[Qemu-devel] [RFC] Coalesced MMIO support

[Anthony Liguori]

The images I need to test this against are not on this computer, so I 
can't commit until tomorrow when I have access to them again.  So I 
figured that was a good reason to post first as an RFC.

MMIO exits are more expensive in KVM or Xen than in QEMU because they 
involve, at least, privilege transitions.  However, MMIO write 
operations can be effectively batched if those writes do not have side 
effects.

Good examples of this include VGA pixel operations when in a planar 
mode.  As it turns out, we can get a nice boost in other areas too.  
Laurent mentioned a 9.7% performance boost in iperf with the coalesced 
MMIO changes for the e1000 when he originally posted this work for KVM.

Regards,

Anthony Liguori

Index: kvm-all.c
===================================================================
--- kvm-all.c   (revision 5874)
+++ kvm-all.c   (working copy)
@@ -24,6 +24,9 @@
 #include "sysemu.h"
 #include "kvm.h"
 
+/* KVM uses PAGE_SIZE in it's definition of COALESCED_MMIO_MAX */
+#define PAGE_SIZE TARGET_PAGE_SIZE
+
 //#define DEBUG_KVM
 
 #ifdef DEBUG_KVM
@@ -52,6 +55,7 @@
     KVMSlot slots[32];
     int fd;
     int vmfd;
+    int coalesced_mmio;
 };
 
 static KVMState *kvm_state;
@@ -228,6 +232,44 @@
     qemu_free(d.dirty_bitmap);
 }
 
+int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
+{
+    int ret = -ENOSYS;
+#ifdef KVM_CAP_COALESCED_MMIO
+    KVMState *s = kvm_state;
+
+    if (s->coalesced_mmio) {
+        struct kvm_coalesced_mmio_zone zone;
+
+        zone.addr = start;
+        zone.size = size;
+
+        ret = kvm_vm_ioctl(s, KVM_REGISTER_COALESCED_MMIO, &zone);
+    }
+#endif
+
+    return ret;
+}
+
+int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size)
+{
+    int ret = -ENOSYS;
+#ifdef KVM_CAP_COALESCED_MMIO
+    KVMState *s = kvm_state;
+
+    if (s->coalesced_mmio) {
+        struct kvm_coalesced_mmio_zone zone;
+
+        zone.addr = start;
+        zone.size = size;
+
+        ret = kvm_vm_ioctl(s, KVM_UNREGISTER_COALESCED_MMIO, &zone);
+    }
+#endif
+
+    return ret;
+}
+
 int kvm_init(int smp_cpus)
 {
     KVMState *s;
@@ -283,6 +325,13 @@
         goto err;
     }
 
+    s->coalesced_mmio = 0;
+#ifdef KVM_CAP_COALESCED_MMIO
+    ret = kvm_ioctl(s, KVM_CHECK_EXTENSION, KVM_CAP_COALESCED_MMIO);
+    if (ret > 0)
+        s->coalesced_mmio = ret;
+#endif
+
     ret = kvm_arch_init(s, smp_cpus);
     if (ret < 0)
         goto err;
@@ -342,6 +391,27 @@
     return 1;
 }
 
+static void kvm_run_coalesced_mmio(CPUState *env, struct kvm_run *run)
+{
+#ifdef KVM_CAP_COALESCED_MMIO
+    KVMState *s = kvm_state;
+    if (s->coalesced_mmio) {
+        struct kvm_coalesced_mmio_ring *ring;
+
+        ring = (void *)run + (s->coalesced_mmio * TARGET_PAGE_SIZE);
+        while (ring->first != ring->last) {
+            struct kvm_coalesced_mmio *ent;
+
+            ent = &ring->coalesced_mmio[ring->first];
+
+            cpu_physical_memory_write(ent->phys_addr, ent->data, ent->len);
+            /* FIXME smp_wmb() */
+            ring->first = (ring->first + 1) % KVM_COALESCED_MMIO_MAX;
+        }
+    }
+#endif
+}
+
 int kvm_cpu_exec(CPUState *env)
 {
     struct kvm_run *run = env->kvm_run;
@@ -372,6 +442,8 @@
             abort();
         }
 
+        kvm_run_coalesced_mmio(env, run);
+
         ret = 0; /* exit loop */
         switch (run->exit_reason) {
         case KVM_EXIT_IO:
Index: kvm.h
===================================================================
--- kvm.h       (revision 5874)
+++ kvm.h       (working copy)
@@ -45,6 +45,9 @@
 
 int kvm_has_sync_mmu(void);
 
+int kvm_coalesce_mmio_region(target_phys_addr_t start, ram_addr_t size);
+int kvm_uncoalesce_mmio_region(target_phys_addr_t start, ram_addr_t size);
+
 /* internal API */
 
 struct KVMState;
Index: exec.c
===================================================================
--- exec.c      (revision 5868)
+++ exec.c      (working copy)
@@ -2345,6 +2345,18 @@
     return p->phys_offset;
 }
 
