On 9/22/2026 8:52 PM, Brian Cain wrote:
On 9/15/2026 8:07 AM, Peter Maydell wrote:
On Mon, 14 Sept 2026 at 23:32, Brian Cain
<[email protected]> wrote:
From: Matheus Tavares Bernardino <[email protected]>
Bare-metal Hexagon programs use OPENDIR, READDIR, and CLOSEDIR
semihosting calls to enumerate host directories. Directory handles
are shared
by CPUs in a cluster, allowing guest indices to resolve across calls.
Use lock_user_string()/unlock_user() to read the OPENDIR path from
guest
memory instead of copying it into a fixed-size buffer.
CLOSEDIR clears its slot in the directory list, so a stale index
reports
EBADF instead of dereferencing a freed pointer.
Also add a functional test for the new semihosting operations.
Signed-off-by: Matheus Tavares Bernardino
<[email protected]>
Signed-off-by: Brian Cain <[email protected]>
---
include/hw/hexagon/hexagon.h | 14 ++-
hw/hexagon/hex-subsys.c | 32 ++++---
target/hexagon/hexswi.c | 108
++++++++++++++++++++++
tests/functional/hexagon/test_systests.py | 11 +++
4 files changed, 153 insertions(+), 12 deletions(-)
diff --git a/include/hw/hexagon/hexagon.h
b/include/hw/hexagon/hexagon.h
index 62398eeb359..58bdcabe87e 100644
--- a/include/hw/hexagon/hexagon.h
+++ b/include/hw/hexagon/hexagon.h
@@ -11,6 +11,7 @@
#include "system/memory.h"
#include "hw/core/boards.h"
+#include "hw/cpu/cluster.h"
struct hexagon_board_boot_info {
uint64_t ram_size;
@@ -159,6 +160,17 @@ struct hexagon_machine_config {
union hexagon_config_table cfgtable;
};
+#define TYPE_HEXAGON_CLUSTER_STATE "hexagon-cluster-state"
+OBJECT_DECLARE_SIMPLE_TYPE(HexagonClusterState, HEXAGON_CLUSTER_STATE)
+
+struct HexagonClusterState {
+ CPUClusterState parent_obj;
+
+ struct {
+ GList *dir_list;
+ } semihosting;
+};
This looks very odd. Why do we need to track open directory handles
in this special per-cluster structure, but we don't need to do
that for e.g. open file handles returned by SYS_OPEN ? I can't
see anything in the semihosting specification that indicates
that SYS_OPENDIR has special behaviour like that.
Here's where it was previously discussed:
https://lore.kernel.org/qemu-devel/[email protected]/
It's my understanding that unlike with file descriptors, we can't have
a host-portable way to provide directory entries spanning distinct
semihosting calls without keeping state like this. If all host OSs
kept this state, we wouldn't need to.
If there's concerns/objections to the design here then I'll drop this
patch from the series so it can focus on mitigating just the coverity
finding(s). And then I'd put the {open,read,close}dir revisions in its
own series.
thanks
-- PMM
