- 16 Jan, 2019 40 commits
-
-
Matthias Kaehlcke authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit d77698df upstream. For gcc stack alignment is configured with -mpreferred-stack-boundary=N, clang has the option -mstack-alignment=N for that purpose. Use the same alignment as with gcc. If the alignment is not specified clang assumes an alignment of 16 bytes, as required by the standard ABI. However as mentioned in d9b0cde9 ("x86-64, gcc: Use -mpreferred-stack-boundary=3 if supported") the standard kernel entry on x86-64 leaves the stack on an 8-byte boundary, as a consequence clang will keep the stack misaligned. Signed-off-by:
Matthias Kaehlcke <mka@chromium.org> Acked-by:
Ingo Molnar <mingo@kernel.org> Signed-off-by:
Masahiro Yamada <yamada.masahiro@socionext.com> Signed-off-by:
Nathan Chancellor <natechancellor@gmail.com> Signed-off-by:
Greg Kroah-Hartman <gregkh@linuxfoundation.org> Signed-off-by:
Juerg Haefliger <juergh@canonical.com> Signed-off-by:
Kleber Sacilotto de Souza <kleber.souza@canonical.com>
-
Matthias Kaehlcke authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 032a2c4f upstream. cc-option is used to enable compiler options for the boot code if they are available. The macro uses KBUILD_CFLAGS and KBUILD_CPPFLAGS for the check, however these flags aren't used to build the boot code, in consequence cc-option can yield wrong results. For example -mpreferred-stack-boundary=2 is never set with a 64-bit compiler, since the setting is only valid for 16 and 32-bit binaries. This is also the case for 32-bit kernel builds, because the option -m32 is added to KBUILD_CFLAGS after the assignment of REALMODE_CFLAGS. Use __cc-option instead of cc-option for the boot mode options. The macro receives the compiler options as parameter instead of using KBUILD_C*FLAGS, for the boot code we pass REALMODE_CFLAGS. Also use separate statements for the __cc-option checks instead of performing them in the initial assignment of REALMODE_CFLAGS since the variable is an input of the macro. Signed-off-by:
-
Matthias Kaehlcke authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 9f3f1fd2 upstream. cc-option uses KBUILD_CFLAGS and KBUILD_CPPFLAGS when it determines whether an option is supported or not. This is fine for options used to build the kernel itself, however some components like the x86 boot code use a different set of flags. Add the new macro __cc-option which is a more generic version of cc-option with additional parameters. One parameter is the compiler with which the check should be performed, the other the compiler options to be used instead KBUILD_C*FLAGS. Refactor cc-option and hostcc-option to use __cc-option and move hostcc-option to scripts/Kbuild.include. Suggested-by:
Arnd Bergmann <arnd@arndb.de> Suggested-by:
Michal Marek <mmarek@suse.com> Signed-off-by:
-
Matthias Kaehlcke authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 121843eb upstream. The constraint "rm" allows the compiler to put mix_const into memory. When the input operand is a memory location then MUL needs an operand size suffix, since Clang can't infer the multiplication width from the operand. Add and use the _ASM_MUL macro which determines the operand size and resolves to the NUL instruction with the corresponding suffix. This fixes the following error when building with clang: CC arch/x86/lib/kaslr.o /tmp/kaslr-dfe1ad.s: Assembler messages: /tmp/kaslr-dfe1ad.s:182: Error: no instruction mnemonic suffix given and no register operands; can't size instruction Signed-off-by:
-
Michael Davidson authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit fdb2726f upstream. aes_ctrby8_avx-x86_64.S uses the C preprocessor for token pasting of character sequences that are not valid preprocessor tokens. While this is allowed when preprocessing assembler files it exposes an incompatibilty between the clang and gcc preprocessors where clang does not strip leading white space from macro parameters, leading to the CONCAT(%xmm, i) macro expansion on line 96 resulting in a token with a space character embedded in it. While this could be resolved by deleting the offending space character, the assembler is perfectly capable of doing the token pasting correctly for itself so we can just get rid of the preprocessor macros. Signed-off-by:
