LoongArch: Add vector extensions support

LoongArch inclusion
category: feature
bugzilla: https://gitee.com/openeuler/kernel/issues/I6BWFP

--------------------------------

Add LoongArch's vector extensions support, which including 128bit LSX
(i.e., Loongson SIMD eXtension) and 256bit LASX (i.e., Loongson Advanced
SIMD eXtension).
Signed-off-by: NHuacai Chen <chenhuacai@loongson.cn>
Change-Id: Icbdcc8c0657d85198accb9050f90dd41c9fa456c

arch/loongarch/Kconfig

@@ -75,6 +75,8 @@ config LOONGARCH
 	select GENERIC_SCHED_CLOCK
 	select GENERIC_SMP_IDLE_THREAD
 	select GENERIC_TIME_VSYSCALL
+	select CPU_SUPPORTS_LSX
+	select CPU_SUPPORTS_LASX
 	select GPIOLIB
 	select HAVE_ARCH_AUDITSYSCALL
 	select HAVE_ARCH_MMAP_RND_BITS if MMU
@@ -145,6 +147,36 @@ config CPU_HAS_PREFETCH
 	bool
 	default y
 
+config CPU_HAS_LSX
+	bool "Support for the Loongson SIMD Extension"
+	depends on CPU_SUPPORTS_LSX
+	depends on 64BIT
+	help
+	  Loongson SIMD Extension (LSX) introduces 128 bit wide vector registers
+	  and a set of SIMD instructions to operate on them. When this option
+	  is enabled the kernel will support allocating & switching LSX
+	  vector register contexts. If you know that your kernel will only be
+	  running on CPUs which do not support LSX or that your userland will
+	  not be making use of it then you may wish to say N here to reduce
+	  the size & complexity of your kernel.
+
+	  If unsure, say Y.
+
+config CPU_HAS_LASX
+	bool "Support for the Loongson Advanced SIMD Extension"
+	depends on CPU_SUPPORTS_LASX
+	depends on 64BIT && CPU_HAS_LSX
+	help
+	  Loongson Advanced SIMD Extension is 256 bit wide SIMD extension.
+
+	  If unsure, say Y.
+
+config CPU_SUPPORTS_LSX
+	bool
+
+config CPU_SUPPORTS_LASX
+	bool
+
 config GENERIC_CALIBRATE_DELAY
 	def_bool y
 

arch/loongarch/configs/loongson3_defconfig

@@ -35,6 +35,8 @@ CONFIG_BPF_SYSCALL=y
 CONFIG_USERFAULTFD=y
 CONFIG_PERF_EVENTS=y
 # CONFIG_COMPAT_BRK is not set
+CONFIG_CPU_HAS_LSX=y
+CONFIG_CPU_HAS_LASX=y
 CONFIG_HOTPLUG_CPU=y
 CONFIG_NUMA=y
 CONFIG_ACPI_SPCR_TABLE=y

arch/loongarch/include/asm/asmmacro.h

@@ -117,6 +117,212 @@
 	.endif
 	.endm
 
+	.macro	parse_vr var vr
+	\var	= -1
+	.ifc	\vr, $vr0
+	\var	= 0
+	.endif
+	.ifc	\vr, $vr1
+	\var	= 1
+	.endif
+	.ifc	\vr, $vr2
+	\var	= 2
+	.endif
+	.ifc	\vr, $vr3
+	\var	= 3
+	.endif
+	.ifc	\vr, $vr4
+	\var	= 4
+	.endif
+	.ifc	\vr, $vr5
+	\var	= 5
+	.endif
+	.ifc	\vr, $vr6
+	\var	= 6
+	.endif
+	.ifc	\vr, $vr7
+	\var	= 7
+	.endif
+	.ifc	\vr, $vr8
+	\var	= 8
+	.endif
+	.ifc	\vr, $vr9
+	\var	= 9
+	.endif
+	.ifc	\vr, $vr10
+	\var	= 10
+	.endif
+	.ifc	\vr, $vr11
+	\var	= 11
+	.endif
+	.ifc	\vr, $vr12
+	\var	= 12
+	.endif
+	.ifc	\vr, $vr13
+	\var	= 13
+	.endif
+	.ifc	\vr, $vr14
+	\var	= 14
+	.endif
+	.ifc	\vr, $vr15
+	\var	= 15
+	.endif
+	.ifc	\vr, $vr16
+	\var	= 16
+	.endif
+	.ifc	\vr, $vr17
+	\var	= 17
+	.endif
+	.ifc	\vr, $vr18
+	\var	= 18
+	.endif
+	.ifc	\vr, $vr19
+	\var	= 19
+	.endif
+	.ifc	\vr, $vr20
+	\var	= 20
+	.endif
+	.ifc	\vr, $vr21
+	\var	= 21
+	.endif
+	.ifc	\vr, $vr22
+	\var	= 22
+	.endif
+	.ifc	\vr, $vr23
+	\var	= 23
+	.endif
+	.ifc	\vr, $vr24
+	\var	= 24
+	.endif
+	.ifc	\vr, $vr25
+	\var	= 25
+	.endif
+	.ifc	\vr, $vr26
+	\var	= 26
+	.endif
+	.ifc	\vr, $vr27
+	\var	= 27
+	.endif
+	.ifc	\vr, $vr28
+	\var	= 28
+	.endif
+	.ifc	\vr, $vr29
+	\var	= 29
+	.endif
+	.ifc	\vr, $vr30
+	\var	= 30
+	.endif
+	.ifc	\vr, $vr31
+	\var	= 31
+	.endif
+	.iflt	\var
+	.error	"Unable to parse register name \r"
+	.endif
+	.endm
+
+	.macro	parse_xr var xr
+	\var	= -1
+	.ifc	\xr, $xr0
+	\var	= 0
+	.endif
+	.ifc	\xr, $xr1
+	\var	= 1
+	.endif
+	.ifc	\xr, $xr2
+	\var	= 2
+	.endif
+	.ifc	\xr, $xr3
+	\var	= 3
+	.endif
+	.ifc	\xr, $xr4
+	\var	= 4
+	.endif
+	.ifc	\xr, $xr5
+	\var	= 5
+	.endif
+	.ifc	\xr, $xr6
+	\var	= 6
+	.endif
+	.ifc	\xr, $xr7
+	\var	= 7
+	.endif
+	.ifc	\xr, $xr8
+	\var	= 8
+	.endif
+	.ifc	\xr, $xr9
+	\var	= 9
+	.endif
+	.ifc	\xr, $xr10
+	\var	= 10
+	.endif
+	.ifc	\xr, $xr11
+	\var	= 11
+	.endif
+	.ifc	\xr, $xr12
+	\var	= 12
+	.endif
+	.ifc	\xr, $xr13
+	\var	= 13
+	.endif
+	.ifc	\xr, $xr14
+	\var	= 14
+	.endif
+	.ifc	\xr, $xr15
+	\var	= 15
+	.endif
+	.ifc	\xr, $xr16
+	\var	= 16
+	.endif
+	.ifc	\xr, $xr17
+	\var	= 17
+	.endif
+	.ifc	\xr, $xr18
+	\var	= 18
+	.endif
+	.ifc	\xr, $xr19
+	\var	= 19
+	.endif
+	.ifc	\xr, $xr20
+	\var	= 20
+	.endif
+	.ifc	\xr, $xr21
+	\var	= 21
+	.endif
+	.ifc	\xr, $xr22
+	\var	= 22
+	.endif
+	.ifc	\xr, $xr23
+	\var	= 23
+	.endif
+	.ifc	\xr, $xr24
+	\var	= 24
+	.endif
+	.ifc	\xr, $xr25
+	\var	= 25
+	.endif
+	.ifc	\xr, $xr26
+	\var	= 26
+	.endif
+	.ifc	\xr, $xr27
+	\var	= 27
+	.endif
+	.ifc	\xr, $xr28
+	\var	= 28
+	.endif
+	.ifc	\xr, $xr29
+	\var	= 29
+	.endif
+	.ifc	\xr, $xr30
+	\var	= 30
+	.endif
+	.ifc	\xr, $xr31
+	\var	= 31
+	.endif
+	.iflt	\var
+	.error	"Unable to parse register name \r"
+	.endif
+	.endm
+
 	.macro	cpu_save_nonscratch thread
 	stptr.d	s0, \thread, THREAD_REG23
 	stptr.d	s1, \thread, THREAD_REG24
@@ -270,6 +476,424 @@
 	fld.d	$f31, \tmp, THREAD_FPR31 - THREAD_FPR0
 	.endm
 
