Line data Source code
1 : //! The Page Service listens for client connections and serves their GetPage@LSN
2 : //! requests.
3 :
4 : use std::borrow::Cow;
5 : use std::num::NonZeroUsize;
6 : use std::os::fd::AsRawFd;
7 : use std::str::FromStr;
8 : use std::sync::Arc;
9 : use std::time::{Duration, Instant, SystemTime};
10 : use std::{io, str};
11 :
12 : use anyhow::{Context, bail};
13 : use async_compression::tokio::write::GzipEncoder;
14 : use bytes::Buf;
15 : use futures::FutureExt;
16 : use itertools::Itertools;
17 : use once_cell::sync::OnceCell;
18 : use pageserver_api::config::{
19 : PageServicePipeliningConfig, PageServicePipeliningConfigPipelined,
20 : PageServiceProtocolPipelinedExecutionStrategy,
21 : };
22 : use pageserver_api::key::rel_block_to_key;
23 : use pageserver_api::models::{
24 : self, PageTraceEvent, PagestreamBeMessage, PagestreamDbSizeRequest, PagestreamDbSizeResponse,
25 : PagestreamErrorResponse, PagestreamExistsRequest, PagestreamExistsResponse,
26 : PagestreamFeMessage, PagestreamGetPageRequest, PagestreamGetSlruSegmentRequest,
27 : PagestreamGetSlruSegmentResponse, PagestreamNblocksRequest, PagestreamNblocksResponse,
28 : PagestreamProtocolVersion, PagestreamRequest, TenantState,
29 : };
30 : use pageserver_api::reltag::SlruKind;
31 : use pageserver_api::shard::TenantShardId;
32 : use postgres_backend::{
33 : AuthType, PostgresBackend, PostgresBackendReader, QueryError, is_expected_io_error,
34 : };
35 : use postgres_ffi::BLCKSZ;
36 : use postgres_ffi::pg_constants::DEFAULTTABLESPACE_OID;
37 : use pq_proto::framed::ConnectionError;
38 : use pq_proto::{BeMessage, FeMessage, FeStartupPacket, RowDescriptor};
39 : use strum_macros::IntoStaticStr;
40 : use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt, BufWriter};
41 : use tokio::task::JoinHandle;
42 : use tokio_util::sync::CancellationToken;
43 : use tracing::*;
44 : use utils::auth::{Claims, Scope, SwappableJwtAuth};
45 : use utils::failpoint_support;
46 : use utils::id::{TenantId, TimelineId};
47 : use utils::logging::log_slow;
48 : use utils::lsn::Lsn;
49 : use utils::simple_rcu::RcuReadGuard;
50 : use utils::sync::gate::{Gate, GateGuard};
51 : use utils::sync::spsc_fold;
52 :
53 : use crate::auth::check_permission;
54 : use crate::basebackup::BasebackupError;
55 : use crate::config::PageServerConf;
56 : use crate::context::{DownloadBehavior, RequestContext};
57 : use crate::metrics::{
58 : self, COMPUTE_COMMANDS_COUNTERS, ComputeCommandKind, LIVE_CONNECTIONS, SmgrOpTimer,
59 : TimelineMetrics,
60 : };
61 : use crate::pgdatadir_mapping::Version;
62 : use crate::span::{
63 : debug_assert_current_span_has_tenant_and_timeline_id,
64 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id,
65 : };
66 : use crate::task_mgr::{self, COMPUTE_REQUEST_RUNTIME, TaskKind};
67 : use crate::tenant::mgr::{
68 : GetActiveTenantError, GetTenantError, ShardResolveResult, ShardSelector, TenantManager,
69 : };
70 : use crate::tenant::storage_layer::IoConcurrency;
71 : use crate::tenant::timeline::{self, WaitLsnError};
72 : use crate::tenant::{GetTimelineError, PageReconstructError, Timeline};
73 : use crate::{basebackup, timed_after_cancellation};
74 :
75 : /// How long we may wait for a [`crate::tenant::mgr::TenantSlot::InProgress`]` and/or a [`crate::tenant::Tenant`] which
76 : /// is not yet in state [`TenantState::Active`].
77 : ///
78 : /// NB: this is a different value than [`crate::http::routes::ACTIVE_TENANT_TIMEOUT`].
79 : const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(30000);
80 :
81 : /// Threshold at which to log slow GetPage requests.
82 : const LOG_SLOW_GETPAGE_THRESHOLD: Duration = Duration::from_secs(30);
83 :
84 : ///////////////////////////////////////////////////////////////////////////////
85 :
86 : pub struct Listener {
87 : cancel: CancellationToken,
88 : /// Cancel the listener task through `listen_cancel` to shut down the listener
89 : /// and get a handle on the existing connections.
90 : task: JoinHandle<Connections>,
91 : }
92 :
93 : pub struct Connections {
94 : cancel: CancellationToken,
95 : tasks: tokio::task::JoinSet<ConnectionHandlerResult>,
96 : gate: Gate,
97 : }
98 :
99 0 : pub fn spawn(
100 0 : conf: &'static PageServerConf,
101 0 : tenant_manager: Arc<TenantManager>,
102 0 : pg_auth: Option<Arc<SwappableJwtAuth>>,
103 0 : tcp_listener: tokio::net::TcpListener,
104 0 : ) -> Listener {
105 0 : let cancel = CancellationToken::new();
106 0 : let libpq_ctx = RequestContext::todo_child(
107 0 : TaskKind::LibpqEndpointListener,
108 0 : // listener task shouldn't need to download anything. (We will
109 0 : // create a separate sub-contexts for each connection, with their
110 0 : // own download behavior. This context is used only to listen and
111 0 : // accept connections.)
112 0 : DownloadBehavior::Error,
113 0 : );
114 0 : let task = COMPUTE_REQUEST_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
115 0 : "libpq listener",
116 0 : libpq_listener_main(
117 0 : conf,
118 0 : tenant_manager,
119 0 : pg_auth,
120 0 : tcp_listener,
121 0 : conf.pg_auth_type,
122 0 : conf.page_service_pipelining.clone(),
123 0 : libpq_ctx,
124 0 : cancel.clone(),
125 0 : )
126 0 : .map(anyhow::Ok),
127 0 : ));
128 0 :
129 0 : Listener { cancel, task }
130 0 : }
131 :
132 : impl Listener {
133 0 : pub async fn stop_accepting(self) -> Connections {
134 0 : self.cancel.cancel();
135 0 : self.task
136 0 : .await
137 0 : .expect("unreachable: we wrap the listener task in task_mgr::exit_on_panic_or_error")
138 0 : }
139 : }
140 : impl Connections {
141 0 : pub(crate) async fn shutdown(self) {
142 0 : let Self {
143 0 : cancel,
144 0 : mut tasks,
145 0 : gate,
146 0 : } = self;
147 0 : cancel.cancel();
148 0 : while let Some(res) = tasks.join_next().await {
149 0 : Self::handle_connection_completion(res);
150 0 : }
151 0 : gate.close().await;
152 0 : }
153 :
154 0 : fn handle_connection_completion(res: Result<anyhow::Result<()>, tokio::task::JoinError>) {
155 0 : match res {
156 0 : Ok(Ok(())) => {}
157 0 : Ok(Err(e)) => error!("error in page_service connection task: {:?}", e),
158 0 : Err(e) => error!("page_service connection task panicked: {:?}", e),
159 : }
160 0 : }
161 : }
162 :
163 : ///
164 : /// Main loop of the page service.
165 : ///
166 : /// Listens for connections, and launches a new handler task for each.
167 : ///
168 : /// Returns Ok(()) upon cancellation via `cancel`, returning the set of
169 : /// open connections.
170 : ///
171 : #[allow(clippy::too_many_arguments)]
172 0 : pub async fn libpq_listener_main(
173 0 : conf: &'static PageServerConf,
174 0 : tenant_manager: Arc<TenantManager>,
175 0 : auth: Option<Arc<SwappableJwtAuth>>,
176 0 : listener: tokio::net::TcpListener,
177 0 : auth_type: AuthType,
178 0 : pipelining_config: PageServicePipeliningConfig,
179 0 : listener_ctx: RequestContext,
180 0 : listener_cancel: CancellationToken,
181 0 : ) -> Connections {
182 0 : let connections_cancel = CancellationToken::new();
183 0 : let connections_gate = Gate::default();
184 0 : let mut connection_handler_tasks = tokio::task::JoinSet::default();
185 :
186 : loop {
187 0 : let gate_guard = match connections_gate.enter() {
188 0 : Ok(guard) => guard,
189 0 : Err(_) => break,
190 : };
191 :
192 0 : let accepted = tokio::select! {
193 : biased;
194 0 : _ = listener_cancel.cancelled() => break,
195 0 : next = connection_handler_tasks.join_next(), if !connection_handler_tasks.is_empty() => {
196 0 : let res = next.expect("we dont poll while empty");
197 0 : Connections::handle_connection_completion(res);
198 0 : continue;
199 : }
200 0 : accepted = listener.accept() => accepted,
201 0 : };
202 0 :
203 0 : match accepted {
204 0 : Ok((socket, peer_addr)) => {
205 0 : // Connection established. Spawn a new task to handle it.
206 0 : debug!("accepted connection from {}", peer_addr);
207 0 : let local_auth = auth.clone();
208 0 : let connection_ctx = listener_ctx
209 0 : .detached_child(TaskKind::PageRequestHandler, DownloadBehavior::Download);
210 0 : connection_handler_tasks.spawn(page_service_conn_main(
211 0 : conf,
212 0 : tenant_manager.clone(),
213 0 : local_auth,
214 0 : socket,
215 0 : auth_type,
216 0 : pipelining_config.clone(),
217 0 : connection_ctx,
218 0 : connections_cancel.child_token(),
219 0 : gate_guard,
220 0 : ));
221 : }
222 0 : Err(err) => {
223 0 : // accept() failed. Log the error, and loop back to retry on next connection.
224 0 : error!("accept() failed: {:?}", err);
225 : }
226 : }
227 : }
228 :
229 0 : debug!("page_service listener loop terminated");
230 :
231 0 : Connections {
232 0 : cancel: connections_cancel,
233 0 : tasks: connection_handler_tasks,
234 0 : gate: connections_gate,
235 0 : }
236 0 : }
237 :
238 : type ConnectionHandlerResult = anyhow::Result<()>;
239 :
240 : #[instrument(skip_all, fields(peer_addr, application_name))]
241 : #[allow(clippy::too_many_arguments)]
242 : async fn page_service_conn_main(
243 : conf: &'static PageServerConf,
244 : tenant_manager: Arc<TenantManager>,
245 : auth: Option<Arc<SwappableJwtAuth>>,
246 : socket: tokio::net::TcpStream,
247 : auth_type: AuthType,
248 : pipelining_config: PageServicePipeliningConfig,
249 : connection_ctx: RequestContext,
250 : cancel: CancellationToken,
251 : gate_guard: GateGuard,
252 : ) -> ConnectionHandlerResult {
253 : let _guard = LIVE_CONNECTIONS
254 : .with_label_values(&["page_service"])
255 : .guard();
256 :
257 : socket
258 : .set_nodelay(true)
259 : .context("could not set TCP_NODELAY")?;
260 :
261 : let socket_fd = socket.as_raw_fd();
262 :
263 : let peer_addr = socket.peer_addr().context("get peer address")?;
264 : tracing::Span::current().record("peer_addr", field::display(peer_addr));
265 :
266 : // setup read timeout of 10 minutes. the timeout is rather arbitrary for requirements:
267 : // - long enough for most valid compute connections
268 : // - less than infinite to stop us from "leaking" connections to long-gone computes
269 : //
270 : // no write timeout is used, because the kernel is assumed to error writes after some time.
271 : let mut socket = tokio_io_timeout::TimeoutReader::new(socket);
272 :
273 : let default_timeout_ms = 10 * 60 * 1000; // 10 minutes by default
274 0 : let socket_timeout_ms = (|| {
275 0 : fail::fail_point!("simulated-bad-compute-connection", |avg_timeout_ms| {
276 : // Exponential distribution for simulating
277 : // poor network conditions, expect about avg_timeout_ms to be around 15
278 : // in tests
279 0 : if let Some(avg_timeout_ms) = avg_timeout_ms {
280 0 : let avg = avg_timeout_ms.parse::<i64>().unwrap() as f32;
281 0 : let u = rand::random::<f32>();
282 0 : ((1.0 - u).ln() / (-avg)) as u64
283 : } else {
284 0 : default_timeout_ms
285 : }
286 0 : });
287 0 : default_timeout_ms
288 : })();
289 :
290 : // A timeout here does not mean the client died, it can happen if it's just idle for
291 : // a while: we will tear down this PageServerHandler and instantiate a new one if/when
292 : // they reconnect.
293 : socket.set_timeout(Some(std::time::Duration::from_millis(socket_timeout_ms)));
294 : let socket = Box::pin(socket);
295 :
296 : fail::fail_point!("ps::connection-start::pre-login");
297 :
298 : // XXX: pgbackend.run() should take the connection_ctx,
299 : // and create a child per-query context when it invokes process_query.
300 : // But it's in a shared crate, so, we store connection_ctx inside PageServerHandler
301 : // and create the per-query context in process_query ourselves.
302 : let mut conn_handler = PageServerHandler::new(
303 : conf,
304 : tenant_manager,
305 : auth,
306 : pipelining_config,
307 : connection_ctx,
308 : cancel.clone(),
309 : gate_guard,
310 : );
311 : let pgbackend = PostgresBackend::new_from_io(socket_fd, socket, peer_addr, auth_type, None)?;
312 :
313 : match pgbackend.run(&mut conn_handler, &cancel).await {
314 : Ok(()) => {
315 : // we've been requested to shut down
316 : Ok(())
317 : }
318 : Err(QueryError::Disconnected(ConnectionError::Io(io_error))) => {
319 : if is_expected_io_error(&io_error) {
320 : info!("Postgres client disconnected ({io_error})");
321 : Ok(())
322 : } else {
323 : let tenant_id = conn_handler.timeline_handles.as_ref().unwrap().tenant_id();
324 : Err(io_error).context(format!(
325 : "Postgres connection error for tenant_id={:?} client at peer_addr={}",
326 : tenant_id, peer_addr
327 : ))
328 : }
329 : }
330 : other => {
331 : let tenant_id = conn_handler.timeline_handles.as_ref().unwrap().tenant_id();
332 : other.context(format!(
333 : "Postgres query error for tenant_id={:?} client peer_addr={}",
334 : tenant_id, peer_addr
335 : ))
336 : }
337 : }
338 : }
339 :
340 : struct PageServerHandler {
341 : conf: &'static PageServerConf,
342 : auth: Option<Arc<SwappableJwtAuth>>,
343 : claims: Option<Claims>,
344 :
345 : /// The context created for the lifetime of the connection
346 : /// services by this PageServerHandler.