Michael Davidson <md@google.com> Signed-off-by:
Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by:
-
Matthias Kaehlcke authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 2c4fd1ac upstream. clang currently does not support these optimizations, only enable them when they are available. Signed-off-by:
-
Matthias Kaehlcke authored
BugLink: https://bugs.launchpad.net/bugs/1810958 clang raises 'asm-operand-widths' warnings in inline assembly code when the size of an operand is < 64 bits and the operand width is unspecified. Most warnings are raised in macros, i.e. the datatype of the operand may vary. Signed-off-by:
-
Stefan Agner authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit ef8c4ed9 upstream. When using a GCC cross toolchain which is not in a compiled in Clang search path, Clang reverts to the system assembler and linker. This leads to assembler or linker errors, depending on which tool is first used for a given architecture. It seems that Clang is not searching $PATH for a matching assembler or linker. Make sure that Clang picks up the correct assembler or linker by passing the cross compilers bin directory as search path. This allows to use Clang provided by distributions with GCC toolchains not in /usr/bin. Link: https://github.com/ClangBuiltLinux/linux/issues/78Signed-off-by:
Stefan Agner <stefan@agner.ch> Reviewed-and-tested-by:
Nick Desaulniers <ndesaulniers@google.com> Signed-off-by:
-
Stefan Agner authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 0f0e8de3 upstream. In order to make sure compiler flag detection for ARM works correctly the no-integrated-as flags need to be set before including the arch specific Makefile. Fixes: cfe17c9b ("kbuild: move cc-option and cc-disable-warning after incl. arch Makefile") Signed-off-by:
-
Sodagudi Prasad authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 0a5f4176 upstream. Currently, GCC disables -Wunused-const-variable, but not -Wunused-variable, so warns unused variables if they are non-constant. While, Clang does not warn unused variables at all regardless of the const qualifier because -Wno-unused-const-variable is implied by the stronger option -Wno-unused-variable. Disable -Wunused-const-variable instead of -Wunused-variable so that GCC and Clang work in the same way. Signed-off-by:
Prasad Sodagudi <psodagud@codeaurora.org> Signed-off-by:
-
Nick Desaulniers authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit df16aaac upstream. When compiling with `make CC=clang HOSTCC=clang`, I was seeing warnings that clang did not recognize -fno-delete-null-pointer-checks for HOSTCC targets. These were added in commit 61163efa ("kbuild: LLVMLinux: Add Kbuild support for building kernel with Clang"). Clang does not support -fno-delete-null-pointer-checks, so adding it to HOSTCFLAGS if HOSTCC is clang does not make sense. It's not clear why the other warnings were disabled, and just for HOSTCFLAGS, but I can remove them, add -Werror to HOSTCFLAGS and compile with clang just fine. Suggested-by:
Nick Desaulniers <nick.desaulniers@gmail.com> Signed-off-by:
-
David Lin authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit bb3f38c3 upstream. We should avoid using the space character when passing arguments to clang, because static code analysis check tool such as sparse may misinterpret the arguments followed by spaces as build targets hence cause the build to fail. Signed-off-by:
David Lin <dtwlin@google.com> Signed-off-by:
-
Masahiro Yamada authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit cfe17c9b upstream. Geert reported commit ae6b289a ("kbuild: Set KBUILD_CFLAGS before incl. arch Makefile") broke cross-compilation using a cross-compiler that supports less compiler options than the host compiler. For example, cc1: error: unrecognized command line option "-Wno-unused-but-set-variable" This problem happens on architectures that setup CROSS_COMPILE in their arch/*/Makefile. Move the cc-option and cc-disable-warning back to the original position, but keep the Clang target options untouched. Fixes: ae6b289a ("kbuild: Set KBUILD_CFLAGS before incl. arch Makefile") Reported-by:
Geert Uytterhoeven <geert@linux-m68k.org> Signed-off-by:
-
Chris Fries authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit ae6b289a upstream. Set the clang KBUILD_CFLAGS up before including arch/ Makefiles, so that ld-options (etc.) can work correctly. This fixes errors with clang such as ld-options trying to CC against your host architecture, but LD trying to link against your target architecture. Signed-off-by:
Chris Fries <cfries@google.com> Signed-off-by:
-
Nick Desaulniers authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 86a9df59 upstream. I was not seeing my linker flags getting added when using ld-option when cross compiling with Clang. Upon investigation, this seems to be due to a difference in how GCC vs Clang handle cross compilation. GCC is configured at build time to support one backend, that is implicit when compiling. Clang is explicit via the use of `-target <triple>` and ships with all supported backends by default. GNU Make feature test macros that compile then link will always fail when cross compiling with Clang unless Clang's triple is passed along to the compiler. For example: $ clang -x c /dev/null -c -o temp.o $ aarch64-linux-android/bin/ld -E temp.o aarch64-linux-android/bin/ld: unknown architecture of input file `temp.o' is incompatible with aarch64 output aarch64-linux-android/bin/ld: warning: cannot find entry symbol _start; defaulting to 0000000000400078 $ echo $? 1 $ clang -target aarch64-linux-android- -x c /dev/null -c -o temp.o $ aarch64-linux-android/bin/ld -E temp.o aarch64-linux-android/bin/ld: warning: cannot find entry symbol _start; defaulting to 00000000004002e4 $ echo $? 0 This causes conditional checks that invoke $(CC) without the target triple, then $(LD) on the result, to always fail. Suggested-by:
-
Ard Biesheuvel authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 91ee5b21 upstream. Clang may emit absolute symbol references when building in non-PIC mode, even when using the default 'small' code model, which is already mostly position independent to begin with, due to its use of adrp/add pairs that have a relative range of +/- 4 GB. The remedy is to pass the -fpie flag, which can be done safely now that the code has been updated to avoid GOT indirections (which may be emitted due to the compiler assuming that the PIC/PIE code may end up in a shared library that is subject to ELF symbol preemption) Passing -fpie when building code that needs to execute at an a priori unknown offset is arguably an improvement in any case, and given that the recent visibility changes allow the PIC build to pass with GCC as well, let's add -fpie for all arm64 builds rather than only for Clang. Tested-by:
Ard Biesheuvel <ard.biesheuvel@linaro.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-efi@vger.kernel.org Link: http://lkml.kernel.org/r/20170818194947.19347-5-ard.biesheuvel@linaro.orgSigned-off-by:
-
Ard Biesheuvel authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 0426a4e6 upstream. To prevent the compiler from emitting absolute references to the section markers when running in PIC mode, override the visibility to 'hidden' for all contents of asm/sections.h Tested-by:
-
Ard Biesheuvel authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit f4857f4c upstream. Replace the inline asm which exports struct offsets as ELF symbols with proper const variables exposing the same values. This works around an issue with Clang which does not interpret the "i" (or "I") constraints in the same way as GCC. Signed-off-by:
-
Matthias Kaehlcke authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit bfb38988 upstream. clang generates plenty of these warnings in different parts of the code, to an extent that the warnings are little more than noise. Disable the 'address-of-packed-member' warning. Signed-off-by:
Douglas Anderson <dianders@chromium.org> Signed-off-by:
-
