On some systems, Linux will log this:
mtrr_cleanup: can not find optimal value
...
please specify mtrr_gran_size/mtrr_chunk_size
This primarily happens on legacy BIOS boot systems with both CPU integrated and discrete graphics chipsets with muxless graphics switching. Historically, this was mostly observed on x86 systems with Intel CPUs and integrated graphics, when combined with NVidia graphics cards ("Optimus"). Nowadays, this could also happen with AMD CPUs with integrated graphics (APU) and a separate, dedicated graphics card.
Adam Helbing wrote a nice article on this topic, originally published on his blog under CC-BY-NC-SA 3.0 license. It is no longer available at the original location but preserved in archives (Archive.today, Internet Archive).
Adam kindly re-licensed it to me (Tom Reynolds) under the CC-BY-SA 4.0 license. You are welcome to create derivative works - see the license deed for details.
A copy of the original article, converted to Markdown, with only minimal changes applied (formatting, Wikipedia link, ortography) is available below. Comments published with the original article were useful to me, but were contributions under the articles' original CC-BY-NC-SA 3.0 license. I have paraphrased them here to make them reusable under the new license.
by Adam Helbing (originally published on May 23, 2012)
If you find one or several of the following lines in your kernel message buffer, your system suffers from memory management problems:
mtrr_cleanup: can not find optimal value
please specify mtrr_gran_size/mtrr_chunk_size
MTRR allocation failed. Graphics performance may suffer.
Keep on reading for some background info and a few steps to solve that problem.
Memory Type Range Registers are basically a table which tells the system how to cache which ranges of installed memory. It is set up by bios initially, but can be altered anytime by the OS if needed. Whenever you change the amount of memory installed or flash a new bios, a new mtrr table is computed. Here is a sane mtrr table from my Lenovo T400 notebook with 6 GiB of ram:
$ cat /proc/mtrr
reg00: base=0x000000000 ( 0MB), size= 2048MB, count=1: write-back
reg01: base=0x080000000 ( 2048MB), size= 1024MB, count=1: write-back
reg02: base=0x0bc000000 ( 3008MB), size= 64MB, count=1: uncachable
reg03: base=0x100000000 ( 4096MB), size= 2048MB, count=1: write-back
reg04: base=0x180000000 ( 6144MB), size= 1024MB, count=1: write-back
reg05: base=0x1bc000000 ( 7104MB), size= 64MB, count=1: uncachable
reg06: base=0x0d0000000 ( 3328MB), size= 256MB, count=1: write-combining
As you see, my Laptop has 7 registers available, there are some small uncachable areas defined (I don't know why, some bios stuff I suppose..), but the big chunks are set to write back. The last is a special one: It is the Intel onboard graphic cards memory, which supports write-combining for best performance.
Propably a lot. In my case, I found the following deeply disturbing lines inside the kernel buffer:
mtrr: no more MTRRs available
[drm] MTRR allocation failed. Graphics performance may suffer.
What happened? The drm module tried to grab a free mtrr to set up a write-combined cache area for the graphic cards memory, but all 7 register were occupied by some other cache-declarations already, so that failed. However, the machine works fine without, 3D and compositing work, as do games and everything, just not as fast as they could.
Luckily, the kernel has some options to control mtrr setup:
$ gunzip < /proc/config.gz | grep -i MTRR_SANITIZER
CONFIG_MTRR_SANITIZER=y
CONFIG_MTRR_SANITIZER_ENABLE_DEFAULT=0
CONFIG_MTRR_SANITIZER_SPARE_REG_NR_DEFAULT=1
Entry MTRR_SANITIZER! The first option switches it on or off, if its off in your kernel, you need to rebuild it with sanitizer support enabled. The following lines configure the sanitizer itself, the first one switches it on or off by default, the next one tells the sanitizer, whether he shall left a number of registers empty on boot for later use. Bingo! We need one register left untouched for drm to grab it. One can enable the sanitizer by default, set the spare reg to 1 and reboot, but as its unusual to build kernel for yourself these times, there are 2 kernel parameter which overload the kernel config. just add
enable_mtrr_cleanup mtrr_spare_reg_nr=1
to your grub, syslinux, lilo, whatever kernel line and reboot to sanitize your mtrrs.
