~mirabilos/klibc/master : changes

~mirabilos/klibc/master » Changes from revision 2240

From Revision 2237 to 2218

Rev	Summary	Authors	Tags	Date
2237	[klibc] syscalls: Make more system calls non-optional [klibc] syscalls: Make more system calls non-optional Currently many system calls are marked as optional in SYSCALLS.def, but should be defined on all architectures. Make them non-optional. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		3 months ago
2236	[klibc] syscalls: Clean up the fork() and clone() definition... [klibc] syscalls: Clean up the fork() and clone() definitions Currently we're defining __clone() on all architectures, but it's only needed where fork() is missing and it's broken on s390 and s390x where the parameter order is incompatible. We're also building a list of new architectures which don't implement fork(), when we could automatically detect that. To fix these problems: - If __NR_fork is defined, generate a fork() system call wrapper, with only sparc and sparc64 as special cases. - Otherwise, generate __clone() as a clone() system call wrapper. Comment why this works despite the architecture differences. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		3 months ago
2235	[klibc] syscalls: Define mlock() etc. on all architectures [klibc] syscalls: Define mlock() etc. on all architectures Due to an error in editing preprocessor conditionals long ago, the m{,un}lock{,all}() system call wrappers are only generated on architectures where we use mmap() and not mmap2(). Move the #endif to the right place. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		3 months ago
2234	[klibc] Remove unused no-MMU support [klibc] Remove unused no-MMU support There is some conditional logic to handle no-MMU architectures, but no such architectures are defined. We are now having to override the automatic detection for each new architecture because it only checks for fork() and not clone(). Since this is currently unused and therefore untested, it's probably broken. Remove it. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		3 months ago
2233	[klibc] parisc: Fix build with Linux 6.10+ [klibc] parisc: Fix build with Linux 6.10+ The parisc version of <asm/signal.h> did not define _NSIG or struct sigset until recently, so we did it in <klibc/archsignal.h>. In Linux 6.10 and on some stable branches these definitions have been added, conflicting with our definitions. So make ours conditional on !defined(_NSIG). Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		3 months ago
2232	[klibc] Delete the ia64 port [klibc] Delete the ia64 port The Linux ia64 port was finally deleted in v6.7, so there's no point in maintaining support in klibc. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		4 months ago
2231	[klibc] inet: Stricter IPv6 field parsing in inet_pton() [klibc] inet: Stricter IPv6 field parsing in inet_pton() We currently don't range-check the fields of an IPv6 address, so the following strings are wrongly accepted: "10000::" "::10000" Since we currently only support hexadecimal fields, implement the range check by limiting the number of digits to 4. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		9 months ago
2230	[klibc] inet: Stricter byte parsing in inet_aton() [klibc] inet: Stricter byte parsing in inet_aton() sscanf() skips spaces in some places, and accepts values that are out-of-range for a byte, so the following strings are wrongly accepted: "1.2.3.256" "256.1.2.3" "-1.2.3.4" "1.2.3.-4" " 1.2.3.4" "1. 2.3.4" "1.2.3.4 " "1.2.3.0x4" Implement our own strict conversion of decimal digits to bytes. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		9 months ago
2229	[klibc] inet: Fix character type test in inet_pton() [klibc] inet: Fix character type test in inet_pton() The argument to ixdigit() and other ctype functions must be a char value converted to unsigned char, or -1; otherwise the results are undefined. In practice, if char is signed we could passs a value < -1 which would result in an out-of-bounds read. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		9 months ago
2228	[klibc] tests: Add test programs for inet_aton() and inet_pt... [klibc] tests: Add test programs for inet_aton() and inet_pton() The preceding bug fix provides strong evidence that these functions haven't been tested, so add test programs. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		9 months ago
2227	[klibc] Fix the inet_pton implementation [klibc] Fix the inet_pton implementation The function inet_pton should read from src but it actually reads from dst. Signed-off-by: Donny Xia <xiadong.main@gmail.com> Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Donny Xia		9 months ago
2226	[klibc] 2.0.13 released, next version is 2.0.14 [klibc] 2.0.13 released, next version is 2.0.14	Ben Hutchings		1 year ago
