| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
This allows compiling a full 128-bit product of two 64-bit integers as a
mul/mulh pair, instead of a nasty long sequence of 20+ instructions.
However, after selecting ARCH_SUPPORTS_INT128, when optimizing for size
the compiler generates calls to __ashlti3, __ashrti3, and __lshrti3 for
shifting __int128 values, causing a link failure:
loongarch64-unknown-linux-gnu-ld: kernel/sched/fair.o: in
function `mul_u64_u32_shr':
<PATH>/include/linux/math64.h:161:(.text+0x5e4): undefined
reference to `__lshrti3'
So provide the implementation of these functions if ARCH_SUPPORTS_INT128.
Closes: https://lore.kernel.org/loongarch/CAAhV-H5EZ=7OF7CSiYyZ8_+wWuenpo=K2WT8-6mAT4CvzUC_4g@mail.gmail.com/
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
The kernel CONFIG_UNWINDER_ORC option enables the ORC unwinder, which is
similar in concept to a DWARF unwinder. The difference is that the format
of the ORC data is much simpler than DWARF, which in turn allows the ORC
unwinder to be much simpler and faster.
The ORC data consists of unwind tables which are generated by objtool.
After analyzing all the code paths of a .o file, it determines information
about the stack state at each instruction address in the file and outputs
that information to the .orc_unwind and .orc_unwind_ip sections.
The per-object ORC sections are combined at link time and are sorted and
post-processed at boot time. The unwinder uses the resulting data to
correlate instruction addresses with their stack states at run time.
Most of the logic are similar with x86, in order to get ra info before ra
is saved into stack, add ra_reg and ra_offset into orc_entry. At the same
time, modify some arch-specific code to silence the objtool warnings.
Co-developed-by: Jinyang He <hejinyang@loongson.cn>
Signed-off-by: Jinyang He <hejinyang@loongson.cn>
Co-developed-by: Youling Tang <tangyouling@loongson.cn>
Signed-off-by: Youling Tang <tangyouling@loongson.cn>
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
1/8 of kernel addresses reserved for shadow memory. But for LoongArch,
There are a lot of holes between different segments and valid address
space (256T available) is insufficient to map all these segments to kasan
shadow memory with the common formula provided by kasan core, saying
(addr >> KASAN_SHADOW_SCALE_SHIFT) + KASAN_SHADOW_OFFSET
So LoongArch has a arch-specific mapping formula, different segments are
mapped individually, and only limited space lengths of these specific
segments are mapped to shadow.
At early boot stage the whole shadow region populated with just one
physical page (kasan_early_shadow_page). Later, this page is reused as
readonly zero shadow for some memory that kasan currently don't track.
After mapping the physical memory, pages for shadow memory are allocated
and mapped.
Functions like memset()/memcpy()/memmove() do a lot of memory accesses.
If bad pointer passed to one of these function it is important to be
caught. Compiler's instrumentation cannot do this since these functions
are written in assembly.
KASan replaces memory functions with manually instrumented variants.
Original functions declared as weak symbols so strong definitions in
mm/kasan/kasan.c could replace them. Original functions have aliases
with '__' prefix in names, so we could call non-instrumented variant
if needed.
Signed-off-by: Qing Zhang <zhangqing@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Add LSX and LASX implementations of xor operations, operating on 64
bytes (one L1 cache line) at a time, for a balance between memory
utilization and instruction mix. Huacai confirmed that all future
LoongArch implementations by Loongson (that we care) will likely also
feature 64-byte cache lines, and experiments show no throughput
improvement with further unrolling.
Performance numbers measured during system boot on a 3A5000 @ 2.5GHz:
> 8regs : 12702 MB/sec
> 8regs_prefetch : 10920 MB/sec
> 32regs : 12686 MB/sec
> 32regs_prefetch : 10918 MB/sec
> lsx : 17589 MB/sec
> lasx : 26116 MB/sec
Acked-by: Song Liu <song@kernel.org>
Signed-off-by: WANG Xuerui <git@xen0n.name>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
The {copy, clear}_user function should returns number of bytes that
could not be {copied, cleared}. So, try to {copy, clear} byte by byte
when ld.{d,w,h} and st.{d,w,h} trapped into an exception.
Reviewed-by: WANG Rui <wangrui@loongson.cn>
Signed-off-by: Weihao Li <liweihao@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Commit ddb5cdbafaaad ("kbuild: generate KSYMTAB entries by modpost")
deprecated <asm/export.h>, which is now a wrapper of <linux/export.h>.
