Add support for the fpu_kern(9) KPI on arm64. It hooks into the existing

VFP code to store the old context, with lazy loading of the new context
when needed.

FPU_KERN_NOCTX is missing as this is unused in the crypto code this has
been tested with, and I am unsure on the requirements of the UEFI
Runtime Services.

Reviewed by:	kib
Obtained from:	ABT Systeems Ltd
Sponsored by:	The FreeBSD Foundation
Differential Revision:	https://reviews.freebsd.org/D8276
This commit is contained in:
Andrew Turner 2016-10-20 09:22:10 +00:00
parent 61f481fb7e
commit 6ed982a221
Notes: svn2git 2020-12-20 02:59:44 +00:00
svn path=/head/; revision=307667
6 changed files with 162 additions and 15 deletions

View File

@ -23,7 +23,7 @@
.\"
.\" $FreeBSD$
.\"
.Dd October 23, 2014
.Dd October 20, 2016
.Dt FPU_KERN 9
.Os
.Sh NAME
@ -134,11 +134,11 @@ of increased system latency.
.El
.Pp
The function does not sleep or block.
It could cause the
It could cause an FPU trap during execution, and on the first FPU access
after the function returns, as well as after each context switch.
On i386 and amd64 this will be the
.Nm Device Not Available
exception during execution, and on the first FPU access after the
function returns, as well as after each context switch
(see Intel Software Developer Manual for the reference).
exception (see Intel Software Developer Manual for the reference).
Currently, no errors are defined which can be returned by
.Fn fpu_kern_enter
to the caller.
@ -190,7 +190,7 @@ and false otherwise.
.Sh NOTES
The
.Nm
is currently implemented only for the i386 and amd64 architectures.
is currently implemented only for the i386, amd64, and arm64 architectures.
.Pp
There is no way to handle floating point exceptions raised from
kernel mode.
@ -208,3 +208,5 @@ The
.Nm
facitily and this manual page were written by
.An Konstantin Belousov Aq Mt kib@FreeBSD.org .
The arm64 support was added by
.An Andrew Turner Aq Mt andrew@FreeBSD.org .

View File

@ -282,9 +282,17 @@ do_el1h_sync(struct trapframe *frame)
switch(exception) {
case EXCP_FP_SIMD:
case EXCP_TRAP_FP:
print_registers(frame);
printf(" esr: %.8lx\n", esr);
panic("VFP exception in the kernel");
#ifdef VFP
if ((curthread->td_pcb->pcb_fpflags & PCB_FP_KERN) != 0) {
vfp_restore_state();
} else
#endif
{
print_registers(frame);
printf(" esr: %.8lx\n", esr);
panic("VFP exception in the kernel");
}
break;
case EXCP_INSN_ABORT:
case EXCP_DATA_ABORT:
far = READ_SPECIALREG(far_el1);
@ -412,6 +420,9 @@ do_el0_sync(struct trapframe *frame)
KASSERT((curthread->td_pcb->pcb_fpflags & ~PCB_FP_USERMASK) == 0,
("Kernel VFP flags set while entering userspace"));
KASSERT(
curthread->td_pcb->pcb_fpusaved == &curthread->td_pcb->pcb_fpustate,
("Kernel VFP state in use when entering userspace"));
}
void

