files cannot overflow the vm_pindex_t type. This is not an MFC,
since increasing the size of vm_pindex_t as was done in -CURRENT
is not useful or appropriate for RELENG_4.

Note that on the alpha platform this change is not necessary because
vm_pindex_t is 64-bit, but limiting files to 16TB there is unlikely
to cause any problems.

Submitted by:	tegge

2000.  Setting B_CLRBUF for the BALLOC call unexpectedly caused BALLOC
to issue a read-before-write for each fs block, resulting in a huge
(4x) performance penalty due to extra I/Os and lack of clustering.
This workaround avoids the unnecessary read resulting in full write
clustering performance to media.  It is expected to have major visible
effects only on large news systems (e.g. INN cyclic buffers), and databases,
especially on hardware RAID systems (10x), and has already been shown to
have a huge positive effect on INN.  Most other programs use write-append
which is already optimal.

This patch does not fix the double-memory-copy to the filesystem buffer.
That is a more difficult problem and is being worked on as a separate
issue.

Reviewed by:	mckusick, tegge

si_iosize_max correctly in ffs, and don't forget to set it in cd9660
and ext2fs (the original log message mistakenly said msdosfs instead of
cd9660)).

incorrectly update the free-space statistics in some cases.

that b_resid is the expected 0 value in preparation for a future commit
that will remove the b_resid calculation entirely.  It turns out that
b_resid WILL always be 0 except for the case where a write fails.  But
in the write failure case the buffer is still valid for reading and
we still want to treat it as such (as if b_resid were 0, even though it
isn't).  The filesystem code should not use b_resid in any calculations
except those involving short-read situations, and even then b_resid
will not be persistent for a VMIO backed buffer since the buffer may be
destroyed and later reconstituted, resetting b_resid to 0.  short-reads
are not used by most filesystems.  They are used by NFS only when reading
softlinks (using a non VMIO buffer). Joy.

another process.

Approved by:	mckusick

operatings involving file removal and file update were not always being
fully comitted to disk, resulting in corruption.
Approved by:		mckusick