This is an automated email from Gerrit.
"Antonio Borneo <borneo.anto...@gmail.com>" just uploaded a new patch set to
Gerrit, which you can find at https://review.openocd.org/c/openocd/+/7434
-- gerrit
commit e3c05d54d5f118f88f7ccbbc4f1a74cd8028a204
Author: Antonio Borneo <borneo.anto...@gmail.com>
Date: Wed Jan 4 23:42:00 2023 +0100
server: gdb: ignore mem read during gdb_sync
ARMv8-A provides the feature of Exclusive access instructions [1],
that allows a memory address to be accessed exclusively within a
block of instructions.
The block starts with a 'Load Exclusive' instruction and ends with
a 'Store Exclusive' instruction.
While executing "stepi" command in GDB, GDB detects the Exclusive
access block and executes it atomically by adding a breakpoint at
the end of the block.
GDB detects the Exclusive access blocks by reading from memory the
next instruction, checking if it is a 'Load Exclusive' and then
reading ahead to look for the end of the block [2].
Such GDB memory accesses can cause the target to crash when GDB
uses invalid addresses, e.g. by executing this usual sequence of
commands:
monitor reset halt
monitor gdb_sync
stepi
The command 'gdb_sync' is used to inform OpenOCD that GDB is not
anymore in sync with the target and that a following 'stepi' will
be used to re-sync GDB with target.
But at the very beginning of 'stepi', GDB will issue memory reads
at addresses that were valid 'before' the 'reset halt' but which
are probably not valid anymore.
If by chance the invalid address contains a value corresponding to
a 'Load Exclusive' instruction, GDB will use breakpoints and will
not send the 'stepi' command to OpenOCD, invalidating the sequence
'gdb_sync' + "stepi'.
Preventing the memory access from GDB side, by temporarily setting
some memory attribute, e.g.
set mem inaccessible-by-default
mem 0 4 ro 8 nocache
doesn't work as it causes 'stepi' to return a read error
Cannot access memory at address 0xffff800008e6ad7c
that blocks the execution of 'stepi'.
Make OpenOCD able to ignore the memory read from GDB after the
command 'gdb_sync', as every address known by GDB has now to be
considered as invalid.
Log a debug message and gracefully return zeros to GDB.
Apply this change for every target, not only for aarch64, as the
GDB addresses are always suppose to be invalid after a 'gdb_sync'.
So far, there is no reason to extend this to memory write too.
Link: [1]
https://developer.arm.com/documentation/100934/0100/Exclusive-access-instructions
Link: [2]
https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=gdb/aarch64-tdep.c;hb=refs/tags/gdb-12.1-release#l2876
Change-Id: I6e0554cfbf475ef34d483ee944e1df70ad417e92
Signed-off-by: Antonio Borneo <borneo.anto...@gmail.com>
diff --git a/src/server/gdb_server.c b/src/server/gdb_server.c
index d8dbc2c8b7..ed0f5853c1 100644
--- a/src/server/gdb_server.c
+++ b/src/server/gdb_server.c
@@ -1492,6 +1492,7 @@ static int gdb_read_memory_packet(struct connection
*connection,
char const *packet, int packet_size)
{
struct target *target = get_target_from_connection(connection);
+ struct gdb_connection *gdb_connection = connection->priv;
char *separator;
uint64_t addr = 0;
uint32_t len = 0;
@@ -1519,6 +1520,16 @@ static int gdb_read_memory_packet(struct connection
*connection,
return ERROR_OK;
}
+ if (gdb_connection->sync) {
+ LOG_TARGET_DEBUG(target, "gdb_sync dummy read: addr: 0x%"
PRIx64 ", len: 0x%8.8" PRIx32, addr, len);
+ hex_buffer = malloc(len * 2 + 1);
+ memset(hex_buffer, '0', len * 2);
+ hex_buffer[len * 2] = 0;
+ gdb_put_packet(connection, hex_buffer, len * 2);
+ free(hex_buffer);
+ return ERROR_OK;
+ }
+
buffer = malloc(len);
LOG_DEBUG("addr: 0x%16.16" PRIx64 ", len: 0x%8.8" PRIx32 "", addr, len);
--