+	.macro lsx_save_data thread tmp
+	parse_r __tmp, \tmp
+	li.w		\tmp, THREAD_FPR0
+	PTR_ADD 	\tmp, \thread, \tmp
+	/* vst opcode is 0xb1 */
+	.word (0xb1 << 22 | ((THREAD_FPR0-THREAD_FPR0) << 10) | __tmp << 5 | 0)
+	.word (0xb1 << 22 | ((THREAD_FPR1-THREAD_FPR0) << 10) | __tmp << 5 | 1)
+	.word (0xb1 << 22 | ((THREAD_FPR2-THREAD_FPR0) << 10) | __tmp << 5 | 2)
+	.word (0xb1 << 22 | ((THREAD_FPR3-THREAD_FPR0) << 10) | __tmp << 5 | 3)
+	.word (0xb1 << 22 | ((THREAD_FPR4-THREAD_FPR0) << 10) | __tmp << 5 | 4)
+	.word (0xb1 << 22 | ((THREAD_FPR5-THREAD_FPR0) << 10) | __tmp << 5 | 5)
+	.word (0xb1 << 22 | ((THREAD_FPR6-THREAD_FPR0) << 10) | __tmp << 5 | 6)
+	.word (0xb1 << 22 | ((THREAD_FPR7-THREAD_FPR0) << 10) | __tmp << 5 | 7)
+	.word (0xb1 << 22 | ((THREAD_FPR8-THREAD_FPR0) << 10) | __tmp << 5 | 8)
+	.word (0xb1 << 22 | ((THREAD_FPR9-THREAD_FPR0) << 10) | __tmp << 5 | 9)
+	.word (0xb1 << 22 | ((THREAD_FPR10-THREAD_FPR0) << 10) | __tmp << 5 | 10)
+	.word (0xb1 << 22 | ((THREAD_FPR11-THREAD_FPR0) << 10) | __tmp << 5 | 11)
+	.word (0xb1 << 22 | ((THREAD_FPR12-THREAD_FPR0) << 10) | __tmp << 5 | 12)
+	.word (0xb1 << 22 | ((THREAD_FPR13-THREAD_FPR0) << 10) | __tmp << 5 | 13)
+	.word (0xb1 << 22 | ((THREAD_FPR14-THREAD_FPR0) << 10) | __tmp << 5 | 14)
+	.word (0xb1 << 22 | ((THREAD_FPR15-THREAD_FPR0) << 10) | __tmp << 5 | 15)
+	.word (0xb1 << 22 | ((THREAD_FPR16-THREAD_FPR0) << 10) | __tmp << 5 | 16)
+	.word (0xb1 << 22 | ((THREAD_FPR17-THREAD_FPR0) << 10) | __tmp << 5 | 17)
+	.word (0xb1 << 22 | ((THREAD_FPR18-THREAD_FPR0) << 10) | __tmp << 5 | 18)
+	.word (0xb1 << 22 | ((THREAD_FPR19-THREAD_FPR0) << 10) | __tmp << 5 | 19)
+	.word (0xb1 << 22 | ((THREAD_FPR20-THREAD_FPR0) << 10) | __tmp << 5 | 20)
+	.word (0xb1 << 22 | ((THREAD_FPR21-THREAD_FPR0) << 10) | __tmp << 5 | 21)
+	.word (0xb1 << 22 | ((THREAD_FPR22-THREAD_FPR0) << 10) | __tmp << 5 | 22)
+	.word (0xb1 << 22 | ((THREAD_FPR23-THREAD_FPR0) << 10) | __tmp << 5 | 23)
+	.word (0xb1 << 22 | ((THREAD_FPR24-THREAD_FPR0) << 10) | __tmp << 5 | 24)
+	.word (0xb1 << 22 | ((THREAD_FPR25-THREAD_FPR0) << 10) | __tmp << 5 | 25)
+	.word (0xb1 << 22 | ((THREAD_FPR26-THREAD_FPR0) << 10) | __tmp << 5 | 26)
+	.word (0xb1 << 22 | ((THREAD_FPR27-THREAD_FPR0) << 10) | __tmp << 5 | 27)
+	.word (0xb1 << 22 | ((THREAD_FPR28-THREAD_FPR0) << 10) | __tmp << 5 | 28)
+	.word (0xb1 << 22 | ((THREAD_FPR29-THREAD_FPR0) << 10) | __tmp << 5 | 29)
+	.word (0xb1 << 22 | ((THREAD_FPR30-THREAD_FPR0) << 10) | __tmp << 5 | 30)
+	.word (0xb1 << 22 | ((THREAD_FPR31-THREAD_FPR0) << 10) | __tmp << 5 | 31)
+	.endm
+
+	.macro lsx_restore_data thread tmp
+	parse_r __tmp, \tmp
+	li.w		\tmp, THREAD_FPR0
+	PTR_ADD		\tmp, \thread, \tmp
+	/* vld opcode is 0xb0 */
+	.word (0xb0 << 22 | ((THREAD_FPR0-THREAD_FPR0) << 10) | __tmp << 5 | 0)
+	.word (0xb0 << 22 | ((THREAD_FPR1-THREAD_FPR0) << 10) | __tmp << 5 | 1)
+	.word (0xb0 << 22 | ((THREAD_FPR2-THREAD_FPR0) << 10) | __tmp << 5 | 2)
+	.word (0xb0 << 22 | ((THREAD_FPR3-THREAD_FPR0) << 10) | __tmp << 5 | 3)
+	.word (0xb0 << 22 | ((THREAD_FPR4-THREAD_FPR0) << 10) | __tmp << 5 | 4)
+	.word (0xb0 << 22 | ((THREAD_FPR5-THREAD_FPR0) << 10) | __tmp << 5 | 5)
+	.word (0xb0 << 22 | ((THREAD_FPR6-THREAD_FPR0) << 10) | __tmp << 5 | 6)
+	.word (0xb0 << 22 | ((THREAD_FPR7-THREAD_FPR0) << 10) | __tmp << 5 | 7)
+	.word (0xb0 << 22 | ((THREAD_FPR8-THREAD_FPR0) << 10) | __tmp << 5 | 8)
+	.word (0xb0 << 22 | ((THREAD_FPR9-THREAD_FPR0) << 10) | __tmp << 5 | 9)
+	.word (0xb0 << 22 | ((THREAD_FPR10-THREAD_FPR0) << 10) | __tmp << 5 | 10)
+	.word (0xb0 << 22 | ((THREAD_FPR11-THREAD_FPR0) << 10) | __tmp << 5 | 11)
+	.word (0xb0 << 22 | ((THREAD_FPR12-THREAD_FPR0) << 10) | __tmp << 5 | 12)
+	.word (0xb0 << 22 | ((THREAD_FPR13-THREAD_FPR0) << 10) | __tmp << 5 | 13)
+	.word (0xb0 << 22 | ((THREAD_FPR14-THREAD_FPR0) << 10) | __tmp << 5 | 14)
+	.word (0xb0 << 22 | ((THREAD_FPR15-THREAD_FPR0) << 10) | __tmp << 5 | 15)
+	.word (0xb0 << 22 | ((THREAD_FPR16-THREAD_FPR0) << 10) | __tmp << 5 | 16)
+	.word (0xb0 << 22 | ((THREAD_FPR17-THREAD_FPR0) << 10) | __tmp << 5 | 17)
+	.word (0xb0 << 22 | ((THREAD_FPR18-THREAD_FPR0) << 10) | __tmp << 5 | 18)
+	.word (0xb0 << 22 | ((THREAD_FPR19-THREAD_FPR0) << 10) | __tmp << 5 | 19)
+	.word (0xb0 << 22 | ((THREAD_FPR20-THREAD_FPR0) << 10) | __tmp << 5 | 20)
+	.word (0xb0 << 22 | ((THREAD_FPR21-THREAD_FPR0) << 10) | __tmp << 5 | 21)
+	.word (0xb0 << 22 | ((THREAD_FPR22-THREAD_FPR0) << 10) | __tmp << 5 | 22)
+	.word (0xb0 << 22 | ((THREAD_FPR23-THREAD_FPR0) << 10) | __tmp << 5 | 23)
+	.word (0xb0 << 22 | ((THREAD_FPR24-THREAD_FPR0) << 10) | __tmp << 5 | 24)
+	.word (0xb0 << 22 | ((THREAD_FPR25-THREAD_FPR0) << 10) | __tmp << 5 | 25)
+	.word (0xb0 << 22 | ((THREAD_FPR26-THREAD_FPR0) << 10) | __tmp << 5 | 26)
+	.word (0xb0 << 22 | ((THREAD_FPR27-THREAD_FPR0) << 10) | __tmp << 5 | 27)
+	.word (0xb0 << 22 | ((THREAD_FPR28-THREAD_FPR0) << 10) | __tmp << 5 | 28)
+	.word (0xb0 << 22 | ((THREAD_FPR29-THREAD_FPR0) << 10) | __tmp << 5 | 29)
+	.word (0xb0 << 22 | ((THREAD_FPR30-THREAD_FPR0) << 10) | __tmp << 5 | 30)
+	.word (0xb0 << 22 | ((THREAD_FPR31-THREAD_FPR0) << 10) | __tmp << 5 | 31)
+	.endm
+
+	.macro	lsx_save_all	thread tmp0 tmp1
+	fpu_save_cc	\thread, \tmp0, \tmp1
+	fpu_save_csr	\thread, \tmp0
+	lsx_save_data	\thread, \tmp0
+	.endm
+
+	.macro	lsx_restore_all	thread tmp0 tmp1
+	lsx_restore_data	\thread, \tmp0
+	fpu_restore_cc	\thread, \tmp0, \tmp1
+	fpu_restore_csr	\thread, \tmp0
+	.endm
+
+	.macro lsx_save_upper vd base tmp off
+	parse_vr __vd, \vd
+	parse_r __tmp, \tmp
+	/* vpickve2gr opcode is 0xe5dfe */
+	.word (0xe5dfe << 11 | 1 << 10 | __vd << 5 | __tmp)
+	st.d	\tmp, \base, (\off+8)
+	.endm
+
+	.macro lsx_save_all_upper thread base tmp
+	li.w	\tmp, THREAD_FPR0
+	PTR_ADD	\base, \thread, \tmp
+	lsx_save_upper $vr0,  \base, \tmp, (THREAD_FPR0-THREAD_FPR0)
+	lsx_save_upper $vr1,  \base, \tmp, (THREAD_FPR1-THREAD_FPR0)
+	lsx_save_upper $vr2,  \base, \tmp, (THREAD_FPR2-THREAD_FPR0)
+	lsx_save_upper $vr3,  \base, \tmp, (THREAD_FPR3-THREAD_FPR0)
+	lsx_save_upper $vr4,  \base, \tmp, (THREAD_FPR4-THREAD_FPR0)
+	lsx_save_upper $vr5,  \base, \tmp, (THREAD_FPR5-THREAD_FPR0)
+	lsx_save_upper $vr6,  \base, \tmp, (THREAD_FPR6-THREAD_FPR0)
+	lsx_save_upper $vr7,  \base, \tmp, (THREAD_FPR7-THREAD_FPR0)
+	lsx_save_upper $vr8,  \base, \tmp, (THREAD_FPR8-THREAD_FPR0)
+	lsx_save_upper $vr9,  \base, \tmp, (THREAD_FPR9-THREAD_FPR0)
+	lsx_save_upper $vr10, \base, \tmp, (THREAD_FPR10-THREAD_FPR0)
+	lsx_save_upper $vr11, \base, \tmp, (THREAD_FPR11-THREAD_FPR0)
+	lsx_save_upper $vr12, \base, \tmp, (THREAD_FPR12-THREAD_FPR0)
+	lsx_save_upper $vr13, \base, \tmp, (THREAD_FPR13-THREAD_FPR0)
+	lsx_save_upper $vr14, \base, \tmp, (THREAD_FPR14-THREAD_FPR0)
+	lsx_save_upper $vr15, \base, \tmp, (THREAD_FPR15-THREAD_FPR0)
+	lsx_save_upper $vr16, \base, \tmp, (THREAD_FPR16-THREAD_FPR0)
+	lsx_save_upper $vr17, \base, \tmp, (THREAD_FPR17-THREAD_FPR0)
+	lsx_save_upper $vr18, \base, \tmp, (THREAD_FPR18-THREAD_FPR0)
+	lsx_save_upper $vr19, \base, \tmp, (THREAD_FPR19-THREAD_FPR0)
+	lsx_save_upper $vr20, \base, \tmp, (THREAD_FPR20-THREAD_FPR0)
