[v2,3/3] kvm: Atomic memslot updates

Commit Message

Emanuele Giuseppe Esposito Nov. 10, 2022, 4:48 p.m. UTC

From: David Hildenbrand <[e-mail]>

If we update an existing memslot (e.g., resize, split), we temporarily
remove the memslot to re-add it immediately afterwards. These updates
are not atomic, especially not for KVM VCPU threads, such that we can
get spurious faults.

Let's inhibit most KVM ioctls while performing relevant updates, such
that we can perform the update just as if it would happen atomically
without additional kernel support.

We capture the add/del changes and apply them in the notifier commit
stage instead. There, we can check for overlaps and perform the ioctl
inhibiting only if really required (-> overlap).

To keep things simple we don't perform additional checks that wouldn't
actually result in an overlap -- such as !RAM memory regions in some
cases (see kvm_set_phys_mem()).

To minimize cache-line bouncing, use a separate indicator
(in_ioctl_lock) per CPU.  Also, make sure to hold the kvm_slots_lock
while performing both actions (removing+re-adding).

We have to wait until all IOCTLs were exited and block new ones from
getting executed.

This approach cannot result in a deadlock as long as the inhibitor does
not hold any locks that might hinder an IOCTL from getting finished and
exited - something fairly unusual. The inhibitor will always hold the BQL.

AFAIKs, one possible candidate would be userfaultfd. If a page cannot be
placed (e.g., during postcopy), because we're waiting for a lock, or if the
userfaultfd thread cannot process a fault, because it is waiting for a
lock, there could be a deadlock. However, the BQL is not applicable here,
because any other guest memory access while holding the BQL would already
result in a deadlock.

Nothing else in the kernel should block forever and wait for userspace
intervention.

Note: pause_all_vcpus()/resume_all_vcpus() or
start_exclusive()/end_exclusive() cannot be used, as they either drop
the BQL or require to be called without the BQL - something inhibitors
cannot handle. We need a low-level locking mechanism that is
deadlock-free even when not releasing the BQL.

Signed-off-by: David Hildenbrand <[e-mail]>
Signed-off-by: Emanuele Giuseppe Esposito <[e-mail]>
Tested-by: Emanuele Giuseppe Esposito <[e-mail]>
---
 accel/kvm/kvm-all.c      | 100 ++++++++++++++++++++++++++++++++++-----
 include/sysemu/kvm_int.h |   8 ++++
 2 files changed, 97 insertions(+), 11 deletions(-)

Comments

Paolo Bonzini Nov. 11, 2022, 11:01 a.m. UTC | #1

On 11/10/22 17:48, Emanuele Giuseppe Esposito wrote:
> 
> +    /* Remove all memslots before adding the new ones. */
> +    QSIMPLEQ_FOREACH_SAFE(u1, &kml->transaction_del, next, u2) {
> +        kvm_set_phys_mem(kml, u1->section, false);
> +        memory_region_unref(u1->section->mr);
> +
> +        QSIMPLEQ_REMOVE(&kml->transaction_del, u1, KVMMemoryUpdate, next);
> +        memory_region_section_free_copy(u1->section);
> +        g_free(u1);
> +    }
> +    QSIMPLEQ_FOREACH_SAFE(u1, &kml->transaction_add, next, u2) {
> +        memory_region_ref(u1->section->mr);
> +        kvm_set_phys_mem(kml, u1->section, true);
> +
> +        QSIMPLEQ_REMOVE(&kml->transaction_add, u1, KVMMemoryUpdate, next);
> +        memory_region_section_free_copy(u1->section);
> +        g_free(u1);
> +    }

I'm not a huge fan of new_copy/free_copy, and I don't think it's needed 
here.  The FlatView is certainly alive between begin and commit (see 
address_space_set_flatview()), and the MemoryRegion is kept alive by the 
FlatView.  In other words, unlike other uses of the functions, here the 
lifetime is bound by kvm_region_commit, and that means you can just copy 
by value.  You can just copy it into KVMMemoryUpdate, i.e.

     typedef struct KVMMemoryUpdate {
         QSIMPLEQ_ENTRY(KVMMemoryUpdate) next;
         MemoryRegionSection section;
     } KVMMemoryUpdate;

Also, you can write the loop as

     while (!QSIMPLEQ_EMPTY(&kvm->transaction_add) {
         u = QSIMPLEQ_FIRST(&kvm->transaction_add);
         QSIMPLEQ_REMOVE_HEAD(&kml->transaction_add, next);
         ...
         g_free(u);
     }

This clarifies the invariant that the QSIMPLEQs become empty at the end 
of the loop.  If it were QTAILQ it would be more a matter of personal 
taste, but QSIMPLEQ_REMOVE is technically not constant-time and that 
also tilts in favor of QSIMPLEQ_REMOVE_HEAD.

