Skip to contentSkip to conversation

Fixing record/replay and adding reverse debugging

PATCH Superseded Pavel Dovgalyuk 5 participants 34 messages Oct 10, 2018 – Nov 30, 2018 Block layer

v7 of 13 Superseded Latest is v13 →

lore
Share & tools
lore
#0 lore raw

v7 cover Fixing record/replay and adding reverse debugging

GDB remote protocol supports reverse debugging of the targets. It includes 'reverse step' and 'reverse continue' operations. The first one finds the previous step of the execution, and the second one is intended to stop at the last breakpoint that would happen when the program is executed normally.

Reverse debugging is possible in the replay mode, when at least one snapshot was created at the record or replay phase. QEMU can use these snapshots for travelling back in time with GDB.

Running the execution in replay mode allows using GDB reverse debugging commands:
- reverse-stepi (or rsi): Steps one instruction to the past. QEMU loads on of the prior snapshots and proceeds to the desired instruction forward. When that step is reaches, execution stops.
- reverse-continue (or rc): Runs execution "backwards". QEMU tries to find breakpoint or watchpoint by loaded prior snapshot and replaying the execution. Then QEMU loads snapshots again and replays to the latest breakpoint. When there are no breakpoints in the examined section of the execution, QEMU finds one more snapshot and tries again. After the first snapshot is processed, execution stops at this snapshot.

The set of patches include the following modifications:
- gdbstub update for reverse debugging support
- functions that automatically perform reverse step and reverse continue operations
- hmp/qmp commands for manipulating the replay process
- improvement of the snapshotting for saving the execution step in the snapshot parameters
- other record/replay fixes

The patches are available in the repository:
https://github.com/ispras/qemu/tree/rr-181010

v7 changes:
- rebased to the new master with upstreamed patches from the series
- several improvements in hmp/qmp commands handling (suggested by Markus Armbruster)
- fixed record/replay with '-rtc base' option enabled
- added document with virtual hardware requirements

v6 changes:
- rebased to the new version of master
- fixed build of linux-user configurations
- added new clock for slirp and vnc timers

v5 changes:
- multiple fixes of record/replay bugs appeared after QEMU core update
- changed reverse debugging to 'since 3.1'

v4 changes:
- changed 'since 2.13' to 'since 3.0' in json (as suggested by Eric Blake)

v3 changes:
- Fixed PS/2 bug with save/load vm, which caused failures of the replay.
- Rebased to the new code base.
- Minor fixes.

v2 changes:
- documented reverse debugging
- fixed start vmstate loading in record mode
- documented qcow2 changes (as suggested by Eric Blake)
- made icount SnapshotInfo field optional (as suggested by Eric Blake)
- renamed qmp commands (as suggested by Eric Blake)
- minor changes

---

Pavel Dovgalyuk (19):
block: implement bdrv_snapshot_goto for blkreplay
replay: disable default snapshot for record/replay
replay: update docs for record/replay with block devices
replay: don't drain/flush bdrv queue while RR is working
replay: finish record/replay before closing the disks
qcow2: introduce icount field for snapshots
migration: introduce icount field for snapshots
replay: provide and accessor for rr filename
replay: introduce info hmp/qmp command
replay: introduce breakpoint at the specified step
replay: implement replay-seek command to proceed to the desired step
replay: refine replay-time module
replay: flush rr queue before loading the vmstate
gdbstub: add reverse step support in replay mode
gdbstub: add reverse continue support in replay mode
replay: describe reverse debugging in docs/replay.txt
replay: add BH oneshot event for block layer
replay: init rtc after enabling the replay
replay: document development rules

 accel/tcg/translator.c    |    1 
 block/blkreplay.c         |    8 +
 block/block-backend.c     |    3 
 block/io.c                |   22 +++
 block/qapi.c              |   17 ++
 block/qcow2-snapshot.c    |    9 +
 block/qcow2.h             |    2 
 blockdev.c                |   10 +
 cpus.c                    |   19 ++-
 docs/devel/replay.txt     |   45 ++++++
 docs/interop/qcow2.txt    |    4 +
 docs/replay.txt           |   45 ++++++
 exec.c                    |    6 +
 gdbstub.c                 |   50 +++++++
 hmp-commands-info.hx      |   14 ++
 hmp-commands.hx           |   44 ++++++
 hmp.h                     |    4 +
 include/block/snapshot.h  |    1 
 include/sysemu/replay.h   |   20 +++
 migration/savevm.c        |   15 +-
 qapi/block-core.json      |    7 +
 qapi/block.json           |    3 
 qapi/misc.json            |   77 +++++++++++
 replay/Makefile.objs      |    3 
 replay/replay-debugging.c |  319 +++++++++++++++++++++++++++++++++++++++++++++
 replay/replay-events.c    |   16 ++
 replay/replay-internal.h  |    7 +
 replay/replay-time.c      |   32 ++---
 replay/replay.c           |   26 ++++
 stubs/Makefile.objs       |    1 
 stubs/replay-user.c       |    9 +
 stubs/replay.c            |   10 +
 vl.c                      |   22 ++-
 33 files changed, 818 insertions(+), 53 deletions(-)
 create mode 100644 docs/devel/replay.txt
 create mode 100644 replay/replay-debugging.c
 create mode 100644 stubs/replay-user.c
Signature · 1 line
Pavel Dovgalyuk
#1 lore raw

PATCH v7 1/19 block: implement bdrv_snapshot_goto for blkreplay

From: Pavel Dovgalyuk <[email protected]>

This patch enables making snapshots with blkreplay used in block devices.
This function is required to make bdrv_snapshot_goto without calling .bdrv_open which is not implemented.

---
 block/blkreplay.c |    8 ++++++++
 1 file changed, 8 insertions(+)
block/blkreplay.c +8 -0
index b5d9efd..142dfe3 100644
@@ -126,6 +126,12 @@ static int coroutine_fn blkreplay_co_flush(BlockDriverState *bs)
     return ret;
 }
 
+static int blkreplay_snapshot_goto(BlockDriverState *bs,
+                                   const char *snapshot_id)
+{
+    return bdrv_snapshot_goto(bs->file->bs, snapshot_id, NULL);
+}
+
 static BlockDriver bdrv_blkreplay = {
     .format_name            = "blkreplay",
     .instance_size          = 0,
@@ -140,6 +146,8 @@ static BlockDriver bdrv_blkreplay = {
     .bdrv_co_pwrite_zeroes  = blkreplay_co_pwrite_zeroes,
     .bdrv_co_pdiscard       = blkreplay_co_pdiscard,
     .bdrv_co_flush          = blkreplay_co_flush,
+
+    .bdrv_snapshot_goto     = blkreplay_snapshot_goto,
 };
 
 static void bdrv_blkreplay_init(void)
#2 lore raw

PATCH v7 2/19 replay: disable default snapshot for record/replay

In reply to Pavel Dovgalyuk

From: Pavel Dovgalyuk <[email protected]>

This patch disables setting '-snapshot' option on by default in record/replay mode. This is needed for creating vmstates in record and replay modes.

---
 vl.c |   10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
vl.c +8 -2
index 4e25c78..46238d2 100644
@@ -3124,7 +3124,13 @@ int main(int argc, char **argv, char **envp)
                 drive_add(IF_PFLASH, -1, optarg, PFLASH_OPTS);
                 break;
             case QEMU_OPTION_snapshot:
-                snapshot = 1;
+                {
+                    Error *blocker = NULL;
+                    snapshot = 1;
+                    error_setg(&blocker, QERR_REPLAY_NOT_SUPPORTED,
+                               "-snapshot");
+                    replay_add_blocker(blocker);
+                }
                 break;
             case QEMU_OPTION_numa:
                 opts = qemu_opts_parse_noisily(qemu_find_opts("numa"),
@@ -4416,7 +4422,7 @@ int main(int argc, char **argv, char **envp)
         qapi_free_BlockdevOptions(bdo->bdo);
         g_free(bdo);
     }
-    if (snapshot || replay_mode != REPLAY_MODE_NONE) {
+    if (snapshot) {
         qemu_opts_foreach(qemu_find_opts("drive"), drive_enable_snapshot,
                           NULL, NULL);
     }
#3 lore raw

PATCH v7 3/19 replay: update docs for record/replay with block devices

In reply to Pavel Dovgalyuk

This patch updates the description of the command lines for using record/replay with attached block devices.

---
 docs/replay.txt |   12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)
docs/replay.txt +9 -3
index 3497585..2c2c5f6 100644
@@ -27,7 +27,7 @@ Usage of the record/replay:
  * First, record the execution with the following command line:
     qemu-system-i386 \
      -icount shift=7,rr=record,rrfile=replay.bin \
-     -drive file=disk.qcow2,if=none,id=img-direct \
+     -drive file=disk.qcow2,if=none,snapshot,id=img-direct \
      -drive driver=blkreplay,if=none,image=img-direct,id=img-blkreplay \
      -device ide-hd,drive=img-blkreplay \
      -netdev user,id=net1 -device rtl8139,netdev=net1 \
@@ -35,7 +35,7 @@ Usage of the record/replay:
  * After recording, you can replay it by using another command line:
     qemu-system-i386 \
      -icount shift=7,rr=replay,rrfile=replay.bin \
-     -drive file=disk.qcow2,if=none,id=img-direct \
+     -drive file=disk.qcow2,if=none,snapshot,id=img-direct \
      -drive driver=blkreplay,if=none,image=img-direct,id=img-blkreplay \
      -device ide-hd,drive=img-blkreplay \
      -netdev user,id=net1 -device rtl8139,netdev=net1 \
@@ -223,7 +223,7 @@ Block devices record/replay module intercepts calls of
 bdrv coroutine functions at the top of block drivers stack.
 To record and replay block operations the drive must be configured
 as following:
- -drive file=disk.qcow2,if=none,id=img-direct
+ -drive file=disk.qcow2,if=none,snapshot,id=img-direct
  -drive driver=blkreplay,if=none,image=img-direct,id=img-blkreplay
  -device ide-hd,drive=img-blkreplay
 
@@ -252,6 +252,12 @@ This snapshot is created at start of recording and restored at start
 of replaying. It also can be loaded while replaying to roll back
 the execution.
 
+'snapshot' flag of the disk image must be removed to save the snapshots
+in the overlay (or original image) instead of using the temporary overlay.
+ -drive file=disk.ovl,if=none,id=img-direct
+ -drive driver=blkreplay,if=none,image=img-direct,id=img-blkreplay
+ -device ide-hd,drive=img-blkreplay
+
 Use QEMU monitor to create additional snapshots. 'savevm <name>' command
 created the snapshot and 'loadvm <name>' restores it. To prevent corruption
 of the original disk image, use overlay files linked to the original images.
#4 lore raw

PATCH v7 4/19 replay: don't drain/flush bdrv queue while RR is working

In reply to Pavel Dovgalyuk

In record/replay mode bdrv queue is controlled by replay mechanism. It does not allow saving or loading the snapshots
when bdrv queue is not empty. Stopping the VM is not blocked by nonempty queue, but flushing the queue is still impossible there, because it may cause deadlocks in replay mode.
This patch disables bdrv_drain_all and bdrv_flush_all in record/replay mode.

---
 block/io.c |   22 ++++++++++++++++++++++
 cpus.c     |    2 --
 2 files changed, 22 insertions(+), 2 deletions(-)
block/io.c +22 -0
index bd9d688..96fb245 100644
@@ -32,6 +32,7 @@
 #include "qemu/cutils.h"
 #include "qapi/error.h"
 #include "qemu/error-report.h"
+#include "sysemu/replay.h"
 
 #define NOT_DONE 0x7fffffff /* used while emulated sync operation in progress */
 
@@ -538,6 +539,13 @@ void bdrv_drain_all_begin(void)
         return;
     }
 
+    /* bdrv queue is managed by record/replay,
+       waiting for finishing the I/O requests may
+       be infinite */
+    if (replay_events_enabled()) {
+        return;
+    }
+
     /* AIO_WAIT_WHILE() with a NULL context can only be called from the main
      * loop AioContext, so make sure we're in the main context. */
     assert(qemu_get_current_aio_context() == qemu_get_aio_context());
@@ -566,6 +574,13 @@ void bdrv_drain_all_end(void)
 {
     BlockDriverState *bs = NULL;
 
+    /* bdrv queue is managed by record/replay,
+       waiting for finishing the I/O requests may
+       be endless */
+    if (replay_events_enabled()) {
+        return;
+    }
+
     while ((bs = bdrv_next_all_states(bs))) {
         AioContext *aio_context = bdrv_get_aio_context(bs);
 
@@ -1997,6 +2012,13 @@ int bdrv_flush_all(void)
     BlockDriverState *bs = NULL;
     int result = 0;
 
+    /* bdrv queue is managed by record/replay,
+       creating new flush request for stopping
+       the VM may break the determinism */
+    if (replay_events_enabled()) {
+        return result;
+    }
+
     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
         AioContext *aio_context = bdrv_get_aio_context(bs);
         int ret;
cpus.c +0 -2
index 1c741bc..d60ddf1 100644
@@ -1078,7 +1078,6 @@ static int do_vm_stop(RunState state, bool send_stop)
     }
 
     bdrv_drain_all();
-    replay_disable_events();
     ret = bdrv_flush_all();
 
     return ret;
@@ -2135,7 +2134,6 @@ int vm_prepare_start(void)
     /* We are sending this now, but the CPUs will be resumed shortly later */
     qapi_event_send_resume();
 
-    replay_enable_events();
     cpu_enable_ticks();
     runstate_set(RUN_STATE_RUNNING);
     vm_state_notify(1, RUN_STATE_RUNNING);
#9 lore raw

PATCH v7 5/19 replay: finish record/replay before closing the disks

In reply to Pavel Dovgalyuk

After recent updates block devices cannot be closed on qemu exit. This happens due to the block request polling when replay is not finished. Therefore now we stop execution recording before closing the block devices.

