new(libsinsp): POC #2 for optimized syscalls option

Signed-off-by: Melissa Kilby <melissa.kilby.oss@gmail.com>

userspace/libsinsp/examples/test.cpp

@@ -44,6 +44,7 @@ std::function<void(sinsp& inspector)> dump;
 static bool g_interrupted = false;
 static const uint8_t g_backoff_timeout_secs = 2;
 static bool g_all_threads = false;
+static bool optimized_syscalls_push_down = false;
 static bool json_dump_init_success = false;
 string engine_string = KMOD_ENGINE; /* Default for backward compatibility. */
 string filter_string = "";
@@ -73,17 +74,18 @@ static void usage()
 Overview: Goal of sinsp-example binary is to test and debug sinsp functionality and print events to STDOUT. All drivers are supported.
 
 Options:
-  -h, --help                                 Print this page.
-  -f <filter>, --filter <filter>             Filter string for events (see https://falco.org/docs/rules/supported-fields/ for supported fields).
-  -j, --json                                 Use JSON as the output format.
-  -a, --all-threads                          Output information about all threads, not just the main one.
-  -b <path>, --bpf <path>               	 BPF probe.
-  -m, --modern_bpf               			 modern BPF probe.
-  -k, --kmod								 Kernel module
-  -s <path>, --scap_file <path>   			 Scap file
-  -d <dim>, --buffer_dim <dim>               Dimension in bytes that every per-CPU buffer will have.
-  -o <fields>, --output-fields-json <fields>    [JSON support only, can also use without -j] Output fields string (see <filter> for supported display fields) that overwrites JSON default output fields for all events. * at the beginning prints JSON keys with null values, else no null fields are printed.
-  -E, --exclude-users                        Don't create the user/group tables
+  -h, --help                                        Print this page.
+  -f <filter>, --filter <filter>                    Filter string for events (see https://falco.org/docs/rules/supported-fields/ for supported fields).
+  -j, --json                                        Use JSON as the output format.
+  -a, --all-threads                                 Output information about all threads, not just the main one.
+  -b <path>, --bpf <path>                           BPF probe.
+  -m, --modern_bpf                                  Modern BPF probe.
+  -k, --kmod                                        Kernel module.
+  -s <path>, --scap_file <path>                     Scap file.
+  -d <dim>, --buffer_dim <dim>                      Dimension in bytes that every per-CPU buffer will have.
+  -o <fields>, --output-fields-json <fields>        [JSON support only, can also use without -j] Output fields string (see <filter> for supported display fields) that overwrites JSON default output fields for all events. * at the beginning prints JSON keys with null values, else no null fields are printed.
+  -E, --exclude-users                               Don't create the user/group tables.
+  -x, --optimized-syscalls                          Only trace syscalls and dependent syscalls needed for sinsp state built-up and life-cyle management.
 )";
 	cout << usage << endl;
 }
@@ -104,12 +106,13 @@ void parse_CLI_options(sinsp& inspector, int argc, char** argv)
 		{"buffer_dim", required_argument, 0, 'd'},
 		{"output-fields-json", required_argument, 0, 'o'},
 		{"exclude-users", no_argument, 0, 'E'},
+		{"optimized-syscalls", no_argument, 0, 'x'},
 		{0, 0, 0, 0}};
 
 	int op;
 	int long_index = 0;
 	while((op = getopt_long(argc, argv,
-				"hf:jab:mks:d:o:E",
+				"hf:jab:mks:d:o:E:x",
 				long_options, &long_index)) != -1)
 	{
 		switch(op)
@@ -151,6 +154,9 @@ void parse_CLI_options(sinsp& inspector, int argc, char** argv)
 		case 'E':
 			inspector.set_import_users(false);
 			break;
+		case 'x':
+			optimized_syscalls_push_down = true;
+			break;
 		default:
 			break;
 		}
@@ -160,12 +166,20 @@ void parse_CLI_options(sinsp& inspector, int argc, char** argv)
 
 void open_engine(sinsp& inspector)
 {
-	std::cout << "-- Try to open: '" + engine_string + "' engine." << std::endl;
-
 	/* Get only necessary tracepoints. */
 	std::unordered_set<uint32_t> tp_set = inspector.enforce_sinsp_state_tp();
 	std::unordered_set<uint32_t> ppm_sc;
 
+	if(optimized_syscalls_push_down && !inspector.m_filter_evttypes_names.empty())
+	{
+		ppm_sc = {};
+		printf("-- Setting optimized syscalls pushdown filters:\n");
+		inspector.set_filter_evttypes_ppm_sc_set(inspector.m_filter_evttypes_names, ppm_sc);
+		inspector.enforce_optimized_ppm_sc_set(ppm_sc, true);
+	}
+
+	std::cout << "-- Try to open: '" + engine_string + "' engine." << std::endl;
+
 	if(!engine_string.compare(KMOD_ENGINE))
 	{
 		inspector.open_kmod(buffer_bytes_dim, ppm_sc, tp_set);
@@ -283,10 +297,6 @@ int main(int argc, char** argv)
 	signal(SIGINT, sigint_handler);
 	signal(SIGTERM, sigint_handler);
 
