Menu
▾
▴
user-mode-linux-devel
|
From: Gerd S. <in...@ge...> - 2006-03-26 18:49:36
|
Hi, I recently tried to benchmark the new AIO-based ubd driver (including O_DIRECT), and ran into serious problems with filesystems > 4GB. I think I found the problem; two overflows in ubd_kern.c (see patch). My test system was 2.6.16 plus all patches from user-mode-linux.sf.net minus no_cow_odirect. Symptoms were that ext3 complained that it was tried to write blocks into the system zone, and switched the filesystem to r/o. I also experienced (hard) lockups when doing lots of I/O. But this seems to be unrelated to this problem (but who knows). Gerd --- linux-2.6.16/arch/um/drivers/ubd_kern.c.orig 2006-03-26 16:57:11.000000000 +0200 +++ linux-2.6.16/arch/um/drivers/ubd_kern.c 2006-03-26 18:16:02.000000000 +0200 @@ -1228,7 +1228,7 @@ static void do_ubd_request(request_queue while(dev->start_sg < dev->end_sg){ struct scatterlist *sg = &dev->sg[dev->start_sg]; - prepare_request(req, &io_req, req->sector << 9, + prepare_request(req, &io_req, (unsigned long long) req->sector << 9, sg->offset, sg->length, sg->page); if(do_io(&io_req, req, dev) == -EAGAIN) @@ -1573,7 +1573,7 @@ static int do_io(struct io_thread_req *r } off = req->offsets[bit] + req->offset + - start * req->sectorsize; + (unsigned long long) start * req->sectorsize; len = (end - start) * req->sectorsize; buf = &req->buffer[start * req->sectorsize]; -- ------------------------------------------------------------ Gerd Stolpmann * Viktoriastr. 45 * 64293 Darmstadt * Germany ge...@ge... http://www.gerd-stolpmann.de Phone: +49-6151-153855 Fax: +49-6151-997714 ------------------------------------------------------------ |
|
From: Jeff D. <jd...@ad...> - 2006-03-27 15:37:33
|
On Sun, Mar 26, 2006 at 08:49:25PM +0200, Gerd Stolpmann wrote: > I recently tried to benchmark the new AIO-based ubd driver (including > O_DIRECT), and ran into serious problems with filesystems > 4GB. I think > I found the problem; two overflows in ubd_kern.c (see patch). The first bit looks reasonable. I would change start to an unsigned long long instead of casting, though. > I also experienced (hard) lockups when doing lots of I/O. But this seems > to be unrelated to this problem (but who knows). What was your workload? Jeff |
|
From: Gerd S. <in...@ge...> - 2006-03-27 15:47:42
|
Am Montag, den 27.03.2006, 10:38 -0500 schrieb Jeff Dike: > On Sun, Mar 26, 2006 at 08:49:25PM +0200, Gerd Stolpmann wrote: > > I recently tried to benchmark the new AIO-based ubd driver (including > > O_DIRECT), and ran into serious problems with filesystems > 4GB. I think > > I found the problem; two overflows in ubd_kern.c (see patch). > > The first bit looks reasonable. I would change start to an unsigned > long long instead of casting, though. > > > I also experienced (hard) lockups when doing lots of I/O. But this seems > > to be unrelated to this problem (but who knows). > > What was your workload? I used iozone in various ways. The simplest workload leading to problems was: - Write a 5GB file - Write it again - Read it back - Read it back This benchmark ran fine the first time I started it. When I started it again, UML locked up in the middle. This is reproducible. Gerd -- ------------------------------------------------------------ Gerd Stolpmann * Viktoriastr. 45 * 64293 Darmstadt * Germany ge...@ge... http://www.gerd-stolpmann.de Phone: +49-6151-153855 Fax: +49-6151-997714 ------------------------------------------------------------ |
|
From: Jeff D. <jd...@ad...> - 2006-03-29 19:32:27
|
On Mon, Mar 27, 2006 at 05:47:31PM +0200, Gerd Stolpmann wrote: > I used iozone in various ways. The simplest workload leading to problems > was: > > - Write a 5GB file > - Write it again > - Read it back > - Read it back > > This benchmark ran fine the first time I started it. When I started it > again, UML locked up in the middle. This is reproducible. I'm interpretting that description as being this command: iozone -s 5G -i 0 -i 0 -i 1 -i 1 Yes? And it runs fine here. Jeff |
|
From: Gerd S. <in...@ge...> - 2006-03-29 21:16:43
|
Am Mittwoch, den 29.03.2006, 14:33 -0500 schrieb Jeff Dike: > On Mon, Mar 27, 2006 at 05:47:31PM +0200, Gerd Stolpmann wrote: > > I used iozone in various ways. The simplest workload leading to problems > > was: > > > > - Write a 5GB file > > - Write it again > > - Read it back > > - Read it back > > > > This benchmark ran fine the first time I started it. When I started it > > again, UML locked up in the middle. This is reproducible. > > I'm interpretting that description as being this command: > iozone -s 5G -i 0 -i 0 -i 1 -i 1 > > Yes? Almost: iozone -i 0 -i 1 -r 64k -s 5G -e -f /Data/testfile -U /Data i.e. record size is 64K, and there is a umount/mount between the runs. /Data is an ext3 filesystem with 4k block size, on a 20 G partition. > And it runs fine here. For me it runs fine only the first time I call it. The next time UML locks up. Can I do something to find out more details? Gerd -- ------------------------------------------------------------ Gerd Stolpmann * Viktoriastr. 45 * 64293 Darmstadt * Germany ge...@ge... http://www.gerd-stolpmann.de Phone: +49-6151-153855 Fax: +49-6151-997714 ------------------------------------------------------------ |
|
From: Jeff D. <jd...@ad...> - 2006-03-30 18:48:19
|
On Wed, Mar 29, 2006 at 11:16:27PM +0200, Gerd Stolpmann wrote: > Almost: > > iozone -i 0 -i 1 -r 64k -s 5G -e -f /Data/testfile -U /Data > > i.e. record size is 64K, and there is a umount/mount between the > runs. /Data is an ext3 filesystem with 4k block size, on a 20 G > partition. OK, this doesn't run fine for me even the first time, so I have something I can chew on. BTW, your earlier fixes are now in my patchset. Jeff |
|
From: Jeff D. <jd...@ad...> - 2006-05-08 17:26:59
|
On Wed, Mar 29, 2006 at 11:16:27PM +0200, Gerd Stolpmann wrote: > Almost: > > iozone -i 0 -i 1 -r 64k -s 5G -e -f /Data/testfile -U /Data > > i.e. record size is 64K, and there is a umount/mount between the > runs. /Data is an ext3 filesystem with 4k block size, on a 20 G > partition. > > > And it runs fine here. > > For me it runs fine only the first time I call it. The next time UML > locks up. > > Can I do something to find out more details? I've been playing with this, and iozone can't make UML do anything bad. Is this the only scenario you have? Jeff |
×
Want the latest updates on software, tech news, and AI?
Get latest updates about software, tech news, and AI from SourceForge directly in your inbox once a month.
Thanks for helping keep SourceForge clean.
X