347 : /// For each query received over the connection,
348 : /// `process_query` creates a child context from this one.
349 : connection_ctx: RequestContext,
350 :
351 : cancel: CancellationToken,
352 :
353 : /// None only while pagestream protocol is being processed.
354 : timeline_handles: Option<TimelineHandles>,
355 :
356 : pipelining_config: PageServicePipeliningConfig,
357 :
358 : gate_guard: GateGuard,
359 : }
360 :
361 : struct TimelineHandles {
362 : wrapper: TenantManagerWrapper,
363 : /// Note on size: the typical size of this map is 1. The largest size we expect
364 : /// to see is the number of shards divided by the number of pageservers (typically < 2),
365 : /// or the ratio used when splitting shards (i.e. how many children created from one)
366 : /// parent shard, where a "large" number might be ~8.
367 : handles: timeline::handle::Cache<TenantManagerTypes>,
368 : }
369 :
370 : impl TimelineHandles {
371 0 : fn new(tenant_manager: Arc<TenantManager>) -> Self {
372 0 : Self {
373 0 : wrapper: TenantManagerWrapper {
374 0 : tenant_manager,
375 0 : tenant_id: OnceCell::new(),
376 0 : },
377 0 : handles: Default::default(),
378 0 : }
379 0 : }
380 0 : async fn get(
381 0 : &mut self,
382 0 : tenant_id: TenantId,
383 0 : timeline_id: TimelineId,
384 0 : shard_selector: ShardSelector,
385 0 : ) -> Result<timeline::handle::Handle<TenantManagerTypes>, GetActiveTimelineError> {
386 0 : if *self.wrapper.tenant_id.get_or_init(|| tenant_id) != tenant_id {
387 0 : return Err(GetActiveTimelineError::Tenant(
388 0 : GetActiveTenantError::SwitchedTenant,
389 0 : ));
390 0 : }
391 0 : self.handles
392 0 : .get(timeline_id, shard_selector, &self.wrapper)
393 0 : .await
394 0 : .map_err(|e| match e {
395 0 : timeline::handle::GetError::TenantManager(e) => e,
396 : timeline::handle::GetError::PerTimelineStateShutDown => {
397 0 : trace!("per-timeline state shut down");
398 0 : GetActiveTimelineError::Timeline(GetTimelineError::ShuttingDown)
399 : }
400 0 : })
401 0 : }
402 :
403 0 : fn tenant_id(&self) -> Option<TenantId> {
404 0 : self.wrapper.tenant_id.get().copied()
405 0 : }
406 : }
407 :
408 : pub(crate) struct TenantManagerWrapper {
409 : tenant_manager: Arc<TenantManager>,
410 : // We do not support switching tenant_id on a connection at this point.
411 : // We can can add support for this later if needed without changing
412 : // the protocol.
413 : tenant_id: once_cell::sync::OnceCell<TenantId>,
414 : }
415 :
416 : #[derive(Debug)]
417 : pub(crate) struct TenantManagerTypes;
418 :
419 : impl timeline::handle::Types for TenantManagerTypes {
420 : type TenantManagerError = GetActiveTimelineError;
421 : type TenantManager = TenantManagerWrapper;
422 : type Timeline = TenantManagerCacheItem;
423 : }
424 :
425 : pub(crate) struct TenantManagerCacheItem {
426 : pub(crate) timeline: Arc<Timeline>,
427 : // allow() for cheap propagation through RequestContext inside a task
428 : #[allow(clippy::redundant_allocation)]
429 : pub(crate) metrics: Arc<Arc<TimelineMetrics>>,
430 : #[allow(dead_code)] // we store it to keep the gate open
431 : pub(crate) gate_guard: GateGuard,
432 : }
433 :
434 : impl std::ops::Deref for TenantManagerCacheItem {
435 : type Target = Arc<Timeline>;
436 0 : fn deref(&self) -> &Self::Target {
437 0 : &self.timeline
438 0 : }
439 : }
440 :
441 : impl timeline::handle::Timeline<TenantManagerTypes> for TenantManagerCacheItem {
442 0 : fn shard_timeline_id(&self) -> timeline::handle::ShardTimelineId {
443 0 : Timeline::shard_timeline_id(&self.timeline)
444 0 : }
445 :
446 0 : fn per_timeline_state(&self) -> &timeline::handle::PerTimelineState<TenantManagerTypes> {
447 0 : &self.timeline.handles
448 0 : }
449 :
450 0 : fn get_shard_identity(&self) -> &pageserver_api::shard::ShardIdentity {
451 0 : Timeline::get_shard_identity(&self.timeline)
452 0 : }
453 : }
454 :
455 : impl timeline::handle::TenantManager<TenantManagerTypes> for TenantManagerWrapper {
456 0 : async fn resolve(
457 0 : &self,
458 0 : timeline_id: TimelineId,
459 0 : shard_selector: ShardSelector,
460 0 : ) -> Result<TenantManagerCacheItem, GetActiveTimelineError> {
461 0 : let tenant_id = self.tenant_id.get().expect("we set this in get()");
462 0 : let timeout = ACTIVE_TENANT_TIMEOUT;
463 0 : let wait_start = Instant::now();
464 0 : let deadline = wait_start + timeout;
465 0 : let tenant_shard = loop {
466 0 : let resolved = self
467 0 : .tenant_manager
468 0 : .resolve_attached_shard(tenant_id, shard_selector);
469 0 : match resolved {
470 0 : ShardResolveResult::Found(tenant_shard) => break tenant_shard,
471 : ShardResolveResult::NotFound => {
472 0 : return Err(GetActiveTimelineError::Tenant(
473 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(*tenant_id)),
474 0 : ));
475 : }
476 0 : ShardResolveResult::InProgress(barrier) => {
477 0 : // We can't authoritatively answer right now: wait for InProgress state
478 0 : // to end, then try again
479 0 : tokio::select! {
480 0 : _ = barrier.wait() => {
481 0 : // The barrier completed: proceed around the loop to try looking up again
482 0 : },
483 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
484 0 : return Err(GetActiveTimelineError::Tenant(GetActiveTenantError::WaitForActiveTimeout {
485 0 : latest_state: None,
486 0 : wait_time: timeout,
487 0 : }));
488 : }
489 : }
490 : }
491 : };
492 : };
493 :
494 0 : tracing::debug!("Waiting for tenant to enter active state...");
495 0 : tenant_shard
496 0 : .wait_to_become_active(deadline.duration_since(Instant::now()))
497 0 : .await
498 0 : .map_err(GetActiveTimelineError::Tenant)?;
499 :
500 0 : let timeline = tenant_shard
501 0 : .get_timeline(timeline_id, true)
502 0 : .map_err(GetActiveTimelineError::Timeline)?;
503 :
504 0 : let gate_guard = match timeline.gate.enter() {
505 0 : Ok(guard) => guard,
506 : Err(_) => {
507 0 : return Err(GetActiveTimelineError::Timeline(
508 0 : GetTimelineError::ShuttingDown,
509 0 : ));
510 : }
511 : };
512 :
513 0 : let metrics = Arc::new(Arc::clone(&timeline.metrics));
514 0 :
515 0 : Ok(TenantManagerCacheItem {
516 0 : timeline,
517 0 : metrics,
518 0 : gate_guard,
519 0 : })
520 0 : }
521 : }
522 :
523 : #[derive(thiserror::Error, Debug)]
524 : enum PageStreamError {
525 : /// We encountered an error that should prompt the client to reconnect:
526 : /// in practice this means we drop the connection without sending a response.
527 : #[error("Reconnect required: {0}")]
528 : Reconnect(Cow<'static, str>),
529 :
530 : /// We were instructed to shutdown while processing the query
531 : #[error("Shutting down")]
532 : Shutdown,
533 :
534 : /// Something went wrong reading a page: this likely indicates a pageserver bug
535 : #[error("Read error")]
536 : Read(#[source] PageReconstructError),
537 :
538 : /// Ran out of time waiting for an LSN
539 : #[error("LSN timeout: {0}")]
540 : LsnTimeout(WaitLsnError),
541 :
542 : /// The entity required to serve the request (tenant or timeline) is not found,
543 : /// or is not found in a suitable state to serve a request.
544 : #[error("Not found: {0}")]
545 : NotFound(Cow<'static, str>),
546 :
547 : /// Request asked for something that doesn't make sense, like an invalid LSN
548 : #[error("Bad request: {0}")]
549 : BadRequest(Cow<'static, str>),
550 : }
551 :
552 : impl From<PageReconstructError> for PageStreamError {
553 0 : fn from(value: PageReconstructError) -> Self {
554 0 : match value {
555 0 : PageReconstructError::Cancelled => Self::Shutdown,
556 0 : e => Self::Read(e),
557 : }
558 0 : }
559 : }
560 :
561 : impl From<GetActiveTimelineError> for PageStreamError {
562 0 : fn from(value: GetActiveTimelineError) -> Self {
563 0 : match value {
564 : GetActiveTimelineError::Tenant(GetActiveTenantError::Cancelled)
565 : | GetActiveTimelineError::Tenant(GetActiveTenantError::WillNotBecomeActive(
566 : TenantState::Stopping { .. },
567 : ))
568 0 : | GetActiveTimelineError::Timeline(GetTimelineError::ShuttingDown) => Self::Shutdown,
569 0 : GetActiveTimelineError::Tenant(e) => Self::NotFound(format!("{e}").into()),
570 0 : GetActiveTimelineError::Timeline(e) => Self::NotFound(format!("{e}").into()),
571 : }
572 0 : }
573 : }
574 :
575 : impl From<WaitLsnError> for PageStreamError {
576 0 : fn from(value: WaitLsnError) -> Self {
577 0 : match value {
578 0 : e @ WaitLsnError::Timeout(_) => Self::LsnTimeout(e),
579 0 : WaitLsnError::Shutdown => Self::Shutdown,
580 0 : e @ WaitLsnError::BadState { .. } => Self::Reconnect(format!("{e}").into()),
581 : }
582 0 : }
583 : }
584 :
585 : impl From<WaitLsnError> for QueryError {
586 0 : fn from(value: WaitLsnError) -> Self {
587 0 : match value {
588 0 : e @ WaitLsnError::Timeout(_) => Self::Other(anyhow::Error::new(e)),
589 0 : WaitLsnError::Shutdown => Self::Shutdown,
590 0 : WaitLsnError::BadState { .. } => Self::Reconnect,
591 : }
592 0 : }
593 : }
594 :
595 : #[derive(thiserror::Error, Debug)]
596 : struct BatchedPageStreamError {
597 : req: PagestreamRequest,
598 : err: PageStreamError,
599 : }
600 :
601 : impl std::fmt::Display for BatchedPageStreamError {
602 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
603 0 : self.err.fmt(f)
604 0 : }
605 : }
606 :
607 : struct BatchedGetPageRequest {
608 : req: PagestreamGetPageRequest,
609 : timer: SmgrOpTimer,
610 : }
611 :
612 : #[cfg(feature = "testing")]
613 : struct BatchedTestRequest {
614 : req: models::PagestreamTestRequest,
615 : timer: SmgrOpTimer,
616 : }
617 :
618 : /// NB: we only hold [`timeline::handle::WeakHandle`] inside this enum,
619 : /// so that we don't keep the [`Timeline::gate`] open while the batch
620 : /// is being built up inside the [`spsc_fold`] (pagestream pipelining).
621 : #[derive(IntoStaticStr)]
622 : enum BatchedFeMessage {
623 : Exists {
624 : span: Span,
625 : timer: SmgrOpTimer,
626 : shard: timeline::handle::WeakHandle<TenantManagerTypes>,
627 : req: models::PagestreamExistsRequest,
628 : },
629 : Nblocks {
630 : span: Span,
631 : timer: SmgrOpTimer,
632 : shard: timeline::handle::WeakHandle<TenantManagerTypes>,
633 : req: models::PagestreamNblocksRequest,
634 : },
635 : GetPage {
636 : span: Span,
637 : shard: timeline::handle::WeakHandle<TenantManagerTypes>,
638 : effective_request_lsn: Lsn,
639 : pages: smallvec::SmallVec<[BatchedGetPageRequest; 1]>,
640 : },
641 : DbSize {
642 : span: Span,
643 : timer: SmgrOpTimer,
644 : shard: timeline::handle::WeakHandle<TenantManagerTypes>,
645 : req: models::PagestreamDbSizeRequest,
646 : },
647 : GetSlruSegment {
648 : span: Span,
649 : timer: SmgrOpTimer,
650 : shard: timeline::handle::WeakHandle<TenantManagerTypes>,
651 : req: models::PagestreamGetSlruSegmentRequest,
652 : },
653 : #[cfg(feature = "testing")]
654 : Test {
655 : span: Span,
656 : shard: timeline::handle::WeakHandle<TenantManagerTypes>,
657 : requests: Vec<BatchedTestRequest>,
658 : },
659 : RespondError {
660 : span: Span,
661 : error: BatchedPageStreamError,
662 : },
663 : }
664 :
665 : impl BatchedFeMessage {
666 0 : fn as_static_str(&self) -> &'static str {
667 0 : self.into()
668 0 : }
669 :
670 0 : fn observe_execution_start(&mut self, at: Instant) {
671 0 : match self {
672 0 : BatchedFeMessage::Exists { timer, .. }
673 0 : | BatchedFeMessage::Nblocks { timer, .. }
674 0 : | BatchedFeMessage::DbSize { timer, .. }
675 0 : | BatchedFeMessage::GetSlruSegment { timer, .. } => {
676 0 : timer.observe_execution_start(at);
677 0 : }
678 0 : BatchedFeMessage::GetPage { pages, .. } => {
679 0 : for page in pages {
680 0 : page.timer.observe_execution_start(at);
681 0 : }
682 : }
683 : #[cfg(feature = "testing")]
684 0 : BatchedFeMessage::Test { requests, .. } => {
685 0 : for req in requests {
686 0 : req.timer.observe_execution_start(at);
687 0 : }
688 : }
689 0 : BatchedFeMessage::RespondError { .. } => {}
690 : }
691 0 : }
692 : }
693 :
694 : impl PageServerHandler {
695 0 : pub fn new(
696 0 : conf: &'static PageServerConf,
697 0 : tenant_manager: Arc<TenantManager>,
698 0 : auth: Option<Arc<SwappableJwtAuth>>,
699 0 : pipelining_config: PageServicePipeliningConfig,
700 0 : connection_ctx: RequestContext,
701 0 : cancel: CancellationToken,
702 0 : gate_guard: GateGuard,
703 0 : ) -> Self {
704 0 : PageServerHandler {
705 0 : conf,
706 0 : auth,
707 0 : claims: None,
708 0 : connection_ctx,
709 0 : timeline_handles: Some(TimelineHandles::new(tenant_manager)),
710 0 : cancel,
711 0 : pipelining_config,
712 0 : gate_guard,
713 0 : }
714 0 : }
715 :
716 : /// This function always respects cancellation of any timeline in `[Self::shard_timelines]`. Pass in
717 : /// a cancellation token at the next scope up (such as a tenant cancellation token) to ensure we respect
718 : /// cancellation if there aren't any timelines in the cache.
