[Qemu-devel] [PATCH 4/6] migration: calculate downtime on dst side

[Alexey Perevalov]

This patch provides downtime calculation per vCPU,
as a summary and as a overlapped value for all vCPUs.

This approach just keeps tree with page fault addr as a key,
and t1-t2 interval of pagefault time and page copy time, with
affected vCPU bit mask.
For more implementation details please see comment to
get_postcopy_total_downtime function.

Signed-off-by: Alexey Perevalov <address@hidden>
---
 include/migration/migration.h |  14 +++
 migration/migration.c         | 280 +++++++++++++++++++++++++++++++++++++++++-
 migration/postcopy-ram.c      |  24 +++-
 migration/qemu-file.c         |   1 -
 migration/trace-events        |   9 +-
 5 files changed, 323 insertions(+), 5 deletions(-)

diff --git a/include/migration/migration.h b/include/migration/migration.h
index 5720c88..5d2c628 100644
--- a/include/migration/migration.h
+++ b/include/migration/migration.h
@@ -123,10 +123,24 @@ struct MigrationIncomingState {
 
     /* See savevm.c */
     LoadStateEntry_Head loadvm_handlers;
+
+    /*
+     *  Tree for keeping postcopy downtime,
+     *  necessary to calculate correct downtime, during multiple
+     *  vm suspends, it keeps host page address as a key and
+     *  DowntimeDuration as a data
+     *  NULL means kernel couldn't provide process thread id,
+     *  and QEMU couldn't identify which vCPU raise page fault
+     */
+    GTree *postcopy_downtime;
 };
 
 MigrationIncomingState *migration_incoming_get_current(void);
 void migration_incoming_state_destroy(void);
+void mark_postcopy_downtime_begin(uint64_t addr, int cpu);
+void mark_postcopy_downtime_end(uint64_t addr);
+uint64_t get_postcopy_total_downtime(void);
+void destroy_downtime_duration(gpointer data);
 
 /*
  * An outstanding page request, on the source, having been received
diff --git a/migration/migration.c b/migration/migration.c
index 79f6425..5bac434 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -38,6 +38,8 @@
 #include "io/channel-tls.h"
 #include "migration/colo.h"
 
+#define DEBUG_VCPU_DOWNTIME 1
+
 #define MAX_THROTTLE  (32 << 20)      /* Migration transfer speed throttling */
 
 /* Amount of time to allocate to each "chunk" of bandwidth-throttled
@@ -77,6 +79,19 @@ static NotifierList migration_state_notifiers =
 
 static bool deferred_incoming;
 
+typedef struct {
+    int64_t begin;
+    int64_t end;
+    uint64_t *cpus; /* cpus bit mask array, QEMU bit functions support
+     bit operation on memory regions, but doesn't check out of range */
+} DowntimeDuration;
+
+typedef struct {
+    int64_t tp; /* point in time */
+    bool is_end;
+    uint64_t *cpus;
+} OverlapDowntime;
+
 /*
  * Current state of incoming postcopy; note this is not part of
  * MigrationIncomingState since it's state is used during cleanup
@@ -117,6 +132,13 @@ MigrationState *migrate_get_current(void)
     return &current_migration;
 }
 
+void destroy_downtime_duration(gpointer data)
+{
+    DowntimeDuration *dd = (DowntimeDuration *)data;
+    g_free(dd->cpus);
+    g_free(data);
+}
+
 MigrationIncomingState *migration_incoming_get_current(void)
 {
     static bool once;
@@ -138,10 +160,13 @@ void migration_incoming_state_destroy(void)
     struct MigrationIncomingState *mis = migration_incoming_get_current();
 
     qemu_event_destroy(&mis->main_thread_load_event);
+    if (mis->postcopy_downtime) {
+        g_tree_destroy(mis->postcopy_downtime);
+        mis->postcopy_downtime = NULL;
+    }
     loadvm_free_handlers(mis);
 }
 