In fact I think we should remove QSIMPLEQ_REMOVE and QSLIST_REMOVE, 
changing all users that need it to QTAILQ and QLIST respectively... 
added to https://wiki.qemu.org/Contribute/BiteSizedTasks#API_conversion.

Paolo

Patch

diff --git a/accel/kvm/kvm-all.c b/accel/kvm/kvm-all.c
index dfc6fe76db..ec478dbe69 100644
--- a/accel/kvm/kvm-all.c
+++ b/accel/kvm/kvm-all.c
@@ -31,6 +31,7 @@ 
 #include "sysemu/kvm_int.h"
 #include "sysemu/runstate.h"
 #include "sysemu/cpus.h"
+#include "sysemu/accel-blocker.h"
 #include "qemu/bswap.h"
 #include "exec/memory.h"
 #include "exec/ram_addr.h"
@@ -46,6 +47,7 @@ 
 #include "sysemu/hw_accel.h"
 #include "kvm-cpus.h"
 #include "sysemu/dirtylimit.h"
+#include "qemu/range.h"
 
 #include "hw/boards.h"
 #include "monitor/stats.h"
@@ -1292,6 +1294,7 @@  void kvm_set_max_memslot_size(hwaddr max_slot_size)
     kvm_max_slot_size = max_slot_size;
 }
 
+/* Called with KVMMemoryListener.slots_lock held */
 static void kvm_set_phys_mem(KVMMemoryListener *kml,
                              MemoryRegionSection *section, bool add)
 {
@@ -1326,14 +1329,12 @@  static void kvm_set_phys_mem(KVMMemoryListener *kml,
     ram = memory_region_get_ram_ptr(mr) + mr_offset;
     ram_start_offset = memory_region_get_ram_addr(mr) + mr_offset;
 
-    kvm_slots_lock();
-
     if (!add) {
         do {
             slot_size = MIN(kvm_max_slot_size, size);
             mem = kvm_lookup_matching_slot(kml, start_addr, slot_size);
             if (!mem) {
-                goto out;
+                return;
             }
             if (mem->flags & KVM_MEM_LOG_DIRTY_PAGES) {
                 /*
@@ -1371,7 +1372,7 @@  static void kvm_set_phys_mem(KVMMemoryListener *kml,
             start_addr += slot_size;
             size -= slot_size;
         } while (size);
-        goto out;
+        return;
     }
 
     /* register the new slot */
@@ -1396,9 +1397,6 @@  static void kvm_set_phys_mem(KVMMemoryListener *kml,
         ram += slot_size;
         size -= slot_size;
     } while (size);
-
-out:
-    kvm_slots_unlock();
 }
 
 static void *kvm_dirty_ring_reaper_thread(void *data)
@@ -1455,18 +1453,94 @@  static void kvm_region_add(MemoryListener *listener,
                            MemoryRegionSection *section)
 {
     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
+    KVMMemoryUpdate *update;
+
+    update = g_new0(KVMMemoryUpdate, 1);
+    update->section = memory_region_section_new_copy(section);
 
-    memory_region_ref(section->mr);
-    kvm_set_phys_mem(kml, section, true);
+    QSIMPLEQ_INSERT_TAIL(&kml->transaction_add, update, next);
 }
 
 static void kvm_region_del(MemoryListener *listener,
                            MemoryRegionSection *section)
 {
     KVMMemoryListener *kml = container_of(listener, KVMMemoryListener, listener);
+    KVMMemoryUpdate *update;
 
