On 7/4/2025 1:32 AM, Philippe Mathieu-Daudé wrote:
Commit 8c56c1a592b ("memory: emulate ioeventfd") added a !KVM
check because the only accelerator available back then were TCG,
QTest and KVM. Then commit 126e7f78036 ("kvm: require
KVM_CAP_IOEVENTFD and KVM_CAP_IOEVENTFD_ANY_LENGTH") suggested
'!KVM' check should be '(TCG || QTest)'. Later more accelerator
were added. Implement the suggestion as a safety measure, not
dispatching to eventfd when hardware accelerator is used.
Signed-off-by: Philippe Mathieu-Daudé <phi...@linaro.org>
Reviewed-by: Richard Henderson <richard.hender...@linaro.org>
---
system/memory.c | 9 ++-------
1 file changed, 2 insertions(+), 7 deletions(-)
diff --git a/system/memory.c b/system/memory.c
index 76b44b8220f..4f713889a8e 100644
--- a/system/memory.c
+++ b/system/memory.c
@@ -25,7 +25,7 @@
#include "qom/object.h"
#include "trace.h"
#include "system/ram_addr.h"
-#include "system/kvm.h"
+#include "system/qtest.h"
#include "system/runstate.h"
#include "system/tcg.h"
#include "qemu/accel.h"
@@ -1530,12 +1530,7 @@ MemTxResult memory_region_dispatch_write(MemoryRegion
*mr,
adjust_endianness(mr, &data, op);
- /*
- * FIXME: it's not clear why under KVM the write would be processed
- * directly, instead of going through eventfd. This probably should
- * test "tcg_enabled() || qtest_enabled()", or should just go away.
- */
- if (!kvm_enabled() &&
+ if ((tcg_enabled() || qtest_enabled()) &&
The FIXME provides two options:
1. change to "tcg_enabled() || qtest_enabled()"
2. remove !kvm_enabled()
And as the FIXME said, it's not clear why under KVM the write would be
processed directly. Now, the question becomes why under hardware
accelerator is used the write would be processed directly instead of
going through eventfd. I think it needs to answer this question when we
do such change, and it's better to put the answer as the comment in the
code.
memory_region_dispatch_write_eventfds(mr, addr, data, size, attrs)) {
return MEMTX_OK;
}