-	open_engine(inspector);
-
-	std::cout << "-- Start capture" << std::endl;
-
 	if(!filter_string.empty())
 	{
 		try
@@ -299,6 +309,10 @@ int main(int argc, char** argv)
 		}
 	}
 
+	open_engine(inspector);
+
+	std::cout << "-- Start capture" << std::endl;
+
 	inspector.start_capture();
 	while(!g_interrupted)
 	{

userspace/libsinsp/filter.cpp

@@ -1677,6 +1677,30 @@ static void add_filtercheck_value(gen_event_filter_check *chk, size_t idx, const
 	}
 }
 
+static void add_evttypes_names(gen_event_filter_check *chk, const std::string& value, std::unordered_set<std::string>& m_filter_evttypes_names)
+{
+	// TODO fix: Currently includes negative evt.type checks e.g. "and not evt.type in (connect, accept)"
+	// -> ideally would account for grammar before current check e.g. "and not"
+	// -> at the same time negative evt.type filter should not be allowed and/or
+	// fall back to simple_set in that case
+
+	switch(chk->m_cmpop)
+	{
+		case CO_BCONTAINS:
+		case CO_BSTARTSWITH:
+		case CO_NE:
+		case CO_LT:
+		case CO_LE:
+		case CO_GT:
+		case CO_GE:
+		case CO_NONE:
+			break;
+		default:
+			m_filter_evttypes_names.insert(value.c_str());
+			break;
+	}
+}
+
 void sinsp_filter_compiler::visit(libsinsp::filter::ast::binary_check_expr* e)
 {
 	m_pos = e->get_pos();
@@ -1699,6 +1723,10 @@ void sinsp_filter_compiler::visit(libsinsp::filter::ast::binary_check_expr* e)
 	for (size_t i = 0; i < m_field_values.size(); i++)
 	{
 		add_filtercheck_value(check, i, m_field_values[i]);
+		if (field.compare("evt.type") == 0)
+		{
+			add_evttypes_names(check, m_field_values[i], m_filter_evttypes_names);
+		}
 	}
 }
 

userspace/libsinsp/filter.h

@@ -104,6 +104,7 @@ class SINSP_PUBLIC sinsp_filter_compiler:
 		\note Throws a sinsp_exception if the filter syntax is not valid
 	*/
 	sinsp_filter* compile();
+	std::unordered_set<std::string> m_filter_evttypes_names = {};
 
 	const libsinsp::filter::ast::pos_info& get_pos() const { return m_pos; }
 

userspace/libsinsp/sinsp.cpp

@@ -1777,6 +1777,12 @@ void sinsp::set_filter(const std::string& filter)
 	sinsp_filter_compiler compiler(this, filter);
 	m_filter = compiler.compile();
 	m_filterstring = filter;
+	m_filter_evttypes_names = compiler.m_filter_evttypes_names;
+}
+
+std::unordered_set<std::string> sinsp::get_filter_evttypes_names()
+{
+	return m_filter_evttypes_names;
 }
 
 const std::string sinsp::get_filter()

userspace/libsinsp/sinsp.h

@@ -337,6 +337,13 @@ class SINSP_PUBLIC sinsp : public capture_stats_source
 	*/
 	void set_filter(sinsp_filter* filter);
 
+	/*!
+	  \brief Return unordered set of string names of the syscalls/events defined in the filter as evt.type.
+
+	  \return \ref set_filter() has to be called first to compile and parse the filter, else return empty.
+	*/
+	std::unordered_set<std::string> get_filter_evttypes_names();
+
 	/*!
 	  \brief Return the filter set for this capture.
 
@@ -902,6 +909,17 @@ class SINSP_PUBLIC sinsp : public capture_stats_source
 	*/
 	std::unordered_set<uint32_t> enforce_sys_ppm_sc_set(std::unordered_set<uint32_t> ppm_sc_set = {});
 
+	/*!
+	  \brief Enforce passed set of syscalls with the ones needed for minimal
+	  sinsp state built-up and lifecycle management (useful for Falco use case).
+	*/
+	void enforce_optimized_ppm_sc_set(std::unordered_set<uint32_t>& ppm_sc_set, bool print = false);
+
+	/*!
+	  \brief Set ppm_sc_code for each evt.type in filter(s).
+	*/
+	void set_filter_evttypes_ppm_sc_set(std::unordered_set<std::string> m_filter_evttypes_names, std::unordered_set<uint32_t>& ppm_sc_set);
+
 	/*!
 	  \brief Get all the available ppm_sc.
 	  Does enforce minimum sinsp state set.
@@ -1335,6 +1353,7 @@ VISIBILITY_PRIVATE
 	uint64_t m_firstevent_ts;
 	sinsp_filter* m_filter;
 	std::string m_filterstring;
+	std::unordered_set<std::string> m_filter_evttypes_names;
 	//
 	// Internal stats
 	//

userspace/libsinsp/sinsp_ppm_sc.cpp

@@ -217,6 +217,95 @@ std::unordered_set<uint32_t> sinsp::enforce_sys_ppm_sc_set(std::unordered_set<ui
 	return ppm_sc_set;
 }
 
