RFR: 8312132: Add tracking of multiple address spaces in NMT [v94]

[Johan Sjölen]

On Mon, 20 May 2024 10:20:39 GMT, Johan Sjölen <jsjolen at openjdk.org> wrote:

>> Hi,
>> 
>> This PR introduces a new abstraction to NMT, named `MemoryFileTracker`. Today, NMT does not track any memory outside of the virtual memory address space. This means that if you allocated memory in something such as a memory-backed file and use `mmap` to map into that memory, then you'll have trouble reporting this to NMT. This is the situation that ZGC is in, and that is what this patch attempts to fix.
>> 
>> ## `MemoryFileTracker`
>> 
>> The `MemoryFileTracker` adds the ability of adding new virtual memory address spaces to NMT and committing memory to these, the basic API is:
>> 
>> ```c++
>>     static MemoryFile* make_device(const char* descriptive_name);
>>     static void free_device(MemoryFile* device);
>> 
>>     static void allocate_memory(MemoryFile* device, size_t offset, size_t size,
>>                                 MEMFLAGS flag, const NativeCallStack& stack);
>>     static void free_memory(MemoryFile* device, size_t offset, size_t size);
>> 
>> 
>> It is easiest to see how this is used by looking at what ZGC's `ZNMT` class does:
>> 
>> ```c++
>> void ZNMT::reserve(zaddress_unsafe start, size_t size) {
>>   MemTracker::record_virtual_memory_reserve((address)start, size, CALLER_PC, mtJavaHeap);
>> }
>> void ZNMT::commit(zoffset offset, size_t size) {
>>   MemTracker::allocate_memory_in(ZNMT::_device, static_cast<size_t>(offset), size, mtJavaHeap, CALLER_PC);
>> }
>> void ZNMT::uncommit(zoffset offset, size_t size) {
>>   MemTracker::free_memory_in(ZNMT::_device, (size_t)offset, size);
>> }
>> 
>> void ZNMT::map(zaddress_unsafe addr, size_t size, zoffset offset) {
>>   // NMT doesn't track mappings at the moment.
>> }
>> void ZNMT::unmap(zaddress_unsafe addr, size_t size) {
>>   // NMT doesn't track mappings at the moment.
>> }
>> 
>> 
>> As you can see, any mapping between reserved regions and device-allocated memory is not recorded in NMT. This means that in detailed mode you only get reserved regions printed for the reserved memory, the device-allocated memory is reported separately. When performing summary reporting any memory allocated via these devices is added to the corresponding `MEMFLAGS` as `committed` memory. 
>> 
>> This patch is also acting as a base on which we deploy multiple new backend ideas to NMT. These ideas are:
>> 
>> 1. Implement VMA tracking using a balanced binary tree approach. Today's `VirtualMemoryTracker`'s usage of linked lists is slow and brittle, we'd like to move away from it. Our Treap-based approach in this patch gives a performance bo...
>
> Johan Sjölen has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - Fix memory leak, store the links in NativeCallStackStorage in Arena
>  - Accidental issue with docs, remove

In a future RFE NativeCallStackStorage can store its links and references in a linear free list allocator, cutting down from 16 bytes per link to 8 bytes. We may be able to cut down on the number of buckets significantly, as we'll have better locality on our side.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/18289#issuecomment-2120158986