Add Large pages support in GC

[mjsabby]

Fixes https://github.com/dotnet/coreclr/issues/18371.

[mjsabby]

I'm thinking if a better error message should be presented if SeLockMemoryPrivilege is missing as opposed to returning OutOfMemory. My personal opinion is that documentation should be good enough to use this advanced feature rather than plumbing a specific error message out of the runtime.

But maybe we could tackle it in the host or CLI?

[mjsabby]

@Maoni0 I was wrong, Linux does have a concept of Reserve and Commit. PROT_NONE flag in the mmap call is like VirtualReserve. So I have to fix my code accordingly.

[Maoni0]

@mjsabby one thing I missed was we should simply not do anything for gc_heap::decommit_heap_segment_pages 'cause our premise is you grab large pages upfront and don't let them go. I suspect you will get an error if you try to decommit on windows anyway. you could do

void gc_heap::decommit_heap_segment_pages (heap_segment* seg,
                                           size_t extra_space)
{
    if (use_large_pages_p)
        return;

[mjsabby]

I have to introduce OpenProcessToken, AdjutTokenPrivileges and LookupPrivilegeValue for the Windows build. I'm going to do LoadLibrary advapi32 and then resolve those functions. I could also put them in utilcode but I don't think there is any utilcode dependency in the GC->EE/OS interface today.

[janvorli]

I'm going to do LoadLibrary advapi32 and then resolve those functions.

I don't see a reason for using LoadLibrary for these, you can use them directly. They are supported since Windows XP.

[janvorli]

OK, I can see you have already done it that way.

[mjsabby]

@Maoni0 What are your thoughts on hard limit + numa awareness + now large pages? I'm thinking to start with we keep Large Pages NUMA-unaware, otherwise it gets a bit complicated where and how much should the memory be split and commuted at startup. Already the multiple heaps overcommits by 3X I think.

For 3.0 maybe we start with GCHardLimit + Large Pages does a 3X overcommit at startup then in the future we add configs for splitting the memory across LOH, SOH + NUMA Nodes?

[Maoni0]

For 3.0 maybe we start with GCHardLimit + Large Pages does a 3X overcommit at startup

this sounds fine to me if it's fine with you since you are our 1st customer for this :) also I'm working on reducing this 3x to 2x.

then in the future we add configs for splitting the memory across LOH, SOH + NUMA Nodes?

most likely we wouldn't need another config - it'll just be improved naturally in the future versions.

[mjsabby]

This is ready for review again.

After this change, the situation will be:

Memory Type	Windows	Linux
GC Heap	✔️	❌
Card Table	❌	❌
Handle Table	❌	❌
Jitted Code	❌	❌
Other runtime heaps	❌	❌

Next up will be Large Pages support for the GC Heap on Linux.

Documentation Items:

(1) Noting that GCHeapHardLimit is required (GCHeapHardLimit takes input as hex bytes)
(2) SeLockMemoryPrivelege notes for Windows.
(3) For Linux we may have to talk about /proc/sys/vm/overcommit_memory which can be there on some distros, I'm not sure on the implications yet.
(4) MiniDump will be the size of the preallocated gc heap + some ... so they will be quite a bit larger.

[Maoni0]

LGTM - comments above.

[mjsabby]

@Maoni0 @janvorli I've made the requested updates, although I'm still trying to see why my feature detection for restricting to Linux TLB isn't working as expected and failing the macOS build.

[mjsabby]

/azp run coreclr-ci

[mjsabby]

How does one trigger these CI jobs again?

[mjsabby]

@dotnet-bot test coreclr-ci

[mjsabby]

@dotnet-bot test Ubuntu x64 Checked CoreFX Tests

[mjsabby]

@dotnet-bot test coreclr-ci (Test Pri0 Windows_NT x64 checked)

[mjsabby]

@Maoni0 @janvorli Updated with your feedback.

[Maoni0]

LGTM.

[mjsabby]

The coreclr-ci says the following:

Failed: 0 (0.00 %)
Passed: 35,542 (100.00 %)
Other: 0 (0.00 %)
Total: 35,542

But it still marked as failed.

[mjsabby]

Looks like the error is

##[Error 1]
The agent: Hosted macOS 8 lost communication with the server. Verify the machine is running and has a healthy network connection. For more information, see: https://go.microsoft.com/fwlink/?linkid=846610

[mjsabby]

@dotnet-bot test coreclr-ci (Test Pri0 OSX x64 checked)