Arnd Bergmann authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 1f318a8b upstream. clang warns about unused inline functions by default: arch/arm/crypto/aes-cipher-glue.c:68:1: warning: unused function '__inittest' [-Wunused-function] arch/arm/crypto/aes-cipher-glue.c:69:1: warning: unused function '__exittest' [-Wunused-function] As these appear in every single module, let's just disable the warnings by marking the two functions as __maybe_unused. Signed-off-by:
Miroslav Benes <mbenes@suse.cz> Acked-by:
Rusty Russell <rusty@rustcorp.com.au> Signed-off-by:
Jessica Yu <jeyu@redhat.com> Signed-off-by:
-
Vinícius Tinti authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 433db3e2 upstream. Add rules to kbuild in order to generate LLVM assembly files with the .ll extension when using clang. # from c code make CC=clang kernel/pid.ll Signed-off-by:
Vinícius Tinti <viniciustinti@gmail.com> Signed-off-by:
Behan Webster <behanw@converseincode.com> Signed-off-by:
-
Behan Webster authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 6748cb3c upstream. This generates smaller resulting object code when compiled with clang. Signed-off-by:
-
Mark Charlebois authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit c3f0d0bc upstream. Clang will warn about unknown warnings but will not return false unless -Werror is set. GCC will return false if an unknown warning is passed. Adding -Werror make both compiler behave the same. [arnd: it turns out we need the same patch for testing whether -ffunction-sections works right with gcc. I've build tested extensively with this patch applied, so let's just merge this one now.] Signed-off-by:
Mark Charlebois <charlebm@gmail.com> Signed-off-by:
Jan-Simon Möller <dl9pf@gmx.de> Signed-off-by:
Kees Cook <keescook@chromium.org> Signed-off-by:
-
Masahiro Yamada authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit a0ae981e upstream. Since commit c3f0d0bc ("kbuild, LLVMLinux: Add -Werror to cc-option to support clang"), cc-option and friends work nicely for clang. However, -Wno-unknown-warning-option makes clang happy with any unknown warning options even if -Werror is specified. Once -Wno-unknown-warning-option is added, any succeeding call of cc-disable-warning is evaluated positive, then unknown warning options are accepted. This should be dropped. Signed-off-by:
-
Jeroen Hofstee authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit cf0c3e68 upstream. KBuild abuses the asm statement to write to a file and clang chokes about these invalid asm statements. Hack it even more by fooling this is actual valid asm code. [masahiro: Import Jeroen's work for U-Boot: http://patchwork.ozlabs.org/patch/375026/ Tweak sed script a little to avoid garbage '#' for GCC case, like #define NR_PAGEFLAGS 23 /* __NR_PAGEFLAGS # */ ] Signed-off-by:
Jeroen Hofstee <jeroen@myspectrum.nl> Signed-off-by:
-
Masahiro Yamada authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 7dd47b95 upstream. This part ended up in redundant code after touched by multiple people. [1] Commit 3234282f ("x86, asm: Fix CFI macro invocations to deal with shortcomings in gas") added parentheses for defined expressions to support old gas for x86. [2] Commit a22dcdb0 ("x86, asm: Fix ancient-GAS workaround") split the pattern into two to avoid parentheses for non-numeric expressions. [3] Commit 95a2f6f7 ("Partially revert patch that encloses asm-offset.h numbers in brackets") removed parentheses from numeric expressions as well because parentheses in MN10300 assembly have a special meaning (pointer access). Apparently, there is a conflict between [1] and [3]. After all, [3] took precedence, and a long time has passed since then. Now, merge the two patterns again because the first one is covered by the other. Signed-off-by:
-
Matthias Kaehlcke authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit ebf003f0 upstream. Largely redundant code is used in different places to generate C headers from offset information extracted from assembly language output. Consolidate the code in Makefile.lib and use this instead. Signed-off-by:
-
