Linux-Trace-Devel Archive on lore.kernel.org
 help / color / Atom feed
* [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session
@ 2021-03-15  6:18 Tzvetomir Stoyanov (VMware)
  2021-03-15  6:18 ` [PATCH v32 1/5] trace-cmd: Add dummy function to initialize timestamp sync logic Tzvetomir Stoyanov (VMware)
                   ` (5 more replies)
  0 siblings, 6 replies; 15+ messages in thread
From: Tzvetomir Stoyanov (VMware) @ 2021-03-15  6:18 UTC (permalink / raw)
  To: rostedt; +Cc: linux-trace-devel

Basic infrastructure for host - guest timestamp synchronization and a
PoC implementation of PTP-like and KVM algorithms. 

The KVM plugin is with higher weight than PTP, it is preferred in case both
are active. KVM is used only if x86-tsc clock is selected for the current
trace, PTP is used for all other cases.

These changes depend on:
   "Refactoring and improvements of time sync logic" patch set.
   "TSC trace clock to nanosecond conversion" patch set.

 v32 changes:
  - Reduced the patch set by moving the logic for TSC to nanosecond
    conversion in a separate set.
  - Moved the logic for internal refactoring of existing and already merged
    time sync logic in a separate set.
 v31 changes:
  - Removed POC label from the patch introducing the PTP time sync
    algorithm.
  - Implemented logic for TSC clock to nanoseconds conversion.
  - Select x86-tsc clock for the tracing session, is KVM timesync plugin is
    available and the user did not specify any trace clock.
 v30 changes:
  - Moved all time sync logic in the trace-cmd library.
  - Separate the logic for pthread barrier time sync threads synchronization
    in its own patch.
 v29 changes:
  - Included <linux/types.h> in trace-timesync-kvm.c for __s64 type.
  - Added pthread barrier to host time sync pthread, to ensure times are
    synchronized before starting the trace.
  - Removed already merged patches from the set.
 v28 changes:
  - Fix bugs, introduced by previous trace-cmd commits, related to tracefs
    library usage.
  - Removed unused type defines for trace packets.
 v27 changes:
  - Fixed broken compilation due to usage of old libtracefs APIs.
  - Rewrote complex macros to follow the Linux kernel requirements for macros.
  - Removed pathces from the set, that were already merged upstream.
 v26 changes:
  - Rebased on top of latest master.
  - Abstracted Time sync protocol names in a structure, to avoid using
    triple char pointers.
  - Removed a patch from the set, as it is already merged:
    "trace-cmd: Add new libtrasefs API to get the current trace clock".
  - Minor coding style changes, for better code readability.
 v25 changes:
  - Addressed Steven's comments.
  - Reorded the patches in the plugin - put the PTP and KVM PoC patches
    last and all other fixes and improvements before them.
  - Implemented a new libtrasefs API to get the current trace clock.
  - Replaced IDs of trace timestamps synchronization protocols with strings.
  - Implemented an "accuracy" parameter for timestamps synchronization
    protocols.
  - Added "scaling ratio" correction, calculated by the timestamp
    synchronization protocols.
  - Added flags, declared by the timestamps synchronization protocols,
    and stored in the trace.dat file, together with the calculated
    time offsets.
  - Implemented timestamp offset calculation per virtual CPU.
 v24 changes:
  - trace-cmd internal refactoring, needed for KVM timestamp synchronization
    plugin.
  - Added KVM timestamp synchronization plugin.

 v23 changes:
  - Added flags to PTP start message to control the behaviour of the
    algorithm. This is for development purposes mostly, to compare and
    evaluate how these changes affect the accuracy.
	PTP_FLAG_FASTEST_RESPONSE:  Consider only the probe with fastest
				    response time, otherwise make a histogram
				    from all probes.
	PTP_FLAG_USE_MARKER: Use trace marker to get the clock,
 			     otherwise use directly the system clock

 v22 changes, addressed Steven's comments:
  - Fixed error flow in read_qemu_guests_pids(), free allocated resources
    in case of an error.
  - Fixed initialisation of reallocated memmory in set_vcpu_pid_mapping().
  - Fixed bitmask logic in tracecmd_tsync_proto_select().
  - Fixed usage of wrong structure in make_trace_resp().

 v21 changes:
  - Rebased on top of latest master.
  - Remove these two patches from the set, as they are already merged:
       trace-cmd: Add new library API for local CPU count
       trace-cmd: Add support for negative time offsets in trace.dat file
  - Added more descriptive warning message when fails to extract Time Shift
    information from the trace.dat file.
  - Added a warning message when fails to obtain vcpu - pid mappings.
  - Handle the case with sparse VCPU numbers in VCPU - PID mapping
    array.
  - Fixed broken recording over FIFOs.
 v20 changes:
  - Rebased on top of latest master.
  - Removed the VCPUS_MAX hardcoded limit, reimplemented the cpu_pid[] array
    to be dynamically allocated.
  - Added a warning when reading of TRACECMD_OPTION_TIME_SHIFT option
    fails, due to unexpected option's size.
  - Improved loading of TRACECMD_OPTION_GUEST option data, as Steven
    suggested.
 v19 changes:
  - Rebased on top of latest master. The librtacefs is already merged,
    this allows to drop few patches from the set and use librtacefs APIs
    instead.
  - Reimplemented all new trace.dat options to be in binary format,
    instead of text. Leverage the new "trace-cmd dump" sub command to
    track what is written in the file.
  - Addressed Steven's comments.

 v18 changes: addressed Steven Rostdet comments:
  - Replaced semaphores with pthread mutexes.
  - Made bitmask with time sync protocols unlimited, so we can support
    more than 32 protocols. That required to redesign the trace request
    packet format.
  - A lot of small fixes.
	
 v17 changes:
  - Implemented new PTP logic for calculating the clocks offset, using
    histogram of all PTP samples. It gives better results than the logic
    with the fastest response time, so set the new one as default in the
    POC patch.

 v16 changes:
  - Fixed compilation in case no VSOCK is available (Thanks to Slavomir Kaslev)
  - Fixed a typo in trace-cmd-record.1.txt (Thanks to Slavomir Kaslev)
  - Added forgotten file in the patch "trace-cmd: Add new library APIs for
    ftrace instances." - trace-instance.c
  - Fixed few compilation warnings related to TSYNC_DEBUG code.
  - Removed a blank line at the end of "tsync_readme" file. 

 v15 changes:
  - Removed the patch for "--proc-map" from the series, as it should not
    be part of it.

 v14 changes:
  - Bring back the PTP-like algorithm and removed the ftrace event based logic.
  - Reimplemented the PTP-like algorithm to use raw ftrace markers, instead of clock_gettime() API.
  - Refactored the logic to be algorithm independent and plugin friendly.
  - Implemented continuous timestamps synchronization, while the trace is running.
  - Moved logic from trace-cmd application to libtracecmd, as new library APIs.
  - Implemented new trace id functionality.
  - Implemented new guest section in host trace.dat file.
	
 v13 changes:
  - Remove few patches from the set, as they were merged.
  - Rebased to the latest master, Slavomirs patchest "Add VM kernel tracing over
    vsockets and FIFOs" got merged!
 	
 v12 changes:
  - Rebased on top of Slavomir's v13 "Add VM kernel tracing over vsockets and FIFOs"

 v11 changes:
  - Rebased on top of Slavomir's v10 "Add VM kernel tracing over vsockets and FIFOs"
  - Addressed Slavomir's commnents from version 10 of the patch series.

 v10 changes:
  - Fixed broken compilation, call to timestamp_correction_calc() in timestamp_correct
    was smashed.
  - Replaced deprecated tep_data_event_from_type() API with tep_find_event().
  - Fixed a warning on assignment const to non const.

 v9 changes:
  - Fixed implementation of binary search algorithm in timestamp_correct()

 v8 changes:
  - Added rmdir() call in tracecmd_remove_instance(), to completely remove the instance. 
  However, there is an issue with deleting the instances using rmdir(), which is investigated.
  - Few changes in read_qemu_guests_pids(), timestamp_correct(), tsync_offset_load() 
 tracecmd_clock_context_new() and find_raw_events() suggested by Slavomir. 

 v7 changes:
  - Added warning messages in case time synchronization cannot be negotiated or fails.
  - Few optimizations and checks in read_qemu_guests_pids(), tsync_offset_load(),
    and find_raw_events(), suggested by Slavomir Kaslev.
  - Reworked timestamp_correct() to not use static variables.
  - Check TRACECMD_OPTION_TIME_SHIFT before reading time sync samples from the trace.dat file

 v6 changes:
  - Refactored tracecmd_msg_snd_time_sync() and tracecmd_msg_rcv_time_sync() functions:
    removed any time sync calculations logic as separate functions in trace-timesync.c file
  - Defined TSYNC_PROBE, TSYNC_REQ and TSYNC_RESP messages, in order to make the time sync
    protocol comprehensible.
  - Addressed Steven Rostedt comments.
  - Addressed Slavomir Kaslev commnets.

v5 changes:
  - Rebased to Slavomir's v8 "Add VM kernel tracing over vsockets and FIFOs"
    patch series.
  - Implemented an algorithm for time drift correction.
  - Addressed Slavomir's commnets.
  - Refactored the code: moved all time sync specific implementation in trace-timesync.c
  - Isolated all hardcoded event specific stuff in a structure, so it could be easily
    moved to external plugins.
  - Added a check for VSOCK support: do not perform vsock dependent time synchronisation
    in case there is no VSOCK support.

 v4 changes:
  - Removed the implementation of PTP-like algorithm. The current
    logic relies on matching time stamps of kvm_exit/virtio_transport_recv_pkt
    events on host to virtio_transport_alloc_pkt/vp_notify events on guest.
  - Rebased to Slavomir's v7 "Add VM kernel tracing over vsockets and FIFOs"
    patch series.
  - Decreased the time synch probes from 5000 to 300.
  - Addressed Steven Rostedt comments.
  - Code cleanup.

 v3 changes:
 - Removed any magic constants, used in the PTP-like algorithm,
   as Slavomir Kaslev suggested.
 - Implemented new algorithm, based on mapping kvm_exit events
   in host context to vsock_send events in guest context,
   suggested by Steven Rostedt.

 v2 changes:
  - Addressed Steven Rostedt comments.
  - Modified PTP-like timestamps sync algorithm to gain more accuracy, with the
    help of Yordan Karadzhov and Slavomir Kaslev.
Tzvetomir Stoyanov (VMware) (5):
  trace-cmd: Add dummy function to initialize timestamp sync logic
  trace-cmd: Add timestamp synchronization per vCPU
  trace-cmd: PTP-like algorithm for host - guest timestamp
    synchronization
  trace-cmd: Debug scripts for PTP-like algorithm for host - guest
    timestamp synchronization
  trace-cmd [POC]: Add KVM timestamp synchronization plugin

 lib/trace-cmd/Makefile                        |   4 +
 .../include/private/trace-cmd-private.h       |   2 +-
 lib/trace-cmd/include/trace-tsync-local.h     |  23 +-
 lib/trace-cmd/trace-input.c                   | 175 +++--
 lib/trace-cmd/trace-timesync-kvm.c            | 460 +++++++++++
 lib/trace-cmd/trace-timesync-ptp.c            | 715 ++++++++++++++++++
 lib/trace-cmd/trace-timesync.c                | 294 +++++--
 scripts/debug/tsync_hist.py                   |  57 ++
 scripts/debug/tsync_readme                    |  12 +
 scripts/debug/tsync_res.py                    |  46 ++
 tracecmd/trace-agent.c                        |   2 +
 tracecmd/trace-dump.c                         |  52 +-
 tracecmd/trace-record.c                       |  23 +-
 13 files changed, 1685 insertions(+), 180 deletions(-)
 create mode 100644 lib/trace-cmd/trace-timesync-kvm.c
 create mode 100644 lib/trace-cmd/trace-timesync-ptp.c
 create mode 100644 scripts/debug/tsync_hist.py
 create mode 100644 scripts/debug/tsync_readme
 create mode 100644 scripts/debug/tsync_res.py

-- 
2.29.2


^ permalink raw reply	[flat|nested] 15+ messages in thread

* [PATCH v32 1/5] trace-cmd: Add dummy function to initialize timestamp sync logic
  2021-03-15  6:18 [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Tzvetomir Stoyanov (VMware)
@ 2021-03-15  6:18 ` Tzvetomir Stoyanov (VMware)
  2021-03-15  6:18 ` [PATCH v32 2/5] trace-cmd: Add timestamp synchronization per vCPU Tzvetomir Stoyanov (VMware)
                   ` (4 subsequent siblings)
  5 siblings, 0 replies; 15+ messages in thread
From: Tzvetomir Stoyanov (VMware) @ 2021-03-15  6:18 UTC (permalink / raw)
  To: rostedt; +Cc: linux-trace-devel

A dummy empty function is added, will be used to initialize timestamp
synchronization logic:
  tracecmd_tsync_init()
When this code is implemented as real plugins, this function will be
removed. Then the initializion will be triggered at plugin load time.

Signed-off-by: Tzvetomir Stoyanov (VMware) <tz.stoyanov@gmail.com>
---
 lib/trace-cmd/trace-timesync.c |  8 ++++++++
 tracecmd/trace-agent.c         |  2 ++
 tracecmd/trace-record.c        | 23 +++++++++++++++--------
 3 files changed, 25 insertions(+), 8 deletions(-)

diff --git a/lib/trace-cmd/trace-timesync.c b/lib/trace-cmd/trace-timesync.c
index 85a57f23..483ebd96 100644
--- a/lib/trace-cmd/trace-timesync.c
+++ b/lib/trace-cmd/trace-timesync.c
@@ -55,6 +55,14 @@ static struct tsync_proto *tsync_proto_find(const char *proto_name)
 	return NULL;
 }
 
+/**
+ * tracecmd_tsync_init - Initialize the global, per task, time sync data.
+ */
+void tracecmd_tsync_init(void)
+{
+
+}
+
 int tracecmd_tsync_proto_register(const char *proto_name, int accuracy, int roles,
 				  int supported_clocks, unsigned int flags,
 				  int (*init)(struct tracecmd_time_sync *),
diff --git a/tracecmd/trace-agent.c b/tracecmd/trace-agent.c
index 49f6dd8c..e6fafd2b 100644
--- a/tracecmd/trace-agent.c
+++ b/tracecmd/trace-agent.c
@@ -251,6 +251,8 @@ static void agent_serve(unsigned int port)
 	if (sd < 0)
 		die("Failed to open vsocket");
 
+	tracecmd_tsync_init();
+
 	if (!get_local_cid(&cid))
 		printf("listening on @%u:%u\n", cid, port);
 
diff --git a/tracecmd/trace-record.c b/tracecmd/trace-record.c
index 30d3ed89..e91c2cad 100644
--- a/tracecmd/trace-record.c
+++ b/tracecmd/trace-record.c
@@ -6376,10 +6376,6 @@ static bool has_local_instances(void)
 	return false;
 }
 
-/*
- * This function contains common code for the following commands:
- * record, start, stream, profile.
- */
 static void record_trace(int argc, char **argv,
 			 struct common_record_context *ctx)
 {
@@ -6540,12 +6536,23 @@ static void record_trace(int argc, char **argv,
 	finalize_record_trace(ctx);
 }
 
+/*
+ * This function contains common code for the following commands:
+ * record, start, stream, profile.
+ */
+static void record_trace_command(int argc, char **argv,
+				 struct common_record_context *ctx)
+{
+	tracecmd_tsync_init();
+	record_trace(argc, argv, ctx);
+}
+
 void trace_start(int argc, char **argv)
 {
 	struct common_record_context ctx;
 
 	parse_record_options(argc, argv, CMD_start, &ctx);
-	record_trace(argc, argv, &ctx);
+	record_trace_command(argc, argv, &ctx);
 	exit(0);
 }
 
@@ -6637,7 +6644,7 @@ void trace_stream(int argc, char **argv)
 	struct common_record_context ctx;
 
 	parse_record_options(argc, argv, CMD_stream, &ctx);
-	record_trace(argc, argv, &ctx);
+	record_trace_command(argc, argv, &ctx);
 	exit(0);
 }
 
@@ -6656,7 +6663,7 @@ void trace_profile(int argc, char **argv)
 	if (!buffer_instances)
 		top_instance.flags |= BUFFER_FL_PROFILE;
 
-	record_trace(argc, argv, &ctx);
+	record_trace_command(argc, argv, &ctx);
 	do_trace_profile();
 	exit(0);
 }
@@ -6666,7 +6673,7 @@ void trace_record(int argc, char **argv)
 	struct common_record_context ctx;
 
 	parse_record_options(argc, argv, CMD_record, &ctx);
-	record_trace(argc, argv, &ctx);
+	record_trace_command(argc, argv, &ctx);
 	exit(0);
 }
 
-- 
2.29.2


^ permalink raw reply	[flat|nested] 15+ messages in thread

* [PATCH v32 2/5] trace-cmd: Add timestamp synchronization per vCPU
  2021-03-15  6:18 [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Tzvetomir Stoyanov (VMware)
  2021-03-15  6:18 ` [PATCH v32 1/5] trace-cmd: Add dummy function to initialize timestamp sync logic Tzvetomir Stoyanov (VMware)
@ 2021-03-15  6:18 ` Tzvetomir Stoyanov (VMware)
  2021-03-15  6:18 ` [PATCH v32 3/5] trace-cmd: PTP-like algorithm for host - guest timestamp synchronization Tzvetomir Stoyanov (VMware)
                   ` (3 subsequent siblings)
  5 siblings, 0 replies; 15+ messages in thread
From: Tzvetomir Stoyanov (VMware) @ 2021-03-15  6:18 UTC (permalink / raw)
  To: rostedt; +Cc: linux-trace-devel

Timestamp synchronization logic is changed to work per virtual CPU. Some
hypervisors maintain time offset and scaling per vCPU. The host-guest
communication protocol is changed to request time stamp offset calculation
for particular vCPU. The guest thread, responsible for running that logic
is pinned to the requested CPU. The time sync medata data, saved in the
trace.dat file is changed to an array of vCPUs. When an event time stamp
is corrected, the CPU on that the event happened is used to get the
correct offset.