+void qemu_register_coalesced_mmio(target_phys_addr_t addr, ram_addr_t size)
+{
+    if (kvm_enabled())
+        kvm_coalesce_mmio_region(addr, size);
+}
+
+void qemu_unregister_coalesced_mmio(target_phys_addr_t addr, ram_addr_t size)
+{
+    if (kvm_enabled())
+        kvm_uncoalesce_mmio_region(addr, size);
+}
+
 /* XXX: better than nothing */
 ram_addr_t qemu_ram_alloc(ram_addr_t size)
 {
Index: hw/cirrus_vga.c
===================================================================
--- hw/cirrus_vga.c     (revision 5868)
+++ hw/cirrus_vga.c     (working copy)
@@ -3218,6 +3218,7 @@
                                            cirrus_vga_mem_write, s);
     cpu_register_physical_memory(isa_mem_base + 0x000a0000, 0x20000,
                                  s->vga_io_memory);
+    qemu_register_coalesced_mmio(isa_mem_base + 0x000a0000, 0x20000);
 
     s->sr[0x06] = 0x0f;
     if (device_id == CIRRUS_ID_CLGD5446) {
Index: hw/pci.c
===================================================================
--- hw/pci.c    (revision 5868)
+++ hw/pci.c    (working copy)
@@ -279,6 +279,7 @@
                         cpu_register_physical_memory(pci_to_cpu_addr(r->addr),
                                                      r->size,
                                                      IO_MEM_UNASSIGNED);
+                        qemu_unregister_coalesced_mmio(r->addr, r->size);
                     }
                 }
                 r->addr = new_addr;
Index: hw/vga.c
===================================================================
--- hw/vga.c    (revision 5868)
+++ hw/vga.c    (working copy)
@@ -2256,6 +2256,7 @@
     vga_io_memory = cpu_register_io_memory(0, vga_mem_read, vga_mem_write, s);
     cpu_register_physical_memory(isa_mem_base + 0x000a0000, 0x20000,
                                  vga_io_memory);
+    qemu_register_coalesced_mmio(isa_mem_base + 0x000a0000, 0x20000);
 }
 
 /* Memory mapped interface */
@@ -2330,6 +2331,7 @@
     cpu_register_physical_memory(ctrl_base, 0x100000, s_ioport_ctrl);
     s->bank_offset = 0;
     cpu_register_physical_memory(vram_base + 0x000a0000, 0x20000, 
vga_io_memory);
+    qemu_register_coalesced_mmio(vram_base + 0x000a0000, 0x20000);
 }
 
 int isa_vga_init(DisplayState *ds, uint8_t *vga_ram_base,
Index: hw/e1000.c
===================================================================
--- hw/e1000.c  (revision 5868)
+++ hw/e1000.c  (working copy)
@@ -1001,10 +1001,22 @@
                 uint32_t addr, uint32_t size, int type)
 {
     E1000State *d = (E1000State *)pci_dev;
+    int i;
+    const uint32_t excluded_regs[] = {
+        E1000_MDIC, E1000_ICR, E1000_ICS, E1000_IMS,
+        E1000_IMC, E1000_TCTL, E1000_TDT, PNPMMIO_SIZE
+    };
 
+
     DBGOUT(MMIO, "e1000_mmio_map addr=0x%08x 0x%08x\n", addr, size);
 
     cpu_register_physical_memory(addr, PNPMMIO_SIZE, d->mmio_index);
+    qemu_register_coalesced_mmio(addr, excluded_regs[0]);
+
+    for (i = 0; excluded_regs[i] != PNPMMIO_SIZE; i++)
+        qemu_register_coalesced_mmio(addr + excluded_regs[i] + 4,
+                                     excluded_regs[i + 1] -
+                                     excluded_regs[i] - 4);
 }
 
 void
Index: cpu-all.h
===================================================================
--- cpu-all.h   (revision 5868)
+++ cpu-all.h   (working copy)
@@ -961,6 +961,15 @@
 void dump_exec_info(FILE *f,
                     int (*cpu_fprintf)(FILE *f, const char *fmt, ...));
 
+/* Coalesced MMIO regions are areas where write operations can be reordered.
+ * This usually implies that write operations are side-effect free.  This 
allows
+ * batching which can make a major impact on performance when using
+ * virtualization.
+ */
+void qemu_register_coalesced_mmio(target_phys_addr_t addr, ram_addr_t size);
+
+void qemu_unregister_coalesced_mmio(target_phys_addr_t addr, ram_addr_t size);
+
 /*******************************************/
 /* host CPU ticks (if available) */