---
 replay/replay.c |    2 ++
 vl.c            |    1 +
 2 files changed, 3 insertions(+)
replay/replay.c +2 -0
index 379b51a..ceee20a 100644
@@ -378,6 +378,8 @@ void replay_finish(void)
     g_free(replay_snapshot);
     replay_snapshot = NULL;
 
+    replay_mode = REPLAY_MODE_NONE;
+
     replay_finish_events();
 }
 
vl.c +1 -0
index 46238d2..a3e2e47 100644
@@ -4627,6 +4627,7 @@ int main(int argc, char **argv, char **envp)
 
     /* No more vcpu or device emulation activity beyond this point */
     vm_shutdown();
+    replay_finish();
 
     job_cancel_sync_all();
     bdrv_close_all();
#10 lore raw

PATCH v7 6/19 qcow2: introduce icount field for snapshots

In reply to Pavel Dovgalyuk

This patch introduces the icount field for saving within the snapshot. It is required for navigation between the snapshots in record/replay mode.

--

v2:
- documented format changes in docs/interop/qcow2.txt (suggested by Eric Blake)

---
 block/qcow2-snapshot.c |    7 +++++++
 block/qcow2.h          |    2 ++
 docs/interop/qcow2.txt |    4 ++++
 3 files changed, 13 insertions(+)
block/qcow2-snapshot.c +7 -0
index bb6a5b7..d682946 100644
@@ -103,6 +103,12 @@ int qcow2_read_snapshots(BlockDriverState *bs)
             sn->disk_size = bs->total_sectors * BDRV_SECTOR_SIZE;
         }
 
+        if (extra_data_size >= 24) {
+            sn->icount = be64_to_cpu(extra.icount);
+        } else {
+            sn->icount = -1ULL;
+        }
+
         /* Read snapshot ID */
         sn->id_str = g_malloc(id_str_size + 1);
         ret = bdrv_pread(bs->file, offset, sn->id_str, id_str_size);
@@ -209,6 +215,7 @@ static int qcow2_write_snapshots(BlockDriverState *bs)
         memset(&extra, 0, sizeof(extra));
         extra.vm_state_size_large = cpu_to_be64(sn->vm_state_size);
         extra.disk_size = cpu_to_be64(sn->disk_size);
+        extra.icount = cpu_to_be64(sn->icount);
 
         id_str_size = strlen(sn->id_str);
         name_size = strlen(sn->name);
block/qcow2.h +2 -0
index ba43031..656787d 100644
@@ -153,6 +153,7 @@ typedef struct QEMU_PACKED QCowSnapshotHeader {
 typedef struct QEMU_PACKED QCowSnapshotExtraData {
     uint64_t vm_state_size_large;
     uint64_t disk_size;
+    uint64_t icount;
 } QCowSnapshotExtraData;
 
 
@@ -166,6 +167,7 @@ typedef struct QCowSnapshot {
     uint32_t date_sec;
     uint32_t date_nsec;
     uint64_t vm_clock_nsec;
+    uint64_t icount;
 } QCowSnapshot;
 
 struct Qcow2Cache;
docs/interop/qcow2.txt +4 -0
index 845d40a..aa1101d 100644
@@ -506,6 +506,10 @@ Snapshot table entry:
 
                     Byte 48 - 55:   Virtual disk size of the snapshot in bytes
 
+                    Byte 56 - 63:   icount value which corresponds to
+                                    the record/replay step when the snapshot
+                                    was taken
+
                     Version 3 images must include extra data at least up to
                     byte 55.
 
#12 lore raw

PATCH v7 7/19 migration: introduce icount field for snapshots

In reply to Pavel Dovgalyuk

Saving icount as a parameters of the snapshot allows navigation between them in the execution replay scenario.
This information can be used for finding a specific snapshot for rewinding the recorded execution to the specific moment of the time. E.g., 'reverse step' action needs to load the nearest snapshot which is prior to the current moment of time .

--

v2:
- made icount in SnapshotInfo optional (suggested by Eric Blake)
v7:
- added more comments for icount member (suggested by Markus Armbruster)

---
 block/qapi.c             |   17 +++++++++++++----
 block/qcow2-snapshot.c   |    2 ++
 blockdev.c               |   10 ++++++++++
 include/block/snapshot.h |    1 +
 migration/savevm.c       |    5 +++++
 qapi/block-core.json     |    7 ++++++-
 qapi/block.json          |    3 ++-
 7 files changed, 39 insertions(+), 6 deletions(-)
block/qapi.c +13 -4
index c66f949..9f2ccfa 100644
@@ -210,6 +210,7 @@ int bdrv_query_snapshot_info_list(BlockDriverState *bs,
         info->date_nsec     = sn_tab[i].date_nsec;
         info->vm_clock_sec  = sn_tab[i].vm_clock_nsec / 1000000000;
         info->vm_clock_nsec = sn_tab[i].vm_clock_nsec % 1000000000;
+        info->icount        = sn_tab[i].icount;
 
         info_list = g_new0(SnapshotInfoList, 1);
         info_list->value = info;
@@ -663,14 +664,15 @@ void bdrv_snapshot_dump(fprintf_function func_fprintf, void *f,
                         QEMUSnapshotInfo *sn)
 {
     char buf1[128], date_buf[128], clock_buf[128];
+    char icount_buf[128] = {0};
     struct tm tm;
     time_t ti;
     int64_t secs;
 
     if (!sn) {
         func_fprintf(f,
-                     "%-10s%-20s%7s%20s%15s",
-                     "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK");
+                     "%-10s%-18s%7s%20s%13s%11s",
+                     "ID", "TAG", "VM SIZE", "DATE", "VM CLOCK", "ICOUNT");
     } else {
         ti = sn->date_sec;
         localtime_r(&ti, &tm);
@@ -683,13 +685,18 @@ void bdrv_snapshot_dump(fprintf_function func_fprintf, void *f,
                  (int)((secs / 60) % 60),
                  (int)(secs % 60),
                  (int)((sn->vm_clock_nsec / 1000000) % 1000));
+        if (sn->icount != -1ULL) {
+            snprintf(icount_buf, sizeof(icount_buf),
+                "%"PRId64, sn->icount);
+        }
         func_fprintf(f,
-                     "%-10s%-20s%7s%20s%15s",
+                     "%-10s%-18s%7s%20s%13s%11s",
                      sn->id_str, sn->name,
                      get_human_readable_size(buf1, sizeof(buf1),
                                              sn->vm_state_size),
                      date_buf,
-                     clock_buf);
+                     clock_buf,
+                     icount_buf);
     }
 }
 
@@ -857,6 +864,8 @@ void bdrv_image_info_dump(fprintf_function func_fprintf, void *f,
                 .date_nsec = elem->value->date_nsec,
                 .vm_clock_nsec = elem->value->vm_clock_sec * 1000000000ULL +
                                  elem->value->vm_clock_nsec,
+                .icount = elem->value->has_icount ?
+                          elem->value->icount : -1ULL,
             };
 
             pstrcpy(sn.id_str, sizeof(sn.id_str), elem->value->id);
block/qcow2-snapshot.c +2 -0
index d682946..96b57f4 100644
@@ -379,6 +379,7 @@ int qcow2_snapshot_create(BlockDriverState *bs, QEMUSnapshotInfo *sn_info)
     sn->date_sec = sn_info->date_sec;
     sn->date_nsec = sn_info->date_nsec;
     sn->vm_clock_nsec = sn_info->vm_clock_nsec;
+    sn->icount = sn_info->icount;
 
     /* Allocate the L1 table of the snapshot and copy the current one there. */
     l1_table_offset = qcow2_alloc_clusters(bs, s->l1_size * sizeof(uint64_t));
@@ -698,6 +699,7 @@ int qcow2_snapshot_list(BlockDriverState *bs, QEMUSnapshotInfo **psn_tab)
         sn_info->date_sec = sn->date_sec;
         sn_info->date_nsec = sn->date_nsec;
         sn_info->vm_clock_nsec = sn->vm_clock_nsec;
+        sn_info->icount = sn->icount;
     }
     *psn_tab = sn_tab;
     return s->nb_snapshots;
blockdev.c +10 -0
index a8755bd..1d66380 100644
@@ -57,6 +57,7 @@
 #include "block/trace.h"
 #include "sysemu/arch_init.h"
 #include "sysemu/qtest.h"
+#include "sysemu/replay.h"
 #include "qemu/cutils.h"
 #include "qemu/help_option.h"
 #include "qemu/throttle-options.h"
@@ -1239,6 +1240,10 @@ SnapshotInfo *qmp_blockdev_snapshot_delete_internal_sync(const char *device,
     info->vm_state_size = sn.vm_state_size;
     info->vm_clock_nsec = sn.vm_clock_nsec % 1000000000;
     info->vm_clock_sec = sn.vm_clock_nsec / 1000000000;
+    if (sn.icount != -1ULL) {
+        info->icount = sn.icount;
+        info->has_icount = true;
+    }
 
     return info;
 
@@ -1447,6 +1452,11 @@ static void internal_snapshot_prepare(BlkActionState *common,
     sn->date_sec = tv.tv_sec;
     sn->date_nsec = tv.tv_usec * 1000;
     sn->vm_clock_nsec = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL);
+    if (replay_mode != REPLAY_MODE_NONE) {
+        sn->icount = replay_get_current_step();
+    } else {
+        sn->icount = -1ULL;
+    }
 
     ret1 = bdrv_snapshot_create(bs, sn);
     if (ret1 < 0) {
include/block/snapshot.h +1 -0
index f73d109..c9c8975 100644
@@ -42,6 +42,7 @@ typedef struct QEMUSnapshotInfo {
     uint32_t date_sec; /* UTC date of the snapshot */
     uint32_t date_nsec;
     uint64_t vm_clock_nsec; /* VM clock relative to boot */
+    uint64_t icount; /* record/replay step */
 } QEMUSnapshotInfo;
 
 int bdrv_snapshot_find(BlockDriverState *bs, QEMUSnapshotInfo *sn_info,
migration/savevm.c +5 -0
index 2d10e45..bdf7136 100644
@@ -2462,6 +2462,11 @@ int save_snapshot(const char *name, Error **errp)
     sn->date_sec = tv.tv_sec;
     sn->date_nsec = tv.tv_usec * 1000;
     sn->vm_clock_nsec = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL);
+    if (replay_mode != REPLAY_MODE_NONE) {
+        sn->icount = replay_get_current_step();
+    } else {
+        sn->icount = -1ULL;
+    }
 
     if (name) {
         ret = bdrv_snapshot_find(bs, old_sn, name);
qapi/block-core.json +6 -1
index cfb37f8..34dc9f1 100644
@@ -26,13 +26,18 @@
 #
 # @vm-clock-nsec: fractional part in nano seconds to be used with vm-clock-sec
 #
+# @icount: Current instruction count. Appears when execution record/replay
+#          is enabled. Used for "time-traveling" to match the moment
+#          in the recorded execution with the snapshots (since 3.1)
+#
 # Since: 1.3
 #
 ##
 { 'struct': 'SnapshotInfo',
   'data': { 'id': 'str', 'name': 'str', 'vm-state-size': 'int',
             'date-sec': 'int', 'date-nsec': 'int',
-            'vm-clock-sec': 'int', 'vm-clock-nsec': 'int' } }
+            'vm-clock-sec': 'int', 'vm-clock-nsec': 'int',
+            '*icount': 'int' } }
 
 ##
 # @ImageInfoSpecificQCow2EncryptionBase:
qapi/block.json +2 -1
index 11f01f2..a6396a9 100644
@@ -176,7 +176,8 @@
 #                    "date-sec": 1000012,
 #                    "date-nsec": 10,
 #                    "vm-clock-sec": 100,
-#                    "vm-clock-nsec": 20
+#                    "vm-clock-nsec": 20,
+#                    "icount": 220414
 #      }
 #    }
 #
#14 lore raw

PATCH v7 8/19 replay: provide and accessor for rr filename

In reply to Pavel Dovgalyuk

This patch adds an accessor function for the name of the record/replay log file. Adding an accessor instead of making variable global, prevents accidental modification of this variable by other modules.