+	lsx_save_upper $vr21, \base, \tmp, (THREAD_FPR21-THREAD_FPR0)
+	lsx_save_upper $vr22, \base, \tmp, (THREAD_FPR22-THREAD_FPR0)
+	lsx_save_upper $vr23, \base, \tmp, (THREAD_FPR23-THREAD_FPR0)
+	lsx_save_upper $vr24, \base, \tmp, (THREAD_FPR24-THREAD_FPR0)
+	lsx_save_upper $vr25, \base, \tmp, (THREAD_FPR25-THREAD_FPR0)
+	lsx_save_upper $vr26, \base, \tmp, (THREAD_FPR26-THREAD_FPR0)
+	lsx_save_upper $vr27, \base, \tmp, (THREAD_FPR27-THREAD_FPR0)
+	lsx_save_upper $vr28, \base, \tmp, (THREAD_FPR28-THREAD_FPR0)
+	lsx_save_upper $vr29, \base, \tmp, (THREAD_FPR29-THREAD_FPR0)
+	lsx_save_upper $vr30, \base, \tmp, (THREAD_FPR30-THREAD_FPR0)
+	lsx_save_upper $vr31, \base, \tmp, (THREAD_FPR31-THREAD_FPR0)
+	.endm
+
+	.macro lsx_restore_upper vd base tmp off
+	parse_vr __vd, \vd
+	parse_r __tmp, \tmp
+	ld.d	\tmp, \base, (\off+8)
+	/* vinsgr2vr opcode is 0xe5d7e */
+	.word	(0xe5d7e << 11 | 1 << 10 | __tmp << 5 | __vd)
+	.endm
+
+	.macro lsx_restore_all_upper thread base tmp
+	li.w	\tmp, THREAD_FPR0
+	PTR_ADD	\base, \thread, \tmp
+	lsx_restore_upper $vr0,  \base, \tmp, (THREAD_FPR0-THREAD_FPR0)
+	lsx_restore_upper $vr1,  \base, \tmp, (THREAD_FPR1-THREAD_FPR0)
+	lsx_restore_upper $vr2,  \base, \tmp, (THREAD_FPR2-THREAD_FPR0)
+	lsx_restore_upper $vr3,  \base, \tmp, (THREAD_FPR3-THREAD_FPR0)
+	lsx_restore_upper $vr4,  \base, \tmp, (THREAD_FPR4-THREAD_FPR0)
+	lsx_restore_upper $vr5,  \base, \tmp, (THREAD_FPR5-THREAD_FPR0)
+	lsx_restore_upper $vr6,  \base, \tmp, (THREAD_FPR6-THREAD_FPR0)
+	lsx_restore_upper $vr7,  \base, \tmp, (THREAD_FPR7-THREAD_FPR0)
+	lsx_restore_upper $vr8,  \base, \tmp, (THREAD_FPR8-THREAD_FPR0)
+	lsx_restore_upper $vr9,  \base, \tmp, (THREAD_FPR9-THREAD_FPR0)
+	lsx_restore_upper $vr10, \base, \tmp, (THREAD_FPR10-THREAD_FPR0)
+	lsx_restore_upper $vr11, \base, \tmp, (THREAD_FPR11-THREAD_FPR0)
+	lsx_restore_upper $vr12, \base, \tmp, (THREAD_FPR12-THREAD_FPR0)
+	lsx_restore_upper $vr13, \base, \tmp, (THREAD_FPR13-THREAD_FPR0)
+	lsx_restore_upper $vr14, \base, \tmp, (THREAD_FPR14-THREAD_FPR0)
+	lsx_restore_upper $vr15, \base, \tmp, (THREAD_FPR15-THREAD_FPR0)
+	lsx_restore_upper $vr16, \base, \tmp, (THREAD_FPR16-THREAD_FPR0)
+	lsx_restore_upper $vr17, \base, \tmp, (THREAD_FPR17-THREAD_FPR0)
+	lsx_restore_upper $vr18, \base, \tmp, (THREAD_FPR18-THREAD_FPR0)
+	lsx_restore_upper $vr19, \base, \tmp, (THREAD_FPR19-THREAD_FPR0)
+	lsx_restore_upper $vr20, \base, \tmp, (THREAD_FPR20-THREAD_FPR0)
+	lsx_restore_upper $vr21, \base, \tmp, (THREAD_FPR21-THREAD_FPR0)
+	lsx_restore_upper $vr22, \base, \tmp, (THREAD_FPR22-THREAD_FPR0)
+	lsx_restore_upper $vr23, \base, \tmp, (THREAD_FPR23-THREAD_FPR0)
+	lsx_restore_upper $vr24, \base, \tmp, (THREAD_FPR24-THREAD_FPR0)
+	lsx_restore_upper $vr25, \base, \tmp, (THREAD_FPR25-THREAD_FPR0)
+	lsx_restore_upper $vr26, \base, \tmp, (THREAD_FPR26-THREAD_FPR0)
+	lsx_restore_upper $vr27, \base, \tmp, (THREAD_FPR27-THREAD_FPR0)
+	lsx_restore_upper $vr28, \base, \tmp, (THREAD_FPR28-THREAD_FPR0)
+	lsx_restore_upper $vr29, \base, \tmp, (THREAD_FPR29-THREAD_FPR0)
+	lsx_restore_upper $vr30, \base, \tmp, (THREAD_FPR30-THREAD_FPR0)
+	lsx_restore_upper $vr31, \base, \tmp, (THREAD_FPR31-THREAD_FPR0)
+	.endm
+
+	.macro	lsx_init_upper vd tmp
+	parse_vr __vd, \vd
+	parse_r __tmp, \tmp
+	/* vinsgr2vr opcode is 0xe5d7e */
+	.word	(0xe5d7e << 11 | 1 << 10 | __tmp << 5 | __vd)
+	.endm
+
+	.macro	lsx_init_all_upper tmp
+	not	\tmp, zero
+	lsx_init_upper	$vr0 \tmp
+	lsx_init_upper	$vr1 \tmp
+	lsx_init_upper	$vr2 \tmp
+	lsx_init_upper	$vr3 \tmp
+	lsx_init_upper	$vr4 \tmp
+	lsx_init_upper	$vr5 \tmp
+	lsx_init_upper	$vr6 \tmp
+	lsx_init_upper	$vr7 \tmp
+	lsx_init_upper	$vr8 \tmp
+	lsx_init_upper	$vr9 \tmp
+	lsx_init_upper	$vr10 \tmp
+	lsx_init_upper	$vr11 \tmp
+	lsx_init_upper	$vr12 \tmp
+	lsx_init_upper	$vr13 \tmp
+	lsx_init_upper	$vr14 \tmp
+	lsx_init_upper	$vr15 \tmp
+	lsx_init_upper	$vr16 \tmp
+	lsx_init_upper	$vr17 \tmp
+	lsx_init_upper	$vr18 \tmp
+	lsx_init_upper	$vr19 \tmp
+	lsx_init_upper	$vr20 \tmp
+	lsx_init_upper	$vr21 \tmp
+	lsx_init_upper	$vr22 \tmp
+	lsx_init_upper	$vr23 \tmp
+	lsx_init_upper	$vr24 \tmp
+	lsx_init_upper	$vr25 \tmp
+	lsx_init_upper	$vr26 \tmp
+	lsx_init_upper	$vr27 \tmp
+	lsx_init_upper	$vr28 \tmp
+	lsx_init_upper	$vr29 \tmp
+	lsx_init_upper	$vr30 \tmp
+	lsx_init_upper	$vr31 \tmp
+	.endm
+
+	.macro lasx_save_data thread tmp
+	parse_r __tmp, \tmp
+	li.w            \tmp, THREAD_FPR0
+	PTR_ADD         \tmp, \thread, \tmp
+	/* xvst opcode is 0xb3 */
+	.word (0xb3 << 22 | ((THREAD_FPR0-THREAD_FPR0) << 10) | __tmp << 5 | 0)
+	.word (0xb3 << 22 | ((THREAD_FPR1-THREAD_FPR0) << 10) | __tmp << 5 | 1)
+	.word (0xb3 << 22 | ((THREAD_FPR2-THREAD_FPR0) << 10) | __tmp << 5 | 2)
+	.word (0xb3 << 22 | ((THREAD_FPR3-THREAD_FPR0) << 10) | __tmp << 5 | 3)
+	.word (0xb3 << 22 | ((THREAD_FPR4-THREAD_FPR0) << 10) | __tmp << 5 | 4)
+	.word (0xb3 << 22 | ((THREAD_FPR5-THREAD_FPR0) << 10) | __tmp << 5 | 5)
+	.word (0xb3 << 22 | ((THREAD_FPR6-THREAD_FPR0) << 10) | __tmp << 5 | 6)
+	.word (0xb3 << 22 | ((THREAD_FPR7-THREAD_FPR0) << 10) | __tmp << 5 | 7)
+	.word (0xb3 << 22 | ((THREAD_FPR8-THREAD_FPR0) << 10) | __tmp << 5 | 8)
+	.word (0xb3 << 22 | ((THREAD_FPR9-THREAD_FPR0) << 10) | __tmp << 5 | 9)
+	.word (0xb3 << 22 | ((THREAD_FPR10-THREAD_FPR0) << 10) | __tmp << 5 | 10)
+	.word (0xb3 << 22 | ((THREAD_FPR11-THREAD_FPR0) << 10) | __tmp << 5 | 11)
+	.word (0xb3 << 22 | ((THREAD_FPR12-THREAD_FPR0) << 10) | __tmp << 5 | 12)
+	.word (0xb3 << 22 | ((THREAD_FPR13-THREAD_FPR0) << 10) | __tmp << 5 | 13)
+	.word (0xb3 << 22 | ((THREAD_FPR14-THREAD_FPR0) << 10) | __tmp << 5 | 14)
+	.word (0xb3 << 22 | ((THREAD_FPR15-THREAD_FPR0) << 10) | __tmp << 5 | 15)
+	.word (0xb3 << 22 | ((THREAD_FPR16-THREAD_FPR0) << 10) | __tmp << 5 | 16)
+	.word (0xb3 << 22 | ((THREAD_FPR17-THREAD_FPR0) << 10) | __tmp << 5 | 17)
+	.word (0xb3 << 22 | ((THREAD_FPR18-THREAD_FPR0) << 10) | __tmp << 5 | 18)
+	.word (0xb3 << 22 | ((THREAD_FPR19-THREAD_FPR0) << 10) | __tmp << 5 | 19)
+	.word (0xb3 << 22 | ((THREAD_FPR20-THREAD_FPR0) << 10) | __tmp << 5 | 20)
+	.word (0xb3 << 22 | ((THREAD_FPR21-THREAD_FPR0) << 10) | __tmp << 5 | 21)
+	.word (0xb3 << 22 | ((THREAD_FPR22-THREAD_FPR0) << 10) | __tmp << 5 | 22)
+	.word (0xb3 << 22 | ((THREAD_FPR23-THREAD_FPR0) << 10) | __tmp << 5 | 23)
+	.word (0xb3 << 22 | ((THREAD_FPR24-THREAD_FPR0) << 10) | __tmp << 5 | 24)
+	.word (0xb3 << 22 | ((THREAD_FPR25-THREAD_FPR0) << 10) | __tmp << 5 | 25)
+	.word (0xb3 << 22 | ((THREAD_FPR26-THREAD_FPR0) << 10) | __tmp << 5 | 26)
+	.word (0xb3 << 22 | ((THREAD_FPR27-THREAD_FPR0) << 10) | __tmp << 5 | 27)
+	.word (0xb3 << 22 | ((THREAD_FPR28-THREAD_FPR0) << 10) | __tmp << 5 | 28)
+	.word (0xb3 << 22 | ((THREAD_FPR29-THREAD_FPR0) << 10) | __tmp << 5 | 29)
+	.word (0xb3 << 22 | ((THREAD_FPR30-THREAD_FPR0) << 10) | __tmp << 5 | 30)
+	.word (0xb3 << 22 | ((THREAD_FPR31-THREAD_FPR0) << 10) | __tmp << 5 | 31)
+	.endm
+
+	.macro lasx_restore_data thread tmp
+	parse_r __tmp, \tmp
+	li.w            \tmp, THREAD_FPR0
+	PTR_ADD         \tmp, \thread, \tmp
