man/man9/buf.9

.\" Copyright (c) 1998
.\"	The Regents of the University of California.  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.
.\" 3. Neither the name of the University nor the names of its contributors
.\"    may be used to endorse or promote products derived from this software
.\"    without specific prior written permission.
.\"
.\" THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``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 REGENTS OR CONTRIBUTORS 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: src/share/man/man9/buf.9,v 1.5.2.5 2001/12/17 11:30:18 ru Exp $
.\"
.Dd December 21, 2004
.Dt BUF 9
.Os
.Sh NAME
.Nm buf
.Nd "kernel buffer I/O scheme used in DragonFly VM system"
.Sh DESCRIPTION
The kernel implements a KVM abstraction of the buffer cache which allows it
to map potentially disparate vm_page's into contiguous KVM for use by
(mainly filesystem) devices and device I/O.
This abstraction supports block sizes from
.Dv DEV_BSIZE
(usually 512) to upwards of several pages or more.
It also supports a relatively primitive byte-granular valid range and dirty
range currently hardcoded for use by NFS.
The code implementing the VM Buffer abstraction is mostly concentrated in
.Pa sys/kern/vfs_bio.c
and
.Pa sys/sys/buf.h .
.Pp
One of the most important things to remember when dealing with buffer pointers
.Ft ( struct buf )
is that the underlying pages are mapped directly from the buffer cache.
No data copying occurs in the scheme proper, though some filesystems
such as UFS do have to copy a little when dealing with file fragments.
The second most important thing to remember is that due to the underlying page
mapping, the
.Fa b_data
base pointer in a buf is always
.Em page
aligned, not
.Em block
aligned.
When you have a VM buffer representing some
.Fa b_offset
and
.Fa b_size ,
the actual start of the buffer is
.Fa ( b_data + ( Fa b_offset & Dv PAGE_MASK ) )
and not just
.Fa b_data .
Finally, the VM system's core buffer cache supports valid and dirty bits
.Fa ( m->valid , m->dirty )
for pages in
.Dv DEV_BSIZE chunks.
Thus a platform with a hardware page size of 4096 bytes has 8 valid and 8
dirty bits.
These bits are generally set and cleared in groups based on the device
block size of the device backing the page.
Complete page's worth are often referred to using the
.Dv VM_PAGE_BITS_ALL
bitmask (i.e. 0xFF if the hardware page size is 4096).
.Pp
VM buffers also keep track of a byte-granular dirty range and valid range.
This feature is normally only used by the NFS subsystem.
I'm not sure why it is used at all, actually, since we have
.Dv DEV_BSIZE
valid/dirty granularity within the VM buffer.
If a buffer dirty operation creates a
.Sq hole ,
the dirty range will extend to cover the hole.
If a buffer validation operation creates a
.Sq hole
the byte-granular valid range is left alone and will not take into account
the new extension.
Thus the whole byte-granular abstraction is considered a bad hack and it
would be nice if we could get rid of it completely.
.Pp
A VM buffer is capable of mapping the underlying VM cache pages into KVM in
order to allow the kernel to directly manipulate the data associated with
the
.Ft ( vnode , Fa b_offset , Fa b_size ) .
The kernel typically unmaps VM buffers the moment they are no longer needed
but often keeps the
.Ft struct buf
structure instantiated and even
.Fa bp->b_pages
array instantiated despite having unmapped them from KVM.
If a page making up a VM buffer is about to undergo I/O, the system typically
unmaps it from KVM and replaces the page in the
.Fa b_pages[]
array with a placemarker called
.Fa bogus_page .
The placemarker forces any kernel subsystems referencing the associated
.Ft struct buf
to re-lookup the associated page.
I believe the placemarker hack is used to allow sophisticated devices
such as filesystem devices to remap underlying pages in order to deal with,
for example, remapping a file fragment into a file block.
.Pp
VM buffers are used to track I/O operations within the kernel.
Unfortunately, the I/O implementation is also somewhat of a hack because
the kernel wants to clear the dirty bit on the underlying pages the moment
it queues the I/O to the VFS device, not when the physical I/O is actually
initiated.
This can create confusion within filesystem devices that use delayed-writes
because you wind up with pages marked clean that are actually still dirty.
If not treated carefully, these pages could be thrown away!
Indeed, a number of serious bugs related to this hack were not fixed until
the
.Fx 2.2.8 / 3.0
release.
The kernel uses an instantiated VM buffer (i.e.
.Ft struct buf )
to placemark pages in this special state.
The buffer is typically flagged
.Dv B_DELWRI .
When a device no longer needs a buffer it typically flags it as
.Dv B_RELBUF .
Due to the underlying pages being marked clean, the
.Dv B_DELWRI | B_RELBUF
combination must be interpreted to mean that the buffer is still actually
dirty and must be written to its backing store before it can actually be
released.
In the case where
.Dv B_DELWRI
is not set, the underlying dirty pages are still properly marked as dirty
and the buffer can be completely freed without losing that clean/dirty state
information.
.\"( XXX do we have to check other flags in regards to this situation ??? ).
.Pp
The kernel reserves a portion of its KVM space to hold VM Buffer's data
maps.
Even though this is virtual space (since the buffers are mapped from the
buffer cache), we cannot make it arbitrarily large because instantiated
VM Buffers
.Ft ( struct buf Ap s )
prevent their underlying pages in the buffer cache from being freed.
This can complicate the life of the paging system.
.\" .Sh SEE ALSO
.\" .Xr <fillmein> 9
.Sh HISTORY
The
.Nm
manual page was originally written by
.An Matthew Dillon
and first appeared in
.Fx 3.1 ,
December 1998.