---
 include/sysemu/replay.h |    2 ++
 replay/replay.c         |    5 +++++
 2 files changed, 7 insertions(+)
include/sysemu/replay.h +2 -0
index 3a7c58e..b3f593f 100644
@@ -71,6 +71,8 @@ void replay_start(void);
 void replay_finish(void);
 /*! Adds replay blocker with the specified error description */
 void replay_add_blocker(Error *reason);
+/* Returns name of the replay log file */
+const char *replay_get_filename(void);
 
 /* Processing the instructions */
 
replay/replay.c +5 -0
index ceee20a..e6b1045 100644
@@ -387,3 +387,8 @@ void replay_add_blocker(Error *reason)
 {
     replay_blockers = g_slist_prepend(replay_blockers, reason);
 }
+
+const char *replay_get_filename(void)
+{
+    return replay_filename;
+}
#15 lore raw

PATCH v7 9/19 replay: introduce info hmp/qmp command

In reply to Pavel Dovgalyuk

This patch introduces 'info replay' monitor command and corresponding qmp request.
These commands request the current record/replay mode, replay log file name, and the execution step (number or recorded/replayed instructions). User may use step number for replay_seek/replay_break commands and for controlling the execution of replay.

--

v2:
- renamed info_replay qmp into query-replay (suggested by Eric Blake)
v7:
- added empty line (suggested by Markus Armbruster)

---
 hmp-commands-info.hx      |   14 ++++++++++++++
 hmp.h                     |    1 +
 qapi/misc.json            |   35 +++++++++++++++++++++++++++++++++++
 replay/Makefile.objs      |    3 ++-
 replay/replay-debugging.c |   42 ++++++++++++++++++++++++++++++++++++++++++
 5 files changed, 94 insertions(+), 1 deletion(-)
 create mode 100644 replay/replay-debugging.c
hmp-commands-info.hx +14 -0
index cbee8b9..9f2f35e 100644
@@ -918,6 +918,20 @@ STEXI
 Show SEV information.
 ETEXI
 
+    {
+        .name       = "replay",
+        .args_type  = "",
+        .params     = "",
+        .help       = "show parameters of the record/replay",
+        .cmd        = hmp_info_replay,
+    },
+
+STEXI
+@item info replay
+@findex info replay
+Display the current record/replay mode and the currently executing step.
+ETEXI
+
 STEXI
 @end table
 ETEXI
hmp.h +1 -0
index 5f1addc..d792149 100644
@@ -148,5 +148,6 @@ void hmp_hotpluggable_cpus(Monitor *mon, const QDict *qdict);
 void hmp_info_vm_generation_id(Monitor *mon, const QDict *qdict);
 void hmp_info_memory_size_summary(Monitor *mon, const QDict *qdict);
 void hmp_info_sev(Monitor *mon, const QDict *qdict);
+void hmp_info_replay(Monitor *mon, const QDict *qdict);
 
 #endif
qapi/misc.json +35 -0
index ada9af5..79f6b21 100644
@@ -3105,6 +3105,41 @@
   'data': [ 'none', 'record', 'play' ] }
 
 ##
+# @ReplayInfo:
+#
+# Status of the record/replay mode.
+#
+# @mode: current mode.
+#
+# @filename: name of the record/replay log file.
+#
+# @step: current step number.
+#
+# Since: 3.1
+#
+##
+{ 'struct': 'ReplayInfo',
+  'data': { 'mode': 'ReplayMode', '*filename': 'str', 'step': 'int' } }
+
+##
+# @query-replay:
+#
+# Retrieves the status of the execution record/replay.
+#
+# Returns: structure with the properties of the record/replay.
+#
+# Since: 3.1
+#
+# Example:
+#
+# -> { "execute": "query-replay" }
+# <- { "return": { "mode": "play", "filename": "log.rr", "step": 220414 } }
+#
+##
+{ 'command': 'query-replay',
+  'returns': 'ReplayInfo' }
+
+##
 # @xen-load-devices-state:
 #
 # Load the state of all devices from file. The RAM and the block devices
replay/Makefile.objs +2 -1
index cee6539..6694e3e 100644
@@ -6,4 +6,5 @@ common-obj-y += replay-input.o
 common-obj-y += replay-char.o
 common-obj-y += replay-snapshot.o
 common-obj-y += replay-net.o
-common-obj-y += replay-audio.o
\ No newline at end of file
+common-obj-y += replay-audio.o
+common-obj-y += replay-debugging.o
replay/replay-debugging.c new file +42 -0

42 lines not on this page · raw message

#17 lore raw

PATCH v7 11/19 replay: implement replay-seek command to proceed to the desired step

In reply to Pavel Dovgalyuk

This patch adds hmp/qmp commands replay_seek/replay-seek that proceed the execution to the specified step.
The commands automatically loads nearest snapshot and replay the execution to find the desired step.

--

v2:
- renamed replay_seek qmp command into replay-seek (suggested by Eric Blake)
v7:
- small fixes related to Markus Armbruster's review

---
 hmp-commands.hx           |   15 +++++++
 hmp.h                     |    1 
 qapi/misc.json            |   11 +++++
 replay/replay-debugging.c |   91 +++++++++++++++++++++++++++++++++++++++++++++
 4 files changed, 118 insertions(+)
hmp-commands.hx +15 -0
index 615d725..1988b2b 100644
@@ -1917,6 +1917,21 @@ Removes replay breakpoint which was previously set with replay_break.
 ETEXI
 
     {
+        .name       = "replay_seek",
+        .args_type  = "step:i",
+        .params     = "step",
+        .help       = "rewinds replay to the specified step",
+        .cmd        = hmp_replay_seek,
+    },
+
+STEXI
+@item replay_seek @var{step}
+@findex replay_seek
+Automatically proceeds to the specified step, when replaying
+the execution.
+ETEXI
+
+    {
         .name       = "info",
         .args_type  = "item:s?",
         .params     = "[subcommand]",
hmp.h +1 -0
index c9b9b4f..d6e1d7e 100644
@@ -151,5 +151,6 @@ void hmp_info_sev(Monitor *mon, const QDict *qdict);
 void hmp_info_replay(Monitor *mon, const QDict *qdict);
 void hmp_replay_break(Monitor *mon, const QDict *qdict);
 void hmp_replay_delete_break(Monitor *mon, const QDict *qdict);
+void hmp_replay_seek(Monitor *mon, const QDict *qdict);
 
 #endif
qapi/misc.json +11 -0
index e00a9b0..869c5fa 100644
@@ -3171,6 +3171,17 @@
 { 'command': 'replay-delete-break' }
 
 ##
+# @replay-seek:
+#
+# Automatically proceeds to the specified step when replaying
+# the execution.
+#
+# @step: destination execution step
+# -> { "execute": "replay-seek", "data": { "step": 220414 } }
+##
+{ 'command': 'replay-seek', 'data': { 'step': 'int' } }
+
+##
 # @xen-load-devices-state:
 #
 # Load the state of all devices from file. The RAM and the block devices
replay/replay-debugging.c +91 -0

100 lines not on this page · raw message

#18 lore raw

PATCH v7 12/19 replay: refine replay-time module

In reply to Pavel Dovgalyuk

This patch removes refactoring artifacts from the replay/replay-time.c

---
 replay/replay-time.c |   32 +++++++++++++-------------------
 1 file changed, 13 insertions(+), 19 deletions(-)
replay/replay-time.c +13 -19
index 17caf35..8986054 100644
@@ -17,16 +17,15 @@
 
 int64_t replay_save_clock(ReplayClockKind kind, int64_t clock, int64_t raw_icount)
 {
-    if (replay_file) {
-        g_assert(replay_mutex_locked());
+    g_assert(replay_file);
+    g_assert(replay_mutex_locked());
 
-        /* Due to the caller's locking requirements we get the icount from it instead
-         * of using replay_save_instructions().
-         */
-        replay_advance_current_step(raw_icount);
-        replay_put_event(EVENT_CLOCK + kind);
-        replay_put_qword(clock);
-    }
+    /* Due to the caller's locking requirements we get the icount from it instead
+     * of using replay_save_instructions().
+     */
+    replay_advance_current_step(raw_icount);
+    replay_put_event(EVENT_CLOCK + kind);
+    replay_put_qword(clock);
 
     return clock;
 }
@@ -48,20 +47,15 @@ void replay_read_next_clock(ReplayClockKind kind)
 /*! Reads next clock event from the input. */
 int64_t replay_read_clock(ReplayClockKind kind)
 {
+    int64_t ret;
     g_assert(replay_file && replay_mutex_locked());
 
     replay_account_executed_instructions();
 
-    if (replay_file) {
-        int64_t ret;
-        if (replay_next_event_is(EVENT_CLOCK + kind)) {
-            replay_read_next_clock(kind);
-        }
-        ret = replay_state.cached_clock[kind];
-
-        return ret;
+    if (replay_next_event_is(EVENT_CLOCK + kind)) {
+        replay_read_next_clock(kind);
     }
+    ret = replay_state.cached_clock[kind];
 
-    error_report("REPLAY INTERNAL ERROR %d", __LINE__);
-    exit(1);
+    return ret;
 }
#16 lore raw

PATCH v7 10/19 replay: introduce breakpoint at the specified step

In reply to Pavel Dovgalyuk

This patch introduces replay_break, replay_delete_break qmp and hmp commands.
These commands allow stopping at the specified instruction. It may be useful for debugging when there are some known events that should be investigated.
replay_break command has one argument - number of instructions executed since the start of the replay. replay_delete_break removes previously set breakpoint.

--

v2:
- renamed replay_break qmp command into replay-break (suggested by Eric Blake)
v7:
- introduces replay_delete_break command

---
 hmp-commands.hx           |   29 ++++++++++++++++
 hmp.h                     |    2 +
 qapi/misc.json            |   31 +++++++++++++++++
 replay/replay-debugging.c |   84 +++++++++++++++++++++++++++++++++++++++++++++
 replay/replay-internal.h  |    4 ++
 replay/replay.c           |   17 +++++++++
 6 files changed, 167 insertions(+)
hmp-commands.hx +29 -0
index db0c681..615d725 100644
@@ -1888,6 +1888,35 @@ Set QOM property @var{property} of object at location @var{path} to value @var{v
 ETEXI
 
     {
+        .name       = "replay_break",
+        .args_type  = "step:i",
+        .params     = "step",
+        .help       = "sets breakpoint on the specified step of the replay",
+        .cmd        = hmp_replay_break,
+    },
+
+STEXI
+@item replay_break @var{step}
+@findex replay_break
+Set breakpoint on the specified step of the replay.
+Execution stops when the specified step is reached.
+ETEXI
+
+    {
+        .name       = "replay_delete_break",
+        .args_type  = "",
+        .params     = "",
+        .help       = "removes replay breakpoint",
+        .cmd        = hmp_replay_delete_break,
+    },
+
+STEXI
+@item replay_delete_break
+@findex replay_delete_break
+Removes replay breakpoint which was previously set with replay_break.
+ETEXI
+
+    {
         .name       = "info",
         .args_type  = "item:s?",
         .params     = "[subcommand]",
hmp.h +2 -0
index d792149..c9b9b4f 100644
@@ -149,5 +149,7 @@ void hmp_info_vm_generation_id(Monitor *mon, const QDict *qdict);
 void hmp_info_memory_size_summary(Monitor *mon, const QDict *qdict);
 void hmp_info_sev(Monitor *mon, const QDict *qdict);
 void hmp_info_replay(Monitor *mon, const QDict *qdict);
+void hmp_replay_break(Monitor *mon, const QDict *qdict);
+void hmp_replay_delete_break(Monitor *mon, const QDict *qdict);
 
 #endif
qapi/misc.json +31 -0
index 79f6b21..e00a9b0 100644
@@ -3140,6 +3140,37 @@
   'returns': 'ReplayInfo' }
 
