>>>>
>>>> On Wed, Aug 12, 2015 at 7:17 PM, Rui Ueyama <ruiu at google.com> wrote:
>>>>
>>>>> I didn't test to see if this change ever has a negative impact on
>>>>> memory usage, but my guess is that's very unlikely because this layout is
>>>>> the same as what MSVC linker creates. If this is inefficient, virtually all
>>>>> Windows executables are being suffered by that, which is unlikely. My
>>>>> understanding is that the kernel maps each section separately to a memory
>>>>> address, so file offset of each section can be given independently from
>>>>> other sections.
>>>>>
>>>>
>>>> The mapping is done at the granularity of aligned 4K pages minimum
>>>> (this is just how the x86 hardware page table mechanism works). A piece of
>>>> the file cannot be moved by an amount that is not a multiple of 4K without
>>>> a full copy.
>>>>
>>>> The only way this could (in the usual case) not have a large overhead
>>>> is for the kernel to do a crazy hack like have special paging semantics for
>>>> files that are executables. This means that when LLD finishes working on a
>>>> memory mapped file, if a section is not 4K aligned at least, then the
>>>> kernel has to then do a copy to make the file conform the the actual memory
>>>> layout it needs to have in the paging subsystem.
>>>>
>>>> Or does windows make full copies of sections always? In other words
>>>> processes don't share e.g. readonly text?
>>>>
>>>> In ELF, the offset in the file and the offset in memory are required to
>>>> be congruent modulo the alignment (see the documentation of p_align in
>>>> http://www.sco.com/developers/gabi/latest/ch5.pheader.html), precisely
>>>> to avoid the need to do crazy hacks like this when loading the program.
>>>>
>>>> You can see that Linux will reject the binary:
>>>> http://lxr.free-electrons.com/source/fs/binfmt_elf.c#L664
>>>> (load_elf_binary)
>>>> -> http://lxr.free-electrons.com/source/fs/binfmt_elf.c#L336 (elf_map)
>>>> -> http://lxr.free-electrons.com/source/mm/util.c#L306 (vm_mmap)
>>>> Notice:
>>>> 312 if (unlikely(offset & ~PAGE_MASK))
>>>> 313 return -EINVAL;
>>>>
>>>> FreeBSD is more lenient, but you can see that the kernel does not like
>>>> the situation when this is violated:
>>>>
>>>> http://src.illumos.org/source/xref/freebsd-head/sys/kern/imgact_elf.c#593
>>>> (__elfN(load_file))
>>>> -->
>>>> http://src.illumos.org/source/xref/freebsd-head/sys/kern/imgact_elf.c#467
>>>> (__elfN(load_section))
>>>> -->
>>>> http://src.illumos.org/source/xref/freebsd-head/sys/kern/imgact_elf.c#398
>>>> (__elfN(map_insert))
>>>> 423 /*
>>>> 424 * The mapping is not page aligned. This means we have
>>>> 425 * to copy the data. Sigh.
>>>> 426 */
>>>>
>>>> -- Sean Silva
>>>>
>>>>
>>>>>
>>>>> On Wed, Aug 12, 2015 at 5:31 PM, Sean Silva <chisophugis at gmail.com>
>>>>> wrote:
>>>>>
>>>>>>
>>>>>>
>>>>>> On Tue, Aug 11, 2015 at 4:09 PM, Rui Ueyama via llvm-commits <
>>>>>> llvm-commits at lists.llvm.org> wrote:
>>>>>>
>>>>>>> Author: ruiu
>>>>>>> Date: Tue Aug 11 18:09:00 2015
>>>>>>> New Revision: 244691
>>>>>>>
>>>>>>> URL: http://llvm.org/viewvc/llvm-project?rev=244691&view=rev
>>>>>>> Log:
>>>>>>> COFF: Align sections to 512-byte boundaries on disk.
>>>>>>>
>>>>>>> Sections must start at page boundaries in memory, but they
>>>>>>> can be aligned to sector boundaries (512-bytes) on disk.
>>>>>>> We aligned them to 4096-byte boundaries even on disk, so we
>>>>>>> wasted disk space a bit.
>>>>>>>
>>>>>>
>>>>>> This will likely force the kernel to copy or otherwise do unnecessary
>>>>>> work when loading. Are you sure that isn't happening? The kernel ideally
>>>>>> wants to just create a couple page table entries. But if it needs to move
>>>>>> things around at <4K granularity to make them properly aligned to their
>>>>>> load address when loading (like this patch I think causes) then it will
>>>>>> need to do copies.
>>>>>>
>>>>>> This can likely be checked by looking for an increase in real memory
>>>>>> usage for the system when the new binaries are loaded (vs. the old
>>>>>> page-aligned ones), since the kernel will have a copy sitting in page cache
>>>>>> and a copy for alignment mapped into the process address space;
>>>>>> alternatively, you can check for the slowdown from the kernel copies when
>>>>>> faulting the memory into the process's address space (or (less likely) it
>>>>>> may do the copies eagerly which should be easy to measure too).
