intel_fb: memory allocation failed

[cproc]

I get the following error (with added debug messages) on one of my test machines when I try to run Sculpt:

[init -> runtime -> intel_fb] Lx_kit::Mem_allocator::alloc(): size: 8388608, align: 8388608
[init -> runtime -> intel_fb] Allocator_avl_base::alloc_aligned(): size: 8388608, align: 23, avail: 52093
Allocator_avl_base::alloc_aligned(): size: 8392704, align: 23, avail: 67542831104
Allocator_avl_base::alloc_aligned(): size: 8392704, align: 22, avail: 67542831104
Allocator_avl_base::alloc_aligned(): size: 8392704, align: 21, avail: 67542831104
Allocator_avl_base::alloc_aligned(): size: 8392704, align: 23, avail: 140736137416704
[init -> drivers -> platform] Allocator_avl_base::alloc_aligned(): size: 8396800, align: 23, avail: 4217597952
Allocator_avl_base::alloc_aligned(): size: 8392704, align: 21, avail: 140735321972736
Allocator_avl_base::alloc_aligned(): size: 8392704, align: 21, avail: 140735623946240
[init -> runtime -> intel_fb] Lx_kit::Mem_allocator::alloc() first attempt failed, retrying with 8392703 bytes added
[init -> runtime -> intel_fb] Allocator_avl_base::alloc_aligned(): size: 8388608, align: 23, avail: 8444796
[init -> runtime -> intel_fb] Error: memory allocation failed for 8388608 align 8388608
Error: illegal WRITE at address 0x34 by pager_object: pd='init -> runtime -> intel_fb' thread='ep' ip=0x11d0a04

It works when I add 'align' to the size of the new buffer, but that's another ~8M in this case and depending on how big other allocations can be, might be too much as a general solution.

[cproc]

edit: the first line of the log output was missing. It's

[init -> runtime -> intel_fb] Lx_kit::Mem_allocator::alloc(): size: 8388608, align: 8388608

[alex-ab]

Can you please provide the debugging patch? I don't find any code with "first attempt" or "retrying", so one has to guess what you did.

[cproc]

@alex-ab: commit f9961d6 is the debugging patch

[alex-ab]

Thanks @cproc. Can you please check whether buffer.virt_addr() (line 95) is aligned according to 'align'. If not, then the second attempt can not succeed, since a region is added which does not fulfill the align requirement.

[cproc]

Thanks @cproc. Can you please check whether buffer.virt_addr() (line 95) is aligned according to 'align'. If not, then the second attempt can not succeed, since a region is added which does not fulfill the align requirement.

buffer.virt_addr() is 0x400000 and align is 0x800000.

[alex-ab]

We had a coffee round discussion regarding the topic, I try to summarize. The main issue with the current implementation is this "+ 1" size, which is required to suppress the auto merging of regions, but presumably causes the alignment heuristic in core to not use the intended natural alignment of 23 but something smaller. There exist several options:

• Extend the Genode API such, that the rm_region.attach also consider an explicit align attribute, which currently is not supported. There was mainly the common view, that this should not be done eagerly, as it breaks the ABI. It may fit later better, e.g. when the region management of a component moves from core to the components themself (in form of a utility).
• The other option is to manage the region in lx_kit explicitly, to avoid the auto merging for attached regions as done by the avl implementation currently.
• One may try to add some add_discrete(parameters + align option) in a non ABI breaking way to the avl implementation, which avoids the merging of region.
• If this does not work out, a copy of the avl implementation with this feature in lx_kit is the other option.

At moment no one is working on trying this out. @cproc feel free if you can/want to investigate.

[alex-ab]

Sorry, "should be done eagerly" -> "should not be done eagerly"

[alex-ab]

I updated the text again with a bit more information.