719 : ///
720 : /// If calling from a function that doesn't use the `[Self::shard_timelines]` cache, then pass in the
721 : /// timeline cancellation token.
722 0 : async fn flush_cancellable<IO>(
723 0 : &self,
724 0 : pgb: &mut PostgresBackend<IO>,
725 0 : cancel: &CancellationToken,
726 0 : ) -> Result<(), QueryError>
727 0 : where
728 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
729 0 : {
730 0 : tokio::select!(
731 0 : flush_r = pgb.flush() => {
732 0 : Ok(flush_r?)
733 : },
734 0 : _ = cancel.cancelled() => {
735 0 : Err(QueryError::Shutdown)
736 : }
737 : )
738 0 : }
739 :
740 : #[allow(clippy::too_many_arguments)]
741 0 : async fn pagestream_read_message<IO>(
742 0 : pgb: &mut PostgresBackendReader<IO>,
743 0 : tenant_id: TenantId,
744 0 : timeline_id: TimelineId,
745 0 : timeline_handles: &mut TimelineHandles,
746 0 : cancel: &CancellationToken,
747 0 : ctx: &RequestContext,
748 0 : protocol_version: PagestreamProtocolVersion,
749 0 : parent_span: Span,
750 0 : ) -> Result<Option<BatchedFeMessage>, QueryError>
751 0 : where
752 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
753 0 : {
754 0 : let msg = tokio::select! {
755 : biased;
756 0 : _ = cancel.cancelled() => {
757 0 : return Err(QueryError::Shutdown)
758 : }
759 0 : msg = pgb.read_message() => { msg }
760 0 : };
761 0 :
762 0 : let received_at = Instant::now();
763 :
764 0 : let copy_data_bytes = match msg? {
765 0 : Some(FeMessage::CopyData(bytes)) => bytes,
766 : Some(FeMessage::Terminate) => {
767 0 : return Ok(None);
768 : }
769 0 : Some(m) => {
770 0 : return Err(QueryError::Other(anyhow::anyhow!(
771 0 : "unexpected message: {m:?} during COPY"
772 0 : )));
773 : }
774 : None => {
775 0 : return Ok(None);
776 : } // client disconnected
777 : };
778 0 : trace!("query: {copy_data_bytes:?}");
779 :
780 0 : fail::fail_point!("ps::handle-pagerequest-message");
781 :
782 : // parse request
783 0 : let neon_fe_msg =
784 0 : PagestreamFeMessage::parse(&mut copy_data_bytes.reader(), protocol_version)?;
785 :
786 : // TODO: turn in to async closure once available to avoid repeating received_at
787 0 : async fn record_op_start_and_throttle(
788 0 : shard: &timeline::handle::Handle<TenantManagerTypes>,
789 0 : op: metrics::SmgrQueryType,
790 0 : received_at: Instant,
791 0 : ) -> Result<SmgrOpTimer, QueryError> {
792 0 : // It's important to start the smgr op metric recorder as early as possible
793 0 : // so that the _started counters are incremented before we do
794 0 : // any serious waiting, e.g., for throttle, batching, or actual request handling.
795 0 : let mut timer = shard.query_metrics.start_smgr_op(op, received_at);
796 0 : let now = Instant::now();
797 0 : timer.observe_throttle_start(now);
798 0 : let throttled = tokio::select! {
799 0 : res = shard.pagestream_throttle.throttle(1, now) => res,
800 0 : _ = shard.cancel.cancelled() => return Err(QueryError::Shutdown),
801 : };
802 0 : timer.observe_throttle_done(throttled);
803 0 : Ok(timer)
804 0 : }
805 :
806 0 : let batched_msg = match neon_fe_msg {
807 0 : PagestreamFeMessage::Exists(req) => {
808 0 : let shard = timeline_handles
809 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
810 0 : .await?;
811 0 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
812 0 : let span = tracing::info_span!(parent: &parent_span, "handle_get_rel_exists_request", rel = %req.rel, req_lsn = %req.hdr.request_lsn, shard_id = %shard.tenant_shard_id.shard_slug());
813 0 : let timer = record_op_start_and_throttle(
814 0 : &shard,
815 0 : metrics::SmgrQueryType::GetRelExists,
816 0 : received_at,
817 0 : )
818 0 : .await?;
819 0 : BatchedFeMessage::Exists {
820 0 : span,
821 0 : timer,
822 0 : shard: shard.downgrade(),
823 0 : req,
824 0 : }
825 : }
826 0 : PagestreamFeMessage::Nblocks(req) => {
827 0 : let shard = timeline_handles
828 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
829 0 : .await?;
830 0 : let span = tracing::info_span!(parent: &parent_span, "handle_get_nblocks_request", rel = %req.rel, req_lsn = %req.hdr.request_lsn, shard_id = %shard.tenant_shard_id.shard_slug());
831 0 : let timer = record_op_start_and_throttle(
832 0 : &shard,
833 0 : metrics::SmgrQueryType::GetRelSize,
834 0 : received_at,
835 0 : )
836 0 : .await?;
837 0 : BatchedFeMessage::Nblocks {
838 0 : span,
839 0 : timer,
840 0 : shard: shard.downgrade(),
841 0 : req,
842 0 : }
843 : }
844 0 : PagestreamFeMessage::DbSize(req) => {
845 0 : let shard = timeline_handles
846 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
847 0 : .await?;
848 0 : let span = tracing::info_span!(parent: &parent_span, "handle_db_size_request", dbnode = %req.dbnode, req_lsn = %req.hdr.request_lsn, shard_id = %shard.tenant_shard_id.shard_slug());
849 0 : let timer = record_op_start_and_throttle(
850 0 : &shard,
851 0 : metrics::SmgrQueryType::GetDbSize,
852 0 : received_at,
853 0 : )
854 0 : .await?;
855 0 : BatchedFeMessage::DbSize {
856 0 : span,
857 0 : timer,
858 0 : shard: shard.downgrade(),
859 0 : req,
860 0 : }
861 : }
862 0 : PagestreamFeMessage::GetSlruSegment(req) => {
863 0 : let shard = timeline_handles
864 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
865 0 : .await?;
866 0 : let span = tracing::info_span!(parent: &parent_span, "handle_get_slru_segment_request", kind = %req.kind, segno = %req.segno, req_lsn = %req.hdr.request_lsn, shard_id = %shard.tenant_shard_id.shard_slug());
867 0 : let timer = record_op_start_and_throttle(
868 0 : &shard,
869 0 : metrics::SmgrQueryType::GetSlruSegment,
870 0 : received_at,
871 0 : )
872 0 : .await?;
873 0 : BatchedFeMessage::GetSlruSegment {
874 0 : span,
875 0 : timer,
876 0 : shard: shard.downgrade(),
877 0 : req,
878 0 : }
879 : }
880 0 : PagestreamFeMessage::GetPage(req) => {
881 : // avoid a somewhat costly Span::record() by constructing the entire span in one go.
882 : macro_rules! mkspan {
883 : (before shard routing) => {{
884 : tracing::info_span!(parent: &parent_span, "handle_get_page_request", rel = %req.rel, blkno = %req.blkno, req_lsn = %req.hdr.request_lsn)
885 : }};
886 : ($shard_id:expr) => {{
887 : tracing::info_span!(parent: &parent_span, "handle_get_page_request", rel = %req.rel, blkno = %req.blkno, req_lsn = %req.hdr.request_lsn, shard_id = %$shard_id)
888 : }};
889 : }
890 :
891 : macro_rules! respond_error {
892 : ($span:expr, $error:expr) => {{
893 : let error = BatchedFeMessage::RespondError {
894 : span: $span,
895 : error: BatchedPageStreamError {
896 : req: req.hdr,
897 : err: $error,
898 : },
899 : };
900 : Ok(Some(error))
901 : }};
902 : }
903 :
904 0 : let key = rel_block_to_key(req.rel, req.blkno);
905 0 : let shard = match timeline_handles
906 0 : .get(tenant_id, timeline_id, ShardSelector::Page(key))
907 0 : .await
908 : {
909 0 : Ok(tl) => tl,
910 0 : Err(e) => {
911 0 : let span = mkspan!(before shard routing);
912 0 : match e {
913 : GetActiveTimelineError::Tenant(GetActiveTenantError::NotFound(_)) => {
914 : // We already know this tenant exists in general, because we resolved it at
915 : // start of connection. Getting a NotFound here indicates that the shard containing
916 : // the requested page is not present on this node: the client's knowledge of shard->pageserver
917 : // mapping is out of date.
918 : //
919 : // Closing the connection by returning ``::Reconnect` has the side effect of rate-limiting above message, via
920 : // client's reconnect backoff, as well as hopefully prompting the client to load its updated configuration
921 : // and talk to a different pageserver.
922 0 : return respond_error!(
923 0 : span,
924 0 : PageStreamError::Reconnect(
925 0 : "getpage@lsn request routed to wrong shard".into()
926 0 : )
927 0 : );
928 : }
929 0 : e => {
930 0 : return respond_error!(span, e.into());
931 : }
932 : }
933 : }
934 : };
935 0 : let span = mkspan!(shard.tenant_shard_id.shard_slug());
936 :
937 0 : let timer = record_op_start_and_throttle(
938 0 : &shard,
939 0 : metrics::SmgrQueryType::GetPageAtLsn,
940 0 : received_at,
941 0 : )
942 0 : .await?;
943 :
944 : // We're holding the Handle
945 0 : let effective_request_lsn = match Self::wait_or_get_last_lsn(
946 0 : &shard,
947 0 : req.hdr.request_lsn,
948 0 : req.hdr.not_modified_since,
949 0 : &shard.get_applied_gc_cutoff_lsn(),
950 0 : ctx,
951 0 : )
952 0 : // TODO: if we actually need to wait for lsn here, it delays the entire batch which doesn't need to wait
953 0 : .await
954 : {
955 0 : Ok(lsn) => lsn,
956 0 : Err(e) => {
957 0 : return respond_error!(span, e);
958 : }
959 : };
960 : BatchedFeMessage::GetPage {
961 0 : span,
962 0 : shard: shard.downgrade(),
963 0 : effective_request_lsn,
964 0 : pages: smallvec::smallvec![BatchedGetPageRequest { req, timer }],
965 : }
966 : }
967 : #[cfg(feature = "testing")]
968 0 : PagestreamFeMessage::Test(req) => {
969 0 : let shard = timeline_handles
970 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
971 0 : .await?;
972 0 : let span = tracing::info_span!(parent: &parent_span, "handle_test_request", shard_id = %shard.tenant_shard_id.shard_slug());
973 0 : let timer =
974 0 : record_op_start_and_throttle(&shard, metrics::SmgrQueryType::Test, received_at)
975 0 : .await?;
976 0 : BatchedFeMessage::Test {
977 0 : span,
978 0 : shard: shard.downgrade(),
979 0 : requests: vec![BatchedTestRequest { req, timer }],
980 0 : }
981 : }
982 : };
983 0 : Ok(Some(batched_msg))
984 0 : }
985 :
986 : /// Post-condition: `batch` is Some()
987 : #[instrument(skip_all, level = tracing::Level::TRACE)]
988 : #[allow(clippy::boxed_local)]
989 : fn pagestream_do_batch(
990 : max_batch_size: NonZeroUsize,
991 : batch: &mut Result<BatchedFeMessage, QueryError>,
992 : this_msg: Result<BatchedFeMessage, QueryError>,
993 : ) -> Result<(), Result<BatchedFeMessage, QueryError>> {
994 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
995 :
996 : let this_msg = match this_msg {
997 : Ok(this_msg) => this_msg,
998 : Err(e) => return Err(Err(e)),
999 : };
1000 :
1001 : match (&mut *batch, this_msg) {
1002 : // something batched already, let's see if we can add this message to the batch
1003 : (
1004 : Ok(BatchedFeMessage::GetPage {
1005 : span: _,
1006 : shard: accum_shard,
1007 : pages: accum_pages,
1008 : effective_request_lsn: accum_lsn,
1009 : }),
1010 : BatchedFeMessage::GetPage {
1011 : span: _,
1012 : shard: this_shard,
1013 : pages: this_pages,
1014 : effective_request_lsn: this_lsn,
1015 : },
1016 0 : ) if (|| {
1017 0 : assert_eq!(this_pages.len(), 1);
1018 0 : if accum_pages.len() >= max_batch_size.get() {
1019 0 : trace!(%accum_lsn, %this_lsn, %max_batch_size, "stopping batching because of batch size");
1020 0 : assert_eq!(accum_pages.len(), max_batch_size.get());
1021 0 : return false;
1022 0 : }
1023 0 : if !accum_shard.is_same_handle_as(&this_shard) {
1024 0 : trace!(%accum_lsn, %this_lsn, "stopping batching because timeline object mismatch");
1025 : // TODO: we _could_ batch & execute each shard seperately (and in parallel).
1026 : // But the current logic for keeping responses in order does not support that.
1027 0 : return false;
1028 0 : }
1029 0 : // the vectored get currently only supports a single LSN, so, bounce as soon
1030 0 : // as the effective request_lsn changes
1031 0 : if *accum_lsn != this_lsn {
1032 0 : trace!(%accum_lsn, %this_lsn, "stopping batching because LSN changed");
1033 0 : return false;
1034 0 : }
1035 0 : true
1036 : })() =>
1037 : {
1038 : // ok to batch
1039 : accum_pages.extend(this_pages);
1040 : Ok(())
1041 : }
1042 : #[cfg(feature = "testing")]
1043 : (
1044 : Ok(BatchedFeMessage::Test {
1045 : shard: accum_shard,
1046 : requests: accum_requests,
1047 : ..
1048 : }),
1049 : BatchedFeMessage::Test {
1050 : shard: this_shard,
1051 : requests: this_requests,
1052 : ..
1053 : },
1054 0 : ) if (|| {
1055 0 : assert!(this_requests.len() == 1);
1056 0 : if accum_requests.len() >= max_batch_size.get() {
1057 0 : trace!(%max_batch_size, "stopping batching because of batch size");
1058 0 : assert_eq!(accum_requests.len(), max_batch_size.get());
1059 0 : return false;
1060 0 : }
1061 0 : if !accum_shard.is_same_handle_as(&this_shard) {
1062 0 : trace!("stopping batching because timeline object mismatch");
1063 : // TODO: we _could_ batch & execute each shard seperately (and in parallel).
1064 : // But the current logic for keeping responses in order does not support that.
