7/23 If I have data in my process address space on Solaris,
and I want to get that memory onto disk without doing
a copy, is there a way to do this? --jwm
\_ What does "without doing a copy" mean? As always, why don't
you tell us what you're trying to do instead of asking for a
specific technical answer to do it your way.
\_ "Without a copy" means without making a copy in memory.
(i.e. no copy from the userland buffer to a kernel buffer)
I have a chunk of data in memory, I want to get it onto
a particular disk via a raw device.
a particular disk via a raw device. Why I want to do this
is not of any great importance to the question.
\_ i think what he's asking is pretty straightforward, what is
your problem? jwm, do you know if mmapped files in solaris
go through the block buffers?
\_ my problem is people on the motd tend to ask vague technical
questions which don't always mean what they seem to. what's
your problem with my asking what he's trying to do so he can
get helped better?
\_ Yeah, I thought of mmap(), but here is the problem with
that. I map the file, then when I write to a page,
I get a fault that reads in the the data off disk.
This is data I am just about to overwrite, plus this
also has a copy from my buffer to the mmap'ed region.
If I could take my buffer and map it onto the file
that would work, but I don't think this is supported.
The 2 methods I can think of that may work are write()
and aio_write() because they both "own" the buffer during
the time the write is happening. If I had solaris source
this would be easy to check, but I don't. The idea is
that Solaris would simply point the DMA engine at my
buffer and never look at my data. --jwm
\_ my point was that you could mmap your buffer instead of
allocating it with malloc() (or on the stack).
i don't see why write() helps, as it does a write to
the block buffers, not to disk. i don't know anything
about aio_write. is there no way in solaris to lock
a block buffer and write to it directly? that's how it
fucking works in the kernel.
\_ It's a raw device I'm writing to, so no block buffers.
I can't mmap instead, because I'm not generating the
data. I guess the best way to describe what I want is
to say I want to avoid a copyin() and have DMA go
directly from my memory to the device. --jwm
\_ why do you suspect write() does a copyin() when
writing to block device? when i write a driver,
i mlock() then setup dma from the user's address
space.
\_ I didn't know, but I suspected that write may not
have the copy (See above). However as a
block devices are probably doing the copyin() in
all cases because they interact with the buffer
cache, it is raw/character devices that don't. |
http://csua.com/feed/