trace-writes

/probe:trace

Find what code reads or writes to a specific memory address.

Arguments

• address (required): Memory address to monitor (hex)
• access (optional): "w" for writes only (default), "rw" for reads and writes
• size (optional): Watch size in bytes: 1, 2, 4, or 8 (default: 4)

Steps

1. Set a hardware watchpoint:

   probe_hwbp_set address=<addr> access=<w|rw> size=<size>
   

   You get back a breakpoint ID. Hardware breakpoints use debug registers DR0-DR3 (4 slots max).

2. Wait for the target action -- Tell the user what to do in the application to trigger the write. For example: "Take damage in-game" or "Click the button" or "Send a message".

3. Read the breakpoint log:

   probe_breakpoint_log id=<bp_id>
   

   Each hit gives you a full register snapshot: RIP (the instruction that accessed the address), RAX-R15, RSP, RFLAGS.

4. Disassemble the writer:

   probe_disassemble address=<rip> count=10
   

   Look at the instruction at RIP — it's the exact instruction that wrote (or read) the address.

5. Get the full function:

   probe_disassemble_function address=<function_start>
   

   Walk backwards from RIP to find the function prologue (look for push rbp / sub rsp / push rbx patterns). Or subtract 0x40-0x100 from RIP and disassemble — the prologue is usually within that range.

6. Identify the caller:

   • RSP from the breakpoint log points to the return address on the stack
   • probe_read_pointer address=<rsp_value> reads the return address
   • Disassemble there to see who called this function

7. Generate a signature for the function so it can be found again:

   probe_generate_signature address=<function_start>
   

8. Clean up:

   probe_hwbp_remove id=<bp_id>
   

   Always remove hardware breakpoints when done — there are only 4 slots.

If the breakpoint doesn't fire

• The value might be written by a different thread — hardware BPs work across all threads, so this is unlikely. More likely the value is written once during initialization and you missed it.
• The value might be computed, not stored — the address might be read from somewhere else and cached. Try watching the source address instead.
• The write might happen via DMA or kernel mode — hardware breakpoints only catch usermode access. If you suspect kernel writes, this approach won't work.
• Try polling instead: probe_watch address=<addr> size=4 interval_ms=100 to catch when the value changes, then narrow down the timing.

Common patterns you'll see

Direct field write:

mov [rcx+0x354], eax    ; rcx = object pointer, 0x354 = field offset, eax = new value

This tells you: the function receives the object in RCX, and offset 0x354 is the field.

Indirect write via pointer chain:

mov rax, [rcx+0x30]     ; load sub-object pointer
mov [rax+0x10], edx     ; write to sub-object field

This tells you: the struct has a pointer at +0x30 to a sub-struct, and the actual field is at +0x10 in the sub-struct.

Atomic/interlocked write:

lock xchg [rcx+0x40], eax   ; thread-safe write

This means the field is accessed from multiple threads — important to know for reading it safely.