1065 0 : return false;
1066 0 : }
1067 0 : let this_batch_key = this_requests[0].req.batch_key;
1068 0 : let accum_batch_key = accum_requests[0].req.batch_key;
1069 0 : if this_requests[0].req.batch_key != accum_requests[0].req.batch_key {
1070 0 : trace!(%accum_batch_key, %this_batch_key, "stopping batching because batch key changed");
1071 0 : return false;
1072 0 : }
1073 0 : true
1074 : })() =>
1075 : {
1076 : // ok to batch
1077 : accum_requests.extend(this_requests);
1078 : Ok(())
1079 : }
1080 : // something batched already but this message is unbatchable
1081 : (_, this_msg) => {
1082 : // by default, don't continue batching
1083 : Err(Ok(this_msg))
1084 : }
1085 : }
1086 : }
1087 :
1088 0 : #[instrument(level = tracing::Level::DEBUG, skip_all)]
1089 : async fn pagesteam_handle_batched_message<IO>(
1090 : &mut self,
1091 : pgb_writer: &mut PostgresBackend<IO>,
1092 : batch: BatchedFeMessage,
1093 : io_concurrency: IoConcurrency,
1094 : cancel: &CancellationToken,
1095 : protocol_version: PagestreamProtocolVersion,
1096 : ctx: &RequestContext,
1097 : ) -> Result<(), QueryError>
1098 : where
1099 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
1100 : {
1101 : let started_at = Instant::now();
1102 : let batch = {
1103 : let mut batch = batch;
1104 : batch.observe_execution_start(started_at);
1105 : batch
1106 : };
1107 :
1108 : // Dispatch the batch to the appropriate request handler.
1109 : let (mut handler_results, span) = log_slow(
1110 : batch.as_static_str(),
1111 : LOG_SLOW_GETPAGE_THRESHOLD,
1112 : self.pagestream_dispatch_batched_message(batch, io_concurrency, ctx),
1113 : )
1114 : .await?;
1115 :
1116 : // We purposefully don't count flush time into the smgr operation timer.
1117 : //
1118 : // The reason is that current compute client will not perform protocol processing
1119 : // if the postgres backend process is doing things other than `->smgr_read()`.
1120 : // This is especially the case for prefetch.
1121 : //
1122 : // If the compute doesn't read from the connection, eventually TCP will backpressure
1123 : // all the way into our flush call below.
1124 : //
1125 : // The timer's underlying metric is used for a storage-internal latency SLO and
1126 : // we don't want to include latency in it that we can't control.
1127 : // And as pointed out above, in this case, we don't control the time that flush will take.
1128 : //
1129 : // We put each response in the batch onto the wire in a separate pgb_writer.flush()
1130 : // call, which (all unmeasured) adds syscall overhead but reduces time to first byte
1131 : // and avoids building up a "giant" contiguous userspace buffer to hold the entire response.
1132 : // TODO: vectored socket IO would be great, but pgb_writer doesn't support that.
1133 : let flush_timers = {
1134 : let flushing_start_time = Instant::now();
1135 : let mut flush_timers = Vec::with_capacity(handler_results.len());
1136 : for handler_result in &mut handler_results {
1137 : let flush_timer = match handler_result {
1138 : Ok((_, timer)) => Some(
1139 : timer
1140 : .observe_execution_end(flushing_start_time)
1141 : .expect("we are the first caller"),
1142 : ),
1143 : Err(_) => {
1144 : // TODO: measure errors
1145 : None
1146 : }
1147 : };
1148 : flush_timers.push(flush_timer);
1149 : }
1150 : assert_eq!(flush_timers.len(), handler_results.len());
1151 : flush_timers
1152 : };
1153 :
1154 : // Map handler result to protocol behavior.
1155 : // Some handler errors cause exit from pagestream protocol.
1156 : // Other handler errors are sent back as an error message and we stay in pagestream protocol.
1157 : for (handler_result, flushing_timer) in handler_results.into_iter().zip(flush_timers) {
1158 : let response_msg = match handler_result {
1159 : Err(e) => match &e.err {
1160 : PageStreamError::Shutdown => {
1161 : // If we fail to fulfil a request during shutdown, which may be _because_ of
1162 : // shutdown, then do not send the error to the client. Instead just drop the
1163 : // connection.
1164 0 : span.in_scope(|| info!("dropping connection due to shutdown"));
1165 : return Err(QueryError::Shutdown);
1166 : }
1167 : PageStreamError::Reconnect(reason) => {
1168 0 : span.in_scope(|| info!("handler requested reconnect: {reason}"));
1169 : return Err(QueryError::Reconnect);
1170 : }
1171 : PageStreamError::Read(_)
1172 : | PageStreamError::LsnTimeout(_)
1173 : | PageStreamError::NotFound(_)
1174 : | PageStreamError::BadRequest(_) => {
1175 : // print the all details to the log with {:#}, but for the client the
1176 : // error message is enough. Do not log if shutting down, as the anyhow::Error
1177 : // here includes cancellation which is not an error.
1178 : let full = utils::error::report_compact_sources(&e.err);
1179 0 : span.in_scope(|| {
1180 0 : error!("error reading relation or page version: {full:#}")
1181 0 : });
1182 :
1183 : PagestreamBeMessage::Error(PagestreamErrorResponse {
1184 : req: e.req,
1185 : message: e.err.to_string(),
1186 : })
1187 : }
1188 : },
1189 : Ok((response_msg, _op_timer_already_observed)) => response_msg,
1190 : };
1191 :
1192 : //
1193 : // marshal & transmit response message
1194 : //
1195 :
1196 : pgb_writer.write_message_noflush(&BeMessage::CopyData(
1197 : &response_msg.serialize(protocol_version),
1198 : ))?;
1199 :
1200 : failpoint_support::sleep_millis_async!("before-pagestream-msg-flush", cancel);
1201 :
1202 : // what we want to do
1203 : let socket_fd = pgb_writer.socket_fd;
1204 : let flush_fut = pgb_writer.flush();
1205 : // metric for how long flushing takes
1206 : let flush_fut = match flushing_timer {
1207 : Some(flushing_timer) => futures::future::Either::Left(flushing_timer.measure(
1208 : Instant::now(),
1209 : flush_fut,
1210 : socket_fd,
1211 : )),
1212 : None => futures::future::Either::Right(flush_fut),
1213 : };
1214 : // do it while respecting cancellation
1215 0 : let _: () = async move {
1216 0 : tokio::select! {
1217 : biased;
1218 0 : _ = cancel.cancelled() => {
1219 : // We were requested to shut down.
1220 0 : info!("shutdown request received in page handler");
1221 0 : return Err(QueryError::Shutdown)
1222 : }
1223 0 : res = flush_fut => {
1224 0 : res?;
1225 : }
1226 : }
1227 0 : Ok(())
1228 0 : }
1229 : .await?;
1230 : }
1231 : Ok(())
1232 : }
1233 :
1234 : /// Helper which dispatches a batched message to the appropriate handler.
1235 : /// Returns a vec of results, along with the extracted trace span.
1236 0 : async fn pagestream_dispatch_batched_message(
1237 0 : &mut self,
1238 0 : batch: BatchedFeMessage,
1239 0 : io_concurrency: IoConcurrency,
1240 0 : ctx: &RequestContext,
1241 0 : ) -> Result<
1242 0 : (
1243 0 : Vec<Result<(PagestreamBeMessage, SmgrOpTimer), BatchedPageStreamError>>,
1244 0 : Span,
1245 0 : ),
1246 0 : QueryError,
1247 0 : > {
1248 : macro_rules! upgrade_handle_and_set_context {
1249 : ($shard:ident) => {{
1250 : let weak_handle = &$shard;
1251 : let handle = weak_handle.upgrade()?;
1252 : let ctx = ctx.with_scope_page_service_pagestream(&handle);
1253 : (handle, ctx)
1254 : }};
1255 : }
1256 0 : Ok(match batch {
1257 : BatchedFeMessage::Exists {
1258 0 : span,
1259 0 : timer,
1260 0 : shard,
1261 0 : req,
1262 0 : } => {
1263 0 : fail::fail_point!("ps::handle-pagerequest-message::exists");
1264 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1265 : (
1266 0 : vec![
1267 0 : self.handle_get_rel_exists_request(&shard, &req, &ctx)
1268 0 : .instrument(span.clone())
1269 0 : .await
1270 0 : .map(|msg| (msg, timer))
1271 0 : .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
1272 0 : ],
1273 0 : span,
1274 : )
1275 : }
1276 : BatchedFeMessage::Nblocks {
1277 0 : span,
1278 0 : timer,
1279 0 : shard,
1280 0 : req,
1281 0 : } => {
1282 0 : fail::fail_point!("ps::handle-pagerequest-message::nblocks");
1283 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1284 : (
1285 0 : vec![
1286 0 : self.handle_get_nblocks_request(&shard, &req, &ctx)
1287 0 : .instrument(span.clone())
1288 0 : .await
1289 0 : .map(|msg| (msg, timer))
1290 0 : .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
1291 0 : ],
1292 0 : span,
1293 : )
1294 : }
1295 : BatchedFeMessage::GetPage {
1296 0 : span,
1297 0 : shard,
1298 0 : effective_request_lsn,
1299 0 : pages,
1300 0 : } => {
1301 0 : fail::fail_point!("ps::handle-pagerequest-message::getpage");
1302 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1303 : (
1304 : {
1305 0 : let npages = pages.len();
1306 0 : trace!(npages, "handling getpage request");
1307 0 : let res = self
1308 0 : .handle_get_page_at_lsn_request_batched(
1309 0 : &shard,
1310 0 : effective_request_lsn,
1311 0 : pages,
1312 0 : io_concurrency,
1313 0 : &ctx,
1314 0 : )
1315 0 : .instrument(span.clone())
1316 0 : .await;
1317 0 : assert_eq!(res.len(), npages);
1318 0 : res
1319 0 : },
1320 0 : span,
1321 : )
1322 : }
1323 : BatchedFeMessage::DbSize {
1324 0 : span,
1325 0 : timer,
1326 0 : shard,
1327 0 : req,
1328 0 : } => {
1329 0 : fail::fail_point!("ps::handle-pagerequest-message::dbsize");
1330 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1331 : (
1332 0 : vec![
1333 0 : self.handle_db_size_request(&shard, &req, &ctx)
1334 0 : .instrument(span.clone())
1335 0 : .await
1336 0 : .map(|msg| (msg, timer))
1337 0 : .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
1338 0 : ],
1339 0 : span,
1340 : )
1341 : }
1342 : BatchedFeMessage::GetSlruSegment {
1343 0 : span,
1344 0 : timer,
1345 0 : shard,
1346 0 : req,
1347 0 : } => {
1348 0 : fail::fail_point!("ps::handle-pagerequest-message::slrusegment");
1349 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1350 : (
1351 0 : vec![
1352 0 : self.handle_get_slru_segment_request(&shard, &req, &ctx)
1353 0 : .instrument(span.clone())
1354 0 : .await
1355 0 : .map(|msg| (msg, timer))
1356 0 : .map_err(|err| BatchedPageStreamError { err, req: req.hdr }),
1357 0 : ],
1358 0 : span,
1359 : )
1360 : }
1361 : #[cfg(feature = "testing")]
1362 : BatchedFeMessage::Test {
1363 0 : span,
1364 0 : shard,
1365 0 : requests,
1366 0 : } => {
1367 0 : fail::fail_point!("ps::handle-pagerequest-message::test");
1368 0 : let (shard, ctx) = upgrade_handle_and_set_context!(shard);
1369 : (
1370 : {
1371 0 : let npages = requests.len();
1372 0 : trace!(npages, "handling getpage request");
1373 0 : let res = self
1374 0 : .handle_test_request_batch(&shard, requests, &ctx)
1375 0 : .instrument(span.clone())
1376 0 : .await;
1377 0 : assert_eq!(res.len(), npages);
1378 0 : res
1379 0 : },
1380 0 : span,
1381 : )
1382 : }
1383 0 : BatchedFeMessage::RespondError { span, error } => {
1384 0 : // We've already decided to respond with an error, so we don't need to
1385 0 : // call the handler.
1386 0 : (vec![Err(error)], span)
1387 : }
1388 : })
1389 0 : }
1390 :
1391 : /// Pagestream sub-protocol handler.
1392 : ///
1393 : /// It is a simple request-response protocol inside a COPYBOTH session.
1394 : ///
1395 : /// # Coding Discipline
1396 : ///
1397 : /// Coding discipline within this function: all interaction with the `pgb` connection
1398 : /// needs to be sensitive to connection shutdown, currently signalled via [`Self::cancel`].
1399 : /// This is so that we can shutdown page_service quickly.
