memory writes via gdb don't work for memory mapped hardware

[Andreas Rasmusson]

When I remote-debug a qemu-guest and attempt to write to a memory mapped location, the
write-handler for the concerned device will not be called. All write-requiests from
gdb are delegated to cpu_physical_memory_write_rom(...). a function that writes to the
underlying ram-block.

I believe requests to memory mapped hardware should be delegated to
address_space_rw().

example:
;; a memory mapped device. No effect, the write-handler is not called
(gdb) set *0xfff3c000 = 48

;; a ram or rom-block. Thos works. The value is changed.
(gdb) set *0x100000 = 48

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

Here's my suggested patch. As noted in the comment, it could perhaps be
improved for the (rare) case when the write-request from gdb spans multiple
memory regions.

$ git diff 85bc2a15121e8bcd9f15eb75794a1eacca9d84bd HEAD ../exec.c
diff --git a/exec.c b/exec.c
index c4f9036..45ef896 100644
--- a/exec.c
+++ b/exec.c
@@ -3676,6 +3676,7 @@ int cpu_memory_rw_debug(CPUState *cpu, target_ulong addr,
     int l;
     hwaddr phys_addr;
     target_ulong page;
+ bool is_memcpy_access;

     while (len > 0) {
         int asidx;
@@ -3691,13 +3692,32 @@ int cpu_memory_rw_debug(CPUState *cpu, target_ulong addr,
         if (l > len)
             l = len;
         phys_addr += (addr & ~TARGET_PAGE_MASK);
+
         if (is_write) {
+ /* if ram/rom region we access the memory
+ via memcpy instead of via the cpu */
+ hwaddr mr_len, addr1;
+ AddressSpace *as = cpu->cpu_ases[asidx].as;
+ MemoryRegion *mr = address_space_translate(as, phys_addr, &addr1, &mr_len, is_write);
+ is_memcpy_access = memory_region_is_ram(mr) || memory_region_is_romd(mr);
+ if(mr_len < len) {
+ /* TODO, mimic more of the loop over mr chunks as
+ done in cpu_physical_memory_write_internal */
+ printf("warning: we dont know whether all bytes "
+ "to be written are ram/rom or io\n");
+ }
+ }
+ else {
+ is_memcpy_access = false;
+ }
+
+ if (is_write && is_memcpy_access) {
             cpu_physical_memory_write_rom(cpu->cpu_ases[asidx].as,
                                           phys_addr, buf, l);
         } else {
             address_space_rw(cpu->cpu_ases[asidx].as, phys_addr,
                              MEMTXATTRS_UNSPECIFIED,
- buf, l, 0);
+ buf, l, is_write);
         }
         len -= l;
         buf += l;

[Peter Maydell (pmaydell)]

The code has moved around somewhat, but it's still true that writes by gdb don't go to devices -- cpu_memory_rw_debug() calls address_space_write_rom() which calls address_space_write_rom_internal() which simply skips writing for non-ram/rom regions.

I'm not sure if the gdb accesses should be special cased or if we should just make address_space_write_rom() write to devices (which would also affect eg ELF file loading, which is useful in some odd corner cases).

[Thomas Huth (th-huth)]

This is an automated cleanup. This bug report has been moved to QEMU's
new bug tracker on gitlab.com and thus gets marked as 'expired' now.
Please continue with the discussion here:

 https://gitlab.com/qemu-project/qemu/-/issues/213