| Patch | Description | Author | Forwarded | Bugs | Origin | Last update |
|---|---|---|---|---|---|---|
| backports/bugfixes.diff | backport a number of general bugfixes from upstream some bounds checking, out of bounds writes, nil pointer dereference in error path, uninitialised variable, plus a few warning fixes for pointer casts and fallthroughs |
mirabilos <tg@debian.org> | no | |||
| backports/ia64-fix-ftbfs.diff | fix IA64 build by backporting enough from upstream | mirabilos <tg@debian.org> | no | |||
| workarounds/sparc64-tls-errno.diff | sparc64: Work around compiler bug Work around sparc64 gcc6 bug that mis-optimizes TLS variable access to errno in certain situations. . When errno is accessed multiple times in a function, the relative offset of the variable to the thread pointer may be stored on the stack in some cases (because loading it from the stack takes only one instruction, whereas loading it from scratch takes 2-4). Unfortunately, in some cases the compiler will only generate 32 bit loads/stores (and not 64 bit), and while the offset is usually much smaller than 2^31, it is negative, and the load instruction the compiler generates does not sign-extend. This causes the compiler to load (tls-base + 2^31 - offset) instead of (tls-base - offset), resulting in a segfault. . The precise bug has not been tracked down so far, because in most cases the compiler will either re-use another register or indeed generate 64 bit loads/stores, making debugging the issue quite tricky. . This patch replaces errno with a define to (*__errno_location()), which circumvents this issue. But as this is slightly slower and typically produces larger code, only enable it on sparc64. . As soon as the compiler bug is tracked down further, this workaround should be removed. Shouldn't be forwarded, as it only works around the errno case, but not other cases of TLS variables. |
Christian Seiler <christian@iwakd.de> | no | 2016-05-29 | ||
| workarounds/tcb-on-stack-atexit.diff | Work around problem of atexit() functions accessing the TCB atexit() functions might want to access the TCB (such as libpthread's own internal atexit function), and they might be called from the startup code after the end of stackgap(), on whose stack the TCB is allocated. . This is a temporary workaround to make this work for most use cases, but this needs to be followed up with a proper fix. |
Christian Seiler <christian@iwakd.de> | no | 2017-01-05 | ||
| bugfixes/missing-socket-syscalls.diff | Add missing socket-related syscalls on various platforms Various platforms were missng socket-related syscalls (e.g. send(), recv()); make sure they are available. . - define accept4() in sys/socket.h - generic: define ppoll() syscall, been available since 2.6.16 (a decade ago) - aarch64, mips64: re-use x86_64's recv/send implementation that defers to recvfrom()/sendto() - parisc: fix weak syscalls (wrong order in macro definition), declare sendto in __sendto.S not sendto.S to properly override the generic version (need syscall6 here) - arm, ia64: define __NR_ppoll - add generic __accept4 implementation via socketcall if that's what the architecture supports (currently s390[x] only) |
Christian Seiler <christian@iwakd.de> | invalid | 2018-10-06 | ||
| bugfixes/pthread-dont-override-errno-location.diff | Don't override errno_location in libpthread if TLS is enabled If TLS is enabled, errno is already thread-safe. Don't override errno_location in that case, otherwise the errno external code sees is not the same as the errno syscalls see (that call __errno_location directly). |
Christian Seiler <christian@iwakd.de> | no | debian | 2017-01-05 | |
| bugfixes/aarch64-define-__clone.diff | aarch64: define __clone, make clone weak libpthread requires an internal __clone to be defined and clone to be a weak alias against it. Otherwise linking against -lpthread would fail in many cases. |
Christian Seiler <christian@iwakd.de> | no | 2017-01-05 | ||
| bugfixes/x32-define-munmap-and-exit.diff | Build __munmap_and_exit on x32 __munmap_and_exit is required by libpthread, so build it. (The x86_64 variant thereof is compatible with x32, so just reference it from Makefile.add.) |
Christian Seiler <christian@iwakd.de> | no | 2017-01-05 | ||
| bugfixes/libc_waitpid.diff | Fix waitpid with -lpthread on s390{,x}, mips64, ia64 On these four architectures waitpid() was defined as a strong function in libc directly, instead of having a weak alias to __libc_waitpid. Due to the unavailability of __libc_waitpid for this reason, linking against -lpthread was not possible in many cases. Follow the scheme for waitpid that is used on other architectures to make linking against -lpthread work again. |
