Damien Zammit, le lun. 30 mars 2020 22:22:12 +1100, a ecrit:
> I cleaned up rumpdisk and used your new api. See patch below.
I'm amazed how very straighforward this implementation looks, rump rocks
:)
Concerning the link line, I don't understand why hardcoding everything?
- For a start, are the _pic.a versions really needed? The .a versions
should just work.
- Then, you don't need to define a link rule by hand, you can put the
names of the hurd libraries to be linked in in HURDLIBS, and the -l
flags for the system libraries to be linked in in LDLIBS. See for
instance ext2fs/Makefile.
- For the rump+machdev piece with "whole-archive", you can stuff that in
LDLIBS.
That way, rumpdisk will benefit from all the common make bits.
About IOC macros in block-rump.c, is rump not providing a header that
defines them, instead of hardcoding them here?
Concerning translate_name, instead of malloc+snprintf, you can use
asprintf. That being said, since it doesn't actually need to be kept
allocated, better just make this a function that takes the output
buffer, which you can allocate on the stack in device_open.
Take care of indentation, the GNU Coding Style is not
if (foo)
{
bla();
}
but
if (foo)
{
bla();
}
There are a couple mach_print debugging calls left.
Please add a /* FIXME: make multithreaded */ in device_write/read. The
way it is currently shaped is completely synchronous: device_read will
be stuck inside rump_sys_read() until the data is retrieved from the
disk. To fix this, we need to add a machdev_multithread_server function
in libmachdev, which calls ports_manage_port_operations_multithread
instead of ports_manage_port_operations_one_thread, and use that here
instead of machdev_server. That way there will be one thread per
device_read/write request, not blocking each other. That will however
pose a problem with the lseek-then-read/write way currently used in
device_read/write: two concurrent device_read calls with call lseek
concurrently before calling read concurrently. You can already fix this
now (even before moving to ports_manage_port_operations_multithread) by
calling rump_sys_pread/pwrite instead.
That being said, on even longer run, we may probably not want
to keep one thread per device_read/write request. We'll want to
call rump_sys_aio_read/write instead and return MIG_NO_REPLY from
device_read/write, and send the mig reply once the aio request has
completed. That way, only the aio request will be kept in rumpdisk
memory instead of a whole thread structure. Again, please add this as a
FIXME note, but for much later. Our current linux glue disk driver does
not even do that currently :)
On rump_sys_read error, you need to unmap buf.
Samuel