+	/* xvld opcode is 0xb2 */
+	.word (0xb2 << 22 | ((THREAD_FPR0-THREAD_FPR0) << 10) | __tmp << 5 | 0)
+	.word (0xb2 << 22 | ((THREAD_FPR1-THREAD_FPR0) << 10) | __tmp << 5 | 1)
+	.word (0xb2 << 22 | ((THREAD_FPR2-THREAD_FPR0) << 10) | __tmp << 5 | 2)
+	.word (0xb2 << 22 | ((THREAD_FPR3-THREAD_FPR0) << 10) | __tmp << 5 | 3)
+	.word (0xb2 << 22 | ((THREAD_FPR4-THREAD_FPR0) << 10) | __tmp << 5 | 4)
+	.word (0xb2 << 22 | ((THREAD_FPR5-THREAD_FPR0) << 10) | __tmp << 5 | 5)
+	.word (0xb2 << 22 | ((THREAD_FPR6-THREAD_FPR0) << 10) | __tmp << 5 | 6)
+	.word (0xb2 << 22 | ((THREAD_FPR7-THREAD_FPR0) << 10) | __tmp << 5 | 7)
+	.word (0xb2 << 22 | ((THREAD_FPR8-THREAD_FPR0) << 10) | __tmp << 5 | 8)
+	.word (0xb2 << 22 | ((THREAD_FPR9-THREAD_FPR0) << 10) | __tmp << 5 | 9)
+	.word (0xb2 << 22 | ((THREAD_FPR10-THREAD_FPR0) << 10) | __tmp << 5 | 10)
+	.word (0xb2 << 22 | ((THREAD_FPR11-THREAD_FPR0) << 10) | __tmp << 5 | 11)
+	.word (0xb2 << 22 | ((THREAD_FPR12-THREAD_FPR0) << 10) | __tmp << 5 | 12)
+	.word (0xb2 << 22 | ((THREAD_FPR13-THREAD_FPR0) << 10) | __tmp << 5 | 13)
+	.word (0xb2 << 22 | ((THREAD_FPR14-THREAD_FPR0) << 10) | __tmp << 5 | 14)
+	.word (0xb2 << 22 | ((THREAD_FPR15-THREAD_FPR0) << 10) | __tmp << 5 | 15)
+	.word (0xb2 << 22 | ((THREAD_FPR16-THREAD_FPR0) << 10) | __tmp << 5 | 16)
+	.word (0xb2 << 22 | ((THREAD_FPR17-THREAD_FPR0) << 10) | __tmp << 5 | 17)
+	.word (0xb2 << 22 | ((THREAD_FPR18-THREAD_FPR0) << 10) | __tmp << 5 | 18)
+	.word (0xb2 << 22 | ((THREAD_FPR19-THREAD_FPR0) << 10) | __tmp << 5 | 19)
+	.word (0xb2 << 22 | ((THREAD_FPR20-THREAD_FPR0) << 10) | __tmp << 5 | 20)
+	.word (0xb2 << 22 | ((THREAD_FPR21-THREAD_FPR0) << 10) | __tmp << 5 | 21)
+	.word (0xb2 << 22 | ((THREAD_FPR22-THREAD_FPR0) << 10) | __tmp << 5 | 22)
+	.word (0xb2 << 22 | ((THREAD_FPR23-THREAD_FPR0) << 10) | __tmp << 5 | 23)
+	.word (0xb2 << 22 | ((THREAD_FPR24-THREAD_FPR0) << 10) | __tmp << 5 | 24)
+	.word (0xb2 << 22 | ((THREAD_FPR25-THREAD_FPR0) << 10) | __tmp << 5 | 25)
+	.word (0xb2 << 22 | ((THREAD_FPR26-THREAD_FPR0) << 10) | __tmp << 5 | 26)
+	.word (0xb2 << 22 | ((THREAD_FPR27-THREAD_FPR0) << 10) | __tmp << 5 | 27)
+	.word (0xb2 << 22 | ((THREAD_FPR28-THREAD_FPR0) << 10) | __tmp << 5 | 28)
+	.word (0xb2 << 22 | ((THREAD_FPR29-THREAD_FPR0) << 10) | __tmp << 5 | 29)
+	.word (0xb2 << 22 | ((THREAD_FPR30-THREAD_FPR0) << 10) | __tmp << 5 | 30)
+	.word (0xb2 << 22 | ((THREAD_FPR31-THREAD_FPR0) << 10) | __tmp << 5 | 31)
+	.endm
+
+	.macro	lasx_save_all	thread tmp0 tmp1
+	fpu_save_cc	\thread, \tmp0, \tmp1
+	fpu_save_csr	\thread, \tmp0
+	lasx_save_data	\thread, \tmp0
+	.endm
+
+	.macro	lasx_restore_all thread tmp0 tmp1
+	lasx_restore_data	\thread, \tmp0
+	fpu_restore_cc	\thread, \tmp0, \tmp1
+	fpu_restore_csr	\thread, \tmp0
+	.endm
+
+	.macro lasx_save_upper xd base tmp off
+	/* Nothing */
+	.endm
+
+	.macro lasx_save_all_upper thread base tmp
+	/* Nothing */
+	.endm
+
+	.macro lasx_restore_upper xd base tmp off
+	parse_xr __xd, \xd
+	parse_xr __xt, \tmp
+	parse_r __base, \base
+	/* vld opcode is 0xb0 */
+	.word (0xb0 << 22 | (\off+16) << 10 | __base << 5 | __xt)
+	/* xvpermi.q opcode is 0x1dfb */
+	.word (0x1dfb << 18 | 0x2 << 10 | __xt << 5 | __xd)
+	.endm
+
+	.macro lasx_restore_all_upper thread base tmp
+	li.w	\tmp, THREAD_FPR0
+	PTR_ADD	\base, \thread, \tmp
+	/* Save $vr31, xvpickve2gr opcode is 0x76efe */
+	.word (0x76efe << 12 | 0 << 10 | 31 << 5 | 0x11)
+	.word (0x76efe << 12 | 1 << 10 | 31 << 5 | 0x12)
+	lasx_restore_upper $xr0, \base, $xr31, (THREAD_FPR0-THREAD_FPR0)
+	lasx_restore_upper $xr1, \base, $xr31, (THREAD_FPR1-THREAD_FPR0)
+	lasx_restore_upper $xr2, \base, $xr31, (THREAD_FPR2-THREAD_FPR0)
+	lasx_restore_upper $xr3, \base, $xr31, (THREAD_FPR3-THREAD_FPR0)
+	lasx_restore_upper $xr4, \base, $xr31, (THREAD_FPR4-THREAD_FPR0)
+	lasx_restore_upper $xr5, \base, $xr31, (THREAD_FPR5-THREAD_FPR0)
+	lasx_restore_upper $xr6, \base, $xr31, (THREAD_FPR6-THREAD_FPR0)
+	lasx_restore_upper $xr7, \base, $xr31, (THREAD_FPR7-THREAD_FPR0)
+	lasx_restore_upper $xr8, \base, $xr31, (THREAD_FPR8-THREAD_FPR0)
+	lasx_restore_upper $xr9, \base, $xr31, (THREAD_FPR9-THREAD_FPR0)
+	lasx_restore_upper $xr10, \base, $xr31, (THREAD_FPR10-THREAD_FPR0)
+	lasx_restore_upper $xr11, \base, $xr31, (THREAD_FPR11-THREAD_FPR0)
+	lasx_restore_upper $xr12, \base, $xr31, (THREAD_FPR12-THREAD_FPR0)
+	lasx_restore_upper $xr13, \base, $xr31, (THREAD_FPR13-THREAD_FPR0)
+	lasx_restore_upper $xr14, \base, $xr31, (THREAD_FPR14-THREAD_FPR0)
+	lasx_restore_upper $xr15, \base, $xr31, (THREAD_FPR15-THREAD_FPR0)
+	lasx_restore_upper $xr16, \base, $xr31, (THREAD_FPR16-THREAD_FPR0)
+	lasx_restore_upper $xr17, \base, $xr31, (THREAD_FPR17-THREAD_FPR0)
+	lasx_restore_upper $xr18, \base, $xr31, (THREAD_FPR18-THREAD_FPR0)
+	lasx_restore_upper $xr19, \base, $xr31, (THREAD_FPR19-THREAD_FPR0)
+	lasx_restore_upper $xr20, \base, $xr31, (THREAD_FPR20-THREAD_FPR0)
+	lasx_restore_upper $xr21, \base, $xr31, (THREAD_FPR21-THREAD_FPR0)
+	lasx_restore_upper $xr22, \base, $xr31, (THREAD_FPR22-THREAD_FPR0)
+	lasx_restore_upper $xr23, \base, $xr31, (THREAD_FPR23-THREAD_FPR0)
+	lasx_restore_upper $xr24, \base, $xr31, (THREAD_FPR24-THREAD_FPR0)
+	lasx_restore_upper $xr25, \base, $xr31, (THREAD_FPR25-THREAD_FPR0)
+	lasx_restore_upper $xr26, \base, $xr31, (THREAD_FPR26-THREAD_FPR0)
+	lasx_restore_upper $xr27, \base, $xr31, (THREAD_FPR27-THREAD_FPR0)
+	lasx_restore_upper $xr28, \base, $xr31, (THREAD_FPR28-THREAD_FPR0)
+	lasx_restore_upper $xr29, \base, $xr31, (THREAD_FPR29-THREAD_FPR0)
+	lasx_restore_upper $xr30, \base, $xr31, (THREAD_FPR30-THREAD_FPR0)
+	lasx_restore_upper $xr31, \base, $xr31, (THREAD_FPR31-THREAD_FPR0)
+	/* Restore $vr31, xvinsgr2vr opcode is 0x76ebe */
+	.word (0x76ebe << 12 | 0 << 10 | 0x11 << 5 | 31)
+	.word (0x76ebe << 12 | 1 << 10 | 0x12 << 5 | 31)
+	.endm
+
+	.macro	lasx_init_upper xd tmp
+	parse_xr __xd, \xd
+	parse_r __tmp, \tmp
+	/* xvinsgr2vr opcode is 0x76ebe */
+	.word	(0x76ebe << 12 | 2 << 10 | __tmp << 5 | __xd)
+	.word	(0x76ebe << 12 | 3 << 10 | __tmp << 5 | __xd)
+	.endm
+
+	.macro	lasx_init_all_upper tmp
+	not	\tmp, zero
+	lasx_init_upper	$xr0 \tmp
+	lasx_init_upper	$xr1 \tmp
+	lasx_init_upper	$xr2 \tmp
+	lasx_init_upper	$xr3 \tmp
+	lasx_init_upper	$xr4 \tmp
+	lasx_init_upper	$xr5 \tmp
+	lasx_init_upper	$xr6 \tmp
+	lasx_init_upper	$xr7 \tmp
+	lasx_init_upper	$xr8 \tmp
+	lasx_init_upper	$xr9 \tmp
+	lasx_init_upper	$xr10 \tmp
+	lasx_init_upper	$xr11 \tmp
+	lasx_init_upper	$xr12 \tmp
+	lasx_init_upper	$xr13 \tmp
+	lasx_init_upper	$xr14 \tmp
+	lasx_init_upper	$xr15 \tmp
+	lasx_init_upper	$xr16 \tmp
+	lasx_init_upper	$xr17 \tmp
+	lasx_init_upper	$xr18 \tmp
+	lasx_init_upper	$xr19 \tmp
+	lasx_init_upper	$xr20 \tmp
+	lasx_init_upper	$xr21 \tmp
+	lasx_init_upper	$xr22 \tmp
+	lasx_init_upper	$xr23 \tmp
+	lasx_init_upper	$xr24 \tmp
+	lasx_init_upper	$xr25 \tmp
+	lasx_init_upper	$xr26 \tmp
+	lasx_init_upper	$xr27 \tmp
+	lasx_init_upper	$xr28 \tmp
+	lasx_init_upper	$xr29 \tmp
+	lasx_init_upper	$xr30 \tmp
+	lasx_init_upper	$xr31 \tmp
+	.endm
+
 .macro not dst src
 	nor	\dst, \src, zero
 .endm