 ##
+# @replay-break:
+#
+# Set breakpoint on the specified step of the replay.
+# Execution stops when the specified step is reached.
+#
+# @step: execution step to stop at
+#
+# Since: 3.1
+#
+# Example:
+#
+# -> { "execute": "replay-break", "data": { "step": 220414 } }
+#
+##
+{ 'command': 'replay-break', 'data': { 'step': 'int' } }
+
+##
+# @replay-delete-break:
+#
+# Removes replay breakpoint.
+#
+# Since: 3.1
+#
+# Example:
+#
+# -> { "execute": "replay-delete-break" }
+#
+##
+{ 'command': 'replay-delete-break' }
+
+##
 # @xen-load-devices-state:
 #
 # Load the state of all devices from file. The RAM and the block devices
replay/replay-debugging.c +84 -0

93 lines not on this page · raw message

replay/replay-internal.h +4 -0
index af6f4d5..94b7e9b 100644
@@ -91,6 +91,10 @@ extern ReplayState replay_state;
 
 /* File for replay writing */
 extern FILE *replay_file;
+/* Step of the replay breakpoint */
+extern uint64_t replay_break_step;
+/* Timer for the replay breakpoint callback */
+extern QEMUTimer *replay_break_timer;
 
 void replay_put_byte(uint8_t byte);
 void replay_put_event(uint8_t event);
replay/replay.c +17 -0
index e6b1045..c6d5de9 100644
@@ -34,6 +34,10 @@ static char *replay_filename;
 ReplayState replay_state;
 static GSList *replay_blockers;
 
+/* Replay breakpoints */
+uint64_t replay_break_step = -1ULL;
+QEMUTimer *replay_break_timer;
+
 bool replay_next_event_is(int event)
 {
     bool res = false;
@@ -73,6 +77,13 @@ int replay_get_instructions(void)
     replay_mutex_lock();
     if (replay_next_event_is(EVENT_INSTRUCTION)) {
         res = replay_state.instructions_count;
+        if (replay_break_step != -1LL) {
+            uint64_t current = replay_get_current_step();
+            assert(replay_break_step >= current);
+            if (current + res > replay_break_step) {
+                res = replay_break_step - current;
+            }
+        }
     }
     replay_mutex_unlock();
     return res;
@@ -99,6 +110,12 @@ void replay_account_executed_instructions(void)
                    will be read from the log. */
                 qemu_notify_event();
             }
+            /* Execution reached the break step */
+            if (replay_break_step == replay_state.current_step) {
+                /* Cannot make callback directly from the vCPU thread */
+                timer_mod_ns(replay_break_timer,
+                    qemu_clock_get_ns(QEMU_CLOCK_REALTIME));
+            }
         }
     }
 }
#19 lore raw

PATCH v7 13/19 replay: flush rr queue before loading the vmstate

In reply to Pavel Dovgalyuk

Non-empty record/replay queue prevents saving and loading the VM state, because it includes pending bottom halves and block coroutines. But when the new VM state is loaded, we don't have to preserve the consistency of the current state anymore. Therefore this patch just flushes the queue allowing the coroutines to finish.

---
 include/sysemu/replay.h  |    2 ++
 migration/savevm.c       |   10 ++++------
 replay/replay-internal.h |    2 --
 3 files changed, 6 insertions(+), 8 deletions(-)
include/sysemu/replay.h +2 -0
index b3f593f..cc5e9b0 100644
@@ -140,6 +140,8 @@ void replay_disable_events(void);
 void replay_enable_events(void);
 /*! Returns true when saving events is enabled */
 bool replay_events_enabled(void);
+/* Flushes events queue */
+void replay_flush_events(void);
 /*! Adds bottom half event to the queue */
 void replay_bh_schedule_event(QEMUBH *bh);
 /*! Adds input event to the queue */
migration/savevm.c +4 -6
index bdf7136..57a297c 100644
@@ -2615,12 +2615,6 @@ int load_snapshot(const char *name, Error **errp)
     AioContext *aio_context;
     MigrationIncomingState *mis = migration_incoming_get_current();
 
-    if (!replay_can_snapshot()) {
-        error_report("Record/replay does not allow loading snapshot "
-                     "right now. Try once more later.");
-        return -EINVAL;
-    }
-
     if (!bdrv_all_can_snapshot(&bs)) {
         error_setg(errp,
                    "Device '%s' is writable but does not support snapshots",
@@ -2654,6 +2648,10 @@ int load_snapshot(const char *name, Error **errp)
         return -EINVAL;
     }
 
+    /* Flush the record/replay queue. Now the VM state is going
+       to change. Therefore we don't need to preserve its consistency */
+    replay_flush_events();
+
     /* Flush all IO requests so they don't interfere with the new state.  */
     bdrv_drain_all_begin();
 
replay/replay-internal.h +0 -2
index 94b7e9b..e37b201 100644
@@ -146,8 +146,6 @@ void replay_read_next_clock(unsigned int kind);
 void replay_init_events(void);
 /*! Clears internal data structures for events handling */
 void replay_finish_events(void);
-/*! Flushes events queue */
-void replay_flush_events(void);
 /*! Returns true if there are any unsaved events in the queue */
 bool replay_has_events(void);
 /*! Saves events from queue into the file */
#20 lore raw

PATCH v7 14/19 gdbstub: add reverse step support in replay mode

In reply to Pavel Dovgalyuk

GDB remote protocol supports two reverse debugging commands:
reverse step and reverse continue.
This patch adds support of the first one to the gdbstub. Reverse step is intended to step one instruction in the backwards direction. This is not possible in regular execution.
But replayed execution is deterministic, therefore we can load one of the prior snapshots and proceed to the desired step. It is equivalent to stepping one instruction back.
There should be at least one snapshot preceding the debugged part of the replay log.

---
 accel/tcg/translator.c    |    1 +
 cpus.c                    |   14 +++++++++++---
 exec.c                    |    5 +++++
 gdbstub.c                 |   42 +++++++++++++++++++++++++++++++++++++++---
 include/sysemu/replay.h   |    7 +++++++
 replay/replay-debugging.c |   33 +++++++++++++++++++++++++++++++++
 stubs/replay.c            |    5 +++++
 7 files changed, 101 insertions(+), 6 deletions(-)
accel/tcg/translator.c +1 -0
index afd0a49..33a543e 100644
@@ -17,6 +17,7 @@
 #include "exec/gen-icount.h"
 #include "exec/log.h"
 #include "exec/translator.h"
+#include "sysemu/replay.h"
 
 /* Pairs with tcg_clear_temp_count.
    To be called by #TranslatorOps.{translate_insn,tb_stop} if
cpus.c +11 -3
index d60ddf1..5189a7e 100644
@@ -1104,9 +1104,17 @@ static bool cpu_can_run(CPUState *cpu)
 
 static void cpu_handle_guest_debug(CPUState *cpu)
 {
-    gdb_set_stop_cpu(cpu);
-    qemu_system_debug_request();
-    cpu->stopped = true;
+    if (!replay_running_debug()) {
+        gdb_set_stop_cpu(cpu);
+        qemu_system_debug_request();
+        cpu->stopped = true;
+    } else {
+        if (!cpu->singlestep_enabled) {
+            cpu_single_step(cpu, SSTEP_ENABLE);
+        } else {
+            cpu_single_step(cpu, 0);
+        }
+    }
 }
 
 #ifdef CONFIG_LINUX
exec.c +5 -0
index d0821e6..7e794cb 100644
@@ -2738,6 +2738,11 @@ static void check_watchpoint(int offset, int len, MemTxAttrs attrs, int flags)
     QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
         if (cpu_watchpoint_address_matches(wp, vaddr, len)
             && (wp->flags & flags)) {
+            if (replay_running_debug()) {
+                /* Don't process the watchpoints when we are
+                   in a reverse debugging operation. */
+                return;
+            }
             if (flags == BP_MEM_READ) {
                 wp->flags |= BP_WATCHPOINT_HIT_READ;
             } else {
gdbstub.c +39 -3

80 lines not on this page · raw message

include/sysemu/replay.h +7 -0
index cc5e9b0..73bf5fc 100644
@@ -73,6 +73,13 @@ void replay_finish(void);
 void replay_add_blocker(Error *reason);
 /* Returns name of the replay log file */
 const char *replay_get_filename(void);
+/* Start making one step in backward direction.
+   Used by gdbstub for backwards debugging.
+   Returns true on success. */
+bool replay_reverse_step(void);
+/* Returns true if replay module is processing
+   reverse_continue or reverse_step request */
+bool replay_running_debug(void);
 
 /* Processing the instructions */
 
replay/replay-debugging.c +33 -0
index d744106..6334d82 100644
@@ -21,6 +21,13 @@
 #include "block/snapshot.h"
 #include "migration/snapshot.h"
 
+static bool replay_is_debugging;
+
+bool replay_running_debug(void)
+{
+    return replay_is_debugging;
+}
+
 void hmp_info_replay(Monitor *mon, const QDict *qdict)
 {
     if (replay_mode == REPLAY_MODE_NONE) {
@@ -215,3 +222,29 @@ void hmp_replay_seek(Monitor *mon, const QDict *qdict)
         return;
     }
 }
+
+static void replay_stop_vm_debug(void *opaque)
+{
+    replay_is_debugging = false;
+    vm_stop(RUN_STATE_DEBUG);
+    replay_break(-1LL, NULL, NULL);
+}
+
+bool replay_reverse_step(void)
+{
+    Error *err = NULL;
+
+    assert(replay_mode == REPLAY_MODE_PLAY);
+
+    if (replay_get_current_step() != 0) {
+        replay_seek(replay_get_current_step() - 1, replay_stop_vm_debug, &err);
+        if (err) {
+            error_free(err);
+            return false;
+        }
+        replay_is_debugging = true;
+        return true;
+    }
+
+    return false;
+}
stubs/replay.c +5 -0
index 4ac6078..521552f 100644
@@ -80,3 +80,8 @@ void replay_mutex_lock(void)
 void replay_mutex_unlock(void)
 {
 }
+
+bool replay_reverse_step(void)
+{
+    return false;
+}
#21 lore raw

PATCH v7 15/19 gdbstub: add reverse continue support in replay mode

In reply to Pavel Dovgalyuk

This patch adds support of the reverse continue operation for gdbstub. Reverse continue finds the last breakpoint that would happen in normal execution from the beginning to the current moment. Implementation of the reverse continue replays the execution twice:
to find the breakpoints that were hit and to seek to the last breakpoint. Reverse continue loads the previous snapshot and tries to find the breakpoint since that moment. If there are no such breakpoints, it proceeds to the earlier snapshot, and so on. When no breakpoints or watchpoints were hit at all, execution stops at the beginning of the replay log.

---
 cpus.c                    |    3 ++
 exec.c                    |    1 +
 gdbstub.c                 |   10 ++++++-
 include/sysemu/replay.h   |    6 ++++
 replay/replay-debugging.c |   69 +++++++++++++++++++++++++++++++++++++++++++++
 stubs/replay.c            |    5 +++
 6 files changed, 93 insertions(+), 1 deletion(-)
cpus.c +3 -0
index 5189a7e..8c0d7c4 100644
@@ -1110,6 +1110,9 @@ static void cpu_handle_guest_debug(CPUState *cpu)
         cpu->stopped = true;
     } else {
         if (!cpu->singlestep_enabled) {
+            /* Report about the breakpoint and
+               make a single step to skip it */
+            replay_breakpoint();
             cpu_single_step(cpu, SSTEP_ENABLE);
         } else {
             cpu_single_step(cpu, 0);
exec.c +1 -0
index 7e794cb..879b51b 100644
@@ -2741,6 +2741,7 @@ static void check_watchpoint(int offset, int len, MemTxAttrs attrs, int flags)
             if (replay_running_debug()) {
                 /* Don't process the watchpoints when we are
                    in a reverse debugging operation. */
+                replay_breakpoint();
                 return;
             }
             if (flags == BP_MEM_READ) {
gdbstub.c +9 -1
index 6af98998..b17bbe5 100644
@@ -1111,6 +1111,14 @@ static int gdb_handle_packet(GDBState *s, const char *line_buf)
                     put_packet(s, "E14");
                     break;
                 }
+            case 'c':
+                if (replay_reverse_continue()) {
+                    gdb_continue(s);
+                    return RS_IDLE;
+                } else {
+                    put_packet(s, "E14");
+                    break;
+                }
             default:
                 goto unknown_command;
             }
@@ -1381,7 +1389,7 @@ static int gdb_handle_packet(GDBState *s, const char *line_buf)
                 pstrcat(buf, sizeof(buf), ";qXfer:features:read+");
             }
             if (replay_mode == REPLAY_MODE_PLAY) {
-                pstrcat(buf, sizeof(buf), ";ReverseStep+");
+                pstrcat(buf, sizeof(buf), ";ReverseStep+;ReverseContinue+");
             }
             put_packet(s, buf);
             break;
include/sysemu/replay.h +6 -0
index 73bf5fc..05cb40a 100644
@@ -77,9 +77,15 @@ const char *replay_get_filename(void);
    Used by gdbstub for backwards debugging.
    Returns true on success. */
 bool replay_reverse_step(void);
+/* Start searching the last breakpoint/watchpoint.
+   Used by gdbstub for backwards debugging.
+   Returns true if the process successfully started. */
+bool replay_reverse_continue(void);
 /* Returns true if replay module is processing
    reverse_continue or reverse_step request */
 bool replay_running_debug(void);
+/* Called in reverse debugging mode to collect breakpoint information */
+void replay_breakpoint(void);
 
 /* Processing the instructions */
 
replay/replay-debugging.c +69 -0

78 lines not on this page · raw message

stubs/replay.c +5 -0
index 521552f..ee64972 100644
@@ -85,3 +85,8 @@ bool replay_reverse_step(void)
 {
     return false;
 }
+
+bool replay_reverse_continue(void)
+{
+    return false;
+}
#22 lore raw

PATCH v7 16/19 replay: describe reverse debugging in docs/replay.txt

In reply to Pavel Dovgalyuk

This patch updates the documentation and describes usage of the reverse debugging in QEMU+GDB.