Replace #include <asm/export.h> with #include <linux/export.h>.
After all the <asm/export.h> lines are converted, <asm/export.h> and
<asm-generic/export.h> will be removed.
Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
There is no EXPORT_SYMBOL() line there, hence #include <asm/export.h>
is unneeded.
Signed-off-by: Masahiro Yamada <masahiroy@kernel.org>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
This patch fixes an underflow issue in the return value within the
exception path, specifically at .Llt8 when the remaining length is less
than 8 bytes.
Cc: stable@vger.kernel.org
Fixes: 8941e93ca590 ("LoongArch: Optimize memory ops (memset/memcpy/memmove)")
Reported-by: Weihao Li <liweihao@loongson.cn>
Signed-off-by: WANG Rui <wangrui@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
In addition to less visual clutter, this also makes Clang happy
regarding the const-ness of arguments. In the original approach, all
Clang gets to see is the incoming arguments whose const-ness cannot be
proven without first being inlined; so Clang errors out here while GCC
is fine.
While at it, tweak several printk format strings because the return type
of csr_read64 becomes effectively unsigned long, instead of unsigned
long long.
Signed-off-by: WANG Xuerui <git@xen0n.name>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Inspired by the commit 42d038c4fb00f ("arm64: Add support for function
error injection") and the commit ee55ff803b383 ("riscv: Add support for
function error injection"), this patch supports function error injection
for LoongArch.
Mainly implement two functions:
(1) regs_set_return_value() which is used to overwrite the return value,
(2) override_function_with_return() which is used to override the probed
function returning and jump to its caller.
Here is a simple test under CONFIG_FUNCTION_ERROR_INJECTION and
CONFIG_FAIL_FUNCTION:
# echo sys_clone > /sys/kernel/debug/fail_function/inject
# echo 100 > /sys/kernel/debug/fail_function/probability
# dmesg
bash: fork: Invalid argument
# dmesg
...
FAULT_INJECTION: forcing a failure.
name fail_function, interval 1, probability 100, space 0, times 1
...
Call Trace:
[<90000000002238f4>] show_stack+0x5c/0x180
[<90000000012e384c>] dump_stack_lvl+0x60/0x88
[<9000000000b1879c>] should_fail_ex+0x1b0/0x1f4
[<900000000032ead4>] fei_kprobe_handler+0x28/0x6c
[<9000000000230970>] kprobe_breakpoint_handler+0xf0/0x118
[<90000000012e3e60>] do_bp+0x2c4/0x358
[<9000000002241924>] exception_handlers+0x1924/0x10000
[<900000000023b7d0>] sys_clone+0x0/0x4
[<90000000012e4744>] do_syscall+0x7c/0x94
[<9000000000221e44>] handle_syscall+0xc4/0x160
Tested-by: Hengqi Chen <hengqi.chen@gmail.com>
Acked-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
LoongArch platform is 64-bit system, which supports 8-bytes memory
accessing, but generic checksum functions use 4-byte memory access.
So add 8-bytes memory access optimization for checksum functions on
LoongArch. And the code comes from arm64 system.
When network hw checksum is disabled, iperf performance improves about
10% with this patch.
Signed-off-by: Bibo Mao <maobibo@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
To optimize memset()/memcpy()/memmove() and so on, we use a jump table
to dispatch cases for short data lengths; and for long data lengths, we
split the destination into head part (first 8 bytes), tail part (last 8
bytes) and middle part. The head part and tail part may be at unaligned
addresses, while the middle part is always aligned (the middle part is
allowed to overlap the head/tail part). In this way, the first and last
8 bytes may be unaligned accesses, but we can make sure the data in the
middle is processed at an aligned destination address.