arch/loongarch/include/asm/fpu.h

@@ -25,6 +25,26 @@ extern void _init_fpu(unsigned int);
 extern void _save_fp(struct loongarch_fpu *);
 extern void _restore_fp(struct loongarch_fpu *);
 
+extern void _save_lsx(struct loongarch_fpu *fpu);
+extern void _restore_lsx(struct loongarch_fpu *fpu);
+extern void _init_lsx_upper(void);
+extern void _restore_lsx_upper(struct loongarch_fpu *fpu);
+
+extern void _save_lasx(struct loongarch_fpu *fpu);
+extern void _restore_lasx(struct loongarch_fpu *fpu);
+extern void _init_lasx_upper(void);
+extern void _restore_lasx_upper(struct loongarch_fpu *fpu);
+
+static inline void enable_lsx(void);
+static inline void disable_lsx(void);
+static inline void save_lsx(struct task_struct *t);
+static inline void restore_lsx(struct task_struct *t);
+
+static inline void enable_lasx(void);
+static inline void disable_lasx(void);
+static inline void save_lasx(struct task_struct *t);
+static inline void restore_lasx(struct task_struct *t);
+
 /*
  * Mask the FCSR Cause bits according to the Enable bits, observing
  * that Unimplemented is always enabled.
@@ -41,6 +61,29 @@ static inline int is_fp_enabled(void)
 		1 : 0;
 }
 
+static inline int is_lsx_enabled(void)
+{
+	if (!cpu_has_lsx)
+		return 0;
+
+	return (csr_read32(LOONGARCH_CSR_EUEN) & CSR_EUEN_LSXEN) ?
+		1 : 0;
+}
+
+static inline int is_lasx_enabled(void)
+{
+	if (!cpu_has_lasx)
+		return 0;
+
+	return (csr_read32(LOONGARCH_CSR_EUEN) & CSR_EUEN_LASXEN) ?
+		1 : 0;
+}
+
+static inline int is_simd_enabled(void)
+{
+	return is_lsx_enabled() | is_lasx_enabled();
+}
+
 #define enable_fpu()		set_csr_euen(CSR_EUEN_FPEN)
 
 #define disable_fpu()		clear_csr_euen(CSR_EUEN_FPEN)
@@ -78,9 +121,22 @@ static inline void own_fpu(int restore)
 static inline void lose_fpu_inatomic(int save, struct task_struct *tsk)
 {
 	if (is_fpu_owner()) {
+		if (is_simd_enabled()) {
+			if (save) {
+				if (is_lasx_enabled())
+					save_lasx(tsk);
+				else
+					save_lsx(tsk);
+			}
+			disable_fpu();
+			disable_lsx();
+			disable_lasx();
+			clear_tsk_thread_flag(tsk, TIF_USEDSIMD);
+		} else {
 			if (save)
 				_save_fp(&tsk->thread.fpu);
 			disable_fpu();
+		}
 		clear_tsk_thread_flag(tsk, TIF_USEDFPU);
 	}
 	KSTK_EUEN(tsk) &= ~(CSR_EUEN_FPEN | CSR_EUEN_LSXEN | CSR_EUEN_LASXEN);
@@ -126,4 +182,140 @@ static inline union fpureg *get_fpu_regs(struct task_struct *tsk)
 	return tsk->thread.fpu.fpr;
 }
 
+enum {
+	CTX_LSX = 1,
+	CTX_LASX = 2,
+};
+
+static inline int is_simd_owner(void)
+{
+	return test_thread_flag(TIF_USEDSIMD);
+}
+
+#ifdef CONFIG_CPU_HAS_LSX
+
+static inline void enable_lsx(void)
+{
+	if (cpu_has_lsx)
+		csr_xchg32(CSR_EUEN_LSXEN, CSR_EUEN_LSXEN, LOONGARCH_CSR_EUEN);
+}
+
+static inline void disable_lsx(void)
+{
+	if (cpu_has_lsx)
+		csr_xchg32(0, CSR_EUEN_LSXEN, LOONGARCH_CSR_EUEN);
+}
+
+static inline void save_lsx(struct task_struct *t)
+{
+	if (cpu_has_lsx)
+		_save_lsx(&t->thread.fpu);
+}
+
+static inline void restore_lsx(struct task_struct *t)
+{
+	if (cpu_has_lsx)
+		_restore_lsx(&t->thread.fpu);
+}
+
+static inline void init_lsx_upper(void)
+{
+	/*
+	 * Check cpu_has_lsx only if it's a constant. This will allow the
+	 * compiler to optimise out code for CPUs without LSX without adding
+	 * an extra redundant check for CPUs with LSX.
+	 */
+	if (__builtin_constant_p(cpu_has_lsx) && !cpu_has_lsx)
+		return;
+
+	_init_lsx_upper();
+}
+
+static inline void restore_lsx_upper(struct task_struct *t)
+{
+	if (cpu_has_lsx)
+		_restore_lsx_upper(&t->thread.fpu);
+}
+
+#else
+static inline void enable_lsx(void) {}
+static inline void disable_lsx(void) {}
+static inline void save_lsx(struct task_struct *t) {}
+static inline void restore_lsx(struct task_struct *t) {}
+static inline void init_lsx_upper(void) {}
+static inline void restore_lsx_upper(struct task_struct *t) {}
+#endif
+
+#ifdef CONFIG_CPU_HAS_LASX
+
+static inline void enable_lasx(void)
+{
+
+	if (cpu_has_lasx)
+		csr_xchg32(CSR_EUEN_LASXEN, CSR_EUEN_LASXEN, LOONGARCH_CSR_EUEN);
+}
+
+static inline void disable_lasx(void)
+{
+	if (cpu_has_lasx)
+		csr_xchg32(0, CSR_EUEN_LASXEN, LOONGARCH_CSR_EUEN);
+}
+
+static inline void save_lasx(struct task_struct *t)
+{
+	if (cpu_has_lasx)
+		_save_lasx(&t->thread.fpu);
+}
+
+static inline void restore_lasx(struct task_struct *t)
+{
+	if (cpu_has_lasx)
+		_restore_lasx(&t->thread.fpu);
+}
+
+static inline void init_lasx_upper(void)
+{
+	if (cpu_has_lasx)
+		_init_lasx_upper();
+}
+
+static inline void restore_lasx_upper(struct task_struct *t)
+{
+	if (cpu_has_lasx)
+		_restore_lasx_upper(&t->thread.fpu);
+}
+
+#else
+static inline void enable_lasx(void) {}
+static inline void disable_lasx(void) {}
+static inline void save_lasx(struct task_struct *t) {}
+static inline void restore_lasx(struct task_struct *t) {}
+static inline void init_lasx_upper(void) {}
+static inline void restore_lasx_upper(struct task_struct *t) {}
+#endif
+
+static inline int thread_lsx_context_live(void)
+{
+	int ret = 0;
+
+	if (__builtin_constant_p(cpu_has_lsx) && !cpu_has_lsx)
+		goto  out;
+
+	ret =  test_thread_flag(TIF_LSX_CTX_LIVE) ? CTX_LSX : 0;
+out:
+	return ret;
+}
+
+static inline int thread_lasx_context_live(void)
+{
+	int ret = 0;
+
+	if (__builtin_constant_p(cpu_has_lasx) && !cpu_has_lasx)
+		goto out;
+
+	ret = test_thread_flag(TIF_LASX_CTX_LIVE) ? CTX_LASX : 0;
+out:
+	return ret;
+}
+
 #endif /* _ASM_FPU_H */

arch/loongarch/kernel/cpu-probe.c

@@ -111,6 +111,18 @@ static void cpu_probe_common(struct cpuinfo_loongarch *c)
 		c->options |= LOONGARCH_CPU_FPU;
 		elf_hwcap |= HWCAP_LOONGARCH_FPU;
 	}
+#ifdef CONFIG_CPU_HAS_LSX
+	if (config & CPUCFG2_LSX) {
+		c->options |= LOONGARCH_CPU_LSX;
+		elf_hwcap |= HWCAP_LOONGARCH_LSX;
+	}
+#endif
+#ifdef CONFIG_CPU_HAS_LASX
+	if (config & CPUCFG2_LASX) {
+		c->options |= LOONGARCH_CPU_LASX;
+		elf_hwcap |= HWCAP_LOONGARCH_LASX;
+	}
+#endif
 	if (config & CPUCFG2_COMPLEX) {
 		c->options |= LOONGARCH_CPU_COMPLEX;
 		elf_hwcap |= HWCAP_LOONGARCH_COMPLEX;

arch/loongarch/kernel/fpu.S

@@ -26,6 +26,32 @@
 	.previous
 	.endm
 
+	.macro	EX_V insn, reg, src, offs
+	parse_v __insn, \insn
+	parse_v __offs, \offs
+	parse_r __src, \src
+	parse_vr __reg, \reg
+
+.ex\@:
+	.word __insn << 22 | __offs << 10 | __src << 5 | __reg
+	.section __ex_table,"a"
+	PTR	.ex\@, fault
+	.previous
+	.endm
+
+	.macro	EX_XV insn, reg, src, offs
+	parse_v __insn, \insn
+	parse_v __offs, \offs
+	parse_r __src, \src
+	parse_xr __reg, \reg
+
+.ex\@:
+	.word __insn << 22 | __offs << 10 | __src << 5 | __reg
+	.section __ex_table,"a"
+	PTR	.ex\@, fault
+	.previous
+	.endm
+
 	.macro sc_save_fp base
 	EX	fst.d	$f0,  \base, (0 * FPU_REG_WIDTH)
 	EX	fst.d	$f1,  \base, (1 * FPU_REG_WIDTH)
@@ -146,6 +172,146 @@
 	movgr2fcsr	fcsr0, \tmp0
 	.endm
 
