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.
.\"
.Dd December 22, 1998
.Dt BUF 9
.Os
.Sh NAME
.Nm buf
.Nd "kernel buffer I/O scheme used in FreeBSD 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 file system) 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
in the
.Fx
source tree.
.Pp
One of the most important things to remember when dealing with buffer pointers
.Pq Vt struct buf
is that the underlying pages are mapped directly from the buffer
cache.
No data copying occurs in the scheme proper, though some file systems
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
.Va b_data
base pointer in a buf is always
.Em page Ns -aligned ,
not
.Em block Ns -aligned .
When you have a VM buffer representing some
.Va b_offset
and
.Va b_size ,
the actual start of the buffer is
.Ql b_data + (b_offset & PAGE_MASK)
and not just
.Ql b_data .
Finally, the VM system's core buffer cache supports
valid and dirty bits
.Pq Va 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 am 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
.Dq hole ,
the dirty range will extend to cover the hole.
If a buffer validation
operation creates a
.Dq 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
.Pq Va vnode , b_offset , b_size .
The kernel typically unmaps VM buffers the moment
they are no longer needed but often keeps the
.Vt struct buf
structure
instantiated and even
.Va 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
.Va b_pages[]
array with a place-marker called bogus_page.
The place-marker forces any kernel
subsystems referencing the associated
.Vt struct buf
to re-lookup the associated
page.
I believe the place-marker hack is used to allow sophisticated devices
such as file system devices to remap underlying pages in order to deal with,
for example, re-mapping 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 file system 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 /
.Fx 3.0
release.
The kernel uses an instantiated VM buffer (i.e.,
.Vt struct buf )
to place-mark 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
.Ql 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
.Pq Vt 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 HISTORY
The
.Nm
manual page was originally written by
.An Matthew Dillon
and first appeared in
.Fx 3.1 ,
December 1998.