>>>>>>
>>>>>> -- Sean Silva
>>>>>>
>>>>>>
>>>>>>>
>>>>>>> Modified:
>>>>>>> lld/trunk/COFF/Writer.cpp
>>>>>>> lld/trunk/test/COFF/baserel.test
>>>>>>> lld/trunk/test/COFF/hello32.test
>>>>>>>
>>>>>>> Modified: lld/trunk/COFF/Writer.cpp
>>>>>>> URL:
>>>>>>> http://llvm.org/viewvc/llvm-project/lld/trunk/COFF/Writer.cpp?rev=244691&r1=244690&r2=244691&view=diff
>>>>>>>
>>>>>>> ==============================================================================
>>>>>>> --- lld/trunk/COFF/Writer.cpp (original)
>>>>>>> +++ lld/trunk/COFF/Writer.cpp Tue Aug 11 18:09:00 2015
>>>>>>> @@ -37,8 +37,7 @@ using namespace lld;
>>>>>>> using namespace lld::coff;
>>>>>>>
>>>>>>> static const int PageSize = 4096;
>>>>>>> -static const int FileAlignment = 512;
>>>>>>> -static const int SectionAlignment = 4096;
>>>>>>> +static const int SectorSize = 512;
>>>>>>> static const int DOSStubSize = 64;
>>>>>>> static const int NumberfOfDataDirectory = 16;
>>>>>>>
>>>>>>> @@ -174,7 +173,7 @@ void OutputSection::addChunk(Chunk *C) {
>>>>>>> Off += C->getSize();
>>>>>>> Header.VirtualSize = Off;
>>>>>>> if (C->hasData())
>>>>>>> - Header.SizeOfRawData = RoundUpToAlignment(Off, FileAlignment);
>>>>>>> + Header.SizeOfRawData = RoundUpToAlignment(Off, SectorSize);
>>>>>>> }
>>>>>>>
>>>>>>> void OutputSection::addPermissions(uint32_t C) {
>>>>>>> @@ -507,15 +506,14 @@ void Writer::createSymbolAndStringTable(
>>>>>>> // We position the symbol table to be adjacent to the end of the
>>>>>>> last section.
>>>>>>> uint64_t FileOff =
>>>>>>> LastSection->getFileOff() +
>>>>>>> - RoundUpToAlignment(LastSection->getRawSize(), FileAlignment);
>>>>>>> + RoundUpToAlignment(LastSection->getRawSize(), SectorSize);
>>>>>>> if (!OutputSymtab.empty()) {
>>>>>>> PointerToSymbolTable = FileOff;
>>>>>>> FileOff += OutputSymtab.size() * sizeof(coff_symbol16);
>>>>>>> }
>>>>>>> if (!Strtab.empty())
>>>>>>> FileOff += Strtab.size() + 4;
>>>>>>> - FileSize = SizeOfHeaders +
>>>>>>> - RoundUpToAlignment(FileOff - SizeOfHeaders,
>>>>>>> FileAlignment);
>>>>>>> + FileSize = RoundUpToAlignment(FileOff, SectorSize);
>>>>>>> }
>>>>>>>
>>>>>>> // Visits all sections to assign incremental, non-overlapping RVAs
>>>>>>> and
>>>>>>> @@ -526,9 +524,9 @@ void Writer::assignAddresses() {
>>>>>>> sizeof(coff_section) * OutputSections.size();
>>>>>>> SizeOfHeaders +=
>>>>>>> Config->is64() ? sizeof(pe32plus_header) :
>>>>>>> sizeof(pe32_header);
>>>>>>> - SizeOfHeaders = RoundUpToAlignment(SizeOfHeaders, PageSize);
>>>>>>> + SizeOfHeaders = RoundUpToAlignment(SizeOfHeaders, SectorSize);
>>>>>>> uint64_t RVA = 0x1000; // The first page is kept unmapped.
>>>>>>> - uint64_t FileOff = SizeOfHeaders;
>>>>>>> + FileSize = SizeOfHeaders;
>>>>>>> // Move DISCARDABLE (or non-memory-mapped) sections to the end of
>>>>>>> file because
>>>>>>> // the loader cannot handle holes.