Signed-off-by: Tzvetomir Stoyanov (VMware) <tz.stoyanov@gmail.com>
---
 .../include/private/trace-cmd-private.h       |   2 +-
 lib/trace-cmd/include/trace-tsync-local.h     |  22 +-
 lib/trace-cmd/trace-input.c                   | 175 +++++++----
 lib/trace-cmd/trace-timesync.c                | 283 +++++++++++++-----
 tracecmd/trace-dump.c                         |  52 ++--
 5 files changed, 362 insertions(+), 172 deletions(-)

diff --git a/lib/trace-cmd/include/private/trace-cmd-private.h b/lib/trace-cmd/include/private/trace-cmd-private.h
index 98616c76..fa024d10 100644
--- a/lib/trace-cmd/include/private/trace-cmd-private.h
+++ b/lib/trace-cmd/include/private/trace-cmd-private.h
@@ -453,7 +453,7 @@ tracecmd_tsync_with_guest(unsigned long long trace_id, int loop_interval,
 			  unsigned int cid, unsigned int port, int guest_pid,
 			  int guest_cpus, const char *proto_name, const char *clock);
 int tracecmd_tsync_with_guest_stop(struct tracecmd_time_sync *tsync);
-int tracecmd_tsync_get_offsets(struct tracecmd_time_sync *tsync,
+int tracecmd_tsync_get_offsets(struct tracecmd_time_sync *tsync, int cpu,
 				int *count, long long **ts,
 				long long **offsets, long long **scalings);
 int tracecmd_tsync_get_session_params(struct tracecmd_time_sync *tsync,
diff --git a/lib/trace-cmd/include/trace-tsync-local.h b/lib/trace-cmd/include/trace-tsync-local.h
index 5cf4f843..b3a3bdfc 100644
--- a/lib/trace-cmd/include/trace-tsync-local.h
+++ b/lib/trace-cmd/include/trace-tsync-local.h
@@ -24,12 +24,7 @@ struct tracecmd_time_sync {
 	int				vcpu_count;
 };
 
-struct clock_sync_context {
-	void				*proto_data;	/* time sync protocol specific data */
-	bool				is_server;	/* server side time sync role */
-	bool				is_guest;	/* guest or host time sync role */
-	struct tracefs_instance		*instance;	/* ftrace buffer, used for time sync events */
-
+struct clock_sync_offsets {
 	/* Arrays with calculated time offsets at given time */
 	int				sync_size;	/* Allocated size of sync_ts,
 							 * sync_offsets and sync_scalings
@@ -40,6 +35,18 @@ struct clock_sync_context {
 	long long			*sync_ts;
 	long long			*sync_offsets;
 	long long			*sync_scalings;
+};
+
+struct clock_sync_context {
+	void				*proto_data;	/* time sync protocol specific data */
+	bool				is_server;	/* server side time sync role */
+	bool				is_guest;	/* guest or host time sync role */
+	struct tracefs_instance		*instance;	/* ftrace buffer, used for time sync events */
+
+	int				cpu_count;
+	struct clock_sync_offsets	*offsets;	/* Array of size cpu_count
+							 * calculated offsets per CPU
+							 */
 
 	/* Identifiers of local and remote time sync peers: cid and port */
 	unsigned int			local_cid;
@@ -53,7 +60,8 @@ int tracecmd_tsync_proto_register(const char *proto_name, int accuracy, int role
 				  int (*init)(struct tracecmd_time_sync *),
 				  int (*free)(struct tracecmd_time_sync *),
 				  int (*calc)(struct tracecmd_time_sync *,
-					      long long *, long long *, long long *));
+					      long long *, long long *,
+					      long long *, unsigned int));
 int tracecmd_tsync_proto_unregister(char *proto_name);
 
 int ptp_clock_sync_register(void);
diff --git a/lib/trace-cmd/trace-input.c b/lib/trace-cmd/trace-input.c
index b53c1932..9969fc7a 100644
--- a/lib/trace-cmd/trace-input.c
+++ b/lib/trace-cmd/trace-input.c
@@ -91,12 +91,19 @@ struct guest_trace_info {
 	int			*cpu_pid;
 };
 
+struct timesync_offsets {
+	int	ts_samples_count;
+	struct ts_offset_sample	*ts_samples;
+};
+
 struct host_trace_info {
 	unsigned long long	peer_trace_id;
 	unsigned int		flags;
 	bool			sync_enable;
 	int			ts_samples_count;
 	struct ts_offset_sample	*ts_samples;
+	int			cpu_count;
+	struct timesync_offsets	*ts_offsets;
 };
 
 struct tsc2nsec {
@@ -1224,47 +1231,51 @@ timestamp_correction_calc(unsigned long long ts, unsigned int flags,
 	return ts + tscor;
 }
 
-static unsigned long long timestamp_host_sync(unsigned long long ts,
+static unsigned long long timestamp_host_sync(unsigned long long ts, int cpu,
 					      struct tracecmd_input *handle)
 {
-	struct host_trace_info	*host = &handle->host;
+	struct timesync_offsets *tsync;
 	int min, mid, max;
 
+	if (cpu >= handle->host.cpu_count)
+		return ts;
+	tsync = &handle->host.ts_offsets[cpu];
+
 	/* We have one sample, nothing to calc here */
-	if (host->ts_samples_count == 1)
-		return ts + host->ts_samples[0].offset;
+	if (tsync->ts_samples_count == 1)
+		return ts + tsync->ts_samples[0].offset;
 
 	/* We have two samples, nothing to search here */
-	if (host->ts_samples_count == 2)
-		return timestamp_correction_calc(ts, host->flags,
-						 &host->ts_samples[0],
-						 &host->ts_samples[1]);
+	if (tsync->ts_samples_count == 2)
+		return timestamp_correction_calc(ts, handle->host.flags,
+						 &tsync->ts_samples[0],
+						 &tsync->ts_samples[1]);
 
 	/* We have more than two samples */
-	if (ts <= host->ts_samples[0].time)
-		return timestamp_correction_calc(ts, host->flags,
-						 &host->ts_samples[0],
-						 &host->ts_samples[1]);
-	else if (ts >= host->ts_samples[host->ts_samples_count-1].time)
-		return timestamp_correction_calc(ts, host->flags,
-						 &host->ts_samples[host->ts_samples_count-2],
-						 &host->ts_samples[host->ts_samples_count-1]);
+	if (ts <= tsync->ts_samples[0].time)
+		return timestamp_correction_calc(ts, handle->host.flags,
+						 &tsync->ts_samples[0],
+						 &tsync->ts_samples[1]);
+	else if (ts >= tsync->ts_samples[tsync->ts_samples_count-1].time)
+		return timestamp_correction_calc(ts, handle->host.flags,
+						 &tsync->ts_samples[tsync->ts_samples_count-2],
+						 &tsync->ts_samples[tsync->ts_samples_count-1]);
 	min = 0;
-	max = host->ts_samples_count-1;
+	max = tsync->ts_samples_count-1;
 	mid = (min + max)/2;
 	while (min <= max) {
-		if (ts < host->ts_samples[mid].time)
+		if (ts < tsync->ts_samples[mid].time)
 			max = mid - 1;
-		else if (ts > host->ts_samples[mid].time)
+		else if (ts > tsync->ts_samples[mid].time)
 			min = mid + 1;
 		else
 			break;
 		mid = (min + max)/2;
 	}
 
-	return timestamp_correction_calc(ts, host->flags,
-					 &host->ts_samples[mid],
-					 &host->ts_samples[mid+1]);
+	return timestamp_correction_calc(ts, handle->host.flags,
+					 &tsync->ts_samples[mid],
+					 &tsync->ts_samples[mid+1]);
 }
 
 static unsigned long long mul_u64_u32_shr(unsigned long long a,
@@ -1283,7 +1294,7 @@ static unsigned long long mul_u64_u32_shr(unsigned long long a,
 	return ret;
 }
 
-static unsigned long long timestamp_calc(unsigned long long ts,
+static unsigned long long timestamp_calc(unsigned long long ts, int cpu,
 					 struct tracecmd_input *handle)
 {
 	unsigned long long t = ts;
@@ -1294,7 +1305,7 @@ static unsigned long long timestamp_calc(unsigned long long ts,
 
 	/* Guest trace file, sync with host timestamps */
 	if (handle->host.sync_enable)
-		t = timestamp_host_sync(ts, handle);
+		t = timestamp_host_sync(ts, cpu, handle);
 
 	if (handle->ts2secs) {
 		/* user specified clock frequency */
@@ -1334,7 +1345,7 @@ static int update_page_info(struct tracecmd_input *handle, int cpu)
 		return -1;
 	}
 	handle->cpu_data[cpu].timestamp = timestamp_calc(kbuffer_timestamp(kbuf),
-							 handle);
+							 cpu, handle);
 
 	return 0;
 }
@@ -1964,7 +1975,7 @@ read_again:
 		goto read_again;
 	}
 
-	handle->cpu_data[cpu].timestamp = timestamp_calc(ts, handle);
+	handle->cpu_data[cpu].timestamp = timestamp_calc(ts, cpu, handle);
 
 	index = kbuffer_curr_offset(kbuf);
 
@@ -2294,34 +2305,79 @@ static int tsync_offset_cmp(const void *a, const void *b)
 	return 0;
 }
 
-static void tsync_offset_load(struct tracecmd_input *handle, char *buf)
+#define safe_read(R, C)					\
+	do {						\
+		if ((C) > size)				\
+			return -EFAULT;			\
+		(R) = tep_read_number(tep, buf, (C));	\
+		buf += (C);				\
+		size -= (C);				\
+	} while (0)
+
+#define safe_read_loop(type)						\
+	do {								\
+		int i;							\
+		for (i = 0; i < ts_offsets->ts_samples_count; i++)	\
+			safe_read(ts_offsets->ts_samples[i].type, 8);	\
+	} while (0)
+
+static int tsync_offset_load(struct tep_handle	*tep,
+			     struct timesync_offsets *ts_offsets, char *buf, int size)
 {
-	struct host_trace_info *host = &handle->host;
-	long long *buf8 = (long long *)buf;
+	int start_size = size;
 	int i, j;
 
-	for (i = 0; i < host->ts_samples_count; i++) {
-		host->ts_samples[i].time = tep_read_number(handle->pevent,
-							   buf8 + i, 8);
-		host->ts_samples[i].offset = tep_read_number(handle->pevent,
-						buf8 + host->ts_samples_count + i, 8);
-		host->ts_samples[i].scaling = tep_read_number(handle->pevent,
-						buf8 + (2 * host->ts_samples_count) + i, 8);
-	}
-	qsort(host->ts_samples, host->ts_samples_count,
+	safe_read_loop(time);
+	safe_read_loop(offset);
+	safe_read_loop(scaling);
+	qsort(ts_offsets->ts_samples, ts_offsets->ts_samples_count,
 	      sizeof(struct ts_offset_sample), tsync_offset_cmp);
 	/* Filter possible samples with equal time */
-	for (i = 0, j = 0; i < host->ts_samples_count; i++) {
-		if (i == 0 || host->ts_samples[i].time != host->ts_samples[i-1].time)
-			host->ts_samples[j++] = host->ts_samples[i];
+	for (i = 0, j = 0; i < ts_offsets->ts_samples_count; i++) {
+		if (i == 0 || ts_offsets->ts_samples[i].time != ts_offsets->ts_samples[i-1].time)
+			ts_offsets->ts_samples[j++] = ts_offsets->ts_samples[i];
 	}
-	host->ts_samples_count = j;
+	ts_offsets->ts_samples_count = j;
+
+	return start_size - size;
+}
+
+static int tsync_cpu_offsets_load(struct tracecmd_input *handle, char *buf, int size)
+{
+	struct tep_handle *tep = handle->pevent;
+	int ret;
+	int i;
+
+	safe_read(handle->host.cpu_count, 4);
+	handle->host.ts_offsets = calloc(handle->host.cpu_count,
+					 sizeof(struct timesync_offsets));
+	if (!handle->host.ts_offsets)
+		return -ENOMEM;
+	for (i = 0; i < handle->host.cpu_count; i++) {
+		safe_read(handle->host.ts_offsets[i].ts_samples_count, 4);
+		handle->host.ts_offsets[i].ts_samples = calloc(handle->host.ts_offsets[i].ts_samples_count,
+							       sizeof(struct ts_offset_sample));
+		if (!handle->host.ts_offsets[i].ts_samples)
+			return -ENOMEM;
+		ret = tsync_offset_load(tep, &handle->host.ts_offsets[i], buf, size);
+		if (ret <= 0)
+			return -EFAULT;
+		size -= ret;
+		buf += ret;
+	}
+	return 0;
 }
 
 static void trace_tsync_offset_free(struct host_trace_info *host)
 {
-	free(host->ts_samples);
-	host->ts_samples = NULL;
+	int i;
+
+	if (host->ts_offsets) {
+		for (i = 0; i < host->cpu_count; i++)
+			free(host->ts_offsets[i].ts_samples);
+		free(host->ts_offsets);
+		host->ts_offsets = NULL;
+	}
 }
 
 static int trace_pid_map_cmp(const void *a, const void *b)
@@ -2575,8 +2631,8 @@ static int handle_options(struct tracecmd_input *handle)
 	struct input_buffer_instance *buffer;
 	struct hook_list *hook;
 	char *buf;
-	int samples_size;
 	int cpus;
+	int ret;
 
 	/* By default, use usecs, unless told otherwise */
 	handle->flags |= TRACECMD_FL_IN_USECS;
@@ -2628,11 +2684,15 @@ static int handle_options(struct tracecmd_input *handle)
 			/*
 			 * long long int (8 bytes) trace session ID
 			 * int (4 bytes) protocol flags.
-			 * int (4 bytes) count of timestamp offsets.
-			 * long long array of size [count] of times,
+			 * int (4 bytes) CPU count.
+			 * array of size [CPU count]:
+			 * [
+			 *  int (4 bytes) count of timestamp offsets.
+			 *  long long array of size [count] of times,
 			 *      when the offsets were calculated.
-			 * long long array of size [count] of timestamp offsets.
-			 * long long array of size [count] of timestamp scaling ratios.*
+			 *  long long array of size [count] of timestamp offsets.
+			 *  long long array of size [count] of timestamp scaling ratios.*
+			 * ]
 			 */
 			if (size < 16 || (handle->flags & TRACECMD_FL_RAW_TS))
 				break;
@@ -2640,18 +2700,9 @@ static int handle_options(struct tracecmd_input *handle)
 								     buf, 8);
 			handle->host.flags = tep_read_number(handle->pevent,
 							     buf + 8, 4);
-			handle->host.ts_samples_count = tep_read_number(handle->pevent,
-									buf + 12, 4);
-			samples_size = (8 * handle->host.ts_samples_count);
-			if (size != (16 + (2 * samples_size))) {
-				warning("Failed to extract Time Shift information from the file: found size %d, expected is %d",
-					size, 16 + (2 * samples_size));
-				break;
-			}
-			handle->host.ts_samples = malloc(2 * samples_size);
-			if (!handle->host.ts_samples)
-				return -ENOMEM;
-			tsync_offset_load(handle, buf + 16);
+			ret = tsync_cpu_offsets_load(handle, buf + 12, size - 12);
+			if (ret < 0)
+				return ret;
 			tracecmd_enable_tsync(handle, true);
 			break;
 		case TRACECMD_OPTION_CPUSTAT:
@@ -4054,7 +4105,7 @@ int tracecmd_get_guest_cpumap(struct tracecmd_input *handle,
 int tracecmd_enable_tsync(struct tracecmd_input *handle, bool enable)
 {
 	if (enable &&
-	    (!handle->host.ts_samples || !handle->host.ts_samples_count))
+	    (!handle->host.ts_offsets || !handle->host.cpu_count))
 		return -1;
 
 	handle->host.sync_enable = enable;
diff --git a/lib/trace-cmd/trace-timesync.c b/lib/trace-cmd/trace-timesync.c
index 483ebd96..b0ea8f6d 100644
--- a/lib/trace-cmd/trace-timesync.c
+++ b/lib/trace-cmd/trace-timesync.c
@@ -36,9 +36,13 @@ struct tsync_proto {
 	int (*clock_sync_free)(struct tracecmd_time_sync *clock_context);
 	int (*clock_sync_calc)(struct tracecmd_time_sync *clock_context,
 			       long long *offset, long long *scaling,
-			       long long *timestamp);
+			       long long *timestamp, unsigned int cpu);
 };
 
+struct tsync_probe_request_msg {
+	unsigned short	cpu;
+} __packed;
+
 static struct tsync_proto *tsync_proto_list;
 
 static struct tsync_proto *tsync_proto_find(const char *proto_name)
@@ -68,7 +72,8 @@ int tracecmd_tsync_proto_register(const char *proto_name, int accuracy, int role
 				  int (*init)(struct tracecmd_time_sync *),
 				  int (*free)(struct tracecmd_time_sync *),
 				  int (*calc)(struct tracecmd_time_sync *,
-					      long long *, long long *, long long *))
+					      long long *, long long *,
+					      long long *, unsigned int))
 {
 	struct tsync_proto *proto = NULL;
 