View File

@ -1,5 +1,5 @@
/*-
* Copyright (c) 2015 The FreeBSD Foundation
* Copyright (c) 2015-2016 The FreeBSD Foundation
* All rights reserved.
*
* This software was developed by Andrew Turner under
@ -48,6 +48,14 @@ CTASSERT(sizeof(((struct pcb *)0)->pcb_fpustate.vfp_regs) == 16 * 32);
static MALLOC_DEFINE(M_FPUKERN_CTX, "fpukern_ctx",
"Kernel contexts for VFP state");
struct fpu_kern_ctx {
struct vfpstate *prev;
#define FPU_KERN_CTX_DUMMY 0x01 /* avoided save for the kern thread */
#define FPU_KERN_CTX_INUSE 0x02
uint32_t flags;
struct vfpstate state;
};
static void
vfp_enable(void)
{
@ -71,9 +79,10 @@ vfp_disable(void)
}
/*
* Called when the thread is dying. If the thread was the last to use the
* VFP unit mark it as unused to tell the kernel the fp state is unowned.
* Ensure the VFP unit is off so we get an exception on the next access.
* Called when the thread is dying or when discarding the kernel VFP state.
* If the thread was the last to use the VFP unit mark it as unused to tell
* the kernel the fp state is unowned. Ensure the VFP unit is off so we get
* an exception on the next access.
*/
void
vfp_discard(struct thread *td)
@ -226,4 +235,111 @@ vfp_init(void)
SYSINIT(vfp, SI_SUB_CPU, SI_ORDER_ANY, vfp_init, NULL);
struct fpu_kern_ctx *
fpu_kern_alloc_ctx(u_int flags)
{
struct fpu_kern_ctx *res;
size_t sz;
sz = sizeof(struct fpu_kern_ctx);
res = malloc(sz, M_FPUKERN_CTX, ((flags & FPU_KERN_NOWAIT) ?
M_NOWAIT : M_WAITOK) | M_ZERO);
return (res);
}
void
fpu_kern_free_ctx(struct fpu_kern_ctx *ctx)
{
KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("free'ing inuse ctx"));
/* XXXAndrew clear the memory ? */
free(ctx, M_FPUKERN_CTX);
}
int
fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags)
{
struct pcb *pcb;
pcb = td->td_pcb;
KASSERT(ctx == NULL || (ctx->flags & FPU_KERN_CTX_INUSE) == 0,
("using inuse ctx"));
if ((flags & FPU_KERN_KTHR) != 0 && is_fpu_kern_thread(0)) {
ctx->flags = FPU_KERN_CTX_DUMMY | FPU_KERN_CTX_INUSE;
return (0);
}
/*
* Check either we are already using the VFP in the kernel, or
* the the saved state points to the default user space.
*/
KASSERT((pcb->pcb_fpflags & PCB_FP_KERN) != 0 ||
pcb->pcb_fpusaved == &pcb->pcb_fpustate,
("Mangled pcb_fpusaved %x %p %p", pcb->pcb_fpflags, pcb->pcb_fpusaved, &pcb->pcb_fpustate));
ctx->flags = FPU_KERN_CTX_INUSE;
vfp_save_state(curthread, pcb);
ctx->prev = pcb->pcb_fpusaved;
pcb->pcb_fpusaved = &ctx->state;
pcb->pcb_fpflags |= PCB_FP_KERN;
pcb->pcb_fpflags &= ~PCB_FP_STARTED;
return (0);
}
int
fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx)
{
struct pcb *pcb;
pcb = td->td_pcb;
KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0,
("FPU context not inuse"));
ctx->flags &= ~FPU_KERN_CTX_INUSE;
if (is_fpu_kern_thread(0) &&
(ctx->flags & FPU_KERN_CTX_DUMMY) != 0)
return (0);
KASSERT((ctx->flags & FPU_KERN_CTX_DUMMY) == 0, ("dummy ctx"));
critical_enter();
vfp_discard(td);
critical_exit();
pcb->pcb_fpflags &= ~PCB_FP_STARTED;
pcb->pcb_fpusaved = ctx->prev;
if (pcb->pcb_fpusaved == &pcb->pcb_fpustate) {
pcb->pcb_fpflags &= ~PCB_FP_KERN;
} else {
KASSERT((pcb->pcb_fpflags & PCB_FP_KERN) != 0,
("unpaired fpu_kern_leave"));
}
return (0);
}
int
fpu_kern_thread(u_int flags)
{
struct pcb *pcb = curthread->td_pcb;
KASSERT((curthread->td_pflags & TDP_KTHREAD) != 0,
("Only kthread may use fpu_kern_thread"));
KASSERT(pcb->pcb_fpusaved == &pcb->pcb_fpustate,
("Mangled pcb_fpusaved"));
KASSERT((pcb->pcb_fpflags & PCB_FP_KERN) == 0,
("Thread already setup for the VFP"));
pcb->pcb_fpflags |= PCB_FP_KERN;
return (0);
}
int
is_fpu_kern_thread(u_int flags)
{
struct pcb *curpcb;
if ((curthread->td_pflags & TDP_KTHREAD) == 0)
return (0);
curpcb = curthread->td_pcb;
return ((curpcb->pcb_fpflags & PCB_FP_KERN) != 0);
}
#endif

View File

@ -54,6 +54,7 @@ struct pcb {
struct vfpstate *pcb_fpusaved;
int pcb_fpflags;
#define PCB_FP_STARTED 0x01
#define PCB_FP_KERN 0x02
/* The bits passed to userspace in get_fpcontext */
#define PCB_FP_USERMASK (PCB_FP_STARTED)
u_int pcb_vfpcpu; /* Last cpu this thread ran VFP code */

View File

@ -45,6 +45,23 @@ void vfp_init(void);
void vfp_discard(struct thread *);
void vfp_restore_state(void);
void vfp_save_state(struct thread *, struct pcb *);
struct fpu_kern_ctx;
/*
* Flags for fpu_kern_alloc_ctx(), fpu_kern_enter() and fpu_kern_thread().
*/
#define FPU_KERN_NORMAL 0x0000
#define FPU_KERN_NOWAIT 0x0001
#define FPU_KERN_KTHR 0x0002
struct fpu_kern_ctx *fpu_kern_alloc_ctx(u_int);
void fpu_kern_free_ctx(struct fpu_kern_ctx *);
int fpu_kern_enter(struct thread *, struct fpu_kern_ctx *, u_int);
int fpu_kern_leave(struct thread *, struct fpu_kern_ctx *);
int fpu_kern_thread(u_int);
int is_fpu_kern_thread(u_int);
#endif
#endif

View File

@ -81,7 +81,7 @@ __FBSDID("$FreeBSD$");
#include <sys/bus.h>
#include "cryptodev_if.h"
#if defined(__i386__) || defined(__amd64__)
#if defined(__i386__) || defined(__amd64__) || defined(__aarch64__)
#include <machine/pcb.h>
#endif
@ -1246,7 +1246,7 @@ crypto_proc(void)
u_int32_t hid;
int result, hint;
#if defined(__i386__) || defined(__amd64__)
#if defined(__i386__) || defined(__amd64__) || defined(__aarch64__)
fpu_kern_thread(FPU_KERN_NORMAL);
#endif