>>>>>>> std::stable_partition(
>>>>>>> @@ -539,13 +537,11 @@ void Writer::assignAddresses() {
>>>>>>> if (Sec->getName() == ".reloc")
>>>>>>> addBaserels(Sec);
>>>>>>> Sec->setRVA(RVA);
>>>>>>> - Sec->setFileOffset(FileOff);
>>>>>>> + Sec->setFileOffset(FileSize);
>>>>>>> RVA += RoundUpToAlignment(Sec->getVirtualSize(), PageSize);
>>>>>>> - FileOff += RoundUpToAlignment(Sec->getRawSize(), FileAlignment);
>>>>>>> + FileSize += RoundUpToAlignment(Sec->getRawSize(), SectorSize);
>>>>>>> }
>>>>>>> SizeOfImage = SizeOfHeaders + RoundUpToAlignment(RVA - 0x1000,
>>>>>>> PageSize);
>>>>>>> - FileSize = SizeOfHeaders +
>>>>>>> - RoundUpToAlignment(FileOff - SizeOfHeaders,
>>>>>>> FileAlignment);
>>>>>>> }
>>>>>>>
>>>>>>> template <typename PEHeaderTy> void Writer::writeHeader() {
>>>>>>> @@ -584,8 +580,8 @@ template <typename PEHeaderTy> void Writ
>>>>>>> Buf += sizeof(*PE);
>>>>>>> PE->Magic = Config->is64() ? PE32Header::PE32_PLUS :
>>>>>>> PE32Header::PE32;
>>>>>>> PE->ImageBase = Config->ImageBase;
>>>>>>> - PE->SectionAlignment = SectionAlignment;
>>>>>>> - PE->FileAlignment = FileAlignment;
>>>>>>> + PE->SectionAlignment = PageSize;
>>>>>>> + PE->FileAlignment = SectorSize;
>>>>>>> PE->MajorImageVersion = Config->MajorImageVersion;
>>>>>>> PE->MinorImageVersion = Config->MinorImageVersion;
>>>>>>> PE->MajorOperatingSystemVersion = Config->MajorOSVersion;
>>>>>>>
>>>>>>> Modified: lld/trunk/test/COFF/baserel.test
>>>>>>> URL:
>>>>>>> http://llvm.org/viewvc/llvm-project/lld/trunk/test/COFF/baserel.test?rev=244691&r1=244690&r2=244691&view=diff
>>>>>>>
>>>>>>> ==============================================================================
>>>>>>> --- lld/trunk/test/COFF/baserel.test (original)
>>>>>>> +++ lld/trunk/test/COFF/baserel.test Tue Aug 11 18:09:00 2015
>>>>>>> @@ -61,7 +61,7 @@
>>>>>>> # BASEREL-HEADER-NEXT: VirtualSize: 0x20
>>>>>>> # BASEREL-HEADER-NEXT: VirtualAddress: 0x5000
>>>>>>> # BASEREL-HEADER-NEXT: RawDataSize: 512
>>>>>>> -# BASEREL-HEADER-NEXT: PointerToRawData: 0x1800
>>>>>>> +# BASEREL-HEADER-NEXT: PointerToRawData: 0xC00
>>>>>>> # BASEREL-HEADER-NEXT: PointerToRelocations: 0x0
>>>>>>> # BASEREL-HEADER-NEXT: PointerToLineNumbers: 0x0
>>>>>>> # BASEREL-HEADER-NEXT: RelocationCount: 0
>>>>>>>
>>>>>>> Modified: lld/trunk/test/COFF/hello32.test
>>>>>>> URL:
>>>>>>> http://llvm.org/viewvc/llvm-project/lld/trunk/test/COFF/hello32.test?rev=244691&r1=244690&r2=244691&view=diff
>>>>>>>
>>>>>>> ==============================================================================
>>>>>>> --- lld/trunk/test/COFF/hello32.test (original)
>>>>>>> +++ lld/trunk/test/COFF/hello32.test Tue Aug 11 18:09:00 2015
>>>>>>> @@ -38,8 +38,8 @@ HEADER-NEXT: MajorImageVersion: 0
>>>>>>> HEADER-NEXT: MinorImageVersion: 0
>>>>>>> HEADER-NEXT: MajorSubsystemVersion: 6
>>>>>>> HEADER-NEXT: MinorSubsystemVersion: 0
>>>>>>> -HEADER-NEXT: SizeOfImage: 20480
>>>>>>> -HEADER-NEXT: SizeOfHeaders: 4096
>>>>>>> +HEADER-NEXT: SizeOfImage: 16896
>>>>>>> +HEADER-NEXT: SizeOfHeaders: 512
>>>>>>> HEADER-NEXT: Subsystem: IMAGE_SUBSYSTEM_WINDOWS_CUI (0x3)
>>>>>>> HEADER-NEXT: Characteristics [ (0x8140)
>>>>>>> HEADER-NEXT: IMAGE_DLL_CHARACTERISTICS_DYNAMIC_BASE (0x40)
>>>>>>>
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> llvm-commits mailing list
>>>>>>> llvm-commits at lists.llvm.org
>>>>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150813/e015b4b0/attachment.html>
More information about the llvm-commits
mailing list