@@ -80,6 +85,7 @@ int tracecmd_tsync_proto_register(const char *proto_name, int accuracy, int role
 	strncpy(proto->proto_name, proto_name, TRACECMD_TSYNC_PNAME_LENGTH);
 	proto->accuracy = accuracy;
 	proto->roles = roles;
+	proto->flags = flags;
 	proto->supported_clocks = supported_clocks;
 	proto->clock_sync_init = init;
 	proto->clock_sync_free = free;
@@ -122,6 +128,7 @@ bool tsync_proto_is_supported(const char *proto_name)
  * tracecmd_tsync_get_offsets - Return the calculated time offsets
  *
  * @tsync: Pointer to time sync context
+ * @cpu: CPU for which to get the calculated offsets
  * @count: Returns the number of calculated time offsets
  * @ts: Array of size @count containing timestamps of callculated offsets
  * @offsets: array of size @count, containing offsets for each timestamp
@@ -129,23 +136,25 @@ bool tsync_proto_is_supported(const char *proto_name)
  *
  * Retuns -1 in case of an error, or 0 otherwise
  */
-int tracecmd_tsync_get_offsets(struct tracecmd_time_sync *tsync,
-				int *count, long long **ts,
-				long long **offsets, long long **scalings)
+int tracecmd_tsync_get_offsets(struct tracecmd_time_sync *tsync, int cpu,
+			       int *count, long long **ts,
+			       long long **offsets, long long **scalings)
 {
 	struct clock_sync_context *tsync_context;
 
 	if (!tsync || !tsync->context)
 		return -1;
 	tsync_context = (struct clock_sync_context *)tsync->context;
+	if (cpu >= tsync_context->cpu_count || !tsync_context->offsets)
+		return -1;
 	if (count)
-		*count = tsync_context->sync_count;
+		*count = tsync_context->offsets[cpu].sync_count;
 	if (ts)
-		*ts = tsync_context->sync_ts;
+		*ts = tsync_context->offsets[cpu].sync_ts;
 	if (offsets)
-		*offsets = tsync_context->sync_offsets;
+		*offsets = tsync_context->offsets[cpu].sync_offsets;
 	if (scalings)
-		*scalings = tsync_context->sync_scalings;
+		*scalings = tsync_context->offsets[cpu].sync_scalings;
 
 	return 0;
 }
@@ -499,6 +508,13 @@ static int clock_context_init(struct tracecmd_time_sync *tsync,
 	if (!clock->instance)
 		goto error;
 
+	clock->cpu_count = tsync->vcpu_count;
+	if (clock->cpu_count) {
+		clock->offsets = calloc(clock->cpu_count, sizeof(struct clock_sync_offsets));
+		if (!clock->offsets)
+			goto error;
+	}
+
 	tsync->context = clock;
 	if (protocol->clock_sync_init && protocol->clock_sync_init(tsync) < 0)
 		goto error;
@@ -508,6 +524,9 @@ static int clock_context_init(struct tracecmd_time_sync *tsync,
 	return 0;
 error:
 	tsync->context = NULL;
+	if (clock->instance)
+		clock_synch_delete_instance(clock->instance);
+	free(clock->offsets);
 	free(clock);
 	return -1;
 }
@@ -523,6 +542,7 @@ void tracecmd_tsync_free(struct tracecmd_time_sync *tsync)
 {
 	struct clock_sync_context *tsync_context;
 	struct tsync_proto *proto;
+	int i;
 
 	if (!tsync->context)
 		return;
@@ -535,14 +555,20 @@ void tracecmd_tsync_free(struct tracecmd_time_sync *tsync)
 	clock_synch_delete_instance(tsync_context->instance);
 	tsync_context->instance = NULL;
 
-	free(tsync_context->sync_ts);
-	free(tsync_context->sync_offsets);
-	free(tsync_context->sync_scalings);
-	tsync_context->sync_ts = NULL;
-	tsync_context->sync_offsets = NULL;
-	tsync_context->sync_scalings = NULL;
-	tsync_context->sync_count = 0;
-	tsync_context->sync_size = 0;
+	if (tsync_context->cpu_count && tsync_context->offsets) {
+		for (i = 0; i < tsync_context->cpu_count; i++) {
+			free(tsync_context->offsets[i].sync_ts);
+			free(tsync_context->offsets[i].sync_offsets);
+			free(tsync_context->offsets[i].sync_scalings);
+			tsync_context->offsets[i].sync_ts = NULL;
+			tsync_context->offsets[i].sync_offsets = NULL;
+			tsync_context->offsets[i].sync_scalings = NULL;
+			tsync_context->offsets[i].sync_count = 0;
+			tsync_context->offsets[i].sync_size = 0;
+		}
+		free(tsync_context->offsets);
+		tsync_context->offsets = NULL;
+	}
 	pthread_mutex_destroy(&tsync->lock);
 	pthread_cond_destroy(&tsync->cond);
 	pthread_barrier_destroy(&tsync->first_sync);
@@ -551,15 +577,69 @@ void tracecmd_tsync_free(struct tracecmd_time_sync *tsync)
 	free(tsync);
 }
 
+static cpu_set_t *pin_to_cpu(int cpu)
+{
+	static size_t size;
+	static int cpus;
+	cpu_set_t *mask = NULL;
+	cpu_set_t *old = NULL;
+
+	if (!cpus) {
+		cpus = tracecmd_count_cpus();
+		size = CPU_ALLOC_SIZE(cpus);
+	}
+	if (cpu >= cpus)
+		goto error;
+
+	mask = CPU_ALLOC(cpus);
+	if (!mask)
+		goto error;
+	old = CPU_ALLOC(cpus);
+	if (!old)
+		goto error;
+
+	CPU_ZERO_S(size, mask);
+	CPU_SET_S(cpu, size, mask);
+	if (pthread_getaffinity_np(pthread_self(), size, old))
+		goto error;
+	if (pthread_setaffinity_np(pthread_self(), size, mask))
+		goto error;
+
+	CPU_FREE(mask);
+	return old;
+
+error:
+	if (mask)
+		CPU_FREE(mask);
+	if (old)
+		CPU_FREE(old);
+	return NULL;
+}
+
+static void restore_pin_to_cpu(cpu_set_t *mask)
+{
+	static size_t size;
+
+	if (!size)
+		size = CPU_ALLOC_SIZE(tracecmd_count_cpus());
+
+	pthread_setaffinity_np(pthread_self(), size, mask);
+	CPU_FREE(mask);
+}
+
 static int tsync_send(struct tracecmd_time_sync *tsync,
-		      struct tsync_proto *proto)
+		      struct tsync_proto *proto, unsigned int cpu)
 {
+	cpu_set_t *old_set = NULL;
 	long long timestamp = 0;
 	long long scaling = 0;
 	long long offset = 0;
 	int ret;
 
-	ret = proto->clock_sync_calc(tsync, &offset, &scaling, &timestamp);
+	old_set = pin_to_cpu(cpu);
+	ret = proto->clock_sync_calc(tsync, &offset, &scaling, &timestamp, cpu);
+	if (old_set)
+		restore_pin_to_cpu(old_set);
 
 	return ret;
 }
@@ -567,55 +647,48 @@ static int tsync_send(struct tracecmd_time_sync *tsync,
 static void tsync_with_host(struct tracecmd_time_sync *tsync)
 {
 	char protocol[TRACECMD_TSYNC_PNAME_LENGTH];
+	struct tsync_probe_request_msg probe;
 	struct tsync_proto *proto;
 	unsigned int command;
+	unsigned int size;
+	char *msg;
 	int ret;
 
 	clock_context_init(tsync, &proto, true);
 	if (!tsync->context)
 		return;
 
+	msg = (char *)&probe;
+	size = sizeof(probe);
 	while (true) {
+		memset(&probe, 0, size);
 		ret = tracecmd_msg_recv_time_sync(tsync->msg_handle,
 						  protocol, &command,
-						  NULL, NULL);
+						  &size, &msg);
 
 		if (ret || strncmp(protocol, TRACECMD_TSYNC_PROTO_NONE, TRACECMD_TSYNC_PNAME_LENGTH) ||
 		    command != TRACECMD_TIME_SYNC_CMD_PROBE)
 			break;
-		ret = tsync_send(tsync, proto);
+		ret = tsync_send(tsync, proto, probe.cpu);
 		if (ret)
 			break;
 	}
 }
 
-static int tsync_get_sample(struct tracecmd_time_sync *tsync,
-			    struct tsync_proto *proto, int array_step)
+static int record_sync_sample(struct clock_sync_offsets *offsets, int array_step,
+			      long long offset, long long scaling, long long ts)
 {
-	struct clock_sync_context *clock;
 	long long *sync_scalings = NULL;
 	long long *sync_offsets = NULL;
 	long long *sync_ts = NULL;
-	long long timestamp = 0;
-	long long scaling = 0;
-	long long offset = 0;
-	int ret;
 
-	ret = proto->clock_sync_calc(tsync, &offset, &scaling, &timestamp);
-	if (ret) {
-		warning("Failed to synchronize timestamps with guest");
-		return -1;
-	}
-	if (!offset || !timestamp || !scaling)
-		return 0;
-	clock = tsync->context;
-	if (clock->sync_count >= clock->sync_size) {
-		sync_ts = realloc(clock->sync_ts,
-				  (clock->sync_size + array_step) * sizeof(long long));
-		sync_offsets = realloc(clock->sync_offsets,
-				       (clock->sync_size + array_step) * sizeof(long long));
-		sync_scalings = realloc(clock->sync_scalings,
-				       (clock->sync_size + array_step) * sizeof(long long));
+	if (offsets->sync_count >= offsets->sync_size) {
+		sync_ts = realloc(offsets->sync_ts,
+				  (offsets->sync_size + array_step) * sizeof(long long));
+		sync_offsets = realloc(offsets->sync_offsets,
+				       (offsets->sync_size + array_step) * sizeof(long long));
+		sync_scalings = realloc(offsets->sync_scalings,
+				       (offsets->sync_size + array_step) * sizeof(long long));
 
 		if (!sync_ts || !sync_offsets || !sync_scalings) {
 			free(sync_ts);
@@ -623,20 +696,43 @@ static int tsync_get_sample(struct tracecmd_time_sync *tsync,
 			free(sync_scalings);
 			return -1;
 		}
-		clock->sync_size += array_step;
-		clock->sync_ts = sync_ts;
-		clock->sync_offsets = sync_offsets;
-		clock->sync_scalings = sync_scalings;
+		offsets->sync_size += array_step;
+		offsets->sync_ts = sync_ts;
+		offsets->sync_offsets = sync_offsets;
+		offsets->sync_scalings = sync_scalings;
 	}
 
-	clock->sync_ts[clock->sync_count] = timestamp;
-	clock->sync_offsets[clock->sync_count] = offset;
-	clock->sync_scalings[clock->sync_count] = scaling;
-	clock->sync_count++;
+	offsets->sync_ts[offsets->sync_count] = ts;
+	offsets->sync_offsets[offsets->sync_count] = offset;
+	offsets->sync_scalings[offsets->sync_count] = scaling;
+	offsets->sync_count++;
 
 	return 0;
 }
 
+static int tsync_get_sample(struct tracecmd_time_sync *tsync, unsigned int cpu,
+			    struct tsync_proto *proto, int array_step)
+{
+	struct clock_sync_context *clock;
+	long long timestamp = 0;
+	long long scaling = 0;
+	long long offset = 0;
+	int ret;
+
+	ret = proto->clock_sync_calc(tsync, &offset, &scaling, &timestamp, cpu);
+	if (ret) {
+		warning("Failed to synchronize timestamps with guest");
+		return -1;
+	}
+	if (!offset || !timestamp || !scaling)
+		return 0;
+	clock = tsync->context;
+	if (!clock || cpu >= clock->cpu_count || !clock->offsets)
+		return -1;
+	return record_sync_sample(&clock->offsets[cpu], array_step,
+				  offset, scaling, timestamp);
+}
+
 #define TIMER_SEC_NANO 1000000000LL
 static inline void get_ts_loop_delay(struct timespec *timeout, int delay_ms)
 {
@@ -654,12 +750,14 @@ static inline void get_ts_loop_delay(struct timespec *timeout, int delay_ms)
 #define CLOCK_TS_ARRAY 5
 static int tsync_with_guest(struct tracecmd_time_sync *tsync)
 {
+	struct tsync_probe_request_msg probe;
 	int ts_array_size = CLOCK_TS_ARRAY;
 	struct tsync_proto *proto;
 	struct timespec timeout;
 	bool first = true;
 	bool end = false;
 	int ret;
+	int i;
 
 	clock_context_init(tsync, &proto, false);
 	if (!tsync->context)
@@ -671,16 +769,21 @@ static int tsync_with_guest(struct tracecmd_time_sync *tsync)
 
 	while (true) {
 		pthread_mutex_lock(&tsync->lock);
-		ret = tracecmd_msg_send_time_sync(tsync->msg_handle,
-						  TRACECMD_TSYNC_PROTO_NONE,
-						  TRACECMD_TIME_SYNC_CMD_PROBE,
-						  0, NULL);
-		ret = tsync_get_sample(tsync, proto, ts_array_size);
+		for (i = 0; i < tsync->vcpu_count; i++) {
+			probe.cpu = i;
+			ret = tracecmd_msg_send_time_sync(tsync->msg_handle,
+							  TRACECMD_TSYNC_PROTO_NONE,
+							  TRACECMD_TIME_SYNC_CMD_PROBE,
+							  sizeof(probe), (char *)&probe);
+			ret = tsync_get_sample(tsync, i, proto, ts_array_size);
+			if (ret)
+				break;
+		}
 		if (first) {
 			first = false;
 			pthread_barrier_wait(&tsync->first_sync);
 		}
-		if (ret || end)
+		if (end || i < tsync->vcpu_count)
 			break;
 		if (tsync->loop_interval > 0) {
 			get_ts_loop_delay(&timeout, tsync->loop_interval);
@@ -819,41 +922,63 @@ error:
 int tracecmd_write_guest_time_shift(struct tracecmd_output *handle,
 				    struct tracecmd_time_sync *tsync)
 {
-	struct iovec vector[6];
+	struct iovec *vector = NULL;
 	unsigned int flags;
 	long long *scalings = NULL;
 	long long *offsets = NULL;
 	long long *ts = NULL;
+	int vcount;
 	int count;
-	int ret;
+	int i, j;
+	int ret = -1;
 
-	ret = tracecmd_tsync_get_offsets(tsync, &count,
-					 &ts, &offsets, &scalings);
-	if (ret < 0 || !count || !ts || !offsets || !scalings)
+	if (!tsync->vcpu_count)
+		return -1;
+	vcount = 3 + (4 * tsync->vcpu_count);
+	vector = calloc(vcount, sizeof(struct iovec));
+	if (!vector)
 		return -1;
 	ret = tsync_get_proto_flags(tsync, &flags);
 	if (ret < 0)
-		return -1;
-
-	vector[0].iov_len = 8;
-	vector[0].iov_base =  &(tsync->trace_id);
-	vector[1].iov_len = 4;
-	vector[1].iov_base = &flags;
-	vector[2].iov_len = 4;
-	vector[2].iov_base = &count;
-	vector[3].iov_len = 8 * count;
-	vector[3].iov_base = ts;
-	vector[4].iov_len = 8 * count;
-	vector[4].iov_base = offsets;
-	vector[5].iov_len = 8 * count;
-	vector[5].iov_base = scalings;
-	tracecmd_add_option_v(handle, TRACECMD_OPTION_TIME_SHIFT, vector, 6);
+		goto out;
+
+	j = 0;
+	vector[j].iov_len = 8;
+	vector[j++].iov_base = &tsync->trace_id;
+	vector[j].iov_len = 4;
+	vector[j++].iov_base = &flags;
+	vector[j].iov_len = 4;
+	vector[j++].iov_base = &tsync->vcpu_count;
+	for (i = 0; i < tsync->vcpu_count; i++) {
+		if (j >= vcount)
+			break;
+		ret = tracecmd_tsync_get_offsets(tsync, i, &count,
+						 &ts, &offsets, &scalings);
+		if (ret < 0 || !count || !ts || !offsets || !scalings)
+			break;
+		vector[j].iov_len = 4;
+		vector[j++].iov_base = &count;
+		vector[j].iov_len = 8 * count;
+		vector[j++].iov_base = ts;
+		vector[j].iov_len = 8 * count;
+		vector[j++].iov_base = offsets;
+		vector[j].iov_len = 8 * count;
+		vector[j++].iov_base = scalings;
+	}
+	if (i < tsync->vcpu_count) {
+		ret = -1;
+		goto out;
+	}
+	tracecmd_add_option_v(handle, TRACECMD_OPTION_TIME_SHIFT, vector, vcount);
 #ifdef TSYNC_DEBUG
 	if (count > 1)
 		printf("Got %d timestamp synch samples in %lld ns trace\n\r",
 			count, ts[count - 1] - ts[0]);
 #endif
-	return 0;
+	ret = 0;
+out:
+	free(vector);
+	return ret;
 }
 
 /**
@@ -938,7 +1063,7 @@ tracecmd_tsync_with_host(struct tracecmd_tsync_protos *tsync_protos,
 		return NULL;
 
 	proto = tsync_proto_select(tsync_protos, clock,
-					       TRACECMD_TIME_SYNC_ROLE_GUEST);
+				   TRACECMD_TIME_SYNC_ROLE_GUEST);
 	if (!proto)
 		goto error;
 	tsync->proto_name = strdup(proto);
diff --git a/tracecmd/trace-dump.c b/tracecmd/trace-dump.c
index 86524d0f..31b13af4 100644
--- a/tracecmd/trace-dump.c
+++ b/tracecmd/trace-dump.c
@@ -375,7 +375,8 @@ static void dump_option_timeshift(int fd, int size)
 	long long trace_id;
 	unsigned int count;
 	unsigned int flags;
-	int i;
+	unsigned int cpus;
+	int i, j;
 
