memory: reintroduce BQL-free fine-grained PIO/MMIO

This patch brings back Jan's idea [1] of BQL-free IO access

This will let us make access to ACPI PM/HPET timers cheaper,
and prevent BQL contention in case of workload that heavily
uses the timers with a lot of vCPUs.

1) 196ea13104f (memory: Add global-locking property to memory regions)
   ... de7ea885c539 (kvm: Switch to unlocked MMIO)

Signed-off-by: Igor Mammedov <imammedo@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Link: https://lore.kernel.org/r/20250814160600.2327672-2-imammedo@redhat.com
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

3 files changed, 28 insertions, 1 deletions

diff --git a/include/system/memory.h b/include/system/memory.h
index e2cd6ed..aa85fc2 100644
--- a/include/system/memory.h
+++ b/include/system/memory.h
@@ -833,6 +833,7 @@ struct MemoryRegion {
     bool nonvolatile;
     bool rom_device;
     bool flush_coalesced_mmio;
+    bool lockless_io;
     bool unmergeable;
     uint8_t dirty_log_mask;
     bool is_iommu;
@@ -2342,6 +2343,17 @@ void memory_region_set_flush_coalesced(MemoryRegion *mr);
 void memory_region_clear_flush_coalesced(MemoryRegion *mr);
 
 /**
+ * memory_region_enable_lockless_io: Enable lockless (BQL free) acceess.
+ *
+ * Enable BQL-free access for devices that are well prepared to handle
+ * locking during I/O themselves: either by doing fine grained locking or
+ * by providing lock-free I/O schemes.
+ *
+ * @mr: the memory region to be updated.
+ */
+void memory_region_enable_lockless_io(MemoryRegion *mr);
+
+/**
  * memory_region_add_eventfd: Request an eventfd to be triggered when a word
  *                            is written to a location.
  *
diff --git a/system/memory.c b/system/memory.c
index 5646547..44701c4 100644
--- a/system/memory.c
+++ b/system/memory.c
@@ -2546,6 +2546,21 @@ void memory_region_clear_flush_coalesced(MemoryRegion *mr)
     }
 }
 
+void memory_region_enable_lockless_io(MemoryRegion *mr)
+{
+    mr->lockless_io = true;
+    /*
+     * reentrancy_guard has per device scope, that when enabled
+     * will effectively prevent concurrent access to device's IO
+     * MemoryRegion(s) by not calling accessor callback.
+     *
+     * Turn it off for lock-less IO enabled devices, to allow
+     * concurrent IO.
+     * TODO: remove this when reentrancy_guard becomes per transaction.
+     */
+    mr->disable_reentrancy_guard = true;
+}
+
 void memory_region_add_eventfd(MemoryRegion *mr,
                                hwaddr addr,
                                unsigned size,
diff --git a/system/physmem.c b/system/physmem.c
index e5dd760..f498572 100644
--- a/system/physmem.c
+++ b/system/physmem.c
@@ -2900,7 +2900,7 @@ bool prepare_mmio_access(MemoryRegion *mr)
 {
     bool release_lock = false;
 
-    if (!bql_locked()) {
+    if (!bql_locked() && !mr->lockless_io) {
         bql_lock();
         release_lock = true;
     }