+	.macro sc_save_lsx base
+	EX_V 0xb1 $vr0,  \base,	(0 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr1,  \base,	(1 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr2,  \base,	(2 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr3,  \base,	(3 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr4,  \base,	(4 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr5,  \base,	(5 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr6,  \base,	(6 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr7,  \base,	(7 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr8,  \base,	(8 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr9,  \base,	(9 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr10,  \base, (10 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr11,  \base, (11 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr12,  \base, (12 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr13,  \base, (13 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr14,  \base, (14 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr15,  \base, (15 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr16,  \base, (16 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr17,  \base, (17 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr18,  \base, (18 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr19,  \base, (19 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr20,  \base, (20 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr21,  \base, (21 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr22,  \base, (22 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr23,  \base, (23 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr24,  \base, (24 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr25,  \base, (25 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr26,  \base, (26 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr27,  \base, (27 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr28,  \base, (28 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr29,  \base, (29 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr30,  \base, (30 * LSX_REG_WIDTH)
+	EX_V 0xb1 $vr31,  \base, (31 * LSX_REG_WIDTH)
+	.endm
+
+	.macro sc_restore_lsx base
+	EX_V 0xb0 $vr0,  \base, (0 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr1,  \base, (1 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr2,  \base, (2 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr3,  \base, (3 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr4,  \base, (4 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr5,  \base, (5 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr6,  \base, (6 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr7,  \base, (7 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr8,  \base, (8 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr9,  \base, (9 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr10,  \base, (10 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr11,  \base, (11 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr12,  \base, (12 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr13,  \base, (13 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr14,  \base, (14 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr15,  \base, (15 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr16,  \base, (16 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr17,  \base, (17 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr18,  \base, (18 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr19,  \base, (19 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr20,  \base, (20 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr21,  \base, (21 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr22,  \base, (22 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr23,  \base, (23 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr24,  \base, (24 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr25,  \base, (25 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr26,  \base, (26 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr27,  \base, (27 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr28,  \base, (28 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr29,  \base, (29 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr30,  \base, (30 * LSX_REG_WIDTH)
+	EX_V 0xb0 $vr31,  \base, (31 * LSX_REG_WIDTH)
+	.endm
+
+	.macro sc_save_lasx base
+	EX_XV	0xb3 $xr0,  \base, (0 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr1,  \base, (1 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr2,  \base, (2 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr3,  \base, (3 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr4,  \base, (4 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr5,  \base, (5 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr6,  \base, (6 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr7,  \base, (7 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr8,  \base, (8 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr9,  \base, (9 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr10, \base, (10 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr11, \base, (11 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr12, \base, (12 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr13, \base, (13 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr14, \base, (14 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr15, \base, (15 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr16, \base, (16 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr17, \base, (17 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr18, \base, (18 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr19, \base, (19 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr20, \base, (20 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr21, \base, (21 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr22, \base, (22 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr23, \base, (23 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr24, \base, (24 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr25, \base, (25 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr26, \base, (26 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr27, \base, (27 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr28, \base, (28 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr29, \base, (29 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr30, \base, (30 * LASX_REG_WIDTH)
+	EX_XV	0xb3 $xr31, \base, (31 * LASX_REG_WIDTH)
+	.endm
+
+	.macro sc_restore_lasx base
+	EX_XV	0xb2 $xr0,  \base, (0 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr1,  \base, (1 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr2,  \base, (2 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr3,  \base, (3 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr4,  \base, (4 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr5,  \base, (5 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr6,  \base, (6 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr7,  \base, (7 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr8,  \base, (8 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr9,  \base, (9 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr10, \base, (10 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr11, \base, (11 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr12, \base, (12 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr13, \base, (13 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr14, \base, (14 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr15, \base, (15 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr16, \base, (16 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr17, \base, (17 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr18, \base, (18 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr19, \base, (19 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr20, \base, (20 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr21, \base, (21 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr22, \base, (22 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr23, \base, (23 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr24, \base, (24 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr25, \base, (25 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr26, \base, (26 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr27, \base, (27 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr28, \base, (28 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr29, \base, (29 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr30, \base, (30 * LASX_REG_WIDTH)
+	EX_XV	0xb2 $xr31, \base, (31 * LASX_REG_WIDTH)
+	.endm
+
 /*
  * Save a thread's fp context.
  */
@@ -167,6 +333,76 @@ SYM_FUNC_START(_restore_fp)
 	jr			ra
 SYM_FUNC_END(_restore_fp)
 
+#ifdef CONFIG_CPU_HAS_LSX
+
+/*
+ * Save a thread's LSX vector context.
+ */
+SYM_FUNC_START(_save_lsx)
+	lsx_save_all	a0 t1 t2
+	jirl zero, ra, 0
+SYM_FUNC_END(_save_lsx)
+EXPORT_SYMBOL(_save_lsx)
+
+/*
+ * Restore a thread's LSX vector context.
+ */
+SYM_FUNC_START(_restore_lsx)
+	lsx_restore_all	a0 t1 t2
+	jirl zero, ra, 0
+SYM_FUNC_END(_restore_lsx)
+
+SYM_FUNC_START(_save_lsx_upper)
+	lsx_save_all_upper a0 t0 t1
+	jirl zero, ra, 0
+SYM_FUNC_END(_save_lsx_upper)
+
+SYM_FUNC_START(_restore_lsx_upper)
+	lsx_restore_all_upper a0 t0 t1
+	jirl zero, ra, 0
+SYM_FUNC_END(_restore_lsx_upper)
+
+SYM_FUNC_START(_init_lsx_upper)
+	lsx_init_all_upper t1
+	jirl zero, ra, 0
+SYM_FUNC_END(_init_lsx_upper)
+#endif
+
+#ifdef CONFIG_CPU_HAS_LASX
+
+/*
+ * Save a thread's LASX vector context.
+ */
+SYM_FUNC_START(_save_lasx)
+	lasx_save_all	a0 t1 t2
+	jirl zero, ra, 0
+SYM_FUNC_END(_save_lasx)
+EXPORT_SYMBOL(_save_lasx)
+
+/*
+ * Restore a thread's LASX vector context.
+ */
+SYM_FUNC_START(_restore_lasx)
+	lasx_restore_all a0 t1 t2
+	jirl zero, ra, 0
+SYM_FUNC_END(_restore_lasx)
+
+SYM_FUNC_START(_save_lasx_upper)
+	lasx_save_all_upper a0 t0 t1
+	jirl zero, ra, 0
+SYM_FUNC_END(_save_lasx_upper)
+
+SYM_FUNC_START(_restore_lasx_upper)
+	lasx_restore_all_upper a0 t0 t1
+	jirl	zero, ra, 0
+SYM_FUNC_END(_restore_lasx_upper)
+
+SYM_FUNC_START(_init_lasx_upper)
+	lasx_init_all_upper t1
+	jirl	zero, ra, 0
+SYM_FUNC_END(_init_lasx_upper)
+#endif
+
 /*
  * Load the FPU with signalling NANS.  This bit pattern we're using has
  * the property that no matter whether considered as single or as double
@@ -245,6 +481,58 @@ SYM_FUNC_START(_restore_fp_context)
 	jr		ra
 SYM_FUNC_END(_restore_fp_context)
 
+/*
+ * a0: fpregs
+ * a1: fcc
+ * a2: fcsr
+ */
+SYM_FUNC_START(_save_lsx_context)
+	sc_save_fcc a1, t0, t1
+	sc_save_fcsr a2, t0
+	sc_save_lsx a0
+	li.w	a0, 0					# success
+	jirl zero, ra, 0
+SYM_FUNC_END(_save_lsx_context)
+
+/*
+ * a0: fpregs
+ * a1: fcc
+ * a2: fcsr
+ */
+SYM_FUNC_START(_restore_lsx_context)
+	sc_restore_lsx a0
+	sc_restore_fcc a1, t1, t2
+	sc_restore_fcsr a2, t1
+	li.w	a0, 0					# success
+	jirl zero, ra, 0
+SYM_FUNC_END(_restore_lsx_context)
+
+/*
+ * a0: fpregs
+ * a1: fcc
+ * a2: fcsr
+ */
+SYM_FUNC_START(_save_lasx_context)
+	sc_save_fcc a1, t0, t1
+	sc_save_fcsr a2, t0
+	sc_save_lasx a0
+	li.w	a0, 0					# success
+	jirl zero, ra, 0
+SYM_FUNC_END(_save_lasx_context)
+
+/*
+ * a0: fpregs
+ * a1: fcc
+ * a2: fcsr
+ */
+SYM_FUNC_START(_restore_lasx_context)
+	sc_restore_lasx a0
+	sc_restore_fcc a1, t1, t2
+	sc_restore_fcsr a2, t1
+	li.w	a0, 0					# success
+	jirl zero, ra, 0
+SYM_FUNC_END(_restore_lasx_context)
+
 SYM_FUNC_START(fault)
 	li.w	a0, -EFAULT				# failure
 	jr	ra

arch/loongarch/kernel/process.c

@@ -105,8 +105,14 @@ int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
 	 */
 	preempt_disable();
 
-	if (is_fpu_owner())
+	if (is_fpu_owner()) {
+		if (is_lasx_enabled())
+			save_lasx(current);
+		else if (is_lsx_enabled())
+			save_lsx(current);
+		else
 			save_fp(current);
+	}
 
 	preempt_enable();
 

arch/loongarch/kernel/ptrace.c

@@ -246,6 +246,90 @@ static int cfg_set(struct task_struct *target,
 	return 0;
 }
 
+#ifdef CONFIG_CPU_HAS_LSX
+
+static void copy_pad_fprs(struct task_struct *target,
+			 const struct user_regset *regset,
+			 struct membuf *to, unsigned int live_sz)
+{
+	int i, j;
+	unsigned long long fill = ~0ull;
+	unsigned int cp_sz, pad_sz;
+
+	cp_sz = min(regset->size, live_sz);
+	pad_sz = regset->size - cp_sz;
+	WARN_ON(pad_sz % sizeof(fill));
+
+	for (i = 0; i < NUM_FPU_REGS; i++) {
+		membuf_write(to, &target->thread.fpu.fpr[i], cp_sz);
+		for (j = 0; j < (pad_sz / sizeof(fill)); j++) {
+			membuf_store(to, fill);
+		}
+	}
+}
+
+static int simd_get(struct task_struct *target,
+		    const struct user_regset *regset,
+		    struct membuf to)
+{
+	const unsigned int wr_size = NUM_FPU_REGS * regset->size;
+
+	if (!tsk_used_math(target)) {
+		/* The task hasn't used FP or LSX, fill with 0xff */
+		copy_pad_fprs(target, regset, &to, 0);
+	} else if (!test_tsk_thread_flag(target, TIF_LSX_CTX_LIVE)) {
+		/* Copy scalar FP context, fill the rest with 0xff */
+		copy_pad_fprs(target, regset, &to, 8);
+#ifdef CONFIG_CPU_HAS_LASX
+	} else if (!test_tsk_thread_flag(target, TIF_LASX_CTX_LIVE)) {
+		/* Copy LSX 128 Bit context, fill the rest with 0xff */
+		copy_pad_fprs(target, regset, &to, 16);
+#endif
+	} else if (sizeof(target->thread.fpu.fpr[0]) == regset->size) {
+		/* Trivially copy the vector registers */
+		membuf_write(&to, &target->thread.fpu.fpr, wr_size);
+	} else {
+		/* Copy as much context as possible, fill the rest with 0xff */
+		copy_pad_fprs(target, regset, &to, sizeof(target->thread.fpu.fpr[0]));
+	}
+
+	return 0;
+}
+
+static int simd_set(struct task_struct *target,
+		    const struct user_regset *regset,
+		    unsigned int pos, unsigned int count,
+		    const void *kbuf, const void __user *ubuf)
+{
+	const unsigned int wr_size = NUM_FPU_REGS * regset->size;
+	unsigned int cp_sz;
+	int i, err, start;
+
+	init_fp_ctx(target);
+
+	if (sizeof(target->thread.fpu.fpr[0]) == regset->size) {
+		/* Trivially copy the vector registers */
+		err = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
+					 &target->thread.fpu.fpr,
+					 0, wr_size);
+	} else {
+		/* Copy as much context as possible */
+		cp_sz = min_t(unsigned int, regset->size,
+			      sizeof(target->thread.fpu.fpr[0]));
+
+		i = start = err = 0;
+		for (; i < NUM_FPU_REGS; i++, start += regset->size) {
+			err |= user_regset_copyin(&pos, &count, &kbuf, &ubuf,
+						  &target->thread.fpu.fpr[i],
+						  start, start + cp_sz);
+		}
+	}
+
+	return err;
+}
+
+#endif /* CONFIG_CPU_HAS_LSX */
+
 struct pt_regs_offset {
 	const char *name;
 	int offset;
@@ -319,6 +403,12 @@ enum loongarch_regset {
 	REGSET_GPR,
 	REGSET_FPR,
 	REGSET_CPUCFG,
+#ifdef CONFIG_CPU_HAS_LSX
+	REGSET_LSX,
+#endif
+#ifdef CONFIG_CPU_HAS_LASX
+	REGSET_LASX,
+#endif
 };
 