 	/*
 	 * long long int (8 bytes) trace session ID
@@ -393,29 +394,34 @@ static void dump_option_timeshift(int fd, int size)
 	do_print(OPTIONS, "0x%llX [peer's trace id]\n", trace_id);
 	read_file_number(fd, &flags, 4);
 	do_print(OPTIONS, "0x%llX [peer's protocol flags]\n", flags);
-	read_file_number(fd, &count, 4);
-	do_print(OPTIONS, "%lld [samples count]\n", count);
-	times = calloc(count, sizeof(long long));
-	if (!times)
-		goto out;
-	offsets = calloc(count, sizeof(long long));
-	if (!offsets)
-		goto out;
-	scalings = calloc(count, sizeof(long long));
-	if (!scalings)
-		goto out;
-
-	for (i = 0; i < count; i++)
-		read_file_number(fd, times + i, 8);
-	for (i = 0; i < count; i++)
-		read_file_number(fd, offsets + i, 8);
-	for (i = 0; i < count; i++)
-		read_file_number(fd, scalings + i, 8);
-
-	for (i = 0; i < count; i++)
-		do_print(OPTIONS, "\t%lld * %lld %lld [offset * scaling @ time]\n",
-			 offsets[i], scalings[1], times[i]);
+	read_file_number(fd, &cpus, 4);
+	do_print(OPTIONS, "0x%llX [peer's CPU count]\n", cpus);
+	for (j = 0; j < cpus; j++) {
+		read_file_number(fd, &count, 4);
+		do_print(OPTIONS, "%lld [samples count for CPU %d]\n", count, j);
+		times = calloc(count, sizeof(long long));
+		offsets = calloc(count, sizeof(long long));
+		scalings = calloc(count, sizeof(long long));
+		if (!times || !offsets || !scalings)
+			goto out;
+		for (i = 0; i < count; i++)
+			read_file_number(fd, times + i, 8);
+		for (i = 0; i < count; i++)
+			read_file_number(fd, offsets + i, 8);
+		for (i = 0; i < count; i++)
+			read_file_number(fd, scalings + i, 8);
+
+		for (i = 0; i < count; i++)
+			do_print(OPTIONS, "\t%lld %lld %lld [offset * scaling @ time]\n",
+				 offsets[i], scalings[1], times[i]);
+		free(times);
+		free(offsets);
+		free(scalings);
+		times = NULL;
+		offsets = NULL;
+		scalings = NULL;
 
+	}
 out:
 	free(times);
 	free(offsets);
-- 
2.29.2


^ permalink raw reply	[flat|nested] 15+ messages in thread

* [PATCH v32 3/5] trace-cmd: PTP-like algorithm for host - guest timestamp synchronization
  2021-03-15  6:18 [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Tzvetomir Stoyanov (VMware)
  2021-03-15  6:18 ` [PATCH v32 1/5] trace-cmd: Add dummy function to initialize timestamp sync logic Tzvetomir Stoyanov (VMware)
  2021-03-15  6:18 ` [PATCH v32 2/5] trace-cmd: Add timestamp synchronization per vCPU Tzvetomir Stoyanov (VMware)
@ 2021-03-15  6:18 ` Tzvetomir Stoyanov (VMware)
  2021-03-15  6:18 ` [PATCH v32 4/5] trace-cmd: Debug scripts for " Tzvetomir Stoyanov (VMware)
                   ` (2 subsequent siblings)
  5 siblings, 0 replies; 15+ messages in thread
From: Tzvetomir Stoyanov (VMware) @ 2021-03-15  6:18 UTC (permalink / raw)
  To: rostedt; +Cc: linux-trace-devel

PTP protocol is designed for synchronizing clocks of machines in a local network.
The same approach can be used for host - guest timestamp synchronization.
This implementation uses ftrace raw markers to track trace timestamps of PTP events.
The patch is a POC, two different algorithms for PTP calculations are proposed:
  - Choosing the sample with the fastest response time for calculating the clocks offset.
  - Calculating the clocks offset using the average of all PTP samples.

The implementation can be tuned using those parameters:
 - #define FASTEST_RESPONSE - is defined, the sample with the fastest response time
    is used for calculating the clocks offset. Otherwise the histogram of all samples is used.
 - #define PTP_SYNC_LOOP 340 - defines the number of samples, used for one calculation.
 - --tsync-interval - a trace-cmd argument, choose the intervals between offset calculations,
	performed continuously during the trace.
 - #define TSYNC_DEBUG - if defined, a debug information is collected and stored in files,
   in the guest machine:
     s-cid*.txt - For each offset calculation: host and guest clocks and calculated offset.
     res-cid*.txt - For each tracing session: all calculated clock offsets.

Signed-off-by: Tzvetomir Stoyanov (VMware) <tz.stoyanov@gmail.com>
---
 lib/trace-cmd/Makefile             |   3 +
 lib/trace-cmd/trace-timesync-ptp.c | 715 +++++++++++++++++++++++++++++
 lib/trace-cmd/trace-timesync.c     |   4 +-
 3 files changed, 721 insertions(+), 1 deletion(-)
 create mode 100644 lib/trace-cmd/trace-timesync-ptp.c

diff --git a/lib/trace-cmd/Makefile b/lib/trace-cmd/Makefile
index 03d1a30c..6d4477a8 100644
--- a/lib/trace-cmd/Makefile
+++ b/lib/trace-cmd/Makefile
@@ -19,6 +19,9 @@ OBJS += trace-msg.o
 OBJS += trace-plugin.o
 OBJS += trace-perf.o
 OBJS += trace-timesync.o
+ifeq ($(VSOCK_DEFINED), 1)
+OBJS += trace-timesync-ptp.o
+endif
 
