LinuxKPI: Make FPU sections thread-safe and use the NOCTX flag.

Reviewed by:	kib
Submitted by:	greg@unrelenting.technology
Differential Revision:	https://reviews.freebsd.org/D29921
MFC after:	1 week
Sponsored by:	NVIDIA Networking
This commit is contained in:
Hans Petter Selasky 2021-07-31 15:32:52 +02:00
parent f7f76c200a
commit 469884cf04
7 changed files with 43 additions and 56 deletions

View File

@ -28,41 +28,13 @@
#ifndef _FPU_API_H_
#define _FPU_API_H_
#if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
#define kernel_fpu_begin() \
lkpi_kernel_fpu_begin()
#include <machine/fpu.h>
#define kernel_fpu_end() \
lkpi_kernel_fpu_end()
extern struct fpu_kern_ctx *__lkpi_fpu_ctx;
extern unsigned int __lkpi_fpu_ctx_level;
static inline void
kernel_fpu_begin()
{
if (__lkpi_fpu_ctx_level++ == 0) {
fpu_kern_enter(curthread, __lkpi_fpu_ctx, FPU_KERN_NORMAL);
}
}
static inline void
kernel_fpu_end()
{
if (--__lkpi_fpu_ctx_level == 0) {
fpu_kern_leave(curthread, __lkpi_fpu_ctx);
}
}
#else
static inline void
kernel_fpu_begin()
{
}
static inline void
kernel_fpu_end()
{
}
#endif
extern void lkpi_kernel_fpu_begin(void);
extern void lkpi_kernel_fpu_end(void);
#endif /* _FPU_API_H_ */

View File

@ -82,7 +82,8 @@ struct task_struct {
int bsd_interrupt_value;
struct work_struct *work; /* current work struct, if set */
struct task_struct *group_leader;
unsigned rcu_section[TS_RCU_TYPE_MAX];
unsigned rcu_section[TS_RCU_TYPE_MAX];
unsigned int fpu_ctx_level;
};
#define current ({ \

View File

@ -30,21 +30,44 @@
#include <sys/proc.h>
#include <sys/kernel.h>
#include <linux/sched.h>
#include <asm/fpu/api.h>
#if defined(__aarch64__) || defined(__amd64__) || defined(__i386__)
#include <machine/fpu.h>
struct fpu_kern_ctx *__lkpi_fpu_ctx;
unsigned int __lkpi_fpu_ctx_level = 0;
/*
* Technically the Linux API isn't supposed to allow nesting sections
* either, but currently used versions of GPU drivers rely on nesting
* working, so we only enter the section on the outermost level.
*/
static void
linux_fpu_init(void *arg __unused)
void
lkpi_kernel_fpu_begin(void)
{
__lkpi_fpu_ctx = fpu_kern_alloc_ctx(0);
if ((current->fpu_ctx_level)++ == 0)
fpu_kern_enter(curthread, NULL, FPU_KERN_NOCTX);
}
SYSINIT(linux_fpu, SI_SUB_EVENTHANDLER, SI_ORDER_SECOND, linux_fpu_init, NULL);
static void
linux_fpu_uninit(void *arg __unused)
void
lkpi_kernel_fpu_end(void)
{
fpu_kern_free_ctx(__lkpi_fpu_ctx);
if (--(current->fpu_ctx_level) == 0)
fpu_kern_leave(curthread, NULL);
}
SYSUNINIT(linux_fpu, SI_SUB_EVENTHANDLER, SI_ORDER_SECOND, linux_fpu_uninit, NULL);
#else
void
lkpi_kernel_fpu_begin(void)
{
}
void
lkpi_kernel_fpu_end(void)
{
}
#endif

View File

@ -398,9 +398,6 @@ x86/x86/mptable_pci.c optional mptable pci
x86/x86/msi.c optional pci
x86/xen/pv.c optional xenhvm
compat/linuxkpi/common/src/linux_fpu.c optional compat_linuxkpi \
compile-with "${LINUXKPI_C}"
contrib/openzfs/module/zcommon/zfs_fletcher_avx512.c optional zfs compile-with "${ZFS_C}"
contrib/openzfs/module/zcommon/zfs_fletcher_intel.c optional zfs compile-with "${ZFS_C}"
contrib/openzfs/module/zcommon/zfs_fletcher_sse.c optional zfs compile-with "${ZFS_C}"

View File

@ -560,10 +560,6 @@ arm64/rockchip/clk/rk3399_pmucru.c optional fdt soc_rockchip_rk3399
# Xilinx
arm/xilinx/uart_dev_cdnc.c optional uart soc_xilinx_zynq
# Linuxkpi
compat/linuxkpi/common/src/linux_fpu.c optional compat_linuxkpi \
compile-with "${LINUXKPI_C}"
# Cloudabi
arm64/cloudabi32/cloudabi32_sysvec.c optional compat_cloudabi32
arm64/cloudabi64/cloudabi64_sysvec.c optional compat_cloudabi64

View File

@ -182,6 +182,3 @@ x86/x86/local_apic.c optional apic
x86/x86/mptable.c optional apic
x86/x86/mptable_pci.c optional apic pci
x86/x86/msi.c optional apic pci
compat/linuxkpi/common/src/linux_fpu.c optional compat_linuxkpi \
compile-with "${LINUXKPI_C}"

View File

@ -8,6 +8,7 @@ SRCS= linux_compat.c \
linux_dmi.c \
linux_domain.c \
linux_firmware.c \
linux_fpu.c \
linux_hrtimer.c \
linux_idr.c \
linux_kmod.c \
@ -29,7 +30,7 @@ SRCS= linux_compat.c \
.if ${MACHINE_CPUARCH} == "aarch64" || ${MACHINE_CPUARCH} == "amd64" || \
${MACHINE_CPUARCH} == "i386"
SRCS+= opt_acpi.h acpi_if.h linux_acpi.c linux_fpu.c
SRCS+= opt_acpi.h acpi_if.h linux_acpi.c
.endif
SRCS+= ${LINUXKPI_GENSRCS}