-
 typedef struct {
     bool optional;
     uint32_t size;
@@ -1754,7 +1779,6 @@ static int postcopy_start(MigrationState *ms, bool 
*old_vm_running)
      */
     ms->postcopy_after_devices = true;
     notifier_list_notify(&migration_state_notifiers, ms);
-
     ms->downtime =  qemu_clock_get_ms(QEMU_CLOCK_REALTIME) - time_at_stop;
 
     qemu_mutex_unlock_iothread();
@@ -2117,3 +2141,255 @@ PostcopyState postcopy_state_set(PostcopyState 
new_state)
     return atomic_xchg(&incoming_postcopy_state, new_state);
 }
 
+#define SIZE_TO_KEEP_CPUBITS (1 + smp_cpus/sizeof(guint64))
+
+void mark_postcopy_downtime_begin(uint64_t addr, int cpu)
+{
+    MigrationIncomingState *mis = migration_incoming_get_current();
+    DowntimeDuration *dd;
+    if (!mis->postcopy_downtime) {
+        return;
+    }
+
+    dd = g_tree_lookup(mis->postcopy_downtime, (gpointer)addr); /* !!! cast */
+    if (!dd) {
+        dd = (DowntimeDuration *)g_new0(DowntimeDuration, 1);
+        dd->cpus = g_new0(guint64, SIZE_TO_KEEP_CPUBITS);
+        g_tree_insert(mis->postcopy_downtime, (gpointer)addr, (gpointer)dd);
+    }
+
+    if (cpu < 0) {
+        /* assume in this situation all vCPUs are sleeping */
+        int i;
+        for (i = 0; i < SIZE_TO_KEEP_CPUBITS; i++) {
+            dd->cpus[i] = ~(uint64_t)0u;
+        }
+    } else
+        set_bit(cpu, dd->cpus);
+
+    /*
+     *  overwrite previously set dd->begin, if that page already was
+     *     faulted on another cpu
+     */
+    dd->begin = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
+    trace_mark_postcopy_downtime_begin(addr, dd, dd->begin, cpu);
+}
+
+void mark_postcopy_downtime_end(uint64_t addr)
+{
+    MigrationIncomingState *mis = migration_incoming_get_current();
+    DowntimeDuration *dd;
+    if (!mis->postcopy_downtime) {
+        return;
+    }
+
+    dd = g_tree_lookup(mis->postcopy_downtime, (gpointer)addr);
+    if (!dd) {
+        /* error_report("Could not populate downtime duration completion time 
\n\
+                        There is no downtime duration for 0x%"PRIx64, addr); */
+        return;
+    }
+
+    dd->end = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
+    trace_mark_postcopy_downtime_end(addr, dd, dd->end);
+}
+
+struct downtime_overlay_cxt {
+    GPtrArray *downtime_points;
+    size_t number_of_points;
+};
+/*
+ * This function split each DowntimeDuration, which represents as start/end
+ * pointand makes a points of it, then fill array with points,
+ * to sort it in future.
+ */
+static gboolean split_duration_and_fill_points(gpointer key, gpointer value,
+                                        gpointer data)
+{
+    struct downtime_overlay_cxt *ctx = (struct downtime_overlay_cxt *)data;
+    DowntimeDuration *dd = (DowntimeDuration *)value;
+    GPtrArray *interval = ctx->downtime_points;
+    if (dd->begin) {
+        OverlapDowntime *od_begin = g_new0(OverlapDowntime, 1);
+        od_begin->cpus = g_memdup(dd->cpus, sizeof(uint64_t) * 
SIZE_TO_KEEP_CPUBITS);
+        od_begin->tp = dd->begin;
+        od_begin->is_end = false;
+        g_ptr_array_add(interval, od_begin);
+        ctx->number_of_points += 1;
+    }
+
+    if (dd->end) {
+        OverlapDowntime *od_end = g_new0(OverlapDowntime, 1);
+        od_end->cpus = g_memdup(dd->cpus, sizeof(uint64_t) * 
SIZE_TO_KEEP_CPUBITS);
+        od_end->tp = dd->end;
+        od_end->is_end = true;
+        g_ptr_array_add(interval, od_end);
+        ctx->number_of_points += 1;
+    }
+
+    if (dd->end && dd->begin)
+        trace_split_duration_and_fill_points(dd->end - dd->begin, 
(uint64_t)key);
+    return FALSE;
+}
+
+#ifdef DEBUG_VCPU_DOWNTIME
+static gboolean calculate_per_cpu(gpointer key, gpointer value,
+                                  gpointer data)
+{
+    int *downtime_cpu = (int *)data;
+    DowntimeDuration *dd = (DowntimeDuration *)value;
+    int cpu_iter;
+    for (cpu_iter = 0; cpu_iter < smp_cpus; cpu_iter++) {
+        if (test_bit(cpu_iter, dd->cpus) && dd->end && dd->begin)
+            downtime_cpu[cpu_iter] += dd->end - dd->begin;
+    }
+    return FALSE;
+}
+#endif /* DEBUG_VCPU_DOWNTIME */
+
+static gint compare_downtime(gconstpointer a, gconstpointer b)
+{
+    DowntimeDuration *dda = (DowntimeDuration *)a;
+    DowntimeDuration *ddb = (DowntimeDuration *)b;
+    return dda->begin - ddb->begin;
+}
+
+static void destroy_overlap_downtime(gpointer data)
+{
+    OverlapDowntime *od = (OverlapDowntime *)data;
+    g_free(od->cpus);
+    g_free(data);
+}
+
+static int check_overlap(uint64_t *b)
+{
+    unsigned long zero_bit = find_first_zero_bit(b, BITS_PER_LONG * 
SIZE_TO_KEEP_CPUBITS);
+    return zero_bit >= smp_cpus;
+}
+
+/*
+ * This function calculates downtime per cpu and trace it
+ *
+ *  Also it calculates total downtime as an interval's overlap,
+ *  for many vCPU.
+ *
+ *  The approach is following:
+ *  Initially intervals are represented in tree where key is
+ *  pagefault address, and values:
+ *   begin - page fault time
+ *   end   - page load time