 # Additional util objects
 OBJS += trace-blk-hack.o
diff --git a/lib/trace-cmd/trace-timesync-ptp.c b/lib/trace-cmd/trace-timesync-ptp.c
new file mode 100644
index 00000000..b05f1cd0
--- /dev/null
+++ b/lib/trace-cmd/trace-timesync-ptp.c
@@ -0,0 +1,715 @@
+// SPDX-License-Identifier: LGPL-2.1
+/*
+ * Copyright (C) 2019, VMware, Tzvetomir Stoyanov tz.stoyanov@gmail.com>
+ *
+ */
+
+#include <fcntl.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <arpa/inet.h>
+#include <linux/vm_sockets.h>
+#include <sys/types.h>
+#include <linux/types.h>
+#include <time.h>
+#include <sched.h>
+#include <limits.h>
+
+#include "trace-cmd.h"
+#include "trace-cmd-private.h"
+#include "tracefs.h"
+#include "trace-tsync-local.h"
+#include "trace-msg.h"
+#include "trace-cmd-local.h"
+
+typedef __be32 be32;
+typedef __u64 u64;
+typedef __s64 s64;
+
+#define PTP_SYNC_LOOP	339
+
+#define PTP_SYNC_PKT_START	1
+#define PTP_SYNC_PKT_PROBE	2
+#define PTP_SYNC_PKT_PROBES	3
+#define PTP_SYNC_PKT_OFFSET	4
+#define PTP_SYNC_PKT_END	5
+
+/* print time sync debug messages */
+#define TSYNC_DEBUG
+
+struct ptp_clock_sync {
+	struct tep_handle	*tep;
+	struct tep_format_field	*id;
+	int			raw_id;
+	int			marker_fd;
+	int			series_id;
+	int			flags;
+	int			debug_fd;
+};
+
+enum {
+/*
+ * Consider only the probe with fastest response time,
+ * otherwise make a histogram from all probes.
+ */
+	PTP_FLAG_FASTEST_RESPONSE		= (1 << 0),
+/*
+ * Use trace marker to get the clock,
+ * otherwise use the system clock directly.
+ */
+	PTP_FLAG_USE_MARKER			= (1 << 1),
+};
+static int ptp_flags = PTP_FLAG_FASTEST_RESPONSE | PTP_FLAG_USE_MARKER;
+
+/*
+ * Calculated using formula [CPU rate]*[calculated offset deviation]
+ * tested on 3GHz CPU, with x86-tsc trace clock and compare the calculated
+ * offset with /sys/kernel/debug/kvm/<VM ID>/vcpu0/tsc-offset
+ * measured 2000ns deviation
+ * using PTP flags PTP_FLAG_FASTEST_RESPONSE | PTP_FLAG_USE_MARKER
+ */
+#define PTP_ACCURACY	6000
+#define PTP_NAME	"ptp"
+
+struct ptp_clock_start_msg {
+	be32	series_id;
+	be32	flags;
+} __packed;
+
+struct ptp_clock_sample {
+	s64		ts;
+	be32		id;
+} __packed;
+
+struct ptp_clock_result_msg {
+	be32			series_id;
+	be32			count;
+	struct ptp_clock_sample	samples[2*PTP_SYNC_LOOP];
+} __packed;
+
+struct ptp_clock_offset_msg {
+	s64	ts;
+	s64	offset;
+};
+
+struct ptp_markers_context {
+	struct clock_sync_context	*clock;
+	struct ptp_clock_sync		*ptp;
+	struct ptp_clock_result_msg	msg;
+	int				size;
+};
+
+struct ptp_marker_buf {
+	int local_cid;
+	int remote_cid;
+	int count;
+	int packet_id;
+} __packed;
+
+struct ptp_marker {
+	int series_id;
+	struct ptp_marker_buf data;
+} __packed;
+
+static int ptp_clock_sync_init(struct tracecmd_time_sync *tsync)
+{
+	const char *systems[] = {"ftrace", NULL};
+	struct clock_sync_context *clock_context;
+	struct ptp_clock_sync *ptp;
+	struct tep_event *raw;
+	char *path;
+
+	if (!tsync || !tsync->context)
+		return -1;
+	clock_context = (struct clock_sync_context *)tsync->context;
+	if (clock_context->proto_data)
+		return 0;
+
+	ptp = calloc(1, sizeof(struct ptp_clock_sync));
+	if (!ptp)
+		return -1;
+
+	ptp->marker_fd = -1;
+	ptp->debug_fd = -1;
+
+	path = tracefs_instance_get_dir(clock_context->instance);
+	if (!path)
+		goto error;
+	ptp->tep = tracefs_local_events_system(path, systems);
+	tracefs_put_tracing_file(path);
+	if (!ptp->tep)
+		goto error;
+	raw = tep_find_event_by_name(ptp->tep, "ftrace", "raw_data");
+	if (!raw)
+		goto error;
+	ptp->id = tep_find_field(raw, "id");
+	if (!ptp->id)
+		goto error;
+	ptp->raw_id = raw->id;
+
+	tep_set_file_bigendian(ptp->tep, tracecmd_host_bigendian());
+	tep_set_local_bigendian(ptp->tep, tracecmd_host_bigendian());
+
+	path = tracefs_instance_get_file(clock_context->instance, "trace_marker_raw");
+	if (!path)
+		goto error;
+	ptp->marker_fd = open(path, O_WRONLY);
+	tracefs_put_tracing_file(path);
+
+	clock_context->proto_data = ptp;
+
+#ifdef TSYNC_DEBUG
+	if (clock_context->is_server) {
+		char buff[256];
+		int res_fd;
+
+		sprintf(buff, "res-cid%d.txt", clock_context->remote_cid);
+
+		res_fd = open(buff, O_CREAT|O_WRONLY|O_TRUNC, 0644);
+		if (res_fd > 0)
+			close(res_fd);
+	}
+#endif
+
+	return 0;
+
+error:
+	if (ptp) {
+		tep_free(ptp->tep);
+		if (ptp->marker_fd >= 0)
+			close(ptp->marker_fd);
+	}
+	free(ptp);
+	return -1;
+}
+
+static int ptp_clock_sync_free(struct tracecmd_time_sync *tsync)
+{
+	struct clock_sync_context *clock_context;
+	struct ptp_clock_sync *ptp;
+
+	if (!tsync || !tsync->context)
+		return -1;
+	clock_context = (struct clock_sync_context *)tsync->context;
+
+	if (clock_context && clock_context->proto_data) {
+		ptp = (struct ptp_clock_sync *)clock_context->proto_data;
+		tep_free(ptp->tep);
+		if (ptp->marker_fd >= 0)
+			close(ptp->marker_fd);
+		if (ptp->debug_fd >= 0)
+			close(ptp->debug_fd);
+		free(clock_context->proto_data);
+		clock_context->proto_data = NULL;
+	}
+	return 0;
+}
+
+/* Save the timestamps of sent ('s') and returned ('r') probes in the
+ * ctx->msg.samples[] array. Depending of the context (server or client), there
+ * may be only returned probes, or both sent and returned probes. The returned
+ * probes are saved first in the array, after them are the sent probes.
+ * Depending of the context, the array can be with size:
+ *  [0 .. max data.count] - holds only returned probes
+ *  [0 .. 2 * max data.count] - holds both returned and sent probes
+ */
+static void ptp_probe_store(struct ptp_markers_context *ctx,
+			    struct ptp_marker *marker,
+			    unsigned long long ts)
+{
+	int index = -1;
+
+	if (marker->data.packet_id == 'r' &&
+	    marker->data.count <= ctx->size) {
+		index = marker->data.count - 1;
+	} else if (marker->data.packet_id == 's' &&
+		  marker->data.count * 2 <= ctx->size){
+		index = ctx->size / 2 + marker->data.count - 1;
+	}
+
+	if (index >= 0) {
+		ctx->msg.samples[index].id = marker->data.count;
+		ctx->msg.samples[index].ts = ts;
+		ctx->msg.count++;
+	}
+}
+
+static int ptp_marker_find(struct tep_event *event, struct tep_record *record,
+			   int cpu, void *context)
+{
+	struct ptp_markers_context *ctx;
+	struct ptp_marker *marker;
+
+	ctx = (struct ptp_markers_context *)context;
+
+	/* Make sure this is our event */
+	if (event->id != ctx->ptp->raw_id || !ctx->ptp->id)
+		return 0;
+	if (record->size >= (ctx->ptp->id->offset + sizeof(struct ptp_marker))) {
+		marker = (struct ptp_marker *)(record->data + ctx->ptp->id->offset);
+		if (marker->data.local_cid == ctx->clock->local_cid &&
+		    marker->data.remote_cid == ctx->clock->remote_cid &&
+		    marker->series_id == ctx->ptp->series_id &&
+		    marker->data.count)
+			ptp_probe_store(ctx, marker, record->ts);
+	}
+
+	return 0;
+}
+
+static inline bool good_probe(struct ptp_clock_sample *server_sample,
+			      struct ptp_clock_sample *send_sample,
+			      struct ptp_clock_sample *client_sample,
+			      int *bad_probes)
+{
+	if (server_sample->ts && send_sample->ts && client_sample->ts &&
+	    server_sample->id == send_sample->id &&
+	    server_sample->id == client_sample->id)
+		return true;
+	(*bad_probes)++;
+	return false;
+}
+
+static int ptp_calc_offset_fastest(struct clock_sync_context *clock,
+			   struct ptp_clock_result_msg *server,
+			   struct ptp_clock_result_msg *client,
+			   long long *offset_ret, long long *ts_ret,
+			   int *bad_probes)
+{
+	struct ptp_clock_sample *sample_send;
+	long long delta_min = LLONG_MAX;
+	long long offset = 0;
+	long long delta = 0;
+	long long ts = 0;
+	int max_i;
+	int i;
+
+	*bad_probes = 0;
+	sample_send = server->samples + (server->count / 2);
+	max_i = server->count / 2 < client->count ?
+		server->count / 2 : client->count;
+	for (i = 0; i < max_i; i++) {
+		if (!good_probe(&server->samples[i], &sample_send[i],
+		    &client->samples[i], bad_probes))
+			continue;
+		ts = (sample_send[i].ts + server->samples[i].ts) / 2;
+		offset = client->samples[i].ts - ts;
+
+		delta = server->samples[i].ts - sample_send[i].ts;
+		if (delta_min > delta) {
+			delta_min = delta;
+			*offset_ret = offset;
+			*ts_ret = ts;
+		}
+#ifdef TSYNC_DEBUG
+		{
+			struct ptp_clock_sync *ptp;
+
+			ptp = (struct ptp_clock_sync *)clock->proto_data;
+			if (ptp && ptp->debug_fd > 0) {
+				char buff[256];
+
+				sprintf(buff, "%lld %lld %lld\n",
+					ts, client->samples[i].ts, offset);
+				write(ptp->debug_fd, buff, strlen(buff));
+			}
+		}
+#endif
+	}
+
+	return 0;
+}
+
+static int ptp_calc_offset_hist(struct clock_sync_context *clock,
+			   struct ptp_clock_result_msg *server,
+			   struct ptp_clock_result_msg *client,
+			   long long *offset_ret, long long *ts_ret,
+			   int *bad_probes)
+{
+	struct ptp_clock_sample *sample_send;
+	long long timestamps[PTP_SYNC_LOOP];
+	long long offsets[PTP_SYNC_LOOP];
+	long long offset_min = LLONG_MAX;
+	long long offset_max = 0;
+	int hist[PTP_SYNC_LOOP];
+	int ind, max = 0;
+	long long bin;
+	int i, k = 0;
+
+	*bad_probes = 0;
+	memset(hist, 0, sizeof(int) * PTP_SYNC_LOOP);
+	sample_send = server->samples + (server->count / 2);
+	for (i = 0; i * 2 < server->count && i < client->count; i++) {
+		if (!good_probe(&server->samples[i], &sample_send[i],
+		    &client->samples[i], bad_probes))
+			continue;
+		timestamps[k] = (sample_send[i].ts + server->samples[i].ts) / 2;
+		offsets[k] = client->samples[i].ts - timestamps[k];
+		if (offset_max < llabs(offsets[k]))
+			offset_max = llabs(offsets[k]);
+		if (offset_min > llabs(offsets[k]))
+			offset_min = llabs(offsets[k]);
+#ifdef TSYNC_DEBUG
+		{
+			struct ptp_clock_sync *ptp;
+
+			ptp = (struct ptp_clock_sync *)clock->proto_data;
+
+			if (ptp && ptp->debug_fd > 0) {
+				char buff[256];
+
+				sprintf(buff, "%lld %lld %lld\n",
+					timestamps[k],
+					client->samples[i].ts, offsets[k]);
+				write(ptp->debug_fd, buff, strlen(buff));
+			}
+		}
+#endif
+		k++;
+	}
+
+	bin = (offset_max - offset_min) / PTP_SYNC_LOOP;
+	for (i = 0; i < k; i++) {
+		ind = (llabs(offsets[i]) - offset_min) / bin;
+		if (ind < PTP_SYNC_LOOP) {
+			hist[ind]++;
+			if (max < hist[ind]) {
+				max = hist[ind];
+				*offset_ret = offsets[i];
+				*ts_ret = timestamps[i];
+			}
+		}
+	}
+
+	return 0;
+}
+
+static void ntoh_ptp_results(struct ptp_clock_result_msg *msg)
+{
+	int i;
+
+	msg->count = ntohl(msg->count);
+	for (i = 0; i < msg->count; i++) {
+		msg->samples[i].id = ntohl(msg->samples[i].id);
+		msg->samples[i].ts = ntohll(msg->samples[i].ts);
+	}
+	msg->series_id = ntohl(msg->series_id);
+}
+
+
+static void hton_ptp_results(struct ptp_clock_result_msg *msg)
+{
+	int i;
+
+	for (i = 0; i < msg->count; i++) {
+		msg->samples[i].id = htonl(msg->samples[i].id);
+		msg->samples[i].ts = htonll(msg->samples[i].ts);
+	}
+	msg->series_id = htonl(msg->series_id);
+	msg->count = htonl(msg->count);
+}
+
+static inline void ptp_track_clock(struct ptp_markers_context *ctx,
+				   struct ptp_marker *marker)
+{
+	if (ctx->ptp->flags & PTP_FLAG_USE_MARKER) {
+		write(ctx->ptp->marker_fd, marker, sizeof(struct ptp_marker));
+	} else {
+		struct timespec clock;
+		unsigned long long ts;
+
+		clock_gettime(CLOCK_MONOTONIC_RAW, &clock);
+		ts = clock.tv_sec * 1000000000LL;
+		ts += clock.tv_nsec;
+		ptp_probe_store(ctx, marker, ts);
+	}
+}
+
+static int ptp_clock_client(struct tracecmd_time_sync *tsync,
+			    long long *offset, long long *timestamp)
+{
+	char sync_proto[TRACECMD_TSYNC_PNAME_LENGTH];
+	struct clock_sync_context *clock_context;
+	struct ptp_clock_offset_msg res_offset;
+	struct ptp_clock_start_msg start;
+	struct ptp_markers_context ctx;
+	struct ptp_clock_sync *ptp;
+	struct ptp_marker marker;
+	unsigned int sync_msg;
+	unsigned int size;
+	char *msg;
+	int count;
+	int ret;
+
+	if (!tsync || !tsync->context || !tsync->msg_handle)
+		return -1;
+
+	clock_context = (struct clock_sync_context *)tsync->context;
+	if (clock_context->proto_data == NULL)
+		return -1;
+
+	ptp = (struct ptp_clock_sync *)clock_context->proto_data;
+	size = sizeof(start);
+	msg = (char *)&start;
+	ret = tracecmd_msg_recv_time_sync(tsync->msg_handle,
+					  sync_proto, &sync_msg,
+					  &size, &msg);
+	if (ret || strncmp(sync_proto, PTP_NAME, TRACECMD_TSYNC_PNAME_LENGTH) ||
+	    sync_msg != PTP_SYNC_PKT_START)
+		return -1;
+	ret = tracecmd_msg_send_time_sync(tsync->msg_handle, PTP_NAME,
+					  PTP_SYNC_PKT_START, sizeof(start),
+					  (char *)&start);
+	marker.data.local_cid = clock_context->local_cid;
+	marker.data.remote_cid = clock_context->remote_cid;
+	marker.series_id = ntohl(start.series_id);
+	marker.data.packet_id = 'r';
+	ptp->series_id = marker.series_id;
+	ptp->flags = ntohl(start.flags);
+	msg = (char *)&count;
+	size = sizeof(count);
+	ctx.msg.count = 0;
+	ctx.size = PTP_SYNC_LOOP;
+	ctx.ptp = ptp;
+	ctx.clock = clock_context;
+	ctx.msg.series_id = ptp->series_id;
+	while (true) {
+		count = 0;
+		ret = tracecmd_msg_recv_time_sync(tsync->msg_handle,
+						  sync_proto, &sync_msg,
+						  &size, &msg);
+		if (ret || strncmp(sync_proto, PTP_NAME, TRACECMD_TSYNC_PNAME_LENGTH) ||
+		    sync_msg != PTP_SYNC_PKT_PROBE || !ntohl(count))
+			break;
+		marker.data.count = ntohl(count);
+		ptp_track_clock(&ctx, &marker);
+		ret = tracecmd_msg_send_time_sync(tsync->msg_handle, PTP_NAME,
+						  PTP_SYNC_PKT_PROBE,
+						  sizeof(count), (char *)&count);
+		if (ret)
+			break;
+	}
+
+	if (strncmp(sync_proto, PTP_NAME, TRACECMD_TSYNC_PNAME_LENGTH) ||
+	    sync_msg != PTP_SYNC_PKT_END)
+		return -1;
+
+	if (ptp->flags & PTP_FLAG_USE_MARKER)
+		tracefs_iterate_raw_events(ptp->tep, clock_context->instance,
+					   NULL, 0, ptp_marker_find, &ctx);
+
+	hton_ptp_results(&ctx.msg);
+	ret = tracecmd_msg_send_time_sync(tsync->msg_handle, PTP_NAME,
+					  PTP_SYNC_PKT_PROBES,
+					  sizeof(ctx.msg), (char *)&ctx.msg);
+
+	msg = (char *)&res_offset;
+	size = sizeof(res_offset);
+	ret = tracecmd_msg_recv_time_sync(tsync->msg_handle,
+					  sync_proto, &sync_msg,
+					  &size, (char **)&msg);
+	if (ret || strncmp(sync_proto, PTP_NAME, TRACECMD_TSYNC_PNAME_LENGTH) ||
+	    sync_msg != PTP_SYNC_PKT_OFFSET)
+		return -1;
+
+	*offset = ntohll(res_offset.offset);
+	*timestamp = ntohll(res_offset.ts);
+
+	return 0;
+}
+
+
+static int ptp_clock_server(struct tracecmd_time_sync *tsync,
+			    long long *offset, long long *timestamp)
+{
+	char sync_proto[TRACECMD_TSYNC_PNAME_LENGTH];
+	struct ptp_clock_result_msg *results = NULL;
+	struct clock_sync_context *clock_context;
+	struct ptp_clock_offset_msg res_offset;
+	struct ptp_clock_start_msg start;
+	struct ptp_markers_context ctx;
+	int sync_loop = PTP_SYNC_LOOP;
+	struct ptp_clock_sync *ptp;
+	struct ptp_marker marker;
+	unsigned int sync_msg;
+	unsigned int size;
+	int bad_probes;
+	int count = 1;
+	int msg_count;
+	int msg_ret;
+	char *msg;
+	int ret;
+
+	if (!tsync || !tsync->context || !tsync->msg_handle)
+		return -1;
+
+	clock_context = (struct clock_sync_context *)tsync->context;
+	if (clock_context->proto_data == NULL)
+		return -1;
+
+	ptp = (struct ptp_clock_sync *)clock_context->proto_data;
+	ptp->flags = ptp_flags;
+	memset(&start, 0, sizeof(start));
+	start.series_id = htonl(ptp->series_id + 1);
+	start.flags = htonl(ptp->flags);
+	ret = tracecmd_msg_send_time_sync(tsync->msg_handle, PTP_NAME,
+					 PTP_SYNC_PKT_START, sizeof(start),
+					 (char *)&start);
+	if (!ret)
+		ret = tracecmd_msg_recv_time_sync(tsync->msg_handle,
+						  sync_proto, &sync_msg,
+						  NULL, NULL);
+	if (ret || strncmp(sync_proto, PTP_NAME, TRACECMD_TSYNC_PNAME_LENGTH) ||
+	    sync_msg != PTP_SYNC_PKT_START)
+		return -1;
+
+	tracefs_instance_file_write(clock_context->instance, "trace", "\0");
+
+	ptp->series_id++;
+	marker.data.local_cid = clock_context->local_cid;
+	marker.data.remote_cid = clock_context->remote_cid;
+	marker.series_id = ptp->series_id;
+	msg = (char *)&msg_ret;
+	size = sizeof(msg_ret);
+	ctx.size = 2*PTP_SYNC_LOOP;
+	ctx.ptp = ptp;
+	ctx.clock = clock_context;
+	ctx.msg.count = 0;
+	ctx.msg.series_id = ptp->series_id;
+	do {
+		marker.data.count = count++;
+		marker.data.packet_id = 's';
+		msg_count = htonl(marker.data.count);
+		ptp_track_clock(&ctx, &marker);
+		ret = tracecmd_msg_send_time_sync(tsync->msg_handle, PTP_NAME,
+						 PTP_SYNC_PKT_PROBE,
+						 sizeof(msg_count),
+						 (char *)&msg_count);
+		if (!ret)
+			ret = tracecmd_msg_recv_time_sync(tsync->msg_handle,
+							  sync_proto, &sync_msg,
+							  &size, &msg);
+
+		marker.data.packet_id = 'r';
+		ptp_track_clock(&ctx, &marker);
+		if (ret || strncmp(sync_proto, PTP_NAME, TRACECMD_TSYNC_PNAME_LENGTH) ||
+		    sync_msg != PTP_SYNC_PKT_PROBE ||
+		    ntohl(msg_ret) != marker.data.count)
+			break;
+	} while (--sync_loop);
+
+	if (sync_loop)
+		return -1;
+
+	ret = tracecmd_msg_send_time_sync(tsync->msg_handle, PTP_NAME,
+					  PTP_SYNC_PKT_END, 0, NULL);
+
+	size = 0;
+	ret = tracecmd_msg_recv_time_sync(tsync->msg_handle,
+					  sync_proto, &sync_msg,
+					  &size, (char **)&results);
+	if (ret || strncmp(sync_proto, PTP_NAME, TRACECMD_TSYNC_PNAME_LENGTH) ||
+	    sync_msg != PTP_SYNC_PKT_PROBES || size == 0 || results == NULL)
+		return -1;
+
+	ntoh_ptp_results(results);
+	if (ptp->flags & PTP_FLAG_USE_MARKER)
+		tracefs_iterate_raw_events(ptp->tep, clock_context->instance,
+					   NULL, 0, ptp_marker_find, &ctx);
+	if (ptp->flags & PTP_FLAG_FASTEST_RESPONSE)
+		ptp_calc_offset_fastest(clock_context, &ctx.msg, results, offset,
+					timestamp, &bad_probes);
+	else
+		ptp_calc_offset_hist(clock_context, &ctx.msg, results, offset,
+				     timestamp, &bad_probes);
+#ifdef TSYNC_DEBUG
+	{
+		char buff[256];
+		int res_fd;
+
+		sprintf(buff, "res-cid%d.txt", clock_context->remote_cid);
+
+		res_fd = open(buff, O_WRONLY|O_APPEND, 0644);
+		if (res_fd > 0) {
+			if (*offset && *timestamp) {
+				sprintf(buff, "%d %lld %lld\n",
+					ptp->series_id, *offset, *timestamp);
+				write(res_fd, buff, strlen(buff));
+			}
+			close(res_fd);
+		}
+
+		printf("\n calculated offset %d: %lld, %d probes, filtered out %d, PTP flags 0x%X\n\r",
+			ptp->series_id, *offset, results->count, bad_probes, ptp->flags);
+		if (ptp && ptp->debug_fd > 0) {
+			sprintf(buff, "%lld %lld 0\n", *offset, *timestamp);
+			write(ptp->debug_fd, buff, strlen(buff));
+			close(ptp->debug_fd);
+			ptp->debug_fd = -1;
+		}
+
+	}
+#endif
+
+	res_offset.offset = htonll(*offset);
+	res_offset.ts = htonll(*timestamp);
+	ret = tracecmd_msg_send_time_sync(tsync->msg_handle, PTP_NAME,
+					  PTP_SYNC_PKT_OFFSET,
+					  sizeof(res_offset),
+					  (char *)&res_offset);
+
+	free(results);
+	return 0;
+}
+
+static int ptp_clock_sync_calc(struct tracecmd_time_sync *tsync,
+			       long long *offset, long long *scaling,
+			       long long *timestamp, unsigned int cpu)
+{
+	struct clock_sync_context *clock_context;
+	int ret;
+
+	if (!tsync || !tsync->context)
+		return -1;
+	clock_context = (struct clock_sync_context *)tsync->context;
+
+#ifdef TSYNC_DEBUG
+	if (clock_context->is_server) {
+		struct ptp_clock_sync *ptp;
+		char buff[256];
+
+		ptp = (struct ptp_clock_sync *)clock_context->proto_data;
+		if (ptp->debug_fd > 0)
+			close(ptp->debug_fd);
+		sprintf(buff, "s-cid%d_%d.txt",
+				clock_context->remote_cid, ptp->series_id+1);
+		ptp->debug_fd = open(buff, O_CREAT|O_WRONLY|O_TRUNC, 0644);
+	}
+#endif
+
+	if (scaling)
+		*scaling = 1;
+	if (clock_context->is_server)
+		ret = ptp_clock_server(tsync, offset, timestamp);
+	else
+		ret = ptp_clock_client(tsync, offset, timestamp);
+
+	return ret;
+}
+
+int ptp_clock_sync_register(void)
+{
+	return tracecmd_tsync_proto_register(PTP_NAME, PTP_ACCURACY,
+					     TRACECMD_TIME_SYNC_ROLE_GUEST |
+					     TRACECMD_TIME_SYNC_ROLE_HOST,
+					     0, TRACECMD_TSYNC_FLAG_INTERPOLATE,
+					     ptp_clock_sync_init,
+					     ptp_clock_sync_free,
+					     ptp_clock_sync_calc);
+
+}
+
+int ptp_clock_sync_unregister(void)
+{
+	return tracecmd_tsync_proto_unregister(PTP_NAME);
+}
diff --git a/lib/trace-cmd/trace-timesync.c b/lib/trace-cmd/trace-timesync.c
index b0ea8f6d..588484e7 100644
--- a/lib/trace-cmd/trace-timesync.c
+++ b/lib/trace-cmd/trace-timesync.c
@@ -64,7 +64,9 @@ static struct tsync_proto *tsync_proto_find(const char *proto_name)
  */
 void tracecmd_tsync_init(void)
 {
-
+#ifdef VSOCK
+	ptp_clock_sync_register();
+#endif
 }
 
 int tracecmd_tsync_proto_register(const char *proto_name, int accuracy, int roles,
-- 
2.29.2


^ permalink raw reply	[flat|nested] 15+ messages in thread

* [PATCH v32 4/5] trace-cmd: Debug scripts for PTP-like algorithm for host - guest timestamp synchronization
  2021-03-15  6:18 [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Tzvetomir Stoyanov (VMware)
                   ` (2 preceding siblings ...)
  2021-03-15  6:18 ` [PATCH v32 3/5] trace-cmd: PTP-like algorithm for host - guest timestamp synchronization Tzvetomir Stoyanov (VMware)
@ 2021-03-15  6:18 ` Tzvetomir Stoyanov (VMware)
  2021-03-15  6:18 ` [PATCH v32 5/5] trace-cmd [POC]: Add KVM timestamp synchronization plugin Tzvetomir Stoyanov (VMware)
  2021-03-19 10:08 ` [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Stefano De Venuto
  5 siblings, 0 replies; 15+ messages in thread
From: Tzvetomir Stoyanov (VMware) @ 2021-03-15  6:18 UTC (permalink / raw)
  To: rostedt; +Cc: linux-trace-devel

These scripts can be used to visualise debug files, written when the PTP-like algorithm
is compiled with TSYNC_DEBUG defined. The files are located in the guest machine:
    s-cid*.txt - For each offset calculation: host and guest clocks and calculated offset.
    res-cid*.txt - For each tracing session: all calculated clock offsets.

tsync_hist.py plots a histogram, using data from a s-cid*.txt file:
	"python tsync_hist.py s-cid2_1.txt"
tsync_res.py plots a line, using data from res-cid*.txt file:
	"python tsync_res.py res-cid2.txt"

