freebsd-skq/share/man/man9/fpu_kern.9
kib cb80ddd115 Add FPU_KERN_NOCTX flag to the fpu_kern_enter() function on amd64.
The flag specifies that the block which uses FPU must be executed in
critical section, i.e. take no context switches, and does not need an
FPU save area during the execution.

It is intended to be applied around fast and short code pathes where
save area allocation is impossible or undesirable, due to context or
due to the relative cost of calculation vs. allocation.

Sponsored by:	The FreeBSD Foundation
MFC after:	2 weeks
2016-09-11 09:14:07 +00:00

211 lines
6.7 KiB
Groff

.\" Copyright (c) 2014
.\" Konstantin Belousov <kib@FreeBSD.org>. All rights reserved.
.\"
.\" Redistribution and use in source and binary forms, with or without
.\" modification, are permitted provided that the following conditions
.\" are met:
.\" 1. Redistributions of source code must retain the above copyright
.\" notice, this list of conditions and the following disclaimer.
.\" 2. Redistributions in binary form must reproduce the above copyright
.\" notice, this list of conditions and the following disclaimer in the
.\" documentation and/or other materials provided with the distribution.
.\"
.\" THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
.\" IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
.\" OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
.\" IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
.\" INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
.\" NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
.\" DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
.\" THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
.\" (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
.\" THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
.\"
.\" $FreeBSD$
.\"
.Dd October 23, 2014
.Dt FPU_KERN 9
.Os
.Sh NAME
.Nm fpu_kern
.Nd "facility to use the FPU in the kernel"
.Sh SYNOPSIS
.Ft struct fpu_kern_ctx *
.Fn fpu_kern_alloc_ctx "u_int flags"
.Ft void
.Fn fpu_kern_free_ctx "struct fpu_kern_ctx *ctx"
.Ft int
.Fn fpu_kern_enter "struct thread *td" "struct fpu_kern_ctx *ctx" "u_int flags"
.Ft int
.Fn fpu_kern_leave "struct thread *td" "struct fpu_kern_ctx *ctx"
.Ft int
.Fn fpu_kern_thread "u_int flags"
.Ft int
.Fn is_fpu_kern_thread "u_int flags"
.Sh DESCRIPTION
The
.Nm
family of functions allows the use of FPU hardware in kernel code.
Modern FPUs are not limited to providing hardware implementation for
floating point arithmetic; they offer advanced accelerators for cryptography
and other computational-intensive algorithms.
These facilities share registers with the FPU hardware.
.Pp
Typical kernel code does not need access to the FPU.
Saving a large register file on each entry to the kernel would waste
time.
When kernel code uses the FPU, the current FPU state must be saved to
avoid corrupting the user-mode state, and vice versa.
.Pp
The management of the save and restore is automatic.
The processor catches accesses to the FPU registers
when the non-current context tries to access them.
Explicit calls are required for the allocation of the save area and
the notification of the start and end of the code using the FPU.
.Pp
The
.Fn fpu_kern_alloc_ctx
function allocates the memory used by
.Nm
to track the use of the FPU hardware state and the related software state.
The
.Fn fpu_kern_alloc_ctx
function requires the
.Fa flags
argument, which currently accepts the following flags:
.Bl -tag -width ".Dv FPU_KERN_NOWAIT" -offset indent
.It Dv FPU_KERN_NOWAIT
Do not wait for the available memory if the request could not be satisfied
without sleep.
.It 0
No special handling is required.
.El
.Pp
The function returns the allocated context area, or
.Va NULL
if the allocation failed.
.Pp
The
.Fn fpu_kern_free_ctx
function frees the context previously allocated by
.Fn fpu_kern_alloc_ctx .
.Pp
The
.Fn fpu_kern_enter
function designates the start of the region of kernel code where the
use of the FPU is allowed.
Its arguments are:
.Bl -tag -width ".Fa ctx" -offset indent
.It Fa td
Currently must be
.Va curthread .
.It Fa ctx
The context save area previously allocated by
.Fn fpu_kern_alloc_ctx
and not currently in use by another call to
.Fn fpu_kern_enter .
.It Fa flags
This argument currently accepts the following flags:
.Bl -tag -width ".Dv FPU_KERN_NORMAL" -offset indent
.It Dv FPU_KERN_NORMAL
Indicates that the caller intends to access the full FPU state.
Must be specified currently.
.It Dv FPU_KERN_KTHR
Indicates that no saving of the current FPU state should be performed,
if the thread called
.Xr fpu_kern_thread 9
function.
This is intended to minimize code duplication in callers which
could be used from both kernel thread and syscall contexts.
The
.Fn fpu_kern_leave
function correctly handles such contexts.
.It Dv FPU_KERN_NOCTX
Avoid nesting save area.
If the flag is specified, the
.Fa ctx
must be passed as
.Va NULL .
The flag should only be used for really short code blocks
which can be executed in a critical section.
It avoids the need to allocate the FPU context by the cost
of increased system latency.
.El
.El
.Pp
The function does not sleep or block.
It could cause 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).
Currently, no errors are defined which can be returned by
.Fn fpu_kern_enter
to the caller.
.Pp
The
.Fn fpu_kern_leave
function ends the region started by
.Fn fpu_kern_enter .
The uses of FPU in the kernel after the call to
.Fn fpu_kern_leave
are erroneous until the next call to
.Fn fpu_kern_enter
is performed.
The function takes the
.Fa td
thread argument, which currently must be
.Va curthread ,
and the
.Fa ctx
context pointer, previously passed to
.Fn fpu_kern_enter .
After the function returns, the context may be freed or reused
by other invocation of
.Fn fpu_kern_enter .
There are no errors defined for the function, it always returns 0.
.Pp
The
.Fn fpu_kern_thread
function enables an optimization for threads which never leave to
the usermode.
The current thread will reuse the usermode save area for the kernel FPU state
instead of requiring an explicitly allocated context.
There are no flags defined for the function, and no error states
that the function returns.
Once this function has been called, neither
.Fn fpu_kern_enter
nor
.Fn fpu_kern_leave
is required to be called and the fpu is available for use in the calling thread.
.Pp
The
.Fn is_fpu_kern_thread
function returns the boolean indicating whether the current thread
entered the mode enabled by
.Fn fpu_kern_thread .
There is currently no flags defined for the function, the return
value is true if the current thread have the permanent FPU save area,
and false otherwise.
.Sh NOTES
The
.Nm
is currently implemented only for the i386 and amd64 architectures.
.Pp
There is no way to handle floating point exceptions raised from
kernel mode.
.Pp
The unused
.Fa flags
arguments
to the
.Nm
functions are to be extended to allow specification of the
set of the FPU hardware state used by the code region.
This would allow optimizations of saving and restoring the state.
.Sh AUTHORS
The
.Nm
facitily and this manual page were written by
.An Konstantin Belousov Aq Mt kib@FreeBSD.org .