+void sinsp::enforce_optimized_ppm_sc_set(std::unordered_set<uint32_t>& ppm_sc_set, bool print)
+{
+	std::unordered_set<uint32_t> ppm_sc_set_filter_evttypes = ppm_sc_set;
+
+	/* Enforce syscalls solely needed for minimal sinsp state built-up and life-cycle management (Falco use case). */
+	std::unordered_set<uint32_t> base_thread_state_set = {
+		PPM_SC_CLONE,	/* These syscalls are all used to build up or modify info of the basic process (tinfo) struct. */
+		PPM_SC_CLONE3,
+		PPM_SC_FORK,
+		PPM_SC_VFORK,
+		PPM_SC_EXECVE,
+		PPM_SC_EXECVEAT,
+		PPM_SC_FCHDIR,
+		PPM_SC_CHDIR,
+		PPM_SC_CHROOT,
+		PPM_SC_CAPSET,
+		PPM_SC_SETGID,
+		PPM_SC_SETPGID,
+		PPM_SC_SETRESGID,
+		PPM_SC_SETRESUID,
+		PPM_SC_SETSID,
+		PPM_SC_SETUID,
+		PPM_SC_CLOSE,	/* Close syscall is crucial for destroying fds associated w/ a thread, given it's generic nature include in base set for simplicity. */
+	};
+
+	/* Crucial minimal PPME events that affect sinsp state other than syscalls are added later w/ the sinsp_ppm_sc APIs. */
+	ppm_sc_set.insert(base_thread_state_set.begin(), base_thread_state_set.end());
+
+	/* Add more dependent syscalls conditioned by syscalls / evttypes passed over the filter. */
+	std::unordered_set<uint32_t> base_network_set = {
+		PPM_SC_SOCKET,
+		PPM_SC_GETSOCKOPT,
+	};
+
+	std::unordered_set<uint32_t> base_accept_listen_set = {
+		PPM_SC_BIND,
+	};
+
+	for(const auto& ppm_sc_code : ppm_sc_set_filter_evttypes)
+	{
+		// TODO finalize
+		if (ppm_sc_code == PPM_SC_CONNECT ||
+			ppm_sc_code == PPM_SC_SOCKET ||
+			ppm_sc_code == PPM_SC_GETSOCKOPT ||
+			ppm_sc_code == PPM_SC_SETSOCKOPT ||
+			ppm_sc_code == PPM_SC_RECVMSG ||
+			ppm_sc_code == PPM_SC_RECVFROM ||
+			ppm_sc_code == PPM_SC_SENDMSG ||
+			ppm_sc_code == PPM_SC_SENDTO ||
+			ppm_sc_code == PPM_SC_BIND ||
+			ppm_sc_code == PPM_SC_SHUTDOWN ||
+			ppm_sc_code == PPM_SC_GETSOCKNAME ||
+			ppm_sc_code == PPM_SC_GETPEERNAME ||
+			ppm_sc_code == PPM_SC_SOCKETPAIR)
+		{
+			ppm_sc_set.insert(base_network_set.begin(), base_network_set.end());
+		} else if ( ppm_sc_code == PPM_SC_ACCEPT ||
+					ppm_sc_code == PPM_SC_ACCEPT4 ||
+					ppm_sc_code == PPM_SC_LISTEN)
+		{
+			ppm_sc_set.insert(base_network_set.begin(), base_network_set.end());
+			ppm_sc_set.insert(base_accept_listen_set.begin(), base_accept_listen_set.end());
+		}
+	}
+
+	if (print)
+	{
+		for(const auto& ppm_sc_code : ppm_sc_set)
+		{
+			std::string ppm_sc_name = g_infotables.m_syscall_info_table[ppm_sc_code].name;
+			printf("- %-25s ppm_code: (%d)\n", ppm_sc_name.c_str(), ppm_sc_code);
+		}
+	}
+
+}
+
+void sinsp::set_filter_evttypes_ppm_sc_set(std::unordered_set<std::string> m_filter_evttypes_names, std::unordered_set<uint32_t>& ppm_sc_set)
+{
+	// TODO possibly change lookup table / lookup procedure as tables get refactored
+	for (int ppm_sc_code = 0; ppm_sc_code < PPM_SC_MAX; ++ppm_sc_code)
+	{
+		std::string ppm_sc_name = g_infotables.m_syscall_info_table[ppm_sc_code].name;
+		if (m_filter_evttypes_names.find(ppm_sc_name) != m_filter_evttypes_names.end())
+		{
+			ppm_sc_set.insert(ppm_sc_code);
+		}
+	}
+}
+
 std::unordered_set<uint32_t> sinsp::get_event_set_from_ppm_sc_set(const std::unordered_set<uint32_t> &ppm_sc_set)
 {
 	std::vector<uint32_t> events_array(PPM_EVENT_MAX, 0);