We have tested micro-bench[1] on a Loongson-3C5000 16-core machine (2.2GHz):
1. memset
| length | src offset | dst offset | speed before | speed after | % |
|--------|------------|------------|--------------|-------------|---------|
| 8 | 0 | 0 | 696.191 | 1518.785 | 118.16% |
| 8 | 0 | 1 | 696.325 | 1518.937 | 118.14% |
| 50 | 0 | 0 | 969.976 | 8053.902 | 730.32% |
| 50 | 0 | 1 | 970.034 | 8058.475 | 730.74% |
| 300 | 0 | 0 | 5876.612 | 16544.703 | 181.53% |
| 300 | 0 | 1 | 5030.849 | 16549.011 | 228.95% |
| 1200 | 0 | 0 | 11797.077 | 16752.137 | 42.00% |
| 1200 | 0 | 1 | 5687.141 | 16645.233 | 192.68% |
| 4000 | 0 | 0 | 15723.27 | 16761.557 | 6.60% |
| 4000 | 0 | 1 | 5906.114 | 16732.316 | 183.30% |
| 8000 | 0 | 0 | 16751.403 | 16770.002 | 0.11% |
| 8000 | 0 | 1 | 5995.449 | 16754.07 | 179.45% |
2. memcpy
| length | src offset | dst offset | speed before | speed after | % |
|--------|------------|------------|--------------|-------------|---------|
| 8 | 0 | 0 | 696.2 | 1670.605 | 139.96% |
| 8 | 0 | 1 | 696.325 | 1671.138 | 139.99% |
| 50 | 0 | 0 | 969.974 | 8724.999 | 799.51% |
| 50 | 0 | 1 | 970.032 | 8730.138 | 799.98% |
| 300 | 0 | 0 | 5564.662 | 16272.652 | 192.43% |
| 300 | 0 | 1 | 4670.436 | 14972.842 | 220.59% |
| 1200 | 0 | 0 | 10740.23 | 16751.728 | 55.97% |
| 1200 | 0 | 1 | 5027.741 | 14874.564 | 195.85% |
| 4000 | 0 | 0 | 15122.367 | 16737.642 | 10.68% |
| 4000 | 0 | 1 | 5536.918 | 14890.397 | 168.93% |
| 8000 | 0 | 0 | 16505.453 | 16553.543 | 0.29% |
| 8000 | 0 | 1 | 5821.619 | 14841.804 | 154.94% |
3. memmove
| length | src offset | dst offset | speed before | speed after | % |
|--------|------------|------------|--------------|-------------|---------|
| 8 | 0 | 0 | 982.693 | 1670.568 | 70.00% |
| 8 | 0 | 1 | 983.023 | 1671.174 | 70.00% |
| 50 | 0 | 0 | 1230.87 | 8727.625 | 609.06% |
| 50 | 0 | 1 | 1232.515 | 8730.138 | 608.32% |
| 300 | 0 | 0 | 6490.375 | 16296.993 | 151.09% |
| 300 | 0 | 1 | 4282.687 | 14972.842 | 249.61% |
| 1200 | 0 | 0 | 11742.755 | 16752.546 | 42.66% |
| 1200 | 0 | 1 | 5039.338 | 14872.951 | 195.14% |
| 4000 | 0 | 0 | 15467.786 | 16737.09 | 8.21% |
| 4000 | 0 | 1 | 5009.905 | 14890.542 | 197.22% |
| 8000 | 0 | 0 | 16489.664 | 16553.273 | 0.39% |
| 8000 | 0 | 1 | 5823.786 | 14858.646 | 155.14% |
* speed: MB/s
* length: byte
[1] https://github.com/heiher/mem-bench
Signed-off-by: WANG Rui <wangrui@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Some assembler symbols are not kprobe safe, such as handle_syscall (used
as syscall exception handler), *memset*/*memcpy*/*memmove* (may cause
recursive exceptions), they can not be instrumented, just blacklist them
for kprobing.
Here is a related problem and discussion:
Link: https://lore.kernel.org/lkml/20230114143859.7ccc45c1c5d9ce302113ab0a@kernel.org/
Tested-by: Jeff Xie <xiehuan09@gmail.com>
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Use the alternative to optimize common libraries according whether CPU
has UAL (hardware unaligned access support) feature, including memset(),
memcopy(), memmove(), copy_user() and clear_user().
