On 1/15/07, David Brownell <david-b@xxxxxxxxxxx> wrote:
On Monday 15 January 2007 5:54 pm, Nate Diller wrote: > This removes the aio implementation from the usb gadget file system. NAK. I see a deep mis-understanding here. > Aside > from making very creative (!) use of the aio retry path, it can't be of any > use performance-wise Other than the basic win of letting one userspace thread keep an I/O stream active while at the same time processing the data it reads or writes?? That's the "async" part of AIO. There's a not-so-little thing called "I/O overlap" ... which is the only way to prevent wasting bandwidth between (non-cacheable) I/O requests, and thus is the only way to let userspace code achieve anything close to the maximum I/O bandwidth the hardware can achieve. We want to see the host side "usbfs" evolve to support AIO like this too, for the same reasons. (Currently it has fairly ugly AIO code that looks unlike any other AIO code in Linux. Recent updates to support a file-per-endpoint device model are a necessary precursor to switching over to standard AIO syscalls.) > because it always kmalloc()s a bounce buffer for the > *whole* I/O size. By and large that's a negligible factor compared to being able to achieve I/O overlap. ISTR the reason for not doing fancy DMA magic was that the cost of this style AIO was under 1 KByte object code on ARM, which was easy to justify ... while DMA magic to do that sort of stuff would be much fatter, as well as more error prone. (And that's why the "creative" use of the retry path. As I've observed before, "retry" is a misnomer in the general sense of an async I/O framework. It's more of a semi-completion callback; I/O can't in general be "retried" on error or fault, and even in the current usage it's not really a "retry".) Now that high speed peripheral hardware is becoming more common on embedded Linuxes -- TI has DaVinci, OMAP 2430, TUSB6010 (as found in the new Nokia 800 tablets); Atmel AVR32 AP7000; at least a couple parts that should be able to use the same musb_hdrc driver as those TI parts; and a few other chips I've heard of -- there may be some virtue in eliminating the memcpy, since those CPUs don't have many MIPS to waste. (Iff the memcpy turns out to be a real issue...) > Perhaps the only reason to keep it around is the ability > to cancel I/O requests, which only applies when using the user space async > I/O interface. It's good to have almost the complete kernel API functionality exposed to userspace, and having I/O cancelation is an inevitable consequence of a complete AIO framework ... but that particular issue was not a driving concern. The reason for AIO is to have a *STANDARD* userspace interface for *ASYNC I/O* which otherwise can't exist. You know, the kind of I/O interface that can't be implemented with read() and write() syscalls, which for non-buffered I/O necessarily preclude all I/O overlap. AIO itself is a direct match to most I/O frameworks' primitives. (AIOCB being directly analagous to peripheral side "struct usb_request" and host side "struct urb".) You know, I've always thought that one reason the AIO discussions seemed strange is that they weren't really focussed on I/O (the lowlevel after-the-caches stuff) so much as filesystems (several layers up in the stack, with intervening caching frameworks). The first several implementations of AIO that I saw were restricted to "real" I/O and not applicable to disk backed files. So while I was glad the Linux approach didn't make that mistake, it's seemed that it might be wanting to make a converse mistake: neglecting I/O that isn't aimed at data stored on disks. > I highly doubt that is enough incentive to justify the extra > complexity here or in user-space, so I think it's a safe bet to remove this. > If that feature still desired, it would be possible to implement a sync > interface that does an interruptible sleep. What's needed is an async, non-sleeeping, interface ... with I/O overlap. That's antithetical to using read()/write() calls, so your proposed approach couldn't possibly work.
haha, wow ok you convinced me :) I got a bit impatient when I was working on this, it took some time just to figure out the intention of the code, and I'm trying to hold to a bit of a schedule here. Without any clear (to me) reason, I didn't want to spend a lot of effort fixing this up. There's really no big difference between the usb drivers here and the disk I/O scheduler queue, AFAICT, so it seems like the solution I want is to do a kmap() on the user buffer and then do the I/O straight out of that. That will eliminate the need for the bounce buffer. I'll post a new version along with the iodesc changes later this week. NATE - To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