Michael Davidson authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit a37c45cd upstream. The Linux Kernel relies on GCC's acceptance of inline assembly as an opaque object which will not have any validation performed on the content. The current behaviour in LLVM is to perform validation of the contents by means of parsing the input if the MC layer can handle it. Disable clangs integrated assembler and use the GNU assembler instead. Wording-mostly-from: Saleem Abdulrasool <compnerd@compnerd.org> Signed-off-by:
-
Behan Webster authored
BugLink: https://bugs.launchpad.net/bugs/1810958 commit 785f11aa upstream. Add cross target to CC if using clang. Also add custom gcc toolchain path for fallback gcc tools. Clang will fallback to using things like ld, as, and libgcc if (respectively) one of the llvm linkers isn't available, the integrated assembler is turned off, or an appropriately cross-compiled version of compiler-rt isn't available. To this end, you can specify the path to this fallback gcc toolchain with GCC_TOOLCHAIN. Signed-off-by:
Greg Hackmann <ghackmann@google.com> Signed-off-by:
-
David Ahern authored
BugLink: https://bugs.launchpad.net/bugs/1810958 [ Upstream commit 7ddacfa5 ] Preethi reported that PMTU discovery for UDP/raw applications is not working in the presence of VRF when the socket is not bound to a device. The problem is that ip6_sk_update_pmtu does not consider the L3 domain of the skb device if the socket is not bound. Update the function to set oif to the L3 master device if relevant. Fixes: ca254490 ("net: Add VRF support to IPv6 stack") Reported-by:
Preethi Ramachandra <preethir@juniper.net> Signed-off-by:
David Ahern <dsahern@gmail.com> Signed-off-by:
David S. Miller <davem@davemloft.net> Signed-off-by:
-
Siva Reddy Kallam authored
BugLink: https://bugs.launchpad.net/bugs/1810958 [ Upstream commit 59663e42 ] This patch has the fix to avoid PHY lockup with 5717/5719/5720 in change ring and flow control paths. This patch solves the RX hang while doing continuous ring or flow control parameters with heavy traffic from peer. Signed-off-by:
Siva Reddy Kallam <siva.kallam@broadcom.com> Acked-by:
Michael Chan <michael.chan@broadcom.com> Signed-off-by:
-
Eric Dumazet authored
BugLink: https://bugs.launchpad.net/bugs/1810958 [ Upstream commit 33d9a2c7 ] eth_type_trans() assumes initial value for skb->pkt_type is PACKET_HOST. This is indeed the value right after a fresh skb allocation. However, it is possible that GRO merged a packet with a different value (like PACKET_OTHERHOST in case macvlan is used), so we need to make sure napi->skb will have pkt_type set back to PACKET_HOST. Otherwise, valid packets might be dropped by the stack because their pkt_type is not PACKET_HOST. napi_reuse_skb() was added in commit 96e93eab ("gro: Add internal interfaces for VLAN"), but this bug always has been there. Fixes: 96e93eab ("gro: Add internal interfaces for VLAN") Signed-off-by:
Eric Dumazet <edumazet@google.com> Signed-off-by:
-
Sabrina Dubroca authored
BugLink: https://bugs.launchpad.net/bugs/1810958 [ Upstream commit 16f7eb2b ] The various types of tunnels running over IPv4 can ask to set the DF bit to do PMTU discovery. However, PMTU discovery is subject to the threshold set by the net.ipv4.route.min_pmtu sysctl, and is also disabled on routes with "mtu lock". In those cases, we shouldn't set the DF bit. This patch makes setting the DF bit conditional on the route's MTU locking state. This issue seems to be older than git history. Signed-off-by:
Sabrina Dubroca <sd@queasysnail.net> Reviewed-by:
Stefano Brivio <sbrivio@redhat.com> Signed-off-by:
-
배석진 authored
BugLink: https://bugs.launchpad.net/bugs/1810958 [ Upstream commit 62230715 ] Only first fragment has the sport/dport information, not the following ones. If we want consistent hash for all fragments, we need to ignore ports even for first fragment. This bug is visible for IPv6 traffic, if incoming fragments do not have a flow label, since skb_get_hash() will give different results for first fragment and following ones. It is also visible if any routing rule wants dissection and sport or dport. See commit 5e5d6fed ("ipv6: route: dissect flow in input path if fib rules need it") for details. [edumazet] rewrote the changelog completely. Fixes: 06635a35 ("flow_dissect: use programable dissector in skb_flow_dissect and friends") Signed-off-by:
배석진 <soukjin.bae@samsung.com> Signed-off-by:
-
Greg Kroah-Hartman authored
BugLink: https://bugs.launchpad.net/bugs/1810947Signed-off-by:
-
Clint Taylor authored
BugLink: https://bugs.launchpad.net/bugs/1810947 commit 65034931 upstream. HDMI 2.0 594Mhz modes were incorrectly selecting 25.200Mhz Automatic N value mode instead of HDMI specification values. V2: Fix 88.2 Hz N value Cc: Jani Nikula <jani.nikula@linux.intel.com> Cc: stable@vger.kernel.org Signed-off-by:
Clint Taylor <clinton.a.taylor@intel.com> Signed-off-by:
Jani Nikula <jani.nikula@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/1540493521-1746-2-git-send-email-clinton.a.taylor@intel.com (cherry picked from commit 5a400aa3) Signed-off-by:
Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Signed-off-by:
-
Stanislav Lisovskiy authored
BugLink: https://bugs.launchpad.net/bugs/1810947 commit 23d80039 upstream. Unfortunately drm_dp_get_mst_branch_device which is called from both drm_dp_mst_handle_down_rep and drm_dp_mst_handle_up_rep seem to rely on that mgr->mst_primary is not NULL, which seem to be wrong as it can be cleared with simultaneous mode set, if probing fails or in other case. mgr->lock mutex doesn't protect against that as it might just get assigned to NULL right before, not simultaneously. There are currently bugs 107738, 108616 bugs which crash in drm_dp_get_mst_branch_device, caused by this issue. v2: Refactored the code, as it was nicely noticed. Fixed Bugzilla bug numbers(second was 108616, but not 108816) and added links. [changed title and added stable cc] Signed-off-by:
Lyude Paul <lyude@redhat.com> Signed-off-by:
Stanislav Lisovskiy <stanislav.lisovskiy@intel.com> Cc: stable@vger.kernel.org Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=108616 Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=107738 Link: https://patchwork.freedesktop.org/patch/msgid/20181109090012.24438-1-stanislav.lisovskiy@intel.comSigned-off-by:
-
Marc Zyngier authored
BugLink: https://bugs.launchpad.net/bugs/1810947 commit 7f3ef5de upstream. Leaving the DRM driver enabled on reboot or kexec has the annoying effect of leaving the display generating transactions whilst the IOMMU has been shut down. In turn, the IOMMU driver (which shares its interrupt line with the VOP) starts warning either on shutdown or when entering the secondary kernel in the kexec case (nothing is expected on that front). A cheap way of ensuring that things are nicely shut down is to register a shutdown callback in the platform driver. Signed-off-by:
Marc Zyngier <marc.zyngier@arm.com> Tested-by:
Vicente Bergas <vicencb@gmail.com> Signed-off-by:
Heiko Stuebner <heiko@sntech.de> Link: https://patchwork.freedesktop.org/patch/msgid/20180805124807.18169-1-marc.zyngier@arm.com Cc: stable@vger.kernel.org Signed-off-by:
-
Mike Kravetz authored
BugLink: https://bugs.launchpad.net/bugs/1810947 commit 017b1660 upstream. The page migration code employs try_to_unmap() to try and unmap the source page. This is accomplished by using rmap_walk to find all vmas where the page is mapped. This search stops when page mapcount is zero. For shared PMD huge pages, the page map count is always 1 no matter the number of mappings. Shared mappings are tracked via the reference count of the PMD page. Therefore, try_to_unmap stops prematurely and does not completely unmap all mappings of the source page. This problem can result is data corruption as writes to the original source page can happen after contents of the page are copied to the target page. Hence, data is lost. This problem was originally seen as DB corruption of shared global areas after a huge page was soft offlined due to ECC memory errors. DB developers noticed they could reproduce the issue by (hotplug) offlining memory used to back huge pages. A simple testcase can reproduce the problem by creating a shared PMD mapping (note that this must be at least PUD_SIZE in size and PUD_SIZE aligned (1GB on x86)), and