2225	[klibc] Install command links as appropriate [klibc] Install command links as appropriate We build gzip to support being called as "gunzip" or "zcat", and similarly we build halt to support being called as "poweroff" or "reboot". In the build directory, we create the aliases as links to the same binary. When installing, however, we pass all the names for the same binary to the "install" command, creating multiple copies of it in the install directory. To fix this: - Define an install_links command that creates links as specified in an install-links-y variable. - Invoke install_links after installing binaries. - In usr/gzip/Kbuild and usr/utils/Kbuild, specify the links in install-links-y instead of install-y. Reported-by: Benjamin Drung <benjamin.drung@canonical.com> References: https://bugs.launchpad.net/ubuntu/+source/klibc/+bug/2028571 Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings	klibc-2.0.13	1 year ago
2224	[klibc] kinit: Support specifying root with PARTLABEL [klibc] kinit: Support specifying root with PARTLABEL Allow specifying the root device to be mounted as root=PARTLABEL=label. The label is the GPT partition label of the intended root partition, as presented by the kernel with the uevent entry PARTNAME. Signed-off-by: Rob Vandermeulen <rvandermeulen@google.com> Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Rob Vandermeulen		1 year ago
2223	[klibc] riscv64: Move shared library below executables [klibc] riscv64: Move shared library below executables The default base address for riscv64 executables is 64 kiB, and I originally selected a base address for the shared library of 2 MiB. This limits the size of executables using it to be < 2 MiB, and it also interacts with a bug in QEMU's user-mode emulation, causing executables to immediately crash. It doesn't help code generation, because single-instruction jumps have a range of only ±1 MiB and two-instruction jumps have a range of ±2 GiB. Move the shared library to 64 kiB and executables to 576 kiB, to allow for generation of single-instruction jumps. (klibc.so is currently much smaller than the 512 kiB this allows for.) References: https://bugs.debian.org/1040981 Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		1 year ago
2222	[klibc] arm: Move shared library below executables [klibc] arm: Move shared library below executables The default base address for arm executables is 64 kiB, and the base address for the shared library is 3.5 MiB when generating Thumb instructions. This limits the size of executables using it to be < 3.5 MiB, and it also interacts with a bug in QEMU's user-mode emulation, causing executables to immediately crash. Moving the shared library to the end of the 16 MiB range didn't help. Instead, move the shared library to 2 MiB and executables to 4 MiB. This works for both Thumb and non-Thumb configurations, so make it unconditional. References: https://bugs.debian.org/1040981 Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		1 year ago
2221	[klibc] mips64: Compatibility with R6 compact branches [klibc] mips64: Compatibility with R6 compact branches MIPS R6 introduced compact branch instructions, including BALC as a replacement for JAL. Both BALC and JAL have a 26-bit offset field that's shifted 2 bits left. However, in BALC this is treated as a PC-relative signed offset (±128 MiB) rather than a segment-relative unsigned offset (0-256 MiB). The base address of the klibc shared library therefore needs to change depending on whether the executable calls it with JAL or BALC. gcc (but not Clang, currently) generates BALC instructions for external function calls on R6 targets, which breaks shared library builds of klibc for these targets. Add a config symbol to control whether compact branches are used and to change the shared library base address accordingly. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		1 year ago
2220	[klibc] ipconfig: Retry failed send after 1 second [klibc] ipconfig: Retry failed send after 1 second Currently we check whether a device has link down before attempting to send a packet, but the send can still fail with ENETDOWN - perhaps due to racing with the kernel's linkwatch work. After such an error we'll wait about 10 seconds before retrying. Since the link check isn't reliable, remove it. Instead, wait only 1 second if we fail to send a packet. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		1 year ago
2219	[klibc] ipconfig: Handle error events from poll [klibc] ipconfig: Handle error events from poll When poll returns events for a socket other than POLLRDNORM, such as POLLERR, we do not read from the socket but keep polling on it. This results in busy-polling while the error state persists. In case some other event is returned, put the device into DEVST_ERROR state for 1 second. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		1 year ago
2218	[klibc] ipconfig: Do not poll sockets we don't intend to rea... [klibc] ipconfig: Do not poll sockets we don't intend to read from When a device is in state DEVST_COMPLETE or DEVST_ERROR, we never read packets from its socket, but we still poll it. If the socket is readable or in an error state, this results in busy-polling until a timeout is reached. Only add a device's socket to the fds array when the device is in some other state. Signed-off-by: Ben Hutchings <ben@decadent.org.uk>	Ben Hutchings		1 year ago

Older »