Line data Source code
1 : //! Part of Safekeeper pretending to be Postgres, i.e. handling Postgres
2 : //! protocol commands.
3 :
4 : use anyhow::Context;
5 : use std::str::FromStr;
6 : use std::str::{self};
7 : use std::sync::Arc;
8 : use tokio::io::{AsyncRead, AsyncWrite};
9 : use tracing::{debug, info, info_span, Instrument};
10 :
11 : use crate::auth::check_permission;
12 : use crate::json_ctrl::{handle_json_ctrl, AppendLogicalMessage};
13 :
14 : use crate::metrics::{TrafficMetrics, PG_QUERIES_GAUGE};
15 : use crate::safekeeper::Term;
16 : use crate::timeline::TimelineError;
17 : use crate::wal_service::ConnectionId;
18 : use crate::{GlobalTimelines, SafeKeeperConf};
19 : use postgres_backend::QueryError;
20 : use postgres_backend::{self, PostgresBackend};
21 : use postgres_ffi::PG_TLI;
22 : use pq_proto::{BeMessage, FeStartupPacket, RowDescriptor, INT4_OID, TEXT_OID};
23 : use regex::Regex;
24 : use utils::auth::{Claims, JwtAuth, Scope};
25 : use utils::{
26 : id::{TenantId, TenantTimelineId, TimelineId},
27 : lsn::Lsn,
28 : };
29 :
30 : /// Safekeeper handler of postgres commands
31 : pub struct SafekeeperPostgresHandler {
32 : pub conf: SafeKeeperConf,
33 : /// assigned application name
34 : pub appname: Option<String>,
35 : pub tenant_id: Option<TenantId>,
36 : pub timeline_id: Option<TimelineId>,
37 : pub ttid: TenantTimelineId,
38 : /// Unique connection id is logged in spans for observability.
39 : pub conn_id: ConnectionId,
40 : /// Auth scope allowed on the connections and public key used to check auth tokens. None if auth is not configured.
41 : auth: Option<(Scope, Arc<JwtAuth>)>,
42 : claims: Option<Claims>,
43 : io_metrics: Option<TrafficMetrics>,
44 : }
45 :
46 : /// Parsed Postgres command.
47 : enum SafekeeperPostgresCommand {
48 : StartWalPush,
49 : StartReplication { start_lsn: Lsn, term: Option<Term> },
50 : IdentifySystem,
51 : TimelineStatus,
52 : JSONCtrl { cmd: AppendLogicalMessage },
53 : }
54 :
55 0 : fn parse_cmd(cmd: &str) -> anyhow::Result<SafekeeperPostgresCommand> {
56 0 : if cmd.starts_with("START_WAL_PUSH") {
57 0 : Ok(SafekeeperPostgresCommand::StartWalPush)
58 0 : } else if cmd.starts_with("START_REPLICATION") {
59 0 : let re = Regex::new(
60 0 : // We follow postgres START_REPLICATION LOGICAL options to pass term.
61 0 : r"START_REPLICATION(?: SLOT [^ ]+)?(?: PHYSICAL)? ([[:xdigit:]]+/[[:xdigit:]]+)(?: \(term='(\d+)'\))?",
62 0 : )
63 0 : .unwrap();
64 0 : let caps = re
65 0 : .captures(cmd)
66 0 : .context(format!("failed to parse START_REPLICATION command {}", cmd))?;
67 0 : let start_lsn =
68 0 : Lsn::from_str(&caps[1]).context("parse start LSN from START_REPLICATION command")?;
69 0 : let term = if let Some(m) = caps.get(2) {
70 0 : Some(m.as_str().parse::<u64>().context("invalid term")?)
