Provided by: freebsd-manpages_8.2-1_all bug


     buf — kernel buffer I/O scheme used in FreeBSD VM system


     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 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

     One of the most important things to remember when dealing with buffer
     pointers (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 b_data base pointer in a
     buf is always *page* aligned, not *block* aligned.  When you have a VM
     buffer representing some b_offset and b_size, the actual start of the
     buffer is (b_data + (b_offset & PAGE_MASK)) and not just b_data.
     Finally, the VM system's core buffer cache supports valid and dirty bits
     (m->valid, m->dirty) for pages in 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 VM_PAGE_BITS_ALL bitmask (i.e., 0xFF if the
     hardware page size is 4096).

     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 DEV_BSIZE
     valid/dirty granularity within the VM buffer.  If a buffer dirty
     operation creates a 'hole', the dirty range will extend to cover the
     hole.  If a buffer validation operation creates a '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.

     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 (vnode,b_offset,b_size).  The kernel typically unmaps VM buffers
     the moment they are no longer needed but often keeps the 'struct buf'
     structure instantiated and even 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 b_pages[] array with a place-marker called bogus_page.  The
     place-marker forces any kernel subsystems referencing the associated
     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.

     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 2.2.8/3.0 release.  The kernel uses an instantiated
     VM buffer (i.e., struct buf) to place-mark pages in this special state.
     The buffer is typically flagged B_DELWRI.  When a device no longer needs
     a buffer it typically flags it as B_RELBUF.  Due to the underlying pages
     being marked clean, the 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
     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 ???)

     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 (struct buf's) prevent their underlying pages in
     the buffer cache from being freed.  This can complicate the life of the
     paging system.


     The buf manual page was originally written by Matthew Dillon and first
     appeared in FreeBSD 3.1, December 1998.