Re: [Help] : RSS/PSS showing 0 during smaps for Xorg

[Hugh Dickins <hughd@google.com>]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 2331 bytes --]

On Mon, 30 Jan 2012, PINTU KUMAR wrote:
>  
> >If these are ordinary pages with struct pages, then you could probably
> >use a loop of vm_insert_page()s to insert them at mmap time, or a fault
> >routine to insert them on fault.  But as I said, I don't know if this
> >memory is part of the ordinary page pool or not.
>  
> You suggestion about using vm_insert_page() instead of remap_pfn_range worked for me and I got the Rss/Pss information for my driver.

Oh, I'm glad that happened to work for you.

> But still there is one problem related to page fault. 
> If I remove remap_pfn_range then I get a page fault in the beginning. 
> I tried to use the same vm_insert_page() during page_fault_handler for each vmf->virtual_address but it did not work.
> So for time being I remove the page fault handler from my vm_operations.
> But with these my menu screen(LCD screen) is not behaving properly (I get colorful lines on my LCD).
> So I need to handle the page fault properly.
>  
> But I am not sure what is that I need to do inside page fault handler. Do you have any example or references or suggestions?

Sounds like you're not using vm_insert_page() properly: I would not expect
you to get a page fault there once you've set up the area with a loop of
vm_insert_page()s.

Check the comments above it in mm/memory.c ("Your vma protection will
have to be set up correctly" might be relevant).

Compare how you're using it with other users of vm_insert_page() in
the kernel tree.  Sorry, I don't have time to do your debugging.

>  
> >Really, the question has to be, why do you need to see non-0s there?
> I want Rss/Pss value to account for how much video memory is used by the driver for the menu-screen,Xorg processes.

So, userspace does an mmap for a large-enough window, but only some part of
that is filled by the driver (whether by remap_pfn_range or vm_insert_pages),
and you'd like to communicate back how much via the Rss, instead of adding
some ioctl or sysfs interface to the driver?  Fair enough.

I expect userspace could also work it out by touching pages of the area
until it gets a SIGBUS, but that might be too dirty a way of finding out.

Hmm, SIGBUS: maybe that's related to the faults that are puzzling you:
perhaps you're mapping less than you need to.

Hugh