71 : } else {
72 0 : None
73 : };
74 0 : Ok(SafekeeperPostgresCommand::StartReplication { start_lsn, term })
75 0 : } else if cmd.starts_with("IDENTIFY_SYSTEM") {
76 0 : Ok(SafekeeperPostgresCommand::IdentifySystem)
77 0 : } else if cmd.starts_with("TIMELINE_STATUS") {
78 0 : Ok(SafekeeperPostgresCommand::TimelineStatus)
79 0 : } else if cmd.starts_with("JSON_CTRL") {
80 0 : let cmd = cmd.strip_prefix("JSON_CTRL").context("invalid prefix")?;
81 : Ok(SafekeeperPostgresCommand::JSONCtrl {
82 0 : cmd: serde_json::from_str(cmd)?,
83 : })
84 : } else {
85 0 : anyhow::bail!("unsupported command {cmd}");
86 : }
87 0 : }
88 :
89 0 : fn cmd_to_string(cmd: &SafekeeperPostgresCommand) -> &str {
90 0 : match cmd {
91 0 : SafekeeperPostgresCommand::StartWalPush => "START_WAL_PUSH",
92 0 : SafekeeperPostgresCommand::StartReplication { .. } => "START_REPLICATION",
93 0 : SafekeeperPostgresCommand::TimelineStatus => "TIMELINE_STATUS",
94 0 : SafekeeperPostgresCommand::IdentifySystem => "IDENTIFY_SYSTEM",
95 0 : SafekeeperPostgresCommand::JSONCtrl { .. } => "JSON_CTRL",
96 : }
97 0 : }
98 :
99 : #[async_trait::async_trait]
100 : impl<IO: AsyncRead + AsyncWrite + Unpin + Send> postgres_backend::Handler<IO>
101 : for SafekeeperPostgresHandler
102 : {
103 : // tenant_id and timeline_id are passed in connection string params
104 0 : fn startup(
105 0 : &mut self,
106 0 : _pgb: &mut PostgresBackend<IO>,
107 0 : sm: &FeStartupPacket,
108 0 : ) -> Result<(), QueryError> {
109 0 : if let FeStartupPacket::StartupMessage { params, .. } = sm {
110 0 : if let Some(options) = params.options_raw() {
111 0 : for opt in options {
112 : // FIXME `ztenantid` and `ztimelineid` left for compatibility during deploy,
113 : // remove these after the PR gets deployed:
114 : // https://github.com/neondatabase/neon/pull/2433#discussion_r970005064
115 0 : match opt.split_once('=') {
116 0 : Some(("ztenantid", value)) | Some(("tenant_id", value)) => {
117 0 : self.tenant_id = Some(value.parse().with_context(|| {
118 0 : format!("Failed to parse {value} as tenant id")
119 0 : })?);
120 : }
121 0 : Some(("ztimelineid", value)) | Some(("timeline_id", value)) => {
122 0 : self.timeline_id = Some(value.parse().with_context(|| {
123 0 : format!("Failed to parse {value} as timeline id")
124 0 : })?);
125 : }
126 0 : Some(("availability_zone", client_az)) => {
127 0 : if let Some(metrics) = self.io_metrics.as_ref() {
128 0 : metrics.set_client_az(client_az)
129 0 : }
130 : }
131 0 : _ => continue,
132 : }
133 : }
134 0 : }
135 :
136 0 : if let Some(app_name) = params.get("application_name") {
137 0 : self.appname = Some(app_name.to_owned());
138 0 : if let Some(metrics) = self.io_metrics.as_ref() {
139 0 : metrics.set_app_name(app_name)
140 0 : }
141 0 : }
142 :
143 0 : let ttid = TenantTimelineId::new(
144 0 : self.tenant_id.unwrap_or(TenantId::from([0u8; 16])),
145 0 : self.timeline_id.unwrap_or(TimelineId::from([0u8; 16])),
146 0 : );
147 0 : tracing::Span::current().record("ttid", tracing::field::display(ttid));
148 0 :
149 0 : Ok(())
150 : } else {
151 0 : Err(QueryError::Other(anyhow::anyhow!(
152 0 : "Safekeeper received unexpected initial message: {sm:?}"
153 0 : )))
154 : }
155 0 : }
156 :
157 0 : fn check_auth_jwt(
158 0 : &mut self,
159 0 : _pgb: &mut PostgresBackend<IO>,
160 0 : jwt_response: &[u8],
161 0 : ) -> Result<(), QueryError> {
162 0 : // this unwrap is never triggered, because check_auth_jwt only called when auth_type is NeonJWT
163 0 : // which requires auth to be present
164 0 : let (allowed_auth_scope, auth) = self
165 0 : .auth
166 0 : .as_ref()
167 0 : .expect("auth_type is configured but .auth of handler is missing");
168 0 : let data = auth
169 0 : .decode(str::from_utf8(jwt_response).context("jwt response is not UTF-8")?)
170 0 : .map_err(|e| QueryError::Unauthorized(e.0))?;
171 :
172 : // The handler might be configured to allow only tenant scope tokens.