1400 : #[instrument(skip_all)]
1401 : async fn handle_pagerequests<IO>(
1402 : &mut self,
1403 : pgb: &mut PostgresBackend<IO>,
1404 : tenant_id: TenantId,
1405 : timeline_id: TimelineId,
1406 : protocol_version: PagestreamProtocolVersion,
1407 : ctx: RequestContext,
1408 : ) -> Result<(), QueryError>
1409 : where
1410 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
1411 : {
1412 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
1413 :
1414 : // switch client to COPYBOTH
1415 : pgb.write_message_noflush(&BeMessage::CopyBothResponse)?;
1416 : tokio::select! {
1417 : biased;
1418 : _ = self.cancel.cancelled() => {
1419 : return Err(QueryError::Shutdown)
1420 : }
1421 : res = pgb.flush() => {
1422 : res?;
1423 : }
1424 : }
1425 :
1426 : let io_concurrency = IoConcurrency::spawn_from_conf(
1427 : self.conf,
1428 : match self.gate_guard.try_clone() {
1429 : Ok(guard) => guard,
1430 : Err(_) => {
1431 : info!("shutdown request received in page handler");
1432 : return Err(QueryError::Shutdown);
1433 : }
1434 : },
1435 : );
1436 :
1437 : let pgb_reader = pgb
1438 : .split()
1439 : .context("implementation error: split pgb into reader and writer")?;
1440 :
1441 : let timeline_handles = self
1442 : .timeline_handles
1443 : .take()
1444 : .expect("implementation error: timeline_handles should not be locked");
1445 :
1446 : let request_span = info_span!("request");
1447 : let ((pgb_reader, timeline_handles), result) = match self.pipelining_config.clone() {
1448 : PageServicePipeliningConfig::Pipelined(pipelining_config) => {
1449 : self.handle_pagerequests_pipelined(
1450 : pgb,
1451 : pgb_reader,
1452 : tenant_id,
1453 : timeline_id,
1454 : timeline_handles,
1455 : request_span,
1456 : pipelining_config,
1457 : protocol_version,
1458 : io_concurrency,
1459 : &ctx,
1460 : )
1461 : .await
1462 : }
1463 : PageServicePipeliningConfig::Serial => {
1464 : self.handle_pagerequests_serial(
1465 : pgb,
1466 : pgb_reader,
1467 : tenant_id,
1468 : timeline_id,
1469 : timeline_handles,
1470 : request_span,
1471 : protocol_version,
1472 : io_concurrency,
1473 : &ctx,
1474 : )
1475 : .await
1476 : }
1477 : };
1478 :
1479 : debug!("pagestream subprotocol shut down cleanly");
1480 :
1481 : pgb.unsplit(pgb_reader)
1482 : .context("implementation error: unsplit pgb")?;
1483 :
1484 : let replaced = self.timeline_handles.replace(timeline_handles);
1485 : assert!(replaced.is_none());
1486 :
1487 : result
1488 : }
1489 :
1490 : #[allow(clippy::too_many_arguments)]
1491 0 : async fn handle_pagerequests_serial<IO>(
1492 0 : &mut self,
1493 0 : pgb_writer: &mut PostgresBackend<IO>,
1494 0 : mut pgb_reader: PostgresBackendReader<IO>,
1495 0 : tenant_id: TenantId,
1496 0 : timeline_id: TimelineId,
1497 0 : mut timeline_handles: TimelineHandles,
1498 0 : request_span: Span,
1499 0 : protocol_version: PagestreamProtocolVersion,
1500 0 : io_concurrency: IoConcurrency,
1501 0 : ctx: &RequestContext,
1502 0 : ) -> (
1503 0 : (PostgresBackendReader<IO>, TimelineHandles),
1504 0 : Result<(), QueryError>,
1505 0 : )
1506 0 : where
1507 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
1508 0 : {
1509 0 : let cancel = self.cancel.clone();
1510 0 : let err = loop {
1511 0 : let msg = Self::pagestream_read_message(
1512 0 : &mut pgb_reader,
1513 0 : tenant_id,
1514 0 : timeline_id,
1515 0 : &mut timeline_handles,
1516 0 : &cancel,
1517 0 : ctx,
1518 0 : protocol_version,
1519 0 : request_span.clone(),
1520 0 : )
1521 0 : .await;
1522 0 : let msg = match msg {
1523 0 : Ok(msg) => msg,
1524 0 : Err(e) => break e,
1525 : };
1526 0 : let msg = match msg {
1527 0 : Some(msg) => msg,
1528 : None => {
1529 0 : debug!("pagestream subprotocol end observed");
1530 0 : return ((pgb_reader, timeline_handles), Ok(()));
1531 : }
1532 : };
1533 :
1534 0 : let result = self
1535 0 : .pagesteam_handle_batched_message(
1536 0 : pgb_writer,
1537 0 : msg,
1538 0 : io_concurrency.clone(),
1539 0 : &cancel,
1540 0 : protocol_version,
1541 0 : ctx,
1542 0 : )
1543 0 : .await;
1544 0 : match result {
1545 0 : Ok(()) => {}
1546 0 : Err(e) => break e,
1547 : }
1548 : };
1549 0 : ((pgb_reader, timeline_handles), Err(err))
1550 0 : }
1551 :
1552 : /// # Cancel-Safety
1553 : ///
1554 : /// May leak tokio tasks if not polled to completion.
1555 : #[allow(clippy::too_many_arguments)]
1556 0 : async fn handle_pagerequests_pipelined<IO>(
1557 0 : &mut self,
1558 0 : pgb_writer: &mut PostgresBackend<IO>,
1559 0 : pgb_reader: PostgresBackendReader<IO>,
1560 0 : tenant_id: TenantId,
1561 0 : timeline_id: TimelineId,
1562 0 : mut timeline_handles: TimelineHandles,
1563 0 : request_span: Span,
1564 0 : pipelining_config: PageServicePipeliningConfigPipelined,
1565 0 : protocol_version: PagestreamProtocolVersion,
1566 0 : io_concurrency: IoConcurrency,
1567 0 : ctx: &RequestContext,
1568 0 : ) -> (
1569 0 : (PostgresBackendReader<IO>, TimelineHandles),
1570 0 : Result<(), QueryError>,
1571 0 : )
1572 0 : where
1573 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
1574 0 : {
1575 0 : //
1576 0 : // Pipelined pagestream handling consists of
1577 0 : // - a Batcher that reads requests off the wire and
1578 0 : // and batches them if possible,
1579 0 : // - an Executor that processes the batched requests.
1580 0 : //
1581 0 : // The batch is built up inside an `spsc_fold` channel,
1582 0 : // shared betwen Batcher (Sender) and Executor (Receiver).
1583 0 : //
1584 0 : // The Batcher continously folds client requests into the batch,
1585 0 : // while the Executor can at any time take out what's in the batch
1586 0 : // in order to process it.
1587 0 : // This means the next batch builds up while the Executor
1588 0 : // executes the last batch.
1589 0 : //
1590 0 : // CANCELLATION
1591 0 : //
1592 0 : // We run both Batcher and Executor futures to completion before
1593 0 : // returning from this function.
1594 0 : //
1595 0 : // If Executor exits first, it signals cancellation to the Batcher
1596 0 : // via a CancellationToken that is child of `self.cancel`.
1597 0 : // If Batcher exits first, it signals cancellation to the Executor
1598 0 : // by dropping the spsc_fold channel Sender.
1599 0 : //
1600 0 : // CLEAN SHUTDOWN
1601 0 : //
1602 0 : // Clean shutdown means that the client ends the COPYBOTH session.
1603 0 : // In response to such a client message, the Batcher exits.
1604 0 : // The Executor continues to run, draining the spsc_fold channel.
1605 0 : // Once drained, the spsc_fold recv will fail with a distinct error
1606 0 : // indicating that the sender disconnected.
1607 0 : // The Executor exits with Ok(()) in response to that error.
1608 0 : //
1609 0 : // Server initiated shutdown is not clean shutdown, but instead
1610 0 : // is an error Err(QueryError::Shutdown) that is propagated through
1611 0 : // error propagation.
1612 0 : //
1613 0 : // ERROR PROPAGATION
1614 0 : //
1615 0 : // When the Batcher encounter an error, it sends it as a value
1616 0 : // through the spsc_fold channel and exits afterwards.
1617 0 : // When the Executor observes such an error in the channel,
1618 0 : // it exits returning that error value.
1619 0 : //
1620 0 : // This design ensures that the Executor stage will still process
1621 0 : // the batch that was in flight when the Batcher encountered an error,
1622 0 : // thereby beahving identical to a serial implementation.
1623 0 :
1624 0 : let PageServicePipeliningConfigPipelined {
1625 0 : max_batch_size,
1626 0 : execution,
1627 0 : } = pipelining_config;
1628 :
1629 : // Macro to _define_ a pipeline stage.
1630 : macro_rules! pipeline_stage {
1631 : ($name:literal, $cancel:expr, $make_fut:expr) => {{
1632 : let cancel: CancellationToken = $cancel;
1633 : let stage_fut = $make_fut(cancel.clone());
1634 0 : async move {
1635 0 : scopeguard::defer! {
1636 0 : debug!("exiting");
1637 0 : }
1638 0 : timed_after_cancellation(stage_fut, $name, Duration::from_millis(100), &cancel)
1639 0 : .await
1640 0 : }
1641 : .instrument(tracing::info_span!($name))
1642 : }};
1643 : }
1644 :
1645 : //
1646 : // Batcher
1647 : //
1648 :
1649 0 : let cancel_batcher = self.cancel.child_token();
1650 0 : let (mut batch_tx, mut batch_rx) = spsc_fold::channel();
1651 0 : let batcher = pipeline_stage!("batcher", cancel_batcher.clone(), move |cancel_batcher| {
1652 0 : let ctx = ctx.attached_child();
1653 0 : async move {
1654 0 : let mut pgb_reader = pgb_reader;
1655 0 : let mut exit = false;
1656 0 : while !exit {
1657 0 : let read_res = Self::pagestream_read_message(
1658 0 : &mut pgb_reader,
1659 0 : tenant_id,
1660 0 : timeline_id,
1661 0 : &mut timeline_handles,
1662 0 : &cancel_batcher,
1663 0 : &ctx,
1664 0 : protocol_version,
1665 0 : request_span.clone(),
1666 0 : )
1667 0 : .await;
1668 0 : let Some(read_res) = read_res.transpose() else {
1669 0 : debug!("client-initiated shutdown");
1670 0 : break;
1671 : };
1672 0 : exit |= read_res.is_err();
1673 0 : let could_send = batch_tx
1674 0 : .send(read_res, |batch, res| {
1675 0 : Self::pagestream_do_batch(max_batch_size, batch, res)
1676 0 : })
1677 0 : .await;
1678 0 : exit |= could_send.is_err();
1679 : }
1680 0 : (pgb_reader, timeline_handles)
1681 0 : }
1682 0 : });
1683 :
1684 : //
1685 : // Executor
1686 : //
1687 :
1688 0 : let executor = pipeline_stage!("executor", self.cancel.clone(), move |cancel| {
1689 0 : let ctx = ctx.attached_child();
1690 0 : async move {
1691 0 : let _cancel_batcher = cancel_batcher.drop_guard();
1692 : loop {
1693 0 : let maybe_batch = batch_rx.recv().await;
1694 0 : let batch = match maybe_batch {
1695 0 : Ok(batch) => batch,
1696 : Err(spsc_fold::RecvError::SenderGone) => {
1697 0 : debug!("upstream gone");
1698 0 : return Ok(());
1699 : }
1700 : };
1701 0 : let batch = match batch {
1702 0 : Ok(batch) => batch,
1703 0 : Err(e) => {
1704 0 : return Err(e);
1705 : }
1706 : };
1707 0 : self.pagesteam_handle_batched_message(
1708 0 : pgb_writer,
1709 0 : batch,
1710 0 : io_concurrency.clone(),
1711 0 : &cancel,
1712 0 : protocol_version,
1713 0 : &ctx,
1714 0 : )
1715 0 : .await?;
1716 : }
1717 0 : }
1718 0 : });
1719 :
1720 : //
1721 : // Execute the stages.
1722 : //
1723 :
1724 0 : match execution {
1725 : PageServiceProtocolPipelinedExecutionStrategy::ConcurrentFutures => {
1726 0 : tokio::join!(batcher, executor)
1727 : }
1728 : PageServiceProtocolPipelinedExecutionStrategy::Tasks => {
1729 : // These tasks are not tracked anywhere.
1730 0 : let read_messages_task = tokio::spawn(batcher);
1731 0 : let (read_messages_task_res, executor_res_) =
1732 0 : tokio::join!(read_messages_task, executor,);
1733 0 : (
1734 0 : read_messages_task_res.expect("propagated panic from read_messages"),
1735 0 : executor_res_,
1736 0 : )
1737 : }
1738 : }
1739 0 : }
1740 :
1741 : /// Helper function to handle the LSN from client request.
1742 : ///
1743 : /// Each GetPage (and Exists and Nblocks) request includes information about
1744 : /// which version of the page is being requested. The primary compute node
1745 : /// will always request the latest page version, by setting 'request_lsn' to
1746 : /// the last inserted or flushed WAL position, while a standby will request
1747 : /// a version at the LSN that it's currently caught up to.
1748 : ///
1749 : /// In either case, if the page server hasn't received the WAL up to the
1750 : /// requested LSN yet, we will wait for it to arrive. The return value is
1751 : /// the LSN that should be used to look up the page versions.
1752 : ///
1753 : /// In addition to the request LSN, each request carries another LSN,
1754 : /// 'not_modified_since', which is a hint to the pageserver that the client
1755 : /// knows that the page has not been modified between 'not_modified_since'
1756 : /// and the request LSN. This allows skipping the wait, as long as the WAL
1757 : /// up to 'not_modified_since' has arrived. If the client doesn't have any
1758 : /// information about when the page was modified, it will use
1759 : /// not_modified_since == lsn. If the client lies and sends a too low
1760 : /// not_modified_hint such that there are in fact later page versions, the
1761 : /// behavior is undefined: the pageserver may return any of the page versions
1762 : /// or an error.
1763 0 : async fn wait_or_get_last_lsn(
1764 0 : timeline: &Timeline,
1765 0 : request_lsn: Lsn,
1766 0 : not_modified_since: Lsn,
1767 0 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1768 0 : ctx: &RequestContext,
1769 0 : ) -> Result<Lsn, PageStreamError> {
1770 0 : let last_record_lsn = timeline.get_last_record_lsn();
1771 0 :
1772 0 : // Sanity check the request
1773 0 : if request_lsn < not_modified_since {
1774 0 : return Err(PageStreamError::BadRequest(
1775 0 : format!(
1776 0 : "invalid request with request LSN {} and not_modified_since {}",
1777 0 : request_lsn, not_modified_since,
1778 0 : )
1779 0 : .into(),
1780 0 : ));
1781 0 : }
1782 0 :
1783 0 : // Check explicitly for INVALID just to get a less scary error message if the request is obviously bogus
1784 0 : if request_lsn == Lsn::INVALID {
1785 0 : return Err(PageStreamError::BadRequest(
1786 0 : "invalid LSN(0) in request".into(),
1787 0 : ));
1788 0 : }
1789 0 :
1790 0 : // Clients should only read from recent LSNs on their timeline, or from locations holding an LSN lease.
1791 0 : //
1792 0 : // We may have older data available, but we make a best effort to detect this case and return an error,
1793 0 : // to distinguish a misbehaving client (asking for old LSN) from a storage issue (data missing at a legitimate LSN).
1794 0 : if request_lsn < **latest_gc_cutoff_lsn && !timeline.is_gc_blocked_by_lsn_lease_deadline() {
1795 0 : let gc_info = &timeline.gc_info.read().unwrap();
1796 0 : if !gc_info.lsn_covered_by_lease(request_lsn) {
1797 0 : return Err(
1798 0 : PageStreamError::BadRequest(format!(
1799 0 : "tried to request a page version that was garbage collected. requested at {} gc cutoff {}",
1800 0 : request_lsn, **latest_gc_cutoff_lsn
1801 0 : ).into())
1802 0 : );
1803 0 : }
1804 0 : }
1805 :
1806 : // Wait for WAL up to 'not_modified_since' to arrive, if necessary
1807 0 : if not_modified_since > last_record_lsn {
1808 0 : timeline
1809 0 : .wait_lsn(
1810 0 : not_modified_since,
1811 0 : crate::tenant::timeline::WaitLsnWaiter::PageService,
1812 0 : timeline::WaitLsnTimeout::Default,
1813 0 : ctx,
1814 0 : )
1815 0 : .await?;
1816 : // Since we waited for 'not_modified_since' to arrive, that is now the last
1817 : // record LSN. (Or close enough for our purposes; the last-record LSN can
1818 : // advance immediately after we return anyway)
1819 0 : Ok(not_modified_since)
1820 : } else {
1821 : // It might be better to use max(not_modified_since, latest_gc_cutoff_lsn)
1822 : // here instead. That would give the same result, since we know that there
1823 : // haven't been any modifications since 'not_modified_since'. Using an older
1824 : // LSN might be faster, because that could allow skipping recent layers when
1825 : // finding the page. However, we have historically used 'last_record_lsn', so
1826 : // stick to that for now.