---
 docs/replay.txt |   33 +++++++++++++++++++++++++++++++++
 1 file changed, 33 insertions(+)
docs/replay.txt +33 -0
index 2c2c5f6..8447fdd 100644
@@ -293,6 +293,39 @@ for recording and replaying must contain identical number of ports in record
 and replay modes, but their backends may differ.
 E.g., '-serial stdio' in record mode, and '-serial null' in replay mode.
 
+Reverse debugging
+-----------------
+
+Reverse debugging allows "executing" the program in reverse direction.
+GDB remote protocol supports "reverse step" and "reverse continue"
+commands. The first one steps single instruction backwards in time,
+and the second one finds the last breakpoint in the past.
+
+Recorded executions may be used to enable reverse debugging. QEMU can't
+execute the code in backwards direction, but can load a snapshot and
+replay forward to find the desired position or breakpoint.
+
+The following GDB commands are supported:
+ - reverse-stepi (or rsi) - step one instruction backwards
+ - reverse-continue (or rc) - find last breakpoint in the past
+
+Reverse step loads the nearest snapshot and replays the execution until
+the required instruction is met.
+
+Reverse continue may include several passes of examining the execution
+between the snapshots. Each of the passes include the following steps:
+ 1. loading the snapshot
+ 2. replaying to examine the breakpoints
+ 3. if breakpoint or watchpoint was met
+    - loading the snaphot again
+    - replaying to the required breakpoint
+ 4. else
+    - proceeding to the p.1 with the earlier snapshot
+
+Therefore usage of the reverse debugging requires at least one snapshot
+created in advance. See the "Snapshotting" section to learn about running
+record/replay and creating the snapshot in these modes.
+
 Replay log format
 -----------------
 
#23 lore raw

PATCH v7 17/19 replay: add BH oneshot event for block layer

In reply to Pavel Dovgalyuk

Replay is capable of recording normal BH events, but sometimes there are single use callbacks scheduled with aio_bh_schedule_oneshot function. This patch enables recording and replaying such callbacks. Block layer uses these events for calling the completion function. Replaying these calls makes the execution deterministic.

--

v6:
- moved stub function to the separate file for fixing linux-user build

---
 block/block-backend.c    |    3 ++-
 include/sysemu/replay.h  |    3 +++
 replay/replay-events.c   |   16 ++++++++++++++++
 replay/replay-internal.h |    1 +
 replay/replay.c          |    2 +-
 stubs/Makefile.objs      |    1 +
 stubs/replay-user.c      |    9 +++++++++
 7 files changed, 33 insertions(+), 2 deletions(-)
 create mode 100644 stubs/replay-user.c
block/block-backend.c +2 -1
index dc0cd57..04f5554 100644
@@ -17,6 +17,7 @@
 #include "block/throttle-groups.h"
 #include "sysemu/blockdev.h"
 #include "sysemu/sysemu.h"
+#include "sysemu/replay.h"
 #include "qapi/error.h"
 #include "qapi/qapi-events-block.h"
 #include "qemu/id.h"
@@ -1379,7 +1380,7 @@ static BlockAIOCB *blk_aio_prwv(BlockBackend *blk, int64_t offset, int bytes,
 
     acb->has_returned = true;
     if (acb->rwco.ret != NOT_DONE) {
-        aio_bh_schedule_oneshot(blk_get_aio_context(blk),
+        replay_bh_schedule_oneshot_event(blk_get_aio_context(blk),
                                 blk_aio_complete_bh, acb);
     }
 
include/sysemu/replay.h +3 -0
index 05cb40a..5c71858 100644
@@ -157,6 +157,9 @@ bool replay_events_enabled(void);
 void replay_flush_events(void);
 /*! Adds bottom half event to the queue */
 void replay_bh_schedule_event(QEMUBH *bh);
+/* Adds oneshot bottom half event to the queue */
+void replay_bh_schedule_oneshot_event(AioContext *ctx,
+    QEMUBHFunc *cb, void *opaque);
 /*! Adds input event to the queue */
 void replay_input_event(QemuConsole *src, InputEvent *evt);
 /*! Adds input sync event to the queue */
replay/replay-events.c +16 -0
index 0964a82..83a7d81 100644
@@ -37,6 +37,9 @@ static void replay_run_event(Event *event)
     case REPLAY_ASYNC_EVENT_BH:
         aio_bh_call(event->opaque);
         break;
+    case REPLAY_ASYNC_EVENT_BH_ONESHOT:
+        ((QEMUBHFunc*)event->opaque)(event->opaque2);
+        break;
     case REPLAY_ASYNC_EVENT_INPUT:
         qemu_input_event_send_impl(NULL, (InputEvent *)event->opaque);
         qapi_free_InputEvent((InputEvent *)event->opaque);
@@ -132,6 +135,17 @@ void replay_bh_schedule_event(QEMUBH *bh)
     }
 }
 
+void replay_bh_schedule_oneshot_event(AioContext *ctx,
+    QEMUBHFunc *cb, void *opaque)
+{
+    if (events_enabled) {
+        uint64_t id = replay_get_current_step();
+        replay_add_event(REPLAY_ASYNC_EVENT_BH_ONESHOT, cb, opaque, id);
+    } else {
+        aio_bh_schedule_oneshot(ctx, cb, opaque);
+    }
+}
+
 void replay_add_input_event(struct InputEvent *event)
 {
     replay_add_event(REPLAY_ASYNC_EVENT_INPUT, event, NULL, 0);
@@ -162,6 +176,7 @@ static void replay_save_event(Event *event, int checkpoint)
         /* save event-specific data */
         switch (event->event_kind) {
         case REPLAY_ASYNC_EVENT_BH:
+        case REPLAY_ASYNC_EVENT_BH_ONESHOT:
             replay_put_qword(event->id);
             break;
         case REPLAY_ASYNC_EVENT_INPUT:
@@ -216,6 +231,7 @@ static Event *replay_read_event(int checkpoint)
     /* Events that has not to be in the queue */
     switch (replay_state.read_event_kind) {
     case REPLAY_ASYNC_EVENT_BH:
+    case REPLAY_ASYNC_EVENT_BH_ONESHOT:
         if (replay_state.read_event_id == -1) {
             replay_state.read_event_id = replay_get_qword();
         }
replay/replay-internal.h +1 -0
index e37b201..af2b941 100644
@@ -51,6 +51,7 @@ enum ReplayEvents {
 
 enum ReplayAsyncEventKind {
     REPLAY_ASYNC_EVENT_BH,
+    REPLAY_ASYNC_EVENT_BH_ONESHOT,
     REPLAY_ASYNC_EVENT_INPUT,
     REPLAY_ASYNC_EVENT_INPUT_SYNC,
     REPLAY_ASYNC_EVENT_CHAR_READ,
replay/replay.c +1 -1
index c6d5de9..93d2573 100644
@@ -22,7 +22,7 @@
 
 /* Current version of the replay mechanism.
    Increase it when file format changes. */
-#define REPLAY_VERSION              0xe02007
+#define REPLAY_VERSION              0xe02008
 /* Size of replay log header */
 #define HEADER_SIZE                 (sizeof(uint32_t) + sizeof(uint64_t))
 
stubs/Makefile.objs +1 -0
index 53d3f32..2acbac3 100644
@@ -24,6 +24,7 @@ stub-obj-y += monitor.o
 stub-obj-y += notify-event.o
 stub-obj-y += qtest.o
 stub-obj-y += replay.o
+stub-obj-y += replay-user.o
 stub-obj-y += runstate-check.o
 stub-obj-y += set-fd-handler.o
 stub-obj-y += slirp.o
stubs/replay-user.c new file +9 -0
new file mode 100644
index 0000000..2ad9e27
@@ -0,0 +1,9 @@
+#include "qemu/osdep.h"
+#include "sysemu/replay.h"
+#include "sysemu/sysemu.h"
+
+void replay_bh_schedule_oneshot_event(AioContext *ctx,
+    QEMUBHFunc *cb, void *opaque)
+{
+    aio_bh_schedule_oneshot(ctx, cb, opaque);
+}
#28 lore raw

PATCH v7 18/19 replay: init rtc after enabling the replay

In reply to Pavel Dovgalyuk

This patch postpones the call of 'configure_rtc' function. This call uses host clock to configure the rtc, but host clock access should be recorded when using icount record/replay mode. Therefore now rtc is configured after switching record/replay mode on.

---
 vl.c |   11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
vl.c +7 -4
index a3e2e47..f910072 100644
@@ -2932,6 +2932,7 @@ int main(int argc, char **argv, char **envp)
     DisplayState *ds;
     QemuOpts *opts, *machine_opts;
     QemuOpts *icount_opts = NULL, *accel_opts = NULL;
+    QemuOpts *rtc_opts = NULL;
     QemuOptsList *olist;
     int optind;
     const char *optarg;
@@ -3738,12 +3739,11 @@ int main(int argc, char **argv, char **envp)
                 warn_report("This option is ignored and will be removed soon");
                 break;
             case QEMU_OPTION_rtc:
-                opts = qemu_opts_parse_noisily(qemu_find_opts("rtc"), optarg,
-                                               false);
-                if (!opts) {
+                rtc_opts = qemu_opts_parse_noisily(qemu_find_opts("rtc"),
+                                                   optarg, false);
+                if (!rtc_opts) {
                     exit(1);
                 }
-                configure_rtc(opts);
                 break;
             case QEMU_OPTION_tb_size:
 #ifndef CONFIG_TCG
@@ -3954,6 +3954,9 @@ int main(int argc, char **argv, char **envp)
     loc_set_none();
 
     replay_configure(icount_opts);
+    if (rtc_opts) {
+        configure_rtc(rtc_opts);
+    }
 
     if (incoming && !preconfig_exit_requested) {
         error_report("'preconfig' and 'incoming' options are "
#30 lore raw

PATCH v7 19/19 replay: document development rules

In reply to Pavel Dovgalyuk

This patch introduces docs/devel/replay.txt which describes the rules that should be followed to make virtual devices usable in record/replay mode.

---
 docs/devel/replay.txt |   45 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 45 insertions(+)
 create mode 100644 docs/devel/replay.txt
docs/devel/replay.txt new file +45 -0
new file mode 100644
index 0000000..61dac1b
@@ -0,0 +1,45 @@
+Record/replay mechanism, that could be enabled through icount mode, expects
+the virtual devices to satisfy the following requirements.
+
+The main idea behind this document is that everything that affects
+the guest state during execution in icount mode should be deterministic.
+
+Timers
+======
+
+All virtual devices should use virtual clock for timers that change the guest
+state. Virtual clock is deterministic, therefore such timers are deterministic
+too.
+
+Virtual devices can also use realtime clock for the events that do not change
+the guest state directly. When the clock ticking should depend on VM execution
+speed, use virtual ext clock. It is not deterministic, but its speed depends
+on the guest execution. This clock is used by the virtual devices (e.g.,
+slirp routing device) that lie outside the replayed guest.
+
+Bottom halves
+=============
+
+Bottom half callbacks, that affect the guest state, should be invoked through
+replay_bh_schedule_event or replay_bh_schedule_oneshot_event functions.
+Their invocations are saved in record mode and synchronized with the existing
+log in replay mode.
+
+Saving/restoring the VM state
+=============================
+
+All fields in the device state structure (including virtual timers)
+should be restored by loadvm to the same values they had before savevm.
+
+Avoid accessing other devices' state, because the order of saving/restoring
+is not defined. It means that you should not call functions like
+'update_irq' in post_load callback. Save everything explicitly to avoid
+the dependencies that may make restoring the VM state non-deterministic.
+
+Stopping the VM
+===============
+
+Stopping the guest should not interfere with its state (with the exception
+of the network connections, that could be broken by the remote timeouts).
+VM can be stopped at any moment of replay by the user. Restarting the VM
+after that stop should not break the replay by the unneeded guest state change.
#32 lore raw
In reply to Pavel Dovgalyuk

Hi Pavel,

Please do not CC me in this series, this isn’t anyhow relevant for me anymore.

