On 11/18/07, Mark Lord <liml@xxxxxx> wrote: > Fajun Chen wrote: > >.. > > I verified your program works in my system and my application works as > > well if changed accordingly. However, this change (indirect IO in sg > > term) may come at a performance cost for IO intensive applications > > since it does NOT utilize mmaped buffer managed by sg driver. Please > > see relevant sg document below: > > http://sg.torque.net/sg/p/sg_v3_ho.html#id2495330 > > http://sg.torque.net/sg/p/sg_v3_ho.html#dmmio > > As an example, sg_rbuf.c in sg3_util package uses SG_FLAG_MMAP_IO flag > > in SG_IO. Please see source code attached. I also noticed that > > MAP_ANONYMOUS is NOT used in mmap() call in sg_rbuf.c, which may not > > be desirable as you pointed out in previous emails. So this brings up > > an interesting sg usage issue: can we use MAP_ANONYMOUS with > > SG_FLAG_MMAP_IO flag in SG_IO? > .. > > The SG_FLAG_MMAP works only with /dev/sg* devices, not /dev/sd* devices. > I don't know which kind you were trying to use, since you still have > not provided your source code for examination. > > If you are using /dev/sg*, then you should be able to get your original mmap() > code to work. But the behaviour described thus far seems to indicate that > your secret program must have been using /dev/sd* instead. > As a matter of fact, I'm using /dev/sg*. Due to the size of my test application, I have not be able to compress it into a small and publishable form. However, this issue can be easily reproduced on my ARM XScale target using sg3_util code as follows: 1. Run printtime.c attached, which prints message to console in a loop. 2. Run sgm_dd (part of sg3_util package, source code attached) on the same system as follows: >sgm_dd if=/dev/sg0 of=/dev/null count=10M bpt=1 The print task can be delayed for as many as 25 seconds. Surprisingly, I can't reproduce the problem in an i386 test system with a more powerful processor. Some clarification to MAP_ANONYMOUS option in mmap(). After fixing a bug and more testing, this option seems to make no difference to cpu load. Sorry about previous report. Back to the drawing board now :-) Thanks, Fajun
Attachment:
printtime.c
Description: Binary data
Attachment:
sgm_dd.c
Description: Binary data