1827 0 : Ok(std::cmp::min(last_record_lsn, request_lsn))
1828 : }
1829 0 : }
1830 :
1831 : /// Handles the lsn lease request.
1832 : /// If a lease cannot be obtained, the client will receive NULL.
1833 : #[instrument(skip_all, fields(shard_id, %lsn))]
1834 : async fn handle_make_lsn_lease<IO>(
1835 : &mut self,
1836 : pgb: &mut PostgresBackend<IO>,
1837 : tenant_shard_id: TenantShardId,
1838 : timeline_id: TimelineId,
1839 : lsn: Lsn,
1840 : ctx: &RequestContext,
1841 : ) -> Result<(), QueryError>
1842 : where
1843 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
1844 : {
1845 : let timeline = self
1846 : .timeline_handles
1847 : .as_mut()
1848 : .unwrap()
1849 : .get(
1850 : tenant_shard_id.tenant_id,
1851 : timeline_id,
1852 : ShardSelector::Known(tenant_shard_id.to_index()),
1853 : )
1854 : .await?;
1855 : set_tracing_field_shard_id(&timeline);
1856 :
1857 : let lease = timeline
1858 : .renew_lsn_lease(lsn, timeline.get_lsn_lease_length(), ctx)
1859 0 : .inspect_err(|e| {
1860 0 : warn!("{e}");
1861 0 : })
1862 : .ok();
1863 0 : let valid_until_str = lease.map(|l| {
1864 0 : l.valid_until
1865 0 : .duration_since(SystemTime::UNIX_EPOCH)
1866 0 : .expect("valid_until is earlier than UNIX_EPOCH")
1867 0 : .as_millis()
1868 0 : .to_string()
1869 0 : });
1870 :
1871 : info!(
1872 : "acquired lease for {} until {}",
1873 : lsn,
1874 : valid_until_str.as_deref().unwrap_or("<unknown>")
1875 : );
1876 :
1877 0 : let bytes = valid_until_str.as_ref().map(|x| x.as_bytes());
1878 :
1879 : pgb.write_message_noflush(&BeMessage::RowDescription(&[RowDescriptor::text_col(
1880 : b"valid_until",
1881 : )]))?
1882 : .write_message_noflush(&BeMessage::DataRow(&[bytes]))?;
1883 :
1884 : Ok(())
1885 : }
1886 :
1887 : #[instrument(skip_all, fields(shard_id))]
1888 : async fn handle_get_rel_exists_request(
1889 : &mut self,
1890 : timeline: &Timeline,
1891 : req: &PagestreamExistsRequest,
1892 : ctx: &RequestContext,
1893 : ) -> Result<PagestreamBeMessage, PageStreamError> {
1894 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
1895 : let lsn = Self::wait_or_get_last_lsn(
1896 : timeline,
1897 : req.hdr.request_lsn,
1898 : req.hdr.not_modified_since,
1899 : &latest_gc_cutoff_lsn,
1900 : ctx,
1901 : )
1902 : .await?;
1903 :
1904 : let exists = timeline
1905 : .get_rel_exists(req.rel, Version::Lsn(lsn), ctx)
1906 : .await?;
1907 :
1908 : Ok(PagestreamBeMessage::Exists(PagestreamExistsResponse {
1909 : req: *req,
1910 : exists,
1911 : }))
1912 : }
1913 :
1914 : #[instrument(skip_all, fields(shard_id))]
1915 : async fn handle_get_nblocks_request(
1916 : &mut self,
1917 : timeline: &Timeline,
1918 : req: &PagestreamNblocksRequest,
1919 : ctx: &RequestContext,
1920 : ) -> Result<PagestreamBeMessage, PageStreamError> {
1921 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
1922 : let lsn = Self::wait_or_get_last_lsn(
1923 : timeline,
1924 : req.hdr.request_lsn,
1925 : req.hdr.not_modified_since,
1926 : &latest_gc_cutoff_lsn,
1927 : ctx,
1928 : )
1929 : .await?;
1930 :
1931 : let n_blocks = timeline
1932 : .get_rel_size(req.rel, Version::Lsn(lsn), ctx)
1933 : .await?;
1934 :
1935 : Ok(PagestreamBeMessage::Nblocks(PagestreamNblocksResponse {
1936 : req: *req,
1937 : n_blocks,
1938 : }))
1939 : }
1940 :
1941 : #[instrument(skip_all, fields(shard_id))]
1942 : async fn handle_db_size_request(
1943 : &mut self,
1944 : timeline: &Timeline,
1945 : req: &PagestreamDbSizeRequest,
1946 : ctx: &RequestContext,
1947 : ) -> Result<PagestreamBeMessage, PageStreamError> {
1948 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
1949 : let lsn = Self::wait_or_get_last_lsn(
1950 : timeline,
1951 : req.hdr.request_lsn,
1952 : req.hdr.not_modified_since,
1953 : &latest_gc_cutoff_lsn,
1954 : ctx,
1955 : )
1956 : .await?;
1957 :
1958 : let total_blocks = timeline
1959 : .get_db_size(DEFAULTTABLESPACE_OID, req.dbnode, Version::Lsn(lsn), ctx)
1960 : .await?;
1961 : let db_size = total_blocks as i64 * BLCKSZ as i64;
1962 :
1963 : Ok(PagestreamBeMessage::DbSize(PagestreamDbSizeResponse {
1964 : req: *req,
1965 : db_size,
1966 : }))
1967 : }
1968 :
1969 : #[instrument(skip_all)]
1970 : async fn handle_get_page_at_lsn_request_batched(
1971 : &mut self,
1972 : timeline: &Timeline,
1973 : effective_lsn: Lsn,
1974 : requests: smallvec::SmallVec<[BatchedGetPageRequest; 1]>,
1975 : io_concurrency: IoConcurrency,
1976 : ctx: &RequestContext,
1977 : ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer), BatchedPageStreamError>> {
1978 : debug_assert_current_span_has_tenant_and_timeline_id();
1979 :
1980 : timeline
1981 : .query_metrics
1982 : .observe_getpage_batch_start(requests.len());
1983 :
1984 : // If a page trace is running, submit an event for this request.
1985 : if let Some(page_trace) = timeline.page_trace.load().as_ref() {
1986 : let time = SystemTime::now();
1987 : for batch in &requests {
1988 : let key = rel_block_to_key(batch.req.rel, batch.req.blkno).to_compact();
1989 : // Ignore error (trace buffer may be full or tracer may have disconnected).
1990 : _ = page_trace.try_send(PageTraceEvent {
1991 : key,
1992 : effective_lsn,
1993 : time,
1994 : });
1995 : }
1996 : }
1997 :
1998 : let results = timeline
1999 : .get_rel_page_at_lsn_batched(
2000 0 : requests.iter().map(|p| (&p.req.rel, &p.req.blkno)),
2001 : effective_lsn,
2002 : io_concurrency,
2003 : ctx,
2004 : )
2005 : .await;
2006 : assert_eq!(results.len(), requests.len());
2007 :
2008 : // TODO: avoid creating the new Vec here
2009 : Vec::from_iter(
2010 : requests
2011 : .into_iter()
2012 : .zip(results.into_iter())
2013 0 : .map(|(req, res)| {
2014 0 : res.map(|page| {
2015 0 : (
2016 0 : PagestreamBeMessage::GetPage(models::PagestreamGetPageResponse {
2017 0 : req: req.req,
2018 0 : page,
2019 0 : }),
2020 0 : req.timer,
2021 0 : )
2022 0 : })
2023 0 : .map_err(|e| BatchedPageStreamError {
2024 0 : err: PageStreamError::from(e),
2025 0 : req: req.req.hdr,
2026 0 : })
2027 0 : }),
2028 : )
2029 : }
2030 :
2031 : #[instrument(skip_all, fields(shard_id))]
2032 : async fn handle_get_slru_segment_request(
2033 : &mut self,
2034 : timeline: &Timeline,
2035 : req: &PagestreamGetSlruSegmentRequest,
2036 : ctx: &RequestContext,
2037 : ) -> Result<PagestreamBeMessage, PageStreamError> {
2038 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
2039 : let lsn = Self::wait_or_get_last_lsn(
2040 : timeline,
2041 : req.hdr.request_lsn,
2042 : req.hdr.not_modified_since,
2043 : &latest_gc_cutoff_lsn,
2044 : ctx,
2045 : )
2046 : .await?;
2047 :
2048 : let kind = SlruKind::from_repr(req.kind)
2049 : .ok_or(PageStreamError::BadRequest("invalid SLRU kind".into()))?;
2050 : let segment = timeline.get_slru_segment(kind, req.segno, lsn, ctx).await?;
2051 :
2052 : Ok(PagestreamBeMessage::GetSlruSegment(
2053 : PagestreamGetSlruSegmentResponse { req: *req, segment },
2054 : ))
2055 : }
2056 :
2057 : // NB: this impl mimics what we do for batched getpage requests.
2058 : #[cfg(feature = "testing")]
2059 : #[instrument(skip_all, fields(shard_id))]
2060 : async fn handle_test_request_batch(
2061 : &mut self,
2062 : timeline: &Timeline,
2063 : requests: Vec<BatchedTestRequest>,
2064 : _ctx: &RequestContext,
2065 : ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer), BatchedPageStreamError>> {
2066 : // real requests would do something with the timeline
2067 : let mut results = Vec::with_capacity(requests.len());
2068 : for _req in requests.iter() {
2069 : tokio::task::yield_now().await;
2070 :
2071 : results.push({
2072 : if timeline.cancel.is_cancelled() {
2073 : Err(PageReconstructError::Cancelled)
2074 : } else {
2075 : Ok(())
2076 : }
2077 : });
2078 : }
2079 :
2080 : // TODO: avoid creating the new Vec here
2081 : Vec::from_iter(
2082 : requests
2083 : .into_iter()
2084 : .zip(results.into_iter())
2085 0 : .map(|(req, res)| {
2086 0 : res.map(|()| {
2087 0 : (
2088 0 : PagestreamBeMessage::Test(models::PagestreamTestResponse {
2089 0 : req: req.req.clone(),
2090 0 : }),
2091 0 : req.timer,
2092 0 : )
2093 0 : })
2094 0 : .map_err(|e| BatchedPageStreamError {
2095 0 : err: PageStreamError::from(e),
2096 0 : req: req.req.hdr,
2097 0 : })
2098 0 : }),
2099 : )
2100 : }
2101 :
2102 : /// Note on "fullbackup":
2103 : /// Full basebackups should only be used for debugging purposes.
2104 : /// Originally, it was introduced to enable breaking storage format changes,
2105 : /// but that is not applicable anymore.
2106 : ///
2107 : /// # Coding Discipline
2108 : ///
2109 : /// Coding discipline within this function: all interaction with the `pgb` connection
2110 : /// needs to be sensitive to connection shutdown, currently signalled via [`Self::cancel`].
2111 : /// This is so that we can shutdown page_service quickly.
2112 : ///
2113 : /// TODO: wrap the pgb that we pass to the basebackup handler so that it's sensitive
2114 : /// to connection cancellation.
2115 : #[allow(clippy::too_many_arguments)]
2116 : #[instrument(skip_all, fields(shard_id, ?lsn, ?prev_lsn, %full_backup))]
2117 : async fn handle_basebackup_request<IO>(
2118 : &mut self,
2119 : pgb: &mut PostgresBackend<IO>,
2120 : tenant_id: TenantId,
2121 : timeline_id: TimelineId,
2122 : lsn: Option<Lsn>,
2123 : prev_lsn: Option<Lsn>,
2124 : full_backup: bool,
2125 : gzip: bool,
2126 : replica: bool,
2127 : ctx: &RequestContext,
2128 : ) -> Result<(), QueryError>
2129 : where
2130 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
2131 : {
2132 0 : fn map_basebackup_error(err: BasebackupError) -> QueryError {
2133 0 : match err {
2134 : // TODO: passthrough the error site to the final error message?
2135 0 : BasebackupError::Client(e, _) => QueryError::Disconnected(ConnectionError::Io(e)),
2136 0 : BasebackupError::Server(e) => QueryError::Other(e),
2137 0 : BasebackupError::Shutdown => QueryError::Shutdown,
2138 : }
2139 0 : }
2140 :
2141 : let started = std::time::Instant::now();
2142 :
2143 : let timeline = self
2144 : .timeline_handles
2145 : .as_mut()
2146 : .unwrap()
2147 : .get(tenant_id, timeline_id, ShardSelector::Zero)
2148 : .await?;
2149 : set_tracing_field_shard_id(&timeline);
2150 : let ctx = ctx.with_scope_timeline(&timeline);
2151 :
2152 : if timeline.is_archived() == Some(true) {
2153 : tracing::info!(
2154 : "timeline {tenant_id}/{timeline_id} is archived, but got basebackup request for it."
2155 : );
2156 : return Err(QueryError::NotFound("timeline is archived".into()));
2157 : }
2158 :
2159 : let latest_gc_cutoff_lsn = timeline.get_applied_gc_cutoff_lsn();
2160 : if let Some(lsn) = lsn {
2161 : // Backup was requested at a particular LSN. Wait for it to arrive.
2162 : info!("waiting for {}", lsn);
2163 : timeline
2164 : .wait_lsn(
2165 : lsn,
2166 : crate::tenant::timeline::WaitLsnWaiter::PageService,
2167 : crate::tenant::timeline::WaitLsnTimeout::Default,
2168 : &ctx,
2169 : )
2170 : .await?;
2171 : timeline
2172 : .check_lsn_is_in_scope(lsn, &latest_gc_cutoff_lsn)
2173 : .context("invalid basebackup lsn")?;
2174 : }
2175 :
2176 : let lsn_awaited_after = started.elapsed();
2177 :
2178 : // switch client to COPYOUT
2179 : pgb.write_message_noflush(&BeMessage::CopyOutResponse)
2180 : .map_err(QueryError::Disconnected)?;
2181 : self.flush_cancellable(pgb, &self.cancel).await?;
2182 :
2183 : // Send a tarball of the latest layer on the timeline. Compress if not
2184 : // fullbackup. TODO Compress in that case too (tests need to be updated)
2185 : if full_backup {
2186 : let mut writer = pgb.copyout_writer();
2187 : basebackup::send_basebackup_tarball(
2188 : &mut writer,
2189 : &timeline,
2190 : lsn,
2191 : prev_lsn,
2192 : full_backup,
2193 : replica,
2194 : &ctx,
2195 : )
2196 : .await
2197 : .map_err(map_basebackup_error)?;
2198 : } else {
2199 : let mut writer = BufWriter::new(pgb.copyout_writer());
2200 : if gzip {
2201 : let mut encoder = GzipEncoder::with_quality(
2202 : &mut writer,
2203 : // NOTE using fast compression because it's on the critical path
2204 : // for compute startup. For an empty database, we get
2205 : // <100KB with this method. The Level::Best compression method
2206 : // gives us <20KB, but maybe we should add basebackup caching
2207 : // on compute shutdown first.