+ *   cpus  - bit mask shows affected cpus
+ *
+ *  To calculate overlap on all cpus, intervals converted into
+ *  array of points in time (downtime_points), the size of
+ *  array is 2 * number of nodes in tree of intervals (2 array
+ *  elements per one in element of interval).
+ *  Each element is marked as end (E) or as start (S) of interval.
+ *  The overlap downtime will be calculated for SE, only in case
+ *  there is sequence S(0..N)E(M) for every vCPU.
+ *
+ * As example we have 3 CPU
+ *
+ *      S1        E1           S1               E1
+ * -----***********------------xxx***************------------------------> CPU1
+ *
+ *             S2                E2
+ * ------------****************xxx---------------------------------------> CPU2
+ *
+ *                         S3            E3
+ * ------------------------****xxx********-------------------------------> CPU3
+ *
+ * We have sequence S1,S2,E1,S3,S1,E2,E3,E1
+ * S2,E1 - doesn't match condition due to sequence S1,S2,E1 doesn't include 
CPU3
+ * S3,S1,E2 - sequenece includes all CPUs, in this case overlap will be S1,E2
+ * Legend of picture is following: * - means downtime per vCPU
+ *                                 x - means overlapped downtime
+ */
+uint64_t get_postcopy_total_downtime(void)
+{
+    MigrationIncomingState *mis = migration_incoming_get_current();
+    uint64_t total_downtime = 0; /* for total overlapped downtime */
+    const int intervals = g_tree_nnodes(mis->postcopy_downtime);
+    int point_iter, start_point_iter, i;
+    struct downtime_overlay_cxt dp_ctx = { 0 };
+    /*
+     * array will contain 2 * interval points or less, if
+     * it was not page fault finalization for page,
+     * real count will be in ctx.number_of_points
+     */
+    dp_ctx.downtime_points = g_ptr_array_new_full(2 * intervals,
+                                                     destroy_overlap_downtime);
+    if (!mis->postcopy_downtime) {
+        goto out;
+    }
+
+#ifdef DEBUG_VCPU_DOWNTIME
+    {
+        gint *downtime_cpu = g_new0(int, smp_cpus);
+        g_tree_foreach(mis->postcopy_downtime, calculate_per_cpu, 
downtime_cpu);
+        for (point_iter = 0; point_iter < smp_cpus; point_iter++)
+        {
+            trace_downtime_per_cpu(point_iter, downtime_cpu[point_iter]);
+        }
+        g_free(downtime_cpu);
+    }
+#endif /* DEBUG_VCPU_DOWNTIME */
+
+    /* make downtime points S/E from interval */
+    g_tree_foreach(mis->postcopy_downtime, split_duration_and_fill_points,
+                   &dp_ctx);
+    g_ptr_array_sort(dp_ctx.downtime_points, compare_downtime);
+
+    for (point_iter = 1; point_iter < dp_ctx.number_of_points;
+         point_iter++) {
+        OverlapDowntime *od = g_ptr_array_index(dp_ctx.downtime_points,
+                point_iter);
+        uint64_t *cur_cpus;
+        int smp_cpus_i = smp_cpus;
+        OverlapDowntime *prev_od = g_ptr_array_index(dp_ctx.downtime_points,
+                                                     point_iter - 1);
+        if (!od || !prev_od)
+            continue;
+        /* we need sequence SE */
+        if (!od->is_end || prev_od->is_end)
+            continue;
+
+        cur_cpus = g_memdup(od->cpus, sizeof(uint64_t) * SIZE_TO_KEEP_CPUBITS);
+        for (start_point_iter = point_iter - 1;
+             start_point_iter >= 0 && smp_cpus_i;
+             start_point_iter--, smp_cpus_i--) {
+            OverlapDowntime *t_od = g_ptr_array_index(dp_ctx.downtime_points,
+                                                      start_point_iter);
+            if (!t_od)
+                break;
+            /* should be S */
+            if (t_od->is_end)
+                break;
+
+            /* points were sorted, it's possible when
+             * end is not occured, but this points were ommited
+             * in split_duration_and_fill_points */
+            if (od->tp <= prev_od->tp) {
+                break;
+            }
+
+            for (i = 0; i < SIZE_TO_KEEP_CPUBITS; i++) {
+                cur_cpus[i] |= t_od->cpus[i];
+            }
+
+            /* check_overlap - just count number of bits in cur_cpus,
+             * and compare it with smp_cpus */
+            if (check_overlap(cur_cpus)) {
+                total_downtime += od->tp - prev_od->tp;
+                /* situation when one S point represents all vCPU is possible 
*/
+                break;
+            }
+        }
+        g_free(cur_cpus);
+    }
+    trace_get_postcopy_total_downtime(g_tree_nnodes(mis->postcopy_downtime),
+        total_downtime);
+out:
+    g_ptr_array_free(dp_ctx.downtime_points, TRUE);
+    return total_downtime;
+}
diff --git a/migration/postcopy-ram.c b/migration/postcopy-ram.c
index 70f0480..ea89f4e 100644
--- a/migration/postcopy-ram.c
+++ b/migration/postcopy-ram.c
@@ -23,8 +23,10 @@
 #include "migration/postcopy-ram.h"
 #include "sysemu/sysemu.h"
 #include "sysemu/balloon.h"
+#include <sys/param.h>
 #include "qemu/error-report.h"
 #include "trace.h"
+#include "glib/glib-helper.h"
 