Regards, Alex

ср, 10 окт. 2018 г. в 16:30, Pavel Dovgalyuk <[email protected]>:

Quoted 133 lines

GDB remote protocol supports reverse debugging of the targets. It includes 'reverse step' and 'reverse continue' operations. The first one finds the previous step of the execution, and the second one is intended to stop at the last breakpoint that would happen when the program is executed normally.

Reverse debugging is possible in the replay mode, when at least one snapshot was created at the record or replay phase. QEMU can use these snapshots for travelling back in time with GDB.

Running the execution in replay mode allows using GDB reverse debugging commands:
- reverse-stepi (or rsi): Steps one instruction to the past. QEMU loads on of the prior snapshots and proceeds to the desired instruction forward. When that step is reaches, execution stops.
- reverse-continue (or rc): Runs execution "backwards". QEMU tries to find breakpoint or watchpoint by loaded prior snapshot and replaying the execution. Then QEMU loads snapshots again and replays to the latest breakpoint. When there are no breakpoints in the examined section of the execution, QEMU finds one more snapshot and tries again. After the first snapshot is processed, execution stops at this snapshot.

The set of patches include the following modifications:
- gdbstub update for reverse debugging support
- functions that automatically perform reverse step and reverse continue operations
- hmp/qmp commands for manipulating the replay process
- improvement of the snapshotting for saving the execution step in the snapshot parameters
- other record/replay fixes

The patches are available in the repository:
https://github.com/ispras/qemu/tree/rr-181010

v7 changes:
- rebased to the new master with upstreamed patches from the series
- several improvements in hmp/qmp commands handling (suggested by Markus
Armbruster)
- fixed record/replay with '-rtc base' option enabled
- added document with virtual hardware requirements

v6 changes:
- rebased to the new version of master
- fixed build of linux-user configurations
- added new clock for slirp and vnc timers

v5 changes:
- multiple fixes of record/replay bugs appeared after QEMU core update
- changed reverse debugging to 'since 3.1'

v4 changes:
- changed 'since 2.13' to 'since 3.0' in json (as suggested by Eric Blake)

v3 changes:
- Fixed PS/2 bug with save/load vm, which caused failures of the replay.
- Rebased to the new code base.
- Minor fixes.

v2 changes:
- documented reverse debugging
- fixed start vmstate loading in record mode
- documented qcow2 changes (as suggested by Eric Blake)
- made icount SnapshotInfo field optional (as suggested by Eric Blake)
- renamed qmp commands (as suggested by Eric Blake)
- minor changes

---

Pavel Dovgalyuk (19):
block: implement bdrv_snapshot_goto for blkreplay
replay: disable default snapshot for record/replay
replay: update docs for record/replay with block devices
replay: don't drain/flush bdrv queue while RR is working
replay: finish record/replay before closing the disks
qcow2: introduce icount field for snapshots
migration: introduce icount field for snapshots
replay: provide and accessor for rr filename
replay: introduce info hmp/qmp command
replay: introduce breakpoint at the specified step
replay: implement replay-seek command to proceed to the desired step
replay: refine replay-time module
replay: flush rr queue before loading the vmstate
gdbstub: add reverse step support in replay mode
gdbstub: add reverse continue support in replay mode
replay: describe reverse debugging in docs/replay.txt
replay: add BH oneshot event for block layer
replay: init rtc after enabling the replay
replay: document development rules

 accel/tcg/translator.c    |    1
 block/blkreplay.c         |    8 +
 block/block-backend.c     |    3
 block/io.c                |   22 +++
 block/qapi.c              |   17 ++
 block/qcow2-snapshot.c    |    9 +
 block/qcow2.h             |    2
 blockdev.c                |   10 +
 cpus.c                    |   19 ++-
 docs/devel/replay.txt     |   45 ++++++
 docs/interop/qcow2.txt    |    4 +
 docs/replay.txt           |   45 ++++++
 exec.c                    |    6 +
 gdbstub.c                 |   50 +++++++
 hmp-commands-info.hx      |   14 ++
 hmp-commands.hx           |   44 ++++++
 hmp.h                     |    4 +
 include/block/snapshot.h  |    1
 include/sysemu/replay.h   |   20 +++
 migration/savevm.c        |   15 +-
 qapi/block-core.json      |    7 +
 qapi/block.json           |    3
 qapi/misc.json            |   77 +++++++++++
 replay/Makefile.objs      |    3
 replay/replay-debugging.c |  319
+++++++++++++++++++++++++++++++++++++++++++++
 replay/replay-events.c    |   16 ++
 replay/replay-internal.h  |    7 +
 replay/replay-time.c      |   32 ++---
 replay/replay.c           |   26 ++++
 stubs/Makefile.objs       |    1
 stubs/replay-user.c       |    9 +
 stubs/replay.c            |   10 +
 vl.c                      |   22 ++-
 33 files changed, 818 insertions(+), 53 deletions(-)
 create mode 100644 docs/devel/replay.txt
 create mode 100644 replay/replay-debugging.c
 create mode 100644 stubs/replay-user.c

--
Pavel Dovgalyuk

Signature · 1 line
Aleksandr Bezzubikov
#29 lore raw

Re: [Qemu-devel] [PATCH v7 18/19] replay: init rtc after enabling the replay

In reply to Pavel Dovgalyuk

I guess 'configure_rtc' function is not the only one which call should be postponed. Since record/replay uses its 'hooks' at very low levels of qemu core, any option processing is subject to potential indirect dependence on effects of 'replay_configure' call. The most apparent example would be 'qemu_clock_get_*(QEMU_CLOCK_HOST)' calls, which are used widely. Furthermore, right now I submitting a patch which adds one such call in place before any option parsed yet, so it will also need to be moved to follow this patch. Even if such calls aren't involved in current version, no guarantees that they will not appear in future (like I already doing it), because there are nothing which prevents developers from doing so.

One way, how this could be enforced, is to add something like 'REPLAY_MODE_INVALID' to 'ReplayMode' enum, replace initialization of global 'replay_mode' variable with this value and add corresponding checks/assertions to 'REPLAY_CLOCK' macro, etc.

But it would be more reliable to move 'icount_opts' processing out of common processing loop (marked as 'second pass of option parsing') and place it to new pass, located before current 'second' one (like it already done for '-nouserconfig' option), and also move 'replace_configure' right after it. Probably it will also require to tune 'icount_opts' processing and record/replay initialization. Since record/replay feature is very specific, I consider it's worth to make such exception for it.

ср, 10 окт. 2018 г. в 19:32, Pavel Dovgalyuk <[email protected]>:

Quoted 42 lines

This patch postpones the call of 'configure_rtc' function. This call uses host clock to configure the rtc, but host clock access should be recorded when using icount record/replay mode. Therefore now rtc is configured after switching record/replay mode on.

---
 vl.c |   11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
vl.c +4 -4
index a3e2e47..f910072 100644
@@ -2932,6 +2932,7 @@ int main(int argc, char **argv, char **envp)
     DisplayState *ds;
     QemuOpts *opts, *machine_opts;
     QemuOpts *icount_opts = NULL, *accel_opts = NULL;
+    QemuOpts *rtc_opts = NULL;
     QemuOptsList *olist;
     int optind;
     const char *optarg;
@@ -3738,12 +3739,11 @@ int main(int argc, char **argv, char **envp)
                 warn_report("This option is ignored and will be removed
 soon");
                 break;
             case QEMU_OPTION_rtc:
-                opts = qemu_opts_parse_noisily(qemu_find_opts("rtc"),
 optarg,
-                                               false);
-                if (!opts) {
+                rtc_opts = qemu_opts_parse_noisily(qemu_find_opts("rtc"),
+                                                   optarg, false);
+                if (!rtc_opts) {
                     exit(1);
                 }
-                configure_rtc(opts);
                 break;

case QEMU_OPTION_tb_size:
#ifndef CONFIG_TCG

vl.c +0 -0
@@ -3954,6 +3954,9 @@ int main(int argc, char **argv, char **envp)
     loc_set_none();

vl.c int main(int argc, char **argv, char **envp)

Quoted 10 lines
vl.c +3 -1
 
     replay_configure(icount_opts);
+    if (rtc_opts) {
+        configure_rtc(rtc_opts);
+    }
 
     if (incoming && !preconfig_exit_requested) {
         error_report("'preconfig' and 'incoming' options are "
 
--

С уважением,
Артем Писаренко

#31 lore raw

Re: [Qemu-devel] [PATCH v7 19/19] replay: document development rules

In reply to Pavel Dovgalyuk

Great! I'm voting with all my fingers up for such rules. But I would suggest even more generic rules which prevent breaking determinism in a more wide sense. At least, where such breakage is trivial to avoid.

Currently I'm working on modification, which extends conditions where guest execution is kept deterministic, above such narrow set like "only rtc clock=vm, no serial devices, no network, no external communication interfaces at all, etc...". Also I'm dealing with bugs (features?), which prevents advertised determinism even in such restricted conditions. I found that my work involves very similar efforts as Pavel's work. And this is extra hard work. I'm feeling like fighting with hundreds of maintainers and contributors, whose efforts are in opposite direction. They starting from qemu underlying core, encouraging asyncronous processing (aio, threads, virtio ioeventfd, etc.), and ending with particular modules or hardware models, which negligently uses any kind of non-blocking calls, callbacks, and even inappropriate QEMUClockType. Nobody cares about synchronization with vcpu thread, except Pavel, me and, possibly, 1-2 more persons in a whole world. I can understand why. Main reason is that it might greatly degradate performance of emulation, which might be avoided by introducing very high complexity. So, my words shouldn't be treated as any kind of criticism. I perfectly understand that it's a complex issue.

Key difference of record/replay is that it solves problem by hooking calls of any source of asynchrony at low level and just replaying it. In other words, it deals with end effects, whereas non-record/replay use case doesn't allow such solution and have to eliminate source of undesired asynchrony by design. As such, record/replay have strong immunity to violation of 'generic determinism' rules and even to hidden and tricky bugs in any module which affects guest state. And that's why development rules Pavel imposes are so democratic (relative to generic ones, I would like to exist).

Anyway, it's just generic idea for discussion. I know, it needs to be more specific. But, if nobody will express interest, I see no reason to continue.

P.S. Trivial example of how qemu could extend conditions for deterministic execution. Сhardev would perform writing to backend using blocking calls, thus making possible deterministic execution in use cases, where guest has only one serial port which outputs data to console and have no interaction with user. At least it would provide user with option, selecting between better performance and determinism.

ср, 10 окт. 2018 г. в 19:32, Pavel Dovgalyuk <[email protected]>:

Quoted 72 lines

This patch introduces docs/devel/replay.txt which describes the rules that should be followed to make virtual devices usable in record/replay mode.

---
 docs/devel/replay.txt |   45 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 45 insertions(+)
 create mode 100644 docs/devel/replay.txt
docs/devel/replay.txt new file +1 -0
new file mode 100644
index 0000000..61dac1b
@@ -0,0 +1,45 @@
+Record/replay mechanism, that could be enabled through icount mode,

expects

docs/devel/replay.txt new file +9 -0
+the virtual devices to satisfy the following requirements.
+
+The main idea behind this document is that everything that affects
+the guest state during execution in icount mode should be deterministic.
+
+Timers
+======
+
+All virtual devices should use virtual clock for timers that change the

guest
+state. Virtual clock is deterministic, therefore such timers are deterministic

docs/devel/replay.txt new file +3 -0
+too.
+
+Virtual devices can also use realtime clock for the events that do not

change
+the guest state directly. When the clock ticking should depend on VM execution
+speed, use virtual ext clock. It is not deterministic, but its speed depends

docs/devel/replay.txt new file +7 -0
+on the guest execution. This clock is used by the virtual devices (e.g.,
+slirp routing device) that lie outside the replayed guest.
+
+Bottom halves
+=============
+
+Bottom half callbacks, that affect the guest state, should be invoked

through

docs/devel/replay.txt new file +2 -0
+replay_bh_schedule_event or replay_bh_schedule_oneshot_event functions.
+Their invocations are saved in record mode and synchronized with the

existing

docs/devel/replay.txt new file +9 -0
+log in replay mode.
+
+Saving/restoring the VM state
+=============================
+
+All fields in the device state structure (including virtual timers)
+should be restored by loadvm to the same values they had before savevm.
+
+Avoid accessing other devices' state, because the order of

saving/restoring

docs/devel/replay.txt new file +11 -0
+is not defined. It means that you should not call functions like
+'update_irq' in post_load callback. Save everything explicitly to avoid
+the dependencies that may make restoring the VM state non-deterministic.
+
+Stopping the VM
+===============
+
+Stopping the guest should not interfere with its state (with the exception
+of the network connections, that could be broken by the remote timeouts).
+VM can be stopped at any moment of replay by the user. Restarting the VM
+after that stop should not break the replay by the unneeded guest state

change.

