[RFC] should block layer export bd_set_size() or equivalent?

[RFC] should block layer export bd_set_size() or equivalent?

[Lou Langholtz]

Background:

Several drivers (like drivers/block/nbd.c or drivers/block/rd.c) need to 
be able to update their devices' size after the first open. The way 
fs/block_dev.c is currently implemented (as recently as 2.5.73-mm2) 
however, updating the size as seen by user processes is not possible 
without directly updating the size stored in various locations of the 
struct block_device and the bd_inode as well. As an example, we have the 
code in drivers/block/rd.c rd_open() which essentially calls what 
fs/block_dev.c calls in its bd_set_size() function. But things are 
changing within the block layer still as evidenced by changes to 
bd_set_size() between 2.5.73 and 2.5.73-mm2 and consequently 
drivers/block/rd.c rd_open() may be incorrect now.

Suggestion:

To have fs/block_dev.c share bd_set_size(struct block_device *bdev, 
loff_t size) or something like this to better abstract the block layer 
from drivers and save us from essentially re-writing the code from here.

Next step:

Discuss this. Why isn't it already exported? What are the cons of just 
exporting bd_set_size()? Should we even let drivers change a device's 
size after the first open? An alternative for nbd at least, is to have 
its user space tool (nbd-client) simply close the device after it has 
changed the size and then re-open it before continuing on. This way 
fs/block_dev.c do_open() re-checks the disk capacity and calls 
bd_set_size() again with the new correct size. Also, bd_set_size() is 
very similar to set_blocksize() (the former though changes the byte size 
of the block device which is what's most critical really to nbd at 
least), perhaps these two functions can be rolled into one; at least 
set_blocksize() is already exported.