On Wednesday July 3, [email protected] wrote:
>
> Now we are in a grey area. The 'usual' stacked drivers work like this:
>
> some fs path
> submit_bh(bh_orig);
>
> ...
>
> stacked driver make_request_fn:
> bh_new = alloc_bh
> bh_new->b_private = bh_orig;
> ...
> submit_bh(bh_new);
>
> if you are just modifying b_private, how exactly is your stacking
> working? ie what about lvm2 on lvm2?
I think this can work sanely and is something I have considered for
raid1-read and multipath in md.
struct privatebit {
bio_end_io_t *oldend;
void *oldprivate;
...other...stuff;
};
make_request(struct request_queue_t *q, struct buffer_head *bh, int rw)
{
struct privatebit *pb = kmalloc(...);
pb->oldend = bh->b_end_io;
pb->oldprivate = bh->b_private;
bh->b_private = pb;
bh->b_end_io = my_end_handler;
..remap b_rdev, b_rsector ...
generic_make_request(bh, rw);
}
Then my_end_handler have do some local cleanup,
re-instate oldend and oldprivate, and pass the bh back up.
For raid1/multipath it would arrange to resubmit the request if there
as an error.
This stacks nicely and allows for the extra bit to be alloced to be
minimal.
We just want ext3/jbd to make sure that it only calls bh2jh on
an unlocked buffer... is that easy?
Ofcourse this ceases to be an issue in 2.5 because the filesys uses
pages or buffer_heads and the device driver uses bios.
NeilBrown
On Thu, Jul 04, 2002 at 02:46:20PM +1000, Neil Brown wrote:
> I think this can work sanely and is something I have considered for
> raid1-read and multipath in md.
>
> struct privatebit {
> bio_end_io_t *oldend;
> void *oldprivate;
> ...other...stuff;
> };
>
> make_request(struct request_queue_t *q, struct buffer_head *bh, int rw)
> {
>
> struct privatebit *pb = kmalloc(...);
>
> pb->oldend = bh->b_end_io;
> pb->oldprivate = bh->b_private;
> bh->b_private = pb;
> bh->b_end_io = my_end_handler;
>
> ..remap b_rdev, b_rsector ...
>
> generic_make_request(bh, rw);
>
> }
>
> Then my_end_handler have do some local cleanup,
> re-instate oldend and oldprivate, and pass the bh back up.
> For raid1/multipath it would arrange to resubmit the request if there
> as an error.
>
> This stacks nicely and allows for the extra bit to be alloced to be
> minimal.
This is exactly what I'm doing in device-mapper :)
> Ofcourse this ceases to be an issue in 2.5 because the filesys uses
> pages or buffer_heads and the device driver uses bios.
y, 2.5 is fine.
- Joe
On Thu, Jul 04 2002, Neil Brown wrote:
> On Wednesday July 3, [email protected] wrote:
> >
> > Now we are in a grey area. The 'usual' stacked drivers work like this:
> >
> > some fs path
> > submit_bh(bh_orig);
> >
> > ...
> >
> > stacked driver make_request_fn:
> > bh_new = alloc_bh
> > bh_new->b_private = bh_orig;
> > ...
> > submit_bh(bh_new);
> >
> > if you are just modifying b_private, how exactly is your stacking
> > working? ie what about lvm2 on lvm2?
>
> I think this can work sanely and is something I have considered for
> raid1-read and multipath in md.
>
> struct privatebit {
> bio_end_io_t *oldend;
> void *oldprivate;
> ...other...stuff;
> };
>
> make_request(struct request_queue_t *q, struct buffer_head *bh, int rw)
> {
>
> struct privatebit *pb = kmalloc(...);
>
> pb->oldend = bh->b_end_io;
> pb->oldprivate = bh->b_private;
> bh->b_private = pb;
> bh->b_end_io = my_end_handler;
>
> ..remap b_rdev, b_rsector ...
>
> generic_make_request(bh, rw);
>
> }
Yes this works fine, as long as you assert that stacking is an entity of
the b_end_io functions, ie you dont rely on the state of the buffer_head
before i/o completion. But it breaks for ext3.
I'm inclined to say that I think 2.4 mappers should go the full length
like I did in loop, which is always safe (although it does eat more
memory, of course, buffer_heads aren't exactly lite :)
int make_request(struct request_queue_t *q, struct buffer_head *bh, int rw)
{
struct buffer_head *stack_bh = some_pool_alloc(...);
stack_bh->b_private = bh;
stack_bh->b_end_io = my_end_handler;
/* setup b_rdev and b_rsector, etc */
generic_make_request(stack_bh, rw);
return 0;
}
> We just want ext3/jbd to make sure that it only calls bh2jh on
> an unlocked buffer... is that easy?
That's the question indeed, someone with a good grasp of jbd should make
that call. If that is the only 'violator' (depending on your point of
view), then yes lets just fix that up and say that the above is pb
private is valid.
> Ofcourse this ceases to be an issue in 2.5 because the filesys uses
> pages or buffer_heads and the device driver uses bios.
Yep, no such problem in 2.5
--
Jens Axboe