Quoted 1 line

--

С уважением,
Артем Писаренко

#33 lore raw
In reply to Pavel Dovgalyuk

Ping?

Paolo, what do you think about upstreaming this before soft freeze?

Pavel Dovgalyuk

Quoted 142 lines

-----Original Message-----
From: Pavel Dovgalyuk [mailto:[email protected]]
Sent: Wednesday, October 10, 2018 4:34 PM
To: [email protected]
Cc: [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]; [email protected]
Subject: [PATCH v7 00/19] Fixing record/replay and adding reverse debugging

GDB remote protocol supports reverse debugging of the targets. It includes 'reverse step' and 'reverse continue' operations. The first one finds the previous step of the execution, and the second one is intended to stop at the last breakpoint that would happen when the program is executed normally.

Reverse debugging is possible in the replay mode, when at least one snapshot was created at the record or replay phase. QEMU can use these snapshots for travelling back in time with GDB.

Running the execution in replay mode allows using GDB reverse debugging commands:
- reverse-stepi (or rsi): Steps one instruction to the past. QEMU loads on of the prior snapshots and proceeds to the desired instruction forward. When that step is reaches, execution stops.
- reverse-continue (or rc): Runs execution "backwards". QEMU tries to find breakpoint or watchpoint by loaded prior snapshot and replaying the execution. Then QEMU loads snapshots again and replays to the latest breakpoint. When there are no breakpoints in the examined section of the execution, QEMU finds one more snapshot and tries again. After the first snapshot is processed, execution stops at this snapshot.

The set of patches include the following modifications:
- gdbstub update for reverse debugging support
- functions that automatically perform reverse step and reverse continue operations
- hmp/qmp commands for manipulating the replay process
- improvement of the snapshotting for saving the execution step in the snapshot parameters
- other record/replay fixes

The patches are available in the repository:
https://github.com/ispras/qemu/tree/rr-181010

v7 changes:
- rebased to the new master with upstreamed patches from the series
- several improvements in hmp/qmp commands handling (suggested by Markus Armbruster)
- fixed record/replay with '-rtc base' option enabled
- added document with virtual hardware requirements

v6 changes:
- rebased to the new version of master
- fixed build of linux-user configurations
- added new clock for slirp and vnc timers

v5 changes:
- multiple fixes of record/replay bugs appeared after QEMU core update
- changed reverse debugging to 'since 3.1'

v4 changes:
- changed 'since 2.13' to 'since 3.0' in json (as suggested by Eric Blake)

v3 changes:
- Fixed PS/2 bug with save/load vm, which caused failures of the replay.
- Rebased to the new code base.
- Minor fixes.

v2 changes:
- documented reverse debugging
- fixed start vmstate loading in record mode
- documented qcow2 changes (as suggested by Eric Blake)
- made icount SnapshotInfo field optional (as suggested by Eric Blake)
- renamed qmp commands (as suggested by Eric Blake)
- minor changes

---

Pavel Dovgalyuk (19):
block: implement bdrv_snapshot_goto for blkreplay
replay: disable default snapshot for record/replay
replay: update docs for record/replay with block devices
replay: don't drain/flush bdrv queue while RR is working
replay: finish record/replay before closing the disks
qcow2: introduce icount field for snapshots
migration: introduce icount field for snapshots
replay: provide and accessor for rr filename
replay: introduce info hmp/qmp command
replay: introduce breakpoint at the specified step
replay: implement replay-seek command to proceed to the desired step
replay: refine replay-time module
replay: flush rr queue before loading the vmstate
gdbstub: add reverse step support in replay mode
gdbstub: add reverse continue support in replay mode
replay: describe reverse debugging in docs/replay.txt
replay: add BH oneshot event for block layer
replay: init rtc after enabling the replay
replay: document development rules

 accel/tcg/translator.c    |    1
 block/blkreplay.c         |    8 +
 block/block-backend.c     |    3
 block/io.c                |   22 +++
 block/qapi.c              |   17 ++
 block/qcow2-snapshot.c    |    9 +
 block/qcow2.h             |    2
 blockdev.c                |   10 +
 cpus.c                    |   19 ++-
 docs/devel/replay.txt     |   45 ++++++
 docs/interop/qcow2.txt    |    4 +
 docs/replay.txt           |   45 ++++++
 exec.c                    |    6 +
 gdbstub.c                 |   50 +++++++
 hmp-commands-info.hx      |   14 ++
 hmp-commands.hx           |   44 ++++++
 hmp.h                     |    4 +
 include/block/snapshot.h  |    1
 include/sysemu/replay.h   |   20 +++
 migration/savevm.c        |   15 +-
 qapi/block-core.json      |    7 +
 qapi/block.json           |    3
 qapi/misc.json            |   77 +++++++++++
 replay/Makefile.objs      |    3
 replay/replay-debugging.c |  319 +++++++++++++++++++++++++++++++++++++++++++++
 replay/replay-events.c    |   16 ++
 replay/replay-internal.h  |    7 +
 replay/replay-time.c      |   32 ++---
 replay/replay.c           |   26 ++++
 stubs/Makefile.objs       |    1
 stubs/replay-user.c       |    9 +
 stubs/replay.c            |   10 +
 vl.c                      |   22 ++-
 33 files changed, 818 insertions(+), 53 deletions(-)
 create mode 100644 docs/devel/replay.txt
 create mode 100644 replay/replay-debugging.c
 create mode 100644 stubs/replay-user.c

--
Pavel Dovgalyuk

Acked-by on v7 6/19
#11 lore raw

Re: [Qemu-devel] [PATCH v7 06/19] qcow2: introduce icount field for snapshots

In reply to Pavel Dovgalyuk

Am 10.10.2018 um 15:34 hat Pavel Dovgalyuk geschrieben:

Quoted 10 lines

This patch introduces the icount field for saving within the snapshot. It is required for navigation between the snapshots in record/replay mode.

--

v2:
- documented format changes in docs/interop/qcow2.txt (suggested by Eric Blake)

#5 lore raw

Re: [Qemu-devel] [PATCH v7 04/19] replay: don't drain/flush bdrv queue while RR is working

In reply to Pavel Dovgalyuk

Am 10.10.2018 um 15:33 hat Pavel Dovgalyuk geschrieben:

Quoted 9 lines

In record/replay mode bdrv queue is controlled by replay mechanism. It does not allow saving or loading the snapshots
when bdrv queue is not empty. Stopping the VM is not blocked by nonempty queue, but flushing the queue is still impossible there, because it may cause deadlocks in replay mode.
This patch disables bdrv_drain_all and bdrv_flush_all in record/replay mode.

Disabling bdrv_flush_all() shouldn't make a big difference in replay mode, so I agree with that.

But does the bdrv_drain_all() change mean that bdrv_drain_all_begin() can return while there are still requests in flight? This sounds like something that could break some existing code, because the meaning of the function is that if it returns success, no requests are in flight any more.

What would move the request forward in the replay case once it has been created? Does this also involve some user interaction? Or may we process the next event in a drain callback of blkreplay or something?

Kevin

#13 lore raw

Re: [Qemu-devel] [PATCH v7 07/19] migration: introduce icount field for snapshots

In reply to Pavel Dovgalyuk

Am 10.10.2018 um 15:34 hat Pavel Dovgalyuk geschrieben:

Quoted 14 lines

Saving icount as a parameters of the snapshot allows navigation between them in the execution replay scenario.
This information can be used for finding a specific snapshot for rewinding the recorded execution to the specific moment of the time. E.g., 'reverse step' action needs to load the nearest snapshot which is prior to the current moment of time .

block/qapi.c +0 -0
index c66f949..9f2ccfa 100644

block/qapi.c int bdrv_query_snapshot_info_list(BlockDriverState *bs, · line 213

Quoted 4 lines
block/qapi.c +1 -0
@@ -210,6 +210,7 @@ int bdrv_query_snapshot_info_list(BlockDriverState *bs,
         info->date_nsec     = sn_tab[i].date_nsec;
         info->vm_clock_sec  = sn_tab[i].vm_clock_nsec / 1000000000;
         info->vm_clock_nsec = sn_tab[i].vm_clock_nsec % 1000000000;
+        info->icount        = sn_tab[i].icount;

Now that it's optional, info->has_icount needs to be set, too.

Quoted 8 lines
block/qapi.c +0 -0
         info_list = g_new0(SnapshotInfoList, 1);
         info_list->value = info;
 
qapi/block-core.json +0 -0
index cfb37f8..34dc9f1 100644

qapi/block-core.json line 31

Quoted 6 lines
qapi/block-core.json +3 -0
@@ -26,13 +26,18 @@
 #
 # @vm-clock-nsec: fractional part in nano seconds to be used with vm-clock-sec
 #
+# @icount: Current instruction count. Appears when execution record/replay
+#          is enabled. Used for "time-traveling" to match the moment
+#          in the recorded execution with the snapshots (since 3.1)

"(since 4.0)" now.

Kevin

#24 lore raw

Re: [Qemu-devel] [PATCH v7 17/19] replay: add BH oneshot event for block layer

In reply to Pavel Dovgalyuk

Am 10.10.2018 um 15:35 hat Pavel Dovgalyuk geschrieben:

Quoted 7 lines

Replay is capable of recording normal BH events, but sometimes there are single use callbacks scheduled with aio_bh_schedule_oneshot function. This patch enables recording and replaying such callbacks. Block layer uses these events for calling the completion function. Replaying these calls makes the execution deterministic.

What are the rules for using aio_bh_schedule_oneshot() vs. replay_bh_schedule_oneshot_event()? You are modifying one place, but if I grep for aio_bh_schedule_oneshot(), I find many other places that use it. Why is this one place special?

Quoted 28 lines

--

v6:
- moved stub function to the separate file for fixing linux-user build

---
 block/block-backend.c    |    3 ++-
 include/sysemu/replay.h  |    3 +++
 replay/replay-events.c   |   16 ++++++++++++++++
 replay/replay-internal.h |    1 +
 replay/replay.c          |    2 +-
 stubs/Makefile.objs      |    1 +
 stubs/replay-user.c      |    9 +++++++++
 7 files changed, 33 insertions(+), 2 deletions(-)
 create mode 100644 stubs/replay-user.c
block/block-backend.c +1 -0
index dc0cd57..04f5554 100644
@@ -17,6 +17,7 @@
 #include "block/throttle-groups.h"
 #include "sysemu/blockdev.h"
 #include "sysemu/sysemu.h"
+#include "sysemu/replay.h"
 #include "qapi/error.h"
 #include "qapi/qapi-events-block.h"
 #include "qemu/id.h"

block/block-backend.c static BlockAIOCB *blk_aio_prwv(BlockBackend *blk, int64_t offset, int bytes, · line 1384

Quoted 6 lines
block/block-backend.c +1 -1
@@ -1379,7 +1380,7 @@ static BlockAIOCB *blk_aio_prwv(BlockBackend *blk, int64_t offset, int bytes,
 
     acb->has_returned = true;
     if (acb->rwco.ret != NOT_DONE) {
-        aio_bh_schedule_oneshot(blk_get_aio_context(blk),
+        replay_bh_schedule_oneshot_event(blk_get_aio_context(blk),
                                 blk_aio_complete_bh, acb);

Indentation is off.

Kevin

#6 lore raw

Re: [Qemu-devel] [PATCH v7 04/19] replay: don't drain/flush bdrv queue while RR is working

In reply to Kevin Wolf
Quoted 21 lines

From: Kevin Wolf [mailto:[email protected]]
Am 10.10.2018 um 15:33 hat Pavel Dovgalyuk geschrieben:

Quoted 9 lines

In record/replay mode bdrv queue is controlled by replay mechanism. It does not allow saving or loading the snapshots
when bdrv queue is not empty. Stopping the VM is not blocked by nonempty queue, but flushing the queue is still impossible there, because it may cause deadlocks in replay mode.
This patch disables bdrv_drain_all and bdrv_flush_all in record/replay mode.

Disabling bdrv_flush_all() shouldn't make a big difference in replay mode, so I agree with that.

But does the bdrv_drain_all() change mean that bdrv_drain_all_begin() can return while there are still requests in flight? This sounds like something that could break some existing code, because the meaning of the function is that if it returns success, no requests are in flight any more.