-    kvm_set_phys_mem(kml, section, false);
-    memory_region_unref(section->mr);
+    update = g_new0(KVMMemoryUpdate, 1);
+    update->section = memory_region_section_new_copy(section);
+
+    QSIMPLEQ_INSERT_TAIL(&kml->transaction_del, update, next);
+}
+
+static void kvm_region_commit(MemoryListener *listener)
+{
+    KVMMemoryListener *kml = container_of(listener, KVMMemoryListener,
+                                          listener);
+    KVMMemoryUpdate *u1, *u2;
+    bool need_inhibit = false;
+
+    if (QSIMPLEQ_EMPTY(&kml->transaction_add) &&
+        QSIMPLEQ_EMPTY(&kml->transaction_del)) {
+        return;
+    }
+
+    /*
+     * We have to be careful when regions to add overlap with ranges to remove.
+     * We have to simulate atomic KVM memslot updates by making sure no ioctl()
+     * is currently active.
+     *
+     * The lists are order by addresses, so it's easy to find overlaps.
+     */
+    u1 = QSIMPLEQ_FIRST(&kml->transaction_del);
+    u2 = QSIMPLEQ_FIRST(&kml->transaction_add);
+    while (u1 && u2) {
+        Range r1, r2;
+
+        range_init_nofail(&r1, u1->section->offset_within_address_space,
+                          int128_get64(u1->section->size));
+        range_init_nofail(&r2, u2->section->offset_within_address_space,
+                          int128_get64(u2->section->size));
+
+        if (range_overlaps_range(&r1, &r2)) {
+            need_inhibit = true;
+            break;
+        }
+        if (range_lob(&r1) < range_lob(&r2)) {
+            u1 = QSIMPLEQ_NEXT(u1, next);
+        } else {
+            u2 = QSIMPLEQ_NEXT(u2, next);
+        }
+    }
+
+
+    kvm_slots_lock();
+    if (need_inhibit) {
+        accel_ioctl_inhibit_begin();
+    }
+
+    /* Remove all memslots before adding the new ones. */
+    QSIMPLEQ_FOREACH_SAFE(u1, &kml->transaction_del, next, u2) {
+        kvm_set_phys_mem(kml, u1->section, false);
+        memory_region_unref(u1->section->mr);
+
+        QSIMPLEQ_REMOVE(&kml->transaction_del, u1, KVMMemoryUpdate, next);
+        memory_region_section_free_copy(u1->section);
+        g_free(u1);
+    }
+    QSIMPLEQ_FOREACH_SAFE(u1, &kml->transaction_add, next, u2) {
+        memory_region_ref(u1->section->mr);
+        kvm_set_phys_mem(kml, u1->section, true);
+
+        QSIMPLEQ_REMOVE(&kml->transaction_add, u1, KVMMemoryUpdate, next);
+        memory_region_section_free_copy(u1->section);
+        g_free(u1);
+    }
+
+    if (need_inhibit) {
+        accel_ioctl_inhibit_end();
+    }
+    kvm_slots_unlock();
 }
 
 static void kvm_log_sync(MemoryListener *listener,
@@ -1610,8 +1684,12 @@  void kvm_memory_listener_register(KVMState *s, KVMMemoryListener *kml,
         kml->slots[i].slot = i;
     }
 
+    QSIMPLEQ_INIT(&kml->transaction_add);
+    QSIMPLEQ_INIT(&kml->transaction_del);
+
     kml->listener.region_add = kvm_region_add;
     kml->listener.region_del = kvm_region_del;
+    kml->listener.commit = kvm_region_commit;
     kml->listener.log_start = kvm_log_start;
     kml->listener.log_stop = kvm_log_stop;
     kml->listener.priority = 10;
diff --git a/include/sysemu/kvm_int.h b/include/sysemu/kvm_int.h
index 3b4adcdc10..5ea2d7924b 100644
--- a/include/sysemu/kvm_int.h
+++ b/include/sysemu/kvm_int.h
@@ -12,6 +12,7 @@ 
 #include "exec/memory.h"
 #include "qapi/qapi-types-common.h"
 #include "qemu/accel.h"
+#include "qemu/queue.h"
 #include "sysemu/kvm.h"
 
 typedef struct KVMSlot
@@ -31,10 +32,17 @@  typedef struct KVMSlot
     ram_addr_t ram_start_offset;
 } KVMSlot;
 
+typedef struct KVMMemoryUpdate {
+    QSIMPLEQ_ENTRY(KVMMemoryUpdate) next;
+    MemoryRegionSection *section;
+} KVMMemoryUpdate;
+
 typedef struct KVMMemoryListener {
     MemoryListener listener;
     KVMSlot *slots;
     int as_id;
+    QSIMPLEQ_HEAD(, KVMMemoryUpdate) transaction_add;
+    QSIMPLEQ_HEAD(, KVMMemoryUpdate) transaction_del;
 } KVMMemoryListener;
 
 #define KVM_MSI_HASHTAB_SIZE    256