I think it's something with how 896 + 64 + 4193371 just barely goes over 4 MB, but only because of the 64 bytes alignment. For some reason it then blows up.

Actually, this no longer happens in latest version of Odin. It was fixed by this: 98b3a9e

Actually, now I got this again, with the exact same repro as above. Either I messed something up or it's slightly system-dependent.

Actually, this no longer happens in latest version of Odin. It was fixed by this: 98b3a9e

That change only affects resizes. There are no resizes in your repro, so I wouldn't expect that change to have any impact.

Actually, this no longer happens in latest version of Odin. It was fixed by this: 98b3a9e

That change only affects resizes. There are no resizes in your repro, so I wouldn't expect that change to have any impact.

Yeah, I think it only made it slightly more rare in my real-world case. But like you say, in my repro it shouldn't matter.

It does seem to reproduce consistently for me using the current repro code.

This fixes it:

diff --git a/core/mem/virtual/arena.odin b/core/mem/virtual/arena.odin
index 5191505cf..e2fb36688 100644
--- a/core/mem/virtual/arena.odin
+++ b/core/mem/virtual/arena.odin
@@ -108,6 +108,14 @@ arena_alloc :: proc(arena: ^Arena, size: uint, alignment: uint, loc := #caller_l
        switch arena.kind {
        case .Growing:
                needed := mem.align_forward_uint(size, alignment)
+               if arena.curr_block != nil {
+                       ptr := uintptr(arena.curr_block.base[arena.curr_block.used:])
+                       mask := uintptr(alignment)-1
+                       if ptr & mask != 0 {
+                               needed += uint(uintptr(alignment) - (ptr & mask))
+                       }
+               }
+
                if arena.curr_block == nil || (safe_add(arena.curr_block.used, needed) or_else 0) > arena.curr_block.reserved {
                        if arena.minimum_block_size == 0 {
                                arena.minimum_block_size = DEFAULT_ARENA_GROWING_MINIMUM_BLOCK_SIZE

That is essentially calc_alignment_offset that is embedded into alloc_from_memory_block in virtual.odin.

E: The logic to decide to allocate a new block does not account for aligning the current block forward.

Thank you @jasonKercher! Would you mind making a PR with that fix?

I could. That was just proof. Do we think it would be cleaner to expose calc_alignment_offset to avoid duplicating the logic?