mem: Improve memory issue when page size is not 4KiB

[KhoraLee]

On devices that have larger page size than 4KiB(such as aarch64 macOS), sometimes page is overlapped when allocating at specific page(alloc_at) and caused memory issue because overlapped page's memory is zero-fiiled. So if region contains already allocated page then it will only allocate remain parts and update allocation table's info to largest one.

If there is better way then I'll close this PR.

[Macdu]

I think it would be better to make the whole allocator use blocks which are the host page size instead of trying to use workaround like that.

[KhoraLee]

I think it would be better to make the whole allocator use blocks which are the host page size instead of trying to use workaround like that.

I don't understand what you mean. Would you describe it a bit more?

[Macdu]

I think it would be better to make the whole allocator use blocks which are the host page size instead of trying to use workaround like that.

I don't understand what you mean. Would you describe it a bit more?

Ah, I kind of see, so the issue is that some games are trying to allocate two contiguous memory segment with the second one not being 8KB-aligned, right?

[KhoraLee]

Ah, I kind of see, so the issue is that some games are trying to allocate two contiguous memory segment with the second one not being 8KB-aligned, right?

For example, on Uncharted

[16:32:25.181] |D| [load_self]: Loading SELF at app0:eboot.bin... (ELF type: 0xfe00, self_filesize: 0x325b, self_offset: 0x4, module_info_offset: 0x4a4)
[16:32:25.181] |D| [load_self]:     [LOAD] (p_type: 0x1): p_offset: 0x1000, p_vaddr: 0x81000000, p_paddr: 0x0, p_filesz: 0x838, p_memsz: 0x838, p_flags: 0x5, p_align: 0x1000
[16:32:25.181] |D| [load_self]:     [LOAD] (p_type: 0x1): p_offset: 0x2000, p_vaddr: 0x81001000, p_paddr: 0x0, p_filesz: 0xbc, p_memsz: 0xc0, p_flags: 0x6, p_align: 0x1000
[16:32:25.181] |C| [alloc_inner]: Failed to allocate at specific page

4KiB Page
┌──────────────────────┬──────────────────────┐
│    Seg0    │ Padding │    Seg1    │ Padding │
└──────────────────────┴──────────────────────┘
├──── Page 528,384 ────┼──── Page 528,385 ────┤

16KiB Page
┌───────────────────────────────────────────────...──┐
│    Seg0    │ Padding │    Seg1    │ Padding │ ...
└───────────────────────────────────────────────...──┘
├────────────────── Page #132,096 ──────────────...──┤

On 4KiB page size, seg0 and seg1 are allocated in different page.
On 16KiB, Seg0 and Seg1 is in same page and it is calling alloc_inner with forcing same page, Page 132,096(containing Seg0 region) is zero-filled when allocating Seg1 and caused memory issue.

[Macdu]

Yes, I see. This issue is kind of annoying to tackle. Your approach looks kind of fine but there are some issues:

• we can alloc_at at the same position multiple time (fine for the time being because alloc_at is only use for elf non-relocatable segments)
• this only works when alloc_at-ing from left to right
• free does not work with your approach (fine for the time being because we never free elf segments)

[KhoraLee]

@Macdu How about current(93e9e5b) implementation?

[Macdu]

Given that your fix only affects aarch64 mac, which for the time being does not work because of the underlying problem, I am fine with it.