Signed-off-by: Tzvetomir Stoyanov (VMware) <tz.stoyanov@gmail.com>
---
 scripts/debug/tsync_hist.py | 57 +++++++++++++++++++++++++++++++++++++
 scripts/debug/tsync_readme  | 12 ++++++++
 scripts/debug/tsync_res.py  | 46 ++++++++++++++++++++++++++++++
 3 files changed, 115 insertions(+)
 create mode 100644 scripts/debug/tsync_hist.py
 create mode 100644 scripts/debug/tsync_readme
 create mode 100644 scripts/debug/tsync_res.py

diff --git a/scripts/debug/tsync_hist.py b/scripts/debug/tsync_hist.py
new file mode 100644
index 00000000..819d1e8f
--- /dev/null
+++ b/scripts/debug/tsync_hist.py
@@ -0,0 +1,57 @@
+# SPDX-License-Identifier: GPL-2.0
+#
+# Copyright (C) 2019, VMware Inc, Tzvetomir Stoyanov <tz.stoyanov@gmail.com>
+# Copyright (C) 2019, VMware Inc, Yordan Karadzhov <ykaradzhov@vmware.com>
+
+
+import matplotlib.pyplot as plt
+import matplotlib.lines as mlines
+import numpy as np
+import sys
+
+def newline(p1, p2):
+    ax = plt.gca()
+    xmin, xmax = ax.get_xbound()
+
+    if(p2[0] == p1[0]):
+        xmin = xmax = p1[0]
+        ymin, ymax = ax.get_ybound()
+    else:
+        ymax = p1[1]+(p2[1]-p1[1])/(p2[0]-p1[0])*(xmax-p1[0])
+        ymin = p1[1]+(p2[1]-p1[1])/(p2[0]-p1[0])*(xmin-p1[0])
+
+    l = mlines.Line2D([xmin,xmax], [ymin,ymax], color='red')
+    ax.add_line(l)
+    return l
+
+
+data = np.loadtxt(fname = sys.argv[1])
+selected_ts  = data[-1, 1]
+selected_ofs = data[-1, 0]
+data = data[:-1,:]
+
+x = data[:, 1] - data[:, 0]
+
+mean = x.mean()
+std = x.std()
+
+num_bins = 500
+min = x.min() #+ .4 * (x.max() - x.min())
+max = x.max() #- .4 * (x.max() - x.min())
+bins = np.linspace(min, max, num_bins, endpoint = False, dtype=int)
+
+fig, ax = plt.subplots()
+
+# the histogram of the data
+n, bins, patches = ax.hist(x, bins, histtype=u'step');
+
+ax.set_xlabel('clock offset [$\mu$s]')
+ax.set_ylabel('entries')
+ax.set_title("$\sigma$=%i" % std)
+
+x1, y1 = [selected_ofs, min], [selected_ofs, max]
+newline(x1, y1)
+
+# Tweak spacing to prevent clipping of ylabel
+fig.tight_layout()
+plt.show()
diff --git a/scripts/debug/tsync_readme b/scripts/debug/tsync_readme
new file mode 100644
index 00000000..f3ebb25d
--- /dev/null
+++ b/scripts/debug/tsync_readme
@@ -0,0 +1,12 @@
+PTP-like algorithm debug
+========================
+
+tsync_*.py scripts can be used to visualise debug files, written when the PTP-like algorithm
+is compiled with TSYNC_DEBUG defined. The files are located in the guest machine:
+    s-cid*.txt - For each offset calculation: host and guest clocks and calculated offset.
+    res-cid*.txt - For each tracing session: all calculated clock offsets.
+
+tsync_hist.py plots a histogram, using data from a s-cid*.txt file:
+	"python tsync_hist.py s-cid2_1.txt"
+tsync_res.py plots a line, using data from res-cid*.txt file:
+	"python tsync_res.py res-cid2.txt"
diff --git a/scripts/debug/tsync_res.py b/scripts/debug/tsync_res.py
new file mode 100644
index 00000000..7d109863
--- /dev/null
+++ b/scripts/debug/tsync_res.py
@@ -0,0 +1,46 @@
+# SPDX-License-Identifier: GPL-2.0
+#
+# Copyright (C) 2019, VMware Inc, Tzvetomir Stoyanov <tz.stoyanov@gmail.com>
+# Copyright (C) 2019, VMware Inc, Yordan Karadzhov <ykaradzhov@vmware.com>
+
+
+import matplotlib.pyplot as plt
+import matplotlib.lines as mlines
+import numpy as np
+import sys
+
+def newline(p1, p2):
+    ax = plt.gca()
+    xmin, xmax = ax.get_xbound()
+
+    if(p2[0] == p1[0]):
+        xmin = xmax = p1[0]
+        ymin, ymax = ax.get_ybound()
+    else:
+        ymax = p1[1]+(p2[1]-p1[1])/(p2[0]-p1[0])*(xmax-p1[0])
+        ymin = p1[1]+(p2[1]-p1[1])/(p2[0]-p1[0])*(xmin-p1[0])
+
+    l = mlines.Line2D([xmin,xmax], [ymin,ymax], color='red')
+    ax.add_line(l)
+    return l
+
+data = np.loadtxt(fname = sys.argv[1])
+x = data[:, 0]
+y = data[:, 1]
+
+fig, ax = plt.subplots()
+
+ax.set_xlabel('samples (t)')
+ax.set_ylabel('clock offset')
+ax.set_title("$\delta$=%i ns" % (max(y) - min(y)))
+
+l = mlines.Line2D(x, y)
+ax.add_line(l)
+ax.set_xlim(min(x), max(x))
+ax.set_ylim(min(y), max(y) )
+
+print(min(y), max(y), max(y) - min(y))
+
+# Tweak spacing to prevent clipping of ylabel
+fig.tight_layout()
+plt.show()
-- 
2.29.2


^ permalink raw reply	[flat|nested] 15+ messages in thread

* [PATCH v32 5/5] trace-cmd [POC]: Add KVM timestamp synchronization plugin
  2021-03-15  6:18 [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Tzvetomir Stoyanov (VMware)
                   ` (3 preceding siblings ...)
  2021-03-15  6:18 ` [PATCH v32 4/5] trace-cmd: Debug scripts for " Tzvetomir Stoyanov (VMware)
@ 2021-03-15  6:18 ` Tzvetomir Stoyanov (VMware)
  2021-03-19 10:08 ` [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Stefano De Venuto
  5 siblings, 0 replies; 15+ messages in thread
From: Tzvetomir Stoyanov (VMware) @ 2021-03-15  6:18 UTC (permalink / raw)
  To: rostedt; +Cc: linux-trace-devel

Added new timestamp synchronization plugin for KVM hosts. It reads the
clock offsets directly from the KVM debug filesystem, if available.
The plugin works only with x86-tsc ftrace clock.

Signed-off-by: Tzvetomir Stoyanov (VMware) <tz.stoyanov@gmail.com>
---
 lib/trace-cmd/Makefile                    |   1 +
 lib/trace-cmd/include/trace-tsync-local.h |   1 +
 lib/trace-cmd/trace-timesync-kvm.c        | 460 ++++++++++++++++++++++
 lib/trace-cmd/trace-timesync.c            |   1 +
 4 files changed, 463 insertions(+)
 create mode 100644 lib/trace-cmd/trace-timesync-kvm.c

diff --git a/lib/trace-cmd/Makefile b/lib/trace-cmd/Makefile
index 6d4477a8..bee6d1f1 100644
--- a/lib/trace-cmd/Makefile
+++ b/lib/trace-cmd/Makefile
@@ -21,6 +21,7 @@ OBJS += trace-perf.o
 OBJS += trace-timesync.o
 ifeq ($(VSOCK_DEFINED), 1)
 OBJS += trace-timesync-ptp.o
+OBJS += trace-timesync-kvm.o
 endif
 
 # Additional util objects
diff --git a/lib/trace-cmd/include/trace-tsync-local.h b/lib/trace-cmd/include/trace-tsync-local.h
index b3a3bdfc..7e56d187 100644
--- a/lib/trace-cmd/include/trace-tsync-local.h
+++ b/lib/trace-cmd/include/trace-tsync-local.h
@@ -65,5 +65,6 @@ int tracecmd_tsync_proto_register(const char *proto_name, int accuracy, int role
 int tracecmd_tsync_proto_unregister(char *proto_name);
 
 int ptp_clock_sync_register(void);
+int kvm_clock_sync_register(void);
 
 #endif /* _TRACE_TSYNC_LOCAL_H */
diff --git a/lib/trace-cmd/trace-timesync-kvm.c b/lib/trace-cmd/trace-timesync-kvm.c
new file mode 100644
index 00000000..03a4fd73
--- /dev/null
+++ b/lib/trace-cmd/trace-timesync-kvm.c
@@ -0,0 +1,460 @@
+// SPDX-License-Identifier: LGPL-2.1
+/*
+ * Copyright (C) 2020, VMware, Tzvetomir Stoyanov tz.stoyanov@gmail.com>
+ *
+ */
+
+#include <fcntl.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <sys/stat.h>
+#include <dirent.h>
+
+#include "trace-cmd.h"
+#include "trace-cmd-private.h"
+#include "tracefs.h"
+#include "trace-tsync-local.h"
+
+#define KVM_DEBUG_FS "/sys/kernel/debug/kvm"
+#define KVM_DEBUG_OFFSET_FILE	"tsc-offset"
+#define KVM_DEBUG_SCALING_FILE	"tsc-scaling-ratio"
+#define KVM_DEBUG_VCPU_DIR	"vcpu"
+
+#define KVM_SYNC_PKT_REQUEST	1
+#define KVM_SYNC_PKT_RESPONSE	2
+
+typedef __s64 s64;
+
+// equal to /sys/kernel/debug/kvm/<VM ID>/vcpu0/tsc-offset
+#define KVM_ACCURACY	0
+#define KVM_NAME	"kvm"
+
+struct kvm_clock_sync {
+	int vcpu_count;
+	char **vcpu_offsets;
+	char **vcpu_scalings;
+	int marker_fd;
+	struct tep_handle *tep;
+	int raw_id;
+	unsigned long long ts;
+};
+
+struct kvm_clock_offset_msg {
+	s64	ts;
+	s64	offset;
+	s64	scaling;
+};
+
+static bool kvm_support_check(bool guest)
+{
+	struct stat st;
+	int ret;
+
+	if (guest)
+		return true;
+
+	ret = stat(KVM_DEBUG_FS, &st);
+	if (ret < 0)
+		return false;
+
+	if (!S_ISDIR(st.st_mode))
+		return false;
+	return true;
+}
+
+static int kvm_open_vcpu_dir(struct kvm_clock_sync *kvm, int cpu, char *dir_str)
+{
+	struct dirent *entry;
+	char path[PATH_MAX];
+	DIR *dir;
+
+	dir = opendir(dir_str);
+	if (!dir)
+		goto error;
+	while ((entry = readdir(dir))) {
+		if (entry->d_type != DT_DIR) {
+			if (!strncmp(entry->d_name, KVM_DEBUG_OFFSET_FILE,
+				     strlen(KVM_DEBUG_OFFSET_FILE))) {
+				snprintf(path, sizeof(path), "%s/%s",
+					 dir_str, entry->d_name);
+				kvm->vcpu_offsets[cpu] = strdup(path);
+			}
+			if (!strncmp(entry->d_name, KVM_DEBUG_SCALING_FILE,
+				     strlen(KVM_DEBUG_SCALING_FILE))) {
+				snprintf(path, sizeof(path), "%s/%s",
+					 dir_str, entry->d_name);
+				kvm->vcpu_scalings[cpu] = strdup(path);
+			}
+		}
+	}
+	if (!kvm->vcpu_offsets[cpu])
+		goto error;
+	closedir(dir);
+	return 0;
+
+error:
+	if (dir)
+		closedir(dir);
+	free(kvm->vcpu_offsets[cpu]);
+	kvm->vcpu_offsets[cpu] = NULL;
+	free(kvm->vcpu_scalings[cpu]);
+	kvm->vcpu_scalings[cpu] = NULL;
+	return -1;
+}
+
+static int kvm_open_debug_files(struct kvm_clock_sync *kvm, int pid)
+{
+	char *vm_dir_str = NULL;
+	struct dirent *entry;
+	char *pid_str = NULL;
+	char path[PATH_MAX];
+	long vcpu;
+	DIR *dir;
+	int i;
+
+	dir = opendir(KVM_DEBUG_FS);
+	if (!dir)
+		goto error;
+	if (asprintf(&pid_str, "%d-", pid) <= 0)
+		goto error;
+	while ((entry = readdir(dir))) {
+		if (!(entry->d_type == DT_DIR &&
+		    !strncmp(entry->d_name, pid_str, strlen(pid_str))))
+			continue;
+		asprintf(&vm_dir_str, "%s/%s", KVM_DEBUG_FS, entry->d_name);
+		break;
+	}
+	closedir(dir);
+	dir = NULL;
+	if (!vm_dir_str)
+		goto error;
+	dir = opendir(vm_dir_str);
+	if (!dir)
+		goto error;
+	while ((entry = readdir(dir))) {
+		if (!(entry->d_type == DT_DIR &&
+		    !strncmp(entry->d_name, KVM_DEBUG_VCPU_DIR, strlen(KVM_DEBUG_VCPU_DIR))))
+			continue;
+		vcpu =  strtol(entry->d_name + strlen(KVM_DEBUG_VCPU_DIR), NULL, 10);
+		if (vcpu < 0 || vcpu >= kvm->vcpu_count)
+			continue;
+		snprintf(path, sizeof(path), "%s/%s", vm_dir_str, entry->d_name);
+		if (kvm_open_vcpu_dir(kvm, vcpu, path) < 0)
+			goto error;
+	}
+	for (i = 0; i < kvm->vcpu_count; i++) {
+		if (!kvm->vcpu_offsets[i])
+			goto error;
+	}
+	closedir(dir);
+	free(pid_str);
+	free(vm_dir_str);
+	return 0;
+error:
+	free(pid_str);
+	free(vm_dir_str);
+	if (dir)
+		closedir(dir);
+	return -1;
+}
+
+static int kvm_clock_sync_init_host(struct tracecmd_time_sync *tsync,
+				    struct kvm_clock_sync *kvm)
+{
+	kvm->vcpu_count = tsync->vcpu_count;
+	kvm->vcpu_offsets = calloc(kvm->vcpu_count, sizeof(char *));
+	kvm->vcpu_scalings = calloc(kvm->vcpu_count, sizeof(char *));
+	if (!kvm->vcpu_offsets || !kvm->vcpu_scalings)
+		goto error;
+	if (kvm_open_debug_files(kvm, tsync->guest_pid) < 0)
+		goto error;
+	return 0;
+
+error:
+	free(kvm->vcpu_offsets);
+	free(kvm->vcpu_scalings);
+	return -1;
+}
+
+static int kvm_clock_sync_init_guest(struct tracecmd_time_sync *tsync,
+				     struct kvm_clock_sync *kvm)
+{
+	const char *systems[] = {"ftrace", NULL};
+	struct clock_sync_context *clock_context;
+	struct tep_event *raw;
+	char *path;
+
+	clock_context = (struct clock_sync_context *)tsync->context;
+	path = tracefs_instance_get_dir(clock_context->instance);
+	if (!path)
+		goto error;
+	kvm->tep = tracefs_local_events_system(path, systems);
+	tracefs_put_tracing_file(path);
+	if (!kvm->tep)
+		goto error;
+	raw = tep_find_event_by_name(kvm->tep, "ftrace", "raw_data");
+	if (!raw)
+		goto error;
+
+	kvm->raw_id = raw->id;
+	tep_set_file_bigendian(kvm->tep, tracecmd_host_bigendian());
+	tep_set_local_bigendian(kvm->tep, tracecmd_host_bigendian());
+
+	path = tracefs_instance_get_file(clock_context->instance, "trace_marker_raw");
+	if (!path)
+		goto error;
+	kvm->marker_fd = open(path, O_WRONLY);
+	tracefs_put_tracing_file(path);
+
+	return 0;
+
+error:
+	if (kvm->tep)
+		tep_free(kvm->tep);
+	if (kvm->marker_fd >= 0)
+		close(kvm->marker_fd);
+
+	return -1;
+}
+
+static int kvm_clock_sync_init(struct tracecmd_time_sync *tsync)
+{
+	struct clock_sync_context *clock_context;
+	struct kvm_clock_sync *kvm;
+	int ret;
+
+	if (!tsync || !tsync->context)
+		return -1;
+	clock_context = (struct clock_sync_context *)tsync->context;
+
+	if (!kvm_support_check(clock_context->is_guest))
+		return -1;
+	kvm = calloc(1, sizeof(struct kvm_clock_sync));
+	if (!kvm)
+		return -1;
+	kvm->marker_fd = -1;
+	if (clock_context->is_guest)
+		ret = kvm_clock_sync_init_guest(tsync, kvm);
+	else
+		ret = kvm_clock_sync_init_host(tsync, kvm);
+	if (ret < 0)
+		goto error;
+
+	clock_context->proto_data = kvm;
+	return 0;
+
+error:
+	free(kvm);
+	return -1;
+}
+
+static int kvm_clock_sync_free(struct tracecmd_time_sync *tsync)
+{
+	struct clock_sync_context *clock_context;
+	struct kvm_clock_sync *kvm = NULL;
+	int i;
+
+	clock_context = (struct clock_sync_context *)tsync->context;
+	if (clock_context)
+		kvm = (struct kvm_clock_sync *)clock_context->proto_data;
+	if (kvm) {
+		for (i = 0; i < kvm->vcpu_count; i++) {
+			free(kvm->vcpu_offsets[i]);
+			kvm->vcpu_offsets[i] = NULL;
+			free(kvm->vcpu_scalings[i]);
+			kvm->vcpu_scalings[i] = NULL;
+		}
+		if (kvm->tep)
+			tep_free(kvm->tep);
+		if (kvm->marker_fd >= 0)
+			close(kvm->marker_fd);
+		free(kvm);
+	}
+	return -1;
+}
+
+static int read_ll_form_file(char *file, long long *res)
+{
+	char buf[32];
+	int ret;
+	int fd;
+
+	if (!file)
+		return -1;
+	fd = open(file, O_RDONLY | O_NONBLOCK);
+	if (fd < 0)
+		return -1;
+	ret = read(fd, buf, 32);
+	close(fd);
+	if (ret <= 0)
+		return -1;
+
+	*res = strtoll(buf, NULL, 10);
+
+	return 0;
+}
+
+static int kvm_clock_host(struct tracecmd_time_sync *tsync,
+			  long long *offset, long long *scaling,
+			  long long *timestamp, unsigned int cpu)
+{
+	char sync_proto[TRACECMD_TSYNC_PNAME_LENGTH];
+	struct clock_sync_context *clock_context;
+	struct kvm_clock_offset_msg packet;
+	struct kvm_clock_sync *kvm = NULL;
+	long long kvm_scaling = 1;
+	unsigned int sync_msg;
+	long long kvm_offset;
+	unsigned int size;
+	char *msg;
+	int ret;
+
+	clock_context = (struct clock_sync_context *)tsync->context;
+	if (clock_context)
+		kvm = (struct kvm_clock_sync *)clock_context->proto_data;
+	if (!kvm || !kvm->vcpu_offsets || !kvm->vcpu_offsets[0])
+		return -1;
+	if (cpu >= kvm->vcpu_count)
+		return -1;
+	ret = read_ll_form_file(kvm->vcpu_offsets[cpu], &kvm_offset);
+	if (ret < 0)
+		return -1;
+	if (kvm->vcpu_scalings && kvm->vcpu_scalings[cpu])
+		read_ll_form_file(kvm->vcpu_scalings[cpu], &kvm_scaling);
+	msg = (char *)&packet;
+	size = sizeof(packet);
+	ret = tracecmd_msg_recv_time_sync(tsync->msg_handle,
+					  sync_proto, &sync_msg,
+					  &size, &msg);
+	if (ret || strncmp(sync_proto, KVM_NAME, TRACECMD_TSYNC_PNAME_LENGTH) ||
+	    sync_msg != KVM_SYNC_PKT_REQUEST)
+		return -1;
+
+	packet.offset = -kvm_offset;
+	packet.scaling = kvm_scaling;
+	ret = tracecmd_msg_send_time_sync(tsync->msg_handle, KVM_NAME,
+					  KVM_SYNC_PKT_RESPONSE, sizeof(packet),
+					  (char *)&packet);
+	if (ret)
+		return -1;
+
+	*scaling = packet.scaling;
+	*offset = packet.offset;
+	*timestamp = packet.ts;
+
+	return 0;
+}
+
+#define KVM_EVENT_MARKER	"kvm sync event"
+static int kvm_marker_find(struct tep_event *event, struct tep_record *record,
+			   int cpu, void *context)
+{
+	struct kvm_clock_sync *kvm = (struct kvm_clock_sync *)context;
+	struct tep_format_field *field;
+	struct tep_format_field *id;
+	char *marker;
+
+	/* Make sure this is our event */
+	if (event->id != kvm->raw_id)
+		return 0;
+	id = tep_find_field(event, "id");
+	field = tep_find_field(event, "buf");
+	if (field && id &&
+	    record->size >= (id->offset + strlen(KVM_EVENT_MARKER) + 1)) {
+		marker = (char *)(record->data + id->offset);
+		if (!strcmp(marker, KVM_EVENT_MARKER)) {
+			kvm->ts = record->ts;
+			return 1;
+		}
+	}
+
+	return 0;
+}
+
+
+static int kvm_clock_guest(struct tracecmd_time_sync *tsync,
+			   long long *offset,
+			   long long *scaling,
+			   long long *timestamp)
+{
+	char sync_proto[TRACECMD_TSYNC_PNAME_LENGTH];
+	struct clock_sync_context *clock_context;
+	struct kvm_clock_offset_msg packet;
+	struct kvm_clock_sync *kvm = NULL;
+	unsigned int sync_msg;
+	unsigned int size;
+	char *msg;
+	int ret;
+
+	clock_context = (struct clock_sync_context *)tsync->context;
+	if (clock_context)
+		kvm = (struct kvm_clock_sync *)clock_context->proto_data;
+	if (!kvm)
+		return -1;
+	kvm->ts = 0;
+	memset(&packet, 0, sizeof(packet));
+	tracefs_instance_file_write(clock_context->instance, "trace", "\0");
+	write(kvm->marker_fd, KVM_EVENT_MARKER, strlen(KVM_EVENT_MARKER) + 1);
+	kvm->ts = 0;
+	tracefs_iterate_raw_events(kvm->tep, clock_context->instance,
+				   NULL, 0, kvm_marker_find, kvm);
+	packet.ts = kvm->ts;
+	ret = tracecmd_msg_send_time_sync(tsync->msg_handle, KVM_NAME,
+					  KVM_SYNC_PKT_REQUEST, sizeof(packet),
+					  (char *)&packet);
+	if (ret)
+		return -1;
+	msg = (char *)&packet;
+	size = sizeof(packet);
+	ret = tracecmd_msg_recv_time_sync(tsync->msg_handle,
+					  sync_proto, &sync_msg,
+					  &size, &msg);
+	if (ret || strncmp(sync_proto, KVM_NAME, TRACECMD_TSYNC_PNAME_LENGTH) ||
+	    sync_msg != KVM_SYNC_PKT_RESPONSE)
+		return -1;
+
+	*scaling = packet.scaling;
+	*offset = packet.offset;
+	*timestamp = packet.ts;
+	return 0;
+}
+
+static int kvm_clock_sync_calc(struct tracecmd_time_sync *tsync,
+			       long long *offset, long long *scaling,
+			       long long *timestamp, unsigned int cpu)
+{
+	struct clock_sync_context *clock_context;
+	int ret;
+
+	if (!tsync || !tsync->context)
+		return -1;
+
+	clock_context = (struct clock_sync_context *)tsync->context;
+
+	if (clock_context->is_guest)
+		ret = kvm_clock_guest(tsync, offset, scaling, timestamp);
+	else
+		ret = kvm_clock_host(tsync, offset, scaling, timestamp, cpu);
+	return ret;
+}
+
+int kvm_clock_sync_register(void)
+{
+	int role = TRACECMD_TIME_SYNC_ROLE_GUEST;
+	int clock = 0;
+
+	if (kvm_support_check(false)) {
+		role |= TRACECMD_TIME_SYNC_ROLE_HOST;
+		clock = TRACECMD_CLOCK_X86_TSC;
+	}
+	return tracecmd_tsync_proto_register(KVM_NAME, KVM_ACCURACY,
+					     role, clock, 0,
+					     kvm_clock_sync_init,
+					     kvm_clock_sync_free,
+					     kvm_clock_sync_calc);
+}
+
+int kvm_clock_sync_unregister(void)
+{
+	return tracecmd_tsync_proto_unregister(KVM_NAME);
+}
diff --git a/lib/trace-cmd/trace-timesync.c b/lib/trace-cmd/trace-timesync.c
index 588484e7..81d6acb4 100644
--- a/lib/trace-cmd/trace-timesync.c
+++ b/lib/trace-cmd/trace-timesync.c
@@ -66,6 +66,7 @@ void tracecmd_tsync_init(void)
 {
 #ifdef VSOCK
 	ptp_clock_sync_register();
+	kvm_clock_sync_register();
 #endif
 }
 