Christian Seiler <christian@iwakd.de> | no | debian | 2017-01-05 | |
| bugfixes/s390x-testandset.diff | Fix __testandset on s390x. __testandset was broken on s390x, because it assumed 32bit as well (the code was identical to s390) - this would cause lockups on spin locks. |
Christian Seiler <christian@iwakd.de> | no | 2017-01-05 | ||
| bugfixes/parisc-getsockopt.diff | Fix getsockopt() on hppa PA-RISC overrides getsockopt() because it requires the last parameter to be passed on the stack to the function, and that needs to be translated to the kernel system call - and a recent renaming of getsockopt.S to __getsockopt.S (including a change to a weak symbol) was not done for PA-RISC, causing getsockopt() to always fail with "Bad Address". |
Christian Seiler <christian@iwakd.de> | no | 2017-01-25 | ||
| bugfixes/parisc-atfork.diff | Properly initialize pthread_atfork's spinlock On PA-RISC (due to some quirk of the instruction set) an unlocked spinlock is initialized with a 1, not a 0, as on other platforms. But pthread_atfork's spinlock wasn't properly initialized with PTHREAD_SPIN_UNLOCKED, instead it was just zero'd out, causing the code to loop endlessly on fork() whenever -lpthread was linked in. |
Christian Seiler <christian@iwakd.de> | no | 2017-01-25 | ||
| bugfixes/thread-self-vs-tcb.diff | Fix libpthread's __thread_self on most non-x86 platforms The thread descriptor is (in general) not to be found at the thread pointer register (that's where the TCB is). The code for x86 (both 32 and 64 bit) does this correctly because it directly reads the ->self data member (via a segment register based memory access). However, the other platforms would just return the thread pointer directly - and hence the wrong data structure. . Since there are open questions w.r.t. the TCB layout on msot platforms fall back to the less efficient (but working) global search on non-x86 for now. |
Christian Seiler <christian@iwakd.de> | no | 2017-01-06 | ||
| bugfixes/alpha-testandset.diff | Fix return value of __testandset on alpha __testandset would always return 1, irrespective of whether the lock was previously set or not, causing an endless loop even with an unlocked spinlock. Fix this by returning 0 in the case where the lock could be taken. |
Christian Seiler <christian@iwakd.de> | no | 2017-01-25 | ||
| bugfixes/arm64-testandset.diff | Fix register reuse in __testandset on arm64 In the case where the stlxr instruction fails on arm64 (either due to another thread having accesed that memory location or spuriously) the register x2 (which stores the pointer to the memory location) is now useless because the original implementation reused it (which it should not have). . Just use the unused w3 register for the status result instead. |
Christian Seiler <christian@iwakd.de> | no | debian | 2017-02-18 | |
| bugfixes/alpha-setjmp.diff | alpha: fix broken setjmp() setjmp() on alpha was broken (due to the way the GP was handled). This patch adds a proper prologue to all setjmp variants, and ensures that the jump to __sigjmp_save includes the information that the GP hasn't changed, so __sigjmp_save doesn't erroneously try to reload it in its own prologue, reading a wrong value and crashing. |
Christian Seiler <christian@iwakd.de> | no | 2017-01-25 | ||
| bugfixes/offsetof-typo.diff | # DP: fix typo preventing the compiler builtin from being used | no | ||||
| bugfixes/newer-linux-headers.diff | Support newer Linux headers Copy some definitions from /usr/include/linux/types.h to make applications using them not FTBFS. |
mirabilos <tg@debian.org> | no | debian | ||
| bugfixes/mips-tls.diff | add MIPS-specific stub prototype for build log scanner | mirabilos <tg@debian.org> | no | |||
| bugfixes/use-cflags-also-for-dummy.diff | never compile any file without CPPFLAGS/CFLAGS discovered by build log scanner | mirabilos <tg@debian.org> | no | |||
| bugfixes/utime-test.diff | fix atime comparisons in utime test When I look at the source code in test/utime.c, that test (and the following one) looks backwards. Computers are fast enough that the usual case is st.st_atime == now. |
Larry Doolittle <larry@doolittle.boa.org> | invalid | |||
| bugfixes/regex-stack-overflow.diff | fix regex stack overflow this only caused an FTBFS due to a testsuite failure on alpha but overwrote the stack on any architecture |
mirabilos <tg@debian.org> | no | |||