Propably yes. Just check the message buffer whether the mtrr allocation failed
message is gone and check your /proc/mtrr
whether there is a write-combined area defined for the graphics. In my case, it still was not. Instead, I found these new lines:
mtrr_cleanup: can not find optimal value
please specify mtrr_gran_size/mtrr_chunk_size
The sanitizer was not able to find an optimal mtrr setup, but he kindly printed a long list with possible setups for me into the buffer, here is just the relevant snippet, as that list is really long:
gran_size: 16M chunk_size: 128M num_reg: 7 lose cover RAM: 14M
gran_size: 16M chunk_size: 256M num_reg: 7 lose cover RAM: 14M
gran_size: 16M chunk_size: 512M num_reg: 7 lose cover RAM: 14M
gran_size: 16M chunk_size: 1G num_reg: 7 lose cover RAM: 14M
gran_size: 16M chunk_size: 2G num_reg: 7 lose cover RAM: 14M
gran_size: 32M chunk_size: 32M num_reg: 7 lose cover RAM: 478M
gran_size: 32M chunk_size: 64M num_reg: 7 lose cover RAM: 478M
gran_size: 32M chunk_size: 128M num_reg: 6 lose cover RAM: 30M
gran_size: 32M chunk_size: 256M num_reg: 6 lose cover RAM: 30M
gran_size: 32M chunk_size: 512M num_reg: 6 lose cover RAM: 30M
The problem is, poor sanitizer found no configuration which meets my requirements of using 6 mtrrs only, and still cover all memory. So he decided to let me choose a tradeoff config for myself. Look at the table, gran_size and chunk_size are the config parameters, num_reg shows how many mtrrs are needed for that config and 'lose cover ram' tells us how many ram we lose when this config is chosen. For my setup, num_reg has to be 6 or less, and I want to lose as little ram as possible, so I chose the first line which loses 30M only.
Finally, appending
enable_mtrr_cleanup mtrr_spare_reg_nr=1 mtrr_gran_size=32M mtrr_chunk_size=128M
to the boot parameters made my box happy again :-) Now 24MiB of ram are gone, but my 3d workloads, which are basically webgl content and the hidden flight simulator inside google earth, run both at significant better framerates.
Q: In the examples given in this article, you state that 6 MTRRs are required. How do you calculate this?
A: The num_reg
value logged by Linux provides the MTRRs required:
gran_size: 32M chunk_size: 128M num_reg: 6 lose cover RAM: 30M
Q: *But why does it have to be exactly 6 MTRRs or less, how did you choose this very line from the many possible combinations Linux suggests there?
Here is what it looks like for me:*
$ cat /proc/mtrr
reg00: base=0x23c000000 ( 9152MB), size= 64MB, count=1: uncachable
reg01: base=0x0be000000 ( 3040MB), size= 32MB, count=1: uncachable
reg02: base=0x000000000 ( 0MB), size= 2048MB, count=1: write-back
reg03: base=0x080000000 ( 2048MB), size= 1024MB, count=1: write-back
reg04: base=0x100000000 ( 4096MB), size= 4096MB, count=1: write-back
reg05: base=0x200000000 ( 8192MB), size= 1024MB, count=1: write-back
reg06: base=0x0bde00000 ( 3038MB), size= 2MB, count=1: uncachable
A: In my experience, on computers with an Intel CPU with integrated graphics, when you examine how many MTRRs to set, one register needs to be left available for the Intel DRM (direct rendering module) integrated graphics driver (i915/i965). It will use it as write-back cache for onboard graphics memory.
Your CPU, just like the one in the article, supports 7 MTRRs (reg00
to reg06
). So if you leave one for the integrated GPU, you can go with a 6 MTRR configuration, and mtrr_spare_reg_nr=1
. While running with these settings, check the kernel log (dmesg
) to confirm whether the Intel DRM uses the leftover MTRR.