2208 : async_compression::Level::Fastest,
2209 : );
2210 : basebackup::send_basebackup_tarball(
2211 : &mut encoder,
2212 : &timeline,
2213 : lsn,
2214 : prev_lsn,
2215 : full_backup,
2216 : replica,
2217 : &ctx,
2218 : )
2219 : .await
2220 : .map_err(map_basebackup_error)?;
2221 : // shutdown the encoder to ensure the gzip footer is written
2222 : encoder
2223 : .shutdown()
2224 : .await
2225 0 : .map_err(|e| QueryError::Disconnected(ConnectionError::Io(e)))?;
2226 : } else {
2227 : basebackup::send_basebackup_tarball(
2228 : &mut writer,
2229 : &timeline,
2230 : lsn,
2231 : prev_lsn,
2232 : full_backup,
2233 : replica,
2234 : &ctx,
2235 : )
2236 : .await
2237 : .map_err(map_basebackup_error)?;
2238 : }
2239 0 : writer.flush().await.map_err(|e| {
2240 0 : map_basebackup_error(BasebackupError::Client(
2241 0 : e,
2242 0 : "handle_basebackup_request,flush",
2243 0 : ))
2244 0 : })?;
2245 : }
2246 :
2247 : pgb.write_message_noflush(&BeMessage::CopyDone)
2248 : .map_err(QueryError::Disconnected)?;
2249 : self.flush_cancellable(pgb, &timeline.cancel).await?;
2250 :
2251 : let basebackup_after = started
2252 : .elapsed()
2253 : .checked_sub(lsn_awaited_after)
2254 : .unwrap_or(Duration::ZERO);
2255 :
2256 : info!(
2257 : lsn_await_millis = lsn_awaited_after.as_millis(),
2258 : basebackup_millis = basebackup_after.as_millis(),
2259 : "basebackup complete"
2260 : );
2261 :
2262 : Ok(())
2263 : }
2264 :
2265 : // when accessing management api supply None as an argument
2266 : // when using to authorize tenant pass corresponding tenant id
2267 0 : fn check_permission(&self, tenant_id: Option<TenantId>) -> Result<(), QueryError> {
2268 0 : if self.auth.is_none() {
2269 : // auth is set to Trust, nothing to check so just return ok
2270 0 : return Ok(());
2271 0 : }
2272 0 : // auth is some, just checked above, when auth is some
2273 0 : // then claims are always present because of checks during connection init
2274 0 : // so this expect won't trigger
2275 0 : let claims = self
2276 0 : .claims
2277 0 : .as_ref()
2278 0 : .expect("claims presence already checked");
2279 0 : check_permission(claims, tenant_id).map_err(|e| QueryError::Unauthorized(e.0))
2280 0 : }
2281 : }
2282 :
2283 : /// `basebackup tenant timeline [lsn] [--gzip] [--replica]`
2284 : #[derive(Debug, Clone, Eq, PartialEq)]
2285 : struct BaseBackupCmd {
2286 : tenant_id: TenantId,
2287 : timeline_id: TimelineId,
2288 : lsn: Option<Lsn>,
2289 : gzip: bool,
2290 : replica: bool,
2291 : }
2292 :
2293 : /// `fullbackup tenant timeline [lsn] [prev_lsn]`
2294 : #[derive(Debug, Clone, Eq, PartialEq)]
2295 : struct FullBackupCmd {
2296 : tenant_id: TenantId,
2297 : timeline_id: TimelineId,
2298 : lsn: Option<Lsn>,
2299 : prev_lsn: Option<Lsn>,
2300 : }
2301 :
2302 : /// `pagestream_v2 tenant timeline`
2303 : #[derive(Debug, Clone, Eq, PartialEq)]
2304 : struct PageStreamCmd {
2305 : tenant_id: TenantId,
2306 : timeline_id: TimelineId,
2307 : protocol_version: PagestreamProtocolVersion,
2308 : }
2309 :
2310 : /// `lease lsn tenant timeline lsn`
2311 : #[derive(Debug, Clone, Eq, PartialEq)]
2312 : struct LeaseLsnCmd {
2313 : tenant_shard_id: TenantShardId,
2314 : timeline_id: TimelineId,
2315 : lsn: Lsn,
2316 : }
2317 :
2318 : #[derive(Debug, Clone, Eq, PartialEq)]
2319 : enum PageServiceCmd {
2320 : Set,
2321 : PageStream(PageStreamCmd),
2322 : BaseBackup(BaseBackupCmd),
2323 : FullBackup(FullBackupCmd),
2324 : LeaseLsn(LeaseLsnCmd),
2325 : }
2326 :
2327 : impl PageStreamCmd {
2328 12 : fn parse(query: &str, protocol_version: PagestreamProtocolVersion) -> anyhow::Result<Self> {
2329 12 : let parameters = query.split_whitespace().collect_vec();
2330 12 : if parameters.len() != 2 {
2331 4 : bail!(
2332 4 : "invalid number of parameters for pagestream command: {}",
2333 4 : query
2334 4 : );
2335 8 : }
2336 8 : let tenant_id = TenantId::from_str(parameters[0])
2337 8 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
2338 4 : let timeline_id = TimelineId::from_str(parameters[1])
2339 4 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
2340 4 : Ok(Self {
2341 4 : tenant_id,
2342 4 : timeline_id,
2343 4 : protocol_version,
2344 4 : })
2345 12 : }
2346 : }
2347 :
2348 : impl FullBackupCmd {
2349 8 : fn parse(query: &str) -> anyhow::Result<Self> {
2350 8 : let parameters = query.split_whitespace().collect_vec();
2351 8 : if parameters.len() < 2 || parameters.len() > 4 {
2352 0 : bail!(
2353 0 : "invalid number of parameters for basebackup command: {}",
2354 0 : query
2355 0 : );
2356 8 : }
2357 8 : let tenant_id = TenantId::from_str(parameters[0])
2358 8 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
2359 8 : let timeline_id = TimelineId::from_str(parameters[1])
2360 8 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
2361 : // The caller is responsible for providing correct lsn and prev_lsn.
2362 8 : let lsn = if let Some(lsn_str) = parameters.get(2) {
2363 : Some(
2364 4 : Lsn::from_str(lsn_str)
2365 4 : .with_context(|| format!("Failed to parse Lsn from {lsn_str}"))?,
2366 : )
2367 : } else {
2368 4 : None
2369 : };
2370 8 : let prev_lsn = if let Some(prev_lsn_str) = parameters.get(3) {
2371 : Some(
2372 4 : Lsn::from_str(prev_lsn_str)
2373 4 : .with_context(|| format!("Failed to parse Lsn from {prev_lsn_str}"))?,
2374 : )
2375 : } else {
2376 4 : None
2377 : };
2378 8 : Ok(Self {
2379 8 : tenant_id,
2380 8 : timeline_id,
2381 8 : lsn,
2382 8 : prev_lsn,
2383 8 : })
2384 8 : }
2385 : }
2386 :
2387 : impl BaseBackupCmd {
2388 36 : fn parse(query: &str) -> anyhow::Result<Self> {
2389 36 : let parameters = query.split_whitespace().collect_vec();
2390 36 : if parameters.len() < 2 {
2391 0 : bail!(
2392 0 : "invalid number of parameters for basebackup command: {}",
2393 0 : query
2394 0 : );
2395 36 : }
2396 36 : let tenant_id = TenantId::from_str(parameters[0])
2397 36 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
2398 36 : let timeline_id = TimelineId::from_str(parameters[1])
2399 36 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
2400 : let lsn;
2401 : let flags_parse_from;
2402 36 : if let Some(maybe_lsn) = parameters.get(2) {
2403 32 : if *maybe_lsn == "latest" {
2404 4 : lsn = None;
2405 4 : flags_parse_from = 3;
2406 28 : } else if maybe_lsn.starts_with("--") {
2407 20 : lsn = None;
2408 20 : flags_parse_from = 2;
2409 20 : } else {
2410 : lsn = Some(
2411 8 : Lsn::from_str(maybe_lsn)
2412 8 : .with_context(|| format!("Failed to parse lsn from {maybe_lsn}"))?,
2413 : );
2414 8 : flags_parse_from = 3;
2415 : }
2416 4 : } else {
2417 4 : lsn = None;
2418 4 : flags_parse_from = 2;
2419 4 : }
2420 :
2421 36 : let mut gzip = false;
2422 36 : let mut replica = false;
2423 :
2424 44 : for ¶m in ¶meters[flags_parse_from..] {
2425 44 : match param {
2426 44 : "--gzip" => {
2427 28 : if gzip {
2428 4 : bail!("duplicate parameter for basebackup command: {param}")
2429 24 : }
2430 24 : gzip = true
2431 : }
2432 16 : "--replica" => {
2433 8 : if replica {
2434 0 : bail!("duplicate parameter for basebackup command: {param}")
2435 8 : }
2436 8 : replica = true
2437 : }
2438 8 : _ => bail!("invalid parameter for basebackup command: {param}"),
2439 : }
2440 : }
2441 24 : Ok(Self {
2442 24 : tenant_id,
2443 24 : timeline_id,
2444 24 : lsn,
2445 24 : gzip,
2446 24 : replica,
2447 24 : })
2448 36 : }
2449 : }
2450 :
2451 : impl LeaseLsnCmd {
2452 8 : fn parse(query: &str) -> anyhow::Result<Self> {
2453 8 : let parameters = query.split_whitespace().collect_vec();
2454 8 : if parameters.len() != 3 {
2455 0 : bail!(
2456 0 : "invalid number of parameters for lease lsn command: {}",
2457 0 : query
2458 0 : );
2459 8 : }
2460 8 : let tenant_shard_id = TenantShardId::from_str(parameters[0])
2461 8 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
2462 8 : let timeline_id = TimelineId::from_str(parameters[1])
2463 8 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
2464 8 : let lsn = Lsn::from_str(parameters[2])
2465 8 : .with_context(|| format!("Failed to parse lsn from {}", parameters[2]))?;
2466 8 : Ok(Self {
2467 8 : tenant_shard_id,
2468 8 : timeline_id,
2469 8 : lsn,
2470 8 : })
2471 8 : }
2472 : }
2473 :
2474 : impl PageServiceCmd {
2475 84 : fn parse(query: &str) -> anyhow::Result<Self> {
2476 84 : let query = query.trim();
2477 84 : let Some((cmd, other)) = query.split_once(' ') else {
2478 8 : bail!("cannot parse query: {query}")
2479 : };
2480 76 : match cmd.to_ascii_lowercase().as_str() {
2481 76 : "pagestream_v2" => Ok(Self::PageStream(PageStreamCmd::parse(
2482 12 : other,
2483 12 : PagestreamProtocolVersion::V2,
2484 12 : )?)),
2485 64 : "pagestream_v3" => Ok(Self::PageStream(PageStreamCmd::parse(
2486 0 : other,
2487 0 : PagestreamProtocolVersion::V3,
2488 0 : )?)),
2489 64 : "basebackup" => Ok(Self::BaseBackup(BaseBackupCmd::parse(other)?)),
2490 28 : "fullbackup" => Ok(Self::FullBackup(FullBackupCmd::parse(other)?)),
2491 20 : "lease" => {
2492 12 : let Some((cmd2, other)) = other.split_once(' ') else {
2493 0 : bail!("invalid lease command: {cmd}");
2494 : };
2495 12 : let cmd2 = cmd2.to_ascii_lowercase();
2496 12 : if cmd2 == "lsn" {
2497 8 : Ok(Self::LeaseLsn(LeaseLsnCmd::parse(other)?))
2498 : } else {
2499 4 : bail!("invalid lease command: {cmd}");
2500 : }
2501 : }
2502 8 : "set" => Ok(Self::Set),
2503 0 : _ => Err(anyhow::anyhow!("unsupported command {cmd} in {query}")),
2504 : }
2505 84 : }
2506 : }
2507 :
2508 : impl<IO> postgres_backend::Handler<IO> for PageServerHandler
2509 : where
2510 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
2511 : {
2512 0 : fn check_auth_jwt(
2513 0 : &mut self,
2514 0 : _pgb: &mut PostgresBackend<IO>,
2515 0 : jwt_response: &[u8],
2516 0 : ) -> Result<(), QueryError> {
2517 : // this unwrap is never triggered, because check_auth_jwt only called when auth_type is NeonJWT
2518 : // which requires auth to be present
2519 0 : let data = self
2520 0 : .auth
2521 0 : .as_ref()
2522 0 : .unwrap()
2523 0 : .decode(str::from_utf8(jwt_response).context("jwt response is not UTF-8")?)