 static const struct user_regset loongarch64_regsets[] = {
@@ -346,6 +436,26 @@ static const struct user_regset loongarch64_regsets[] = {
 		.regset_get	= cfg_get,
 		.set		= cfg_set,
 	},
+#ifdef CONFIG_CPU_HAS_LSX
+	[REGSET_LSX] = {
+		.core_note_type	= NT_LOONGARCH_LSX,
+		.n		= NUM_FPU_REGS,
+		.size		= 16,
+		.align		= 16,
+		.regset_get	= simd_get,
+		.set		= simd_set,
+	},
+#endif
+#ifdef CONFIG_CPU_HAS_LASX
+	[REGSET_LASX] = {
+		.core_note_type	= NT_LOONGARCH_LASX,
+		.n		= NUM_FPU_REGS,
+		.size		= 32,
+		.align		= 32,
+		.regset_get	= simd_get,
+		.set		= simd_set,
+	},
+#endif
 };
 
 static const struct user_regset_view user_loongarch64_view = {

arch/loongarch/kernel/signal.c

@@ -50,6 +50,16 @@ extern asmlinkage int
 _save_fp_context(void __user *fpregs, void __user *fcc, void __user *csr);
 extern asmlinkage int
 _restore_fp_context(void __user *fpregs, void __user *fcc, void __user *csr);
+extern asmlinkage int
+_save_lsx_context(void __user *fpregs, void __user *fcc, void __user *fcsr);
+extern asmlinkage int
+_restore_lsx_context(void __user *fpregs, void __user *fcc, void __user *fcsr);
+extern asmlinkage int
+_save_lasx_context(void __user *fpregs, void __user *fcc, void __user *fcsr);
+extern asmlinkage int
+_restore_lasx_context(void __user *fpregs, void __user *fcc, void __user *fcsr);
+extern asmlinkage int _save_lsx_all_upper(void __user *buf);
+extern asmlinkage int _restore_lsx_all_upper(void __user *buf);
 
 struct rt_sigframe {
 	struct siginfo rs_info;
@@ -65,9 +75,29 @@ struct extctx_layout {
 	unsigned long size;
 	unsigned int flags;
 	struct _ctx_layout fpu;
+	struct _ctx_layout lsx;
+	struct _ctx_layout lasx;
 	struct _ctx_layout end;
 };
 
+/* LSX context */
+#define LSX_CTX_MAGIC		0x53580001
+#define LSX_CTX_ALIGN		16
+struct lsx_context {
+	__u64	regs[2*32];
+	__u64	fcc;
+	__u32	fcsr;
+};
+
+/* LASX context */
+#define LASX_CTX_MAGIC		0x41535801
+#define LASX_CTX_ALIGN		32
+struct lasx_context {
+	__u64	regs[4*32];
+	__u64	fcc;
+	__u32	fcsr;
+};
+
 static void __user *get_ctx_through_ctxinfo(struct sctx_info *info)
 {
 	return (void __user *)((char *)info + sizeof(struct sctx_info));
@@ -115,6 +145,96 @@ static int copy_fpu_from_sigcontext(struct fpu_context __user *ctx)
 	return err;
 }
 
+static int copy_lsx_to_sigcontext(struct lsx_context __user *ctx)
+{
+	int i;
+	int err = 0;
+	uint64_t __user *regs	= (uint64_t *)&ctx->regs;
+	uint64_t __user *fcc	= &ctx->fcc;
+	uint32_t __user *fcsr	= &ctx->fcsr;
+
+	for (i = 0; i < NUM_FPU_REGS; i++) {
+		err |= __put_user(get_fpr64(&current->thread.fpu.fpr[i], 0),
+				  &regs[2*i]);
+		err |= __put_user(get_fpr64(&current->thread.fpu.fpr[i], 1),
+				  &regs[2*i+1]);
+	}
+	err |= __put_user(current->thread.fpu.fcc, fcc);
+	err |= __put_user(current->thread.fpu.fcsr, fcsr);
+
+	return err;
+}
+
+static int copy_lsx_from_sigcontext(struct lsx_context __user *ctx)
+{
+	int i;
+	int err = 0;
+	u64 fpr_val;
+	uint64_t __user *regs	= (uint64_t *)&ctx->regs;
+	uint64_t __user *fcc	= &ctx->fcc;
+	uint32_t __user *fcsr	= &ctx->fcsr;
+
+	for (i = 0; i < NUM_FPU_REGS; i++) {
+		err |= __get_user(fpr_val, &regs[2*i]);
+		set_fpr64(&current->thread.fpu.fpr[i], 0, fpr_val);
+		err |= __get_user(fpr_val, &regs[2*i+1]);
+		set_fpr64(&current->thread.fpu.fpr[i], 1, fpr_val);
+	}
+	err |= __get_user(current->thread.fpu.fcc, fcc);
+	err |= __get_user(current->thread.fpu.fcsr, fcsr);
+
+	return err;
+}
+
+static int copy_lasx_to_sigcontext(struct lasx_context __user *ctx)
+{
+	int i;
+	int err = 0;
+	uint64_t __user *regs	= (uint64_t *)&ctx->regs;
+	uint64_t __user *fcc	= &ctx->fcc;
+	uint32_t __user *fcsr	= &ctx->fcsr;
+
+	for (i = 0; i < NUM_FPU_REGS; i++) {
+		err |= __put_user(get_fpr64(&current->thread.fpu.fpr[i], 0),
+				  &regs[4*i]);
+		err |= __put_user(get_fpr64(&current->thread.fpu.fpr[i], 1),
+				  &regs[4*i+1]);
+		err |= __put_user(get_fpr64(&current->thread.fpu.fpr[i], 2),
+				  &regs[4*i+2]);
+		err |= __put_user(get_fpr64(&current->thread.fpu.fpr[i], 3),
+				  &regs[4*i+3]);
+	}
+	err |= __put_user(current->thread.fpu.fcc, fcc);
+	err |= __put_user(current->thread.fpu.fcsr, fcsr);
+
+	return err;
+}
+
+static int copy_lasx_from_sigcontext(struct lasx_context __user *ctx)
+{
+	int i;
+	int err = 0;
+	u64 fpr_val;
+	uint64_t __user *regs	= (uint64_t *)&ctx->regs;
+	uint64_t __user *fcc	= &ctx->fcc;
+	uint32_t __user *fcsr	= &ctx->fcsr;
+
+	for (i = 0; i < NUM_FPU_REGS; i++) {
+		err |= __get_user(fpr_val, &regs[4*i]);
+		set_fpr64(&current->thread.fpu.fpr[i], 0, fpr_val);
+		err |= __get_user(fpr_val, &regs[4*i+1]);
+		set_fpr64(&current->thread.fpu.fpr[i], 1, fpr_val);
+		err |= __get_user(fpr_val, &regs[4*i+2]);
+		set_fpr64(&current->thread.fpu.fpr[i], 2, fpr_val);
+		err |= __get_user(fpr_val, &regs[4*i+3]);
+		set_fpr64(&current->thread.fpu.fpr[i], 3, fpr_val);
+	}
+	err |= __get_user(current->thread.fpu.fcc, fcc);
+	err |= __get_user(current->thread.fpu.fcsr, fcsr);
+
+	return err;
+}
+
 /*
  * Wrappers for the assembly _{save,restore}_fp_context functions.
  */
@@ -136,6 +256,42 @@ static int restore_hw_fpu_context(struct fpu_context __user *ctx)
 	return _restore_fp_context(regs, fcc, fcsr);
 }
 
+static int save_hw_lsx_context(struct lsx_context __user *ctx)
+{
+	uint64_t __user *regs	= (uint64_t *)&ctx->regs;
+	uint64_t __user *fcc	= &ctx->fcc;
+	uint32_t __user *fcsr	= &ctx->fcsr;
+
+	return _save_lsx_context(regs, fcc, fcsr);
+}
+
+static int restore_hw_lsx_context(struct lsx_context __user *ctx)
+{
+	uint64_t __user *regs	= (uint64_t *)&ctx->regs;
+	uint64_t __user *fcc	= &ctx->fcc;
+	uint32_t __user *fcsr	= &ctx->fcsr;
+
+	return _restore_lsx_context(regs, fcc, fcsr);
+}
+
+static int save_hw_lasx_context(struct lasx_context __user *ctx)
+{
+	uint64_t __user *regs	= (uint64_t *)&ctx->regs;
+	uint64_t __user *fcc	= &ctx->fcc;
+	uint32_t __user *fcsr	= &ctx->fcsr;
+
+	return _save_lasx_context(regs, fcc, fcsr);
+}
+
+static int restore_hw_lasx_context(struct lasx_context __user *ctx)
+{
+	uint64_t __user *regs	= (uint64_t *)&ctx->regs;
+	uint64_t __user *fcc	= &ctx->fcc;
+	uint32_t __user *fcsr	= &ctx->fcsr;
+
+	return _restore_lasx_context(regs, fcc, fcsr);
+}
+
 static int fcsr_pending(unsigned int __user *fcsr)
 {
 	int err, sig = 0;
@@ -227,6 +383,146 @@ static int protected_restore_fpu_context(struct extctx_layout *extctx)
 	return err ?: sig;
 }
 