Quoted 3 lines

What would move the request forward in the replay case once it has been created? Does this also involve some user interaction? Or may we process the next event in a drain callback of blkreplay or something?

You are right - there are some operation that can't be performed at any time during the replay, because of unfinished block requests.

One of these is creating the VM snapshot. I've added a check in another patch, which prevents snapshotting while the event queue (which holds the block requests) is not empty.

There could also be other things, that will be fixed when we try to use it.

Moving replay forward until the queue is empty is not a good idea, because step-by-step debugging should be available and work without skipping the instructions.

Pavel Dovgalyuk

#25 lore raw

Re: [Qemu-devel] [PATCH v7 17/19] replay: add BH oneshot event for block layer

In reply to Kevin Wolf
Quoted 8 lines

From: Kevin Wolf [mailto:[email protected]]
Am 10.10.2018 um 15:35 hat Pavel Dovgalyuk geschrieben:

Quoted 6 lines

Replay is capable of recording normal BH events, but sometimes there are single use callbacks scheduled with aio_bh_schedule_oneshot function. This patch enables recording and replaying such callbacks. Block layer uses these events for calling the completion function. Replaying these calls makes the execution deterministic.

Quoted 6 lines
Quoted 1 line

What are the rules for using aio_bh_schedule_oneshot() vs. replay_bh_schedule_oneshot_event()? You are modifying one place, but if I grep for aio_bh_schedule_oneshot(), I find many other places that use it. Why is this one place special?

This one is special, because it caused a record/replay bug.
I can't validate all other places, because I don't understand block layer good enough.

That's why our current strategy is fixing replay, when it is breaks. It's too hard to create the test for verifying the modification. And for the current patch there was the bug which was fixed.

The rule is the following: when the event is asynchronous and its finalization affects the guest state, then this event should be processed by the record/replay queue.

Quoted 31 lines

block/block-backend.c static BlockAIOCB *blk_aio_prwv(BlockBackend *blk, int64_t offset, int

Quoted 29 lines

--

v6:
- moved stub function to the separate file for fixing linux-user build

---
 block/block-backend.c    |    3 ++-
 include/sysemu/replay.h  |    3 +++
 replay/replay-events.c   |   16 ++++++++++++++++
 replay/replay-internal.h |    1 +
 replay/replay.c          |    2 +-
 stubs/Makefile.objs      |    1 +
 stubs/replay-user.c      |    9 +++++++++
 7 files changed, 33 insertions(+), 2 deletions(-)
 create mode 100644 stubs/replay-user.c
block/block-backend.c +1 -0
index dc0cd57..04f5554 100644
@@ -17,6 +17,7 @@
 #include "block/throttle-groups.h"
 #include "sysemu/blockdev.h"
 #include "sysemu/sysemu.h"
+#include "sysemu/replay.h"
 #include "qapi/error.h"
 #include "qapi/qapi-events-block.h"
 #include "qemu/id.h"
@@ -1379,7 +1380,7 @@ static BlockAIOCB *blk_aio_prwv(BlockBackend *blk, int64_t offset, int

bytes,

Quoted 1 line
Quoted 7 lines

block/block-backend.c static BlockAIOCB *blk_aio_prwv(BlockBackend *blk, int64_t offset, int bytes, · line 1384

Quoted 5 lines
     acb->has_returned = true;
     if (acb->rwco.ret != NOT_DONE) {
block/block-backend.c +1 -1
-        aio_bh_schedule_oneshot(blk_get_aio_context(blk),
+        replay_bh_schedule_oneshot_event(blk_get_aio_context(blk),
                                 blk_aio_complete_bh, acb);

Indentation is off.

Thanks.

Pavel Dovgalyuk

#7 lore raw

Re: [Qemu-devel] [PATCH v7 04/19] replay: don't drain/flush bdrv queue while RR is working

In reply to Pavel Dovgalyuk

Am 30.11.2018 um 08:55 hat Pavel Dovgalyuk geschrieben:

Quoted 28 lines
Quoted 21 lines

From: Kevin Wolf [mailto:[email protected]]
Am 10.10.2018 um 15:33 hat Pavel Dovgalyuk geschrieben:

Quoted 9 lines

In record/replay mode bdrv queue is controlled by replay mechanism. It does not allow saving or loading the snapshots
when bdrv queue is not empty. Stopping the VM is not blocked by nonempty queue, but flushing the queue is still impossible there, because it may cause deadlocks in replay mode.
This patch disables bdrv_drain_all and bdrv_flush_all in record/replay mode.

Disabling bdrv_flush_all() shouldn't make a big difference in replay mode, so I agree with that.

But does the bdrv_drain_all() change mean that bdrv_drain_all_begin() can return while there are still requests in flight? This sounds like something that could break some existing code, because the meaning of the function is that if it returns success, no requests are in flight any more.

Quoted 3 lines

What would move the request forward in the replay case once it has been created? Does this also involve some user interaction? Or may we process the next event in a drain callback of blkreplay or something?

You are right - there are some operation that can't be performed at any time during the replay, because of unfinished block requests.

Quoted 8 lines

One of these is creating the VM snapshot. I've added a check in another patch, which prevents snapshotting while the event queue (which holds the block requests) is not empty.

There could also be other things, that will be fixed when we try to use it.

Moving replay forward until the queue is empty is not a good idea, because step-by-step debugging should be available and work without skipping the instructions.

So if the idea is that in replay mode, bdrv_drain_all_begin() is only called when there are no requests in flight anyway (because callers catch this situation earler), can we make this an assertion that the block device is already quiesced?

Skipping drain without knowing that there are no requests in flight feels simply wrong, and I'm almost sure it will result in bugs. On the other hand, if we know that no requests are in flight, there's no real point in skipping.

So I think I still don't understand the situation where skipping a drain is the correct solution.

Kevin

#8 lore raw

Re: [Qemu-devel] [PATCH v7 04/19] replay: don't drain/flush bdrv queue while RR is working

Quoted 36 lines

From: Kevin Wolf [mailto:[email protected]]
Am 30.11.2018 um 08:55 hat Pavel Dovgalyuk geschrieben:

Quoted 34 lines
Quoted 21 lines

From: Kevin Wolf [mailto:[email protected]]
Am 10.10.2018 um 15:33 hat Pavel Dovgalyuk geschrieben:

Quoted 9 lines

In record/replay mode bdrv queue is controlled by replay mechanism. It does not allow saving or loading the snapshots
when bdrv queue is not empty. Stopping the VM is not blocked by nonempty queue, but flushing the queue is still impossible there, because it may cause deadlocks in replay mode.
This patch disables bdrv_drain_all and bdrv_flush_all in record/replay mode.

Disabling bdrv_flush_all() shouldn't make a big difference in replay mode, so I agree with that.

But does the bdrv_drain_all() change mean that bdrv_drain_all_begin() can return while there are still requests in flight? This sounds like something that could break some existing code, because the meaning of the function is that if it returns success, no requests are in flight any more.

Quoted 3 lines

What would move the request forward in the replay case once it has been created? Does this also involve some user interaction? Or may we process the next event in a drain callback of blkreplay or something?

You are right - there are some operation that can't be performed at any time during the replay, because of unfinished block requests.

One of these is creating the VM snapshot. I've added a check in another patch, which prevents snapshotting while the event queue (which holds the block requests) is not empty.

There could also be other things, that will be fixed when we try to use it.

Quoted 7 lines
Quoted 2 lines

Moving replay forward until the queue is empty is not a good idea, because step-by-step debugging should be available and work without skipping the instructions.

So if the idea is that in replay mode, bdrv_drain_all_begin() is only called when there are no requests in flight anyway (because callers catch this situation earler), can we make this an assertion that the block device is already quiesced?

Maybe it's behavior changed in the latest version?
We observed that bdrv_drain_all hangs, when there is a block request in the queue. Therefore we decided that stopping with non-empty queue is better than skipping some steps.

Quoted 4 lines

Skipping drain without knowing that there are no requests in flight feels simply wrong, and I'm almost sure it will result in bugs. On the other hand, if we know that no requests are in flight, there's no real point in skipping.

What bugs can happen?

Pavel Dovgalyuk

#26 lore raw

Re: [Qemu-devel] [PATCH v7 17/19] replay: add BH oneshot event for block layer

In reply to Pavel Dovgalyuk

Am 30.11.2018 um 09:21 hat Pavel Dovgalyuk geschrieben:

Quoted 15 lines
Quoted 8 lines

From: Kevin Wolf [mailto:[email protected]]
Am 10.10.2018 um 15:35 hat Pavel Dovgalyuk geschrieben:

Quoted 6 lines

Replay is capable of recording normal BH events, but sometimes there are single use callbacks scheduled with aio_bh_schedule_oneshot function. This patch enables recording and replaying such callbacks. Block layer uses these events for calling the completion function. Replaying these calls makes the execution deterministic.

Quoted 6 lines
Quoted 1 line

What are the rules for using aio_bh_schedule_oneshot() vs. replay_bh_schedule_oneshot_event()? You are modifying one place, but if I grep for aio_bh_schedule_oneshot(), I find many other places that use it. Why is this one place special?

Quoted 7 lines

This one is special, because it caused a record/replay bug. I can't validate all other places, because I don't understand block layer good enough.

That's why our current strategy is fixing replay, when it is breaks. It's too hard to create the test for verifying the modification. And for the current patch there was the bug which was fixed.

So nobody really understands the code, but we're just fixing symptoms whenever something crashes, without ever thinking about the design? And then the code will organically grow to maybe approximate what we wanted it to do?

Honestly, that's not the way to build reliable and maintainable software.

Quoted 3 lines

The rule is the following: when the event is asynchronous and its finalization affects the guest state, then this event should be processed by the record/replay queue.

Are there any BHs that can never affect the guest state?

Maybe you should really intercept this inside qemu_bh_schedule() and aio_bh_schedule_oneshot() to catch all instances. This would look more likely to address the root cause rather than twenty different special cases and forgetting the other hundred instances.

But why do you even need to record and replay BHs? Aren't they already fully deterministic if the code that schedules them is deterministic? Is this hinting at problems in a different part of the code, so that the caller isn't deterministic even though we expect it to be?

Kevin

#27 lore raw

Re: [Qemu-devel] [PATCH v7 17/19] replay: add BH oneshot event for block layer

Quoted 30 lines

From: Kevin Wolf [mailto:[email protected]]
Am 30.11.2018 um 09:21 hat Pavel Dovgalyuk geschrieben:

Quoted 15 lines
Quoted 8 lines

From: Kevin Wolf [mailto:[email protected]]
Am 10.10.2018 um 15:35 hat Pavel Dovgalyuk geschrieben:

Quoted 6 lines

Replay is capable of recording normal BH events, but sometimes there are single use callbacks scheduled with aio_bh_schedule_oneshot function. This patch enables recording and replaying such callbacks. Block layer uses these events for calling the completion function. Replaying these calls makes the execution deterministic.

Quoted 6 lines
Quoted 1 line

What are the rules for using aio_bh_schedule_oneshot() vs. replay_bh_schedule_oneshot_event()? You are modifying one place, but if I grep for aio_bh_schedule_oneshot(), I find many other places that use it. Why is this one place special?

Quoted 7 lines

This one is special, because it caused a record/replay bug. I can't validate all other places, because I don't understand block layer good enough.

That's why our current strategy is fixing replay, when it is breaks. It's too hard to create the test for verifying the modification. And for the current patch there was the bug which was fixed.

So nobody really understands the code, but we're just fixing symptoms whenever something crashes, without ever thinking about the design? And then the code will organically grow to maybe approximate what we wanted it to do?

Quoted 8 lines

Honestly, that's not the way to build reliable and maintainable software.

Quoted 3 lines

The rule is the following: when the event is asynchronous and its finalization affects the guest state, then this event should be processed by the record/replay queue.

Are there any BHs that can never affect the guest state?

For example, qemu_bh_schedule(qmp_dispatcher_bh) in handle_qmp_command. thread-pool.c also uses BH callbacks, but they are for creating threads, and not for modifying the replayed guest state.

Quoted 9 lines

Maybe you should really intercept this inside qemu_bh_schedule() and aio_bh_schedule_oneshot() to catch all instances. This would look more likely to address the root cause rather than twenty different special cases and forgetting the other hundred instances.

But why do you even need to record and replay BHs? Aren't they already fully deterministic if the code that schedules them is deterministic? Is this hinting at problems in a different part of the code, so that the caller isn't deterministic even though we expect it to be?

BHs may be invoked at random moments, that are not related to the executed instructions. As we replay CPU, memory, and peripheral device state, we must replay all invocations all BHs that affect them.

Replay must be done precisely, with instruction-level granularity. That is why we record the moment of BH invocation for matching it in the replay phase.

Pavel Dovgalyuk