2524 0 : .map_err(|e| QueryError::Unauthorized(e.0))?;
2525 :
2526 0 : if matches!(data.claims.scope, Scope::Tenant) && data.claims.tenant_id.is_none() {
2527 0 : return Err(QueryError::Unauthorized(
2528 0 : "jwt token scope is Tenant, but tenant id is missing".into(),
2529 0 : ));
2530 0 : }
2531 0 :
2532 0 : debug!(
2533 0 : "jwt scope check succeeded for scope: {:#?} by tenant id: {:?}",
2534 : data.claims.scope, data.claims.tenant_id,
2535 : );
2536 :
2537 0 : self.claims = Some(data.claims);
2538 0 : Ok(())
2539 0 : }
2540 :
2541 0 : fn startup(
2542 0 : &mut self,
2543 0 : _pgb: &mut PostgresBackend<IO>,
2544 0 : sm: &FeStartupPacket,
2545 0 : ) -> Result<(), QueryError> {
2546 0 : fail::fail_point!("ps::connection-start::startup-packet");
2547 :
2548 0 : if let FeStartupPacket::StartupMessage { params, .. } = sm {
2549 0 : if let Some(app_name) = params.get("application_name") {
2550 0 : Span::current().record("application_name", field::display(app_name));
2551 0 : }
2552 0 : };
2553 :
2554 0 : Ok(())
2555 0 : }
2556 :
2557 : #[instrument(skip_all, fields(tenant_id, timeline_id))]
2558 : async fn process_query(
2559 : &mut self,
2560 : pgb: &mut PostgresBackend<IO>,
2561 : query_string: &str,
2562 : ) -> Result<(), QueryError> {
2563 0 : fail::fail_point!("simulated-bad-compute-connection", |_| {
2564 0 : info!("Hit failpoint for bad connection");
2565 0 : Err(QueryError::SimulatedConnectionError)
2566 0 : });
2567 :
2568 : fail::fail_point!("ps::connection-start::process-query");
2569 :
2570 : let ctx = self.connection_ctx.attached_child();
2571 : debug!("process query {query_string}");
2572 : let query = PageServiceCmd::parse(query_string)?;
2573 : match query {
2574 : PageServiceCmd::PageStream(PageStreamCmd {
2575 : tenant_id,
2576 : timeline_id,
2577 : protocol_version,
2578 : }) => {
2579 : tracing::Span::current()
2580 : .record("tenant_id", field::display(tenant_id))
2581 : .record("timeline_id", field::display(timeline_id));
2582 :
2583 : self.check_permission(Some(tenant_id))?;
2584 : let command_kind = match protocol_version {
2585 : PagestreamProtocolVersion::V2 => ComputeCommandKind::PageStreamV2,
2586 : PagestreamProtocolVersion::V3 => ComputeCommandKind::PageStreamV3,
2587 : };
2588 : COMPUTE_COMMANDS_COUNTERS.for_command(command_kind).inc();
2589 :
2590 : self.handle_pagerequests(pgb, tenant_id, timeline_id, protocol_version, ctx)
2591 : .await?;
2592 : }
2593 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2594 : tenant_id,
2595 : timeline_id,
2596 : lsn,
2597 : gzip,
2598 : replica,
2599 : }) => {
2600 : tracing::Span::current()
2601 : .record("tenant_id", field::display(tenant_id))
2602 : .record("timeline_id", field::display(timeline_id));
2603 :
2604 : self.check_permission(Some(tenant_id))?;
2605 :
2606 : COMPUTE_COMMANDS_COUNTERS
2607 : .for_command(ComputeCommandKind::Basebackup)
2608 : .inc();
2609 : let metric_recording = metrics::BASEBACKUP_QUERY_TIME.start_recording();
2610 0 : let res = async {
2611 0 : self.handle_basebackup_request(
2612 0 : pgb,
2613 0 : tenant_id,
2614 0 : timeline_id,
2615 0 : lsn,
2616 0 : None,
2617 0 : false,
2618 0 : gzip,
2619 0 : replica,
2620 0 : &ctx,
2621 0 : )
2622 0 : .await?;
2623 0 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
2624 0 : Result::<(), QueryError>::Ok(())
2625 0 : }
2626 : .await;
2627 : metric_recording.observe(&res);
2628 : res?;
2629 : }
2630 : // same as basebackup, but result includes relational data as well
2631 : PageServiceCmd::FullBackup(FullBackupCmd {
2632 : tenant_id,
2633 : timeline_id,
2634 : lsn,
2635 : prev_lsn,
2636 : }) => {
2637 : tracing::Span::current()
2638 : .record("tenant_id", field::display(tenant_id))
2639 : .record("timeline_id", field::display(timeline_id));
2640 :
2641 : self.check_permission(Some(tenant_id))?;
2642 :
2643 : COMPUTE_COMMANDS_COUNTERS
2644 : .for_command(ComputeCommandKind::Fullbackup)
2645 : .inc();
2646 :
2647 : // Check that the timeline exists
2648 : self.handle_basebackup_request(
2649 : pgb,
2650 : tenant_id,
2651 : timeline_id,
2652 : lsn,
2653 : prev_lsn,
2654 : true,
2655 : false,
2656 : false,
2657 : &ctx,
2658 : )
2659 : .await?;
2660 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
2661 : }
2662 : PageServiceCmd::Set => {
2663 : // important because psycopg2 executes "SET datestyle TO 'ISO'"
2664 : // on connect
2665 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
2666 : }
2667 : PageServiceCmd::LeaseLsn(LeaseLsnCmd {
2668 : tenant_shard_id,
2669 : timeline_id,
2670 : lsn,
2671 : }) => {
2672 : tracing::Span::current()
2673 : .record("tenant_id", field::display(tenant_shard_id))
2674 : .record("timeline_id", field::display(timeline_id));
2675 :
2676 : self.check_permission(Some(tenant_shard_id.tenant_id))?;
2677 :
2678 : COMPUTE_COMMANDS_COUNTERS
2679 : .for_command(ComputeCommandKind::LeaseLsn)
2680 : .inc();
2681 :
2682 : match self
2683 : .handle_make_lsn_lease(pgb, tenant_shard_id, timeline_id, lsn, &ctx)
2684 : .await
2685 : {
2686 : Ok(()) => {
2687 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?
2688 : }
2689 : Err(e) => {
2690 : error!("error obtaining lsn lease for {lsn}: {e:?}");
2691 : pgb.write_message_noflush(&BeMessage::ErrorResponse(
2692 : &e.to_string(),
2693 : Some(e.pg_error_code()),
2694 : ))?
2695 : }
2696 : };
2697 : }
2698 : }
2699 :
2700 : Ok(())
2701 : }
2702 : }
2703 :
2704 : impl From<GetActiveTenantError> for QueryError {
2705 0 : fn from(e: GetActiveTenantError) -> Self {
2706 0 : match e {
2707 0 : GetActiveTenantError::WaitForActiveTimeout { .. } => QueryError::Disconnected(
2708 0 : ConnectionError::Io(io::Error::new(io::ErrorKind::TimedOut, e.to_string())),
2709 0 : ),
2710 : GetActiveTenantError::Cancelled
2711 : | GetActiveTenantError::WillNotBecomeActive(TenantState::Stopping { .. }) => {
2712 0 : QueryError::Shutdown
2713 : }
2714 0 : e @ GetActiveTenantError::NotFound(_) => QueryError::NotFound(format!("{e}").into()),
2715 0 : e => QueryError::Other(anyhow::anyhow!(e)),
2716 : }
2717 0 : }
2718 : }
2719 :
2720 : #[derive(Debug, thiserror::Error)]
2721 : pub(crate) enum GetActiveTimelineError {
2722 : #[error(transparent)]
2723 : Tenant(GetActiveTenantError),
2724 : #[error(transparent)]
2725 : Timeline(#[from] GetTimelineError),
2726 : }
2727 :
2728 : impl From<GetActiveTimelineError> for QueryError {
2729 0 : fn from(e: GetActiveTimelineError) -> Self {
2730 0 : match e {
2731 0 : GetActiveTimelineError::Tenant(GetActiveTenantError::Cancelled) => QueryError::Shutdown,
2732 0 : GetActiveTimelineError::Tenant(e) => e.into(),
2733 0 : GetActiveTimelineError::Timeline(e) => QueryError::NotFound(format!("{e}").into()),
2734 : }
2735 0 : }
2736 : }
2737 :
2738 : impl From<crate::tenant::timeline::handle::HandleUpgradeError> for QueryError {
2739 0 : fn from(e: crate::tenant::timeline::handle::HandleUpgradeError) -> Self {
2740 0 : match e {
2741 0 : crate::tenant::timeline::handle::HandleUpgradeError::ShutDown => QueryError::Shutdown,
2742 0 : }
2743 0 : }
2744 : }
2745 :
2746 0 : fn set_tracing_field_shard_id(timeline: &Timeline) {
2747 0 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
2748 0 : tracing::Span::current().record(
2749 0 : "shard_id",
2750 0 : tracing::field::display(timeline.tenant_shard_id.shard_slug()),
2751 0 : );
2752 0 : debug_assert_current_span_has_tenant_and_timeline_id();
2753 0 : }
2754 :
2755 : struct WaitedForLsn(Lsn);
2756 : impl From<WaitedForLsn> for Lsn {
2757 0 : fn from(WaitedForLsn(lsn): WaitedForLsn) -> Self {
2758 0 : lsn
2759 0 : }
2760 : }
2761 :
2762 : #[cfg(test)]
2763 : mod tests {
2764 : use utils::shard::ShardCount;
2765 :
2766 : use super::*;
2767 :
2768 : #[test]
2769 4 : fn pageservice_cmd_parse() {
2770 4 : let tenant_id = TenantId::generate();
2771 4 : let timeline_id = TimelineId::generate();
2772 4 : let cmd =
2773 4 : PageServiceCmd::parse(&format!("pagestream_v2 {tenant_id} {timeline_id}")).unwrap();
2774 4 : assert_eq!(
2775 4 : cmd,
2776 4 : PageServiceCmd::PageStream(PageStreamCmd {
2777 4 : tenant_id,
2778 4 : timeline_id,
2779 4 : protocol_version: PagestreamProtocolVersion::V2,
2780 4 : })
2781 4 : );
2782 4 : let cmd = PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id}")).unwrap();
2783 4 : assert_eq!(
2784 4 : cmd,
2785 4 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2786 4 : tenant_id,
2787 4 : timeline_id,
2788 4 : lsn: None,
2789 4 : gzip: false,
2790 4 : replica: false
2791 4 : })
2792 4 : );
2793 4 : let cmd =
2794 4 : PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id} --gzip")).unwrap();
2795 4 : assert_eq!(
2796 4 : cmd,
2797 4 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2798 4 : tenant_id,
2799 4 : timeline_id,
2800 4 : lsn: None,
2801 4 : gzip: true,
2802 4 : replica: false
2803 4 : })
2804 4 : );
2805 4 : let cmd =
2806 4 : PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id} latest")).unwrap();
2807 4 : assert_eq!(
2808 4 : cmd,
2809 4 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2810 4 : tenant_id,
2811 4 : timeline_id,
2812 4 : lsn: None,
2813 4 : gzip: false,
2814 4 : replica: false
2815 4 : })
2816 4 : );
2817 4 : let cmd = PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id} 0/16ABCDE"))
2818 4 : .unwrap();
2819 4 : assert_eq!(
2820 4 : cmd,
2821 4 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2822 4 : tenant_id,
2823 4 : timeline_id,
2824 4 : lsn: Some(Lsn::from_str("0/16ABCDE").unwrap()),
2825 4 : gzip: false,
2826 4 : replica: false
2827 4 : })
2828 4 : );
2829 4 : let cmd = PageServiceCmd::parse(&format!(
2830 4 : "basebackup {tenant_id} {timeline_id} --replica --gzip"
2831 4 : ))
2832 4 : .unwrap();
2833 4 : assert_eq!(
2834 4 : cmd,
2835 4 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2836 4 : tenant_id,
2837 4 : timeline_id,
2838 4 : lsn: None,
2839 4 : gzip: true,
2840 4 : replica: true
2841 4 : })
2842 4 : );
2843 4 : let cmd = PageServiceCmd::parse(&format!(
2844 4 : "basebackup {tenant_id} {timeline_id} 0/16ABCDE --replica --gzip"
2845 4 : ))
2846 4 : .unwrap();
2847 4 : assert_eq!(
2848 4 : cmd,
2849 4 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2850 4 : tenant_id,
2851 4 : timeline_id,
2852 4 : lsn: Some(Lsn::from_str("0/16ABCDE").unwrap()),
2853 4 : gzip: true,
2854 4 : replica: true
2855 4 : })
2856 4 : );
2857 4 : let cmd = PageServiceCmd::parse(&format!("fullbackup {tenant_id} {timeline_id}")).unwrap();
2858 4 : assert_eq!(
2859 4 : cmd,
2860 4 : PageServiceCmd::FullBackup(FullBackupCmd {
2861 4 : tenant_id,
2862 4 : timeline_id,
2863 4 : lsn: None,
2864 4 : prev_lsn: None
2865 4 : })
2866 4 : );
2867 4 : let cmd = PageServiceCmd::parse(&format!(
2868 4 : "fullbackup {tenant_id} {timeline_id} 0/16ABCDE 0/16ABCDF"
2869 4 : ))
2870 4 : .unwrap();
2871 4 : assert_eq!(
2872 4 : cmd,
2873 4 : PageServiceCmd::FullBackup(FullBackupCmd {
2874 4 : tenant_id,
2875 4 : timeline_id,
2876 4 : lsn: Some(Lsn::from_str("0/16ABCDE").unwrap()),
2877 4 : prev_lsn: Some(Lsn::from_str("0/16ABCDF").unwrap()),
2878 4 : })
2879 4 : );
2880 4 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2881 4 : let cmd = PageServiceCmd::parse(&format!(
2882 4 : "lease lsn {tenant_shard_id} {timeline_id} 0/16ABCDE"
2883 4 : ))
2884 4 : .unwrap();
2885 4 : assert_eq!(
2886 4 : cmd,
2887 4 : PageServiceCmd::LeaseLsn(LeaseLsnCmd {
2888 4 : tenant_shard_id,
2889 4 : timeline_id,
2890 4 : lsn: Lsn::from_str("0/16ABCDE").unwrap(),
2891 4 : })
2892 4 : );
2893 4 : let tenant_shard_id = TenantShardId::split(&tenant_shard_id, ShardCount(8))[1];
2894 4 : let cmd = PageServiceCmd::parse(&format!(
2895 4 : "lease lsn {tenant_shard_id} {timeline_id} 0/16ABCDE"
2896 4 : ))
2897 4 : .unwrap();
2898 4 : assert_eq!(
2899 4 : cmd,
2900 4 : PageServiceCmd::LeaseLsn(LeaseLsnCmd {
2901 4 : tenant_shard_id,
2902 4 : timeline_id,
2903 4 : lsn: Lsn::from_str("0/16ABCDE").unwrap(),
2904 4 : })
2905 4 : );
2906 4 : let cmd = PageServiceCmd::parse("set a = b").unwrap();
2907 4 : assert_eq!(cmd, PageServiceCmd::Set);
2908 4 : let cmd = PageServiceCmd::parse("SET foo").unwrap();
2909 4 : assert_eq!(cmd, PageServiceCmd::Set);
2910 4 : }
2911 :
2912 : #[test]
2913 4 : fn pageservice_cmd_err_handling() {
2914 4 : let tenant_id = TenantId::generate();
2915 4 : let timeline_id = TimelineId::generate();
2916 4 : let cmd = PageServiceCmd::parse("unknown_command");
2917 4 : assert!(cmd.is_err());
2918 4 : let cmd = PageServiceCmd::parse("pagestream_v2");
2919 4 : assert!(cmd.is_err());
2920 4 : let cmd = PageServiceCmd::parse(&format!("pagestream_v2 {tenant_id}xxx"));
2921 4 : assert!(cmd.is_err());
2922 4 : let cmd = PageServiceCmd::parse(&format!("pagestream_v2 {tenant_id}xxx {timeline_id}xxx"));
2923 4 : assert!(cmd.is_err());
2924 4 : let cmd = PageServiceCmd::parse(&format!(
2925 4 : "basebackup {tenant_id} {timeline_id} --gzip --gzip"
2926 4 : ));
2927 4 : assert!(cmd.is_err());
2928 4 : let cmd = PageServiceCmd::parse(&format!(
2929 4 : "basebackup {tenant_id} {timeline_id} --gzip --unknown"
2930 4 : ));
2931 4 : assert!(cmd.is_err());
2932 4 : let cmd = PageServiceCmd::parse(&format!(
2933 4 : "basebackup {tenant_id} {timeline_id} --gzip 0/16ABCDE"
2934 4 : ));
2935 4 : assert!(cmd.is_err());
2936 4 : let cmd = PageServiceCmd::parse(&format!("lease {tenant_id} {timeline_id} gzip 0/16ABCDE"));
2937 4 : assert!(cmd.is_err());
2938 4 : }
2939 : }
|