+static int protected_save_lsx_context(struct extctx_layout *extctx)
+{
+	int err = 0;
+	struct sctx_info __user *info = extctx->lsx.addr;
+	struct lsx_context __user *lsx_ctx = (struct lsx_context *)get_ctx_through_ctxinfo(info);
+	uint64_t __user *regs	= (uint64_t *)&lsx_ctx->regs;
+	uint64_t __user *fcc	= &lsx_ctx->fcc;
+	uint32_t __user *fcsr	= &lsx_ctx->fcsr;
+
+	while (1) {
+		lock_fpu_owner();
+		if (is_lsx_enabled())
+			err = save_hw_lsx_context(lsx_ctx);
+		else
+			err = copy_lsx_to_sigcontext(lsx_ctx);
+		unlock_fpu_owner();
+
+		err |= __put_user(LSX_CTX_MAGIC, &info->magic);
+		err |= __put_user(extctx->lsx.size, &info->size);
+
+		if (likely(!err))
+			break;
+		/* Touch the LSX context and try again */
+		err = __put_user(0, &regs[0]) |
+			__put_user(0, &regs[32*2-1]) |
+			__put_user(0, fcc) |
+			__put_user(0, fcsr);
+		if (err)
+			return err;	/* really bad sigcontext */
+	}
+
+	return err;
+}
+
+static int protected_restore_lsx_context(struct extctx_layout *extctx)
+{
+	int err = 0, sig = 0, tmp __maybe_unused;
+	struct sctx_info __user *info = extctx->lsx.addr;
+	struct lsx_context __user *lsx_ctx = (struct lsx_context *)get_ctx_through_ctxinfo(info);
+	uint64_t __user *regs	= (uint64_t *)&lsx_ctx->regs;
+	uint64_t __user *fcc	= &lsx_ctx->fcc;
+	uint32_t __user *fcsr	= &lsx_ctx->fcsr;
+
+	err = sig = fcsr_pending(fcsr);
+	if (err < 0)
+		return err;
+
+	while (1) {
+		lock_fpu_owner();
+		if (is_lsx_enabled())
+			err = restore_hw_lsx_context(lsx_ctx);
+		else
+			err = copy_lsx_from_sigcontext(lsx_ctx);
+		unlock_fpu_owner();
+
+		if (likely(!err))
+			break;
+		/* Touch the LSX context and try again */
+		err = __get_user(tmp, &regs[0]) |
+			__get_user(tmp, &regs[32*2-1]) |
+			__get_user(tmp, fcc) |
+			__get_user(tmp, fcsr);
+		if (err)
+			break;	/* really bad sigcontext */
+	}
+
+	return err ?: sig;
+}
+
+static int protected_save_lasx_context(struct extctx_layout *extctx)
+{
+	int err = 0;
+	struct sctx_info __user *info = extctx->lasx.addr;
+	struct lasx_context __user *lasx_ctx =
+		(struct lasx_context *)get_ctx_through_ctxinfo(info);
+	uint64_t __user *regs	= (uint64_t *)&lasx_ctx->regs;
+	uint64_t __user *fcc	= &lasx_ctx->fcc;
+	uint32_t __user *fcsr	= &lasx_ctx->fcsr;
+
+	while (1) {
+		lock_fpu_owner();
+		if (is_lasx_enabled())
+			err = save_hw_lasx_context(lasx_ctx);
+		else
+			err = copy_lasx_to_sigcontext(lasx_ctx);
+		unlock_fpu_owner();
+
+		err |= __put_user(LASX_CTX_MAGIC, &info->magic);
+		err |= __put_user(extctx->lasx.size, &info->size);
+
+		if (likely(!err))
+			break;
+		/* Touch the LASX context and try again */
+		err = __put_user(0, &regs[0]) |
+			__put_user(0, &regs[32*4-1]) |
+			__put_user(0, fcc) |
+			__put_user(0, fcsr);
+		if (err)
+			return err;	/* really bad sigcontext */
+	}
+
+	return err;
+}
+
+static int protected_restore_lasx_context(struct extctx_layout *extctx)
+{
+	int err = 0, sig = 0, tmp __maybe_unused;
+	struct sctx_info __user *info = extctx->lasx.addr;
+	struct lasx_context __user *lasx_ctx =
+		(struct lasx_context *)get_ctx_through_ctxinfo(info);
+	uint64_t __user *regs	= (uint64_t *)&lasx_ctx->regs;
+	uint64_t __user *fcc	= &lasx_ctx->fcc;
+	uint32_t __user *fcsr	= &lasx_ctx->fcsr;
+
+	err = sig = fcsr_pending(fcsr);
+	if (err < 0)
+		return err;
+
+	while (1) {
+		lock_fpu_owner();
+		if (is_lasx_enabled())
+			err = restore_hw_lasx_context(lasx_ctx);
+		else
+			err = copy_lasx_from_sigcontext(lasx_ctx);
+		unlock_fpu_owner();
+
+		if (likely(!err))
+			break;
+		/* Touch the LASX context and try again */
+		err = __get_user(tmp, &regs[0]) |
+			__get_user(tmp, &regs[32*4-1]) |
+			__get_user(tmp, fcc) |
+			__get_user(tmp, fcsr);
+		if (err)
+			break;	/* really bad sigcontext */
+	}
+
+	return err ?: sig;
+}
+
 static int setup_sigcontext(struct pt_regs *regs, struct sigcontext __user *sc,
 			    struct extctx_layout *extctx)
 {
@@ -240,7 +536,11 @@ static int setup_sigcontext(struct pt_regs *regs, struct sigcontext __user *sc,
 	for (i = 1; i < 32; i++)
 		err |= __put_user(regs->regs[i], &sc->sc_regs[i]);
 
-	if (extctx->fpu.addr)
+	if (extctx->lasx.addr)
+		err |= protected_save_lasx_context(extctx);
+	else if (extctx->lsx.addr)
+		err |= protected_save_lsx_context(extctx);
+	else if (extctx->fpu.addr)
 		err |= protected_save_fpu_context(extctx);
 
 	/* Set the "end" magic */
@@ -274,6 +574,20 @@ static int parse_extcontext(struct sigcontext __user *sc, struct extctx_layout *
 			extctx->fpu.addr = info;
 			break;
 
+		case LSX_CTX_MAGIC:
+			if (size < (sizeof(struct sctx_info) +
+				    sizeof(struct lsx_context)))
+				goto invalid;
+			extctx->lsx.addr = info;
+			break;
+
+		case LASX_CTX_MAGIC:
+			if (size < (sizeof(struct sctx_info) +
+				    sizeof(struct lasx_context)))
+				goto invalid;
+			extctx->lasx.addr = info;
+			break;
+
 		default:
 			goto invalid;
 		}
@@ -319,7 +633,11 @@ static int restore_sigcontext(struct pt_regs *regs, struct sigcontext __user *sc
 	for (i = 1; i < 32; i++)
 		err |= __get_user(regs->regs[i], &sc->sc_regs[i]);
 
-	if (extctx.fpu.addr)
+	if (extctx.lasx.addr)
+		err |= protected_restore_lasx_context(&extctx);
+	else if (extctx.lsx.addr)
+		err |= protected_restore_lsx_context(&extctx);
+	else if (extctx.fpu.addr)
 		err |= protected_restore_fpu_context(&extctx);
 
 bad:
@@ -375,7 +693,13 @@ static unsigned long setup_extcontext(struct extctx_layout *extctx, unsigned lon
 	extctx->size += extctx->end.size;
 
 	if (extctx->flags & SC_USED_FP) {
-		if (cpu_has_fpu)
+		if (cpu_has_lasx && thread_lasx_context_live())
+			new_sp = extframe_alloc(extctx, &extctx->lasx,
+			  sizeof(struct lasx_context), LASX_CTX_ALIGN, new_sp);
+		else if (cpu_has_lsx && thread_lsx_context_live())
+			new_sp = extframe_alloc(extctx, &extctx->lsx,
+			  sizeof(struct lsx_context), LSX_CTX_ALIGN, new_sp);
+		else if (cpu_has_fpu)
 			new_sp = extframe_alloc(extctx, &extctx->fpu,
 			  sizeof(struct fpu_context), FPU_CTX_ALIGN, new_sp);
 	}

arch/loongarch/kernel/traps.c

@@ -514,12 +514,67 @@ static void init_restore_fp(void)
 	BUG_ON(!is_fp_enabled());
 }
 
+static void init_restore_lsx(void)
+{
+	enable_lsx();
+
+	if (!thread_lsx_context_live()) {
+		/* First time LSX context user */
+		init_restore_fp();
+		init_lsx_upper();
+		set_thread_flag(TIF_LSX_CTX_LIVE);
+	} else {
+		if (!is_simd_owner()) {
+			if (is_fpu_owner()) {
+				restore_lsx_upper(current);
+			} else {
+				__own_fpu();
+				restore_lsx(current);
+			}
+		}
+	}
+
+	set_thread_flag(TIF_USEDSIMD);
+
+	BUG_ON(!is_fp_enabled());
+	BUG_ON(!is_lsx_enabled());
+}
+
+static void init_restore_lasx(void)
+{
+	enable_lasx();
+
+	if (!thread_lasx_context_live()) {
+		/* First time LASX context user */
+		init_restore_lsx();
+		init_lasx_upper();
+		set_thread_flag(TIF_LASX_CTX_LIVE);
+	} else {
+		if (is_fpu_owner() || is_simd_owner()) {
+			init_restore_lsx();
+			restore_lasx_upper(current);
+		} else {
+			__own_fpu();
+			enable_lsx();
+			restore_lasx(current);
+		}
+	}
+
+	set_thread_flag(TIF_USEDSIMD);
+
+	BUG_ON(!is_fp_enabled());
+	BUG_ON(!is_lsx_enabled());
+	BUG_ON(!is_lasx_enabled());
+}
+
 asmlinkage void noinstr do_fpu(struct pt_regs *regs)
 {
 	irqentry_state_t state = irqentry_enter(regs);
 
 	local_irq_enable();
 	die_if_kernel("do_fpu invoked from kernel context!", regs);
+	BUG_ON(is_lsx_enabled());
+	BUG_ON(is_lasx_enabled());
 
 	preempt_disable();
 	init_restore_fp();
@@ -534,7 +589,19 @@ asmlinkage void noinstr do_lsx(struct pt_regs *regs)
 	irqentry_state_t state = irqentry_enter(regs);
 
 	local_irq_enable();
+	if (!cpu_has_lsx) {
 		force_sig(SIGILL);
+		goto out;
+	}
+
+	die_if_kernel("do_lsx invoked from kernel context!", regs);
+	BUG_ON(is_lasx_enabled());
+
+	preempt_disable();
+	init_restore_lsx();
+	preempt_enable();
+
+out:
 	local_irq_disable();
 
 	irqentry_exit(regs, state);
@@ -545,7 +612,18 @@ asmlinkage void noinstr do_lasx(struct pt_regs *regs)
 	irqentry_state_t state = irqentry_enter(regs);
 
 	local_irq_enable();
+	if (!cpu_has_lasx) {
 		force_sig(SIGILL);
+		goto out;
+	}
+
+	die_if_kernel("do_lasx invoked from kernel context!", regs);
+
+	preempt_disable();
+	init_restore_lasx();
+	preempt_enable();
+
+out:
 	local_irq_disable();
 
 	irqentry_exit(regs, state);