-- 
2.29.2


^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session
  2021-03-15  6:18 [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Tzvetomir Stoyanov (VMware)
                   ` (4 preceding siblings ...)
  2021-03-15  6:18 ` [PATCH v32 5/5] trace-cmd [POC]: Add KVM timestamp synchronization plugin Tzvetomir Stoyanov (VMware)
@ 2021-03-19 10:08 ` Stefano De Venuto
  2021-03-19 11:55   ` Tzvetomir Stoyanov
  5 siblings, 1 reply; 15+ messages in thread
From: Stefano De Venuto @ 2021-03-19 10:08 UTC (permalink / raw)
  To: Tzvetomir Stoyanov (VMware), rostedt; +Cc: linux-trace-devel, Dario Faggioli

Hello,

On 3/15/21 7:18 AM, Tzvetomir Stoyanov (VMware) wrote:
> Basic infrastructure for host - guest timestamp synchronization and a
> PoC implementation of PTP-like and KVM algorithms. 
I'm trying to make some analysis on traces generated by host and guest,
so I applied this series and the related dependencies.

However, when the KVM algorithm is used, the merging process
of the two traces doesn't seem to happen properly.

The output of `trace-cmd report -i trace.dat -i trace-tumbleweed.dat` is:

           trace.dat: cpus=4
trace-tumbleweed.dat: cpus=1
           trace.dat:            sleep-23948 [002]61124758463654:
write_msr:            c0000100, value 7f35092875c0
           trace.dat:           <idle>-0     [000]61124760841508:
write_msr:            c0000100, value 7f6136b75640
           trace.dat:        CPU 0/KVM-1640  [000]61124760851620:
write_msr:            c0011020, value 40000000000000
           trace.dat:        CPU 0/KVM-1640  [000]61124760861778:
read_msr:             c0000101, value ffff9b1b76c00000
           trace.dat:        CPU 0/KVM-1640  [000]61124760862832:
read_msr:             c0000081, value 23001000000000
           trace.dat:        CPU 0/KVM-1640  [000]61124760864347:
read_msr:             c0000082, value ffffffffb5400010
           trace.dat:        CPU 0/KVM-1640  [000]61124760865225:
read_msr:             c0000083, value ffffffffb5401670
           .
           .
           .
trace-tumbleweed.dat:   IPC I/O Parent-4313 [000]8446232460209009653:
write_msr:            c0000100, value 7f94535c7640
trace-tumbleweed.dat:    IPC I/O Child-4503 [000]8446348845232796661:
write_msr:            c001011f, value 4
trace-tumbleweed.dat:    IPC I/O Child-4503 [000]8447532508154803189:
write_msr:            c0000100, value 7f945e8d5780
trace-tumbleweed.dat:           <idle>-0    [000]8452336523564659701:
write_msr:            c001011f, value 0
trace-tumbleweed.dat:           <idle>-0    [000]8452515224268944373:
write_msr:            6e0, value 731478a3bf62
trace-tumbleweed.dat:           <idle>-0    [000]8452610851715789813:
write_msr:            6e0, value 731478968f37
trace-tumbleweed.dat:           <idle>-0    [000]8555757214338495477:
write_msr:            6e0, value 731478853ec1
trace-tumbleweed.dat:           <idle>-0    [000]8555904475882173429:
write_msr:            6e0, value 7314778a60e8
trace-tumbleweed.dat:      kworker/0:1-7351 [000]8556355589182174197:
write_msr:            c0000100, value 7f8373573740

The TimeShift information is inside the guest trace file,
so I think this is not a transfer problem.

I also tried to remove the last patch, the one related to the KVM clock,
in order to use PTP and the two traces are merged correctly.

The commands used to record are:

Host:
# trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
msr:* -C x86-tsc sleep 1

Guest:
# echo x86-tsc > /sys/kernel/tracing/trace_clock
# trace-cmd agent

Am I doing something wrong?

If necessary, I can provide more info about my setup, or do more tests.


Thanks and Regards,

Stefano

^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session
  2021-03-19 10:08 ` [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Stefano De Venuto
@ 2021-03-19 11:55   ` Tzvetomir Stoyanov
  2021-03-19 17:44     ` Stefano De Venuto
  0 siblings, 1 reply; 15+ messages in thread
From: Tzvetomir Stoyanov @ 2021-03-19 11:55 UTC (permalink / raw)
  To: Stefano De Venuto; +Cc: Steven Rostedt, Linux Trace Devel, Dario Faggioli

Hi Stefano,

On Fri, Mar 19, 2021 at 12:08 PM Stefano De Venuto
<stefano.devenuto99@gmail.com> wrote:
>
> Hello,
>
> On 3/15/21 7:18 AM, Tzvetomir Stoyanov (VMware) wrote:
> > Basic infrastructure for host - guest timestamp synchronization and a
> > PoC implementation of PTP-like and KVM algorithms.
> I'm trying to make some analysis on traces generated by host and guest,
> so I applied this series and the related dependencies.
>
> However, when the KVM algorithm is used, the merging process
> of the two traces doesn't seem to happen properly.
>
> The output of `trace-cmd report -i trace.dat -i trace-tumbleweed.dat` is:
>
>            trace.dat: cpus=4
> trace-tumbleweed.dat: cpus=1
>            trace.dat:            sleep-23948 [002]61124758463654:
> write_msr:            c0000100, value 7f35092875c0
>            trace.dat:           <idle>-0     [000]61124760841508:
> write_msr:            c0000100, value 7f6136b75640
>            trace.dat:        CPU 0/KVM-1640  [000]61124760851620:
> write_msr:            c0011020, value 40000000000000
>            trace.dat:        CPU 0/KVM-1640  [000]61124760861778:
> read_msr:             c0000101, value ffff9b1b76c00000
>            trace.dat:        CPU 0/KVM-1640  [000]61124760862832:
> read_msr:             c0000081, value 23001000000000
>            trace.dat:        CPU 0/KVM-1640  [000]61124760864347:
> read_msr:             c0000082, value ffffffffb5400010
>            trace.dat:        CPU 0/KVM-1640  [000]61124760865225:
> read_msr:             c0000083, value ffffffffb5401670
>            .
>            .
>            .
> trace-tumbleweed.dat:   IPC I/O Parent-4313 [000]8446232460209009653:
> write_msr:            c0000100, value 7f94535c7640
> trace-tumbleweed.dat:    IPC I/O Child-4503 [000]8446348845232796661:
> write_msr:            c001011f, value 4
> trace-tumbleweed.dat:    IPC I/O Child-4503 [000]8447532508154803189:
> write_msr:            c0000100, value 7f945e8d5780
> trace-tumbleweed.dat:           <idle>-0    [000]8452336523564659701:
> write_msr:            c001011f, value 0
> trace-tumbleweed.dat:           <idle>-0    [000]8452515224268944373:
> write_msr:            6e0, value 731478a3bf62
> trace-tumbleweed.dat:           <idle>-0    [000]8452610851715789813:
> write_msr:            6e0, value 731478968f37
> trace-tumbleweed.dat:           <idle>-0    [000]8555757214338495477:
> write_msr:            6e0, value 731478853ec1
> trace-tumbleweed.dat:           <idle>-0    [000]8555904475882173429:
> write_msr:            6e0, value 7314778a60e8
> trace-tumbleweed.dat:      kworker/0:1-7351 [000]8556355589182174197:
> write_msr:            c0000100, value 7f8373573740
>
> The TimeShift information is inside the guest trace file,
> so I think this is not a transfer problem.
>
> I also tried to remove the last patch, the one related to the KVM clock,
> in order to use PTP and the two traces are merged correctly.
>
> The commands used to record are:
>
> Host:
> # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
> msr:* -C x86-tsc sleep 1

The guest trace clock is set automatically as the host, so this
command should be enough:
# trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
msr:*  sleep 1

>
> Guest:
> # echo x86-tsc > /sys/kernel/tracing/trace_clock

There is no need to set manually the guest clock, it will be
overwritten by trace-cmd agent.

> # trace-cmd agent
>
> Am I doing something wrong?

All your commands and workflow look good, nothing wrong.

>
> If necessary, I can provide more info about my setup, or do more tests.

Yes, please can you send me both host and guest trace files ?
Also, it will be useful to send me the content of the KVM debug files:
    /sys/kernel/debug/kvm/<guest ID>/vcpu<*>/tsc-offset

>
>
> Thanks and Regards,
>
> Stefano

Thanks for testing this code!

-- 
Tzvetomir (Ceco) Stoyanov
VMware Open Source Technology Center

^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session
  2021-03-19 11:55   ` Tzvetomir Stoyanov
@ 2021-03-19 17:44     ` Stefano De Venuto
  2021-03-20  6:25       ` Tzvetomir Stoyanov
  0 siblings, 1 reply; 15+ messages in thread
From: Stefano De Venuto @ 2021-03-19 17:44 UTC (permalink / raw)
  To: Tzvetomir Stoyanov; +Cc: Steven Rostedt, Linux Trace Devel, Dario Faggioli



On 3/19/21 12:55 PM, Tzvetomir Stoyanov wrote:
> Hi Stefano,
>
> On Fri, Mar 19, 2021 at 12:08 PM Stefano De Venuto
> <stefano.devenuto99@gmail.com> wrote:
Hi!
>> The commands used to record are:
>>
>> Host:
>> # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
>> msr:* -C x86-tsc sleep 1
> The guest trace clock is set automatically as the host, so this
> command should be enough:
> # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
> msr:*  sleep 1
>
>> Guest:
>> # echo x86-tsc > /sys/kernel/tracing/trace_clock
> There is no need to set manually the guest clock, it will be
> overwritten by trace-cmd agent.
>
Thanks so much for the proper way to do it, really appreciated.
>> If necessary, I can provide more info about my setup, or do more tests.
> Yes, please can you send me both host and guest trace files ?
Here are the trace files, host and guest respectively:

- http://xenbits.xen.org/people/dariof/tracing-examples/kvm/sync-kvmclock/trace.dat
- http://xenbits.xen.org/people/dariof/tracing-examples/kvm/sync-kvmclock/trace-tumbleweed.dat

> Also, it will be useful to send me the content of the KVM debug files:
>     /sys/kernel/debug/kvm/<guest ID>/vcpu<*>/tsc-offset
The guest has one vcpu (vcpu0) and the content of the file is:

255647917761327
>>
>> Thanks and Regards,
>>
>> Stefano
> Thanks for testing this code!
>
Thanks for your time,

Stefano

^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session
  2021-03-19 17:44     ` Stefano De Venuto
@ 2021-03-20  6:25       ` Tzvetomir Stoyanov
       [not found]         ` <0f6e158c-43a3-4117-2a17-5a13cd6a7970@gmail.com>
  0 siblings, 1 reply; 15+ messages in thread
From: Tzvetomir Stoyanov @ 2021-03-20  6:25 UTC (permalink / raw)
  To: Stefano De Venuto; +Cc: Steven Rostedt, Linux Trace Devel, Dario Faggioli

Hi Stefano,

On Fri, Mar 19, 2021 at 7:44 PM Stefano De Venuto
<stefano.devenuto99@gmail.com> wrote:
>
>
>
> On 3/19/21 12:55 PM, Tzvetomir Stoyanov wrote:
> > Hi Stefano,
> >
> > On Fri, Mar 19, 2021 at 12:08 PM Stefano De Venuto
> > <stefano.devenuto99@gmail.com> wrote:
> Hi!
> >> The commands used to record are:
> >>
> >> Host:
> >> # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
> >> msr:* -C x86-tsc sleep 1
> > The guest trace clock is set automatically as the host, so this
> > command should be enough:
> > # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
> > msr:*  sleep 1
> >
> >> Guest:
> >> # echo x86-tsc > /sys/kernel/tracing/trace_clock
> > There is no need to set manually the guest clock, it will be
> > overwritten by trace-cmd agent.
> >
> Thanks so much for the proper way to do it, really appreciated.
> >> If necessary, I can provide more info about my setup, or do more tests.
> > Yes, please can you send me both host and guest trace files ?
> Here are the trace files, host and guest respectively:
>
> - http://xenbits.xen.org/people/dariof/tracing-examples/kvm/sync-kvmclock/trace.dat
> - http://xenbits.xen.org/people/dariof/tracing-examples/kvm/sync-kvmclock/trace-tumbleweed.dat
>
> > Also, it will be useful to send me the content of the KVM debug files:
> >     /sys/kernel/debug/kvm/<guest ID>/vcpu<*>/tsc-offset
> The guest has one vcpu (vcpu0) and the content of the file is:
>
> 255647917761327

Looks like there is a scaling between host and guest clocks in your
setup, not just a simple offset. We did not test yet our
implementation with scaling, although both offset and scaling are part
of the calculations. That makes your use case very valuable for us, as
we have an opportunity to test it now. And yes, looks like we have a
bug here.
Please, when you have time, can you repeat again the tracing session
and send again both trace files + the content of the KVM debug files:
     /sys/kernel/debug/kvm/<guest ID>/vcpu0/tsc-offset
     /sys/kernel/debug/kvm/<guest ID>/vcpu0/tsc-scaling-ratio
I'm asking to do a new trace, as most probably these offset and
scaling could be different now.

Thanks!