using migrate_pages() to migrate process pages between nodes while continually writing to the huge pages being migrated. To fix, have the try_to_unmap_one routine check for huge PMD sharing by calling huge_pmd_unshare for hugetlbfs huge pages. If it is a shared mapping it will be 'unshared' which removes the page table entry and drops the reference on the PMD page. After this, flush caches and TLB. mmu notifiers are called before locking page tables, but we can not be sure of PMD sharing until page tables are locked. Therefore, check for the possibility of PMD sharing before locking so that notifiers can prepare for the worst possible case. Link: http://lkml.kernel.org/r/20180823205917.16297-2-mike.kravetz@oracle.com [mike.kravetz@oracle.com: make _range_in_vma() a static inline] Link: http://lkml.kernel.org/r/6063f215-a5c8-2f0c-465a-2c515ddc952d@oracle.com Fixes: 39dde65c ("shared page table for hugetlb page") Signed-off-by:
Mike Kravetz <mike.kravetz@oracle.com> Acked-by:
Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Reviewed-by:
Naoya Horiguchi <n-horiguchi@ah.jp.nec.com> Acked-by:
Michal Hocko <mhocko@suse.com> Cc: Vlastimil Babka <vbabka@suse.cz> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: Jerome Glisse <jglisse@redhat.com> Cc: Mike Kravetz <mike.kravetz@oracle.com> Cc: <stable@vger.kernel.org> Signed-off-by:
Andrew Morton <akpm@linux-foundation.org> Signed-off-by:
Jérôme Glisse <jglisse@redhat.com> Signed-off-by:
-
Mike Kravetz authored
BugLink: https://bugs.launchpad.net/bugs/1810947 commit 5e41540c upstream. This bug has been experienced several times by the Oracle DB team. The BUG is in remove_inode_hugepages() as follows: /* * If page is mapped, it was faulted in after being * unmapped in caller. Unmap (again) now after taking * the fault mutex. The mutex will prevent faults * until we finish removing the page. * * This race can only happen in the hole punch case. * Getting here in a truncate operation is a bug. */ if (unlikely(page_mapped(page))) { BUG_ON(truncate_op); In this case, the elevated map count is not the result of a race. Rather it was incorrectly incremented as the result of a bug in the huge pmd sharing code. Consider the following: - Process A maps a hugetlbfs file of sufficient size and alignment (PUD_SIZE) that a pmd page could be shared. - Process B maps the same hugetlbfs file with the same size and alignment such that a pmd page is shared. - Process B then calls mprotect() to change protections for the mapping with the shared pmd. As a result, the pmd is 'unshared'. - Process B then calls mprotect() again to chage protections for the mapping back to their original value. pmd remains unshared. - Process B then forks and process C is created. During the fork process, we do dup_mm -> dup_mmap -> copy_page_range to copy page tables. Copying page tables for hugetlb mappings is done in the routine copy_hugetlb_page_range. In copy_hugetlb_page_range(), the destination pte is obtained by: dst_pte = huge_pte_alloc(dst, addr, sz); If pmd sharing is possible, the returned pointer will be to a pte in an existing page table. In the situation above, process C could share with either process A or process B. Since process A is first in the list, the returned pte is a pointer to a pte in process A's page table. However, the check for pmd sharing in copy_hugetlb_page_range is: /* If the pagetables are shared don't copy or take references */ if (dst_pte == src_pte) continue; Since process C is sharing with process A instead of process B, the above test fails. The code in copy_hugetlb_page_range which follows assumes dst_pte points to a huge_pte_none pte. It copies the pte entry from src_pte to dst_pte and increments this map count of the associated page. This is how we end up with an elevated map count. To solve, check the dst_pte entry for huge_pte_none. If !none, this implies PMD sharing so do not copy. Link: http://lkml.kernel.org/r/20181105212315.14125-1-mike.kravetz@oracle.com Fixes: c5c99429 ("fix hugepages leak due to pagetable page sharing") Signed-off-by:
Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by:
-