 /* Arbitrary limit on size of each discard command,
  * keeps them around ~200 bytes
@@ -81,6 +83,11 @@ static bool ufd_version_check(int ufd, 
MigrationIncomingState *mis)
         return false;
     }
 
+    if (mis && UFFD_FEATURE_THREAD_ID & api_struct.features) {
+        mis->postcopy_downtime = g_tree_new_full(g_int_cmp64,
+                                         NULL, NULL, 
destroy_downtime_duration);
+    }
+
     if (getpagesize() != ram_pagesize_summary()) {
         bool have_hp = false;
         /* We've got a huge page */
@@ -404,6 +411,18 @@ static int ram_block_enable_notify(const char *block_name, 
void *host_addr,
     return 0;
 }
 
+static int get_mem_fault_cpu_index(uint32_t pid)
+{
+    CPUState *cpu_iter;
+
+    CPU_FOREACH(cpu_iter) {
+        if (cpu_iter->thread_id == pid)
+           return cpu_iter->cpu_index;
+    }
+    trace_get_mem_fault_cpu_index(pid);
+    return -1;
+}
+
 /*
  * Handle faults detected by the USERFAULT markings
  */
@@ -481,8 +500,10 @@ static void *postcopy_ram_fault_thread(void *opaque)
         rb_offset &= ~(qemu_ram_pagesize(rb) - 1);
         trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
                                                 qemu_ram_get_idstr(rb),
-                                                rb_offset);
+                                                rb_offset, 
msg.arg.pagefault.feat.ptid);
 