We have tested UnixBench on a Loongson-3A5000 quad-core machine (1.6GHz):
1, One copy, before patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 9566582.0 819.8
Double-Precision Whetstone 55.0 2805.3 510.1
Execl Throughput 43.0 2120.0 493.0
File Copy 1024 bufsize 2000 maxblocks 3960.0 209833.0 529.9
File Copy 256 bufsize 500 maxblocks 1655.0 89400.0 540.2
File Copy 4096 bufsize 8000 maxblocks 5800.0 320036.0 551.8
Pipe Throughput 12440.0 340624.0 273.8
Pipe-based Context Switching 4000.0 109939.1 274.8
Process Creation 126.0 4728.7 375.3
Shell Scripts (1 concurrent) 42.4 2223.1 524.3
Shell Scripts (8 concurrent) 6.0 883.1 1471.9
System Call Overhead 15000.0 518639.1 345.8
========
System Benchmarks Index Score 500.2
2, One copy, after patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 9567674.7 819.9
Double-Precision Whetstone 55.0 2805.5 510.1
Execl Throughput 43.0 2392.7 556.4
File Copy 1024 bufsize 2000 maxblocks 3960.0 417804.0 1055.1
File Copy 256 bufsize 500 maxblocks 1655.0 112909.5 682.2
File Copy 4096 bufsize 8000 maxblocks 5800.0 1255207.4 2164.2
Pipe Throughput 12440.0 555712.0 446.7
Pipe-based Context Switching 4000.0 99964.5 249.9
Process Creation 126.0 5192.5 412.1
Shell Scripts (1 concurrent) 42.4 2302.4 543.0
Shell Scripts (8 concurrent) 6.0 919.6 1532.6
System Call Overhead 15000.0 511159.3 340.8
========
System Benchmarks Index Score 640.1
3, Four copies, before patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 38268610.5 3279.2
Double-Precision Whetstone 55.0 11222.2 2040.4
Execl Throughput 43.0 7892.0 1835.3
File Copy 1024 bufsize 2000 maxblocks 3960.0 235149.6 593.8
File Copy 256 bufsize 500 maxblocks 1655.0 74959.6 452.9
File Copy 4096 bufsize 8000 maxblocks 5800.0 545048.5 939.7
Pipe Throughput 12440.0 1337359.0 1075.0
Pipe-based Context Switching 4000.0 473663.9 1184.2
Process Creation 126.0 17491.2 1388.2
Shell Scripts (1 concurrent) 42.4 6865.7 1619.3
Shell Scripts (8 concurrent) 6.0 1015.9 1693.1
System Call Overhead 15000.0 1899535.2 1266.4
========
System Benchmarks Index Score 1278.3
4, Four copies, after patch:
System Benchmarks Index Values BASELINE RESULT INDEX
Dhrystone 2 using register variables 116700.0 38272815.5 3279.6
Double-Precision Whetstone 55.0 11222.8 2040.5
Execl Throughput 43.0 8839.2 2055.6
File Copy 1024 bufsize 2000 maxblocks 3960.0 313912.9 792.7
File Copy 256 bufsize 500 maxblocks 1655.0 80976.1 489.3
File Copy 4096 bufsize 8000 maxblocks 5800.0 1176594.3 2028.6
Pipe Throughput 12440.0 2100941.9 1688.9
Pipe-based Context Switching 4000.0 476696.4 1191.7
Process Creation 126.0 18394.7 1459.9
Shell Scripts (1 concurrent) 42.4 7172.2 1691.6
Shell Scripts (8 concurrent) 6.0 1058.3 1763.9
System Call Overhead 15000.0 1874714.7 1249.8
========
System Benchmarks Index Score 1488.8
Signed-off-by: Jun Yi <yijun@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Loongson-2 series (Loongson-2K500, Loongson-2K1000) don't support
unaligned access in hardware, while Loongson-3 series (Loongson-3A5000,
Loongson-3C5000) are configurable whether support unaligned access in
hardware. This patch add unaligned access emulation for those LoongArch
processors without hardware support.
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Use the `.L_xxx` label to improve fixup code and then remove the .fixup
section usage.
Signed-off-by: Youling Tang <tangyouling@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Consolidate all the __ex_table constuction code with a _ASM_EXTABLE or
_asm_extable helper.
There should be no functional change as a result of this patch.
Signed-off-by: Youling Tang <tangyouling@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
1, Use nr/nx to replace ri/xi;
2, Add 0x prefix for hexadecimal data.
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
The content of LoongArch's compiler.h is trivial, with some unused
anywhere, so inline the definitions and remove the header.
Signed-off-by: Jun Yi <yijun@loongson.cn>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Support for the syntactic sugar is present in upstream binutils port
from the beginning. Use it for shorter lines and better consistency.
Generated code should be identical.
Signed-off-by: WANG Xuerui <git@xen0n.name>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Add some library functions for LoongArch, including: delay, memset,
memcpy, memmove, copy_user, strncpy_user, strnlen_user and tlb dump
functions.
Reviewed-by: WANG Xuerui <git@xen0n.name>
Reviewed-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
|
Add Kbuild, Makefile, Kconfig and link script for LoongArch build
infrastructure.
Reviewed-by: Guo Ren <guoren@kernel.org>
Reviewed-by: WANG Xuerui <git@xen0n.name>
Reviewed-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
|
