direct-to-BIO for O_DIRECT

After disabling kiobufs in sg I would like such a drop in replacement.

Agreed. Tape devices could be added to your list. Large page support will make for very efficient zero copy IO.

CIOs?

Doug Gilbert

Ben had lightweight sg structures called kvecs' and kveclets'. And library functions to map pages into them. And code to attach them to BIOs. So we'll be looking at getting that happening.

The other common requirement (used in several places in the kernel, and in LVM2) is the ability to perform bulk I/O against a blockdev - simply read and write a chunk of disk into a list of kernel pages. So we'll need a
library function for that. And the O_DIRECT/raw implementation can be bent
around to use those things.

Haven't thought about large pages. We don't seem to have an implementation of them yet, and I'm not sure how the DMA mapping API would get along with them.

-

BIOs are for BLOCK devices we want sth. like this for CHARACTER devices.

I just want sth. along the lines of this:

/* Pin down (COMPLETE!) user pages and put them into a scatter gather list */ int sg_map_user_pages(struct scatterlist *sgl, const unsigned int nr_pages,

		unsigned long uaddr, int rw) {
	int res, i;
	struct page *pages[nr_pages];

	down_read(&current->mm->mmap_sem);
	res = get_user_pages(
			current,
			current->mm,
			uaddr,
			nr_pages,
			rw == READ, /* logic is perversed^Wreversed here :-( */
			0, /* don't force */
			&pages[0],
			NULL);
	up_read(&current->mm->mmap_sem);

	/* Errors and no page mapped should return here */
	if (res <= 0) return res;

	for (i=1; i < res; i++) {
		sgl[i].page = pages[i];	
	}
	return res;
}

/* And unmap them... */

int sg_unmap_user_pages(struct scatterlist *sgl, const unsigned int nr_pages) {
	int i;

	for (i=0; i < nr_pages; i++)
		page_cache_release(sgl[i].page);

	return 0;
}

Possibly more complicated and less error prone, but you get the idea ;-)

Regards

Ingo Oeser

Science is what we can tell a computer. Art is everything else. --- D.E.Knuth

Ok, I've looked at them and they don't help me at all.

A user, who splits its IO into single pages, wants to do DMA and needs bus addresses for that. So he needs "struct scatterlist".

If one doesn't need to DMA, one can do copy_{from,to}_user directly with an immediate buffer, so the splitup isn't needed.

For this I conclude, that using the EXISTING 'struct scatterlist' will be enough for both. Attaching a vector of these to the BIOs is no problem. Neither it is for CHARACTER device IOs (CIOs).

So by using this simple abstraction we MIGHT waste only 4-8 bytes per page submitted, but by page-splitting the IO only for devices, that need DMA (e.g. make the request that explicitly) we don't really waste it and support BIOs and CIOs the same way.

I will refine that code for my own uses anyway, so if nobody with more clues about IO than me implements it, I will submit it later.

Regards

Ingo Oeser

Science is what we can tell a computer. Art is everything else. --- D.E.Knuth

And as I recall, a grand plan was hatched at the kernel summit to slice and dice all the various forms of block IO into that model.

Seeing -> believing

Daniel