+        mark_postcopy_downtime_begin(msg.arg.pagefault.address,
+                            
get_mem_fault_cpu_index(msg.arg.pagefault.feat.ptid));
         /*
          * Send the request to the source - we want to request one
          * of our host page sizes (which is >= TPS)
@@ -577,6 +598,7 @@ int postcopy_place_page(MigrationIncomingState *mis, void 
*host, void *from,
 
         return -e;
     }
+    mark_postcopy_downtime_end((uint64_t)host);
 
     trace_postcopy_place_page(host);
     return 0;
diff --git a/migration/qemu-file.c b/migration/qemu-file.c
index 195fa94..c9f3e47 100644
--- a/migration/qemu-file.c
+++ b/migration/qemu-file.c
@@ -547,7 +547,6 @@ size_t qemu_get_buffer_in_place(QEMUFile *f, uint8_t **buf, 
size_t size)
 int qemu_peek_byte(QEMUFile *f, int offset)
 {
     int index = f->buf_index + offset;
-
     assert(!qemu_file_is_writable(f));
     assert(offset < IO_BUF_SIZE);
 
diff --git a/migration/trace-events b/migration/trace-events
index 7372ce2..ab2e1e4 100644
--- a/migration/trace-events
+++ b/migration/trace-events
@@ -110,6 +110,12 @@ process_incoming_migration_co_end(int ret, int ps) "ret=%d 
postcopy-state=%d"
 process_incoming_migration_co_postcopy_end_main(void) ""
 migration_set_incoming_channel(void *ioc, const char *ioctype) "ioc=%p 
ioctype=%s"
 migration_set_outgoing_channel(void *ioc, const char *ioctype, const char 
*hostname)  "ioc=%p ioctype=%s hostname=%s"
+mark_postcopy_downtime_begin(uint64_t addr, void *dd, int64_t time, int cpu) 
"addr 0x%" PRIx64 " dd %p time %" PRId64 " cpu %d"
+mark_postcopy_downtime_end(uint64_t addr, void *dd, int64_t time) "addr 0x%" 
PRIx64 " dd %p time %" PRId64
+get_postcopy_total_downtime(int num, uint64_t total) "faults %d, total 
downtime %" PRIu64
+split_duration_and_fill_points(int64_t downtime, uint64_t addr) "downtime %" 
PRId64 " addr 0x%" PRIx64
+downtime_per_cpu(int cpu_index, int downtime) "downtime cpu[%d]=%d"
+source_return_path_thread_downtime(uint64_t downtime) "downtime %" PRIu64
 
 # migration/rdma.c
 qemu_rdma_accept_incoming_migration(void) ""
@@ -186,7 +192,7 @@ postcopy_ram_enable_notify(void) ""
 postcopy_ram_fault_thread_entry(void) ""
 postcopy_ram_fault_thread_exit(void) ""
 postcopy_ram_fault_thread_quit(void) ""
-postcopy_ram_fault_thread_request(uint64_t hostaddr, const char *ramblock, 
size_t offset) "Request for HVA=%" PRIx64 " rb=%s offset=%zx"
+postcopy_ram_fault_thread_request(uint64_t hostaddr, const char *ramblock, 
size_t offset, int pid) "Request for HVA=%" PRIx64 " rb=%s offset=%zx %d"
 postcopy_ram_incoming_cleanup_closeuf(void) ""
 postcopy_ram_incoming_cleanup_entry(void) ""
 postcopy_ram_incoming_cleanup_exit(void) ""
@@ -195,6 +201,7 @@ save_xbzrle_page_skipping(void) ""
 save_xbzrle_page_overflow(void) ""
 ram_save_iterate_big_wait(uint64_t milliconds, int iterations) "big wait: %" 
PRIu64 " milliseconds, %d iterations"
 ram_load_complete(int ret, uint64_t seq_iter) "exit_code %d seq iteration %" 
PRIu64
+get_mem_fault_cpu_index(uint32_t pid) "pid %u is not vCPU"
 
 # migration/exec.c
 migration_exec_outgoing(const char *cmd) "cmd=%s"
-- 
1.8.3.1