173 0 : if matches!(allowed_auth_scope, Scope::Tenant)
174 0 : && !matches!(data.claims.scope, Scope::Tenant)
175 : {
176 0 : return Err(QueryError::Unauthorized(
177 0 : "passed JWT token is for full access, but only tenant scope is allowed".into(),
178 0 : ));
179 0 : }
180 :
181 0 : if matches!(data.claims.scope, Scope::Tenant) && data.claims.tenant_id.is_none() {
182 0 : return Err(QueryError::Unauthorized(
183 0 : "jwt token scope is Tenant, but tenant id is missing".into(),
184 0 : ));
185 0 : }
186 0 :
187 0 : debug!(
188 0 : "jwt scope check succeeded for scope: {:#?} by tenant id: {:?}",
189 0 : data.claims.scope, data.claims.tenant_id,
190 0 : );
191 :
192 0 : self.claims = Some(data.claims);
193 0 : Ok(())
194 0 : }
195 :
196 0 : async fn process_query(
197 0 : &mut self,
198 0 : pgb: &mut PostgresBackend<IO>,
199 0 : query_string: &str,
200 0 : ) -> Result<(), QueryError> {
201 0 : if query_string
202 0 : .to_ascii_lowercase()
203 0 : .starts_with("set datestyle to ")
204 : {
205 : // important for debug because psycopg2 executes "SET datestyle TO 'ISO'" on connect
206 0 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
207 0 : return Ok(());
208 0 : }
209 :
210 0 : let cmd = parse_cmd(query_string)?;
211 0 : let cmd_str = cmd_to_string(&cmd);
212 0 :
213 0 : let _guard = PG_QUERIES_GAUGE.with_label_values(&[cmd_str]).guard();
214 0 :
215 0 : info!("got query {:?}", query_string);
216 :
217 0 : let tenant_id = self.tenant_id.context("tenantid is required")?;
218 0 : let timeline_id = self.timeline_id.context("timelineid is required")?;
219 0 : self.check_permission(Some(tenant_id))?;
220 0 : self.ttid = TenantTimelineId::new(tenant_id, timeline_id);
221 0 :
222 0 : match cmd {
223 : SafekeeperPostgresCommand::StartWalPush => {
224 0 : self.handle_start_wal_push(pgb)
225 0 : .instrument(info_span!("WAL receiver"))
226 0 : .await
227 : }
228 0 : SafekeeperPostgresCommand::StartReplication { start_lsn, term } => {
229 0 : self.handle_start_replication(pgb, start_lsn, term)
230 0 : .instrument(info_span!("WAL sender"))
231 0 : .await
232 : }
233 0 : SafekeeperPostgresCommand::IdentifySystem => self.handle_identify_system(pgb).await,
234 0 : SafekeeperPostgresCommand::TimelineStatus => self.handle_timeline_status(pgb).await,
235 0 : SafekeeperPostgresCommand::JSONCtrl { ref cmd } => {
236 0 : handle_json_ctrl(self, pgb, cmd).await
237 : }
238 : }
239 0 : }
240 : }
241 :
242 : impl SafekeeperPostgresHandler {
243 0 : pub fn new(
244 0 : conf: SafeKeeperConf,
245 0 : conn_id: u32,
246 0 : io_metrics: Option<TrafficMetrics>,
247 0 : auth: Option<(Scope, Arc<JwtAuth>)>,
248 0 : ) -> Self {
249 0 : SafekeeperPostgresHandler {
250 0 : conf,
251 0 : appname: None,
252 0 : tenant_id: None,
253 0 : timeline_id: None,
254 0 : ttid: TenantTimelineId::empty(),
255 0 : conn_id,
256 0 : claims: None,
257 0 : auth,
258 0 : io_metrics,
259 0 : }
260 0 : }
261 :
262 : // when accessing management api supply None as an argument
263 : // when using to authorize tenant pass corresponding tenant id
264 0 : fn check_permission(&self, tenant_id: Option<TenantId>) -> Result<(), QueryError> {
265 0 : if self.auth.is_none() {
266 : // auth is set to Trust, nothing to check so just return ok
267 0 : return Ok(());
268 0 : }
269 0 : // auth is some, just checked above, when auth is some
270 0 : // then claims are always present because of checks during connection init
271 0 : // so this expect won't trigger
272 0 : let claims = self
273 0 : .claims
274 0 : .as_ref()
275 0 : .expect("claims presence already checked");
