[Intel-gfx] significant ioremap leak in i915?

[Breton M. Saunders]

Guys,

   This might be covered elsewhere, but help me come up to speed:  I am 
trying to analyze a leak in i915 that occurs on a digital sinage system 
that I've built.  The system basically is doing a lot of 
XCompositeRedirectWindow / glXBindTexImageEXT calls to render web views 
and mplayer output onto opengl textures for subsequent rendering.

   In my testing, I observe that an enormous block is being lost in 
ioremap by looking at /proc/vmallocinfo:

<snip>
0xffffc90000200000-0xffffc90010201000 268439552 
pci_mmcfg_arch_map+0x33/0x90 phys=e0000000 ioremap
<snip>
0xffffc90010f80000-0xffffc90020f81000 268439552 
i915_driver_load+0x20c/0x6d0 [i915] phys=c0000000 ioremap
<snip>

   So in this example 268 megabytes have been lost.

   Now what is even more vexing is if I stop the software, and X11 both 
blocks are still present on the vmalloc list.  If I then proceed to 
rmmod i915 the i915 entry vanishes, however, the pci_mmcfg_arch_map 
mapping remains.

   It is entirely possible that I am doing something stupid from 
userland (opengl) however, my test runs reliably on an NVidia based 
machine - no leaks, runs until I power it off.  I also think that 
whatever wrongness I may be doing in userland I shouldn't be able to 
blow up the system in this way.

   For clarity, my setup is:
      A NUC 2820 (I.e. Haswell N2820 CPU, 2.13GHz).  1Gb physical ram.
      Operating system: Ubuntu 14.04 (as shipped):
        kernel version 3.13.0-37-generic (i.e. Ubuntu's build)
        Xorg: X.Org X Server 1.15.1
                  Intel xorg module: compiled for 1.15.1, module version 
= 2.99.916
                  Acceleration mode: uxa (although SNA shows no 
difference with regards to the leak).

Any suggestions would be welcome on how to address / analyze this - in 
the meantime I will be digging through the i915 source code to try to 
better understand the problem.

Cheers,

     -Brett