| bugfixes/ia64-fix-weaks.diff | fix weak symbols on IA64 | mirabilos <tg@debian.org> | no | |||
| bugfixes/testandset-size-and-alignment.diff | change anything passed to __testandset to at least long some architectures need at least 64 bit (alpha) space, and some need 64 bit alignent (IA64) even if 32 bit space might be enough; make it consistent (except hppa) with the testsuite’s use of a long |
mirabilos <tg@debian.org> | no | |||
| bugfixes/cflags.diff | Do not override main CFLAGS for amd64 or x32 Fixes, at least, stack smashing on x32 due to missing EXTRACFLAGS. . We need the GCC include path restored (main CFLAGS contain -nostdinc) for building a file that uses GCC vector math intrinsics, though. |
mirabilos <tg@debian.org> | invalid | debian | 2019-10-24 | |
| bugfixes/x32-struct-member-sizes.diff | Fix some struct member sizes on x32 These need to be of the same size as amd64 uses. . Also fix signedness of useconds_t. |
mirabilos <tg@debian.org> | invalid | debian | 2019-10-24 | |
| bugfixes/hppa-socketcalls.diff | Fix fifth and sixth syscall argument in setsockopt Also uses implicit space register selection in generic syscall code. | John David Anglin <dave.anglin@bell.net> | invalid | debian | 2019-11-10 | |
| features/unshare.diff | add unshare syscall | Larry Doolittle <larry@doolittle.boa.org> | no | |||
| debian/sparc32-mcpu-v9.diff | sparc: use -mcpu=v9 instead of -mcpu=supersparc This fixes a build failure of libowfat on sparc: . diet -v -Os gcc -g -o t t.o libowfat.a `cat libsocket` -lpthread gcc -nostdlib -static -L/usr/lib/diet/lib-sparc /usr/lib/diet/lib-sparc/start.o -g -o t t.o libowfat.a -lpthread -isystem /usr/lib/diet/include -D__dietlibc__ -Os -mcpu=supersparc /usr/lib/diet/lib-sparc/libc.a -lgcc /usr/lib/diet/lib-sparc/libc.a /usr/lib/diet/lib-sparc/libc.a(vprintf.o): In function `vprintf': vprintf.c:(.text+0x1c): warning: warning: the printf functions add several kilobytes of bloat. t.o: In function `main': t.c:(.text+0x30): undefined reference to `__sync_add_and_fetch_4' t.c:(.text+0x48): undefined reference to `__sync_add_and_fetch_4' t.c:(.text+0x64): undefined reference to `__sync_bool_compare_and_swap_4' t.c:(.text+0x80): undefined reference to `__sync_bool_compare_and_swap_4' collect2: ld returned 1 exit status . According to http://www.debian.org/releases/stable/sparc/ch02s01.html.en supersparc isn't supported anymore anyway, and multisparc and niagara are v9. |
Gerrit Pape <pape@smarden.org> | no | 2010-04-20 | ||
| debian/native-x32.diff | Support native x32 builds | Christian Seiler <christian@iwakd.de> | invalid | |||
| debian/buildflags.diff | allow useful passing-in of external CFLAGS Make it possible to pass in external CFLAGS to the compiler. This way we can use dpkg-buildflags in debian/rules. Maybe we should forward this? |
Christian Seiler <christian@iwakd.de> | no | 2016-10-31 | ||
| debian/multiarch.diff | Add Debian multiarch to diet wrapper In order to properly support Debian-style multiarch, make the diet wrapper use /usr/lib/<triplet>/diet instead of /usr/lib/diet as its default diethome, while still detecting cross compilers. This is far easier than patching dietlibc to use different short architecture names for the cases where the same name is used for multiple Debian archs (powerpc/powerpcspe: ppc, armel/armhf: arm, etc.), all the while still retaining some degree of compatibility with upstream. . If no cross compiler is detected, the native architecture is used, if an unknown cross compiler is detected, /usr/lib/diet is used as the default fallback. Setting the DIETHOME environment variable can override this, of course. |
Christian Seiler <christian@iwakd.de> | invalid | 2016-05-02 | ||
| debian/reproducible-buildpath.diff | Add .file tags to all start.S files When linking raw object files, if they don't have a FILE entry _and_ they contain local symbols, ld.bfd will add FILE entries itself, but with the file name specified on the command line. As the diet wrapper adds an absolute path for the startup file here, this causes an implicit encoding of the build path. Simply add .file "start.S at the beginning of all start.S startup files to avoid this problem. . See also: https://bugs.debian.org/857315 |
Christian Seiler <christian@iwakd.de> | invalid | |||
| security/insecure-default-PATH.diff | Fix insecure default PATH Throsten Glaser <t.glaser@tarent.de> discovered that the default PATH (absent the environment variable) contains the local directory. |
Christian Seiler <christian@iwakd.de> | yes | upstream | 2016-07-20 |