276 0 : check_permission(claims, tenant_id).map_err(|e| QueryError::Unauthorized(e.0))
277 0 : }
278 :
279 0 : async fn handle_timeline_status<IO: AsyncRead + AsyncWrite + Unpin>(
280 0 : &mut self,
281 0 : pgb: &mut PostgresBackend<IO>,
282 0 : ) -> Result<(), QueryError> {
283 : // Get timeline, handling "not found" error
284 0 : let tli = match GlobalTimelines::get(self.ttid) {
285 0 : Ok(tli) => Ok(Some(tli)),
286 0 : Err(TimelineError::NotFound(_)) => Ok(None),
287 0 : Err(e) => Err(QueryError::Other(e.into())),
288 0 : }?;
289 :
290 : // Write row description
291 0 : pgb.write_message_noflush(&BeMessage::RowDescription(&[
292 0 : RowDescriptor::text_col(b"flush_lsn"),
293 0 : RowDescriptor::text_col(b"commit_lsn"),
294 0 : ]))?;
295 :
296 : // Write row if timeline exists
297 0 : if let Some(tli) = tli {
298 0 : let (inmem, _state) = tli.get_state().await;
299 0 : let flush_lsn = tli.get_flush_lsn().await;
300 0 : let commit_lsn = inmem.commit_lsn;
301 0 : pgb.write_message_noflush(&BeMessage::DataRow(&[
302 0 : Some(flush_lsn.to_string().as_bytes()),
303 0 : Some(commit_lsn.to_string().as_bytes()),
304 0 : ]))?;
305 0 : }
306 :
307 0 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"TIMELINE_STATUS"))?;
308 0 : Ok(())
309 0 : }
310 :
311 : ///
312 : /// Handle IDENTIFY_SYSTEM replication command
313 : ///
314 0 : async fn handle_identify_system<IO: AsyncRead + AsyncWrite + Unpin>(
315 0 : &mut self,
316 0 : pgb: &mut PostgresBackend<IO>,
317 0 : ) -> Result<(), QueryError> {
318 0 : let tli = GlobalTimelines::get(self.ttid).map_err(|e| QueryError::Other(e.into()))?;
319 :
320 0 : let lsn = if self.is_walproposer_recovery() {
321 : // walproposer should get all local WAL until flush_lsn
322 0 : tli.get_flush_lsn().await
323 : } else {
324 : // other clients shouldn't get any uncommitted WAL
325 0 : tli.get_state().await.0.commit_lsn
326 : }
327 0 : .to_string();
328 :
329 0 : let sysid = tli.get_state().await.1.server.system_id.to_string();
330 0 : let lsn_bytes = lsn.as_bytes();
331 0 : let tli = PG_TLI.to_string();
332 0 : let tli_bytes = tli.as_bytes();
333 0 : let sysid_bytes = sysid.as_bytes();
334 0 :
335 0 : pgb.write_message_noflush(&BeMessage::RowDescription(&[
336 0 : RowDescriptor {
337 0 : name: b"systemid",
338 0 : typoid: TEXT_OID,
339 0 : typlen: -1,
340 0 : ..Default::default()
341 0 : },
342 0 : RowDescriptor {
343 0 : name: b"timeline",
344 0 : typoid: INT4_OID,
345 0 : typlen: 4,
346 0 : ..Default::default()
347 0 : },
348 0 : RowDescriptor {
349 0 : name: b"xlogpos",
350 0 : typoid: TEXT_OID,
351 0 : typlen: -1,
352 0 : ..Default::default()
353 0 : },
354 0 : RowDescriptor {
355 0 : name: b"dbname",
356 0 : typoid: TEXT_OID,
357 0 : typlen: -1,
358 0 : ..Default::default()
359 0 : },
360 0 : ]))?
361 0 : .write_message_noflush(&BeMessage::DataRow(&[
362 0 : Some(sysid_bytes),
363 0 : Some(tli_bytes),
364 0 : Some(lsn_bytes),
365 0 : None,
366 0 : ]))?
367 0 : .write_message_noflush(&BeMessage::CommandComplete(b"IDENTIFY_SYSTEM"))?;
368 0 : Ok(())
369 0 : }
370 :
371 : /// Returns true if current connection is a replication connection, originating
372 : /// from a walproposer recovery function. This connection gets a special handling:
373 : /// safekeeper must stream all local WAL till the flush_lsn, whether committed or not.
374 0 : pub fn is_walproposer_recovery(&self) -> bool {
375 0 : match &self.appname {
376 0 : None => false,
377 0 : Some(appname) => {
378 0 : appname == "wal_proposer_recovery" ||
379 : // set by safekeeper peer recovery
380 0 : appname.starts_with("safekeeper")
381 : }
382 : }
383 0 : }
384 : }
|