> >>
> >> Thanks and Regards,
> >>
> >> Stefano
> > Thanks for testing this code!
> >
> Thanks for your time,
>
> Stefano



-- 
Tzvetomir (Ceco) Stoyanov
VMware Open Source Technology Center

^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session
       [not found]         ` <0f6e158c-43a3-4117-2a17-5a13cd6a7970@gmail.com>
@ 2021-03-22 10:13           ` Tzvetomir Stoyanov
  2021-03-22 14:56             ` Steven Rostedt
  2021-03-25  7:20           ` Tzvetomir Stoyanov
  1 sibling, 1 reply; 15+ messages in thread
From: Tzvetomir Stoyanov @ 2021-03-22 10:13 UTC (permalink / raw)
  To: Stefano De Venuto; +Cc: Steven Rostedt, Linux Trace Devel, Dario Faggioli

On Sat, Mar 20, 2021 at 6:17 PM Stefano De Venuto
<stefano.devenuto99@gmail.com> wrote:
>
>
>
> On 3/20/21 7:25 AM, Tzvetomir Stoyanov wrote:
> > Hi Stefano,
> >
> > On Fri, Mar 19, 2021 at 7:44 PM Stefano De Venuto
> > <stefano.devenuto99@gmail.com> wrote:
> >>
> >>
> >> On 3/19/21 12:55 PM, Tzvetomir Stoyanov wrote:
> >>> Hi Stefano,
> >>>
> >>> On Fri, Mar 19, 2021 at 12:08 PM Stefano De Venuto
> >>> <stefano.devenuto99@gmail.com> wrote:
> >> Hi!
> >>>> The commands used to record are:
> >>>>
> >>>> Host:
> >>>> # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
> >>>> msr:* -C x86-tsc sleep 1
> >>> The guest trace clock is set automatically as the host, so this
> >>> command should be enough:
> >>> # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
> >>> msr:*  sleep 1
> >>>
> >>>> Guest:
> >>>> # echo x86-tsc > /sys/kernel/tracing/trace_clock
> >>> There is no need to set manually the guest clock, it will be
> >>> overwritten by trace-cmd agent.
> >>>
> >> Thanks so much for the proper way to do it, really appreciated.
> >>>> If necessary, I can provide more info about my setup, or do more tests.
> >>> Yes, please can you send me both host and guest trace files ?
> >> Here are the trace files, host and guest respectively:
> >>
> >> - http://xenbits.xen.org/people/dariof/tracing-examples/kvm/sync-kvmclock/trace.dat
> >> - http://xenbits.xen.org/people/dariof/tracing-examples/kvm/sync-kvmclock/trace-tumbleweed.dat
> >>
> >>> Also, it will be useful to send me the content of the KVM debug files:
> >>>     /sys/kernel/debug/kvm/<guest ID>/vcpu<*>/tsc-offset
> >> The guest has one vcpu (vcpu0) and the content of the file is:
> >>
> >> 255647917761327
> > Looks like there is a scaling between host and guest clocks in your
> > setup, not just a simple offset. We did not test yet our
> > implementation with scaling, although both offset and scaling are part
> > of the calculations. That makes your use case very valuable for us, as
> > we have an opportunity to test it now. And yes, looks like we have a
> > bug here.
> > Please, when you have time, can you repeat again the tracing session
> > and send again both trace files + the content of the KVM debug files:
> >      /sys/kernel/debug/kvm/<guest ID>/vcpu0/tsc-offset
> >      /sys/kernel/debug/kvm/<guest ID>/vcpu0/tsc-scaling-ratio
> > I'm asking to do a new trace, as most probably these offset and
> > scaling could be different now.
> Yes, the trace files are attached to this mail.
>
> The content of tsc-offset is:
> 453568564244284
>
> The content of tsc-scaling-ratio is:
> 4294967296
>

Confirmed, the problem is in the trace-cmd logic that works with the
KVM scaling.  Looks like tsc-scaling-ratio is the default KVM scaling
on your machine, (1 << 32), and in that case the scaling should be
ignored in our calculations. I still have no solution, going to
analyze how the KVM scaling works.

Thanks, Stafano!

>
> Thanks and Regards,
>
> Stefano
>
> >
> > Thanks!
> >
> >>>> Thanks and Regards,
> >>>>
> >>>> Stefano
> >>> Thanks for testing this code!
> >>>
> >> Thanks for your time,
> >>
> >> Stefano
> >
> >
>


-- 
Tzvetomir (Ceco) Stoyanov
VMware Open Source Technology Center

^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session
  2021-03-22 10:13           ` Tzvetomir Stoyanov
@ 2021-03-22 14:56             ` Steven Rostedt
  2021-03-25 20:55               ` Dario Faggioli
  0 siblings, 1 reply; 15+ messages in thread
From: Steven Rostedt @ 2021-03-22 14:56 UTC (permalink / raw)
  To: Tzvetomir Stoyanov; +Cc: Stefano De Venuto, Linux Trace Devel, Dario Faggioli

On Mon, 22 Mar 2021 12:13:09 +0200
Tzvetomir Stoyanov <tz.stoyanov@gmail.com> wrote:

> Confirmed, the problem is in the trace-cmd logic that works with the
> KVM scaling.  Looks like tsc-scaling-ratio is the default KVM scaling
> on your machine, (1 << 32), and in that case the scaling should be
> ignored in our calculations. I still have no solution, going to
> analyze how the KVM scaling works.

I'm guessing that this was run on a AMD CPU, as the default scaling (the
one to ignore) is 1 << 32 for AMD and for Intel it's 1 << 48.

-- Steve

^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session
       [not found]         ` <0f6e158c-43a3-4117-2a17-5a13cd6a7970@gmail.com>
  2021-03-22 10:13           ` Tzvetomir Stoyanov
@ 2021-03-25  7:20           ` Tzvetomir Stoyanov
  2021-03-25 18:01             ` Stefano De Venuto
  1 sibling, 1 reply; 15+ messages in thread
From: Tzvetomir Stoyanov @ 2021-03-25  7:20 UTC (permalink / raw)
  To: Stefano De Venuto; +Cc: Steven Rostedt, Linux Trace Devel, Dario Faggioli

Hi Stefano,

On Sat, Mar 20, 2021 at 6:17 PM Stefano De Venuto
<stefano.devenuto99@gmail.com> wrote:
>
>
>
> On 3/20/21 7:25 AM, Tzvetomir Stoyanov wrote:
> > Hi Stefano,
> >
> > On Fri, Mar 19, 2021 at 7:44 PM Stefano De Venuto
> > <stefano.devenuto99@gmail.com> wrote:
> >>
> >>
> >> On 3/19/21 12:55 PM, Tzvetomir Stoyanov wrote:
> >>> Hi Stefano,
> >>>
> >>> On Fri, Mar 19, 2021 at 12:08 PM Stefano De Venuto
> >>> <stefano.devenuto99@gmail.com> wrote:
> >> Hi!
> >>>> The commands used to record are:
> >>>>
> >>>> Host:
> >>>> # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
> >>>> msr:* -C x86-tsc sleep 1
> >>> The guest trace clock is set automatically as the host, so this
> >>> command should be enough:
> >>> # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
> >>> msr:*  sleep 1
> >>>
> >>>> Guest:
> >>>> # echo x86-tsc > /sys/kernel/tracing/trace_clock
> >>> There is no need to set manually the guest clock, it will be
> >>> overwritten by trace-cmd agent.
> >>>
> >> Thanks so much for the proper way to do it, really appreciated.
> >>>> If necessary, I can provide more info about my setup, or do more tests.
> >>> Yes, please can you send me both host and guest trace files ?
> >> Here are the trace files, host and guest respectively:
> >>
> >> - http://xenbits.xen.org/people/dariof/tracing-examples/kvm/sync-kvmclock/trace.dat
> >> - http://xenbits.xen.org/people/dariof/tracing-examples/kvm/sync-kvmclock/trace-tumbleweed.dat
> >>
> >>> Also, it will be useful to send me the content of the KVM debug files:
> >>>     /sys/kernel/debug/kvm/<guest ID>/vcpu<*>/tsc-offset
> >> The guest has one vcpu (vcpu0) and the content of the file is:
> >>
> >> 255647917761327
> > Looks like there is a scaling between host and guest clocks in your
> > setup, not just a simple offset. We did not test yet our
> > implementation with scaling, although both offset and scaling are part
> > of the calculations. That makes your use case very valuable for us, as
> > we have an opportunity to test it now. And yes, looks like we have a
> > bug here.
> > Please, when you have time, can you repeat again the tracing session
> > and send again both trace files + the content of the KVM debug files:
> >      /sys/kernel/debug/kvm/<guest ID>/vcpu0/tsc-offset
> >      /sys/kernel/debug/kvm/<guest ID>/vcpu0/tsc-scaling-ratio
> > I'm asking to do a new trace, as most probably these offset and
> > scaling could be different now.
> Yes, the trace files are attached to this mail.
>
> The content of tsc-offset is:
> 453568564244284
>
> The content of tsc-scaling-ratio is:
> 4294967296
>
Just submitted the v33 of the patch set, added a check for a
non-default KVM scaling. Now it should work on your setup, as the KVM
scaling there looks to be the default. I didn't test it, as we have no
machine which supports KVM scaling. This is just a workaround, we have
to implement support to KVM scaling in our algorithm.
Please, can you test if it is OK ? Note, you have to apply  these
patch sets also, as there are dependencies:
   v2 Refactoring and improvements of time sync logic
   v4 TSC trace clock to nanosecond conversion

Thanks for testing this code!

>
> Thanks and Regards,
>
> Stefano
>
> >
> > Thanks!
> >
> >>>> Thanks and Regards,
> >>>>
> >>>> Stefano
> >>> Thanks for testing this code!
> >>>
> >> Thanks for your time,
> >>
> >> Stefano
> >
> >
>


-- 
Tzvetomir (Ceco) Stoyanov
VMware Open Source Technology Center

^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session
  2021-03-25  7:20           ` Tzvetomir Stoyanov
@ 2021-03-25 18:01             ` Stefano De Venuto
  0 siblings, 0 replies; 15+ messages in thread
From: Stefano De Venuto @ 2021-03-25 18:01 UTC (permalink / raw)
  To: Tzvetomir Stoyanov; +Cc: Steven Rostedt, Linux Trace Devel, Dario Faggioli



On 3/25/21 8:20 AM, Tzvetomir Stoyanov wrote:
> Hi Stefano,
>
> On Sat, Mar 20, 2021 at 6:17 PM Stefano De Venuto
> <stefano.devenuto99@gmail.com> wrote:
>>
>>
>> On 3/20/21 7:25 AM, Tzvetomir Stoyanov wrote:
>>> Hi Stefano,
>>>
>>> On Fri, Mar 19, 2021 at 7:44 PM Stefano De Venuto
>>> <stefano.devenuto99@gmail.com> wrote:
>>>>
>>>> On 3/19/21 12:55 PM, Tzvetomir Stoyanov wrote:
>>>>> Hi Stefano,
>>>>>
>>>>> On Fri, Mar 19, 2021 at 12:08 PM Stefano De Venuto
>>>>> <stefano.devenuto99@gmail.com> wrote:
>>>> Hi!
>>>>>> The commands used to record are:
>>>>>>
>>>>>> Host:
>>>>>> # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
>>>>>> msr:* -C x86-tsc sleep 1
>>>>> The guest trace clock is set automatically as the host, so this
>>>>> command should be enough:
>>>>> # trace-cmd record -C x86-tsc -e kvm:* -e msr:* -A tumbleweed:823 -e
>>>>> msr:*  sleep 1
>>>>>
>>>>>> Guest:
>>>>>> # echo x86-tsc > /sys/kernel/tracing/trace_clock
>>>>> There is no need to set manually the guest clock, it will be
>>>>> overwritten by trace-cmd agent.
>>>>>
>>>> Thanks so much for the proper way to do it, really appreciated.
>>>>>> If necessary, I can provide more info about my setup, or do more tests.
>>>>> Yes, please can you send me both host and guest trace files ?
>>>> Here are the trace files, host and guest respectively:
>>>>
>>>> - http://xenbits.xen.org/people/dariof/tracing-examples/kvm/sync-kvmclock/trace.dat
>>>> - http://xenbits.xen.org/people/dariof/tracing-examples/kvm/sync-kvmclock/trace-tumbleweed.dat
>>>>
>>>>> Also, it will be useful to send me the content of the KVM debug files:
>>>>>     /sys/kernel/debug/kvm/<guest ID>/vcpu<*>/tsc-offset
>>>> The guest has one vcpu (vcpu0) and the content of the file is:
>>>>
>>>> 255647917761327
>>> Looks like there is a scaling between host and guest clocks in your
>>> setup, not just a simple offset. We did not test yet our
>>> implementation with scaling, although both offset and scaling are part
>>> of the calculations. That makes your use case very valuable for us, as
>>> we have an opportunity to test it now. And yes, looks like we have a
>>> bug here.
>>> Please, when you have time, can you repeat again the tracing session
>>> and send again both trace files + the content of the KVM debug files:
>>>      /sys/kernel/debug/kvm/<guest ID>/vcpu0/tsc-offset
>>>      /sys/kernel/debug/kvm/<guest ID>/vcpu0/tsc-scaling-ratio
>>> I'm asking to do a new trace, as most probably these offset and
>>> scaling could be different now.
>> Yes, the trace files are attached to this mail.
>>
>> The content of tsc-offset is:
>> 453568564244284
>>
>> The content of tsc-scaling-ratio is:
>> 4294967296
>>
> Just submitted the v33 of the patch set, added a check for a
> non-default KVM scaling. Now it should work on your setup, as the KVM
> scaling there looks to be the default. I didn't test it, as we have no
> machine which supports KVM scaling. This is just a workaround, we have
> to implement support to KVM scaling in our algorithm.
> Please, can you test if it is OK ? Note, you have to apply  these
> patch sets also, as there are dependencies:
>    v2 Refactoring and improvements of time sync logic
>    v4 TSC trace clock to nanosecond conversion
>
> Thanks for testing this code!
I tested the patch series and now works properly on my setup!

Thanks,
Stefano
>
>> Thanks and Regards,
>>
>> Stefano
>>
>>> Thanks!
>>>
>>>>>> Thanks and Regards,
>>>>>>
>>>>>> Stefano
>>>>> Thanks for testing this code!
>>>>>
>>>> Thanks for your time,
>>>>
>>>> Stefano
>>>
>


^ permalink raw reply	[flat|nested] 15+ messages in thread

* Re: [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session
  2021-03-22 14:56             ` Steven Rostedt
@ 2021-03-25 20:55               ` Dario Faggioli
  0 siblings, 0 replies; 15+ messages in thread
From: Dario Faggioli @ 2021-03-25 20:55 UTC (permalink / raw)
  To: Steven Rostedt, Tzvetomir Stoyanov; +Cc: Stefano De Venuto, Linux Trace Devel


[-- Attachment #1: Type: text/plain, Size: 1027 bytes --]

On Mon, 2021-03-22 at 10:56 -0400, Steven Rostedt wrote:
> On Mon, 22 Mar 2021 12:13:09 +0200
> Tzvetomir Stoyanov <tz.stoyanov@gmail.com> wrote:
> 
> > Confirmed, the problem is in the trace-cmd logic that works with
> > the
> > KVM scaling.  Looks like tsc-scaling-ratio is the default KVM
> > scaling
> > on your machine, (1 << 32), and in that case the scaling should be
> > ignored in our calculations. I still have no solution, going to
> > analyze how the KVM scaling works.
> 
> I'm guessing that this was run on a AMD CPU, as the default scaling
> (the
> one to ignore) is 1 << 32 for AMD and for Intel it's 1 << 48.
> 
BTW, yes. Stefano could give us details (lscpu or anything) if
necessary, but this indeed is AMD.

Regards
-- 
Dario Faggioli, Ph.D
http://about.me/dario.faggioli
Virtualization Software Engineer
SUSE Labs, SUSE https://www.suse.com/
-------------------------------------------------------------------
<<This happens because _I_ choose it to happen!>> (Raistlin Majere)

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

^ permalink raw reply	[flat|nested] 15+ messages in thread

end of thread, back to index

Thread overview: 15+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-03-15  6:18 [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Tzvetomir Stoyanov (VMware)
2021-03-15  6:18 ` [PATCH v32 1/5] trace-cmd: Add dummy function to initialize timestamp sync logic Tzvetomir Stoyanov (VMware)
2021-03-15  6:18 ` [PATCH v32 2/5] trace-cmd: Add timestamp synchronization per vCPU Tzvetomir Stoyanov (VMware)
2021-03-15  6:18 ` [PATCH v32 3/5] trace-cmd: PTP-like algorithm for host - guest timestamp synchronization Tzvetomir Stoyanov (VMware)
2021-03-15  6:18 ` [PATCH v32 4/5] trace-cmd: Debug scripts for " Tzvetomir Stoyanov (VMware)
2021-03-15  6:18 ` [PATCH v32 5/5] trace-cmd [POC]: Add KVM timestamp synchronization plugin Tzvetomir Stoyanov (VMware)
2021-03-19 10:08 ` [PATCH v32 0/5]Timestamp synchronization of host - guest tracing session Stefano De Venuto
2021-03-19 11:55   ` Tzvetomir Stoyanov
2021-03-19 17:44     ` Stefano De Venuto
2021-03-20  6:25       ` Tzvetomir Stoyanov
     [not found]         ` <0f6e158c-43a3-4117-2a17-5a13cd6a7970@gmail.com>
2021-03-22 10:13           ` Tzvetomir Stoyanov
2021-03-22 14:56             ` Steven Rostedt
2021-03-25 20:55               ` Dario Faggioli
2021-03-25  7:20           ` Tzvetomir Stoyanov
2021-03-25 18:01             ` Stefano De Venuto

Linux-Trace-Devel Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/linux-trace-devel/0 linux-trace-devel/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 linux-trace-devel linux-trace-devel/ https://lore.kernel.org/linux-trace-devel \
		linux-trace-devel@vger.kernel.org
	public-inbox-index linux-trace-devel

Example config snippet for mirrors

Newsgroup available over NNTP:
	nntp://nntp.lore.kernel.org/org.kernel.vger.linux-trace-devel


AGPL code for this site: git clone https://public-inbox.org/public-inbox.git