doc - Fix previous commit.

[dragonfly.git] / notes / porting_drivers.txt

$DragonFly: doc/notes/porting_drivers.txt,v 1.4 2008/04/06 19:08:30 pavalos Exp $

		    PORTING FREEBSD DRIVERS TO DRAGONFLY

* Copy the driver code to the appropriate DragonFly directory.  For example,
  a disk driver /usr/src/sys/dev/blah in FreeBSD would likely be 
  /usr/src/sys/dev/disk/blah in DragonFly.

* Keep all the CVS ids in the files for future reference point. If the files
  don't have it, please add it to them. You can find them out from CVSWeb
  repository or by checking out the specific source files with the FreeBSD
  SVN client.
  This could be automated in the case pkgsrc's SVN client had the set of
  patches that provide this functionality.

* Remove FBSDID declaration and '#include <sys/cdefs.h>' as well.

* Driver local #include's probably use a <dev/blah/blah.h> path.  These
  need to be changed to "blah.h".   '.' is not included in the #include
  path in FreeBSD builds, but it is in DragonFly builds.

* Other #include's may reference things in <dev/...> which in DragonFly
  reside in <bus/...>.  In particular, dev/pccard becomes bus/pccard.
  Note that defines in FreeBSD's pccard_cis.h reside in DragonFly's
  pccardreg.h .

* The following kernel functions have been renamed in DragonFly:

  malloc(), free() etc.	->	kmalloc(), kfree() etc.
  printf() etc.		->	kprintf() etc.
  psignal()		->	ksignal()
  random()		->	krandom()

* MUTEX conversion - mutexes are generally replaced by spinlocks.  However,
  DragonFly spinlocks are more restrictive than FreeBSD mutexes so a
  direct replacement is not necessarily appropriate in all cases.  A lockmgr
  lock should be used when a direct replacement is not appropriate.
  In particular, DragonFly does not allow recursive exclusive spinlocks
  and does not allow multiple exclusive spinlocks to be held by any given
  thread.

  Instances of <sys/mutex.h> should be replaced with <sys/spinlock.h>.

  When replacing mutexes with spinlocks it is a good idea to rename 
  the structural field (typically 'mtx') to something else (typically 'spin').

  The &Giant mutex is typically converted to get_mplock() and rel_mplock().
  However, there are places where FreeBSD unlocks giant around some code and
  then relocks giant... those should simply be removed.

  FreeBSD has weird callout + mutex functions.  DragonFly does not integrate
  the two.  Instead, the driver in DragonFly must obtain the spinlocks
  in question in the callback routine.

  As a rule of thumb, MTX_DEF mutexes should be replaced with exclusive,
  recursive lockmgr locks.

  So, suppose the original code is using 
	struct mtx my_mtx;
  you'd normally rename it to
	struct lock my_lock;

  and change the initialization from something like
	mtx_init(&my_mtx, "mymtx", "whatever", MTX_DEF);
  to
	lockinit(&my_lock, "mylock", 0, LK_CANRECURSE);

  Destroying it is trivial,
	mtx_destroy(&my_mtx);
  becomes
	lockuninit(&my_lock);

  You use the same function for locking and unlocking a lockmgr lock,
  so exchange
	mtx_lock(&my_mtx);
  with
	lockmgr(&my_lock, LK_EXCLUSIVE);
  and
	mtx_unlock(&my_mtx);
  with
	lockmgr(&my_lock, LK_RELEASE);

  For testing the lock status, one would use
	lockstatus(&my_lock, curthread);
  in place of
	mtx_owned(&my_mtx);

  An
	mtx_trylock(&my_mtx);
  call is replaced with
	lockmgr(&my_lock, LK_EXCLUSIVE|LK_NOWAIT);

  As for mtx_assert() calls, translate them like this:

	mtx_assert(&my_mtx, MA_OWNED) -> KKASSERT(lockstatus(&my_lock, curthread) != 0)
	mtx_assert(&my_mtx, MA_NOTOWNED) -> KKASSERT(lockstatus(&my_lock, curthread) == 0)

  In DragonFly, lockstatus() does not return information about whether there have been
  recursive lock acquisitions, so there is no generic way to emulate the

	mtx_assert(&my_mtx, MA_OWNED|MA_RECURSED);
	mtx_assert(&my_mtx, MA_OWNED|MA_NOTRECURSED);

  calls.
	
* UMA conversion - generally speaking UMA should be converted to a standard
  kmalloc.

  Note however that in FreeBSD M_NOWAIT is often used in cases where, in fact,
  the kmalloc cannot fail without blowing something up or causing a fatal
  (and very unexpected) I/O error.  M_INTWAIT should be used for these cases.

* CDEVSW conversion - see other devices.  Generally speaking a major number
  is needed and a function map needs to be specified more explicitly.

  Most calls passing struct cdev pointers are dev_t's in DragonFly.

  All device vectors in DragonFly pass a dev_<name>_args structure pointer
  instead of explicit arguments.

  Strategy calls - we pass BIO's and a lot of BUF fields are in the BIO
  in FreeBSD, but left in the BUF in DragonFly.  FreeBSD for some reason
  names its struct bio pointers 'bp', its a good idea to rename them to 'bio'
  to avoid confusion and have a struct buf *bp = bio->bio_buf; pointer to
  access the buf.

* MSLEEP/TSLEEP conversion.  The DragonFly msleep/tsleep do not have 'PRI'
  priorities.  0 should be used.

* BUS_* FUNCTIONS

  bus_setup_intr() - replace INTR_TYPE_* flags with 0.  There is an extra
  argument for an interrupt interlock using the sys/serializer.h interface.
  This can either be left NULL or you can convert the spinlock(s) for
  the driver into serializer locks and integrate the interrupt service
  routine with a serializer.

* CAM CODE - cam_simq* code refcounts, so shared device queues (raid and
  multi-channel devices) are not freed before all references have gone
  away.